Bierens - Statistical Foundations of Econometrics

1
Introduction to the Mathematical and Statistical Foundations of Econometrics
Herman J. Bierens Pennsylvania State University, USA, and Tilburg University, the Netherlands
Contents
Preface
Chapter 1: Probability and Measure 1.1. The Texas lotto 1.1.1 Introduction 1.1.2 Binomial numbers 1.1.3 Sample space 1.1.4 Algebras and sigma-algebras of events 1.1.5 Probability measure 1.2. Quality control 1.2.1 Sampling without replacement 1.2.2 Quality control in practice 1.2.3 Sampling with replacement 1.2.4 Limits of the hypergeometric and binomial probabilities 1.3. 1.4. Why do we need sigma-algebras of events? Properties of algebras and sigma-algebras 1.4.1 General properties 1.4.2 Borel sets 1.5. 1.6. Properties of probability measures The uniform probability measure 1.6.1 Introduction 1.6.2 Outer measure 1.7. Lebesgue measure and Lebesgue integral 1.7.1 Lebesgue measure 1.7.2 Lebesgue integral 1.8. Random variables and their distributions
3 1.8.1 Random variables and vectors 1.8.2 Distribution functions 1.9. 1.10. Density functions Conditional probability, Bayes' rule, and independence 1.10.1 Conditional probability 1.10.2 Bayes' rule 1.10.3 Independence 1.11. Exercises
Appendices: 1.A. 1.B. Common structure of the proofs of Theorems 6 and 10 Extension of an outer measure to a probability measure
Chapter 2: Borel Measurability, Integration, and Mathematical Expectations 2.1. 2.2. 2.3. 2.4. Introduction Borel measurability Integrals of Borel measurable functions with respect to a probability measure General measurability, and integrals of random variables with respect to probability measures 2.5. 2.6. Mathematical expectation Some useful inequalities involving mathematical expectations 2.6.1 Chebishev's inequality 2.6.2 Holder's inequality 2.6.3 Liapounov's inequality 2.6.4 Minkowski's inequality 2.6.5 Jensen's inequality 2.7. 2.8. Expectations of products of independent random variables Moment generating functions and characteristic functions 2.8.1 Moment generating functions
4 2.8.2 Characteristic functions 2.9. Exercises
Appendix: 2.A. Uniqueness of characteristic functions
Chapter 3: Conditional Expectations 3.1. 3.2. 3.3. 3.4. 3.5. 3.6. Introduction Properties of conditional expectations Conditional probability measures and conditional independence Conditioning on increasing sigma-algebras Conditional expectations as the best forecast schemes Exercises
Appendix: 3.A. Proof of Theorem 3.12
Chapter 4: Distributions and Transformations 4.1. Discrete distributions 4.1.1 The hypergeometric distribution 4.1.2 The binomial distribution 4.1.3 The Poisson distribution 4.1.4 The negative binomial distribution 4.2. 4.3. 4.4. Transformations of discrete random vectors Transformations of absolutely continuous random variables Transformations of absolutely continuous random vectors 4.4.1 The linear case 4.4.2 The nonlinear case 4.5. The normal distribution
5 4.5.1 The standard normal distribution 4.5.2 The general normal distribution 4.6. Distributions related to the normal distribution 4.6.1 The chi-square distribution 4.6.2 The Student t distribution 4.6.3 The standard Cauchy distribution 4.6.4 The F distribution 4.7. 4.8. 4.9. The uniform distribution and its relation to the standard normal distribution The gamma distribution Exercises
Appendices: 4.A: 4.B: Tedious derivations Proof of Theorem 4.4
Chapter 5: The Multivariate Normal Distribution and its Application to Statistical Inference 5.1. 5.2. 5.3. 5.4. Expectation and variance of random vectors The multivariate normal distribution Conditional distributions of multivariate normal random variables Independence of linear and quadratic transformations of multivariate normal random variables 5.5. 5.6. Distribution of quadratic forms of multivariate normal random variables Applications to statistical inference under normality 5.6.1 Estimation 5.6.2 Confidence intervals 5.6.3 Testing parameter hypotheses 5.7. Applications to regression analysis 5.7.1 The linear regression model 5.7.2 Least squares estimation
6 5.7.3 Hypotheses testing 5.8. Exercises
Appendix: 5.A. Proof of Theorem 5.8
Chapter 6: Modes of Convergence 6.1. 6.2. 6.3. 6.4. Introduction Convergence in probability and the weak law of large numbers Almost sure convergence, and the strong law of large numbers The uniform law of large numbers and its applications 6.4.1 The uniform weak law of large numbers 6.4.2 Applications of the uniform weak law of large numbers 6.4.2.1 Consistency of M-estimators 6.4.2.2 Generalized Slutsky's theorem 6.4.3 The uniform strong law of large numbers and its applications 6.5. 6.6. 6.7. 6.8. 6.9. 6.10. 6.11. Convergence in distribution Convergence of characteristic functions The central limit theorem Stochastic boundedness, tightness, and the Op and op notations Asymptotic normality of M-estimators Hypotheses testing Exercises
Appendices: 6.A. 6.B. 6.C. Proof of the uniform weak law of large numbers Almost sure convergence and strong laws of large numbers Convergence of characteristic functions and distributions
7 Chapter 7: Dependent Laws of Large Numbers and Central Limit Theorems 7.1. 7.2. 7.3. 7.4. Stationarity and the Wold decomposition Weak laws of large numbers for stationary processes Mixing conditions Uniform weak laws of large numbers 7.4.1 Random functions depending on finite-dimensional random vectors 7.4.2 Random functions depending on infinite-dimensional random vectors 7.4.3 Consistency of M-estimators 7.5. Dependent central limit theorems 7.5.1 Introduction 7.5.2 A generic central limit theorem 7.5.3 Martingale difference central limit theorems 7.6. Exercises
Appendix: 7.A. Hilbert spaces
Chapter 8: Maximum Likelihood Theory 8.1. 8.2. 8.3. Introduction Likelihood functions Examples 8.3.1 The uniform distribution 8.3.2 Linear regression with normal errors 8.3.3 Probit and Logit models 8.3.4 The Tobit model 8.4. Asymptotic properties of ML estimators 8.4.1 Introduction 8.4.2 First and second-order conditions
8 8.4.3 Generic conditions for consistency and asymptotic normality 8.4.4 Asymptotic normality in the time series case 8.4.5 Asymptotic efficiency of the ML estimator 8.5. Testing parameter restrictions 8.5.1 The pseudo t test and the Wald test 8.5.2 The Likelihood Ratio test 8.5.3 The Lagrange Multiplier test 8.5.4 Which test to use? 8.6. Exercises
Appendix I: Review of Linear Algebra I.1. I.2. I.3. I.4. I.5. I.6. Vectors in a Euclidean space Vector spaces Matrices The inverse and transpose of a matrix Elementary matrices and permutation matrices Gaussian elimination of a square matrix, and the Gauss-Jordan iteration for inverting a matrix I.6.1 Gaussian elimination of a square matrix I.6.2 The Gauss-Jordan iteration for inverting a matrix I.7. I.8. I.9. I.10. I.11. I.12. I.13. I.14. Gaussian elimination of a non-square matrix Subspaces spanned by the columns and rows of a matrix Projections, projection matrices, and idempotent matrices Inner product, orthogonal bases, and orthogonal matrices Determinants: Geometric interpretation and basic properties Determinants of block-triangular matrices Determinants and co-factors Inverse of a matrix in terms of co-factors
9 I.15. Eigenvalues and eigenvectors I.15.1 Eigenvalues I.15.2 Eigenvectors I.15.3 Eigenvalues and eigenvectors of symmetric matrices I.16. I.17. I.18 Positive definite and semi-definite matrices Generalized eigenvalues and eigenvectors Exercises
Appendix II: Miscellaneous Mathematics II.1. Sets and set operations II.1.1 General set operations II.1.2 Sets in Euclidean spaces II.2. II.3. II.4. II.5. II.6. II.7. II.8. II.9. Supremum and infimum Limsup and liminf Continuity of concave and convex functions Compactness Uniform continuity Derivatives of functions of vectors and matrices The mean value theorem Taylor's theorem
II.10. Optimization
Appendix III: A Brief Review of Complex Analysis III.1. III.2. III.3. III.4. The complex number system The complex exponential function The complex logarithm Series expansion of the complex logarithm
10 III.5. Complex integration
References
11
Preface
This book is intended for a rigorous introductory Ph.D. level course in econometrics, or for use in a field course in econometric theory. It is based on lecture notes that I have developed during the period 1997-2003 for the first semester econometrics course Introduction to Econometrics in the core of the Ph.D. program in economics at the Pennsylvania State University. Initially these lecture notes were written as a companion to Gallants (1997) textbook, but have been developed gradually into an alternative textbook. Therefore, the topics that are covered in this book encompass those in Gallants book, but in much more depth. Moreover, in order to make the book also suitable for a field course in econometric theory I have included various advanced topics as well. I used to teach this advanced material in the econometrics field at the Free University of Amsterdam and Southern Methodist University, on the basis of the draft of my previous textbook, Bierens (1994). Some chapters have their own appendices, containing the more advanced topics and/or difficult proofs. Moreover, there are three appendices with material that is supposed to be known, but often is not, or not sufficiently. Appendix I contains a comprehensive review of linear algebra, including all the proofs. This appendix is intended for self-study only, but may serve well in a half-semester or one quarter course in linear algebra. Appendix II reviews a variety of mathematical topics and concepts that are used throughout the main text, and Appendix III reviews the basics of complex analysis which is needed to understand and derive the properties of characteristic functions. At the beginning of the first class I always tell my students: Never ask me how. Only ask me why. In other words, dont be satisfied with recipes. Of course, this applies to other
12 economics fields as well, in particular if the mission of the Ph.D. program is to place its graduates at research universities. First, modern economics is highly mathematical. Therefore, in order to be able to make original contributions to economic theory Ph.D. students need to develop a mathematical mind. Second, students who are going to work in an applied econometrics field like empirical IO or labor need to be able to read the theoretical econometrics literature in order to keep up-to-date with the latest econometric techniques. Needless to say, students interested in contributing to econometric theory need to become professional mathematicians and statisticians first. Therefore, in this book I focus on teaching why, by providing proofs, or at least motivations if proofs are too complicated, of the mathematical and statistical results necessary for understanding modern econometric theory. Probability theory is a branch of measure theory. Therefore, probability theory is introduced, in Chapter 1, in a measure-theoretical way. The same applies to unconditional and conditional expectations in Chapters 2 and 3, which are introduced as integrals with respect to probability measures. These chapters are also beneficial as preparation for the study of economic theory, in particular modern macroeconomic theory. See for example Stokey, Lucas, and Prescott (1989). It usually takes me three weeks (at a schedule of two lectures of one hour and fifteen minutes per week) to get through Chapter 1, skipping all the appendices. Chapters 2 and 3 together, without the appendices, usually take me about three weeks as well. Chapter 4 deals with transformations of random variables and vectors, and also lists the most important univariate continuous distributions, together with their expectations, variances, moment generating functions (if they exist), and characteristic functions. I usually explain only
13 the change-of variables formula for (joint) densities, leaving the rest of Chapter 4 for self-tuition. The multivariate normal distribution is treated in detail in Chapter 5, far beyond the level found in other econometrics textbooks. Statistical inference, i.e., estimation and hypotheses testing, is also introduced in Chapter 5, in the framework of the normal linear regression model. At this point it is assumed that the students have a thorough understanding of linear algebra. This assumption, however, is often more fiction than fact. To tests this hypothesis, and to force the students to refresh their linear algebra, I usually assign all the exercises in Appendix I as homework before starting with Chapter 5. It takes me about three weeks to get through this chapter. Asymptotic theory for independent random variables and vectors, in particular the weak and strong laws of large numbers and the central limit theorem, is discussed in Chapter 6, together with various related convergence results. Moreover, the results in this chapter are applied to M-estimators, including nonlinear regression estimators, as an introduction to asymptotic inference. However, I have never been able to get beyond Chapter 6 in one semester, even after skipping all the appendices and Sections 6.4 and 6.9 which deals with asymptotic inference. Chapter 7 extends the weak law of large numbers and the central limit theorem to stationary time series processes, starting from the Wold (1938) decomposition. In particular, the martingale difference central limit theorem of McLeish (1974) is reviewed together with preliminary results. Maximum likelihood theory is treated in Chapter 8. This chapter is different from the standard treatment of maximum likelihood theory in that special attention is paid to the problem
14 of how to setup the likelihood function in the case that the distribution of the data is neither absolutely continuous nor discrete. In this chapter only a few references to the results in Chapter 7 are made, in particular in Section 8.4.4. Therefore, Chapter 7 is not prerequisite for Chapter 8, provided that the asymptotic inference parts of Chapter 6 (Sections 6.4 and 6.9) have been covered. Finally, the helpful comments of five referees on the draft of this book, and the comments of my colleague Joris Pinkse on Chapter 8, are gratefully acknowledged. My students have pointed out many typos in earlier drafts, and their queries have led to substantial improvements of the exposition. Of course, only I am responsible for any remaining errors.
15
Chapter 1 Probability and Measure
1.1. 1.1.1
The Texas lotto Introduction Texans (used to) play the lotto by selecting six different numbers between 1 and 50,
which cost $1 for each combination1. Twice a week, on Wednesday and Saturday at 10:00 P.M., six ping-pong balls are released without replacement from a rotating plastic ball containing 50 ping-pong balls numbered 1 through 50. The winner of the jackpot (which occasionally accumulates to 60 or more million dollars!) is the one who has all six drawn numbers correct, where the order in which the numbers are drawn does not matter. What are the odds of winning if you play one set of six numbers only? In order to answer this question, suppose first that the order of the numbers does matter. Then the number of ordered sets of 6 out of 50 numbers is: 50 possibilities for the first drawn number, times 49 possibilities for the second drawn number, times 48 possibilities for the third drawn number, times 47 possibilities for the fourth drawn number, times 46 possibilities for the fifth drawn number, times 45 possibilities for the sixth drawn number: kk
50 k'1 50&6 k'1
k (50 & j) ' k k '

5 50 j'0 k'45
k k
'
50! . (50 & 6)!
The notation n!, read: n factorial, stands for the product of the natural numbers 1 through n:
16 n! ' 12.......(n&1)n if n > 0, 0! ' 1 . The reason for defining 0! = 1 will be explained below. Since a set of six given numbers can be permutated in 6! ways, we need to correct the above number for the 6! replications of each unordered set of six given numbers. Therefore, the number of sets of six unordered numbers out of 50 is: 50 6
def.
'
50! ' 15,890,700. 6!(50&6)!
Thus, the probability of winning the Texas lotto if you play only one combination of six numbers is 1/15,890,700. 2
1.1.2
Binomial numbers In general, the number of ways we can draw a set of k unordered objects out of a set of n
objects without replacement is: n k

def.
'
n! . k!(n&k)!
(1.1)
These (binomial) numbers3, read as: n choose k, also appear as coefficients in the binomial expansion (a % b)n ' j
n k'0
n k
a kb n&k .
(1.2)
The reason for defining 0! = 1 is now that the first and last coefficients in this binomial expansion are always equal to 1:
17 n 0 ' n n ' n! 1 ' ' 1. 0!n! 0!
For not too large an n the binomial numbers (1.1) can be computed recursively by hand, using the Triangle of Pascal: 1 1 1 1 1 1 1 5 4 10 3 6 10 2 3 4 5 1 1 1 1 1 1 (1.3)
Except for the 1's on the legs and top of the triangle in (1.3), the entries are the sums of the adjacent numbers on the previous line, which is due to the easy equality: n&1 k&1 n&1 k n k
'
for n $ 2 , k ' 1,....,n&1 .
(1.4)
Thus, the top 1 corresponds to n = 0, the second row corresponds to n = 1, the third row corresponds to n = 2, etc., and for each row n+1, the entries are the binomial numbers (1.1) for k = 0,....,n. For example, for n = 4 the coefficients of a kb n&k in the binomial expansion (1.2) can be found on row 5 in (1.3): (a % b)4 ' 1a 4 % 4a 3b % 6a 2b 2 % 4ab 3 % 1b 4 .
18 1.1.3 Sample space The Texas lotto is an example of a statistical experiment. The set of possible outcomes of this statistical experiment is called the sample space, and is usually denoted by . In the Texas lotto case contains N = 15,890,700 elements: ' {1 ,.....,N} , where each element j is a set itself consisting of six different numbers ranging from 1 to 50, such that for any pair i , condition i j for i j is equivalent to the condition that i _ j . j with i j , i j . Since in this case the elements j of are sets themselves, the
1.1.4
Algebras and sigma-algebras of events A set { j ,...., j } of different number combinations you can bet on is called an event.
1 k
The collection of all these events, denoted by , is a family of subsets of the sample space . In the Texas lotto case the collection consists of all subsets of , including itself and the empty set i .4 In principle you could bet on all number combinations if you are rich enough (it will cost you $15,890,700). Therefore, the sample space itself is included in . You could also decide not to play at all. This event can be identified as the empty set i . For the sake of completeness it is included in as well. Since in the Texas lotto case the collection contains all subsets of , it automatically satisfies the following conditions: ' \A 0 , If A 0 then A (1.5)
' \A is the complement of the set A (relative to the set ), i.e., the set of all elements where A of that are not contained in A; If A , B 0 then A^B 0 . (1.6)
19 By induction, the latter condition extends to any finite union of sets in : If Aj 0 for j = 1,2,...,n, then ^j'1Aj 0 .
n
Definition 1.1: A collection of subsets of a non-empty set satisfying the conditions (1.5) and (1.6) is called an algebra.5
In the Texas lotto example the sample space is finite, and therefore the collection of subsets of is finite as well. Consequently, in this case the condition (1.6) extends to: If Aj 0 for j ' 1,2,.... then ^j'1Aj 0 .
4
(1.7)
However, since in this case the collection of subsets of is finite, there are only a finite number of distinct sets Aj 0 . Therefore, in the Texas lotto case the countable infinite union
4
^j'1Aj in (1.7) involves only a finite number of distinct sets Aj; the other sets are replications of
these distinct sets. Thus, condition (1.7) does not require that all the sets Aj 0 are different.
Definition 1.2: A collection of subsets of a non-empty set satisfying the conditions (1.5) and (1.7) is called a & algebra.6
1.1.5
Probability measure Now let us return to the Texas lotto example. The odds, or probability, of winning is 1/N
for each valid combination j of six numbers, hence if you play n different valid number combinations {j , ...,j } , the probability of winning is n/N: P({j , ...,j }) ' n/N . Thus, in
1 n 1 n
the Texas lotto case the probability P(A) , A 0 , is given by the number n of elements in the
20 set A, divided by the total number N of elements in . In particular we have P() ' 1, and if you do not play at all the probability of winning is zero: P(i) ' 0 . The function P(A) , A 0 , is called a probability measure: it assigns a number P(A) 0 [0,1] to each set A 0 . Not every function which assigns numbers in [0,1] to the sets in is a probability measure, though:
Definition 1.3: A mapping P: 6 [0,1] from a & algebra of subsets of a set into the unit interval is a probability measure on { , } if it satisfies the following three conditions: If A 0 then P(A) $ 0 , P() ' 1 , For disjoint sets Aj 0 , P(^j'1 Aj) ' 'j'1P(Aj) .
4 4
(1.8) (1.9) (1.10)
Recall that sets are disjoint if they have no elements in common: their intersections are the empty set. The conditions (1.8) and (1.9) are clearly satisfied for the case of the Texas lotto. On the other hand, in the case under review the collection of events contains only a finite number of sets, so that any countably infinite sequence of sets in must contain sets that are the same. At first sight this seems to conflict with the implicit assumption that there always exist countably infinite sequences of disjoint sets for which (1.10) holds. It is true indeed that any countably infinite sequence of disjoint sets in a finite collection of sets can only contain a finite number of non-empty sets. This is no problem though, because all the other sets are then equal
to the empty set i . The empty set is disjoint with itself: i _ i ' i , and with any other set: A _ i ' i . Therefore, if is finite then any countable infinite sequence of disjoint sets consists of a finite number of non-empty sets, and an infinite number of replications of the empty set. Consequently, if is finite then it is sufficient for the verification of condition
21
Since in the Texas lotto case P(A1Â2) ' (n1%n2)/N , P(A1) ' n1/N , and P(A2) ' n2/N , where n1 is the number of elements of A1 and n2 is the number of elements of A2 , the latter condition is satisfied, and so is condition (1.10). The statistical experiment is now completely described by the triple { , , P} , called the probability space, consisting of the sample space , i.e., the set of all possible outcomes of the statistical experiment involved, a & algebra of events, i.e., a collection of subsets of the sample space such that the conditions (1.5) and (1.7) are satisfied, and a probability measure P: 6 [0,1] satisfying the conditions (1.8), (1.9), and (1.10). In the Texas lotto case the collection of events is an algebra, but because is finite it is automatically a & algebra.
(1.10) to verify that for any pair of disjoint sets A1 , A2 in , P(A1Â2) = P(A1) + P(A2) .
1.2. 1.2.1
Quality control Sampling without replacement As a second example, consider the following case. Suppose you are in charge of quality
control in a light bulb factory. Each day N light bulbs are produced. But before they are shipped out to the retailers, the bulbs need to meet a minimum quality standard, say: no more than R out of N bulbs are allowed to be defective. The only way to verify this exactly is to try all the N
22 bulbs out, but that will be too costly. Therefore, the way quality control is conducted in practice is to draw randomly n bulbs without replacement, and to check how many bulbs in this sample are defective. Similarly to the Texas lotto case, the number M of different samples sj of size n you can draw out of a set of N elements without replacement is: M ' N n .
Each sample sj is characterized by a number kj of defective bulbs in the sample involved. Let K be the actual number of defective bulbs. Then kj 0 {0,1,...,min(n,K)} . Let ' {0,1,....,n}, and let the & algebra be the collection of all subsets of . The number of samples sj with kj = k # min(n,K) defective bulbs is: K k N&K n&k ,
because there are K choose k ways to draw k unordered numbers out of K numbers without replacement, and N-K choose n-k ways to draw n - k unordered numbers out of N - K numbers without replacement. Of course, in the case that n > K the number of samples sj with kj = k > min (n,K) defective bulbs is zero. Therefore, let: K P({k}) ' k N &K n&k N n
m
if 0 # k # min(n,K) , P({k}) ' 0 elsewhere ,
(1.11)
and let for each set A ' {k1 , ...... , km} 0 , P(A) ' 'j'1P({kj}) . (Exercise: Verify that this function P satisfies all the requirements of a probability measure.) The triple { , , P} is now
23 the probability space corresponding to this statistical experiment. The probabilities (1.11) are known as the Hypergeometric(N,K,n) probabilities.
1.2.2
Quality control in practice7 The problem in applying this result in quality control is that K is unknown. Therefore, in
practice the following decision rule as to whether K # R or not is followed. Given a particular number r # n , to be determined below, assume that the set of N bulbs meets the minimum quality requirement K # R if the number k of defective bulbs in the sample is less or equal to r . Then the set A(r) ' {0,1,...,r} corresponds to the assumption that the set of N bulbs meets the minimum quality requirement K # R , hereafter indicated by accept, with probability P(A(r)) ' 'k'0P({k}) ' pr(n,K) ,
r
(1.12)
(r) ' {r%1,....,n} corresponds to the assumption that this set of say, whereas its complement A N bulbs does not meet this quality requirement, hereafter indicated by reject, with corresponding probability (r)) ' 1 & p (n,K) . P(A r Given r, this decision rule yields two types of errors, a type I error with probability 1 & pr(n,K) if you reject while in reality K # R , and a type II error with probability pr(K,n) if you accept while in reality K > R . The probability of a type I error has upper bound:
p1(r,n) ' 1 & min pr(n,K),

K#R
(1.13)
say, and the probability of a type II error has upper bound p2(r,n) ' max pr(n,K) ,
K >R
(1.14)
24 say. In order to be able to choose r, one has to restrict either p1(r,n) or p2(r,n) , or both. Usually it is former which is restricted, because a type I error may cause the whole stock of N bulbs to be trashed. Thus, allow the probability of a type I error to be maximal , say = 0.05. Then r should be chosen such that p1(r,n) # . Since p1(r,n) is decreasing in r because (1.12) is increasing in r, we could in principle choose r arbitrarily large. But since p2(r,n) is increasing in r, we should not choose r unnecessarily large. Therefore, choose r = r(n|), where r(n|) is the minimum value of r for which p1(r,n) # . Moreover, if we allow the type II error to be maximal , we have to choose the sample size n such that p2(r(n|),n) # . As we will see later, this decision rule is an example of a statistical test, where H0: K # R is called the null hypothesis to be tested at the 100% significance level, against the alternative hypothesis H1: K > R . The number r(n|) is called the critical value of the test, and the number k of defective bulbs in the sample is called the test statistic.
1.2.3
Sampling with replacement As a third example, consider the quality control example in the previous section, except
that now the light bulbs are sampled with replacement: After testing a bulb, it is put back in the stock of N bulbs, even if the bulb involved proves to be defective. The rationale for this behavior may be that the customers will accept maximally a fraction R/N of defective bulbs, so that they will not complain as long as the actual fraction K/N of defective bulbs does not exceed R/N. In other words, why not selling defective light bulbs if it is OK with the customers? The sample space and the & algebra are the same as in the case of sampling
25 without replacement, but the probability measure P is different. Consider again a sample sj of size n containing k defective light bulbs. Since the light bulbs are put back in the stock after being tested, there are K k ways of drawing an ordered set of k defective bulbs, and (N & K)n&k ways of drawing an ordered set of n-k working bulbs. Thus the number of ways we can draw, with replacement, an ordered set of n light bulbs containing k defective bulbs is K k(N & K)n&k . Moreover, similarly to the Texas lotto case it follows that the number of unordered sets of k defective bulbs and n-k working bulbs is: n choose k. Thus, the total number of ways we can choose a sample with replacement containing k defective bulbs and n-k working bulbs in any order is: n k K k(N & K)n&k .
Moreover, the number of ways we can choose a sample of size n with replacement is N n . Therefore, n K k(N & K)n&k n k p (1 & p)n&k , k ' 0,1,2,....,n, ' n k k N
P({k}) '
(1.15)
where p ' K/N , and again for each set A ' {k1 , ...... , km} 0 , P(A) ' 'j'1P({kj}) . Of course, replacing
m
P({k}) in (1.11) by (1.15) the argument in Section 1.2.2 still applies. The probabilities (1.15) are known as the Binomial(n,p) probabilities.
26 1.2.4 Limits of the hypergeometric and binomial probabilities Note that if N and K are large relative to n, the hypergeometric probability (1.11) and the binomial probability (1.15) will be almost the same. This follows from the fact that for fixed k and n: K P({k}) ' k N&K n&k N n K!(N&K)! k!(K&k)!(n&k)!(N&K&n%k)! ' N! n!(N&n)!
K!(N&K)! K! (N&K)! n n! (K&k)!(N&K&n%k)! (K&k)! (N&K&n%k)! ' ' k!(n&k)! N! N! k (N&n)! (N&n)! n k
k
'
(j'1(K&k%j) (j'1 (N&K&n%k%j)

k n&k
(j'1(N&n%j)
n
'
n k
(j'1
K k j K n k j n&k & % (j'1 1& & % % N N N N N N N n j n (j'1 1& % N N
n k
p k(1&p)n&k if N 6 4 and K/N 6 p .
Thus, the binomial probabilities also arise as limits of the hypergeometric probabilities. Moreover, if in the case of the binomial probability (1.15) p is very small and n is very large, the probability (1.15) can be approximated quite well by the Poisson() probability: P({k}) ' exp(&) k , k ' 0,1,2,.......... , k!
(1.16)
27 where ' np . This follows from (1.15) by choosing p ' /n for n > , with > 0 fixed, and letting n 6 4 while keeping k fixed: P({k}) ' ' n k n! p (1 & p)n&k ' /n k 1 & /n n&k k k!(n&k)!
n k
k n! 1 & /n k k! n (n&k)! 1 & /n
6 exp(&)
k for n 6 4 , k!
because n! n k(n&k)! (j'1(n&k%j)

k
'
nk
k j k ' (j'1 1& % n n

k
6 (j'11 ' 1 for n 6 4 ,

k
1 & /n and 1 & /n

n
6 1 for n 6 4 ,
6 exp(&) for n 6 4 .
(1.17)
Since (1.16) is the limit of (1.15) for p ' /n 9 0 as n 6 4 , the Poisson probabilities (1.16) are often used to model the occurrence of rare events. Note that the sample space corresponding to the Poisson probabilities is = {0,1,2,....}, and the & algebra of events involved can be chosen to be the collection of all subsets of , because any non-empty subset A of is either countable infinite or finite. If such a subset A is countable infinite, it takes the form A ' {k1 , k2 , k3 , ..........} , where the kjs are distinct
4
nonnegative integers, hence P(A) ' 'j'1P({kj}) is well-defined. The same applies of course if A is finite: if A = {k1 , .... , km} then P(A) ' 'j'1P({kj}) . This probability measure clearly
m
satisfies the conditions (1.8), (1.9), and (1.10).
28 1.3. Why do we need sigma-algebras of events? In principle we could define a probability measure on an algebra of subsets of the sample space, rather than on a !algebra. We only need to change condition (1.10) to: For disjoint sets Aj 0 such that ^j'1 Aj 0 , P(^j'1 Aj) ' 'j'1P(Aj) . By letting all but a finite
4 4 4
number of these sets are equal to the empty set, this condition then reads: For disjoint sets Aj 0 , j = 1,2,...,n < 4, P(^j'1 Aj) ' 'j'1P(Aj) . However, if we would confine a probability
n n
measure to an algebra all kind of useful results will no longer apply. One of these results is the so-called strong law of large numbers. See Chapter 6. As an example, consider the following game. Toss a fair coin infinitely many times, and assume that after each tossing you will get one dollar if the outcome it head, and nothing if the outcome is tail. The sample space in this case can be expressed in terms of the winnings, i.e., each element of takes the form of a string of infinitely many zeros and ones, for example = (1,1,0,1,0,1......). Now consider the event: After n tosses the winning is k dollars. This event corresponds to the set Ak,n of elements of for which the sum of the first n elements in the string involved is equal to k. For example, the set A1,2 consists of all of the type (1,0,......) and (0,1,......). Similarly to the example in Section 1.2.3 it can be shown that n (1/2)n for k ' 0,1,2,....,n, P(Ak,n) ' 0 for k > n or k < 0 . k
P(Ak,n) '
Next, for q = 1,2,.... consider the events: After n tosses the average winning k/n is contained in the interval [0.5!1/q, 0.5+1/q]. These events correspond to the sets Bq,n ' ^k'[n/2&n/q)]%1Ak,n ,
[n/2%n/q]
where [x] denotes the smallest integer $ x. Then the set _m'nBq,m corresponds to the event:
4
From the n-th tossing onwards the average winning will stay in the interval [0.5!1/q, 0.5+1/q],
and the set ^n'1_m'nBq,m corresponds to the event: There exists an n (possibly depending on )
4 4
29
such that from the n-th tossing onwards the average winning will stay in the interval [0.5!1/q, 0.5+1/q]. Finally, the set _q'1^n'1_m'nBq,m corresponds to the event: The average winning
4 4 4
converges to as n converges to infinity". Now the strong law of large numbers states that the latter event has probability 1: P[_q'1^n'1_m'nBq,m] = 1. However, this probability is only defined
4 4 4 4 4 4
if _q'1^n'1_m'nBq,m 0 . In order to guarantee this, we need to require that is a -algebra.
1.4. 1.4.1
Properties of algebras and sigma-algebras General properties In this section I will review the most important results regarding algebras, & algebras,
and probability measures. Our first result is trivial:
Theorem 1.1: If an algebra contains only a finite number of sets then it is a -algebra. Consequently, an algebra of subsets of a finite set is a & algebra.
However, an algebra of subsets of an infinite set is not necessarily a & algebra. A counter example is the collection ( of all subsets of = (0,1] of the type (a,b], where a < b are rational numbers in [0,1], together with their finite unions and the empty set i . Verify that ( is an algebra. Next, let pn = [10n ]/10n and an = 1/ pn, where [x] means truncation to the nearest integer # x . Note that pn 8 , hence an 9 &1 as n 6 4 . Then for n = 1,2,3,...., (an , 1] 0 ( , but ^n'1(an,1] ' (&1,1] ( because &1 is irrational. Thus, (
4
30 is not a & algebra.
Theorem 1.2: If is an algebra, then A ,B 0 implies A_B 0 , hence by induction, Aj 0 for j = 1,...,n < 4 imply _j'1Aj 0 . A collection of subsets of a nonempty set
n
is an algebra if it satisfies condition (1.5) and the condition that for any pair A ,B 0 , A_B 0 .
Proof: Exercise. Similarly, we have
Theorem 1.3: If is a & algebra, then for any countable sequence of sets Aj 0 ,
4
condition (1.5) and the condition that for any countable sequence of sets Aj 0 , _j'1Aj 0
4
_j'1Aj 0 . A collection of subsets of a nonempty set is a & algebra if it satisfies
These results will be convenient in cases where it is easier to prove that (countable) intersections are included in than to prove that (countable) unions are included If is already an algebra, then condition (1.7) alone would make it a & algebra. However, the condition in the following theorem is easier to verify than (1.7):
Theorem 1.4: If is an algebra and Aj, j =1,2,3,... is a countable sequence of sets in , then there exists a countable sequence of disjoint sets Bj in such that ^j'1Aj ' ^j'1Bj .
4 4
31 Consequently, an algebra is also a & algebra if for any sequence of disjoint sets Bj in , ^j'1Bj 0 .
4 n n Proof: Let Aj 0 . Denote B1 ' A1, Bn%1 ' An%1\(^j'1Aj) ' An%1_(_j'1A j) . It follows
from the properties of an algebra (see Theorem 1.2) that all the Bj s are sets in . Moreover, it is easy to verify that the Bjs are disjoint, and that ^j'1Aj ' ^j'1Bj . Thus, if ^j'1Bj 0 then
4 4 4
^j'1Aj 0 . Q.E.D.
4
where is a possibly uncountable index set. Then ' _0 is a & algebra.
Theorem 1.5: Let , 0 , be a collection of & algebras of subsets of a given set ,
Proof: Exercise. For example, let ' {(0,1] , i , (0,] , (,1]} , 0 ' (0,1] . Then _0 = {(0,1],i} is a & algebra (the trivial algebra). Theorem 1.5 is important, because it guarantees that for any collection of subsets of there exists a smallest & algebra containing . By adding complements and countable unions it is possible to extend to a & algebra. This can always be done, because is contained in the & algebra of all subsets of , but there is often no unique way of doing this, except in the case where is finite. Thus, let , 0 , be the collection of all & algebras
containing . Then = _0 is the smallest & algebra containing .
32 Definition 1.4: The smallest & algebra containing a given collection of sets is called the & algebra generated by , and is usually denoted by () . Note that ' ^0 is not always a & algebra. For example, let = [0,1], and let
4
for n $ 1, n ' {[0,1] , i , [0,1&n &1] , (1&n &1,1]} . Then An ' [0,1&n &1] 0 n d ^n'1n , but the interval [0,1) = ^n'1An is not contained in any of the & algebras n , hence
4
^n'1An ^n'1n .
4 4
by augmenting it with the missing sets. The smallest & algebra containing ^0 is usually denoted by 0 ' ^0 .
def.
However, it is always possible to extend ^0 to a & algebra, often in various ways,
The notion of smallest -algebra of subsets of is always relative to a given collection of subsets of . Without reference to such a given collection the smallest -algebra of subsets of is { , i} , which is called the trivial -algebra. Moreover, similarly to Definition 1.4 we can define the smallest algebra of subsets of containing a given collection of subsets of , which we will denote by () . For example, let = (0,1], and let be the collection of all intervals of the type (a,b] with 0 # a < b # 1 . Then () consists of the sets in together with the empty set i, and all finite unions of disjoint sets in . To see this, check first that this collection () is an algebra, as follows. (a) The complement of (a,b] in is (0,a]^(b,1] . If a = 0 then (0,a] ' (0,0] ' i , and if
b = 1 then (b,1] ' (1,1] ' i , hence (0,a]^(b,1] is a set in or a finite union of disjoint sets in . (b) Let (a,b] in and (c,d] in , where without loss of generality we may assume that a # c. If b < c then (a,b]^(c,d] is a union of disjoint sets in . If c # b # d then (a,b]^(c,d] ' (a,d] is a set in itself, and if b > d then (a,b]^(c,d] ' (a,b] is a set in itself. Thus, finite unions of sets in are either sets in itself or finite unions of disjoint sets in . (c) ' ^n A j'0(bj,aj%1] , where b0 ' 0 and an%1 ' 1 , which is a finite union of disjoint sets in itself. Moreover, similarly to part (b) it is easy to verify that finite unions of sets of the type A can be written as finite unions of disjoint sets in . Thus, the sets in together with the empty set i and all finite unions of disjoint sets in form an algebra of subsets of = (0,1]. Let A ' ^j'1(aj ,b j] , where 0 # a1 < b1 < a2 < b2 < ...... < an < b n # 1 . Then
n
33
In order to verify that this is the smallest algebra containing , remove one of the sets in this algebra that does not belong to itself. Since all sets in the algebra are of the type A in part (c), let us remove this particular set A. But then ^j'1(aj ,bj] is no longer included in the collection,
n
hence we have to remove each of the intervals (aj ,bj] as well, which however is not allowed because they belong to . Note that the algebra () is not a -algebra, because countable infinite unions are not always included in () . For example, ^n'1(0,1&n &1] ' (0,1) is a countable union of sets in
4
() which itself is not included in () . However, we can extend () to (()) , the smallest -algebra containing () , which coincides with () .
34 1.4.2 Borel sets An important special case of Definition 1.4 is where ' , and is the collection of all open intervals: ' {(a,b) : a < b , a,b 0 } . (1.18)
Definition 1.5: The & algebra generated by the collection (1.18) of all open intervals in is called the Euclidean Borel field, denoted by B, and its members are called the Borel sets.
Note, however, that B can be defined in different ways, because the & algebras generated by the collections of open intervals, closed intervals: {[a,b] : a # b , a,b 0 } , and half-open intervals, {(&4,a] : a 0 } , respectively, are all the same! We show this for one case only:
Theorem 1.6: B = ({(&4,a] : a 0 }) .
Proof: Let ( ' {(&4 , a] : a 0 } . (a) (1.19)
If the collection defined by (1.18) is contained in (() , then (() is a & algebra
containing . But B = () is the smallest & algebra containing , hence B = () d ((). In order to prove this, construct an arbitrary set (a,b) in out of countable unions and/or complements of sets in ( , as follows. Let A ' (&4 , a] and B ' (&4 , b] , where a < b are
35 0 ( ) , and thus arbitrary real numbers. Then A , B 0 ( , hence A , B ( This implies that (() contains all sets of the type (a,b] , hence (a,b) = ^n'1(a , b & (b&a)/n]
4
0 ( ) . ~(a,b] ' (&4 , a]^(b , 4) ' A^B (
0 (() . Thus, d (() . (b) If the collection ( defined by (1.19) is contained in B = () , then () is a
& algebra containing ( . But (() is the smallest & algebra containing ( , hence (() d () = B. In order to prove the latter, observe that for m = 1,2,...., Am ' ^n'1(a&n , a%m &1) is a
4
0 () , and consequently (&4 , a] ' _4 countable union of sets in , hence A m'1Am = m

4 ) 0 () . Thus, d () = B. ~(^m'1A m (
We have shown now that B = () d (() and (() d () = B. Thus, B and (() are the same. Q.E.D.8 The notion of Borel set extends to higher dimensions as well:
Definition 1.6: Bk = ({j'1(aj,bj) : aj < bj , aj , b j 0 }) is the k-dimensional Euclidean Borel field. Its members are also called Borel sets (in k ).
Also this is only one of the ways to define higher-dimensional Borel sets. In particular, similarly to Theorem 1.6 we have:
Theorem 1.7: Bk
= ({j'1(&4,aj] : aj 0 }) .
36
1.5.
Properties of probability measures The three axioms (1.8), (1.9), and (1.10) imply a variety of properties of probability
measures. Here we list only the most important ones.
Theorem 1.8: Let { , , P} be a probability space. The following hold for sets in : (a) P(i) ' 0 , ) ' 1 & P(A) , (b) P(A (c) A d B implies P(A) # P(B) , (d) P(A^B) % P(A_B) ' P(A) % P(B) , (e) If An d An%1 for n ' 1,2,..., then P(An) 8 P(^n'1An) ,
4
(f) If An e An%1 for n ' 1,2,..., then P(An) 9 P(_n'1An) ,

4
(g) P(^n'1An) # 'n'1P(An) .

4 4
Proof: (a)-(c): Easy exercises. ) ^ (A_B) ^ (B_A ) is a union of disjoint sets, hence by axiom (1.10), (d) A^B ' (A_B ) % P(A_B) % P(B_A ) . Moreover, A ' (A_B ) ^ (A_B) is a union of P(A^B) = P(A_B ) % P(A_B) , and similarly, P(B) ' P(B_A ) % P(A_B) . disjoint sets , hence P(A) ' P(A_B Combining these results, part (d) follows. (e) Let B1 ' A1 , Bn ' An\An&1 for n $ 2 . Then An ' ^j'1Aj ' ^j'1Bj and ^j'1Aj ' ^j'1Bj .
n n 4 4
Since the Bj s are disjoint, it follows from axiom (1.10) that
Part (e) follows now from the fact that 'j'n%1P(Bj) 9 0 .

4
P(^j'1Aj) ' 'j'1P(Bj) ' 'j'1P(Bj) % 'j'n%1P(Bj) ' P(An) % 'j'n%1P(Bj) .

4 4 n 4 4
37
(f) This part follows from part (e) , using complements. (g) Exercise
1.6. 1.6.1
The uniform probability measure Introduction Fill a bowl with ten balls numbered from zero to nine. Draw randomly a ball from this
bowl, and write down the corresponding number as the first decimal digit of a number between zero and one. For example, if the first drawn number is 4, then write down 0.4. Put the ball back in the bowl, and repeat this experiment. If for example the second ball corresponds to the number 9, then this number becomes the second decimal digit: 0.49. Repeating this experiment infinitely many times yields a random number between zero and one. Clearly, the sample space involved is the unit interval: ' [0,1] . For a given number x 0 [0,1] the probability that this random number is less or equal to x is: x. To see this, suppose that you only draw two balls, and that x = 0.58. If the first ball has a number less than 5, it does not matter what the second number is. There are 5 ways to draw a first number less or equal to 4, and 10 ways to draw the second number. Thus, there are 50 ways to draw a number with a first digit less or equal to 4. There is only one way to draw a first number equal to 5, and 9 ways to draw a second number less or equal to 8. Thus, the total number of ways we can generate a number less or equal to 0.58 is 59, and the total number of ways we can draw two numbers with replacement is 100. Therefore, if we only draw two balls
38 with replacement, and use the numbers involved as the first and second decimal digit, the probability that we get a number less or equal to 0.58 is: 0.59. Similarly, if we draw 10 balls with replacement, the probability that we get a number less or equal to, say, 0.5831420385 is: 0.5831420386. In the limit the difference between x and the corresponding probability disappears. Thus, for x 0 [0,1] we have: P([0,x]) ' x . By the same argument it follows that for x 0 [0,1] , P({x}) ' P([x,x]) ' 0 , i.e., the probability that the random number involved will be exactly equal to a given number x is zero. Therefore, for a given x 0 [0,1] , P((0,x]) = P([0,x)) = P((0,x)) ' x . More generally, for any interval in [0,1] the corresponding probability is the length of the interval involved, regardless as to whether the endpoints are included or not: Thus, for 0 # a < b # 1 we have P([a,b]) ' P((a,b]) ' P([a,b)) ' P((a,b)) = b!a. Any finite union of intervals can be written as a finite union of disjoint intervals by cutting out the overlap. Therefore, this probability measure extends to finite unions of intervals, simply by adding up the lengths of the disjoint intervals involved. Moreover, observe that the collection of all finite unions of sub-intervals in [0,1], including [0,1] itself and the empty set, is closed under the formation of complements and finite unions. Thus, we have derived the probability measure P corresponding to the statistical experiment under review for an algebra 0 of subsets of [0,1], namely 0 ' {(a,b),[a,b],(a,b],[a,b) , a,b0[0,1], a#b, and their finite unions } , (1.20)
where [a,a] is the singleton {a}, and each of the sets (a,a), (a,a] and [a,a) should be interpreted as the empty set i . This probability measure is a special case of the Lebesgue measure, which assigns to each interval its length. If you are only interested in making probability statements about the sets in the algebra
39 (1.20), then your are done. However, although the algebra (1.20) contains a large number of sets, we cannot yet make probability statements involving arbitrary Borel sets in [0,1], because not all the Borel sets in [0,1] are included in (1.20). In particular, for a countable sequence of sets Aj 0 0 the probability P(^j'1Aj) is not always defined, because there is no guarantee that
4 4
^j'1Aj 0 0 . Therefore, if you want to make probability statements about arbitrary Borel set in
[0,1], you need to extend the probability measure P on 0 to a probability measure defined on the Borel sets in [0,1]. The standard approach to do this is to use the outer measure:
1.6.2
Outer measure Any subset A of [0,1] can always be completely covered by a finite or countably infinite
union of sets in the algebra 0 : A d ^j'1Aj , where Aj 0 0 , hence the probability of A is

4
sequences of sets Aj 0 0 such that A d ^j'1Aj then yields the outer measure:
4
bounded from above by 'j'1P(Aj) . Taking the infimum of 'j'1P(Aj) over all countable
4 4
Definition 1.7: Let 0 be an algebra of subsets of . The outer measure of an arbitrary subset A of is: P ((A) ' inf 'j'1P(Aj) .
4
Ad^j'1A j , A j00
4
(1.21)
Note that it is not required in (1.21) that ^j'1Aj 0 0 .

4
Since a union of sets Aj in an algebra 0 can always be written as a union of disjoint sets in the algebra algebra 0 (see Theorem 1.4), we may without loss of generality assume that the
infimum in (1.21) is taken over all disjoint sets Aj in 0 such that such that A d ^j'1Aj . This
4
40
implies that If A 0 0 then P ((A) ' P(A) . The question now arises for which other subsets of the outer measure is a probability measure. Note that the conditions (1.8) and (1.9) are satisfied for the outer measure P ( (Exercise: Why?), but in general condition (1.10) does not hold for arbitrary sets. See for example Royden (1968, pp. 63-64). Nevertheless, it is possible to extend the outer measure to a probability measure on a -algebra containing 0 : (1.22)
Theorem 1.9: Let P be a probability measure on { , 0 }, where 0 is an algebra, and let ' (0) be the smallest & algebra containing the algebra 0 . Then the outer measure P* is a unique probability measure on { , } which coincides with P on 0 .
The proof that the outer measure P* is a probability measure on ' (0) which coincide with P on 0 is lengthy and therefore given in Appendix B. The proof of the uniqueness of P* is even more longer and is therefore omitted. Consequently, for the statistical experiment under review there exists a & algebra of subsets of ' [0,1] , containing the algebra 0 defined in (1.20), for which the outer measure P (: 6 [0,1] is a unique probability measure. This probability measure assigns in this case to each interval in [0,1] its length as probability. It is called the uniform probability measure. It is not hard to verify that the & algebra involved contains all the Borel subsets of [0,1]: {[0,1]_B , for all Borel sets B} d . (Exercise: Why?) This collection of Borel
subsets of [0,1] is usually denoted by [0,1] _ B , and is a & algebra itself (Exercise: Why?). Therefore, we could also describe the probability space of this statistical experiment by the
41
probability space {[0,1], [0,1] _ B, P ( }, where P ( is the same as before. Moreover, defining the probability measure on B as (B) ' P (([0,1]_B) , we can describe this statistical experiment also by the probability space { , B, }, where in particular ((&4,x]) ' 0 if x # 0 , ((&4,x]) ' x if 0 < x # 1 , ((&4,x]) ' 1 if x > 1 , and more generally for intervals with endpoints a < b, ((a,b)) ' ([a,b]) ' ([a,b)) ' ((a,b]) ' ((&4,b]) & ((&4,a]) , whereas for all other Borel sets B, (B) '
B d ^j'1(a j , b j)
4
inf
'j'1((aj , b j)) .
4
(1.23)
1.7. 1.7.1
Lebesgue measure and Lebesgue integral Lebesgue measure Along similar lines as in the construction of the uniform probability measure we can
define the Lebesgue measure, as follows. Consider a function which assigns to each open interval (a,b) its length: ((a,b)) ' b & a , and define for all other Borel sets B in , (B) ' inf
4
B d ^j'1(a j , b j)
'j'1((a j , bj)) '

4
B d ^j'1(a j , b j)
4
inf
'j'1(bj & aj) .

4
This function is called the Lebesgue measure on , which measures the total length of a Borel set, where the measurement is taken from the outside. Similarly, let now (i'1(ai,b i)) ' (i'1(bi &a i) , and define for all other Borel sets B in
k k
k,
42 (B) ' inf

k
B d ^j'1{i'1(ai,j , bi,j)}
4
'j'1 i'1(ai,j , bi,j) '

4 k
B d ^j'1{i'1(ai,j , bi,j)}
4 k
inf
'j'1 (i'1(bi,j & ai,j) .

4 k
This is the Lebesgue measure on k, which measures the area (in the case k = 2) or the volume (in the case k $ 3) of a Borel set in k, where again the measurement is taken from the outside. Note that in general Lebesgue measures are not probability measures, because the Lebesgue measure can be infinite. In particular, ( k) = 4. However, if confined to a set with Lebesgue measure 1 it becomes the uniform probability measure. More generally, for any Borel set A 0 k with positive and finite Lebesgue measure, (B) ' (A_B)/(A) is the uniform probability measure on B k _ A.
1.7.2
Lebesgue integral The Lebesgue measure gives rise to a generalization of the Riemann integral. Recall that
the Riemann integral of a non-negative function f(x) over a finite interval (a,b] is defined as f(x)dx ' sup j inf f(x) (Im) m m'1 x0I m
n a b
where the Im are intervals forming a finite partition of (a,b] , i.e., they are disjoint, and their union is (a,b]: (a,b] ' ^m'1Im , 8(Im ) is the length of Im , hence 8(Im ) is the Lebesgue measure
n
of Im , and the supremum is taken over all finite partitions of (a,b]. Mimicking this definition, the Lebesgue integral of a non-negative function f(x) over a Borel set A can be defined as f(x)dx ' sup j mA m'1
n x0B m
inf f(x) (Bm) ,
43 where now the Bm s are Borel sets forming a finite partition of A, and the supremum is taken over all such partitions. If the function f(x) is not non-negative, we can always write it as the difference of two non-negative functions: f(x) ' f%(x) & f&(x) , where f%(x) ' max[0 , f(x)], f&(x) ' max[0 , &f(x)] . Then the Lebesgue integral over a Borel set A is defined as mA f(x)dx ' mA f%(x)dx & mA f&(x)dx ,
provided that at least one of the right hand side integrals is finite. However, we still need to impose a further condition on the function f in order to be Lebesgue integrable. A sufficient condition is that for each Borel set B in , the set {x: f(x) 0 B} is a Borel set itself. As we will see in the next chapter, this is the condition for Borel measurability of f. Finally, note that if A is an interval and f(x) is Riemann integrable over A, then the Riemann integral and the Lebesgue integral coincide.
1.8. 1.8.1
Random variables and their distributions Random variables and vectors Loosely speaking, a random variable is a numerical translation of the outcomes of a
statistical experiment. For example, flip a fair coin once. Then the sample space is ' {H , T} , where H stands for Head, and T stands for Tail. The & algebra involved is = {,i,{H},{T}}, and the corresponding probability measure is defined by P({H} ) ' P( {T} } ) ' 1/2 . Now
44 define the function X() ' 1 if ' H , X() ' 0 if ' T . Then X is a random variable which takes the value 1 with probability and the value 0 with probability :
(short&hand notation)
P(X ' 1) P(X ' 0)
' '
P({0 : X() ' 1}) ' P({H}) ' 1/2 , P({0 : X() ' 0}) ' P({T}) ' 1/2 .
(short&hand notation)
Moreover, for an arbitrary Borel set B we have ' P({H}) P(X 0 B) ' P({0 : X() 0 B}) ' P({T}) ' P(i) ' 1/2 if 1 0 B and 0 B , ' 1/2 if 1 B and 0 0 B , 1 0 if 1 0 B and 0 0 B , if 1 B and 0 B , '
' P({H , T}) '
where again P(X 0 B) is a short-hand notation9 for P({0 : X() 0 B}) . In this particular case the set {0 : X() 0 B} is automatically equal to one of the elements of , and therefore the probability P(X 0 B) = P( {0 : X() 0 B} ) is welldefined. In general, however, we need to confine the mappings X : 6 to those for which we can make probability statements about events of the type {0 : X() 0 B} , where B is an arbitrary Borel set, which is only possible if these sets are members of :
Definition 1. 8: Let { , , P} be a probability space. A mapping X: 6 is called a random variable defined on { , , P} if X is measurable , which means that for every Borel set B, {0 : X() 0 B} 0 . Similarly, a mapping X: 6 k is called a k-dimensional random vector defined on { , , P} if X is measurable , in the sense that for every Borel set B in B k, {0 : X() 0 B} 0 .
45 In verifying that a real function X: 6 is measurable , it is not necessary to verify that for all Borel sets B, sets of the type (&4 ,x] : {0 : X() 0 B} 0 , but only that this property holds for Borel
Theorem 1.10: A mapping X: 6 is measurable (hence X is a random variable) if and only if for all x 0 the sets {0 : X() # x} are members of . Similarly, a mapping X: 6 k is measurable (hence X is a random vector of dimension k) if and only if for all x ' (x1 ,..... , xk )T 0 k the sets
k
_j'1{0 : Xj() # xj} ' {0 : X() 0 j'1(&4 , xj]}

k
are members of , where the Xjs are the components of X.
Proof: Consider the case k = 1. Suppose that {0 : X() 0 (&4,x]} 0 , x 0 . Let D be the collection of all Borel sets B for which {0 : X() 0 B} 0 . Then D d B, and
D contains the collection of half-open intervals (&4 , x] , x 0 . If D is a & algebra itself, it
is a & algebra containing the half-open intervals. But B is the smallest & algebra containing the half-open intervals (see Theorem 1.6), so that then B d D, hence D ' B. Therefore, it suffices to prove that D is a & algebra: (a) Let B 0 D. Then {0 : X() 0 B} 0 , hence } 0 ~{0 : X() 0 B} ' {0 : X() 0 B 0 D. and thus B (b) Next, let Bj 0 D for j = 1,2,.... Then {0 : X() 0 Bj} 0 , hence
and thus ^j'1Bj 0 D.

4
^j'1{0 : X() 0 Bj} ' {0 : X() 0 ^j'1Bj} 0

4 4
46
The proof of the case k > 1 is similar. Q.E.D.10 The sets {0 : X() 0 B} are usually denoted by X &1(B) : X &1(B) ' { 0 : X() 0 B} . The collection X ' {X &1(B), B 0 B} is a & algebra itself (Exercise: Why?), and is called the & algebra generated by the random variable X. More generally:
def.
Definition 1.9: Let X be a random variable (k=1) or a random vector (k > 1). The & algebra X = {X &1(B), B 0 B k} is called the & algebra generated by X.
In the coin tossing case, the mapping X is one-to-one, and therefore in that case X is the same as , but in general X will be smaller than . For example, roll a dice, and let X = 1 if the outcome is even, and X = 0 if the outcome is odd. Then X ' {{1,2,3,4,5,6} , {2,4,6} , {1,3,5} , i} , whereas in this case consists of all subsets of ' {1,2,3,4,5,6} . Given a k dimensional random vector X, or a random variable X (the case k=1), define for arbitrary Borel sets B 0 Bk : X(B) ' P X &1(B) ' P {0: X() 0 B} . (1.24)
Then X(@) is a probability measure on { k , Bk }:
47 (a) (b) (c) for all B 0 Bk, X(B) $ 0 , X(k) ' 1 , for all disjoint
j
0 B k , X ^j'1Bj ' 'j'1 X(Bj) .

4 4
Thus, the random variable X maps the probability space { , , P} into a new probability space, { , B, X }, which in its turn is mapped back by X &1 into the (possibly smaller) probability space { , X , P} . Similarly for random vectors.
Definition 1.10: The probability measure X(@) defined by (1.24) is called the probability measure induced by X.
1.8.2
Distribution functions For Borel sets of the type (&4 , x] , or j'1(&4 , xj] in the multivariate case, the value of the
k
induced probability measure X is called the distribution function:
Definition 1.11: Let X be a random variable (k=1) or a random vector ( k>1) with induced probability measure X . The function F(x) ' X(j'1(&4 , xj]) , x ' (x1 , .... , xk)T 0 k , is called the distribution function of X.
k
It follows from these definitions, and Theorem 1.8 that
Theorem 1.11: A distribution function of a random variable is always right continuous:
48 x 0 , lim90F(x % ) ' F(x) , and monotonic non-decreasing: F(x1) # F(x2) if x1 < x2 , with limx9&4F(x) ' 0, limx84F(x) ' 1 .
Proof: Exercise. However, a distribution function is not always left continuous. As a counter example, consider the distribution function of the Binomial (n,p) distribution in Section 1.2.2. Recall that the corresponding probability space consists of sample space ' {0,1,2,...,n}, the -algebra of all subsets of , and probability measure P({k}) defined by (1.15) . The random variable X involved is defined as X(k) = k, with distribution function F(x) ' 0 for x < 0 ,
F(x) ' 'k#xP({k}) for x 0 [0,n] , F(x) ' 1 for x > n , Now let for example x ' 1. Then for 0 < < 1, F(1 & ) ' F(0) , and F(1 % ) ' F(1) , hence lim90F(1 % ) ' F(1) , but lim90F(1 & ) ' F(0) < F(1) . The left limit of a distribution function F in x is usually denoted by F(x!): F(x&) ' lim90F(x & ) . Thus if x is a continuity point then F(x-) = F(x), and if x is a discontinuity point then F(x-) < F(x). The Binomial distribution involved is an example of a discrete distribution. The uniform distribution on [0,1] derived in Section 1.5 is an example of a continuous distribution, with distribution function
def.
49 F(x) ' 0 for x < 0 , F(x) ' x for x 0 [0,1] , F(x) ' 1 for x > 1 . In the case of the Binomial distribution (1.15) the number of discontinuity points of F is finite, and in the case of the Poisson distribution (1.16) the number of discontinuity points of F is countable infinite. In general we have: (1.25)
Theorem 1.12: The set of discontinuity points of a distribution function of a random variable is countable.
Proof: Let D be the set of all discontinuity points of the distribution function F(x). Every point x in D is associated with an non-empty open interval (F(x-),F(x)) = (a,b), say, which is contained in [0,1]. For each of these open intervals (a,b) there exists a rational number q such a < q < b , hence the number of open intervals (a,b) involved is countable, because the rational numbers are countable. Therefore, D is countable. Q.E.D. The results of Theorems 1.11-1.12 only hold for distribution functions of random variables, though. It is possible to generalize these results to distribution functions of random vectors, but this generalization is far from trivial and therefore omitted. As follows from Definition 1.11, a distribution function of a random variable or vector X is completely determined by the corresponding induced probability measure X(@) . But what about the other way around, i.e., given a distribution function F(x), is the corresponding induced probability measure X(@) unique? The answer is yes, but we prove the result only for the
50 univariate case:
Theorem 1.13: Given the distribution function F of a random vector X 0 k, there exists a unique probability measure on { k , Bk} such that for x ' (x1,....,xk)T 0 k , F(x) = i'1(&4 ,xi] .
k
Proof: Let k = 1 and let T0 be the collection of all intervals of the type (a,b),[a,b],(a,b],[a,b) , (&4,a) , (4,a] , (b,4) , [b,4) , a#b 0 , (1.26)
together with their finite unions, where [a,a] is the singleton {a}, and (a,a), (a,a] and [a,a) should be interpreted as the empty set i . Then each set in T0 can be written as a finite union of disjoint sets of the type (1.26) (Compare (1.20) ), hence T0 is an algebra. Define for !4 < a < b < 4, ((a,a)) ' ((a,a]) ' ([a,a)) ' (i) ' 0 ({a}) ' F(a) & lim90F(a&) , ((a,b]) ' F(b) & F(a) ([a,b)) ' ((a,b]) & ({b}) % ({a}) , ([a,b]) ' ((a,b]) % ({a}) ((a,b)) ' ((a,b]) & ({b}) , ((&4,a]) ' F(a) ((&4,a]) ' F(a) & ({a}) , ((b,4)) ' 1 & F(b) ([b,4)) ' ((b,4)) % ({b}) and let for disjoint sets A1 , ....... , An of the type (1.26), (^j'1Aj ) ' 'j'1(Aj) . Then the
n n
distribution function F defines a probability measure on T0 , and this probability measure coincides on T0 with the induced probability measure X . It follows now from Theorem 1.9 that there exists a -algebra T containing T0 for which the same applies. This -algebra T may be
51 chosen equal to the -algebra B of Borel sets. Q.E.D. The importance of this result is that there is a one-to-one relationship between the distribution function F of a random variable or vector X and the induced probability measure X . Therefore, the distribution function contains all the information about X .
Definition 1.12: A distribution function F on k and its associated probability measure on { k , Bk} are called absolutely continuous with respect to Lebesgue measure if for every Borel set B in k with zero Lebesgue measure, (B) = 0.
We will need this concept in the next section.
1.9.
Density functions An important concept is that of a density function. Density functions are usually
associated to differentiable distribution functions:
Definition 1.13: The distribution of a random variable X is called absolutely continuous if there exists a non-negative integrable function f, called the density function of X, such that the distribution function F of X can be written as the (Lebesgue) integral F(x) = f(u)du . m&4
x
Similarly, the distribution of a random vector X 0 k is called absolutely continuous if there exists a non-negative integrable function f on k , called the joint density, such that the distribution function F of X can be written as the integral F(x) ' m&4
x1
.....
m&4
xk
f(u1 ,... ,uk)du1....duk ,
52 where x ' (x1 , ...... , xk)T .
Thus, in the case F(x) '
f(u)du the density function f(x) is the derivative of F(x): m&4

x
f(x) ' F )(x) , and in the multivariate case F(x1,...,xk) ' density is f(x1,...,xk) ' (M/Mx1)......(M/Mxk)F(x1,...,xk) .
m&4
x1
.....
m&4
xk
f(u1 ,... ,u k)du1....duk the joint
The reason for calling the distribution functions in Definition 1.13 absolutely continuous is that in this case the distributions involved are absolutely continuous with respect to Lebesgue measure. See Definition 1.12. To see this, consider the case F(x) ' (Exercise) that the corresponding probability measure is: (B) ' mB f(x)dx , (1.27) f(u)du , and verify m&4
x
where the integral is now the Lebesgue integral over a Borel set B. Since the Lebesgue integral over a Borel set with zero Lebesgue measure is zero (Exercise), it follows that (B) = 0 if the Lebesgue measure of B is zero. For example the uniform distribution (1.25) is absolutely continuous, because we can write (1.25) as F(x) ' f(u)du , with density f(u) = 1 for 0 < u < 1 and zero elsewhere. Note m&4
x
that in this case F(x) is not differentiable in 0 and 1, but that does not matter, as long as the set of points for which the distribution function is not differentiable has zero Lebesgue measure. Moreover, a density of a random variable always integrates to 1, because 1 ' limx64F(x) '
4 4 4
f(u)du . Similarly for random vectors X 0 k : m&4

4
..... f(u ,... ,uk)du1....du k ' 1 . m&4m&4 m&4 1 Note that continuity and differentiability of a distribution function are not sufficient
53 conditions for absolute continuity. It is possible to construct a continuous distribution function F(x) that is differentiable on a subset D d , with \D a set with Lebesgue measure zero, such that F )(x) / 0 on D, so that in this case F )(x)dx / 0. Such distributions functions are called m&4
x
singular. See Chung (1974, pp. 12-13) for an example of how to construct a singular distribution function on , and Chapter 5 for singular multivariate normal distributions.
1.10.
Conditional probability, Bayes rule, and independence
1.10.1 Conditional probability Consider statistical experiment with probability space {S,,P}, and suppose that it is known that the outcome of this experiment is contained in a set B with P(B) > 0. What is the probability of an event A, given that the outcome of the experiment is contained in B? For example, roll a dice. Then S = {1,2,3,4,5,6}, is the F-algebra of all subsets of S, and P({T}) = 1/6 for T = 1,2,3,4,5,6. Let B be the event: "the outcome is even": B = {2,4,6}, and let A = {1,2,3}. If we know that the outcome is even, then we know that the outcomes {1,3} in A will not occur: if the outcome in contained in A, it is contained in A1B = {2}. Knowing that the outcome is either 2,4, or 6, the probability that the outcome is contained in A is therefore 1/3 = P(A1B)/P(B). This is the conditional probability of A, given B, denoted by P(A|B). If it is revealed that the outcome of a statistical experiment is contained in a particular set B, then the sample space S is reduced to B, because we then know that the outcomes in the complement of B will not occur, the F-algebra is reduced to 1B, the collection of all intersections of the sets in with B: 1B ={A1B, A0} (Exercise: Is this a F-algebra?), and the probability measure involved becomes P(A|B) = P(A1B)/P(B), hence the probability space becomes
{B , _B , P(@|B)} . See Exercise 19 below.
54
1.10.2 Bayes rule form a partition of the sample space S, Let A and B be sets in . Since the sets A and A ) , hence we have B ' (B _ A) ^ (B _ A ) ' P(B*A)P(A) % P(B*A )P(A ) . P(B) ' P(B_A) % P(B_A Moreover, P(A*B) ' P(A_B) P(B*A)P(A) ' . P(B) P(B)
Combining these two results now yields Bayes' rule: P(A*B) ' P(B*A)P(A) )P(A ) P(B*A)P(A) % P(B*A .
Thus, Bayes rule enables us to compute the conditional probability P(A|B) if P(A) and the ) are given. conditional probabilities P(B*A) and P(B*A More generally, if Aj, j =1,2,.....n (# 4) is a partition of the sample space S, i.e., the Ajs are disjoint sets in such that ' ^j'1Aj , then
n
P(Ai*B) '
'j'1P(B*Aj)P(Aj)
n
P(B*Ai)P(Ai)
Bayes rule plays an important role in a special branch of statistics [and econometrics], called Bayesian statistics [econometrics].
55 1.10.3 Independence If P(A|B) = P(A), then knowing that the outcome is in B does not give us any information about A. In that case the events A and B are called independent. For example, if I tell you that the outcome of the dice experiment is contained in the set {1,2,3,4,5,6} = S, then you know nothing about the outcome: P(A|S) = P(A1S)/P(S) = P(A), hence S is independent of any other event A. Note that P(A|B) = P(A) is equivalent to P(A1B) = P(A)P(B). Thus,
Definition 1.14: Sets A and B in are (pairwise) independent if P(A1B) = P(A)P(B).
If events A and B are independent, and events B and C are independent, are the events A and C independent? The answer is: not necessarily. In order to give a counter example, observe and B , A and B , and A and B , because that if A and B are independent, then so are A _B) ' P(B) & P(A_B) ' P(B) & P(A)P(B) ' (1&P(A))P(B) ' P(A )P(B) , P(A and similarly, ) ' P(A)P(B ) and P(A _B ) ' P(A )P(B ) . P(A_B and 0 < P(A) < 1, then B and C = A are independent if A and B are independent, Now if C = A but ) ' P(i) ' 0 , P(A_C) ' P(A_A whereas ) ' P(A)(1&P(A)) 0 . P(A)P(C) ' P(A)P(A Thus, for more than two events we need a stronger condition for independence than pairwise
56 independence, namely:
Definition 1.15: A sequence Aj of sets in is independent if for every sub-sequence Aj , i = 1,2,..,n, P(_i'1Aj ) ' (i'1P(Aj ) .
n n
i i i
By requiring that the latter holds for all sub-sequences rather than P(_i'1Ai ) ' (i'1P(Ai ) , we
4 4
avoid the problem that a sequence of events would be called independent if one of the events is the empty set. The independence of a pair or sequence of random variables or vectors can now be defined as follows.
Definition 1.16: Let Xj be a sequence of random variables or vectors defined on a common probability space {S,,P}. X1 and X2 are pairwise independent if for all Borel sets B1, B2, the sets A1 ' {0: X1() 0 B1} and A2 ' {0: X2() 0 B2} are independent. The sequence Xj is independent if for all Borel sets Bj the sets Aj ' {0: Xj() 0 Bj} are independent.
As we have seen before, the collection j ' {{0: Xj() 0 B} , B 0 B}} = {Xj (B), B 0 B}} is a sub F-algebra of . Therefore, Definition 1.16 also reads: The sequence of random variables Xj is independent if for arbitrary Aj 0 j the sequence of sets Aj is independent according to Definition 1.15. Independence usually follows from the setup of a statistical experiment. For example, draw randomly with replacement n balls from a bowl containing R red balls and N!R white
&1
57 balls, and let Xj = 1 if the j-th draw is a red ball, and Xj =0 if the j-th draw is a white ball. Then X1,...,Xn are independent (and X1+...+Xn has the Binomial (n,p) distribution, with p = R/N). However, if we would draw these balls without replacement, then X1,...,Xn are not independent. For a sequence of random variables Xj it suffices to verify the condition in Definition 1.16 for Borel sets Bj of the type (!4,xj], xj 0 , only:
Theorem 1.14: Let X1,...,Xn be random variables, and denote for x 0 and j = 1,....,n, Aj(x) ' {0: Xj() # x} . Then X1,...,Xn are independent if and only if for arbitrary (x1,.....,xn)T 0 n the sets A1(x1),......,An(xn) are independent.
The complete proof of Theorem 1.14 is difficult and is therefore omitted, but the result can be motivated as follow. Let j ' {,i,Xj ((&4,x]),Xj ((y,4)), x,y0, together with all finite unions and intersections of the latter two types of sets}. Then j is an algebra such that for arbitrary Aj 0 j the sequence of sets Aj is independent. This is not too hard to prove. Now j ' {Xj (B), B 0 B}} is the smallest -algebra containing j , and is also the smallest monotone class containing j . It can be shown (but this is the hard part), using the properties of monotone class (see Exercise 11 below), that for arbitrary Aj 0 j the sequence of sets Aj is independent as well It follows now from Theorem 1.14 that:
0 &1 0 0 0 0 &1 &1
Theorem 1.15: The random variables X1,...,Xn are independent if and only if the joint distribution function F(x) of X = (X1,...,Xn)T can be written as the product of the distribution
58 functions Fj(xj) of the Xj s, i.e., F(x) ' (j'1Fj(xj) , where x ' (x1 , .... , xn)T .
n
The latter distribution functions Fj(xj) are called the marginal distribution functions. Moreover, it follows straightforwardly from Theorem 1.15 that if the joint distribution of X ' (X1,....,Xn)T is absolutely continuous with joint density function f(x), then X1,...,Xn are independent if and only if f(x) can be written as the product of the density functions fj(xj) of the Xj s: f(x) ' (j'1fj(xj) , where x ' (x1 , .... , xn)T .
n
The latter density functions are called the marginal density functions.
1.11. 1. 2. 3.
Exercises Prove (1.4). Prove (1.17) by proving that ln[(1 & /n)n] ' n ln(1 & /n) 6 & for n 6 4 . Let ( be the collection of all subsets of = (0,1] of the type (a,b], where a < b are
rational numbers in [0,1], together with their finite disjoint unions and the empty set i . Verify that ( is an algebra. 4. 5. 6. Prove Theorem 1.2. Prove Theorem 1.5. Let = (0,1], and let be the collection of all intervals of the type (a,b] with
0 # a < b # 1 . Give as many distinct examples as you can of sets that are contained in () (the smallest -algebra containing this collection ), but not in () (the smallest algebra containing the collection ).
59 7. 8. 9. Show that ({[a,b] : a # b , a,b 0 }) = B. Prove part (g) of Theorem 1.8. Prove that 0 defined by (1.20) is an algebra. Prove (1.22). 11. A collection of subsets of a set is called a monotone class if the following two
conditions hold: An 0 , An d An%1, n ' 1,2,3,..... imply ^n'1An 0 ,

4
An 0 , An e An%1, n ' 1,2,3,..... imply _n'1An 0 .

4
Show that an algebra is a F-algebra if and only if it is a monotone class. 12. 0 , A collection of subsets of a set is called a & system if A 0 implies A
4
and for disjoint sets Aj 0 , ^j'1Aj 0 . A collection of subsets of a set is called a
& system if A,B 0 implies that A_B 0 . Prove that if a & system is also a & system,
then it is a F-algebra . 13. Let be the smallest & algebra of subsets of containing the (countable) collection
of half-open intervals (&4 , q] with rational endpoints q. Prove that contains all the Borel subsets of : B = . 14. Consider the following subset of 2: L ' {(x,y) 0 2: y ' x , 0 # x # 1} . Explain
why L is a Borel set. 15. Consider the following subset of 2: C ' {(x,y) 0 2: x 2 % y 2 # 1} . Explain why C
is a Borel set. 16. Prove Theorem 1.11. Hint: Use Definition 1.12 and Theorem 1.8. Determine first which
60 parts of Theorem 1.8 apply. 17. Let F(x) ' f(u)du be an absolutely continuous distribution function. Prove that m&4
x
corresponding probability measure is given by the Lebesgue integral (1.27). 18. 19. Prove that the Lebesgue integral over a Borel set with zero Lebesgue measure is zero. Let {,,P} be a probability space, and let B 0 with P(B) > 0. Verify that
{B ,_B,P(@|B)} is a probability space. 20. 21. Are disjoint sets in independent? (Application of Bayes rule): Suppose that 1 out of 10,000 people suffer from a certain
disease, say HIV+. Moreover, suppose that there exists a medical test for this disease which is 90% reliable: If you don't have the disease, the test will confirm that with probability 0.9, and the same if you do have the disease. If a randomly selected person is subjected to this test, and the test indicates that this person has the disease, what is the probability that this person actually has this disease? In other words, if you were this person, would you be scared or not? 22. and B are independent (and Let A and B in be pairwise independent. Prove that A
are independent and A and B are independent). therefore A and B 23. Draw randomly without replacement n balls from a bowl containing R red balls and
N!R white balls, and let Xj = 1 if the j-th draw is a red ball, and Xj =0 if the j-th draw is a white ball. Show that X1,...,Xn are not independent.
61
Appendices
1.A. Common structure of the proofs of Theorems 1.6 and 1.10 The proofs of Theorems 1.6 and 1.10 employ a similar argument, namely the following:
Theorem 1.A.1. Let be a collection of subsets of a set S, and let () be the smallest Falgebra containing . Moreover, let D be a Boolean function on () , i.e., D is a set function which takes either the value "True" or "False". Furthermore, let (A) ' True for all sets A in . If the collection D of sets A in () for which (A) ' True is a F-algebra itself, then (A) ' True for all sets A in () .
Proof: Since D is a collection of sets in () we have D d () . Moreover, by assumption, d D , and D is a F-algebra . But () is the smallest F-algebra containing , hence () d D . Thus, D ' () , and consequently, (A) ' True for all sets A in () . Q.E.D. This type of proof will also be used later on. Of course, the hard part is to prove that D is F-algebra. In particular, the collection D is not automatically a F-algebra. Take for example the case where S = [0,1], is the collection of all intervals [a,b] with 0 # a 0, and (A) ' False otherwise. In this case () consists of all the Borel subsets of [0,1], but D does not contain singletons whereas () does, so D is smaller than () , and therefore not a F-algebra.
62 1.B. Extension of an outer measure to a probability measure In order to use the outer measure as a probability measure for more general sets that those in 0 , we have to extend the algebra 0 to a -algebra of events for which the outer measure is a probability measure. In this appendix it will be shown how can be constructed, via the following lemmas..
Lemma 1.B.1: For any sequence Bn of disjoint sets in , P ((^n'1Bn) # 'n'1P ((Bn) .
4 4
Proof: Given an arbitrary g > 0 it follows from (1.21) that there exists a countable sequence of sets An,j in 0 such that Bn d ^j'1An,j and P ((Bn) > 'j'1P(An,j ) & g2&n , hence
4 4
'n'1P ((Bn) > 'n'1'j'1P(An,j ) & g'n'12&n ' 'n'1'j'1P(An,j ) & g .

4 4 4 4 4 4
(1.28)
Moreover, ^n'1Bn d ^n'1^j'1An,j , where the latter is a countable union of sets in 0 , hence it
4 4 4
follows from (1.21) that P ((^n'1Bn ) # 'n'1'j'1P(An,j ) .

4 4 4
(1.29)
Combining (1.28) and (1.29) it follows that for arbitrary g > 0 , 'n'1P ((Bn) > P ((^n'1Bn ) & g .
4 4
(1.30)
Letting g 9 0 , the lemma follows now from (1.30) . Q.E.D. Thus, in order for the outer measure to be a probability measure, we have to impose
63 conditions on the collection of subsets of such that for any sequence Bj of disjoint sets in , P ((^j'1Bj) $ 'j'1P ((Bj) . The latter is satisfied if we choose as follows:
4 4
Lemma 1.B.2: Let be a collection of subsets sets B of such that for any subset A of : ) . P ((A) ' P ((A_B) % P ((A_B Then for all countable sequences of disjoint sets Aj 0 , P ((^j'1Aj) ' 'j'1P ((Aj) .
4 4 4 ' ^4 Proof: Let A ' ^j'1Aj , B ' A1 . Then A_B ' A_A1 ' A1 and A_B j'2Aj are
(1.31)
disjoint, hence ) ' P ((A ) % P ((^j'2A ) . P ((^j'1Aj) ' P ((A) ' P ((A_B) % P ((A_B 1 j
4 4 4
(1.32)
Repeating (1.32) for P ((^j'kAj) with B ' Ak , k=2,...,n, it follows by induction that
4 n 4 n
hence P ((^j'1Aj) $ 'j'1P ((Aj) . Q.E.D.

4 4
P ((^j'1Aj) ' 'j'1P ((Aj) % P ((^j'n%1Aj) $ 'j'1P ((Aj) for all n $ 1 ,
such that 'j'1P(Aj) # P ((A) % g . Moreover, since A_B d ^j'1Aj_B, where Aj_B 0 0 ,
4 4
an arbitrary small number g > 0 . Then there exists an covering A d ^j'1Aj, where Aj 0 0 ,
4
Note that condition (1.31) automatically holds if B 0 0 : Choose an arbitrary set A and
d ^4 and A_B j'1Aj_B, where A j_B 0 0 , we have ) . follows now that P ((A) $ P ((A_B) + P ((A_B We show now that
( ( ( ) # '4 P ((A_B j'1P(Aj_B) , hence P (A_B) % P (A_B) # P (A) % g . Since g is arbitrary, it
P ((A_B) # 'j'1P(Aj_B) and

4
Lemma 1.B.3: The collection in Lemma 1.B.2 is a & algebra of subsets of , containing
64 the algebra 0 .
0 . Now let Proof: First, it follows trivially from (1.31) that B 0 implies B Bj 0 . It remains to show that ^j'1Bj 0 , which I will do in two steps. First, I will show
4
that is an algebra, and then I will use Theorem 1.4 to show that is also a & algebra. (a) B1^B2 0 . We have Proof that is an algebra: We have to show that B1,B2 0 implies that ) ' P ((A_B _B ) % P ((A_B _B ), P ((A_B 1 1 2 1 2 ) A_(B1^B2) ' (A_B1)^(A_B2_B 1 we have ). P ((A_(B1^B2)) # P ((A_B1) % P ((A_B2_B 1 Thus: _B ) # P ((A_B ) % P ((A_B _B ) % P ((A_B _B ) P ((A_(B1^B2)) % P ((A_B 1 2 1 2 1 2 1 ) ' P ((A) . ' P ((A_B1) % P ((A_B 1 (1.33)
and since
from (1.33) that P ((A) ' P ((A_(B1^B2)) % P ((A_(~(B1^B2)) . Thus, B1,B2 0 implies that B1^B2 0 , hence is an algebra (containing the algebra 0 ).
_B and P ((A) # P ((A_(B ^B )) % P ((A_(~(B ^B )) , it follows now Since ~(B1^B2) ' B 1 2 1 2 1 2
(b)
Proof that is a & algebra: Since we have established that is an algebra, it
follows from Theorem 1.4 that in proving that is also a & algebra it suffices to verify that ^j'1Bj 0 for disjoint sets Bj 0 : For such sets we have: A_(^j'1Bj)_Bn ' A_Bn , and
4 n
' A_(^j'1 B ) , hence A_(^j'1Bj)_B n j

n n&1
65
n n n &1 ) ' P ((A_B ) % P ((A_(^n P ((A_(^j'1Bj)) ' P ((A_(^j'1Bj)_Bn) % P ((A_(^j'1Bj)_B j'1 Bj)) . n n
Consequently, P ((A_(^j'1Bj)) ' 'j'1P ((A_Bj ) .

n n 4 n n ' _4 Next, let B ' ^j'1Bj . Then B j'1Bj d _j'1B j ' ~(^j'1Bj) , hence
(1.34)
) # P ((A_(~[^j'1B ])) . P ((A_B j

n
(1.35)
It follows now from (1.34) and (1.35) that for all n $ 1 , ) , P ((A) ' P ((A_(^j'1Bj)) % P ((A_(~[^j'1Bj])) $ 'j'1P ((A_Bj ) % P ((A_B
n n n
hence
4 ) $ P ((A_B) % P ((A_B ) , P ((A) $ 'j'1P ((A_Bj ) % P ((A_B
(1.36)
where the last inequality is due to P ((A_B) ' P ((^j'1(A_Bj)) # 'j'1P ((A_Bj) .
4 4
) (compare Lemma 1.B.1), it follows from Since we always have P ((A) # P ((A_B) % P ((A_B (1.36) that for countable unions B ' ^j'1Bj of disjoint sets Bj 0 ,
4
) , P ((A) ' P ((A_B) % P ((A_B
hence B 0 . Consequently, is a & algebra, and the outer measure P* is a probability measure on { , }. Q.E.D.
Lemma 1.B.4: The & algebra in Lemma 1.B.3 can be chosen such that P ( is unique: any
66 probability measure P( on {,} which coincide with P on 0 is equal to the outer measure P (.
The proof of Lemma 1.B.4 is too difficult and too long [see Billingsley (1986, Theorems 3.2-3.3)], and is therefore omitted. Combining Lemmas 1.B.2-1.B.4, Theorem 1.9 follows.
Endnotes 1. In the Spring of 2000 the Texas Lottery has changed the rules: The number of balls has been increased to 54, in order to create a larger jackpot. The official reason for this change is to make playing the lotto more attractive, because a higher jackpot will make the lotto game more exciting. Of course, the actual reason is to boost the lotto revenues! 2. Under the new rules (see note 1), this probability is: 1/25,827,165.
3. These binomial numbers can be computed using the Tools 6 Discrete distribution tools menu of EasyReg International, the free econometrics software package developed by the author. EasyReg International can be downloaded from web page http://econ.la.psu.edu/~hbierens/EASYREG.HTM 4. Note that the latter phrase is superfluous, because d reads: every element of is included in , which is clearly a true statement, and i d is true because i d i^ ' . 5. 6. 7. 8. Also called a Field. Also called a & Field, or a Borel Field. This section may be skipped. See also Appendix 1.A.
9. In the sequel we will denote the probability of an event involving random variables or vectors X as P(expression involving X), without referring to the corresponding set in . For example, for random variables X and Y defined on a common probability space { , , P} the short-hand notation P(X > Y) should be interpreted as P ({0 : X() > Y()}) . 10. See also Appendix 1.A.
67
Chapter 2 Borel Measurability, Integration, and Mathematical Expectations
2.1.
Introduction Consider the following situation: You are sitting in a bar next to a guy who proposes to
play the following game. He will roll a dice, and pay you a dollar per dot. However, you have to pay him an amount y up-front each time he rolls the dice. The question is: which amount y should you pay him in order for both of you to play even if this game is played indefinitely? Let X be the amount you win in a single play. Then in the long run you will receive X = 1 dollar in 1 out of 6 times, X = 2 dollar in 1 out of 6 times, up to X = 6 dollar in 1 out of 6 times. Thus, in average you will receive (1+2+...+6 )/6 = 3.5 dollar per game, hence the answer is: y = 3.5. Clearly, X is a random variable: X() ' 'j'1 j.I( 0 {j}) , where here and in the sequel
6
I(.) denotes the indicator function: I(true) ' 1 , I(false) ' 0 . This random variable is defined on the probability space {, ,P}, where ={1,2,3,4,5,6}, is the -algebra of all subsets of , and P({}) = 1/6 for each 0 . Moreover, y ' 'j'1 j/6 ' 'j'1 jP({j}) . This amount y is called the mathematical expectation of X, and is
6 6
denoted by E(X). More generally, if X is the outcome of a game with pay-off function g(X), where X is
discrete: pj ' P[X ' xj ] > 0 with 'j'1pj ' 1 (n is possibly infinite), and if this game is
n
68
repeated indefinitely, then the average pay-off will be y ' E[g(X)] ' 'j'1g(xj)pj .
n
(2.1)
Some computer programming languages, such as Fortran, Visual Basic, C++, etc., have a build-in function which generates uniformly distributed random numbers between zero and one. Now suppose that the guy next to you at the bar pulls out his laptop computer, and proposes to generate random numbers and pay you X dollar per game if the random number involved is X, provided you pay him an amount y up front each time. The question is again: which amount y should you pay him in order for both of you to play even if this game is played indefinitely? Since the random variable X involved is uniformly distributed on [0,1], it has distribution function F(x) ' 0 for x # 0 , F(x) ' x for 0 < x < 1 , F(x) ' 1 for x $ 1 , with density function f(x) ' F )(x) ' I(0 < x < 1) . More formally, X ' X() ' is a non-negative random variable defined on the probability space { , , P} , where = [0,1], = [0,1]1B, i.e., the -algebra of all Borel sets in [0,1], and P is the Lebesgue measure on [0,1]. In order to determine y in this case, let X(() ' 'j'1[inf0(b
m
j&1,b j]
X()]I( 0 (bj&1,bj]) ' 'j'1bj&1I( 0 (bj&1,bj]) ,

m
where b0 = 0 and bm =1. Clearly, 0 # X( # X with probability 1, and similarly to the dice game the amount y involved will be greater or equal to 'j'1bj&1P((bj&1,bj]) ' 'j'1bj&1(b j&bj&1) .
m m m
Taking the supremum over all possible partitions ^j'1(bj&1,bj] of (0,1] then yields the integral
69 y ' E(X) ' m0

1
xdx ' 1/2 .
(2.2)
More generally, if X is the outcome of a game with pay-off function g(X), where X has an absolutely continuous distribution with density f(x), then m
4
y ' E[g(X)] '
g(x)f(x)dx .
(2.3)
&4
Now two questions arise. First, under what conditions is g(X) a well-defined random variable? Second, how do we determine E(X) if the distribution of X is neither discrete nor absolutely continuous?
2.2.
Borel measurability Let g be a real function and let X be a random variable defined on the probability space
{ , , P} . In order for g(X) to be a random variable, we must have that: For all Borel sets B, { 0 : g(X()) 0 B} 0 . (2.4)
It is possible to construct a real function g and a random variable X for which this is not the case. But if For all Borel sets B, AB ' {x 0 : g(x) 0 B} is a Borel set itself , then (2.4) is clearly satisfied, because then for any Borel set B, and AB defined in (2.5), {0 : g(X()) 0 B} ' {0 : X() 0 AB} 0 . Moreover, if (2.5) is not satisfied, in the sense that there exists a Borel set B for which AB is not a Borel set itself, then it is possible to construct a random variable X such that the set {0 : g(X()) 0 B} ' {0 : X() 0 AB} , (2.5)
70 hence for such a random variable X, g(X) is not a random variable itself.1 Thus, g(X) is guaranteed to be a random variable if and only if (2.5) is satisfied. Such real functions g(x) are called Borel measurable:
Definition 2.1: A real function g is Borel measurable if and only if for all Borel sets B in the sets AB = {x0 : g(x) 0 B} are Borel sets in . Similarly, a real function g on k is Borel measurable if and only if for all Borel sets B in the sets AB = {x0k : g(x) 0 B} are Borel sets in k .
However, we do not need to verify condition (2.5) for all Borel sets. It suffices to verify it for Borel sets of the type (&4 , y] , y 0 , only:
Theorem 2.1: A real function g on k is Borel measurable if and only if for all y 0 the sets Ay = {x0k : g(x) # y} are Borel sets in k .
Proof: Let D be the collection of all Borel sets B in for which the sets {x0k : g(x) 0 B} are Borel sets in k , including the Borel sets of the type (&4 , y] , y 0 . Then D contains the collection of all intervals of the type (&4 , y] , y 0 . The smallest algebra containing the collection { (&4 , y] , y 0 } is just the Euclidean Borel field B = ({ (&4 , y] , y 0 }), hence if D is a -algebra then B d D. But D is a collection of Borel
71 sets, hence D d B. Thus, if D is a -algebra then B =D. The proof that D is a -algebra is left as an exercise. Q.E.D. The simplest Borel measurable function is the simple function:
Definition 2.2: A real function g on k is called a simple function if it takes the form g(x) ' 'j'1ajI(x 0 Bj) , with m < 4 , aj 0 , where the Bj s are disjoint Borel sets in k .
m
Without loss of generality we may assume that the disjoint Borel sets Bj s form a partition of k : ^j'1Bj ' k , because if not, then let g(x) = 'j'1 ajI(x 0 Bj) , with Bm%1 ' k \(^j'1Bj)
m m%1 m
and am+1 = 0. Moreover, without loss of generality we may assume that the aj s are all different. For example, if g(x) = 'j'1 ajI(x 0 Bj) and am ' am%1 then g(x) ' 'j'1ajI(x 0 Bj ) , where
m%1 m (
Bj ' Bj for j = 1,...,m-1 and Bm ' Bm^Bm%1 .

( (
Theorem 2.1 can be used to prove that:
Theorem 2.2: Simple functions are Borel measurable.
Proof: Let g(x) ' 'j'1ajI(x 0 Bj) be a simple function on k . For arbitrary y 0 ,
m
{x0k: g(x) # y} ' {x0k: 'j'1ajI(x 0 Bj) # y} '

m
aj # y
^ Bj ,
which is a finite union of Borel sets and therefore a Borel set itself. Since y was arbitrary, it follows from Theorem 2.1 that g is Borel measurable. Q.E.D.
72 Theorem 2.3: If f(x) and g(x) are simple functions, then so are f(x) + g(x), f(x)-g(x), and f(x).g(x). If in addition g(x) 0 for all x, then f(x)/g(x) is a simple function.
Proof: Exercise Theorem 2.1 can also be used to prove:
Theorem 2.4: Let gj(x) , j ' 1,2,3,...., be a sequence of Borel measurable functions. Then (a) f1,n(x) ' min{g1(x) , ..... , gn(x)} and f2,n(x) ' max{g1(x) , ..... , g n(x)} are Borel
measurable, (b) (c) (d) f1(x) ' infn$1gn(x) and f2(x) ' supn$1g n(x) are Borel measurable, h1(x) ' liminfn64gn(x) and h2(x) ' limsupn64g n(x) are Borel measurable, if g(x) ' limn64gn(x) exists, then g is Borel measurable.
Proof: First, note that the min, max, inf, sup, liminf, limsup, and lim operations are taken pointwise in x. I will only prove the min, inf and liminf cases, for Borel measurable real functions on . Again, let y 0 be arbitrary. Then, (a) (b) (c) {x0: f1,n(x) # y} ' ^j'1{x0: gj(x) # y} 0 B.
n
{x0: f1(x) # y} ' ^j'1{x0: gj(x) # y} 0 B.

4
{x0: h1(x) # y} ' _n'1^j'n{x0: gj(x) # y} 0 B.

4 4
The max, sup, limsup and lim cases are left as exercises. Q.E.D. Since continuous functions can be written as a pointwise limit of step functions, and step
73 functions with a finite number of steps are simple functions, it follows from Theorems 2.1 and 2.4(d) that:
Theorem 2.5: Continuous real functions are Borel measurable.
Proof: Let g be a continuous function on . Define for natural numbers n, gn(x) ' g(x) if -n < x # n, gn(x) ' 0 elsewhere. Next, define for j = 0,...,m-1 and m = 1,2,..., B( j,m,n) ' (&n % 2n.j/m,&n % 2(j%1)n/m] . Then the Bj(m,n) s are disjoint intervals such that ^j'0 Bj(m,n) ' (&n,n], hence the function
m&1
gn,m(x) ' 'j'0 infx

m&1
(0B( j,m,n)
g(x() I x 0 B( j,m,n)
is a step function with a finite number of steps, and thus a simple function. Since trivially g(x) ' limn64gn(x) pointwise in x, g(x) is Borel measurable if the functions gn(x) are Borel measurable [see Theorem 2.4(d)]. Similarly, the functions gn(x) are Borel measurable if for arbitrary fixed n, gn(x) ' limm64gn,m(x) pointwise in x, because the gn,m(x) s are simple functions and thus Borel measurable. To prove gn(x) ' limm64gn,m(x) , choose an arbitrary fixed x and choose n > |x|. Then there exists a sequence of indices jn,m such that x 0 B( jn,m,m,n) for all m, hence 0 # gn(x) & gn,m(x) # g(x) & infx
(0B( jn,m,m,n)
g(x() # sup*x&x
(*#2n/m
*g(x) & g(x()* 6 0
as m 6 4 . The latter result follows from the continuity of g(x). Q.E.D. Next, I will show that real functions are Borel measurable if and only if they are limits of simple functions, in two steps:
74 Theorem 2.6: A nonnegative real function g(x) is Borel measurable if and only if there exists a non-decreasing sequence gn(x) of nonnegative simple functions such that pointwise in x, 0 # gn(x) # g(x) , and limn64gn(x) ' g(x).
Proof: The if case follows straightforwardly from Theorems 2.2 and 2.4. For proving the only if case, let for 1 # m # n2n, gn(x) ' (m&1)/2n if (m&1)/2n # g(x) < m/2n , g n(x) ' n otherwise. Then gn(x) is a sequence of simple functions, satisfying 0 # gn(x) # g(x) , and limn64gn(x) ' g(x), pointwise in x. Q.E.D. Every real function g(x) can be written as a difference of two non-negative functions: g(x) ' g%(x) & g&(x) , where g%(x) ' max{g(x),0} , g&(x) ' max{&g(x),0} . Moreover, if g is Borel measurable, then so are g% and g& in (2.6). It follows therefore straightforwardly from (2.6) and Theorems 2.3 and 2.6 that: (2.6)
Theorem 2.7: A real function g(x) is Borel measurable if an only if it is the limit of a sequence of simple functions.
Proof: Exercise.
Using Theorem 2.7, Theorem 2.3 can now be generalized to:
Theorem 2.8: If f(x) and g(x) are Borel measurable functions, then so are f(x) + g(x), f(x)-g(x),
75 and f(x).g(x). Moreover, if g(x) 0 for all x, then f(x)/g(x) is a Borel measurable function.
Proof: Exercise
2.3.
Integrals of Borel measurable functions with respect to a probability measure If g is a step function on (0,1], say g(x) = 'j'1ajI(x 0 (bj,bj%1]) , where b0 = 0 and bm+1 =
m
1, then the Riemann integral of g over (0,1] is defined as: m0

1
g(x)dx ' 'j'1aj(bj%1&b j) ' 'j'1aj((b j , bj%1]) ,

m m
where is the uniform probability measure on (0,1]. Mimicking this results for simple functions and more general probability measures , we can define the integral of a simple function with respect to a probability measure as follows:
Definition 2.3: Let be a probability measure on { k ,Bk}, and let g(x) = 'j'1ajI(x0Bj) be a
m
simple function on k . Then the integral of g with respect to is defined as g(x)d(x) ' 'j'1aj(Bj) . 2 m
def. m
For non-negative continuous real functions g on (0,1], the Riemann integral of g over (0,1] is defined as m0
1
g(x)dx ' sup
0 0#g(#gm
g((x)dx ,
where the supremum is taken over all step functions g( satisfying 0 # g((x) # g(x) for all x in (0,1]. Again, we may mimick this result for non-negative Borel measurable functions g and
76 general probability measures :
Definition 2.4: Let be a probability measure on { k ,Bk}, and let g(x) be a non-negative Borel measurable function on k . Then the integral of g with respect to is defined as: g(x)d(x) ' m
def. 0#g(#gm
sup g((x)d(x) ,
where the supremum is taken over all simple functions g( satisfying 0 # g((x) # g(x) for all x in a Borel set B with (B) = 1.
Using the decomposition (2.6), we can now define the integral of an arbitrary Borel measurable function with respect to a probability measure:
Definition 2.5: Let be a probability measure on { k , Bk}, and let g(x) be a Borel measurable function on k . Then the integral of g with respect to is defined as: g(x)d(x) ' g%(x)d(x) & g&(x)d(x) , m m m (2.7)
where g%(x) ' max{g(x),0} , g&(x) ' max{&g(x),0} , provided that at least one of the integrals at the right hand side of (2.7) is finite.3
Definition 2.6: The integral of a Borel measurable function g with respect to a probability measure over a Borel set A is defined as mA
def.
g(x)d(x) '
I(x0A)g(x)d(x) . m
77 All the well-known properties of Riemann integrals carry over to these new integrals. In particular:
Theorem 2.9: Let f(x) and g(x) be Borel measurable functions on k , let be a probability measure on { k ,Bk}, and let A be a Borel set in k . Then (a) (b) (c) (d) (e) (f) (g) mA (g(x) % f(x))d(x) ' g(x)d(x) % f(x)d(x) . mA mA m mA
4 ^j'1A j
For disjoint Borel sets Aj in k , If g(x) $ 0 for all x in A, then
g(x)d(x) ' 'j'1

4
mA j
g(x)d(x) .
g(x)d(x) $ 0 . g(x)d(x) $ mA f(x)d(x) .
If g(x) $ f(x) for all x in A, then
mA
g(x)d(x) # |g(x)|d(x) . /0mA /0 mA If (A) = 0, then g(x)d(x) ' 0 . mA If |g(x)|d(x) < 4 and limn64(An) ' 0 for a sequence of Borel sets An then m
limn64 g(x)d(x) = 0. mA n
Proofs of (a)-(f): Exercise. Proof of (g): Without loss of generality we may assume that g(x) $ 0. Let Ck ' {x0: k # g(x) < k%1} and Bm ' {x0: g(x) $ m} ' ^k'mCk .
4
Then g(x)d(x) ' j g(x)d(x) < 4 , m Ck k'0 m

4
hence g(x)d(x) ' j g(x)d(x) 6 0 for m 6 4 . mB m Ck k'm m

4
(2.8)
Therefore,
78 mA n g(x)d(x) ' # g(x)d(x) % g(x)d(x) mA n_B m mA n_(\B m) g(x)d(x) % m(An) , mB m
hence for fixed m, g(x)d(x) . limsupn64 g(x)d(x) # mA n mB m Letting m 6 4 , part (g) of Theorem 2.9 follows from (2.8). Q.E.D. Moreover, there are two important theorems involving limits of a sequence of Borel measurable functions and their integrals, namely the monotone convergence theorem and the dominated convergence theorem:
Theorem 2.10: (Monotone convergence) Let gn be a non-decreasing sequence of non-negative Borel measurable functions on k , i.e., for any fixed x 0 k , 0 # gn(x) # gn%1(x) for n = 1,2,3,..., and let be a probability measure on { k , Bk}. Then limn64 gn(x)d(x) ' limn64gn(x)d(x) . m m
Proof: First, observe from Theorem 2.9(d) and the monotonicity of gn that gn(x)d(x) m is monotonic non-decreasing, and that therefore limn64 gn(x)d(x) exists (but may be infinite), m and g(x) ' limn64gn(x) exists (but may be infinite), and is Borel measurable. Moreover, since for x 0 k , gn(x) # g(x), it follows easily from Theorem 2.9(d) that gn(x)d(x) # g(x)d(x) , m m hence limn64 gn(x)d(x) # g(x)d(x) . m m Thus, it remains to be shown that
79 limn64 gn(x)d(x) $ g(x)d(x) . m m It follows from the definition on the integral g(x)d(x) that (2.9) is true if for any simple m function f(x) with 0 # f(x) # g(x), limn64 gn(x)d(x) $ f(x)d(x) . m m (2.10) (2.9)
Given such a simple function f(x), let for arbitrary g > 0, An ' {x0k : g n(x) $ (1&g)f(x)} , and let sup xf(x) ' M . Note that, since f(x) is simple, M < 4. Moreover, note that limn64(k \ An) ' limn64 {x0k : g n(x) # (1&g)f(x)} ' 0 . Furthermore, observe that g (x)d(x) $ g (x)d(x) $ (1&g) f(x)d(x) m n mA n n mA n ' (1&g) f(x)d(x) & (1&g) f(x)d(x) $ (1&g) f(x)d(x) & (1&g) M (k\An) . m mk\A n m It follows now from (2.11) and (2.12) that for arbitrary g > 0 , limn64 gn(x)d(x) $ m (1&g) f(x)d(x) , which implies (2.10). Combining (2.9) and (2.10), the theorem follows. Q.E.D. m (2.12) (2.11)
Theorem 2.11: (Dominated convergence) Let gn be sequence of Borel measurable functions on k such that pointwise in x, g(x) ' limn64gn(x) , and let g (x) ' supn$1|gn(x)|. If g (x)d(x) < 4 , where is a probability measure on { k , Bk}, then m limn64 gn(x)d(x) ' g(x)d(x) . m m
Proof: Let fn(x) ' g (x) & supm$ng m(x) . Then fn(x) is non-decreasing and non-negative, and limn64fn(x) = g (x) & g(x) . Thus it follows from the condition g (x)d(x) < 4 and m
80 Theorems 2.9(a,d)-2.10 that g(x)d(x) ' limn64 supm$ngm(x)d(x) $ limn64supm$n gm(x)d(x) m m m ' limsupn64 g n(x)d(x) . m
(2.13)
(x) % infm$ng m(x) . Then hn(x) is non-decreasing and non-negative, and Next, let hn(x) ' g limn64hn(x) = g (x) % g(x) . Thus it follows again from the condition Theorems 2.9(a,d)-2.10 that g(x)d(x) ' limn64 infm$ngm(x)d(x) # limn64infm$n g m(x)d(x) m m m ' liminfn64 g n(x)d(x) . m The theorem now follows from (2.13) and (2.14). Q.E.D. In the statistical and econometric literature you will encounter integrals of the form mA g(x)dF(x) , where F is a distribution function. Since each distribution function F(x) on k is g (x)d(x) < 4 and m
(2.14)
uniquely associated with a probability measure on Bk, one should interpret these integrals as mA
def.
g(x)dF(x) '
mA
g(x)d(x) ,
(2.15)
where is the probability measure on Bk associated with F, g is a Borel measurable function on k , and A is a Borel set in k .
81 2.4. General measurability, and integrals of random variables with respect to probability measures All the definitions and results in the previous sections carry over to mappings X: 6 , where is a nonempty set, with a -algebra of subsets of . Recall that X is a random variable defined on a probability space {, ,P} if for all Borel sets B in , { 0 : X() 0 B} 0 . Moreover, recall that it suffices to verify this condition for Borel sets of the type By = (&4 , y] , y 0 . In this section I will list these generalizations, with all random variables involved defined on a common probability space {, ,P}.
Definition 2.7: A random variable X is called simple if it takes the form X() ' 'j'1bjI( 0 Aj) , with m < 4 , bj 0 , where the Ajs are disjoint sets in .
m
Compare Definition 2.2. (Verify similarly to Theorem 2.2 that a simple random variable is indeed a random variable.) Again, we may assume without loss of generality that the bjs are all different. For example, if X has a hypergeometric or binomial distribution, then X is a simple random variable.
Theorem 2.12: If X and Y are simple random variables, then so are X+Y, X-Y and X.Y. If in addition Y is non-zero with probability 1, then X/Y is a simple random variable.
Proof: Similar to Theorem 2.3.
82 Theorem 2.13: Let Xj be a sequence of random variables. Then max1#j#n Xj, min1#j#n Xj, supn$1 Xn, infn$1 Xn, limsupn64 Xn, and liminfn64 Xn are random variables. If limn64 Xn() ' X() for all in a set A in with P(A) = 1, then X is a random variable.
Theorem 2.14: A mapping X: 6 is a random variable if and only if there exists a sequence Xn of simple random variables such that limn64 Xn() ' X() for all in a set A in with P(A) = 1.
Similarly to Definitions 2.3, 2.4, 2.5 and 2.6, we may define integrals of a random variable X with respect to the probability measure P as follows, in four steps.
Definition 2.8: Let X be a simple random variable: X() ' 'j'1bjI( 0 Aj) , say. Then the
m
integral of X with respect of P is defined as
X()dP() ' 'j'1bjP(Aj) . 4 m

def. m
Definition 2.9: Let X be a non-negative random variable (with probability 1). Then the integral of X with respect of P is defined as X()dP() ' sup0#X #X X()(dP() , where the ( m m
def.
supremum is taken over all simple random variables X* satisfying 0 # X( # X with probability 1.
83 Definition 2.10: Let X be a random variable. Then the integral of X with respect of P is defined as X()dP() ' X%()dP() & X&()dP() , where X% = max{X,0} and X& = m m m
def.
max{&X,0} , provided that at least one of the latter two integrals is finite.
Definition 2.11: The integral of a random variable X with respect to a probability measure P over a set A in is defined as mA X()dP() '
def.
I( 0 A)X()dP() . m
Theorem 2.15: Let X and Y be random variables, and let A be a set in . Then (a) (b) (c) (d) (e) (f) (g) mA (X() % Y())dP() ' X()dP() % Y()dP() . mA mA m
4 ^j'1A j
For disjoint sets Aj in ,
X()dP() ' 'j'1

4
If X () $ 0 for all in A, then
mA
mA j
X()dP() .
X()dP() $ 0 . mA X()dP() $ mA Y()dP() .
If X () $ Y () for all in A, then X()dP() # |X()|dP() . /0mA /0 mA If P(A) = 0, then X()dP() ' 0 . mA If
|X()|dP() < 4 and for a sequence of sets An in , limn64P(An) ' 0 , then m
limn64 X()dP() ' 0 . mA n
Proof: Similar to Theorem 2.9. Also the monotone and dominated convergence theorems carry over:
Theorem 2.16: Let Xn be a monotonic non-decreasing sequence of non-negative random variables defined on the probability space {, ,P}, i.e., there exists a set A 0 with P(A) =
84 1 such that for all 0 A , 0 # Xn() # Xn%1() , n ' 1,2,3,.... Then limn64 Xn()dP() ' limn64Xn()dP() . m m
Theorem 2.17: Let Xn be a sequence of random variables defined on the probability space {, ,P} such that for all in a set A 0 with P(A) ' 1 , Y() ' limn64Xn() . Let ' sup X . If X n$1 n ()dP() < 4 then lim X Y()dP() . n64 Xn()dP() ' m m m
Proof: Similar to Theorem 2.11. Finally, note that the integral of a random variable with respect to the corresponding probability measure P is related to the definition of the integral of a Borel measurable function with respect to a probability measure :
Theorem 2.18: Let X be the probability measure induced by the random variable X. Then X()dP() ' xd X(x) . Moreover, if g is a Borel measurable real function on k, and X is a m m k-dimensional random vector with induced probability measure X , then g(X())dP() = m g(x)d X(x) . Furthermore, denoting in the latter case Y= g(X), with Y the probability m measure induced by Y, we have Y()dP() ' g(X())dP() ' g(x)d X(x) ' yd Y(y) . m m m m
Proof: Let X be a simple random variable: X() ' 'j'1bjI( 0 Aj) , say, and recall that
m
without loss of generality we may assume that the bj s are all different. Each of the disjoint sets
85 Aj are associated with disjoint Borel sets Bj such that Aj ' {0: X() 0 Bj} (for example, let Bj = {bj}). Then X()dP() ' 'j'1bjP(Aj) ' 'j'1bj X(Bj) ' g((x)d X(x) , m m
m m
where g((x) ' 'j'1bjI(x 0 Bj) is a simple function such that

m
g((X()) ' 'j'1bjI(X() 0 Bj) ' 'j'1bjI( 0 Aj) ' X() .

m m
Therefore, in this case the Borel set C ' {x: g((x) x} has X measure zero: X(C) ' 0, and consequently, X()dP() ' g (x)d X(x) % g((x)d X(x) ' xd X(x) ' xd X(x) . m m\C ( mC m\C m The rest of the proof is left as an exercise. Q.E.D. (2.16)
2.5.
Mathematical expectation With these new integrals introduced, we can now answer the second question stated at the
end of the introduction: How to define the mathematical expectation if the distribution of X is neither discrete nor absolutely continuous:
Definition 2.12: The mathematical expectation of a random variable X is defined as: E(X) ' X()dP() , or equivalently as: E(X) ' xdF(x) [cf. (2.15)], where F is the m m
distribution function of X, provided that the integrals involved are defined. Similarly, if g(x) is a Borel measurable function on k and X is a random vector in k then equivalently,
86 E[g(X)] ' g(X())dP() = g(x)dF(x) , provided that the integrals involved are defined. m m
Note that the latter part of Definition 2.12 covers both examples (2.1) and (2.3). As motivated in the introduction, the mathematical expectation E[g(X)] may be interpreted as the limit of the average pay-off of a repeated game with pay-off function g. This is related to the law of large numbers which we will discuss later, in Chapter 7: If X1, X2, X3,.. ...... is a sequence of independent random variables or vectors each distributed the same as X, and g is a Borel measurable function such that E[|g(X)|] < 4 , then P limn64(1/n)'j'1g(Xj) ' E[g(X)] =
n
1. There are a few important special cases of the function g, in particular the variance of X, which measures the variation of X around its expectation E(X), and the covariance of a pair of random variables X and Y, which measures how X and Y fluctuate together around their expectations:
Definition 2.13: The ms moment (m = 1,2,3,.... ) of a random variable X is defined as: E(Xm), and the ms central moment of X is defined by E(*X& x*m) , where x ' E(X) . The second central moment is called the variance of X : var(X) ' E[(X & x)2 ] ' x , say. The covariance of a pair (X,Y) of random variables is defined as: cov(X,Y) ' E[(X & x)(Y & y )] , where x is the same as before, and y ' E(Y) . The correlation (coefficient) of a pair (X,Y) of
2
87 random variables is defined as: corr(X,Y) ' say. cov(X,Y) var(X) var(Y) ' (X,Y),
The correlation coefficient measures the extent to which Y can be approximated by a linear function of X, and vice versa. In particular, If exactly Y ' % X then corr(X,Y) ' 1 if > 0 , corr(X,Y) ' &1 if < 0 . Moreover, (2.17)
Definition 2.14: Random variables X and Y said to be uncorrelated if cov(X,Y) = 0. A sequence of random variables Xj is uncorrelated if for all i j, Xi and Xj are uncorrelated.
Furthermore, it is easy to verify that
Theorem 2.19: If X1,.....,Xn are uncorrelated, then var 'j'1Xj ' 'j'1var(Xj) .
n n
Proof: Exercise.
88 2.6. Some useful inequalities involving mathematical expectations There are a few inequalities that will prove to be useful later on, in particular Chebishevs inequality, Holders inequality, Liapounovs inequality, and Jensens inequality.
2.6.1. Chebishevs inequality Let X be a non-negative random variable with distribution function F(x), and let (x) be a monotonic increasing non-negative Borel measurable function on [0,4). Then for arbitrary g > 0, E[(X)] ' $ (x)dF(x) ' (x)dF(x) % (x)dF(x) m m{(x)>(g)} m{(x)#(g)} (2.18)
(x)dF(x) $ (g) dF(x) ' (g) dF(x) ' (g)(1 & F(g)) , m{(x)>(g)} m{(x)>(g)} m{x>g}
hence P(X > g) ' 1 & F(g) # E[(X)] . (g)
(2.19)
In particular, it follows from (2.19) that for a random variable Y with expected value y ' E(Y) and variance y , P 0: *Y()& y* > y / g
2 2
# g.
(2.20)
2.6.2
Holders inequality Holders inequality is based on the fact that ln(x) is a concave function on (0,4): for 0 < a
< b, and 0 # # 1, ln(a % (1&)b) $ ln(a) % (1&)ln(b) , hence a % (1&)b $ a b 1& . (2.21)
89 Now let X and Y be random variables, and put a ' *X*p / E(*X*p) , b ' *Y*q / E(*Y*q) , where p > 1, and p &1 % q &1 ' 1 . Then it follows from (2.21), with = 1/p and 1- = 1/q , that p &1 *X*p E(*X* )
p
% q &1
*Y*q E(*Y* )
q
*X*p E(*X* )
p
1/p
*Y*q E(*Y* )
q
1/q
'
*X.Y* E(*X* )
p 1/p
E(*Y*q) 1/q
Taking expectations yields Holders inequality: E(*X.Y*) # E(*X*p) 1/p E(*Y*q) 1/q , where p > 1 and 1 1 % ' 1. p q
(2.22)
For the case p = q = 2 inequality (2.22) reads E(*X.Y*) # E(X 2) E(Y 2) , which is known as the Cauchy-Schwartz inequality.
2.6.3
Liapounovs inequality Liapounovs inequality follows from Holders inequality (2.22) by replacing Y with 1: E(*X*) # E(*X*p) 1/p , where p $ 1.
2.6.4
Minkowskis inequality If for some p $ 1, E[|X|p] < 4 and E[|Y|p] < 4 then E(*X % Y*) # E(*X*p) 1/p % E(*Y*p) 1/p . (2.23)
This inequality is due to Minkowski. For p = 1 the result is trivial. Therefore, let p > 1. First note that E[|X % Y|p] # E[(2.max(|X|, |Y|))p] ' 2pE[max(|X|p , |Y|p)] # 2pE[|X|p % |Y|p] < 4 , hence we may apply Liapounovs inequality: E(*X % Y*) # E(*X % Y*p) 1/p . Next, observe that (2.24)
90 E(*X % Y*p) ' E(*X % Y*p&1*X % Y*) # E(*X % Y*p&1*X*) % E(*X % Y*p&1*Y*) . Let q = p /(p-1). Since 1/q + 1/p = 1 it follows from Holders inequality that E(*X % Y*p&1*X*) # E(*X % Y*(p&1)q) 1/q E(|X|p) 1/p # E(*X % Y*p) 1&1/p E(|X|p) 1/p , and similarly, E(*X % Y*p&1*Y*) # E(*X % Y*p) 1&1/p E(|Y|p) 1/p . Combining (2.24), (2.25), (2.26) and (2.27), Minkowskis inequality (2.23) follows. (2.27) (2.26) (2.25)
2.6.5
Jensens inequality A real function (x) on is called convex if for all a, b 0 and 0 # # 1, (a % (1&)b) # (a) % (1&)(b) .
It follows by induction that then also

n 'j'1ja j
# j j(aj) , where j > 0 for j ' 1,..,n , and j j ' 1.

n n j'1 j'1
(2.28)
Consequently, it follows from (2.28) that for a simple random variable X, (E(X)) # E((X)) for all convex real functions on . (2.29)
This is Jensens inequality. Since (2.29) holds for simple random variables, it holds for all random variables. Similarly we have (E(X)) $ E((X)) for all concave real functions on .
2.7.
Expectations of products of independent random variables Let X and Y be independent random variables, and let f and g be Borel measurable
functions on . I will show now at then E[f(X)g(Y)] ' (E[f(X)])(E[g(Y)]) . (2.30)
91 In general (2.30) does not hold, although there are cases where (2.30) holds for dependent X and Y. As an example of a case where (2.30) does not hold, let X = U0.U1 and Y = U0.U2, where U0, U1 and U2 are independent uniformly [0,1] distributed, and let f(x) ' x , g(x) ' x . The joint density of U0, U1 and U2 is: h(u0 , u1 , u2) ' 1 if (u0 , u1 , u2)T 0 [0,1][0,1][0,1] , h(u0 , u1 , u2) ' 0 elsewhere , hence E[f(X)g(Y)] ' E[X.Y] ' E[U0 U1U2] '
2 1 2 u u u du du du m0 m0 m0 0 1 2 0 1 2 1 1
'
1 2 1 1 u0 du0 u1du1 u2du2 m0 m0 m0
' (1/3)(1/2)(1/2) ' 1/12 , whereas E[f(X)] ' E[X] ' m0 m0 m0

1 1 1
u0u1du0du1du2 '
m0
u0du0
m0
u1du1
m0
du2 ' 1/4 ,
and similarly, E[g(Y)] = E[Y] = 1/4. As an example of dependent random variables X and Y for which (2.30) holds, let now X = U0(U1 & 0.5) and Y = U0(U2 & 0.5) , where U0 , U1 , and U2 are the same as before, and again f(x) ' x , g(x) ' x . Then it is easy to show that E[X.Y] ' E[X] ' E[Y] ' 0 . In order to prove (2.30) for independent random variables X and Y, let f and g be simple functions: f(x) ' 'i'1iI(x 0 Ai) , g(x) ' 'j'1jI(x 0 Bj) ,
m n
where the Ais are disjoint Borel sets, and the Bjs are disjoint Borel sets. Then
92 E[f(X)g(Y)] ' E 'i'1'j'1ijI(X 0 Ai and Y 0 Bj)

m n
'
m
' 'i'1'j'1ijP { 0 : X() 0 Ai}_{ 0 : Y() 0 Bj}

n
'i'1'j'1ijI(X() 0 Ai and Y() 0 Bj) dP()

m n
' 'i'1'j'1ijP { 0 : X() 0 Ai} P { 0 : Y() 0 Bj}

m n
' 'i'1iP { 0 : X() 0 Ai} 'j'1jP { 0 : Y() 0 Bj}

m n
' E[f(X)] E[g(Y)] ,
because by the independence of X and Y, P(X 0 Ai and Y 0 Bj) ' P(X 0 Ai)P(Y 0 Bj) . From this result it follows more generally:
Theorem 2.20: Let X and Y be random vectors in p and q , respectively. Then X and Y are independent if and only if E[f(X)g(Y)] ' (E[f(X)])(E[g(Y)]) for all Borel measurable functions f and g on p and q , respectively, for which the expectations involved are defined.
This theorem implies that independent random variables are uncorrelated. The reverse, however, is in general not true. A counter example is the case I have considered before, namely X = U0(U1 & 0.5) and Y = U0(U2 & 0.5) , where U0 , U1 , and U2 are independent uniformly [0,1] distributed. In this case E[X.Y] ' E[X] ' E[Y] ' 0 , hence cov(X,Y) = 0, but X and Y are dependent, due to the common factor U0. The latter can be shown formally in different ways, but the easiest way is to verify that, for example, E[X 2.Y 2] (E[X 2])(E[Y 2]) , so that the dependence of X and Y follows from Theorem 2.20.
93 2.8. 2.8.1 Moment generating functions and characteristic functions Moment generating functions The moment generating function of a bounded random variable X , i.e., P[|X| # M] = 1 for some positive real number M < 4, is defined as the function m(t) ' E[exp(t.X)] , t 0 , where the argument t is non-random. More generally: (2.31)
Definition 2.15: The moment generating function of a random vector X in k is defined by m(t) ' E[exp(t TX)] for t 0 d k , where is the set of non-random vectors t for which the moment generating function exists and is finite.
For bounded random variables the moment generating function exists and is finite for all values of t. In particular, in the univariate bounded case we can write m(t) ' E[exp(t.X)] ' E j
4 k'0
t kX k t kE[X k] ' j . k! k! k'0

4
It is easy to verify that the j-th derivative of m(t) is: d jm(t) (dt) j ' j
4 k'j
m (j)(t) '
t k&jE[X k] t k&jE[X k] ' E[X j] % j (k&j)! (k&j)! k'j%1

4
(2.32)
hence the j-th moment of X is m (j)(0) ' E[X j] . This is the reason for calling m(t) the moment generating function. Although the moment generating function is a handy tool for computing moments of a (2.33)
94 distribution, its actual importance is due to the fact that the shape of the moment generating function in an open neighborhood of zero uniquely characterizes the distribution of a random variable. In order to show this, we need the following result.
Theorem 2.21: The distributions of two random vectors X and Y in k are the same if and only if for all bounded continuous functions on k, E[(X)] ' E[(Y)] .
Proof: I shall only prove this theorem for the case where X and Y are random variables: k = 1. Note that the only if case follows from the definition of expectation. Let F(x) be the distribution function of X and let G(y) be the distribution function of Y. Let a < b be arbitrary continuity points of F(x) and G(y), and define ' 0 (x) ' ' 1 if if x $ b, x < a, (2.34)
b&x ' if a # x < b . b&a
Clearly, (2.34) is a bounded continuous function, and therefore by assumption we have E[(X)] = E[(Y)] . Now observe from (2.34) that E[(X)] ' and E[(X)] ' Similarly, E[(Y)] ' and b&x (y)dG(y) ' G(a) % dG(x) $ G(a) m m b&a
a b
b&x (x)dF(x) ' F(a) % dF(x) $ F(a) m m b&a

a
b&x (x)dF(x) ' F(a) % dF(x) # F(b) . m m b&a

a
95 E[(X)] ' b&x (y)dG(y) ' G(a) % dG(x) # G(b) . m m b&a

a b
Combining these inequalities with E[(X)] ' E[(Y)] it follows that for arbitrary continuity points a < b of F(x) and G(y), G(a) # F(b) , F(a) # G(b) . Letting b 9 a it follows from (2.35) that F(a) ' G(a) . Q.E.D. Now assume that the random variables X and Y are discrete, and take with probability 1 the values x1 , ..... ,xn . Without loss of generality we may assume that xj = j, i.e., P[X 0 {1,2,...,n}] ' P[Y 0 {1,2,...,n}] ' 1 . Suppose that all the moments of X and Y match: For k = 1,2,3,...., E[X k] ' E[Y k] . I will show that then for an arbitrary bounded continuous function on , E[(X)] ' E[(Y)] . Denoting pj = P[X = j], qj = P[Y = j] we can write E[(X)] ' j (j)pj , E[(Y)] ' j (j)qj .
n n j'1 j'1 n&1 'k'0 kt k
(2.35)
It is always possible to construct a polynomial (t) ' 1,...,n, by solving 1 1 1 ! ! ! " 1 ! 0 1 !
such that (j) ' (j) for j =
(1) ' (2) ! (n) .
1 2 22 2n&1 1 n n 2 n n&1
n&1
Then E[(X)] ' j j k j kpj ' j k j j kpj ' j k E[X k]

n n&1 n&1 k'0 n n&1 k'0 j'1 k'0 j'1
and similarly E[(Y)] ' j k E[Y k] .

n&1 k'0
96 Hence, it follows from Theorem 2.21 that if all the corresponding moments of X and Y are the same, then the distributions of X and Y are the same. Thus if the moment generating functions of X and Y coincide on a open neighborhood of zero, and if all the moments of X and Y are finite, then it follows from (2.33) that all the corresponding moments of X and Y are the same:
Theorem 2.22: If the random variables X and Y are discrete, and take with probability 1 only a finite number of values, then the distributions of X and Y are the same if and only if the moment generating functions of X and Y coincide on an arbitrary small open neighborhood of zero.
However, this result also applies without the conditions that X and Y are discrete and take only a finite number of values, and for random vectors as well, but the proof is complicated and therefore omitted:
Theorem 2.23: If the moment generating functions mX(t) and mY(t) of the random vectors X and Y in k are defined and finite in an open neighborhood N0() = {x 0 k : 2x2 < } of the origin of k, then the distributions of X and Y are the same if and only if mX(t) ' mY(t) for all t 0 N0() .
2.8.2
Characteristic functions The disadvantage of the moment generating function is that is may not be finite in an
arbitrarily small open neighborhood of zero. For example, if X has a standard Cauchy distribution, i.e., X has density
97 f(x) ' then m(t) ' m

4
1 (1%x 2)
(2.36)
exp(t.x)f(x)dx
' 4 if t 0 , ' 1 if t ' 0 . (2.37)
&4
There are many other distributions with the same property as (2.37), hence the moment generating functions in these cases are of no use for comparing distributions. The solution to this problem is to replace t in (2.31) with i.t, where i ' &1 . The resulting function n(t) = m(i.t) is called the characteristic function of the random variable X: n(t) ' E[exp(i.t.X)] , t 0 . More generally,
Definition 2.16: The characteristic function of a random vector X in k is defined by n(t) ' E[exp(i.t TX)] , t 0 k , where the argument t is non-random.
The characteristic function is bounded, because exp(i.x) ' cos(x) % i.sin(x) . See Appendix III. Thus, the characteristic function in Definition 2.16 can be written as n(t) ' E[cos(t TX)] % i.E[sin(t TX)] , t 0 k . Note that by the dominated convergence theorem (Theorem 2.11), limt60 n(t) ' 1 ' n(0) , hence a characteristic function is always continuous in t = 0. Replacing moment generating functions with characteristic functions, Theorem 2.23 now becomes:
98 Theorem 2.24: Random variables or vectors have the same distribution if and only if their characteristic functions are identical.
The proof of this theorem is complicated, and is therefore given in Appendix 2.A at the end of this chapter. The same applies to the following useful result, which is known as the inversion formula for characteristic functions:
Theorem 2.25: Let X be a random vector in k with characteristic function n(t). If n(t) is absolutely integrable, i.e., *n(t)*dt < 4 , then the distribution of X is absolutely continuous mk exp(&i.t Tx) n(t)dt . mk
with joint density f(x) ' (2)&k
2.9. 1. 2. 3. 4.
Exercises Prove that the collection D in the proof of Theorem 2.1 is a -algebra. Prove Theorem 2.3. Prove Theorem 2.4 for the max, sup, limsup and lim cases. Complete the proof of Theorem 2.5, by proving that gn(x) ' limm64gn,m(x) pointwise in
x, and g(x) ' limn64gn(x) pointwise in x. 5. 6. 7. 8. Why is it true that if g is Borel measurable, then so are g% and g& in (2.6)? Prove Theorem 2.7. Prove Theorem 2.8. Let g(x) ' x if x is rational, g(x) ' &x if x is irrational. Prove that g(x) is Borel
measurable.
99 9. Prove parts (a)-(f) of Theorem 2.9 for simple functions g(x) ' 'i'1aiI(x 0 Bi) , f(x) ' 'j'1bjI(x 0 C j) .
n m
10.
Why can you conclude from exercise 9 that parts (a)-(f) of Theorem 2.9 hold for arbitrary
non-negative Borel measurable functions? 11. Why can you conclude from exercise 10 that Theorem 2.9 holds for arbitrary Borel
measurable functions, provided that the integrals involved are defined? 12. 13. 14. 15. From which result on probability measures does (2.11) follow? Determine for each inequality in (2.12) which part of Theorem 2.9 has been used. Why do we need the condition in Theorem 2.11 that g (x)d(x) < 4 ? m Note that we cannot generalize Theorem 2.5 to random variables, because something
missing prevents us from defining a continuous mapping X: 6 . What is missing? 16. 17. 18. Verify (2.16), and complete the proof of Theorem 2.18. Prove equality (2.2). Show that var(X) ' E(X 2) & (E(X))2 , cov(X,Y) ' E(X.Y) & (E(X))(E(Y)) , and -1 #
corr(X,Y) # 1. Hint: Derive the latter result from var(Y & X) $ 0 for all . 19. 20. 21. 22. 23 24. 25. Prove (2.17). Which parts of Theorem 2.15 have been used in (2.18)? How does (2.20) follow from (2.19)? Why does it follows from (2.28) that (2.29) holds for simple random variables? Prove Theorem 2.19. Complete the proof of Theorem 2.20 for the case p = q = 1. Let X = U0(U1 & 0.5) and Y = U0(U2 & 0.5) , where U0 , U1 , and U2 are
100 independent uniformly [0,1] distributed. Show that E[X 2.Y 2] (E[X 2])(E[Y 2]) . 26. Prove that if (2.29) holds for simple random variables, it holds for all random variables.
Hint: Use the fact that convex and concave functions are continuous (See Appendix II). 27. 28. Derive the moment generating functions of the Binomial (n,p) distribution. Use the results in exercise 27 to derive the expectation and variance of the Binomial (n,p)
distribution. 29. Show that the moment generating function of the Binomial (n,p) distribution converges
pointwise in t to the moment generating function of the Poisson () distribution if n 6 4 and p 90 such that n.p 6 . 30. Derive the characteristic function of the uniform [0,1] distribution. Is the inversion
formula for characteristic functions applicable in this case ? 31. 32. If the random variable X has characteristic function exp(i.t), what is the distribution of X? Show that the characteristic function of a random variable X is real-valued if and only if
the distribution of X is symmetric, i.e., X and !X have the same distribution. 33. Use the inversion formula for characteristic functions to show that n(t) ' exp(&*t*) is
the characteristic function of the standard Cauchy distribution [see (2.36) for the density involved]. Hints: Show first, using Exercise 32 and the inversion formula, that f(x) ' &1 and then use integration by parts. m0
4
cos(t.x) exp(&t)dt ,
101
Appendix
2.A. Uniqueness of characteristic functions In order to understand characteristic functions, you need to understand the basics of complex analysis, which is provided in Appendix III. Therefore, it is recommended to read Appendix III first. In the univariate case, Theorem 2.24 is a straightforward corollary of the following link between a probability measure and its characteristic function.
Theorem 2.A.1: Let be a probability measure on the Borel sets in with characteristic function n, and let a < b be continuity points of : ({a}) ' ({b}) ' 0 . Then 1 exp(&i.t.a)&exp(&i.t.b) ((a,b]) ' lim n(t)dt . i.t T64 2 m
&T T
(2.38)
Proof: Using the definition of characteristic function, we can write exp(&i.t.a)&exp(&i.t.b) exp(i.t(x&a))&exp(i.t.(x&b)) n(t)dt ' d(x)dt m m m i.t i.t
&T&4 T T 4
&T
(2.39) ' m M64 lim

T
&T
&M
exp(i.t(x&a))&exp(i.t.(x&b)) d(x)dt m i.t
Next, observe that /0 exp(i.t(x&a))&exp(i.t.(x&b)) d(x)/0 # exp(&i.ta)&exp(&i.t.b) ([&M,M]) /0 /0 00 m 00 i.t i.t 0 0 00&M 00
M
|exp(&i.t.a)&exp(&i.t.b)| ' |t|
2(1 & cos(t.(b&a)) t2
# b&a
102
Therefore, it follows from the bounded convergence theorem that exp(&i.t.a)&exp(&i.t.b) exp(i.t(x&a))&exp(i.t.(x&b)) n(t)dt ' lim d(x)dt m m m i.t i.t M64
&T &M M T 4 T T T M
&T
(2.40)
exp(i.t(x&a))&exp(i.t.(x&b)) exp(i.t(x&a))&exp(i.t.(x&b)) dtd(x) ' dt d(x mm i.t i.t M64 m m lim

&M&T &4 &T
The integral between square brackets can be written as exp(i.t(x&a))&exp(i.t.(x&b)) exp(i.t(x&a))&1 exp(i.t.(x&b))&1 dt ' dt & dt m m m i.t i.t i.t
&T &T T T T
&T
'
T
&T
cos(t(x&a))&1%i.sin(t(x&a)) cos(t(x&b))&1%i.sin(t(x&b)) dt & dt m m i.t i.t

&T
(2.41) ' sin(t(x&a)) sin(t(x&b)) sin(t(x&a)) sin(t(x&b)) dt & dt ' 2 dt(x&a) & 2 dt(x&b) m m m t(x&a) m t(x&b) t t
&T 0 0 T(x&a) T(x&b) T|x&a| T|x&b| T T T
&T
sin(t) sin(t) sin(t) sin(t) ' 2 dt & 2 dt ' 2sgn(x&a) dt & 2sgn(x&b) dt , m m m m t t t t
0 0 0 0
where sgn(x) = 1 if x > 0, sgn(0) = 0, and sgn(x) = !1 if x < 0. The last two integrals in (2.41) are of the form sin(t) dt ' sin(t) exp(&t.u)dudt ' sin(t)exp(&t.u)dtdu m t m m mm
0 0 0 0 0 x x 4 4x
'
du exp(&x.u) & [cos(x) % u.sin(x)] du . m 1%u 2 m 1%u 2

0 0
(2.42)
where the last equality follows from integration by parts:
103 dcos(t) x sin(t)exp(&t.u)dt ' & exp(&t.u)dt ' cos(t)exp(&t.u) 0 & u. cos(t)exp(&t.u)dt m m dt m
0 0 0 x x x
' 1 & cos(x)exp(&x.u) & u.
m
0
dsin(t) exp(&t.u)dt dt
x
' 1 & cos(x)exp(&x.u) & u.sin(x)exp(&x.u) & u 2 sin(t)exp(&t.u)dt . m

0
Clearly, the second integral at the right-hand side of (2.42) is bounded in x > 0, and converges to zero as x 64. The first integral at the right-hand side of (2.42) is m 1%u 2 du
0 4
'
darctan(u) ' arctan(4) ' /2 . m

0
Thus, the integral (2.42) is bounded, hence so is (2.41), and exp(i.t(x&a))&exp(i.t.(x&b)) dt' [sgn(x&a) & sgn(x&b)] i.t T64 m
&T T
lim
(2.43)
It follows now from (2.39) , (2.40), (2.43) and the dominated convergence theorem that 1 exp(&i.t.a)&exp(&i.t.b) 1 n(t)dt ' [sgn(x&a) & sgn(x&b)]d(x) m i.t 2m T64 2
&T T
lim
(2.44)
1 1 ' ((a,b)) % ({a}) % ({b}) . 2 2 The last equality in (2.44) follow from the fact that 0 if x < a or x > b , sgn(x&a) & sgn(x&b) ' 1 if x ' a or x ' b , 2 if a < x < b . The result (2.38) now follows from (2.44) and the condition ({a}) ' ({b}) ' 0 . Q.E.D. Note that (2.38) also reads as
104 1 exp(&i.t.a)&exp(&i.t.b) F(b) & F(a) ' lim n(t)dt , i.t T64 2 m
&T T
(2.45)
where F is the distribution function corresponding to the probability measure . Next, suppose that n is absolutely integrable: written as 1 exp(&i.t.a)&exp(&i.t.b) F(b) & F(a) ' n(t)dt , 2 m i.t
&4 4
*n(t)*dt < 4 . Then (2.45) can be m&4

4
and it follows from the dominated convergence theorem that F(b) & F(a) 1 1&exp(&i.t.(b&a)) F (a) ' lim ' lim exp(&i.t.a)n(t)dt b&a 2 m b9a i.t.(b&a) b 9a
) &4 4
1 exp(&i.t.a)n(t)dt . 2 m &4 This proves Theorem 2.25 for the univariate case. ' In the multivariate case Theorem 2.A.1 becomes:
Theorem 2.A.2: Let be a probability measure on the Borel sets in k with characteristic function n. Let B ' j'1(aj ,bj] , where aj < bj for j = 1,2,...,k, and let MB be the border of B, i.e., MB ' {j'1[aj ,bj]}\{j'1(aj ,bj)}. If (MB) ' 0 then m k
k j'1 k k k
(B) ' lim .... lim

T164 T k64
exp(&i.tj.aj)&exp(&i.tj.bj) i.2t j
n(t)dt ,
(2.46)
k j'1(&T j,T j)
where t = (t1,...,tk)T.
This result proves Theorem 2.24 for the general case.
105 Moreover, if *n(t)*dt < 4 then (2.46) becomes mk (B) ' mk j'1
k
k
exp(&i.t j.aj)&exp(&i.t j.bj) i.2tj
n(t)dt ,
and by the dominated convergence theorem we may take partial derivatives inside the integral: Mk(B) 1 ' exp(&i.t Ta) n(t)dt , km Ma1.....Mak (2) k
(2.47)
where a = (a1,...,ak)T. The latter is just the density corresponding to F in point a. Thus, (2.47) proves Theorem 2.25.
Endnotes 1. 2. The actual construction of such a counter example is difficult, though, but not impossible.
The notation g(x)d(x) is somewhat odd, because (x) has no meaning. It would be m better to denote the integral involved by g(x)(dx) (which some authors do), where dx m represents a Borel set. The current notation, however, is the most common, and therefore adopted here too. 3. Because 4 ! 4 is not defined.
Again, the notation X()dP() is odd because P() has no meaning. Some authors use m the notation X()P(d) , where d represents a set in . The former notation is the most m common, and therefore adopted. 4.
106
Chapter 3 Conditional Expectations
3.1.
Introduction Roll a dice, and let the outcome be Y. Define the random variable X = 1 if Y is even, and
X = 0 if Y is odd. The expected value of Y is E[Y] = (1+2+3+4+5+6)/6 = 3.5. But what would the expected value of Y be if it is revealed that the outcome is even: X = 1? The latter information implies that Y is either 2, 4 or 6, with equal probabilities 1/3, hence the expected value of Y, conditional on the event X = 1, is E[Y|X=1] = (2+4+6)/3 = 4. Similarly, if it is revealed that X = 0, then Y is either 1, 3, or 5, with equal probabilities 1/3, hence the expected value of Y, conditional on the event X = 0, is E[Y|X=0] = (1+3+5)/3 = 3. Both results can be captured in a single statement: E[Y|X] ' 3%X . In this example the conditional probability of Y = y, given X = x, is1 P(Y ' y|X'x) ' ' P(Y ' y and X'x) P(X'x) (3.1)
P({y}_{2,4,6}) P({y}) 1/6 1 ' ' ' if x ' 1 and y 0 {2,4,6} P({2,4,6}) P({2,4,6}) 1/2 3 P({y}_{2,4,6}) P(i) ' ' 0 if x ' 1 and y {2,4,6} P({2,4,6}) P({2,4,6}) (3.2)
'
'
P({y}_{1,3,5}) P({y}) 1/6 1 ' ' ' if x ' 0 and y 0 {1,3,5} P({1,3,5}) P({1,3,5}) 1/2 3
107 ' P({y}_{1,3,5}) P(i) ' ' 0 if x ' 0 and y {1,3,5} P({1,3,5}) P({1,3,5})
hence 2%4%6 ' 4 if x '1 3
j yP(Y'y|X'x)
6 y'1
'
1%3%5 ' ' 3 if x '0 3
' 3 % x.
Thus in the case where both Y and X are discrete random variables, the conditional expectation E[Y|X] can be defined as E[Y|X] ' ' yp(y|X) , where p(y|x) ' P(Y'y|X'x) for P(X'x) > 0
y
A second example is where X is uniformly [0,1] distributed, and given the outcome x of X, Y is randomly drawn from the uniform [0,x] distribution. Then the distribution function F(y) of Y is: F(y) ' P(Y # y) ' P(Y # y and X # y) % P(Y # y and X > y) ' P(X # y) % P(Y # y and X > y) ' y % E[I(Y # y)I(X > y)] ' y % m0 m0 ' y %
1 x
I(z # y)x &1dz I(x > y)dx ' y % my
my m0 1 (y/x)dx ' y(1 & ln(y)) for 0# y #1 .

1
min(x,y) &1
x dz dx
Hence, the density of Y is: f(y) ' F )(y) ' &ln(y) for y 0 [0,1] , f(y) ' 0 for y [0,1] . Thus, the expected value of Y is: E[Y] ' m0
1
y(&ln(y))dy ' 1/4 . But what would the expected
value be if it is revealed that X ' x for a given number x 0 (0,1) ? The latter information
108 implies that Y is now uniformly [0,x] distributed, hence the conditional expectation involved is E[Y|X'x] ' x &1 m0
x
ydy ' x/2 .
More generally, the conditional expectation of Y given X is: E[Y|X] ' X &1 m0
X
ydy ' X/2 .
(3.3)
The latter example is a special case of a pair (Y,X) of absolutely continuously distributed random variables with joint density function f(y,x) and marginal density fx(x) . The conditional distribution function of Y given the event X 0 [x , x%] , > 0 , is: 1 f(u,v)dvdu m m
x y x%
P(Y # y* X 0 [x,x%]) '
P(Y # y and X 0 [x,x%]) ' P(X 0 [x,x%])
&4
1 f (v)dv m x
x
x%
Letting 9 0 then yields the conditional distribution function of Y given the event X = x: m
y
F(y|x) ' lim P(Y # y* X 0 [x,x%]) '

90
f(u,x)du /fx(x), provided f x(x) > 0.
&4
Note that we cannot define this conditional distribution function directly as F(y|x) ' P(Y # y and X ' x)/P(X ' x) , because for continuous random variables X, P(X = x) = 0. The conditional density of Y given the event X = x is now f(y|x) ' MF(y|x)/My = f(y,x)/fx(x), and the conditional expectation of Y given the event X = x can therefore be defined as:
109 m
4
E[Y|X'x] '
yf(y|x)dy ' g(x) , say.
&4
Plugging in X for x then yields: m

4
E[Y|X] '
yf(y|X)dy ' g(X) .
(3.4)
&4
These examples demonstrate two fundamental properties of conditional expectations. The first one is that E[Y|X] is a function of X, which can be translated as follows: Let Y and X be two random variables defined on a common probability space {,,P} , and let X be the & algebra generated by X: X ' {X &1(B) , B 0 B}, where X-1(B) is a short-hand notation for the set {0 : X() 0 B} , and B is the Euclidean Borel field. Then: Z ' E[Y|X] is measurable X , . which means that for all Borel sets B, {0: Z() 0 B} 0 X . Secondly, we have E[(Y & E[Y|X])I(X 0 B)] ' 0 for all Borel sets B . In particular in the case (3.4) we have mm
4 4 4
(3.5)
(3.6)
E[(Y & E[Y|X])I(X 0 B)] '

4 4
y & g(x) I(x0B)f(y,x)dydx

4
&4&4
'
&4 &4
m m
yf(y|x)dy I(x0B)fx(x)dx & m

4 4
&4 &4
m m
f(y|x)dy g(x)I(x0B)f x(x)dx
(3.7)
'
g(x)I(x0B)fx(x)dx &
&4
&4
g(x)I(x0B)f x(x)dx ' 0 .
110 Since X ' {X &1(B) , B 0 B}, property (3.6) is equivalent to m

A
Y() & Z() dP() ' 0 for all A 0 X .
(3.8)
Moreover, note that 0 X , so that (3.8) implies E(Y) ' Y()dP() ' Z()dP() ' E(Z) , m m

(3.9)
provided that the expectations involved are defined. A sufficient condition for the existence of E(Y) is that E( |Y| ) < 4 . We will see later that (3.10) is also a sufficient condition for the existence of E(Z) . I will show now that the condition (3.6) also holds for the examples (3.1) and (3.3). Of course, in the case (3.3) I have already shown this in (3.7), but it is illustrative to verify it again for the special case involved. In the case (3.1) the random variable Y.I(X=1) takes the value 0 with probability , and the values 2, 4, or 6 with probability 1/6, and the random variable Y.I(X=0) takes the value 0 with probability , and the values 1, 3, or 5 with probability 1/6, so that E[Y.I(X0B)] ' E[Y.I(X'1)] ' E[Y.I(X0B)] ' E[Y] E[Y.I(X0B)] ' 0 which by (3.1) and (3.6) is equal to 2 if 1 0 B and 0 B , (3.10)
E[Y.I(X0B)] ' E[Y.I(X'0)] ' 1.5 if 1 B and 0 0 B , ' 3.5 if 1 0 B and 0 0 B , if 1 B and 0 B ,
111 E[(E[Y|X])I(X0B)] ' 3E[I(X0B)] % E[X.I(X0B)] ' 3P(X0B) % P(X'1 and X0B) ' 3P(X'1) % P(X'1) ' 3P(X'0 or X'1) % P(X'1) ' 0 ' 2 if 1 0 B and 0 B , ' 3P(X'0) % P(X'1 and X'0) ' 1.5 if 1 B and 0 0 B , ' 3.5 if 1 0 B and 0 0 B , if 1 B and 0 B .
Moreover, in the case (3.3) the distribution function of Y.I(X0B) is: FB(y) ' P(Y.I(X0B) # y) ' P(Y # y and X 0 B) % P(X B) ' P(X 0 B_[0,y]) % P(Y # y and X 0 B_(y,1)) % P(X B) ' I(x0B)dx % y x &1I(x 0 B)dx % 1 & I(x 0 B)dx m m m
0 y 0 y 1 1
' 1 &
I(x0B)dx % y x &1I(x 0 B)dx m m

y y
for 0 # y # 1 ,
hence the density involved is fB(y) ' x &1I(x 0 B)dx for y 0 [0,1] , f B(y) ' 0 for y [0,1] . m
y 1
Thus 1 y.I(y 0 B)dy , E[Y.I(X 0 B)] ' y x I(x 0 B)dx dy ' m m 2m

&1 0 y 0 1 1 1
which is equal to
112 1 1 x.I(x 0 B)dx . E E[Y|X]I(X0B) ' E[X.I(X 0 B)] ' 2 2m

0 1
The two conditions (3.5) and (3.8) uniquely define Z ' E[Y|X] , in the sense that if there exist two versions of E[Y|X] , say Z1 ' E[Y|X] and Z2 ' E[Y|X] , satisfying the conditions (3.5) and (3.8), then P(Z1 ' Z2) ' 1 . To see this, let A ' { 0 : Z1() < Z2()} . Then A 0 X , hence it follows from (3.8) that m
A
(3.11)
Z2() & Z1() dP() ' E (Z2&Z1)I(Z2&Z1 > 0) ' 0 .
The latter equality implies P(Z2 & Z1 > 0) ' 0 , as I will show in Lemma 3.1 below. Replacing the set A by A ' { 0 : Z1() > Z2()} , it follows similarly that P(Z2 & Z1 < 0) ' 0 . Combining these two cases it follows that P(Z2 Z1) ' 0 .
Lemma 3.1: E[Z.I(Z > 0)] ' 0 implies P(Z > 0) = 0.
Proof: Choose g > 0 arbitrary. Then 0 ' E[Z.I(Z > 0)] ' E[Z.I(0 < Z < g)] % E[Z.I(Z $ g)] $ E[Z.I(Z $ g)] $ gE[I(Z $ g)] ' gP(Z $ g) , hence P(Z > g) ' 0 for all g > 0 . Now take g ' 1/n , n ' 1,2,..... , and let Cn ' {0: Z() > n &1} . Then Cn d Cn%1 , hence
113 P(Z > 0) ' P[ ^n'1C n] ' limn64P[Cn] ' 0 .

4
(3.12)
Q.E.D. The conditions (3.5) and (3.8) only depend on the conditioning random variable X via the sub- & algebra X of . Therefore, we can define the conditional expectation of a random variable Y relative to an arbitrary sub- & algebra 0 of , denoted by E[Y| 0 ], as follows:
Definition 3.1: Let Y be a random variable defined on a probability space {,,P} , satisfying E(|Y|) < 4 , and let 0 d be a sub- & algebra of . The conditional expectation of Y relative to the sub- & algebra 0 , denoted by E[Y| 0 ] = Z, say, is a random variable Z which is measurable 0 , and is such that for all sets A 0 0 , mA Y()dP() ' mA Z()dP() .
3.2.
Properties of conditional expectations As said before, the condition E(|Y|) < 4 is also a sufficient condition for the existence of
E(E[Y| 0]). The reason is two-fold. First, I have already established in (3.9) that
Theorem 3.1: E[E(Y|0)] = E(Y).
Second, conditional expectations preserve inequality:
114 Theorem 3.2: If P(X # Y) ' 1 then P(E(X|0) # E(Y|0)) ' 1 .
Proof: Let A ' {0: E(X|0)() > E(Y |0)()} . Then A 0 0 , and X()dP() ' E(X|0)()dP() # Y()dP() ' E(Y|0)()dP() , m m m m
A A A A
hence 0 # m
A
E(Y|0)() & E(X|0)() dP() # 0 .
(3.13)
It follows now from (3.13) and Lemma 3.1 that P({0: E(X|0)() > E(Y|0)()} ) ' 0 . Q.E.D. Theorem 3.2 implies that |E(Y|0)| # E(|Y| |0) with probability 1, and applying Theorem 3.1 it follows that E[|E(Y|0)|] # E(|Y|) . Therefore, the condition E(|Y|) < 4 is a sufficient condition for the existence of E( E [Y| 0] ). Conditional expectations also preserve linearity:
Theorem 3. 3: If E[|X|] < 4 and E[|Y|] < 4 then P[E(X % Y|0) = E(X|0) + E(Y|0)] = 1.
Proof: Let Z0 ' E(X %Y|0) , Z1 ' E(X|0) , Z2 ' E(Y|0) . For every A 0 0 we have: mA Z0()dP() ' mA (X() %Y())dP() ' X()dP() % Y()dP() , mA mA mA Z1()dP() ' mA X()dP() ,
115 and mA hence mA Z0() & Z1() & Z2() dP() ' 0 . (3.14) Z2()dP() ' mA Y()dP() ,
Taking A ' {0: Z0() & Z1() & Z2() > 0} it follows from (3.14) and Lemma 3.1 that P(A) = 0, and taking A ' {0: Z0() & Z1() & Z2() < 0} it follows similarly that P(A) = 0, hence P({0: Z0() & Z1() & Z2() 0}) ' 0 . Q.E.D. If we condition a random variable Y on itself, then intuitively we may expect that E(Y|Y) = Y, because then Y acts as a constant. More formally, this result can be stated as:
Theorem 3.4: Let E[|Y|] < 4 . If Y is measurable , then P(E(Y|) ' Y) ' 1 .
Proof: Let Z ' E(Y|) . For every A 0 we have: mA Y() & Z() dP() ' 0 . (3.15)
Take A ' {0: Y() & Z() > 0} . Then A 0 , hence it follows from (3.15) and Lemma 3.1 that P(A) = 0. Similarly, taking A ' {0: Y() & Z() < 0} it follows that P(A) = 0. Thus P({0: Y() & Z() 0}) ' 0 . Q.E.D. In Theorem 3.4 I have conditioned Y on the largest sub- & algebra of , namely itself. The smallest sub- & algebra of is T = {,i} , which is called the trivial & algebra.
116 Theorem 3.5: Let E[|Y|] < 4 . Then P[ E(Y| T ) = E(Y)] = 1.
Proof: Exercise, along the same lines as the proofs of Theorems 3.2-3.4.
The following theorem, which plays a key-role in regression analysis, follows from combining the results of Theorems 3.3 and 3.4:
Theorem 3.6: Let E[|Y|] < 4 and U ' Y & E[Y|0] . Then P[E(U|0) ' 0] ' 1 .
Proof: Exercise.
Next, let (Y, X, Z) be jointly continuously distributed with joint density function f(y,x,z) and marginal densities fy,x(y,x), fx,z(x,z) and fx(x). Then the conditional expectation of Y given X = x and Z = z is E[Y|X,Z] ' m
4
yf(y|X,Z)dy ' gx,z(X,Z) , say,
&4
where f(y|x,z) = f(y,x,z)/fx,z(x,z) is the conditional density of Y given X = x and Z = z. The conditional expectation of Y given X = x alone is E[Y|X] ' m
4
yf(y|X)dy ' gx(X) , say,
&4
where f(y|x) = fy,x(y,x)/fx(x) is the conditional density of Y given X = x alone. Denoting the conditional density of Z given X = x by fz(z|x) = fz,x(z,x)/fx(x), it follows now that
117 E E[Y|X,Z] |X ' m

4
&4 &4
m m
yf(y|X,z)dy f z(z|X)dz '

4
'
&4
&4
f(y,X,z)dzdy
f (y,X) 1 ' y y,x dy ' yf(y|X)dy ' E[Y|X] . m f x(X) m f x(X)

&4 &4
&4 &4 4
f (X,z) f(y,X,z) dy x,z dz m m fx,z(X,z) fx(X) y
This is one of the versions of the Law of Iterated Expectations. Denoting by X,Z the & algebra generated by (X,Z), and by X the & algebra generated by X, this result can be translated as: E E[Y|X,Z] |X ' E[Y|X] . Note that X d X,Z , because X ' {{0: X() 0 B1} , B1 0 B} = {{0: X() 0 B1 , Z() 0 } , B1 0 B} d {{0: X() 0 B1 , Z() 0 B2} , B1 , B2 0 B} = X,Z . Therefore, the law of iterated expectations can be stated more generally as:
Theorem 3.7: Let E[|Y|] < 4 , and let 0 d 1 be sub- & algebras of . Then P E E[Y|1] |0 ' E(Y|0) ' 1 .
Proof: Let Z0 ' E[Y|0] , Z1 ' E[Y|1] , and Z2 ' E[Z1|0] . It has to be shown that P(Z0 ' Z2) ' 1 . Let A 0 0 . Then also A 0 1 . It follows from Definition 3.1 that Z0 ' E[Y|0] implies Y()dP() ' Z0()dP() , m m
A A
Z1 ' E[Y|1] implies
118 Y()dP() ' Z1()dP() , m m

A A
and Z2 ' E[Z1|0] implies Z ()dP() ' Z1()dP() , m 2 m

A A
Combining these equalities it follows that for all A 0 0 , m

A
Z0() & Z2() dP() ' 0 .
(3.16)
Now choose A ' {0: Z0() & Z2() > 0} . Note that A 0 0 . Then it follows from (3.16) and Lemma 3.1 that P(A) ' 0 . Similarly, if we choose A ' {0: Z0() & Z2() < 0} then again P(A) ' 0 . Therefore, P(Z0 ' Z1) ' 1 . Q.E.D. The following monotone convergence theorem for conditional expectations plays a keyrole in the proofs of Theorems 3.9 and 3.10 below.
Theorem 3.8: (Monotone convergence). Let Xn be a sequence of non-negative random variables defined on a common probability space { , , P} , such that P(Xn # Xn%1) = 1 and E[supn$1Xn] < 4. Then P limn64 E[Xn|0] ' E[limn64Xn|0] ' 1.
Proof: Let Zn ' E[Xn|0] and X ' limn64Xn . It follows from Theorem 3.2 that Zn is monotonic non-decreasing, hence Z ' limn64Zn exists. Let A 0 0 be arbitrary, and denote for 0 , Y n() ' Z n().I( 0 A) , Y() ' Z().I( 0 A) . Then also Yn is nonnegative and monotonic non-decreasing, and Y ' limn64Y n, hence it follows from the monotone convergence
119 theorem that limn64 Y n()dP() ' Y()dP() , which is equivalent to m m limn64 Zn()dP() ' Z()dP() . mA mA Similarly, denoting Un() ' Xn().I(0A) , U() ' X().I(0A) , it follows from the monotone convergence theorem that limn64 Un()dP() ' U()dP() , which is equivalent m m to limn64 Xn()dP() ' X()dP() . mA mA Moreover, it follows from the definition of Zn ' E[Xn|0] that mA Z n()dP() ' mA Xn()dP() . (3.19) (3.18) (3.17)
It follows now from (3.17), (3.18) and (3.19) that mA Z()dP() ' mA X()dP() . (3.20)
Theorem 3.8 easily follows from (3.20). Q.E.D. The following theorem extends the result of Theorem 3.4:
Theorem 3.9: Let X be measurable 0 , and let both E(|Y|) and E(|XY|) be finite. Then P[E(XY|0) ' X.E(Y|0)] ' 1.
Proof: I will prove the theorem involved only for the case that both X and Y are nonnegative with probability 1, leaving the general case as an easy exercise.
120 Denote Z ' E(XY|0) , Z0 ' E(Y|0) . If A0 0: mA Z()dP() ' mA X()Z0()dP(), (3.21)
then the theorem under review holds. (a) First, consider the case that X is discrete: X() ' 'j'1jI( 0 Aj) , say, where the Aj s are
n
disjoint sets in 0 and the j s are non-negative numbers. Let A 0 0 be arbitrary, and observe that A_Aj 0 0 for j = 1,..,n. Then by Definition 3.1,
n
X()Z0()dP() ' j jI(0Aj)Z0()dP() ' j j Z0()dP() m m j'1 m j'1

n A A n A _A j
' j j Y()dP() ' j j I(0Aj)Y()dP() ' jI(0Aj)Y()dP() m m mj j'1 j'1 j'1

n n A _A j A A
' X()Y()dP() ' Z()dP() , m m

A A
which proves the theorem for the case that X is discrete. (b) If X is not discrete then there exists a sequence of discrete random variables Xn such that
for each 0 we have: 0 # Xn() # X() and Xn() 8 X() monotonic, hence also Xn()Y() 8 X()Y() monotonic. Therefore, it follows from Theorem 3.8 and part (a) that, E[XY|0] ' limE[XnY|0] ' limXn E[Y|0] ' X E[Y|0]
n64 n64
with probability 1. Thus the theorem under review holds for the case that both X and Y are nonnegative with probability 1. (c) The rest of the proof is left as an exercise. Q.E.D. We have seen for the case that Y and X are jointly absolutely continuous distributed that
121 the conditional expectation E[Y|X] is a function of X. This holds also more generally:
Theorem 3.10: Let Y and X be random variables defined on the probability space { , , P }, and assume that E(*Y*) < 4 . Then there exists a Borel measurable function g such that P[E(Y|X) ' g(X)] ' 1. This result carries over to the case where X is a finite-dimensional random vector.
Proof: The proof involves the following steps: (a) Suppose that Y is non-negative and bounded: K < 4: P({0: 0 # Y() # K}) = 1,
and let Z = E(Y*X ) , where X is the & algebra generated by X. Then P({0: 0 # Z() # K}) ' 1 . (3.22)
(b)
Under the conditions of part (a) there exists a sequence of discrete random variables Zm,
m m
Zm() ' 'i'1i,mI( 0 Ai,m) , where Ai,m 0 X , Ai,m_Aj,m ' i if i j , î'1Ai,m ' , Borel set Bi,m such that Ai,m ' X &1(Bi,m) . Thus, if we take gm(x) ' 'i'1i,mI(x 0 Bi,m) then Zm =
m
0 # i,m < 4 for i = 1,..,m, such that Zm() 8 Z() monotonic. For each Ai,m we can find a
gm(X) with probability 1. Next, let g(x) ' limsupm64gm(x) . This function is Borel measurable, and Z ' limsupm64Zm ' limsupm64g m(X) = g(X) with probability 1. (c) Let Yn = Y.I(Y < n). Then Y n() 8 Y() monotonic. By part (b) it follows that there
exists a Borel measurable function gn(x) such that E(Y n*X) ' gn(X) . Let g(x) = limsupn64gn(x), which is Borel measurable. It follows now from Theorem 3.8 that
122 E(Y*X) ' limn64 E(Y n*X) ' limsupn64 E(Y n*X) ' limsupn64 gn(X) ' g(X) . (d) Let Y % ' max(Y,0) , Y & ' max(&Y,0) . Then Y ' Y % & Y & , and therefore by part (c),
E(Y %*X) ' g %(X) , say, and E(Y &*X) ' g &(X) , say. Then E(Y*X) = g %(X) & g &(X) = g(X). Q.E.D. If random variables X and Y are independent, then knowing the realization of X will not reveal anything about Y, and vice versa. The following theorem formalizes this fact.
Theorem 3. 11: Let X and Y be independent random variables. If E[|Y|] < 4 then P(E[Y|X] = E[|Y|]) = 1. More generally, let Y be defined on the probability space {S,,P}, let Y be the F!algebra generated by Y, and let 0 be a sub-F!algebra of such that Y and 0 are independent, i.e., for all A 0 Y and B 0 0 , P(A_B) ' P(A)P(B) . If E[|Y|] < 4 then P(E[Y|0] = E[|Y|]) = 1.
Proof: Let X be the & algebra generated by X, and let A 0 X be arbitrary. There exists a Borel set B such that A ' {0: X() 0 B} . Then mA Y()dP() ' m Y()I( 0 A)dP()' m Y()I(X() 0 B)dP()
' E[YI(X0B)] ' E[Y]E[I(X0B)] , where the last equality follows from the independence of Y and X. Moreover E[Y]E[I(X0B)] ' E[Y] I(X()0B)dP() ' E[Y] I(0A)dP() ' E[Y]dP() . m m mA Thus
123 mA Y()dP() ' mA E[Y]dP().
By the definition of conditional expectation, this implies that E[Y|X] = E[Y] with probability 1. Q.E.D.
3.3.
Conditional probability measures and conditional independence The notion of a probability measure relative to a sub-F-algebra can be defined similar to
Definition 3.1, using the conditional expectation of an indicator function:
Definition 3.2: Let {S,,P} be a probability space, and let 0 d be a -algebra. Then for any set A in , P(A | 0) ' E[IA | 0] , where IA() ' I( 0 A) .
In the sequel I will use the shorthand notation P(Y 0 B | X) to indicate the conditional probability P({ 0 : Y() 0 B} | X ) , where B is a Borel set and X is the F-algebra generated by X, and P(Y 0 B | 0) to indicate P({ 0 : Y() 0 B} | 0 ) for any sub-F-algebra 0 of . The event Y 0 B involved may be replaces by any equivalent expression. Similar to the notion of independence of sets and random variables and/or vectors (see Chapter 1) we can now define conditional independence:
algebra 0 of if for any subsequence jn, P(^nAj | 0) ' (nP(Aj | 0) . Moreover, a

n n
Definition 3.3: A sequence of sets Aj 0 is conditional independent relative to a sub-F-
sequence Yj of random variables or vectors defined on a common probability space {S,,P} is
124 conditional independent relative to a sub-F-algebra 0 of if for any sequence Bj of conformable Borel sets the sets Aj ' { 0 : Y j() 0 Bj} are conditional independent relative to 0 .
3.4.
Conditioning on increasing sigma-algebras Consider a random variable Y defined on the probability space {S,,P}, satisfying E[|Y|]
< 4, and let n be an non-decreasing sequence of sub-F-algebras of : n d n%1 d . The question I will address is: What is the limit of E[Y|n] for n 64? As will be shown below, the answer to this question is fundamental for time series econometrics. We have seen in Chapter 1 that the union of F-algebras is not necessarily a F-algebra itself. Thus, ^n'1n may not be a F-algebra. Therefore, let
4
4 ' n'1 n ' ^n'1n ,

4 def. 4 4
(3.23)
contains ^n'1n .
4
i.e., 4 is the smallest F-algebra containing ^n'1n . Clearly, 4 d , because the latter also
The answer to our question is now:
Theorem 3.12: If Y is measurable , E[|Y|] < 4, and {n } is a non-decreasing sequence of sub-F-algebras of , then limn64E[Y|n] ' E[Y|4] with probability 1, where 4 is defined by (3.23).
This result is usually proved by using martingale theory. See Billingsley (1986), Chung
125 (1974) and Chapter 7. However, in Appendix 3.A I will provide an alternative proof of Theorem 3.12 which does not require martingale theory.
3.5.
Conditional expectations as the best forecast schemes I will show now that the conditional expectation of a random variable Y given a random
variable or vector X is the best forecasting scheme for Y, in the sense that the mean square forecast error is minimal. Let (X) be a forecast of Y, where is a Borel measurable function. The mean square forecast error (MSFE) is defined by MSFE ' E[(Y & (X))2] . The question is: for which function is the MSFE minimal. The answer is:
Theorem 3.13: If E[Y 2] < 4 , then E[(Y & (X))2] is minimal for (X) ' E[Y|X] .
Proof: According to Theorem 3.10 there exists a Borel measurable function g such that E[Y|X] ' g(X) with probability 1. Denote U ' Y & E[Y|X] ' Y & g(X) . It follows from Theorems 3.3, 3.4 and 3.9 that E[(Y & (X))2|X] ' E[(U % g(X) & (X))2|X] ' E[U 2|X] % 2E[(g(X) & (X))U|X] % E[(g(X) & (X))2|X] ' E[U 2|X] % 2(g(X) & (X))E[U|X] % (g(X) & (X))2 , (3.24)
where the last equality follows from Theorems 3.9 and 3.4. Since by Theorem 3. 6, E(U|X) ' 0 with probability 1, equation (3.24) becomes E[(Y & (X))2|X] ' E[U 2|X] % (g(X) & (X))2 . (3.25)
126 Applying Theorem 3.1 to (3.25), it follows now that E[(Y & (X))2] ' E[U 2] % E[(g(X) & (X))2] ,
which is minimal if E[(g(X) & (X))2] ' 0 . According to Lemma 3.1, this condition is equivalent to the condition that P[g(X) ' (X)] ' 1 . Q.E.D. Theorem 3.13 is the basis for regression analysis. In parametric regression analysis, a dependent variable Y is "explained" by a vector of explanatory (also called "independent") variables X according to a regression model of the type Y ' g(X , 0) % U , where g(x,) is a known function of x and an unknown vector of parameters, and U is the error term which is assumed to satisfy the condition E[U|X] ' 0 (with probability 1). The problem is then to estimate the unknown parameter vector . For example, a Mincer-type wage equation explains the log of the wage, Y, of a worker out of the years of education, X1, and the years of experience on the job, X2, by a regression model of the type Y ' % X1 % X2 & X2 % U , so that in this case ' (,,,)T , X ' (X1 , X2)T , and g(X,) ' % X1 % X2 & X2 . The condition that E[U|X] ' 0 with probability 1 now implies that E[Y|X] ' g(X,) with probability 1 for some parameter vector . It follows therefore from Theorem 3.12 that minimizes the mean square error function E[(Y & g(X,))2] : ' argmin E[(Y & g(X , ())2] ,
( 2 2
(3.26)
where "argmin" stands for the argument for which the function involved is minimal. Next, consider a strictly stationary time series process Yt.
127 Definition 3.4: A time series process Yt is said to be strictly stationary if for arbitrary integers m1 < m2 <....< mk the joint distribution of Yt&m ,......,Yt&m does not depend on the time index t.
1 k
Consider the problem of forecasting Yt on the basis on the past Yt!j , j $1, of Yt. Usually we do not observe the whole past of Yt , but only Yt!j for j =1,...,t-1, say. It follows from Theorem 3.13 that the optimal MSFE forecast of Yt given the information on Yt!j for j =1,...,m is the conditional expectation of Yt given Yt!j for j =1,...,m. Thus, if E[Y t ] < 4 then E [Y t|Yt&1,.....,Yt&m ] ' argmin E [ (Y t & (Yt&1,.....,Yt&m ) )2].
2
Similarly as before, the minimum is taken over all Borel measurable functions R on m. Moreover, because of the strict stationarity assumption, there exists a Borel measurable function gm on m which does not depend on the time index t such that with probability 1, E [Y t|Yt&1,...,Yt&m ] = gm(Yt&1,...,Yt&m ) for all t. Theorem 3.12 now tells us that limm64E [Y t|Yt&1,.....,Yt&m ] ' limm64gm(Yt&1,.....,Yt&m ) ' E [Y t|Yt&1 , Yt&2 , Yt&3 , ......... ] where the latter is the conditional expectation of Yt given its whole past Yt!j, j $1. More
t&1 t&1 4 t&1
(3.27)
formally, let t&m ' (Yt&1 , .... , Yt&m) and &4 ' m'1t&m . Then (3.27) reads limm64E [Y t|t&m ] ' E [Y t|&4 ] .
t&1 t&1
The latter conditional expectation is also denoted by Et!1[Yt]: Et&1[Y t] ' E [Y t|Yt&1 , Yt&2 , Yt&3 , ......... ] ' E [Y t|&4 ] .
def. def. t&1
(3.28)
In practice we do not observe the whole past of time series processes. However, it follows
128 from Theorem 3.12 that if t is large then approximately, E[Y t|Yt&1 , ... , Y1 ] . Et&1 [Y t ] . In time series econometrics the focus is often on modeling (3.28) as a function of past values of Yt and an unknown parameter vector 2, say. For example, an autoregressive model of order 1, denoted by AR(1), takes the form Et&1[Y t] ' % Yt&1 , ' ( , )T , where || < 1 . Then Y t ' % Yt&1 % Ut , where Ut is called the error term. If this model is true, then Ut ' Y t & Et&1[Y t] , which by Theorem 3. 6 satisfies P(Et&1[Ut] ' 0) ' 1. The condition || < 1 is one of the two necessary conditions for strict stationarity of Yt , the other one being that Ut is strictly stationary. To see this, observe that by backwards substitution we can write Y t ' /(1&) % 'j'0jUt&j , provided that || < 1 . The strict
4
stationarity of Yt follows now from the strict stationarity of Ut.
3.6. 1. 2. 3. 4. 5. 6. 7. 8.
Exercises: Why is property (3.6) equivalent to (3.8)? Why is the set A defined by (3.11) contained in X ? Why does (3.12) hold? Prove Theorem 3.5. Prove Theorem 3.6. Verify (3.20) . Why does Theorem 3.8 follow from (3.20)? Why does (3.21) imply that Theorem 3.9 holds ? Complete the proof of Theorem 3.9 for the general case, by writing X ' max(0,X)
& max(0,&X) = X1 & X2 , say, and Y ' max(0,Y) & max(0,&Y) = Y1 & Y2 , say, and applying the result of part (b) of the proof to each pair Xi , Y j .
129 9. 10. Prove (3.22). Let Y and X be random variables, with E[*Y*] < 4 , and let M be a Borel measurable
one-to-one mapping from into . Prove that E[Y|X] ' E[Y|(X)] with probability 1. 11. Let Y and X be random variables, with E[Y 2] < 4 , P(X ' 1) ' P(X ' 0) ' 0.5 ,
E[Y] ' 0 , and E[X.Y] ' 1 . Derive E[Y|X] . Hint: Use Theorems 3.10 and 3.13.
Appendix
3.A. Proof of Theorem 3.12 Let Zn ' E[Y|n] and Z ' E[Y|4] , and let A 0 ^n'1n be arbitrary. Note that the
4
latter implies A 0 4 . Because of the monotonicity of {n } there exists an index kA (depending on A) such that for all n $ kA, hence limn64 Z n()dP() ' Y()dP() . m m
A A
(3.29)
If Y is bounded: P[|Y| < M] ' 1 for some positive real number M, then Zn is uniformly bounded: |Zn| ' |E[Y|n]| # E[|Y||n] # M , hence it follows from (3.29), the dominated convergence theorem and the definition of Z that lim Z ()dP() ' Z()dP() m n64 n m
A A
(3.30)
for all sets A 0 ^n'1n .

4 4
of {n } that ^n'1n is an algebra. Now let ( be the collection of all subsets of 4

4
Although ^n'1n is not necessarily a F-algebra, it is easy to verify from the monotonicity
130 satisfying the following two conditions: (a) (b) For each set B 0 ( equality (3.30) holds with A = B
Since (3.30) holds for A = S because 0 ^n'1n , it is trivial that (3.30) also holds for the
4
For each pair of sets B1 0 ( and B2 0 ( equality (3.30) holds with A ' B1^B2 .
of A: complement A lim Z ()dP() ' Z()dP() , m n64 n m

A A
0 . Thus, is an algebra. Note that this algebra exists, because hence if B 0 ( then B ( ( ^n'1n is an algebra satisfying the conditions (a) and (b). Thus, ^n'1n d ( d 4 .
4 4
smallest F-algebra containing ^n'1n . For any sequence of disjoint sets Aj 0 ( it follows
4
I will show now that ( is a F-algebra, so that 4 ' ( , because the former is the
from (3.30) that limn64Zn()dP() ' j limn64Zn()dP() ' j Z()dP() ' Z()dP() . m m j'1 m j'1 m
4 4 Aj Aj ^j'1A j
4
^j'1A j
4
hence ^j'1Aj 0 ( . This implies that ( is a F-algebra containing ^n'1n , because we have
4 4
seen in Chapter 1 that an algebra which is closed under countable unions of disjoint sets is a Falgebra. Hence 4 ' ( , and consequently, (3.30) holds for all sets A 0 4 . This implies that P[Z ' limn64Zn] = 1 if Y is bounded. Next, let Y be non-negative: P[|Y $ 0] ' 1 , and denote for natural numbers m $ 1, Bm ' { 0 : m&1 # Y() < m} , Y m ' Y.I(m&1 # Y < m) , Zn
(m)
' E[Y m|n] and Z (m) =
E[Y m|4] . I have just shown that for fixed m $ 1 and arbitrary A 0 4 ,
131 lim Z ()dP() ' Z (m)()dP() ' Y m()dP() ' Y()dP() m n64 n m m m
(m) A A A A _B m
where the last two equalities follow from the definitions of Z (m) and Y m . Since ) ' 0 it follows that Zn(m)()I( 0 B ) ' 0 , hence Y m()I( 0 B m m m limn64Zn ()dP() '
(m)
A _B m
A _B m
Y()dP()
Moreover, it follows from the definition of conditional expectations and Theorem 3.7 that Zn
(m)
' E[Y.I(m&1 # Y < m)|n] ' E[Y|Bm_n] ' E[E(Y|n)|Bm_n] ' E[Zn|Bm_n] ,
4
hence for every set A 0 ^n'1n , limn64 m Z n ()dP() ' limn64

(m)
A _B m
A _B m
Z n()dP() '
A _B m
limn64Zn()dP() '
A _B m
Y()dP() ,
which by the same argument as in the bounded case carries over to the sets A 0 4 . Thus we have m limn64Z n()dP() ' m Y()dP()
A _B m
A _B m
for all sets A 0 4 . Consequently lim Z ()dP() ' j limn64Z n()dP() ' j Y()dP() ' Y()dP() m n64 n m m'1 m m'1 m
4 4 A A _B m A _B m A
for all sets A 0 4 . This proves the theorem for the case P[|Y $ 0] ' 1 . The general case is now easy, using the decomposition Y ' max(0,Y) & max(0,&Y) .
132 Endnote 1. Here and in the sequel the notations P(Y ' y|X'x) , P(Y ' y and X'x) , and P(X'x) (and similar notations involving inequalities) are merely short-hand notations for the probabilities P({0: Y() ' y}|{0: X()'x}) , P({0: Y() ' y}_{0: X()'x}) , and P({0: X()'x}) , respectively.
133
Chapter 4 Distributions and Transformations
In this chapter I will review the most important univariate distributions and derive their expectation, variance, moment generating function (if they exist), and characteristic function. Quite a few distributions arise as transformations of random variables or vectors. Therefore I will also address the problem how for a Borel measure function or mapping g(x) the distribution of Y = g(X) is related to the distribution of X.
4.1.
Discrete distributions In Chapter 1 I have introduced three " natural" discrete distributions, namely the
hypergeometric, binomial, and Poisson distributions. The first two are natural in the sense that they arise from the way the random sample involved is drawn, and the latter because it is a limit of the binomial distribution. A fourth "natural" discrete distribution I will discuss is the negative binomial distribution.
4.1.1
The hypergeometric distribution Recall that a random variable X has a hypergeometric distribution if K k N &K n&k N n
P(X ' k) '
for k ' 0,1,2,.....,min(n,K) , P(X ' k) ' 0 elsewhere ,
(4.1)
134 where 0 < n < N and 0 < K < N are natural numbers. This distribution arises, for example, if we draw randomly without replacement n balls from a bowl containing K red balls and N !K white balls. The random variable X is then the number of red balls in the sample. In almost all applications of this distribution, n < K, so that I will focus on that case only. The moment generating function involved cannot be simplified further than its definition mH(t) = 'k'0exp(t.k)P(X ' k) , and the same applies to the characteristic function. Therefore, we
m
have to derive the expectation directly: K k N&K n&k N n K!(N&K)! (k&1)!(K&k)!(n&k)!(N&K&n%k)! ' j N! k'1 n!(N&n)!
n
E[X] ' j k
n k'0
(K&1)!((N&1)&(K&1))! k!((K&1)&k)!((n&1)&k)!((N&1)&(K&1)&(n&1)%k)! nK ' j N k'0 (N&1)! (n&1)!((N&1)&(n&1))!

n&1
'
nK j N k'0
n&1
K&1 k
(N&1)&(K&1) (n&1)&k N&1 n&1
'
nK . N
Along similar lines it follows that E[X(X&1)] ' n(n&1)K(K&1) , N(N&1)
(4.2)
hence var(X) ' E[X 2] & (E[X])2 ' nK (n&1)(K&1) nK % 1 & . N N&1 N
135 4.1.2 The binomial distribution A random variable X has a binomial distribution if n k p (1 & p)n&k for k ' 0,1,2,....,n, P(X ' k) ' 0 elsewhere , k
P(X ' k) '
(4.3)
were 0 < p < 1. This distribution arises, for example, if we draw randomly with replacement n balls from a bowl containing K red balls and N!K white balls, where K/N = p. The random variable X is then the number of red balls in the sample. We have seen in Chapter 1 that the binomial probabilities are limits of hypergeometric probabilities: If both N and K converge to infinity such that K/N 6 p then for fixed n and k, (4.1) converges to (4.3). This suggests that also the expectation and variance of the binomial distribution are the limits of the expectation and variance of the hypergeometric distribution, respectively: E[X] ' np , var(X) ' np(1&p) . As we will see later, in general convergence of distributions does not imply convergence of expectations and variances, except if the random variables involved are uniformly bounded. Therefore, in this case the conjecture is true because the distributions involved are bounded: P[0 # X < n] = 1. However, it is not hard to verify (4.4) and (4.5) from the moment generating function:
B(t) ' j exp(t.k) n k'0
(4.4) (4.5)
n k p (1 & p)n&k ' j k k'0

n
n (pe t)k(1 & p)n&k ' (p.e t % 1 & p)n . k
(4.6)
Similarly, the characteristic function is nB(t) ' (p.e i.t % 1 & p)n .
136 4.1.3 The Poisson distribution A random variable X is Poisson() distributed if for k = 0,1,2,3,......., and some > 0, P(X ' k) ' exp(&) k . k!
(4.7)
Recall that the Poisson probabilities are limits of the binomial probabilities (4.3) for n 6 4 and p 9 0 such that np 6 . It is left as exercises to show that the expectation, variance, moment generating function, and characteristic function of the Poisson() distribution are: E[X] ' , var(X) ' , mP(t) ' exp[(e t & 1)] , (4.8) (4.9) (4.10)
and nP(t) ' exp[(e i.t & 1)] , (4.11)
respectively.
4.1.4
The negative binomial distribution Consider a sequence of independent repetitions of a random experiment with constant
probability p of success. Let the random variable X be the total number of failures in this sequence before the m-th success, where m $1. Thus, X+m is equal to the number of trials necessary to produce exactly m successes. The probability P(X = k), k = 0,1,2,...., is the product of the probability of obtaining exactly m-1 successes in the first k+m-1 trials, which is equal to
137 the binomial probability k%m&1 m&1 p (1&p)k%m&1&(m&1) m&1 and the probability p of a success on the (k+m)-th trial. Thus k%m&1 m p (1&p)k, k ' 0,1,2,3,.... m&1
P(X ' k) '
This distribution is called the Negative Binomial (m,p) [shortly: NB(m,p)] distribution. It is easy to verify from the above argument that a NB(m,p) distributed random variable can be generated as the sum of m independent NB(1,p) distributed random variables, i.e., if X1,1,.....,X1,m are independent NB(1,p) distributed, then X ' 'j'1X1,j is NB(m,p) distributed. The
n
moment generating function of the NB(1,p) distribution is k mNB(1,p)(t) ' j exp(k.t) p(1&p)k ' p j (1&p)e t 0 k'0 k'0
4 4
'
p 1&(1&p)e t
provided that t < !ln(1!p), hence the moment generating function of the NB(m,p) distribution is p 1&(1&p)e t
m
mNB(m,p)(t) '
, t < &ln(1&p) .
(4.12)
Replacing t by i.t in (4.12) yields the characteristic function: p 1&(1&p)e i.t

m
nNB(m,p)(t) '
'
p(1%(1&p)e i.t) 1%(1&p)2
It is now easy to verify, using the moment generating function, that for a NB(m,p) distributed random variable X, E[X] ' m(1&p) /p ,
138 var(X) ' m(1&p)2/p 2 % m(1&p)/p .
4.2.
Transformations of discrete random variables and vectors In the discrete case the question "Given a random variable or vector X and a Borel
measure function or mapping g(x), how is the distribution of Y = g(X) related to the distribution of X?" is easy to answer. If P[X 0 {x1 , x2 , ....... }] = 1 and g(x1) , g(x2) , .... are all different, the answer is trivial: P(Y ' g(xj)) = P(X ' xj) . If some of the values g(x1) , g(x2) ,...... are the same, let {y1 , y2 , ......} be the set of distinct values of g(x1) , g(x2) ,...... Then P(Y ' yj) 'j I[yj ' g(xi)] P(X ' xi ) .
4 i'1
(4.13)
It is easy to see that (4.13) carries over to the multivariate discrete case. For example, if X is Poisson(8) distributed and g(x) ' sin2(x) ' sin(x) 2 , so that for m = 0,1,2,3,...., g(2m) ' sin2(m) ' 0, g(2m%1) ' sin2(m%/2) ' 1 , then P(Y ' 0) = e &'j'02j/(2j)! and P(Y ' 1) ' e &'4 2j%1/(2j%1)! . j'0
4
As an application, let X ' (X1 , X2)T , where X1 and X2 are independent Poisson(8) distributed, and let Y = X1 + X2 . Then for y = 0,1,2,....... P(Y ' y) ' j j I[y ' i%j] P(X1 ' i , X2 ' j ) ' exp(&2)
4 4 i'0 j'0
(2)y . y!
(4.14)
Hence, Y is Poisson(28) distributed. More generally, we have
Theorem 4.1: If for j = 1,.....,k the random variables Xj are independent Poisson(8j) distributed then 'j'1Xj is Poisson('j'1j) distributed.
k k
139 4.3. Transformations of absolutely continuous random variables If X is absolutely continuously distributed, with distribution function F(x) ' f(u)du , m&4
x
the derivation of the distribution function of Y = g(X) is less trivial. Let us assume first that g is continuous and monotonic increasing: g(x) < g(z) if x < z. Note that these conditions imply that g is differentiable1. Then g is a one-to-one mapping, i.e., for each y 0 [g(&4) ,g(4)] there exists one and only one x 0 ^{&4}^{4} such that y = g(x). This unique x is denoted by x ' g &1(y) . Note that the inverse function g &1(y) is also monotonic increasing and differentiable. Now let H(y) be the distribution function of Y. Then: H(y) ' P(Y # y) ' P(g(X) # y) ' P(X # g &1(y)) ' F(g &1(y)) . (4.15)
Taking the derivative of (4.15) yields the density h(y) of Y: h(y) ' H )(y) ' f(g &1(y)) dg &1(y) . dy
(4.16)
If g is differentiable and monotonic decreasing: g(x) < g(z) if x > z, then g &1(y) is also monotonic decreasing, so that (4.15) becomes H(y) ' P(Y # y) ' P(g(X) # y) ' P(X $ g &1(y)) ' 1 & F(g &1(y)) ,
and (4.16) becomes dg &1(y) . dy
h(y) ' H )(y) ' f(g &1(y)) &
(4.17)
&1 Note that in this case the derivative of g (y) is negative, because g &1(y) is monotonic
decreasing. Therefore, we can combine (4.16) and (4.17) into one expression:
140 h(y) ' f(g &1(y))/0 00 dg &1(y) /. dy 000
(4.18)
Theorem 4.2: If X is absolutely continuously distributed with density f, and Y = g(X), where g is a differentiable monotonic real function on , then Y is absolutely continuously distributed, with density h(y) given by (4.18) if min[g(&4),g(4)] < y < max[g(&4),g(4)] , and h(y) = 0 elsewhere.
4.4. 4.4.1
Transformations of absolutely continuous random vectors The linear case Let X ' (X1 , X2)T be a bivariate random vector, with distribution function
x1 x2
F(x) '
&4&4
mm
f(u1 , u2)du1 du2 '
(&4,x1](&4,x2]
f(u)du , where x ' (x1 , x2)T, u ' (u1 , u2)T
In this section I will derive the joint density of Y = AX + b, where A is a (non-random) nonsingular 22 matrix and b is a non-random 21 vector. Let us first consider the case that A is equal to the unit matrix I, so that Y = X + b with b ' (b1 , b2)T . Then the joint distribution function H(y) of Y is H(y) ' P(Y1 # y1 , Y2 # y2) ' P(X1 # y1&b1 , X2 # y2&b2) ' F(y1&b1 , y2&b2) ,
hence the density if Y is
141 h(y) ' M2H(y) ' f(y1&b1 , y2&b2) ' f(y&b). My1My2
Recall from linear algebra (see Appendix I) that any square matrix A can be decomposed into A ' R &1L.D.U, (4.19)
where R is a permutation matrix (possibly equal to the unit matrix I), L is a lower-triangular matrix with diagonal elements all equal to 1, U is an upper-triangular matrix with diagonal elements all equal to 1, and D is a diagonal matrix. Therefore, I will consider the four cases, A = U, A = D, A = L, and A = R!1 separately, for b = 0, and then apply the results involved sequentially according to the decomposition (4.19) to X+b, which then yields the general result. Consider the case that Y = AX, with A an upper-triangular matrix: 1 a 0 1
A ' Then Y '
(4.20)
Y1 Y2
'
X1%aX2 X2
hence the joint distribution function H(y) of Y is H(y) ' P(Y1 # y1 , Y2 # y2) ' P(X1%aX2 # y1 , X2 # y2) ' E I(X1 # y1 & aX2)I(X2 # y2) ' E E I(X1 # y1 & aX2 )|X2 I(X2 # y2) ' m
y2 y1&ax2
&4
&4
f1|2(x1|x2)dx1 f2(x2)dx2 '
&4
y2
y1&ax2
(4.21) f(x1,x2)dx1 dx2 ,
&4
142 where f1|2(x1|x2) is the conditional density of X1 given X2 = x2, and f2(x2) is the marginal density of X2 . Taking partial derivatives, if follows from (4.21) that for Y = AX with A given by (4.20), M2H(y) M h(y) ' ' f(y &ax2,x2)dx2 ' f(y1&ay2,y2) ' f(A &1y) . My1My2 My2 m 1
&4 y2
Along the same lines it follows that if A is a lower-triangular matrix then the joint density of Y = AX is M2H(y) ' f(y1 , y2&ay1) ' f(A &1y) . My1My2
h(y) '
(4.22)
Next, let A be a nonsingular diagonal matrix a1 0 0 a2
A '
where a1 0 , a2 0 . Then Y1 ' a1X1 and Y2 ' a2X2 , hence the joint distribution function H(y) is: H(y) ' P(Y1 # y1 , Y2 # y2) ' P(a1X1 # y1 , a2X2 # y2) '
y1/a1 y2/a2
P(X1 # y1/a1 , X2 # y2/a2) ' P(X1 # y1/a1 , X2 > y2/a2) ' P(X1 > y1/a1 , X2 # y2/a2) ' P(X1 > y1/a1 , X2 > y2/a2) '
&4 &4 y1/a1 4
m m
f(x1 , x2)dx1dx2 if a1 > 0 , a2 > 0 , f(x1 , x2)dx1dx2 if a1 > 0 , a2 < 0 , (4.23) f(x1 , x2)dx1dx2 if a1 < 0 , a2 > 0 , f(x1 , x2)dx1dx2 if a1 < 0 , a2 < 0 .
&4 y2/a2 4 y2/a2
m m m m
y1/a1 &4 4 4
y1/a1 y2/a2
m m
It is a standard calculus exercise for verify from (4.23) that in all four cases
143 h(y) ' f(y1/a1 , y2/a2) M2H(y) ' ' f(A &1y) det(A &1) . My1My2 |a1a2|
(4.24)
Finally, consider the case where A is the inverse of a permutation matrix ( which is a matrix that permutates the columns of the unit matrix), say: 0 1 1 0
&1
A '
'
0 1 1 0
Then the joint distribution function H(y) of Y =AX is H(y) ' P(Y1 # y1 , Y2 # y2) ' P(X2 # y1 , X1 # y2) ' F(y2 ,y1) ' F(Ay) , and the density involved is h(y) ' M2H(y) ' f(y2 , y1) ' f(Ay) . My1My2
Combining these results, it is not hard to verify, using the decomposition (4.19), that for the bivariate case (k = 2):
Theorem 4.3: Let X be k-variate absolutely continuously distributed with joint density f(x), and let Y = AX + b, where A is a nonsingular square matrix. Then Y is k-variate absolutely continuously distributed, with joint density h(y) ' f(A &1(y&b)) |det(A &1)|.
However, this result holds for the general case as well
4.4.2
The nonlinear case If we denote G(x) ' Ax%b , G &1(y) ' A &1(y&b) , then the result of Theorem 4.3 reads:
h(y) ' f(G &1(y)) |det(MG &1(y)/My)|. This suggests that Theorem 4.3 can be generalized as
144 follows.
Theorem 4.4: Let X be k-variate absolutely continuously distributed with joint density f(x), x ' (x1 , .... , xk)T , and let Y = G(X), where G(x) ' (g1(x) , .... , gk(x))T is a one-to-one mapping with inverse mapping x ' G &1(y) ' (g1 (y) , ..... , gk (y))T , whose components are differentiable in the components of y ' (y1 , .... , yk)T . Let J(y) ' Mx/My ' MG &1(y)/My , i.e., J(y) is the matrix with i,js element Mgi (y)/Myj , which is called the Jacobian. Then Y is k-variate absolutely continuously distributed, with joint density h(y) ' f(G &1(y)) |det(J(y))| for y in the set G(k) ' {y 0 k: y ' G(x) , f(x) > 0 , x 0 k} , and h(y) = 0 elsewhere.
( ( (
This conjecture is indeed true. Its formal proof is given in Appendix 4. B. An application of Theorem 4.4 is the following problem. Consider the function f(x) ' c.exp(&x 2/2) if x $ 0 , ' 0 if x < 0 . For which value of c is this function a density?. In order to solve this problem, consider the joint density f(x1 , x2) ' c 2exp[&(x1 %x2 )/2] , x1 $ 0 , x2 $ 0, which is the joint distribution of X = (X1,X2)T, where X1 and X2 are independent random drawings from the distribution with density (4.25). Next, consider the transformation Y = (Y1,Y2)T = G(X) defined by: Y1 ' X1 %X2
2 2 2 2
(4.25)
0 (0,4)
Y2 ' arctan(X1/X2) 0 (0,/2).
145 The inverse X ' G &1(Y) of this transformation is X1 ' Y1sin(Y2) , X2 ' Y1cos(Y2) , with Jacobian: MX1/MY1 MX1/MY2 MX2/MY1 MX2/MY2 sin(Y2) Y1cos(Y2) cos(Y2) &Y1sin(Y2)
J(Y) '
'
Note that det[J(Y)] ' &Y1 . Consequently, the density h(y) ' h(y1 , y2) ' f(G &1(y)) |det(J(y))| is: h(y1,y2) ' c 2y1exp(&y1 /2) for y1 > 0 and 0 < y2 < /2 , ' 0 elsewhere , hence, mm
0 0 4 /2 2
1 '
2 y1exp(&y1 /2)dy2dy1'
c (/2) y1exp(&y1 /2)dy1' c 2/2 . m

2 2 0
Thus the answer is c ' 2/ : m

0 4
exp(&x 2/2) /2
dx ' 1.
Note that this result implies that m

4
exp(&x 2/2) 2
dx ' 1.
(4.26)
&4
146 4.5. The normal distribution I will now review a number of univariate continuous distributions that play a key role in statistical and econometric inference, starting with the normal distribution. The standard normal distribution emerges as a limiting distribution of an aggregate of random variables. In particular, if X1,....Xn are independent random variables with expectation and finite and positive variance F2 then for large n the random variable Y n ' (1/ n)'j'1(Xj&)/ is approximately standard
n
normally distributed. This result, known as the central limit theorem, will be derived in Chapter 6, and carries over to various types of dependent random variables (see Chapter 7).
4.5.1
The standard normal distribution The standard normal distribution is an absolutely continuous distribution with density
function exp(&x 2/2) 2 , x 0 ,
f(x) '
(4.27)
Compare with (4.26). Its moment generating function is m

4
mN(0,1)(t) ' m
4
exp(t.x)f(x)dx '
&4
&4
exp(t.x)
exp(&x 2/2) 2 m
4
dx
' exp(t /2)
exp[&(x 2&2t.x%t 2)/2] 2 ' exp(t /2)

2
dx ' exp(t /2)
exp[&(x&t)2/2] 2
dx
(4.28)
&4
&4
&4
exp[&u 2/2] 2
du ' exp(t 2/2) ,
which exists for all t 0 , and its characteristic function is
147 nN(0,1)(t) ' m(i.t) ' exp(&t 2/2) .
Consequently, if X is standard normally distributed then E[X] ' m )(t) ' 0 , E[X 2] ' var(X) ' m ))(t) ' 1.
t'0
t'0
Due to this the standard normal distribution is denoted by N(0,1), where the first number is the expectation and the second number is the variance, and the statement "X is standard normally distributed" is usually abbreviated as "X ~ N(0,1)".
4.5.2
The general normal distribution Now let Y = + FX, where X ~ N(0,1). It is left as an easy exercise to verify that the
density of Y takes the form exp &(x&)2/2 2 , x 0 ,
f(x) '
with corresponding moment generating function mN(,2)(t) ' E[exp(t.Y)] ' exp( t)exp(2t 2/2) , t 0 ,
and characteristic function nN(,2)(t) ' E[exp(i.t.Y)] ' exp(i. t)exp(&2t 2/2) .
Consequently, E[Y] ' , var(Y) ' 2 . This distribution is the general normal distribution, denoted by N(,F2). Thus, Y~ N(,F2).
148 4.6. Distributions related to the standard normal distribution The standard normal distribution gives rise, via various transformations, to other distributions, such as the chi-square, t, Cauchy, and F distribution. These distributions are fundamental in testing statistical hypotheses, as we will see later.
4.6.1
The chi-square distribution Let X1,....Xn be independent N(0,1) distributed random variables, and let Y n ' 'j'1Xj .
n 2
(4.29)
The distribution of Yn is called the chi-square distribution with n degrees of freedom, denoted by n or 2(n) . Its distribution and density functions can be derived recursively, starting from the case n = 1:
y y 2
G1(y) ' P[Y1 # y] ' G1(y) ' 0 for y # 0 ,
2 P[X1
# y] ' P[& y # X1 # y] ' for y > 0 ,
f(x)dx ' 2 f(x)dx m

0
& y
hence g1(y) ' G1 (y) ' f

)
y / y '
exp(&y/2) y 2
for y > 0 ,
g1(y) ' 0 for y # 0 ,
where f (x) is defined by (4.27). Thus, g1(y) is the density of the 1 distribution. The corresponding moment generating function is
149 m2(t) '

1
1 1&2t
for t < 1/2 ,
(4.30)
and the characteristic function is n2(t) '

1
1 1&2.i.t
'
1%2.i.t 1%4.t
2
(4.31)
It follows easily from (4.29), (4.36) and (4.31) that the moment generating and characteristic functions of the n distribution are
2
m2(t) '
n
1 1&2t
n/2
for t < 1/2
(4.32)
and 1%2.i.t 1%4.t 2

n/2
n2(t) '
n
respectively. Therefore, the density of the n distribution is
gn(y) ' where for " > 0,
y n/2&1exp(&y/2) (n/2)2n/2
(4.33)
() '
x &1exp(&x)dx . m
0
(4.34)
The result (4.33) can be proved by verifying that for t < 1/2, (4.32) is the moment generating function of (4.33). The function (4.34) is called the Gamma function. Note that
150 (1) ' 1 , (1/2) ' , (%1) ' () for > 0 . Moreover, the expectation and variance of the n distribution are
2
(4.35)
E[Y n] ' n , var(Y n) ' 2n .
(4.36)
4.6.2
The Student t distribution Let X ~ N(0,1) and Yn ~ n , where X and Yn are independent. Then the distribution of the
2
random variable Tn ' X Y n /n
is called the (Student2) t distribution with n degrees of freedom, denoted by tn. The conditional density hn(x|y) of Tn given Yn = y is the density of the N(1,n/y) distribution, hence the unconditional density of Tn is exp(&(x 2/n)y/2) y n/2&1exp(&y/2) dy ' n/2 m ( n /2)2 n / y 2 0
4
h n(x) '
((n%1)/2) n (n/2) (1%x 2/n)(n%1)/2
The expectation of Tn does not exist if n = 1, as we will see below, and is zero for n $ 2, by symmetry. Moreover, the variance of Tn is infinite for n = 2., whereas for n $ 3, var(Tn) ' E[T n ] '
2
n . n&2
(4.37)
See Appendix 4.A . The moment generating function of the tn distribution does not exist, but it characteristic function does, of course:
151 ((n%1)/2) m (1%x 2/n)(n%1)/2 n (n/2) & 4 exp(it.x)

4
nt (t) '
n
dx '
2.((n%1)/2)
cos(t.x) dx . m n (n/2) 0 (1%x 2/n)(n%1)/2
4.6.3
The standard Cauchy distribution The t1 distribution is also known as the standard Cauchy distribution. Its density is: h1(x) ' (1) (1/2) (1%x )
2
'
1 (1%x 2 )
(4.38)
where the second equality follows from (4.35), and its characteristic function is nt (t) ' exp(&|t|) .
1
The latter follows from the inversion formula for characteristic functions: 1 1 exp(&i.t.x)exp(&|t|)dt ' . 2 2 m (1 % x ) &4
4
(4.39)
See Appendix 4.A. Moreover, it is easy to verify from (4.38) that the expectation of the Cauchy distribution does not exist, and that the second moment is infinite.
4.6.4
The F distribution Let Xm ~ m and Yn ~ n , where Xm and Yn are independent. Then the distribution of the
2 2
random variable Xm/m Y n /n
F '
is said to be F with m and n degrees of freedom, denoted by Fm,n. Its distribution function is
152 m
0 4 m.x.y/n
Hm,n(x) ' P[F # x] ' and its density is
m
0
z m/2&1exp(&z/2) (m/2)2m/2
dz
y n/2&1exp(&y/2) (n/2)2n/2
dy , x > 0 ,
hm,n(x) ' See Appendix 4.A
m m/2(m/2%n/2)x m/2&1 n m/2(m/2)(n/2) 1%m.x/n m/2%n/2
, x > 0,
(4.40)
Moreover, it is shown in Appendix 4.A that E[F] ' n/(n&2) if ' 4 var(F) ' n $ 3,
if n ' 1,2 , 2n 2(m%n&4) m(n&2)2(n&4) if n $ 5, (4.41)
' 4 ' not defined
if n ' 3,4 , if n ' 1,2 .
Furthermore, the moment generating function of the Fm,n distribution does not exist, and the computation of the characteristic function is too tedious an exercise, and therefore omitted.
4.7.
The uniform distribution and its relation to the standard normal distribution As we have seen before in Chapter 1, the uniform [0,1] distribution has density f(x) ' 1 for 0 # x # 1, f(x) ' 0 elsewhere.
More generally, the uniform [a,b] distribution (denoted by U[a,b]) has density f(x) ' moment generating function mU[a,b](t) ' exp(t.b)&exp(t.a) , (b&a)t 1 for a # x # b, f(x) ' 0 elsewhere , b&a
153 and characteristic function nU[a,b](t) ' exp(i.b.t)&exp(i.a.t) (sin(b.t)%sin(a.t)) & i.(cos(b.t)%cos(a.t)) ' . i.(b&a)t b&a
Most computer languages such as Fortran, Pascal, and Visual Basic have a build-in function which generates independent random drawings from the uniform [0,1] distribution.3 These random drawings can be converted into independent random drawings from the standard normal distribution via the transformation X1 ' cos(2U1) . &2.ln(U2) , (4.42) X2 ' sin(2U1) . &2.ln(U2) ,
where U1 and U2 are independent U[0,1] distributed. Then X1 and X2 are independent standard normally distributed. This method is called the Box-Muller algorithm.
4.8.
The Gamma distribution The n distribution is a special case of a Gamma distribution. The density of the Gamma
2
distribution is x &1exp(&x/) () , x > 0, > 0 , > 0 .

2
g(x) '
This distribution is denoted by '(",$). Thus, the n distribution is a Gamma distribution with " = n/2 and $ = 2. The Gamma distribution has moment generating function
154 m(,)(t) ' [1&t]& , t < 1/ , (4.43)
and characteristic function n(,)(t) ' [1&.i.t]& . Therefore, the '(",$) distribution has expectation "$ and variance "$2. The '(",$) distribution with " = 1 is called the exponential distribution.
4.9. 1. 2. 3. 4. 5. 6. 7. 8. 9.
Exercises Derive (4.2). Derive (4.4) and (4.5) directly from (4.3). Derive (4.4) and (4.5) from the moment generating function (4.6). Derive (4.8), (4.9), and (4.10). If X is discrete and Y = g(X), do we need to require that g is Borel measurable? Prove the last equality in (4.14). Prove Theorem 4.1, using characteristic functions. Prove that (4.24) holds for all four cases in (4.23). Let X be a random variable with continuous distribution function F(x). Derive the
distribution of Y = F(X). 10. The standard normal distribution has density f(x) ' exp(&x 2 / 2)/ 2 , x 0 . Let X1 and
X2 be independent random drawings from the standard normal distribution involved, and let Y1 = X1 + X2, Y2 = X1 ! X2. Derive the joint density h(y1 , y2) , say, of Y1 and Y2 , and show that Y1 and Y2 are independent. Hint: Use Theorem 4.3. 11. The exponential distribution has density f(x) ' &1exp(&x/) if x $ 0, f(x) = 0 if x < 0,
155 where 2 > 0 is a constant. Let X1 and X2 be independent random drawings from the exponential distribution involved, and let Y1 = X1 + X2, Y2 = X1 ! X2. Derive the joint density h(y1 , y2) , say, of Y1 and Y2 . Hints: Determine first the support {(y1,y2)T 0 2: h(y1 , y2) > 0} of h(y1 , y2) , and then use Theorem 4.3. 12. 13. Let X ~ N(0,1). Derive E[X2k] for k = 2,3,4, using the moment generating function. Let X1,X2,...,Xn be independent standard normally distributed. Show that (1/ n)'j'1Xj is
n
standard normally distributed. 14. 15. 16. 17. 18. 19. Prove (4.30). Show that for t < 1/2, (4.32) is the moment generating function of (4.33). Explain why the moment generating function of the tn distribution does not exist Prove (4.35). Prove (4.36). Let X1,X2,...,Xn be independent standard Cauchy distributed. Show that (1/n)'j'1Xj is
n
standard Cauchy distributed. 20. The class of standard stable distributions consists of distributions with characteristic
functions of the type n(t) ' exp(&|t|/) , where 0 (0 , 2]. Note that the standard normal distribution is stable with " = 2, and the standard Cauchy distribution is stable with " = 1. Show that for a random sample X1,X2,...,Xn from a standard stable distribution with parameter ", the random variable Y n ' n &1/'j'1Xj has the same standard stable distribution (this is the reason
n
for calling these distributions stable). 21. 22. Let X and Y be independent standard normally distributed. Derive the distribution of X/Y. Derive the characteristic function of the distribution with density exp(-|x|) /2, ! 4 < x < 4.
156 23. 24. 25. Explain why the moment generating function of the Fm,n distribution does not exist. Prove (4.43). Show that if U1 and U2 are independent U[0,1] distributed then X1 and X2 in (4.42) are
independent standard normally distributed. 26. If X and Y are independent '(1,1) distributed, what is the distribution of X-Y?
Appendices
4.A. Tedious derivations Derivation of (4.37): n((n%1)/2) m (1%x 2/n)(n%1)/2 n (n/2) & 4 dx & n((n%1)/2)
4 4
2 E[Tn ]
'
x 2 /n
dx 1 dx
' '
n((n%1)/2)
4
m (1%x 2/n)(n%1)/2 n (n/2) & 4 1
1%x 2/n
m (1%x 2/n)(n%1)/2 n (n/2) & 4
n((n%1)/2)
m (1%x 2)(n&1)/2 (n/2) & 4
dx & n '
n((n&1)/2%1) (n/2&1) n & n ' . (n/2) ((n&1)/2) n&2
In this derivation I have used (4.35) and the fact that m

4
1 '
hn&2(x)dx ' '
&4
1 dx 2 m (n&2) ((n&2)/2) &4 (1%x /(n&2))(n&1)/2 1 dx . 2 (n&1)/2 m (1 % x ) ((n&2)/2) &4 ((n&1)/2)

4
((n&1)/2)
157 Derivation of (4.39): For m > 0 we have: 1 1 1 exp(&i.t.x)exp(&|t|)dt ' exp(&i.t.x)exp(&t)dt % exp(i.t.x)exp(&t)dt 2 m 2 m 2 m
&m 0 0 m m m
' '
1 1 exp[&(1%i.x)t]dt % exp[&(1&i.x)t]dt 2 m 2 m 1 exp[&(1%i.x)t] 1 exp[&(1&i.x)t] m % 2 &(1%i.x) /00 0 2 &(1&i.x) /00 0

m 0 0
'
1 1 1 1 1 exp[&(1%i.x)m] 1 exp[&(1&i.x)m] % & & 2 (1%i.x) 2 (1&i.x) 2 (1%i.x) 2 (1&i.x) 1 exp(&m) ' & [cos(m.x)&x.sin(m.x)] . (1%x 2) (1%x 2)
Letting m 6 4, (4.39) follows. Derivation of (4.40): m.y (m.x.y/n)m/2&1exp(&(m.x.y/(2n) y n/2&1exp(&y/2) dy m n (m/2)2m/2 (n/2)2n/2
0 4
hm,n(x) '
) Hm,n(x)
'
' n '
m/2
(m/2)(n/2)2m/2%n/2 m 0 m m/2x m/2&1
m m/2x m/2&1
y m/2%n/2&1exp & 1%m.x/n y/2 dy

4
m/2
(m/2)(n/2) 1%m.x/n
m/2%n/2 m 0
z m/2%n/2&1exp &z dz
'
m m/2(m/2%n/2)x m/2&1 n m/2(m/2)(n/2) 1%m.x/n m/2%n/2
, x > 0,
158 Derivation of (4.41): It follows from (4.40) that m (1%x)

0 4
x m/2&1
m/2%n/2
dx '
(m/2)(n/2) , (m/2%n/2)
hence if k < n/2 then x m/2%k&1 x hm,n(x)dx ' dx m/2 m (1%m.x/n)m/2%n/2 m ( m /2) ( n /2) n 0 0
k 4
m m/2(m/2%n/2)
' (n/m)k
(m/2%n/2) x (m%2k)/2&1 (m/2%k)(n/2&k) dx ' (n/m)k ( m % 2 k )/2 % ( n & 2 k )/2 m (m/2)(n/2) (1%x) (m/2)(n/2)
0
'
k&1 k kj'0 (n/m) k kj'1
(m/2%j) (n/2&j)
k&1
where the last equality follows from the fact that by (4.35), (%k) ' ()(j'0 (%j) for " > 0. Thus, n xh (x)dx ' if n $ 3 , m,n ' 4 if n # 2 , m m,n n&2
0 4
m,n '
(4.45)
n 2(m%2) x 2hm,n(x)dx ' if n $ 5 , m m(n&2)(n&4)

0
(4.46) ' 4 if n # 4 .
The results in (4.41) follow now from (4.45) and (4.46).
159 4.B. Proof of Theorem 4.4 For notational convenience I will prove Theorem 4.4 for the case k = 2 only. First note that the distribution of Y is absolutely continuous, because for arbitrary Borel sets B in 2, P[Y 0 B] ' P[G(X) 0 B] ' P[X 0 G &1(B)] ' f(x)dx. mG &1(B)
If B has Lebesgue measure zero then, since G is a one-to-one mapping, the Borel set A = G!1(B) has Lebesgue measure zero. Therefore, Y has density h(y), say, so that for arbitrary Borel sets B in 2, P[Y 0 B] ' mB h(y)dy .
Choose a fixed y0 ' (y0,1 , y0,2)T in the support G(2) of Y such that x0 ' G &1(y0) is a continuity point of the density f of X and y0 is a continuity point of the density h of Y. Let for some positive numbers 1 and 2 , (1 ,2) ' [y0,1 , y0,1%1][y0,2 , y0,2%2] . Then, with 8 the Lebesgue measure, P[Y 0 (1 ,2)] ' f(x)dx # supx0G &1(( , ))f(x) (G &1((1 ,2))) 1 2 mG &1((1 ,2))
1 ,2)
(4.47)
' supy0( and similarly,
f(G (y)) (G ((1 ,2)))
&1
&1
P[Y 0 (1 ,2)] $ infy0(
1 ,2)
f(G &1(y)) (G &1((1 ,2)))
(4.48)
It follows now from (4.47) and (4.48) that P[Y 0 (1 ,2)] 12 (G &1((1 ,2))) 12
h(y0) ' lim lim
190 290
' f(G (y0)) lim lim
&1
190 290
(4.49)
160 It remains to show that the latter limit is equal to |det[J(y0)]|. Denoting G &1(y) ' (g1 (y) , g2 (y))T , it follows from the mean value theorem that for each element gj (y) there exists a j 0 [0,1] depending on y and y0 such that gj (y) ' gj (y0) + Jj(y0%j(y&y0))(y&y0) , where Jj(y) is the j-th row of J(y). Thus, denoting J1(y0%1(y&y0)) & J1(y0) J2(y0%2(y&y0)) & J2(y0)
( ( ( ( (
D0(y) '
(y) & J(y ) , ' J 0 0
(4.50)
say, we have G &1(y) ' G &1(y0) % J(y0)(y&y0) % D0(y)(y&y0) . Now put A = J(y0)&1 and b = y0 & J(y0)&1G &1(y0) . Then G &1(y) ' A &1(y&b) % D0(y)(y&y0) , hence G &1((1 ,2)) ' {x 0 2 : x ' A &1(y&b) % D0(y)(y&y0) , y 0 (1 ,2)} (4.52) (4.51)
The matrix A maps the set (4.52) onto A[G &1((1 ,2))] ' {x 0 2 : x ' y & b % A.D0(y)(y&y0) , y 0 (1 ,2)}
def.
(4.53)
where for arbitrary Borel sets B conformable with a matrix A, A[B] ' {x: x = Ay, y 0 B}. Since the Lebesgue measure is invariant for location shifts (i.e., the vector b in (4.53)) , it follows that A[G &1((1 ,2))] ' {x 0 2 : x ' y % A.D0(y)(y&y0) , y 0 (1 ,2)} (4.54)
Observe from (4.50) that (y) & I A.D0(y) ' J(y0)&1D0(y) ' J(y0)&1J 0 2 (4.55)
161 and (y) ' I . limy6y J(y0)&1J 0 2

0
(4.56)
Then (y)(y&y ) , y 0 ( , )} A[G &1((1 ,2))] ' {x 0 2 : x ' y0 % J(y0)&1J 0 0 1 2 (4.57)
It can be shown, using (4.56), that A[G &1((1 ,2))] (1 ,2)
limlim
190 290
' 1.
(4.58)
Recall from Appendix I that the matrix A can be written as A = QDU, where Q is an orthogonal matrix, D is a diagonal matrix, and U is an upper-triangular matrix with diagonal. elements all equal to 1. Let B = (0,1)(0,1). Then it is not hard to verify in the 22 case that U maps B onto a parallelogram U[B] with the same area as B, hence (U[B]) ' (B) ' 1 . Consequently, the Lebesgue measure of the rectangle D[B] is the same as the Lebesgue measure of the set D[U[B]]. Moreover, an orthogonal matrix rotates a set of point around the origin, leaving all the angles and distances the same. Therefore, the set A[B] has the same Lebesgue measure as the rectangle D[B]: (A[B]) ' (D[B]) ' |det[D]| = |det[A]|. Along the same lines the following more general result can be shown.
Lemma 4.B.1: For a kk matrix A and a Borel set B in k, (A[B]) ' |det[A]|(B) , where 8 is the Lebesgue measure on the Borel sets in k.
162 Thus, (4.58) now becomes A[G &1((1 ,2))] (1 ,2) G &1((1 ,2)) 12
limlim
190 290
' |det[A]|limlim
190 290
' 1,
hence G &1((1 ,2)) 12 1 ' |det[A &1]| ' |det[J(y0)]|. |det[A]|
limlim
190 290
'
(4.59)
Theorem 4.4 follows now from (4.49) and (4.59).
Endnotes 1. Except perhaps on a set with Lebesgue measure zero.
2. The t distribution was discovered by W. S. Gosset, who published the result under the pseudonym Student. The reason for the latter was that his employer, an Irish brewery, did not want its competitors to know that statistical methods were being used. 3. See for example Section 7.1 in Press, Flannery, Teukolsky, and Vetterling (1989).
163
Chapter 5 The Multivariate Normal Distribution and its Application to Statistical Inference
5.1.
Expectation and variance of random vectors Multivariate distributions employ the concepts of expectation vector and variance matrix.
The expected "value", or more precisely, the expectation vector (sometimes also called the "mean vector") of a random vector X ' (x1 , .... , xn)T is defined as the vector of expected values:
def.
E(X) ' (E(x1) , .... , E(xn))T . Adopting the convention that the expectation of a random matrix is the matrix of the expectations of its elements, we can define the variance matrix of X as:1
def.
Var(X) ' E[(X & E(X))(X & E(X))T ] cov(x1 , x1) cov(x1 , x2) ... cov(x1 , xn) ' cov(x2 , x1) : var(x2) : ... cov(x2 , xn) " : .
(5.1)
cov(xn , x1) cov(xn , x2) ... cov(xn , xn) Recall that the diagonal elements of the matrix (5.1) are variances: cov(xj , xj) ' var(xj) . Obviously, a variance matrix is symmetric and positive (semi-)definite. Moreover, note that (5.1) can be written as Var(X) ' E[X X T] & (E[X])(E[X])T . (5.2
164 Similarly, the covariance matrix of a pair of random vectors X and Y is the matrix of covariances of their components:2
def.
Cov(X,Y) ' E[(X & E(X))(Y & E(Y))T] .
(5.3)
Note that Cov(Y,X) ' Cov(X,Y)T . Thus, for each pair X, Y there are two covariance matrices, one being the transpose of the other.
5.2.
The multivariate normal distribution Now let the components of X = (x1 , .... , xn)T be independent standard normally
distributed random variables. Then E(X) ' 0 (0 n) and Var(X) ' In . Moreover, the joint density f(x) = f(x1,...,xn) of X in this case is the product of the standard normal marginal densities: f(x) ' f(x1 , ... , xn) ' k
n j'1
exp &xj / 2 2
'
exp & 1 'j'1xj

n 2
exp & 1 x Tx '

2
( 2)
( 2)n
The shape of this density for the case n = 2 is displayed in Figure 5.1:
Figure 5.1: The bivariate standard normal density on [-3,3][-3,3]
165 Next, consider the following linear transformations of X: Y = F + AX, where = ( 1 , ... , n)T is a vector of constants and A is a non-singular n n matrix with non-random elements. Because A is nonsingular and therefore invertible, this transformation is a one-to-one mapping, with inverse X ' A &1(Y&) . Then the density function g(y) of Y is equal to: g(y) ' f(x)*det(Mx/My)* ' f(A &1y & A &1)*det M(A &1y&A &1)/My * exp & 1 (y&)T(A &1)TA &1(y&) &1 &1 f ( A y & A ) 2 ' f(A &1y&A &1)*det(A &1)* ' ' *det(A)* ( 2)n*det(A)* exp & 1 (y&)T(AA T)&1(y&) '
2
( 2)n *det( AA T )* Observe that F is the expectation vector of Y: E(Y) ' % A E(X) = F. But what is AAT? We know from (5.2) that Var(Y) = E[YYT] ! FFT. Therefore, substituting Y = F + AX yields: Var(Y) ' E[( %AX)( T%X TA T) & T] ' E(X T) A T % A E(X) T % A E(XX T) A T ' AA T , because E(X) = 0 and E[XX T] ' In . Thus, AAT is the variance matrix of Y. This argument gives rise to the following definition of the n-variate normal distribution:
Definition 5.1: Let Y be an n1 random vector satisfying E(Y) = F and Var(Y) = E, where E is nonsingular. Then Y is distributed Nn(F,E) if the density g(y) of Y is of the form exp & 1 (y&)T&1(y&) g(y) '
2
( 2)
det()
(5.4)
166 In the same way as before we can show that a nonsingular (hence one-to-one) linear transformation of a normal distribution is normal itself:
Theorem 5.1: Let Z = a + BY, where Y is distributed Nn(F,E) and B is a non-singular matrix of constants. Then Z is distributed Nn(a + BF, BEBT).
Proof: First, observe that: Z = a + BY implies Y = B!1(Z!a). Let h(z) be the density of Z and g(y) the density of Y. Then h(z) ' g(y)*det(My/Mz)* ' g(B &1z&B &1a)*det(M(B &1z&B &1a)/Mz)* ' g(B &1(z&a)) det(BB T) '
2
g(B &1z&B &1a) *det(B)*
exp & 1 (B &1(z&a)&)T&1(B &1(z&a)&) '

2
'
( 2)n det() det(BB T) . ( 2)n det(BB T)
exp & 1 (z&a&B)T(BB T)&1(z&a&B)
Q.E.D. I will now relax the assumption in Theorem 5.1 that the matrix B is a nonsingular n n matrix. This more general version of Theorem 5.1 can be proved using the moment generating function or the characteristic function of the multivariate normal distribution.
Theorem 5.2: Let Y be distributed Nn(F,E). Then the moment generating function of Y is m(t) ' exp(t T % t T t / 2) , and the characteristic of Y is n(t) ' exp(i.t T & t T t / 2) .
167 Proof: We have exp & 1 (y&)T&1(y&)

2
m(t) '
exp[t y] m
T
( 2)n det()
dy
'
exp & 1 y T&1y & 2 T&1y % T&1 & 2t Ty

2
( 2) det()
n
dy
'
exp & 1 y T&1y & 2( % t)T&1y % ( % t)T&1( % t)

2
( 2)n det() exp

1 2
dy
( % t)T&1( % t) & T&1
'
exp & 1 (y& & t)T&1(y& & t)

2
dy exp t T %
det()
1 T t t 2
Since the last integral is equal to one, the result for the moment generating function follows. The result for the characteristic function follows from n(t) ' m(i.t) . Q.E.D.
Theorem 5.3: Theorem 5.1 holds for any linear transformation Z = a + BY.
Proof: Let Z = a + BY, where B is m n. It is easy to verify that the characteristic function of Z is: nZ(t) ' E[exp(i.t TZ)] ' E[exp(i.t T(a%BY))] ' exp(i.t Ta)E[exp(i.t TBY)] = exp i.(a%B)Tt & t TBB Tt . Theorem 5.3 follows now from Theorem 5.2.. Q.E.D. Note that this result holds regardless whether the matrix BB T is nonsingular or not. In the latter case the normal distribution involved is called "singular":
168 Definition 5.2: An n 1 random vector Y has a singular Nn(F,E) distribution if its characteristic function is of the form nY(t) ' exp(i.t T & t Tt) with E a singular positive semi-define matrix..
Because of the latter, the distribution of the random vector Y involved is no longer absolutely continuous, but the form of the characteristic function is the same as in the nonsingular case, and that is all that matters. For example, let n = 2 and ' 0 0 , ' 1 0 0 2 ,
where 2 > 0 but small. The density of the corresponding N2(F,E) distribution of Y ' (Y1 , Y2)T is exp(&y1 /2) exp(&y2 /(22)) . f(y1 , y2|) ' 2 2 Then lim90 f(y1 , y2|) ' 0 if y2 0 , lim90 f(y1 , y2|) ' 4 if y2 ' 0 . Thus, a singular multivariate normal distribution does not have a density. In Figure 5.2 the density (5.5) for the near-singular case 2 ' 0.00001 is displayed. The height of the picture is actually rescaled to fit in the the box [-3,3][-3,3][-3,3]. If we let F approach zero the height of the ridge corresponding to the marginal density of Y1 will increase to infinity.
2 2
(5.5)
169
Figure 5.2: Density of a near-singular normal distribution on [-3,3][-3,3]
The next theorem shows that uncorrelated multivariate normal distributed random variables are independent. Thus, while for most distributions uncorrelatedness does not imply independence, for the multivariate normal distribution it does.
Theorem 5.4: Let X be n-variate normally distributed, and let X1 and X2 be sub-vectors of components of X. If X1 and X2 are uncorrelated, i.e., Cov(X1 , X2) ' O , then X1 and X2 are independent.
Proof: Since X1 and X2 cannot have common components, we may without loss of generality assume that X ' (X1 , X2 )T , X1 0 k , X2 0 m . Partition the expectation vector and variance matrix of X conformably as:
T T
170 1 2 11 12 21 22
E(X) '
, Var(X) '
Then E12 = O and E21 = O because they are covariance matrices, and X1 and X2 are uncorrelated, hence the density of X is: x1 x2 1 2 det
T
exp & f(x) ' f(x1 , x2) '
1 2
&
11 0 11 0
0 22
&1
x1 x2
&
1 2
( 2)n
0 22
&1
'
exp & 1 (x1& 1)T11 (x1& 1)

2
&1
( 2) det(11)
k
exp & 1 (x2& 2)T22 (x2& 2)

2
( 2)m det(22)
This implies independence of X1 and X2. Q.E.D.
5.3.
Conditional distributions of multivariate normal random variables Let Y be a scalar random variable and X be a k-dimensional random vector. Assume that Y X Y X YY YX XY XX
- Nk%1
where Y ' E(Y) , X ' E(X) , and YY ' Var(Y) , YX ' Cov(Y , X T) ' E[(Y&E(Y))(X&E(X))T] , XY ' Cov(X , Y) ' E(X&E(X))(Y&E(Y)) ' YX , XX ' Var(X) .
T
171 In order to derive the conditional distribution of Y, given X, let U ' Y & & TX , where " is a scalar constant and $ is a k 1 vector of constants, such that E(U) = 0 and U and X are independent. It follows from Theorem 5.1 that U X & 0 1 &B T 0 Ik , Y X 1 & T 0 Ik YY YX XY XX 1 0T
'
- Nk%1
& % Y & T X X
& Ik
The variance matrix involved can be rewritten as: YY&YX&TXY%TXX XY&XX YX&TXX XX
Var
U X
'
(5.6)
Next, choose $ such that U and X are uncorrelated and hence independent. In view of (5.6) a necessary and sufficient condition for that is: XY & XX ' 0, hence ' XXXY . Moreover, E(U) = 0 if ' Y & T X . Then YY & YX & TXY % TXX ' YY & YXXXXY , YX & TXX ' 0T , XY & XX ' 0 ,
&1 &1
and consequently U X 0 X , YY&YXXXXY 0T 0 XX

&1
- Nk%1
(5.7)
Thus U and X are independent normally distributed, and consequently E(U*X) ' E(U) = 0. Since Y ' % TX % U , we now have E(Y*X) ' % T E(X*X) % E(U*X) ' % TX .
172 Moreover, it is easy to verify from (5.7) that the conditional density of Y, given X = x, is exp & 1 (y&&Tx)2 / u
2 2
f(y*x) '
u 2
where u ' YY & YXXXXY .
&1
Furthermore, note that u is just the conditional variance of Y, given X: u ' var(Y*X) ' E Y & E(Y*X) 2*X . Summarizing:
2 def.
Theorem 5.5: Let Y X Y X YY YX XY XX
- Nk%1
where Y0 , X0k , and EXX is nonsingular. Then conditionally on X, Y is normally distributed with conditional expectation E(Y*X) ' %TX , where ' XXXY and ' Y&T X , and conditional variance var(Y*X) ' YY & YXXXXY .
&1 &1
The result in Theorem 5.5 is the basis for linear regression analysis. Suppose that Y measures an economic activity that is partly caused or influenced by other economic variables, measured by the components of the random vector X. In applied economics the relation between Y, called the dependent variable, and the components of X, called the independent variables or the regressors, is often modeled linearly as Y = " + $TX + U, where " is the intercept, $ is the vector of slope parameters (also called regression coefficients), and U is an error term which is
173 usually assumed to be independent of X and normally N(0,F2) distributed. Theorem 5.5 shows that if Y and X are jointly normally distributed, then such a linear relation between Y and X exists.
5.4.
Independence of linear and quadratic transformations of multivariate normal random variables Let X be distributed Nn(0,In), i.e., X is n-variate standard normally distributed. Consider
the linear transformations Y = BX, where B is a k n matrix of constants, and Z = CX, where C is an m n matrix of constants. It follows from Theorem 5.4 that Y Z 0 0 BB T BC T CB T CC T
- Nk%m
Then Y and Z are uncorrelated and therefore independent if and only if CBT = O. More generally we have:
Theorem 5.6: Let X be distributed Nn(0,In), and consider the linear transformations Y = b + BX, where b is a k 1 vector and B a k n matrix of constants, and Z = c + CX, where c is an m 1 vector and C an m n matrix of constants. Then Y and Z are independent if and only if BCT = O.
This result can be used to set forth conditions for independence of linear and quadratic transformations of standard normal random vectors:
Theorem 5.7: Let X and Y be defined as in Theorem 5.6, and let Z = XTCX, where C is a symmetric n n matrix of constants. Then Y and Z are independent if BC = O.
174 Proof: First, note that the latter condition only makes sense if C is singular, as otherwise B = O. Thus, let rank(C) = m < n. We can write C ' QQ T , where 7 is a diagonal matrix with the eigenvalues of C on the diagonal, and Q is the orthogonal matrix of corresponding eigenvectors. Let V = QTX, which is Nn(0,In) distributed because QQT = In. Since n ! m eigenvalues of C are zero, we can partition Q, 7 and V such that 1 O O O V1 V2 Q1 X
T Q2 X T
Q ' (Q1 , Q2) , '
, V '
'
, Z ' V1 1V1 ,
where 71 is the diagonal matrix with the m nonzero eigenvalues of C on the diagonal. Then 1 O O O Q1
T
BC ' B(Q1 , Q2)
T Q2
' BQ11Q1 ' O
implies BQ11 ' BQ11Q1 Q1 ' O (because Q TQ ' In implies Q1 Q1 ' Im) , which in its turn implies that BQ1 = O. The latter is a sufficient condition for the independence of V1 and Y, hence of the independence of Z and Y. Q.E.D. Finally, consider the conditions for independence of two quadratic forms of standard normal random vectors:
Theorem 5.8: Let X - Nn(0,In) , Z1 ' X TAX , Z2 ' X TBX , where A and B are symmetric n n matrices of constants. Then Z1 and Z2 are independent if and only if AB = O.
The proof of Theorem 5.8 is not difficult but quite lengthy and therefore given in the
175 Appendix 5.A.
5.5.
Distributions of quadratic forms of multivariate normal random variables As we will see in Section 5.6 below, quadratic forms of multivariate normal random
variables play a key-role in statistical testing theory. The two most important results are stated in Theorems 5.9 and 5.10:
Theorem 5.9: Let X be distributed Nn(0,E), where E is nonsingular. Then XTE!1X is distributed as P2 n.
distributed, hence Y1,...,Yn are i.i.d. N(0,1) and thus X T&1X ' Y TY ' 'j'1Y j - n . Q.E.D.
n 2 2
Proof: Denote Y ' (Y1 ,.... , Y n)T ' &X . Then Y is n-variate standard normally
The next theorem employs the concept of an idempotent matrix. Recall from Appendix I that a square matrix M is idempotent if M2 = M. If M is also symmetric, we can write M = Q7QT, where 7 is the diagonal matrix of eigenvalues of M and Q is the corresponding orthogonal matrix of eigenvectors. Then M2 = M implies 72 = 7, hence the eigenvalues of M are either 1 or 0. If all eigenvalues are 1, then 7 = I, hence M = I. Thus the only nonsingular symmetric idempotent matrix is the unit matrix. Consequently, the concept of a symmetric idempotent matrix is only meaningful if the matrix involved is singular. The rank of a symmetric idempotent matrix M equals the number of nonzero eigenvalues, hence trace(M) = trace(Q7QT) = trace(7QTQ) = trace(7) = rank(7) = rank(M), where trace(M) is defined as the sum of the diagonal elements of M. Note that we have used the property trace(AB)
176 = trace(BA) for conformable matrices A and B.
Theorem 5.10: Let X be distributed Nn(0,I), and let M be a symmetric idempotent n n matrix of constants with rank k. Then XTMX is distributed P2 k.
Proof: We can write Ik O O O where Q is the orthogonal matrix of eigenvectors. Since Y ' (Y1 , ... , Y n)T ' Q TX - Nn(0 , I) we now have Ik O O O Q.E.D. Y ' j Y j - k .
k 2 2 j'1
M ' Q
Q T,
X TMX ' Y T
5.6. 5.6.1
Applications to statistical inference under normality Estimation Statistical inference is concerned with parameter estimation and parameter inference. The
latter will be discussed in the next subsections. Loosely speaking, an estimator of a parameter is a function of the data which serves as an approximation of the parameter involved. For example, if X1, X2,...,Xn is a random sample from the N(F,F2) distribution then the sample mean X ' (1 / n)'j'1Xj may serve as an estimator of the
n
unknown parameter (the population mean). More formally, given a data set {X1, X2,...,Xn } for
177 which the joint distribution function depends on an unknown parameter (vector) 2, an estimator = gn(X1,...,Xn) of the data which serves as an of 2 is a Borel measurable function approximation of 2. Of course, the function gn should not depend on unknown parameters itself. In principle we can construct many functions of the data that may serve as an approximation of an unknown parameter. For example, one may consider using X1 only as an estimator of . So the question arises which function of the data should be used. In order to be able to select among the many candidates for an estimator, we need to formulate some desirable properties of estimators. The first one is unbiasedness:
of a parameter (vector) 2 is unbiased if E[ ] = 2. Definition 5.3: An estimator
The unbiasedness property is not specific to a particular value of the parameter involved, but should hold for all possible values of this parameter, in the sense that if we draw a new data set from the same type of distribution but with a different parameter value, the estimator should stay unbiased. In other words, if the joint distribution function of the data is Fn(x1,...,xn|2), where 2 0 1 is an unknown parameter (vector) in a parameter space 1, i.e., the space of all possible values ' g (X ,....,X ) is an unbiased estimator of 2, then g (x ,....,x )dF (x ,....,x |) ' of 2, and n 1 n n n 1 n m n 1 for all 2 0 1. Note that in the above example both X and X1 are unbiased estimators of . Thus, we need a further criterion in order to select an estimator. This criterion is efficiency:
of an unknown scalar parameter 2 is efficient if for all Definition 5.4: An unbiased estimator
178 , var( ) # var( ) . In the case that 2 is a parameter vector the other unbiased estimators ) & Var( ) is a positive semi-definite matrix.. latter reads: Var(
In our example, X1 is not an efficient estimator of , because var(X1) ' 2 and var(X) ' 2 / n . But is X efficient? In order to answer this question, we need to derive the minimum variance of an unbiased estimator, as follows. For notational convenience, stack the data in a vector X. Thus, in the univariate case, X = (X1, X2,...,Xn )T, and in the multivariate case, X = (X1 ,....,Xn )T . Assume that the joint distribution of X is absolutely continuous with density ' g (X) be an fn(x|2), which for each x is twice continuously differentiable in 2. Moreover, let n unbiased estimator of 2. Then g (x)f (x|)dx ' m n n Furthermore, assume for the time being that 2 is a scalar, and let d d g n(x)fn(x|)dx ' gn(x) fn(x|)dx . m d m d Conditions for (5.9) can be derived from the mean-value theorem and the dominated convergence theorem. In particular, (5.9) is true for all 2 in an open set 1 if |g (x)|sup0 |d 2f n(x|)/(d)2|dx < 4. m n Then if follows from (5.8) and (5.9) that d d gn(x) ln(fn(x|)) f n(x|)dx ' gn(x) f n(x|)dx ' 1 m m d d Similarly, if (5.8)
T T
(5.9)
(5.10)
179 d d fn(x|)dx ' f (x|)dx m d n d m which is true for all 2 in an open set 1 for which f (x|)dx = 1, we have mn d d ln(fn(x|)) f n(x|)dx ' f (x|)dx ' 0 . m d m d n (5.12) m (5.11)
sup0 |d 2fn(x|)/(d)2|dx < 4 , then, since
' d ln(f (X|))/d , it follows now from (5.10) that E[ . ] ' 1 and from (5.12) that Denoting n ] = 0. Therefore, cov( , ) ' E[ . ] & E[ ]E[ ] ' 1 . Since by the Cauchy-Schwartz E[ , )| # var( ) var( ) , we now have that var( ) $ 1/var( ): inequality, |cov( ) $ var( 1 E d ln(f n(X|))/d
2
(5.13)
This result is known as the Cramer-Rao inequality, and the right-hand side of (5.13) is called the Cramer-Rao lower bound. More generally we have:
Theorem 5.11: (Cramer-Rao) Let fn(x|2) be the joint density of the data, stacked in a vector X, be an unbiased estimator of 2. Then Var( ) = where 2 is a parameter vector. Let E (Mln(fn(X|)/MT) (Mln(f n(X|)/M)
&1
% D , where D is a positive semi-definite matrix.
Now let us return to our problem whether the sample mean X of a random sample from the N(F,F2) distribution is an efficient estimator of . In this case the joint density of the sample is fn(x |,2) ' (j'1exp(&(xj&)2/2) / 22 , hence Mln(fn(X|,2)) /M ' 'j'1(Xj&)/2 and
n n
thus the Cramer-Rao lower bound is
180 1 E Mln(f n(X|, )) /M

2 2
' 2/n.
(5.14)
This is just the variance of the sample mean X , hence X is an efficient estimator of . This result holds for the multivariate case as well:
sample mean X ' (1/n)'j'1Xj is an unbiased and efficient estimator of .

n
Theorem 5.12: Let X1, X2,...,Xn be a random sample from the Nk[ , ] distribution. Then the
The sample variance of a random sample X1, X2,...,Xn from a univariate distribution with expectation and variance 2 is defined by S 2 ' (1/(n&1))'j'1(Xj&X )2 ,
n
(5.15)
which serves as an estimator of F2. An alternative form of the sample variance is 2 ' (1/n)'j'1(Xj&X )2 '
n
n&1 2 S , n
(5.16)
but as I will show for the case of a random sample from the N(F,F2) distribution, (5.15) is an unbiased estimator, and (5.16) is not:
Theorem 5.13: Let S2 be the sample variance of a random sample X1,...,Xn from the N(F,F2) distribution. Then (n!1)S2/F2 is distributed n&1 .
2
The proof of Theorem 5.13 is left as an exercise. Since the expectation of the n&1 distribution is n!1, this result implies that E(S 2) ' 2 , whereas by (5.16), E( 2) ' 2(n&1)/n . Moreover,
181 since the variance of the n&1 distribution is 2(n!1), it follows from Theorem 5.13 that Var(S 2) ' 24/(n&1) . (5.17)
2
The Cramer-Rao lower bound for an unbiased estimator of 2 is 24 / n , so that S 2 is not efficient, but it is close if n is large. For a random sample X1, X2,...,Xn from a multivariate distribution with expectation vector and variance matrix G the sample variance matrix takes the form
T ' (1/(n&1))'n j'1(Xj&X )(X j&X ) .
(5.18)
This is also an unbiased estimator of ' Var(Xj) , even if the distibution involved is not normal.
5.6.2
Confidence intervals Since estimators are approximations of unknown parameters, the question arises how
close they are. I will answer this question for the sample mean and the sample variance in the case of a random sample X1, X2,...,Xn from the N(F,F2) distribution. It is almost trivial that X - N( , 2/n) , hence n(X&)/ - N(0,1) . Therefore, for given " 0 (0,1) there exists a $ > 0 such that m
(5.19)
P [*X& * # / n] ' P * n(X&)/* # '
exp(&u 2 / 2) 2
du ' 1 & .
(5.20)
&
For example, if we choose " = 0.05 then $ = 1.96, so that in this case P [X&1.96/ n # # X%1.96/ n] ' 0.95 The interval [X&1.96/ n , X%1.96/ n] is called the 95% confidence interval of . If F would be
182 known, then this interval can be computed and will tell us how close X and are, with a margin of error of 5%. But in general F is not known, so how do we proceed then? In order to solve this problem, we need the following corollary of Theorem 5.7:
Theorem 5.14: Let X1, X2,...,Xn be a random sample from the N(F,F2) distribution. Then the sample mean X and the sample variance S 2 are independent.
Proof: Observe that X( ' ((X1&)/ , (X2&)/ , ..... , (Xn&)/)T - N n(0 , In) , X ' + (/n , ... , /n)X( ' b % BX( , say, and )/ (X1&X : )/ (Xn&X The latter implies that (n&1)S 2 / 2 ' ' I & 1 1 1 (1 , 1 , ... , 1) X( ' CX( , say . n : 1
T X( C TCX(
' X( C 2X( ' X( CX( , because C is
symmetric and idempotent, with rank(C) = trace(C) = n ! 1. Therefore, by Theorem 5.7 the sample mean and the sample variance are independent if BC = 0, which in the present case is equivalent to the condition CBT = 0. The latter is easily verified: 1 CB T ' n I & 1 1 (1 , ..... , 1) n : 1 Q.E.D. It follows now from (5.19), Theorems 5.13 and 5.14, and the definition of the Student t distribution that: 1 1 : 1 ' n 1 1 : 1 & 1 1 1 n n : 1 ' 0
183 & ) / S - t . Theorem 5.15: Under the conditions of Theorem 5.14, n(X n&1
Recall from Chapter 4 that the tn!1 distribution has density hn&1(x) ' where (y) ' m0
4 y&1
(n/2) (n&1) ((n&1)/2) (1%x 2/(n&1))n/2
(5.21)
exp(&x)dx , y > 0. Thus, similarly to (5.20), for each " 0 (0,1) and sample
size n there exists a $n > 0 such that P [*X& * # nS/ n] ' m&n
n
hn&1(u)du ' 1 & ,
(5.22)
so that [X&nS/ n , X%nS/ n] is now the (1!")100% confidence interval of Similarly, on the basis of Theorem 5.13 we can construct confidence intervals of 2 . Recall from Chapter 4 that the n&1 distribution has density gn&1(x) ' x (n&1)/2&1exp(&x/2) ((n&1)/2)2(n&1)/2 .
2
For given " 0 (0,1) and sample size n we can choose $1,n < $2,n be such that P[(n&1)S 2/2,n # 2 # (n&1)S 2/1,n] ' P[1,n # (n&1)S 2/2# 2,n] ' m1,n
2,n
gn&1(u)du ' 1 & .
(5.23)
There are different ways to choose $1,n and $2,n such that the last equality in (5.23) holds. Clearly, the optimal choice is such that 1,n&2,n is minimal because it will yield the smallest confidence interval, but that is computationally complicated. Therefore, in practice $1,n and $2,n
&1 &1
184 are often chosen such that m0

1,n
gn&1(u)du ' /2 ,
m2,n
4
gn&1(u)du ' /2 .
(5.24)
A practical point is how to solve the integral equations in (5.20), (5.22) and (5.24). Most statistics and econometrics textbooks contain tables from which you can look op the values of the $s involved, given ". Moreover, there are various web pages from which you can download programs to calculate these values.3
5.6.3
Testing parameter hypotheses Suppose you consider starting up a business to sell a new product in the USA, say a
particular type of European car which is not yet imported in the US. In order to determine whether there is a market for this car in the US, you have selected randomly n persons from the population of potential buyers of this car. Each person j in the sample is asked how much he or she would be willing to pay for this car. Let the answer be Yj. Moreover, suppose that the cost of importing this car is a fixed amount Z per car. Denote Xj ' ln(Y j / Z) , and assume that Xj is N(F,F2) distributed. If F > 0 then your planned car import business will be profitable, otherwise you should forget about this idea. In order to decide whether F > 0 or F # 0, you need a decision rule based on the random sample X = (X1, X2,...,Xn )T. Any decision rule takes the following form. Given a subset C of n, to be determines below, decide that F > 0 if X 0 C, and decide that F # 0 if X C. Thus, you decide that the hypothesis F # 0 is true if I(X 0 C) ' 0 , and you decide that the hypothesis F > 0 is true if I(X 0 C) ' 1 . In this case the hypothesis F # 0 is called the null hypothesis, usually
185 denoted by H0: F # 0, and the hypothesis F > 0 is called the alternative hypothesis, denoted by H1: F > 0. The procedure itself is called a statistical test. This decision rule yields two types of errors. The first one, called the type I error, is that you decide that H1 is true while in reality H0 is true. The other error, called the type II error, is that you decide that H0 is true while in reality H1 is true. Both errors come with costs. If the type I error occurs you will incorrectly assume that your car import business will be profitable, so that you will loose your investment if you start up you business. If the type II error occurs you will forgo a profitable business opportunity. Clearly, the type I error is the more serious of the two. Now choose C such that X 0 C if and only if n ( X / S ) > for some fixed $ > 0. Then n / S > ] (5.25)
P[X 0 C] ' P[ n(X / S ) > ] ' P[ n(X&) / S % ' P[ n (X&) / % ' m&4
4
n / > .S/]
P[S/ < (u % n / )/]exp[&u 2 / 2]/ 2 du ,
where the last equality follows from Theorem 5.14 and (5.19). If F # 0 this probability is the probability of a type I error. Clearly, the probability (5.25) is an increasing function of , hence the maximum probability of a type I error is obtained for = 0. But if = 0 then it follows from Theorem 5.15 that n(X / S ) - tn&1 , hence max # 0P[X 0 C] ' m
4
hn&1(u)du ,
(5.26)
where hn&1 is the density of the tn&1 distribution. See (5.21). The probability (5.26) is called the size of the test of the null hypothesis involved, which is the maximum risk of a type I error, and "100% is called the significance level of the test. Depending on how risk averse you are, you
186 have to choose a size " 0 (0,1), and therefore you have to choose $ = $n such that mn
4
hn&1(u)du ' . This value $n is called the critical value of the test involved, and since it is
based on the distribution of n(X / S ) , the latter is called the test statistic involved. Replacing $ in (5.25) by $n , 1 minus the probability of a type II error is a function of /F > 0: m
4
n(/) '
P[S/ < (u % n / )/n]
exp(&u 2 / 2) 2
du , > 0 .
(5.27)
& n /
This function is called the power function, which is the probability of correctly rejecting the null hypothesis H0 in favor of the alternative hypothesis H1. Consequently, 1 & n(/) , > 0 , is the probability of a type II error. The test in this example is called a t-test, because the critical value $n is derived from the t-distribution. A test is said to be consistent if the power function converges to 1 as n 6 4 for all values of the parameter(s) under the alternative hypothesis. Using the results in the next chapter it can be shown that the above test is consistent: limn64n(/) ' 1 if > 0. (5.28)
Now let us consider the test of the null hypothesis H0: F = 0 against the alternative hypothesis H1: F 0. Under the null hypothesis, n(X / S ) - tn&1 exactly. Given the size " 0
(0,1), choose the critical value $n > 0 as in (5.22). Then H0 is accepted if | n(X / S )| # n and rejected in favor of H1 if | n(X / S )| > n . The power function of this test is
187 n(/) ' P[S/ < | u % m&4

4
n/ |/ n]exp[&u 2 / 2]/ 2 du , 0 .
(5.29)
This test is known as is the two-sided t-test. Also this test is consistent: limn64n(/) ' 1 if 0. (5.30)
5.7. 5.7.1
Applications to regression analysis The linear regression model Consider a random sample Zj ' (Y j , Xj )T , j = 1,2,...,n, from a k-variate nonsingular
T
normal distribution, where Y j 0 , Xj 0 k&1 . We have seen in Section 5.3 that we can write Y j ' % Xj % Uj , Uj - N(0 , 2) , j ' 1,..,n,
T
(5.31)
where Uj ' Y j & E[Y j|Xj] is independent of Xj. This is the classical linear regression model, where Yj is the dependent variable, Xj is the vector of independent variables, also called the regressors, and Uj is the error term. This model is widely used in empirical econometrics, even in the case where Xj is not known to be normally distributed. Denoting
T
Y1 Y ' Yn
1 X1 !
! , X ' !
, 0 '
U1 , U ' ! , Un
1 Xn
model (5.31) can be written in vector/matrix form as Y ' X0 % U , U|X - Nn[0 , 2In] , where U|X is a short-hand notation for " U conditional on X". In the next subsections I will address the problems how to estimate the parameter vector (5.32)
188 20 and how to test various hypotheses about 20 and its components.
5.7.2
Least squares estimation Observe that E[(Y & X)T(Y & X)] ' E[(U % X(0&))T(U % X(0&))] ' E[U TU] % 2(0&)TE X TE[U | X] % (0&)T E[X TX] (0&) ' n.2 % (0&)T E[X TX] (0&) . (5.33)
Hence it follows from (5.33) that4 0 ' argmin E[(Y & X)T(Y & X)] ' E[X TX] &1E[X TY] ,
0k
(5.34)
provided that the matrix E[X TX] is nonsingular. However, the nonsingularity of the distribution of Zj ' (Y j , Xj )T guarantees that E[X TX] is nonsingular, because it follows from Theorem 5.5 that the solution (5.34) is unique if GXX = Var(Xj) is nonsingular. The expression (5.34) suggests to estimate 20 by the Ordinary5 Least Squares (OLS) estimator ' argmin (Y & X)T(Y & X) ' X TX &1X TY .
0k T
(5.35)
It follows easily from (5.32) and (5.35) that & ' X TX &1X TU , 0 (5.36)
is conditionally unbiased: E[ |X] ' , and therefore also unconditionally unbiased: hence 0 ] ' . More generally, E[ 0
189 |X - N [ , 2(X TX)&1] . k 0 is not normal. Of course, the unconditional distribution of Note that the OLS estimator is not efficient, because 2(E[X TX])&1 is the Cramer-Rao ) ' 2E[(X TX)&1] 2(E[X TX])&1. lower bound of an unbiased estimator of (5.37), and Var( of However, the OLS estimator is the most efficient of all conditionally unbiased estimators (5.37) that are linear functions of Y. In other words, the OLS estimator is the Best Linear Unbiased Estimator (BLUE). This result is known as the Gauss-Markov theorem: (5.37)
Theorem 5.16: (Gauss-Markov theorem) Let C(X) be a kn matrix whose elements are Borel ' C(X)Y . If E[ |X] ' then for measurable functions of the random elements of X, and let 0 |X] ' 2C(X)C(X)T ' 2(X TX)&1 + D. some positive semi-definite kk matrix D, Var[
= 20 + Proof: The conditional unbiasedness condition implies that C(X)X = Ik, hence |X) ' 2C(X)C(X)T . Now C(X)U, and thus Var( D ' 2[C(X)C(X)T & (X TX)&1] ' 2[C(X)C(X)T & C(X)X(X TX)&1X TC(X)T] ' 2C(X)[In & X(X TX)&1X T]C(X)T ' 2C(X)MC(X)T , say, where the second equality follows from the unbiasedness condition CX = Ik. The matrix M ' In & X X TX &1X T (5.38)
is idempotent, hence its eigenvalues are either 1 or 0. Since all the eigenvalues are non-negative, M is positive semi-definite, and so is C(X)MC(X)T. Q.E.D. Next, we need an estimator of the error variance F2. If we would observe the errors Uj,
then we could use the sample variance S 2 ' (1/(n&1))'j'1(Uj&U )2 of the Uj s as an unbiased
n
190
estimator. This suggests to use OLS residuals, 1 Xj ,
' Y & X T , where X ' U j j j j
(5.39)
instead of the actual errors Uj in this sample variance. Taking into account that 2 ' (1/(n&1))'n 2 the feasible variance estimator involved takes the form S j'1Uj . However, this estimator is not unbiased, but a minor correction will yield an unbiased estimator of F2, namely , S 2 ' (1/(n&k))'j'1U j
n 2
/ 0, 'j'1U j
n
(5.40)
(5.41)
which is called the OLS estimator of F2: The unbiasedness of this estimator is a by-product of the following more general result, which is related to the result of Theorem 5.13.
Theorem 5.17: Conditional on X and well as unconditionally, (n&k)S 2/2 - n&k , hence E[S 2] ' 2 .
Proof: Observe that

2 n 2 n n T 2 T 'j'1U j ' 'j'1(Y j & X j ) ' 'j'1 Uj & X j (&0 ) n 2 n T ( & ) % ( & )T 'n T ' 'j'1Uj & 2 'j'1UjX j'1X j Xj (&0 ) j 0 0
(5.42)
& ) % ( & ) X TX ( & ) ' U TU & 2U TX( 0 0 0 ' U TU & U TX(X TX)&1X TU ' U TMU ,
where the last two equalities follow from (5.36) and (5.38), respectively. Since the matrix M is
191 idempotent, with rank rank(M) ' trace(M) ' trace(In) & trace X X TX &1X T ' n&k it follows from Theorem 5.10 that conditional on X, (5.42) divided by 2 has a n&k distribution:
n 2 2 2 'j'1U j / | X - n&k . 2
' trace(In) & trace X TX &1X TX
(5.43)
It is left as an exercise to prove that (5.43) implies that also the unconditional distribution of (5.42) divided by 2 is n&k :
2
/ 2 - n&k . 'j'1U j
n 2 2 2
(5.44)
Since the expectation of the n&k distribution is n!k, it follows from (5.44) that the OLS estimator (5.41) of 2 is unbiased. Q.E.D. Next, observe from (5.38) that XTM = O, so that by Theorem 5.7 (XTX)!1XTU and UTMU are independent, conditionally on X, i.e. P[X TU # x and U TMU # z|X] ' P[X TU # x|X].P[U TMU # z|X] , x 0 k , z $ 0 . Consequently,
and S 2 are independent, Theorem 5.18: Conditional on X,
but unconditionally they can be dependent. Theorems 5.17 and 5.18 yield two important corollaries, which I will state in the next theorem. These results play a key role in statistical testing.
192 Theorem 5.19: (a) Let c 0 k be a given nonrandom vector. Then & ) c T( 0 S c T(X TX)&1c (b)
- tn&k .
(5.45)
Let R be a given nonrandom mk matrix with rank m # k. Then & )T R T R X TX &1R T ( 0 m.S 2
&1
& ) R ( 0
- Fm,n&k .
(5.46)
& )|X - N[0 , 2c T(X TX)&1c] , hence Proof of (5.45): It follows from (5.37) that c T( 0 /0 X - N[0 , 1] . c (X X) c 0 & ) c T( 0
T T &1
(5.47)
If follows now from Theorem 5.18 that conditional on X the random variable in (5.47) and S2 are independent, hence it follows from Theorem 5.17 and the definition of the t-distribution that (5.45) is true conditional on X, and therefore also unconditionally. & )|X - N [0 , 2R(X TX)&1R T] , hence Proof of (5.46): It follows from (5.37) that R( 0 m it follows from Theorem 5.9 that & )T R T R X TX &1R T ( 0
2 &1
& ) R ( 0
/0 X - m . 0
2
(5.48)
Again it follows from Theorem 5.18 that conditional on X the random variable in (5.48) and S2 are independent, hence it follows from Theorem 5.17 and the definition of the F-distribution that (5.46) is true conditional on X, and therefore also unconditionally. Q.E.D.
193 Note that the results in Theorem 5.19 do not hinge on the assumption that the vector Xj in model (5.31) has a multivariate normal distributed. The only conditions that matter for the validity of Theorem 5.19 are that in (5.32), U|X - Nn(0,2In) , and P[0 < det(X TX) < 4] ' 1.
5.7.3
Hypotheses testing Theorem 5.19 is the basis for hypotheses testing in linear regression analysis. First,
consider the problem whether a particular components of the vector Xj of explanatory variables in model (5.31) has an effect on Yj or not. If not, the corresponding component of $ is zero. Each component of $ corresponds to a component i,0, i > 0, of 0 . Thus, the null hypothesis involved is H0: i,0 ' 0 . (5.49)
be component i of , and let the vector e be column i of the unit matrix I . Then it Let i i k follows from Theorem 5.19(a) that under the null hypothesis (5.49), i S ei (X X) ei The statistic ti in (5.50) is called the t-statistic or t-value of the coefficient i,0. If it conceivable that i,0 can take negative or positive values, the appropriate alternative hypothesis is H1: i,0 0 . (5.51)
T T &1
ti '
- tn&k .
(5.50)
Given the size " 0 (0,1) of the test, the critical value ( corresponds to P[|T| > ] ' , where T - tn&k . Thus, the null hypothesis (5.49) is accepted if |ti| # , and rejected in favor of the alternative hypothesis (5.51) if |ti| > . In the latter case we say that i,0 is significant at the "100% significance level. This test is called the two-sided t-test.
194 If the possibility that i,0 is negative can be excluded, the appropriate alternative hypothesis is H1 : i,0 > 0 .
%
(5.52)
Given the size " the critical value (+ involved now corresponds to P[T > %] ' , where again T - tn&k . Thus the null hypothesis (5.49) is accepted if ti # % , and rejected in favor of the alternative hypothesis (5.52) if ti > % . This is the right-sided t-test. Similarly, if the possibility that i,0 is positive can be excluded, the appropriate alternative hypothesis is H1 : i,0 < 0 .
&
(5.53)
Then the null hypothesis (5.49) is accepted if ti $ &% , and rejected in favor of the alternative hypothesis (5.53) if ti < &% . This is the left-sided t-test. If the null hypothesis (5.49) is not true, then it can be shown, using the results in the next chapter, that for n 6 4 and arbitrary M > 0, P[ti > M] 6 1 if i,0 > 0 and P[ti < &M] 6 1 if i,0 < 0. Therefore, the t-tests involved are consistent. Finally, consider a null hypothesis of the form H0: R0 ' q , where R is a given mk matrix with rank m#k , and q is a given m1 vector . (5.54)
For example, the null hypothesis that the parameter vector $ in model (5.31) is a zero vector corresponds to R ' (0 , Ik&1) , q ' 0 0 k&1 , m ' k&1 . This hypothesis implies that none on the components of Xj have any effect on Yj. In that case Yj = " + Uj, and since Uj and Xj are independent, so are Yj and Xj. It follows from Theorem 5.19(b) that under the null hypothesis (5.54),
195
T &1 T T ' (R&q) R X X R F m.S 2 &1
&q) (R
- Fm,n&k .
(5.55)
Given the size " the critical value ( is chosen such that P[F > ] ' , where F - Fm,n&k . # , and rejected in favor of the alternative Thus the null hypothesis (5.54) is accepted if F > . For obvious reasons, this test is called the F test. Moreover, it can hypothesis R0 q if F be shown, using the results in the next chapter, that if the null hypothesis (5.54) is false then for > M] ' 1 . Thus the F test is a consistent test. any M > 0, limn64P[F
5.8. 1.
Exercises Let Y X Determine E(Y,X). Determine var(U), where U = Y ! E(Y,X). Why are U and X independent? Let X be n-variate standard normally distributed, and let A be a non-stochastic nk matrix 1 0 4 1 1 1
- N2
(a) (b) (c) 2.
with rank k < n. The projection of X on the column space of A is a vector p such that the following two conditions hold: (1) (2) (a) (b) p is a linear combination of the columns of A; the distance between X and p, 2X&p2 ' (X&p)T(X&p) , is minimal. Show that p = A(ATA)-1ATX. Is it possible to write down the density of p? If yes, do it. If no, why not?
196 (c) (d) (e) 3. 4. Show that 2p22 ' p Tp has a P2 distribution. Determine the degrees of freedom involved. Show that 2X&p22 has a P2 distribution. Determine the degrees of freedom involved. Show that 2p2 and 2X&p2 are independent. Prove Theorem 5.13. Show that (5.11) is true for 2 in an open set 1 if d 2fn(x|)/(d)2 is for each x continuous sup |d 2fn(x|)/(d)2|dx < 4 . Hint. Use the mean value theorem and the dominated m 0
on 1, and
convergence theorem. 5. Show that for a random sample X1, X2,...,Xn from a distribution with expectation and
variance 2 the sample variance (5.15) of is an unbiased estimator of 2 , even if the distribution involved is not normal. 6. 7. Prove (5.17). Show that or a random sample X1, X2,...,Xn from a multivariate distribution with
expectation vector and variance matrix G the sample variance matrix (5.18) is and unbiased estimator of G. 8. Given a random sample of size n from the N( , 2) distribution, prove that the Cramer-
Rao lower bound for an unbiased estimator of 2 is 24 / n . 9. 10. 11. Prove Theorem 5.15. Prove the second equalities in (5.34) and (5.35). Show that the Cramer-Rao lower bound of an unbiased estimator of (5.37) is equal to
2(E[X TX])&1 . 12. Show that the matrix (5.38) is idempotent.
197 13. 14. 15. Why is (5.40) true? Why does (5.43) imply (5.44)? Suppose your econometric software package reports that the OLS estimate of a regression
parameter is 1.5, with corresponding t-value 2.4. However, you are only interested in whether the true parameter value is 1 or not. How would you test these hypotheses? Compute the test statistic involved. Moreover, given that the sample size is n = 30 and that your model has 5 other parameters, conduct the test using size 0.05. You have to look up the critical value involved in one of the statistics or econometrics textbook that contain tables of the t-distribution.6
Appendix
5.A. Proof of Theorem 5.8 Note again that the condition AB = O only makes sense if both A and B are singular, if otherwise either A, B or both are O. Write A ' QAAQA , B ' QBBQB , where QA and QB are orthogonal matrices of eigenvectors and 7A and 7B are diagonal matrices of corresponding eigenvalues. Then Z1 ' X TQAAQA X , Z2 ' X TQBBQB X . Since A and B are both singular, it follows that 7A and 7B are singular. Thus let 1 A ' O O
T T T T
O &2 O , O O O
where 71 is the k k diagonal matrix of positive eigenvalues, and !72 the m m diagonal matrix of negative eigenvalues of A, with k + m < n. Then
198
1 1
0
T
2 1
0
1 2
Ik
0 0 In&k&m
2 1
0
1 2
0 QA X .
T
Z1 ' X TQA 0 &2 0 QA X ' X TQA 0 0 0
0 0
2 0 0 0
0 &Im 0 0
0 0
2 0 0 0
Similarly, denote 1 B '

(
( O &2 O ,
* where 7* 1 is the p p diagonal matrix of positive eigenvalues, and !72 is the q q diagonal
matrix of negative eigenvalues of B, with p + q < n. Then

(
1
(1 ) 2 Z2 ' X TQB 0 0
0
(
1
Ip
0 0 In&p&q
(1 ) 2 0 0
0
(
1
0 QB X .
T
(2 ) 2 0 0 0
0 &Iq 0 0
(2 ) 2 0 0 0
Next, let
1 2 1
1 Y1 '
O O
1 2
( (1 ) 2
O
(
1
O QB X ' M2X , say .

T
O 2 O O O O
QA X ' M1X , say ,
Y2 '
O O
(2 ) 2 O O O
Then
199 Ik O O O In&k&m Y1 ' Y1 D1Y1 , say ,

T
T Z1 ' Y1 O &Im
O and
Ip O
T Z2 ' Y2 O &Iq
O O Y2 ' Y2 D2Y2 , say ,

T
O O In&p&q where the diagonal matrices D1 and D2 are nonsingular but possibly different. Clearly, Z1 and Z2 are independent if Y1 and Y2 are. Now observe that
1 2 1 1 2 1
0 2 0
1 2
0 0 In&k&m
Ik
0 0 In&k&m
1 0 0
0
1 2
0 QA Q B
T
( (1 ) 2
0
( (2 ) 2
1
0 0 0
A B ' QA
0 0
0 &Im 0 0
2 0 0 0
0 0
Ip
0 0 In&p&q
(1 ) 0 0 0
0
( (2 ) 2
1
0 0 In&p&q QB .
T
0 &I q 0 0
The first three matrices are nonsingular, and so are the last three. Therefore, AB = O if and only if 1 M1M2 '
T
1 2
0
1 2
0 Q A QB
T
(1 ) 2 0 0
0
( (2 ) 2
1
0 0 0 ' O.
0 0
2 0 0 0
200
It follows now from Theorem 5.7 that the latter implies that Y1 and Y2 are independent, hence the condition AB = O implies that Y1 and Y2 are independent. Q.E.D.
Endnotes 1. In order to distinguish the variance of a random variable from the variance matrix of a random vector, the latter will be denoted by Var, with capital V. 2. The capital C in Cov indicates that this is a covariance matrix rather than a covariance of two random variables. 3. These calculators are also included in my free econometrics software package EasyReg International, which you can download from http://econ.la.psu.edu/~hbierens/EASYREG.HTM. 4. Recall that "argmin" stands for the argument for which the function involved takes a minimum. 5. The OLS estimator is called "ordinary" to distinguish it from the nonlinear least squares estimator. See Chapter 6 for the latter. 6. Or use the authors free econometrics software package EasyReg International. The tdistribution calculator can be found under "Tools".
201
Chapter 6 Modes of Convergence
6.1.
Introduction Toss a fair coin n times, and let Yj = 1 if the outcome of the j-th tossing is head, and Yj =
!1 if the outcome involved is tail. Denote Xn ' (1/n)'j'1Y j . For the case n = 10 the left-hand
n
side panel of Figure 6.1 displays the distribution function Fn(x)1 of Xn on the interval [!1.5, 1.5], and the right-hand side panel displays a typical plot of Xk for k =1,2,...,10, based on simulated Yjs.2
Figure 6.1. n = 10. Left: Distribution function of Xn . Right: Plot of Xk for k=1,2,...,n.
Now let us see what happens if we increase n: First, consider the case n = 100, in Figure 6.2. The distribution function Fn(x) becomes steeper for x close to zero, and Xn seems to tend towards zero.
Figure 6.2. n = 100. Left: Distribution function of Xn . Right: Plot of Xk for k=1,2,...,n.
202 These phenomena are even more apparent for the case n = 1000, in Figure 6.3.
Figure 6.3. n = 1000. Left: Distribution function of Xn . Right: Plot of Xk for k=1,2,...,n. What you see in Figures 6.1-6.3 is the law of large numbers: Xn ' (1/n)'j'1Y j 6 E[Y1] ' 0 in
n
some sense, to be discussed below, and the related phenomenon that Fn(x) converges pointwise in x 0 to the distribution function F(x) = I(x $ 0) of a "random" variable X satisfying P[X ' 0] = 1. Next, let us have a closer look at the distribution function of nXn : Gn(x) = Fn(x/ n) , with corresponding probabilities P[ nXn ' (2k&n)/ n] , k = 0,1,...,n, and see what happens if n 64. These probabilities can be displayed in the form of a histogram: P 2(k&1)/ n& n < nXn # 2k/ n& n 2/ n if x 0 2(k&1)/ n& n , 2k/ n& n , k '0,1,....,n , Hn(x) ' 0 elsewhere. Figures 6.4-6.6 compare Gn(x) with the distribution function of the standard normal distribution, and Hn(x) with the standard normal density, for n = 10, 100 and 1000.
Hn(x) '
203
Figure 6.4. n = 10: Left: Gn(x), right: Hn(x), compared with the standard normal distribution.
What you see in the left-hand side panels in Figures 6.4-6.6 is the central limit theorem: m
x
lim Gn(x) '

n64
exp[&u 2/2] 2
du ,
&4
pointwise in x, and what you see in the right-hand side panels is the corresponding fact that Gn(x%) & Gn(x) exp[&x 2/2] 2
lim lim
90 n64
'
204
The law of large numbers and the central limit theorem play a key role in statistics and econometrics. In this chapter I will review and explain these laws.
6.2.
Convergence in probability and the weak law of large numbers Let Xn be a sequence of random variables (or vectors) and let X a random or constant
variable (or conformable vector).
Definition 6.1: We say that Xn converges in probability to X, also denoted as plimn64Xn = X or Xn 6p X , if for an arbitrary g > 0 we have limn64P(*Xn & X* > g) = 0, or equivalently, limn64P(*Xn & X* # g) = 1.
In this definition, X may be a random variable or a constant. The latter case, where P(X = c) = 1 for some constant c, is the most common case in econometric applications. Also, this definition carries over to random vectors, provided that the absolute value function *x* is replaced by the Euclidean norm 2x2 ' x Tx . The right-hand side panels of Figures 6.1-6.3 demonstrate the law of large numbers. One of the versions of this law is the Weak Law of Large Numbers (WLLN), which also applies to uncorrelated random variables.
Theorem 6.1: (WLLN for uncorrelated random variables). Let X1 ,...,Xn be a sequence of uncorrelated random variables with E(Xj) = and var(Xj) = 2 < 4 , and let X = (1/n)'j'1Xj .
n
205 Then plimn64X = .
Proof: Since E(X ) ' and Var(X ) ' 2 / n , it follows from Chebishev inequality that P(*X & * > g) # 2/(ng2) 6 0 if n 6 4 . Q.E.D. The condition of a finite variance can be traded in for the i.i.d. condition:
Theorem 6.2: (The WLLN for i.i.d. random variables). Let X1,...,Xn be a sequence of independent identically distributed random variables with E [|Xj|] < 4 and E(Xj) = , and let X ' (1/n)'j'1Xj . Then plimn64X = .
n
Proof: Let Yj = Xj .I(|Xj| # j) and Zj = Xj .I(|Xj| > j), so that Xj = Yj + Zj. Then E*(1 / n)'j'1(Zj & E(Zj)) * # 2(1 / n)'j'1E[*Zj*] ' 2(1 / n)'j'1E[|X1|I(|X1| > j)] 6 0 ,
n n n
(6.1)
and E[*(1 / n)'j'1(Y j & E(Y j)) *2] # (1 / n 2)'j'1E[Y j ] ' (1 / n 2)'j'1E[X1 I(|X1| # j)]
n n 2 n 2
' (1 / n 2)'j'1'k'1E[X1 I(k & 1 < |X1| # k)]

n j 2
# (1 / n 2)'j'1'k'1k.E[|X1|.I(k & 1 < |X1| # k)]

n j
(6.2)
(1 / n 2)'j'1'k'1'i'kE[|X1|.I(i & 1 < |X1| # i)] # (1 / n 2)'j'1'k'1E[|X1|.I(|X1| > k & 1)

n j&1 j n j&1
# (1 / n)'k'1E[|X1|.I(|X1| > k & 1)] 6 0

n
as n 64, where the last equality in (6.2) follows from the easy equality 'k'1 k.k ' 'k'1'i'ki ,
j j&1 j
206 and the convergence results in (6.1) and (6.2) follow from the fact that E[|X1|I(|X1| > j)] 6 0 for j 6 4, because E[*X1*] < 4 . Using Chebishevs inequality it follows now from (6.1) and (6.2) that for arbitrary g > 0, P[*(1/n)'j'1(Xj & E(Xj))* > g] # P[*(1/n)'j'1(Y j & E(Y j))*
n n
% *(1/n)'j'1(Zj & E(Z j))*> g]

n
# P[*(1/n)'j'1(Y j & E(Y j))* > g/2] % P[*(1/n)'j'1(Z j & E(Zj))*> g/2]
n n
(6.3)
# 4E[*(1/n)'j'1(Y j & E(Y j))*2]/g2 % 2E[*(1/n)'j'1(Z j & E(Zj))*]/g 6 0

n n
as n 64. Note that the second inequality in (6.3) follow from the fact that for non-negative random variables X and Y, P[X%Y > g] # P[X > g/2] % P[Y > g/2] . The theorem under review follows now from (6.3), Definition 6.1 and the fact that g is arbitrary. Q.E.D. Note that Theorems 6.1-6.2 carry over to finite-dimensional random vectors Xj, by replacing the absolute values *.* by Euclidean norms: 2x2 ' x Tx , and the variance by the variance matrix. The reformulation of Theorems 6.1-6.2 for random vectors is left as an easy exercise. Convergence in probability carries over after taking continuous transformations. This results is often referred to as Slutky's theorem:
Theorem 6.3: (Slutsky's theorem). Let Xn a sequence of random vectors in k satisfying Xn 6p c, where c is non-random . Let Q(x) be an m -valued function on k which is continuous in c. Then Q(Xn) 6p Q(c).
207 Proof: Consider the case m = k = 1. It follows from the continuity of Q that for an arbitrary g > 0 there exists a * > 0 such that *x & c* # implies *(x) & (c)* # g , hence P(*Xn & c* # ) # P(*(Xn) & (c)* # g) . Since limn64P(*Xn & c* # ) = 1, the theorem follows for the case under review. The more general case with m > 1 and/or k > 1, can be proved along the same lines. Q.E.D. The condition that c is constant is not essential. Theorem 6.3 carries over to the case where c is a random variable or vector, as we will see in Theorem 6.7 below. Convergence in probability does not automatically imply convergence of expectations. A counter-example is Xn = X +1/n, where X has a Cauchy distribution (see Chapter 4). Then E[Xn] and E(X) are not defined, but Xn 6p X. However,
Theorem 6.4: (Bounded convergence theorem) If Xn is bounded: P(*Xn* # M) = 1 for some M < 4 and all n, then Xn 6p X implies limn64E(Xn) = E(X).
Proof: First, X has to be bounded too, with the same bound M, because otherwise Xn 6p X is not possible. Without loss of generality we may now assume that P(X = 0) = 1 and that Xn is a non-negative random variable, by replacing Xn with |Xn ! X|, because E[|Xn ! X|] 6 0 implies limn64E(Xn) = E(X). Next, let Fn(x) be the distribution function of Xn, and let g > 0 be arbitrary. Then 0 # E(Xn) ' m0
M
xdFn(x) '
m0
xdFn(x) %
mg
xdFn(x) # g % M.P(Xn $ g) .
(6.4)
208 Since the latter probability converges to zero (by the definition of convergence in probability and the assumption that Xn is nonnegative, with zero probability limit), we have 0 # limsupn64E(Xn) # g for all g > 0, hence limn64E(Xn) = 0. Q.E.D. The condition that Xn in Theorem 6.4 is bounded can be relaxed, using the concept of uniform integrability:
Definition 6.2: A sequence Xn of random variables is said to be uniformly integrable if limM64supn$1E[|Xn|. I(|Xn| > M)] ' 0 .
Note that this Definition 6.carries over to random vectors by replacing the absolute value |.| with the Euclidean norm ||.||. Moreover, it is easy to verify that if *Xn* # Y with probability 1 for all n $1, where E(Y) < 4, then Xn is uniformly integrable.
Theorem 6.5: (Dominated convergence theorem) Let Xn be uniformly integrable. Then Xn 6p X implies limn64E(Xn) = E(X).
Proof: Again, without loss of generality we may assume that P(X = 0) = 1 and that Xn is a non-negative random variable. Let 0 < g < M be arbitrary. Then similarly to (6.4), 0 # E(Xn) ' m0
4
xdFn(x) '
m0
xdFn(x) %
mg
xdFn(x) %
4
mM
xdFn(x) (6.5)
# g % M.P(Xn $ g) % supn$1
mM
xdFn(x) .
For fixed M the second term at the right-hand side of (6.5) converges to zero. Moreover, by
209 uniform integrability we can choose M so large that the third term is smaller than g. Hence, 0 # limsupn64E(Xn) # 2g for all g > 0, and thus limn64E(Xn) = 0. Q.E.D. Also Theorems 6.4 and 6.5 carry over to random vectors, by replacing the absolute value function *x* by the Euclidean norm 2x2 ' x Tx .
6.3.
Almost sure convergence, and the strong law of large numbers In most (but not all!) cases where convergence in probability and the weak law of large
numbers apply, we actually have a much stronger result:
Definition 6.3: We say that Xn converges almost surely (or: with probability 1) to X, also denoted by Xn 6 X a.s. (or: w.p.1), if for all g > 0 , limn64P(supm$n*Xm & X* # g) ' 1, or equivalently, P(limn64Xn ' X) ' 1 . (6.7) (6.6)
The equivalence of the conditions (6.6) and (6.7) will be proved in Appendix 6.B (Theorem 6.B.1). It follows straightforwardly from (6.6) that almost sure convergence implies convergence in probability. The converse, however, is not true. It is possible that a sequence Xn converges in probability but not almost surely. For example, let Xn = Un /n, where the Uns are i.i.d. nonnegative random variables with distribution function G(u) ' exp(&1/u) for u > 0, G(u) ' 0 for u # 0. Then for arbitrary g > 0,
210 P(|Xn| # g) ' P(Un # ng) ' G(ng) ' exp(&1/(ng)) 6 1 as n 6 4 , hence Xn 6p 0. On the other hand, P(|Xm| # g for all m $ n) ' P(Um # mg for all m $ n) ' m'nG(mg) ' exp &g&1'm'nm &1 ' 0 ,
4 4
where the second equality follows from the independence of the Uns, and the last equality follows from the fact that 'm'1m &1 ' 4 . Consequently, Xn does not converge to 0 almost
4
surely. Theorems 6.2-6.5 carry over to the almost sure convergence case, without additional conditions:
Theorem 6.6: (Kolmogorov's strong law of large numbers). Under the conditions of Theorem 6 a.s. 6.2, X
Proof: See Appendix 6.B. The result of Theorem 6.6 is actually what you see happening in the right-hand side panels of Figures 6.1-6.3.
Theorem 6.7: (Slutsky's theorem). Let Xn a sequence of random vectors in k converging a.s. to a (random or constant) vector X. Let Q(x) be an m -valued function on k which is continuous on an open subset 3 B of k for which P(X 0 B) = 1). Then (Xn) 6 (X) a.s.
211 Proof: See Appendix 6.B. Since a.s. convergence implies convergence in probability, it is trivial that:
Theorem 6.8: If Xn 6 X a.s., then the result of Theorem 6.4 carries over.
Theorem 6.9: If Xn 6 X a.s., then the result of Theorem 6.5 carries over.
6.4. 6.4.1
The uniform law of large numbers and its applications The uniform weak law of large numbers In econometrics we often have to deal with means of random functions. A random
function is a function that is a random variable for each fixed value of its argument. More precisely:
Definition 6.4: Let {S,,P} be the probability space. A random function f(2) on a subset 1 of a Euclidean space is a mapping f(,): 6 such that for each Borel set B in and each 2 0 1, { 0 : f( , ) 0 B} 0 .
Usually random functions take the form of a function g(X,2) of a random vector X and a nonrandom vector 2. For such functions we can extend the weak law of large numbers for i.i.d. random variables to a Uniform Weak Law of Large Numbers (UWLLN):
212 Theorem 6.10: (UWLLN). Let Xj, j = 1,..,n, be a random sample from a k-variate distribution, and let 0 be non-random vectors in a closed and bounded (hence compact4) subset d m . Moreover, let g(x,2) be a Borel measurable function on k such that for each x, g(x,2) is a continuous function on 1. Finally, assume that E[sup0*g(Xj , )*] < 4. Then plimn64sup0*(1/n)'j'1g(Xj , ) & E[g(X1 , )]* ' 0 .
n
Proof: See Appendix 6.A.
6.4.2
Applications of the uniform weak law of large numbers
6.4.2.1 Consistency of M-estimators In Chapter 5 I have introduced the concept of a parameter estimator, and listed a few desirable properties of estimators, i.e., unbiasedness and efficiency. Another obviously desirable property is that the estimator gets closer to the parameter to be estimated if we use more data information. This is the consistency property:
of a parameter , based on a sample of size n, is called consistent Definition 6.5: An estimator ' . if plimn64
Theorem 6.6 is an important tool in proving consistency of parameter estimators. A large class of estimators are obtained by maximizing or minimizing an objective function of the form (1/n)'j'1g(Xj , ) , where g, Xj and are the same as in Theorem 6.10. These estimators are
n
called M-estimators (where the M indicates that the estimator is obtained by Maximizing or
213 Minimizing a Mean of random functions). Suppose that the parameter of interest is 0 = argmax0E[g(X1,)], where is a given closed and bounded set. Note that "argmax" is a shorthand notation for the argument for which the function involved is maximal. Then it seems a ' argmax (1/n)'n natural choice to use j'1g(Xj , ) as an estimator of 0 . Indeed, under some 0 mild conditions the estimator involved is consistent:
= argmax Q Theorem 6.11: (Consistency of M- estimators) Let 0 () , 0 = argmax0Q() ,
() = (1/n)'n where Q j'1g(Xj , ) , and Q() ' E [Q()] ' E[g(X1 , )] , with g, Xj and the same as in Theorem 6.10. If 20 is unique, in the sense that for arbitrary g > 0 there exists a * > 0 ' . such that Q(0) & sup2& 2>g Q() > , then plimn64 0
0
0 and 0 , because g(x,2) is continuous in 2. See Proof: First, note that 0 Appendix II. By the definition of 20, ( ) & Q ( ) ' Q ( ) & Q ( ) % Q ( ) & Q ( ) 0 # Q 0 0 0 0 ( ) & Q ( ) % Q ( ) & Q ( ) # 2sup*Q () & Q ()* , # Q 0 0
0
(6.8)
and it follows from Theorem 6.3 that the right-hand side of (6.8) converges in probability to zero. Thus: ) ' Q( ) . plimn64 Q( 0 (6.9)
Moreover, the uniqueness condition implies that for arbitrary g > 0 there exists a * > 0 such that ) $ if 2 & 2 > g , hence Q(0) & Q( 0 & 2 > g # PQ ( ) & Q ( ) $ . P 2 0 0 (6.10)
214 Combining (6.9) and (6.10), the theorem under review follows from Definition 6.1. Q.E.D. It is easy to verify that Theorem 6.11 carries over to the "argmin" case, simply by replacing g by -g. As an example, let X1 , ... , Xn be a random sample from the non-central Cauchy distribution, with density h(x|0) = 1/[(1%(x&0)2] , and suppose that we know that 0 is contained in a given closed and bounded interval . Let g(x , ) ' f(x&) , where f(x) = exp(&x 2/2)/ 2 is the density of the standard normal distribution. Then
4
E[g(X1 , )] '
&4
exp(&(x%0&)2)/ 2 (1%x 2)
dx ' f(x&%0)h(x|0)dx ' (&0) , m

&4
(6.11)
say, where (y) is a density itself, namely the density of Y ' U % Z , with U and Z independent random drawings form the standard normal and standard Cauchy distribution, respectively. This is called the convolution of the two densities involved. The characteristic function of Y is exp(&|t|&t 2/2) , so that by the inversion formula for characteristic functions 1 (y) ' cos(t.y)exp(&|t|&t 2/2)dt . m 2
&4 4
(6.12)
This function is maximal in y = 0, and this maximum is unique, because for fixed y 0 the set {t 0 : cos(t.y) ' 1} is countable and therefore has Lebesgue measure zero. In particular, it follows from (6.12) that for arbitrary g > 0, 1 sup|y|$g|cos(t.y)|exp(&|t|&t 2/2)dt < (0) . sup|y|$g(y) # 2 m
&4 4
(6.13)
Combining (6.11) and (6.13) yields sup|& |$gE[g(X1 , )] < E[g(X1 , 0)] . Thus, all the
0
215 ' . conditions of Theorem 6.11 are satisfied, hence plimn64 0 Another example is the nonlinear least squares estimator. Consider a random sample Zj ' (Y j , Xj )T , j ' 1,2,....,n , with Y j 0 , Xj 0 k , and assume that:
T
Assumption 6. 1: For a given function f(x,) on k , with 1 a given compact subset of m , there exists a 0 0 such that P[E[Y j|Xj] ' f(Xj , 0)] ' 1 . Moreover, for each x 0 k , f(x,) is a continuous function on 1, and for each 0 , f(x,) is a Borel measurable function on k . Furthermore, let E[Y1 ] < 4 , E[sup0f(X1 ,)2] < 4 , and inf||& ||$E [(f(X1 , ) & f(X1 , 0))2] > 0 for > 0.
0
Denoting Uj ' Y j & E[Y j|Xj] we can write Y j ' f(Xj , 0) % Uj , where P(E[Uj|Xj] ' 0) ' 1. (6.14)
This is the general form of a nonlinear regression model. I will show now that under Assumption 6.1 the nonlinear least squares estimator
2 ' argmin (1/n)'n j'1(Y j & f(X j , )) 0
(6.15)
is a consistent estimator of 0 . Let g(Zj , ) ' (Y j & f(Xj , ))2 . Then it follows from Assumption 6.1 and Theorem 6.10 that plimn64sup0|(1/n)'j'1[g(Zj,)&E[g(Z1,)]| ' 0 .
n
Moreover,
216 E [g(Z1,)] ' E[(Uj % f(Xj , 0) & f(Xj , ))2] ' E[Uj |] % 2E[E(Uj|Xj)(f(Xj , 0) & f(Xj , ))] % E[(f(Xj , 0) & f(Xj , ))2] ' E[Uj |] % E[(f(Xj , 0) & f(Xj , ))2] , hence it follows from Assumption 6.1 that inf||& ||$E [|g(Z1 , )|] > 0 for > 0. Therefore the
0
condition of Theorem 6.11 for the argmin case are satisfied, and consequently, the nonlinear least squares estimator (6.15) is consistent.
6.4.2.2 Generalized Slutskys theorem Another easy but useful corollary of Theorem 6.6 is the following generalization of Theorem 6.3:
Theorem 6.12: (Generalized Slutskys theorem) Let Xn a sequence of random vectors in k converging in probability to a nonrandom vector c. Let Mn(x) be a sequence of random functions on k satisfying plimn64supx0B*n(x) & (x)* = 0, where B is a closed and bounded subset of k containing c, and M is a continuous nonrandom function on B. Then n(Xn) 6p (c) .
Proof: Exercise. This theorem can be further generalized to the case where c = X is a random vector, simply by adding the condition that P[X 0 B] ' 1 , but the current result suffices for the applications of Theorem 6.12. This theorem plays a key-role in deriving the asymptotic distribution of an M-estimator,
217 together with the central limit theorem discussed below.
6.4.3
The uniform strong law of large numbers and its applications The results of Theorems 6.10-6.12 also hold almost surely. See Appendix 6.B for the
proofs.
Theorem 6.13: Under the conditions of Theorem 6.10, sup0*(1/n)'j'1g(Xj , ) & E[g(X1 , )]*
n
6 0 a.s.
6 a.s. Theorem 6.14: Under the conditions of Theorems 6.11 and 6.13, 0
Theorem 6.15: Under the conditions of Theorem 6.12 and the additional condition that Xn 6 c a.s., n(Xn) 6 (c) a.s.
6.5.
Convergence in distribution Let Xn be a sequence of random variables (or vectors) with distribution functions Fn(x),
and let X be a random variable (or conformable random vector) with distribution function F(x).
Definition 6.6: We say that Xn converges to X in distribution (denoted by Xn 6d X ) if limn64Fn(x) = F(x), pointwise in x , possibly except in the discontinuity points of F(x).
218 Alternative notation: If X has a particular distribution, for example N(0,1), then Xn 6d X is also denoted by Xn 6d N(0,1). The reason for excluding discontinuity points of F(x) in the definition of convergence in distribution is that in these discontinuity points, limn64Fn(x) may not be right-continuous. For example, let Xn = X + 1/n. Then Fn(x) = F(x-1/n). Now if F(x) is discontinuous in x0, then limn64F(x0-1/n) < F(x0), hence limn64Fn(x0) < F(x0). Thus, without the exclusion of discontinuity points, X + 1/n would not converge in distribution to the distribution of X, which would be counter-intuitive. If each of the components of a sequence of random vectors converge in distribution, then the random vectors themselves may not converge in distribution. As a counter-example, let X1n X2n 0 0 1 (&1)n/2 (&1)n/2 1
Xn '
- N2
(6.16)
Then X1n 6d N(0,1) and X2n 6d N(0,1), but Xn does not converge in distribution. Moreover, in general Xn 6d X does not imply that Xn 6p X. For example, if we replace X by an independent random drawing Z from the distribution of X, then Xn 6d X and Xn 6d Z are equivalent statements, because these statements only say that the distribution function of Xn converges to the distribution function of X (or Z), pointwise in the continuity points of the latter distribution function. If Xn 6d X would imply Xn 6p X, then Xn 6d Z in distr. would imply that X = Z, which is not possible, because X and Z are independent. The only exception is the case where the distribution of X is degenerated: P(X = c) = 1 for some constant c:
Theorem 6.16: If Xn converges in distribution to X, and P(X = c) = 1, where c is a constant, then
219 Xn converges in probability to c.
Proof: Exercise. Note that this result is demonstrated in the left-hand side panels of Figures 6.1-6.3. On the other hand,
Theorem 6.17: Xn 6p X implies Xn 6d X.
Proof: Theorem 6.17 follows straightforwardly from Theorem 6.3, Theorem 6.4, and Theorem 6.18 below. Q.E.D. There is a one-to-one correspondence between convergence in distribution and convergence of expectations of bounded continuous functions of random variables:
Theorem 6.18: Let Xn and X be random vectors in k . Then Xn 6d X if and only if for all bounded continuous functions on k , limn64E[(Xn)] ' E[(X)] .
Proof: I will only prove this theorem for the case where Xn and X are random variables. Throughout the proof the distribution function of Xn is denoted by Fn(x), and the distribution function of X by F(x). Proof of the "only if" case: Let Xn 6d X. Without loss of generality we may assume that (x) 0 [0,1] for all x. For any g > 0 we can choose continuity points a and b of F(x) such that F(b) - F(a) > 1!g. Moreover, we can choose continuity points a = c1 < c2 <...< cm = b of F(x) such
220 that for j = 1,..,m-1, sup (x) &

x0(c j,cj%1]
inf (x) # g .
x0(c j,cj%1]
(6.17)
Now define (x) ' inf (x) for x 0 (cj,cj%1] , j ' 1,.., m&1 , (x) ' 0 elsewhere. (6.18)
x0(c j,cj%1]
Then 0 # (x) & (x) # g for x 0 (a,b] , 0 # (x) & (x) # 1 for x (a,b] , hence limsup*E[(Xn)] & E[(Xn)]*
n64
# limsup
n64
x0(a,b]
*(x)&(x)*dFn(x) %
x(a,b]
*(x)&(x)*dFn(x)
(6.19)
# g % 1 & lim Fn(b) & Fn(a) ' g % 1 & F(b) & F(a) # 2g .
n64
Moreover, we have *E[(X)] & E[(X)]* # 2g , and limn64E[(Xn)] ' E[(X)] . Combining (6.19), (6.20) and (6.21), the "only if" part easily follows. Proof of the "if" case: Let a < b be arbitrary continuity points of F(x), and let ' 0 (x) ' ' 1 if if x $ b, x < a, (6.22) (6.21) (6.20)
b&x ' if a # x < b . b&a
Then clearly (6.22) is a bounded continuous function. Next, observe that b&x (x)dFn(x) ' Fn(a) % dF (x) $ Fn(a) m m b&a n
a b
E[(Xn)] '
(6.23)
221 hence E[(X)] ' limE[(Xn)] $ limsupFn(a) .

n64 n64
(6.24)
Moreover, b&x (x)dF(x) ' F(a) % dF(x) # F(b) . m m b&a

a b
E[(X)] '
(6.25)
Combining (6.24) and (6.25) yields F(b) $ limsupn64Fn(a) , hence, since b (> a) was arbitrary, letting b 9 a it follows that. F(a) $ limsupFn(a) .
n64
(6.26)
Similarly, for c < a we have F(c) # liminfn64Fn(a) , hence letting c 8 a it follows that F(a) # liminfFn(a) .
n64
(6.27)
Combining (6.26) and (6.27), the "if" part now follows, i.e., F(a) ' limn64Fn(a) . Q.E.D. Note that the "only if" part of Theorem 6.18 implies another version of the bounded convergence theorem:
Theorem 6.19: (Bounded convergence theorem) If Xn is bounded: P(*Xn* # M) = 1 for some M < 4 and all n, then Xn 6d X implies limn64E(Xn) = E(X).
Proof: Easy exercise. Using Theorem 6.18, it is not hard to verify that the following result holds.
Theorem 6.20: (Continuous Mapping Theorem) Let Xn and X be random vectors in k such that
222 Xn 6d X, and let (x) be a continuous mapping from k into m . Then (Xn) 6d (X).
Proof: Exercise. Examples of applications of Theorem 6.20 are: (1) (2) Let Xn 6d X, where X is N(0,1) distributed. Then Xn 6d 1 . Let Xn 6d X, where X is Nk(0,I) distributed. Then Xn Xn 6d k .
T 2 2 2
If Xn 6d X, Yn 6d Y, and (x,y) is a continuous function, then in general it does not follow that (Xn ,Yn) 6d (X,Y), except if either X or Y has a degenerated distribution:
Theorem 6.21: Let X and Xn be random vectors in k such that Xn 6d X, and let Yn be a random vector in m such that plimn64Yn = c, where c 0 m is a nonrandom vector. Moreover, let (x,y) be a continuous function on the set k {y 0 m: 2y & c2 < } for some > 0. 5 Then (Xn,Yn) 6d (X,c).
Proof: Again, we prove the theorem for the case k = m = 1 only. Let Fn(x) and F(x) be the distribution functions of Xn and X, respectively, and let (x,y) be a bounded continuous function on (c&,c%) for some > 0. Without loss of generality we may assume that *(x,y)* # 1. Next, let g > 0 be arbitrary, and choose continuity points a 1!g. Then for any > 0,
223 *E[(Xn,Y n)] & E[(Xn,c)* # E[*(Xn,Y n) & (Xn,c)*I(*Y n&c*#)] % E[*(Xn,Y n) & (Xn,c)*I(*Y n&c*>)] # E[*(Xn,Y n) & (Xn,c)*I(*Y n&c*#)I(Xn0[a,b])] % 2P(Xn[a,b]) % 2P(*Y n&c*>) # sup
x0[a,b], *y&c*#
(6.28)
*(x,y)&(x,c)* % 2(1&Fn(b)%Fn(a)) % 2P(*Y n&c*>) .
Since a continuous function on a closed and bounded subset of an Euclidean space is uniformly continuous on that subset (see Appendix II), we can choose so small that sup
x0[a,b], *y&c*#
*(x,y)&(x,c)* < g .
(6.29)
Moreover, 1 - Fn(b) + Fn(a) 6 1 - F(b) + F(a) < g, and P(*Y n&c*>) 6 0 . Therefore, it follows from (6.28) that: limsup*E[(Xn,Y n)] & E[(Xn,c)* # 3g .
n64
(6.30)
The rest of the proof is left as an exercise. Q.E.D.
Corollary 6.1: Let Zn be t-distributed with n degrees of freedom. Then Zn 6d N(0,1).
Proof: By the definition of the t-distribution with n degrees of freedom we can write
224 Zn ' Uo 1 2 j Uj n j'1

n
, (6.31)
where U0 , U1 ,..,Un are i.i.d. N(0,1). Let Xn = U0 and X = U0, so that trivially Xn 6d X. Let E(U2 1) = 1. Let (x,y) = x/%y. Note that (x,y) is continuous on R (1-g,1+g) for 0 < g < 1. Thus by Theorem 6.21, Zn ' (Xn,Y n) 6 (X,1) ' U0 - N(0,1) in distribution. Q.E.D. Y n ' (1/n)'j'1Uj . Then by the weak law of large numbers (Theorem 6.2) we have: plimn64Yn =
n 2
Corollary 6.2: Let U1...Un be a random sample from Nk(F,), where is non-singular. Denote
n T &1 ' (1/n)'n T U j'1Uj , ' (1/(n&1))'j'1(Uj&U )(U j&U ) , and let Zn = n(U&) (U&) . Then
Zn 6d k .
Proof: For a kk matrix A = (a1,..,ak), let vec(A) be the k21 vector of stacked columns aj, j = 1,...,k, of A: vec(A) ' (a1 , ... , a k )T ' b , say, with inverse vec-1(b) = A. Let c = vec(), Yn = ), Xn = n(U &) , X - N (0,) , and (x,y) ' x T(vec&1(y))&1x . Since is nonsingular, vec( k there exists a neighborhood C() = {y0kk: 2y&c2 < } of c such that for all y in C(), vec-1(y) is nonsingular (Exercise: Why?), and consequently, (x,y) is continuous on kC() (Exercise: Why?). The corollary follows now from Theorem 6.21 (Exercise: Why?). Q.E.D.
T T
225 6.6. Convergence of characteristic functions Recall that the characteristic function of a random vector X in k is defined as (t) ' E[exp(it TX)] ' E[cos(t TX)] % i.E[sin(t TX)] for t 0 k , where i ' &1 . The last equality is due to the fact that exp(i.x) = cos(x) + i.sin(x). Also recall that distributions are the same if and only if their characteristic functions are the same. This property can be extended to sequences of random variables and vectors:
Theorem 6.22: Let Xn and X be random vectors in k with characteristic functions n(t) and (t) , respectively. Then Xn 6d X if and only if (t) ' limn64n(t) for all t 0 k .
Proof: See Appendix 6.C for the case k = 1. Note that the "only if" part of Theorem 6.22 follows from Theorem 6.18: Xn 6d X implies that for any t 0 k , limn64E[cos(t TXn)] ' E[cos(t TX)] , limn64E[sin(t TXn)] ' E[sin(t TX)] , hence limn64n(t) ' limn64E[cos(t TXn)] % i.limn64E[sin(t TXn)] ' E[cos(t TX)] % i.E[sin(t TX)] ' (t) . Theorem 6.22 plays a key-role in the derivation of the central limit theorem, in the next section.
226 6.7. The central limit theorem The prime example of the concept of convergence in distribution is the central limit theorem, which we have seen in action in Figures 6.4-6.6:
2 Theorem 6.23: Let X1,....,Xn be i.i.d. random variables satisfying E(Xj) = , var(Xj) = < 4 , 2 ' (1/n)'n and let X j'1Xj . Then n(X & ) 6d N(0, ) .
Proof: Without loss of generality we may assume that F = 0 and F = 1. Let (t) be the characteristic function of Xj. The assumptions F = 0 and F = 1 imply that the first and second derivatives of (t) at t = 0 are equal to )(0) ' 0, ))(0) ' &1 , respectively, hence by Taylor's theorem, applied to Re[N(t)] and Im[N(t)] separately, there exists numbers 1,t, 2,t 0 [0,1] such that (t) ' (0) % t)(0) % 1 2 1 t Re[))(1,t.t)] % i.Im[))(2,t.t)] ' 1 & t 2 % z(t)t 2 , 2 2
say, where z(t) ' (1 % Re[))(1,t.t)] % i.Im[))(2,t.t)]) / 2 . Note that z(t) is bounded and satisfies limt60z(t) ' 0. . Then Next, let n(t) be the characteristic function of nX
n(t) ' (t/ n) t 2n

2 n
t2 z(t/ n)t 2 ' 1 & % 2n n n m 1& t 2n

2 n&m
(6.32) z(t/ n)t n

2 m
' 1 &
% j
n
m'1
227 For n so large that t2/(2n) < 1 we have

n n /0 j 00 0 m'1 m n&m m n n /0 # j 00 0 m'1 m 2 n
t2 1& 2n
z(t/ n)t 2 n
*z(t/ n)*t 2 n
(6.33)
' 1 %
*z(t/ n)*t n
& 1
Now observe that for any real valued sequence an which converges to a, lim ln (1%a n/n )n ' lim n ln(1%an/n) ' lim a n lim
n64 n64 n64 n64
ln(1%an/n) & ln(1) a n /n
' a lim
60
ln(1%)&ln(1) ' a,
hence limn64a n ' a Y lim 1 % an/n

n64 n
' e a.
(6.34)
Letting an = *z(t/ n)*t 2 , which has limit a = 0, it follows from (6.34) that the right-hand side expression in (6.33) converges to zero, and letting an = a = -t2/2 it follows then from (6.32) that limn(t) ' e &t /2.
n64
2
(6.35)
The right-hand side of (6.35) is the characteristic function of the standard normal distribution. The theorem follows now from Theorem 6.22. Q.E.D. There is also a multivariate version of the central limit theorem:
Theorem 6.24: Let X1,....,Xn be i.i.d. random vectors in k satisfying E(Xj) = , Var(Xj) = ' (1/n)'n , where is finite, and let X j'1Xj . Then n(X & ) 6d N k(0 , ) .
228 Proof: Let 0 k be arbitrary but not a zero vector. Then it follows from Theorem 6.23 that &) 6 N(0 , T) , hence it follows from Theorem 6.22 that for all t 0 , nT(X d
&)]) ' exp(&t 2T / 2) . Choosing t = 1, we thus have that for arbitrary limn64E(exp[i.t nT(X &)]) ' exp(&T / 2) . Since the latter is the characteristic 0 k , limn64E(exp[i.T n(X function of the Nk(0 , ) distribution, Theorem 6.24 follows now from Theorem 6.22. Q.E.D. Next, let M be a continuously differentiable mapping from k to m, and let the conditions of Theorem 6.24 hold. The question is: What is the limiting distribution of n((X) & ()) , if any? In order to answer this question, assume for the time being that k = m = 1, and let var(Xj) = F2, so that n(X & ) 6d N(0 , 2) . It follows from the mean value
theorem (see Appendix II) that there exists a random variable 8 0 [0,1] such that n((X) & ()) ' n ( X & ))( %(X&)) Since n ( X & ) 6d N(0 , 2) implies (X & ) 6d 0 , which by Theorem 6.16 implies that X 6p , it follows that % (X & ) 6p . Moreover, since the derivative ) is continuous in it follows now from Theorem 6.3 that )( % (X & )) 6p )() . Therefore, it follows from Theorem 6.21 that n((X) & ()) 6d N[0,2()())2] . Along similar lines, applying
the mean value theorem to each of the components of M separately, the following more general result can be proved. This approach is known as the *-method.
Theorem 6.25: Let Xn be a random vector in k satisfying
n(Xn & ) 6d Nk[0 , ] , where
0 k is nonrandom. Moreover, let (x) ' (1(x) , .... , m(x))T with x ' (x1 , .... , xk)T be a mapping from k to m such that the mk matrix of partial derivatives
229 M1(x) / Mx1 M1(x) / Mxk (x) ' ! " ! (6.36) Mm(x) / Mx1 Mm(x) / Mxk exists in an arbitrary small open neighborhood of and its elements are continuous in . Then n((Xn) & ()) 6d Nm[0 , ()()T ] .
6.8.
Stochastic boundedness, tightness, and the Op and op notations. The stochastic boundedness and related tightness concepts are important for various
reasons, but one of the most important reasons is that they are necessary conditions for convergence in distribution.
Definition 6.7: A sequence of random variables or vectors Xn is said to be stochastically bounded if for every g 0 (0,1) there exists a finite M > 0 such that infn$1P[||Xn|| # M] > 1&g .
Of course, if Xn is bounded itself, i.e., P[||Xn|| # M] ' 1 for all n, it is stochastically bounded as well, but the other way around may not be true. For example, if the Xn are equally distributed (but not necessarily independent) random variables with common distribution function F, then for every g 0 (0,1) we can choose continuity points !M and M of F such that P[|Xn| # M] = F(M)&F(&M) = 1!g. Thus, the stochastic boundedness condition limits the heterogeneity of the Xn s. Stochastic boundedness is usually denoted by Op(1): Xn = Op(1) means that the sequence Xn is stochastically bounded. More generally:
230 Definition 6.8: Let an be a sequence of positive non-random variables. Then Xn = Op(an) means that Xn /an is stochastically bounded, and Op(an) by itself represents a generic random variable or vector Xn such that Xn = Op(an).
The necessity of stochastic boundedness for convergence in distribution follows from the fact that:
Theorem 6.26: Convergence in distribution implies stochastic boundedness.
Proof: Let Xn and X be random variables with corresponding distribution functions Fn and F, respectively, and assume that Xn 6d X . Given an g 0 (0,1) we can choose continuity points !M1 and M1 of F such that F(M1) > 1&g/4 , F(&M1) < g/4 . Since limn64Fn(M1) ' F(M1) there exists an index n1 such that |Fn(M1) & F(M1)| < g/4 if n $ n1 , hence Fn(M1) > 1&g/2 if n $ n1 . Similarly, there exists an index n2 such that Fn(&M1) < g/2 if n $ n2 . Let m = max(n1 , n2) . Then infn$mP[|Xn| # M1] > 1&g . Finally, we can always choose an M2 so large that min1#n#m&1P[|Xn| # M2] > 1&g . Taking M ' max(M1 , M2) the theorem follows. The proof of the multivariate case is almost the same. Q.E.D. Note that since convergence in probability implies convergence in distribution, it follows trivially from Theorem 6.26 that convergence in probability implies stochastic boundedness. For example, let Sn ' 'j'1Xj , where the Xj's are i.i.d. random variables with
n
expectation and variance F2 < 4. If = 0 then Sn ' Op( n) , because by the central limit theorem, Sn/ n converges in distribution to N(0,F2). However, if 0 then only Sn ' Op(n) ,
231 because then Sn/ n & n 6d N(0 ,2) , hence Sn/ n ' Op(1) % Op( n) and thus Sn ' Op( n) % Op(n) = Op(n) . In Definition 6.2 I have introduced the concept of uniform integrability. It is left as an exercise to prove that
Theorem 6.27: Uniform integrability implies stochastic boundedness.
Tightness is the version of stochastic boundedness for probability measures:
Definition 6.9: A sequence of probability measures n on the Borel sets in k is called tight if for an arbitrary g 0 (0,1) there exists a compact subset K of k such that infn$1 n(K) > 1!g.
Clearly, if Xn = Op(1) then the sequence of corresponding induced probability measures n is tight, because the sets of the type K ' {x 0 k : ||x|| # M} are closed and bounded for M < 4 and therefore compact. For sequences of random variables and vectors the tightness concept does not add much over the stochastic boundedness concept, but the tightness concept is fundamental in proving socalled functional central limit theorems. If Xn = Op(1) then obviously for any * > 0, Xn ' Op(n ) . But Xn/n is now more than stochastically bounded, because then we also have that Xn/n 6p 0 . The latter is denoted by Xn ' op(n ) :
232 Definition 6.10: Let an be a sequence of positive non-random variables. Then Xn = op(an) means that Xn /an converges in probability to zero (or a zero vector if Xn is a vector), and op(an) by itself represents a generic random variable or vector Xn such that Xn = op(an). Moreover, the sequence 1/an represents that rate of convergence of Xn .
Thus, Xn 6p X can also be denoted by Xn ' X % op(1) . This notation is handy if the difference of Xn and X is a complicated expression. For example, the result of Theorem 6.25 is due to the fact that by the mean value theorem + op(1), where M1(x) / Mx () ' n
x' %1,n(X n&)
() n(X & ) ' () n(X & ) n((Xn) & ()) ' n n n
! Mm(x) / Mx
x' %k,n(X n&)
, with j,n 0 [0,1], j ' 1 , .... , k .
() & ()) n(X & ) can now be represented by op(1), because The remainder term ( n n () 6 () and n(X & ) 6 N [0 ,] , hence by Theorem 6.21 this remainder term n p n d k converges in distribution to the zero vector and thus it converges in probability to the zero vector.
6.9.
Asymptotic normality of M-estimators In this section I will set forth conditions for the asymptotic normality of M-estimators, in
of a parameter 0 m is addition to the conditions for consistency. An estimator 0 asymptotically normally distributed if there exist an increasing sequence of positive numbers an & ) 6 N [0,] . Usually, a ' and a positive semi-definite mm matrix G such that an( n 0 d m but there are exceptions to this rule. n,
233 Asymptotic normality is fundamental for econometrics. Most of the econometric tests rely on it. Moreover, the proof of the asymptotic normality theorem below also illustrates nicely the usefulness of the main results in this chapter. Given that the data is a random sample, we only need a few addition conditions over the conditions of Theorems 6.10 and 6.11:
Theorem 6.28: Let in addition to the conditions of Theorems 6.10 and 6.11 the following conditions be satisfied: (a) (b) (c) (d) 1 is convex; 0 is an interior point of 1; For each x 0 k , g(x,2) is twice continuously differentiable on 1;. For each pair i , i of components of 2, E[sup0*M2g(X1 , )/(Mi Mi )*] < 4;
1 2 1 2
(e)
The mm matrix A ' E
M2g(X1 , 0) M0 M0
T
is nonsingular;
(f)
The mm matrix B ' E
Mg(X1 , 0) M0
T
Mg(X1 , 0) M0
is finite.
Then
& ) 6 N [0 , A &1BA &1] . n( 0 d m
Proof: I will prove the theorem for the case m = 1 only, leaving the general case as an exercise. 6 . Since is an interior point of I have already established in Theorem 6.11 that p 0 0
234 is an interior point converges to 1, and consequently the probability that 1, the probability that () = (1/n)'n the first-order condition for a maximum of Q j'1g(Xj , ) in ' holds converges to 1. Thus: )( ) ' 0] ' 1, limn64P[Q (6.37)
)() ' dQ ()/d . Next, observe from the mean value theorem that there exists where as usual, Q 0 [0,1] such that a ( ) ' nQ
)
)( ) % Q ))( % ( & )) n( & ) , nQ 0 0 0 0
(6.38)
))() ' d 2Q ()/(d)2 . Note that by the convexity of 1, where Q ( & ) 0 ] ' 1, P[0% 0 , and by the consistency of ( & )] ' plimn64[0% 0 0 (6.40) (6.39)
Moreover, it follows from Theorem 6.10 and conditions (c) and (d), with the latter adapted to the univariate case, that ))() & Q))()* ' 0 . plimn64sup0*Q (6.41)
where Q))() is the second derivative of Q() ' E [g(X1 , )] . Then it follows from (6.39), (6.40), (6.41) and Theorem 6.12 that ))( % ( & )) ' Q))( ) 0 . plimn64Q 0 0 0 (6.42)
Note that Q))(0) corresponds to the matrix A in condition (e), so that Q))(0) is positive in the "argmin" case and negative in the "argmax" case. Therefore, it follows from (6.42) and Slutskys theorem (Theorem 6.3) that ( % ( & ))&1 ' Q))( )&1 ' A &1. plimn64Q 0 0 0 Now (6.38) can be rewritten as
))
(6.43)
235 & ) ' &Q ))( % ( & ))&1 nQ )( ) % Q ))( % ( & ))&1 nQ )( ) n( 0 0 0 0 0 0 ( & ))&1 nQ )( ) % o (1) , ))( % ' &Q 0 0 0 p (6.44)
where the op(1) term follows from (6.37), (6.43) and Slutskys theorem. Because of condition (b), the first-order condition for 0 applies, i.e., Q)(0) ' E [dg(X1 , 0)/d0] ' 0 . Moreover, condition (f), adapted to the univariate case, now reads as: var [dg(X1 , 0)/d0] ' B 0 (0,4) . )( ) ' (1/ n)'n nQ j'1dg(Xj , 0)/d0 6d N[0,B] . 0 Now it follows from (6.43), (6.47) and Theorem 6.21 that ))( % ( & ))&1 nQ )( ) 6 N[0 , A &1BA &1] , &Q 0 0 0 d (6.48) (6.46) (6.45)
Therefore, it follows from (6.45), (6.46), and the central limit theorem (Theorem 6.23) that (6.47)
hence the result of the theorem under review for the case m = 1 follows from (6.44), (6.48) and Theorem 6.21. Q.E.D. The result of Theorem 6.28 is only useful if we are able to estimate the asymptotic variance matrix A &1BA &1 consistently, because then we will be able to design tests of various hypotheses about the parameter vector 0 .
Theorem 6.29: Let

n M2g(X , 1 ) ' 1 A , j n j'1 M M T
(6.49)
and
236 ' 1 B j n j'1

n
) Mg(X1 , M
T
) Mg(X1 , M
(6.50)
' A , and under the additional condition that Under the conditions of Theorem 6.28, plimn64A ' B . Consequently, plim A &1B A &1 ' A &1BA &1 . E[sup0||Mg(X1 , )/MT||2] < 4, plimn64B n64
Proof: The theorem follows straightforwardly from the uniform weak law of large numbers and various Slutskys theorems, in particular Theorem 6.21.
6.10.
Hypotheses testing As an application of Theorems 6.28 and 6.29, consider the problem of testing a null
hypothesis against an alternative hypothesis of the form H0 : R0 ' q , H1 : R0 q , (6.51)
respectively, where R is a given rm matrix of rank r # m, and q is a given r1 vector. Under the &q) 6 N [0 , RA &1BA &1R T] , null hypothesis in (6.51) and the conditions of Theorem 6.2, n(R d r and if the matrix B is nonsingular then the asymptotic variance matrix involved is nonsingular. Then is follows from Theorem 6.21 that:
Theorem 6.30: Under the conditions of Theorems 6.28 and 6.29, the additional condition that B is nonsingular, and the null hypothesis in (6.51) with R of full rank r, &q)T RA &1B A &1R T &1(R &q) 6 2 Wn ' n(R d r. On the other hand, under the alternative hypothesis in (6.51), Wn/n 6p (R0&q)T RA &1BA &1R T &1(R0&q) > 0. (6.53) (6.52)
237 The statistic Wn is now the test statistic of the Wald test of the null hypothesis in (6.51). Given the size " 0 (0,1), choose a critical value $ such that for a r distributed random variable Z, P[Z > ] ' , so that under the null hypothesis in (6.51), P[Wn > ] 6 ". Then the null hypothesis is accepted if Wn # and rejected in favor of the alternative hypothesis if Wn > . Due to (6.53), this test is consistent. In the case that r = 1, so that R is a row vector, we can modify (6.52) to &1B A &1R T &1/2(R &q) 6 N(0,1) , t n ' n RA d whereas under the alternative hypothesis, (6.53) becomes t n/ n 6p RA &1BA &1R T &1/2(R0&q) 0 . (6.55) (6.54)
2
These results can be used to construct a two-sided or one sided test, similarly to the t-test we have seen before in the previous chapter. In particular,
Theorem 6.31: Assume that the conditions of Theorem 6.30 hold. Let i,0 be component i of 0 ,
( . Consider the hypotheses H : ' ( be component i of and let i,0 , H1 : i,0 i,0 , 0 i,0 i
where i,0 is given (often the value i,0 ' 0 is of special interest). Let the vector ei be column i of the unit matrix Im. Then under H0, ti ' & ( n( i,0) i A e ) ei A B i
T
&1
&1
6d N(0,1) ,
(6.56)
whereas under H1, ti/ n 6p & ( i,0 i,0 ei A BA ei)

T &1 &1
0.
(6.57)
238 Given the size " 0 (0,1), choose a critical value $ such that for a standard normally distributed random variable U, P[|U| > ] ' , so that by (6.56), P[|ti| > ] 6 " if the null hypothesis is true. Then the null hypothesis is accepted if |ti| # and rejected in favor of the alternative hypothesis if |ti| > . It is obvious from (6.57) that this test is consistent. The statistic ti in (6.56) is usually referred to as a t-test statistic because of the similarity of this test with the t-test in the normal random sample case. However, its finite sample distribution under the null hypothesis may not be of the t distribution type at all. Moreover, in the
( , and if case i,0 ' 0 the statistic ti is called the t-value (or pseudo t-value) of the estimator i
the test rejects the null hypothesis this estimator is said to be significant at the "100% significance level.
6.12. 1.
Exercises Let Xn ' (X1,n , ... , Xk,n)T and c ' (c1 , ... , ck)T . Prove that plimn64Xn ' c if and only if
plimn64Xi,n ' ci for i = 1,..,k. 2. 3. 4. 5. 6. 7. 8. 9. Prove that if P(*Xn* # M) = 1 and Xn 6p X then P(*X* # M) = 1. Complete the proof of Theorem 6.5. Prove Theorem 6.12. Explain why the random vector Xn in (6.16) does not converge in distribution. Prove Theorem 6.16. Prove Theorem 6.17. Prove (6.21). Prove Theorem 6.19.
239 10. 11. 12. 13. Prove Theorem 6.20, using Theorem 6.18. Finish the proof of Theorem 6.21. Answer the questions Why? in the proof of Corollary 6.2. Prove that the limit (6.35) is just the characteristic function of the standard normal
distribution. 14. 15. 16. 17. 18. 19. Prove the first and the last equality in (6.32). Prove Theorem 6.25. Prove Theorem 6.27. Hint: Use Chebishevs inequality for first absolute moments. Adapt the proof of Theorem 6.28 for m = 1 to the multivariate case m > 1. Prove Theorem 6.29. Formulate the conditions (additional to Assumption 6.1 ) for the asymptotic normality of
2
the nonlinear least squares estimator (6.15) for the special case that P[E(U1 |X1) ' 2] ' 1 .
Appendices
6.A. Proof of the uniform weak law of large numbers First, recall that "sup" denotes the smallest upper bound of the function involved, and similarly, "inf" is the largest lower bound. Now let for arbitrary * > 0 and ( 0 , (() = {0 : 2&(2 < } . Using the fact that supx*f(x)* # max{|sup xf(x)|,|infxf(x)|} # |supxf(x)| % |infxf(x)|, it follows that
240 sup *(1/n)'j'1g(Xj , ) & E[g(X1 , )*

n n
0(()
# * sup (1/n)'j'1g(Xj , ) & E[g(X1 , )] *

0(()
(6.58)
% * inf (1/n)'j'1g(Xj , ) & E[g(X1 , )] *

n 0(()
Moreover, sup (1/n)'j'1g(Xj , ) & E[g(X1 , )] # (1/n)'j'1 sup g(Xj , ) &

n n 0(()
0(()
0(()
inf E[g(X1 , )] (6.59)
# |(1/n)'j'1 sup g(Xj , ) & E[ sup g(X1 , )]|

n 0(() 0(()
% E[ sup g(X1 , )] & E[ inf g(X1 , )]

0(() 0(()
and similarly, inf (1/n)'j'1g(Xj , ) & E[g(X1 , )] $ (1/n)'j'1 inf g(Xj , ) & sup E[g(X1 , )]
n n 0(() 0(()
0(()
$ &|(1/n)'j'1 inf g(Xj , ) & E[ inf g(X1 , )]|

n 0(() 0(()
(6.60)
% E[ inf g(X1 , )] & E[ sup g(X1 , )]

0(() 0(()
Hence | sup (1/n)'j'1g(Xj , )&E[g(X1 , )] | # |(1/n)'j'1 sup g(Xj , )&E[ sup g(X1 , )]|
n n 0(() 0(() 0(()
% |(1/n)'j'1 inf g(Xj , ) & E[ inf g(X1 , )]|

n 0(() 0(()
(6.61)
% E[ sup g(X1 , )] & E[ inf g(X1 , )]

0(() 0(()
241
and similarly | inf (1/n)'j'1g(Xj , )&E[g(X1 , )] | # |(1/n)'j'1 sup g(Xj , )&E[ sup g(X1 , )]|
n n 0(() 0(() 0(()
% |(1/n)'j'1 inf g(Xj , ) & E[ inf g(X1 , )]|

n 0(() 0(()
(6.62)
% E[ sup g(X1 , )] & E[ inf g(X1 , )]

0(() 0(()
Combining (6.58), (6.61), and (6.62) it follows that sup *(1/n)'j'1g(Xj , )&E[g(X1 , )* # 2|(1/n)'j'1 sup g(Xj , )&E[ sup g(X1 , )]|
n n 0(() 0(()
0(()
% 2|(1/n)'j'1 inf g(Xj , ) & E[ inf g(X1 , )]|

n 0(() 0(()
(6.63)
% 2 E[ sup g(X1 , )] & E[ inf g(X1 , )]

0(() 0(()
It follows from the continuity of g(x,) in and the dominated convergence theorem [Theorem 6.5] that limsup sup E [ sup g(X1 , ) &
90 (0 0(()
0(()
inf g(X1 , )]
# limE sup [ sup g(X1 , ) &

90 (0 0(()
0(()
inf g(X1 , )] ' 0 ,
hence we can choose so small that sup E [ sup g(X1 , ) &

0(()
(0
0(()
inf g(X1 , )] < g/4 .
(6.64)
Furthermore, by the compactness of it follows that there exist a finite number of *'s, say 1,...,N(), such that
242 d ^ (i) .
N() i'1
(6.65)
Therefore, it follows from Theorem 6.2, (6.63), (6.64), and (6.65), that P sup0*(1/n)'j'1g(Xj , ) & E[g(X1 , )]* > g
n
# P max1#i#N() sup0 ( )*(1/n)'j'1g(Xj , ) & E[g(X1 , )]* > g

n
i
# 'i'1 P sup0 ( )*(1/n)'j'1g(Xj , ) & E[g(X1 , )]* > g

N() n
i
# 'i'1 P |(1/n)'j'1sup0 ( )g(Xj , ) & E[sup0 ( )g(X1 , )]|

N() n
( (
n |(1/n)'j'1inf0 ( )g(Xj , ) (
(6.66) & E[inf0 ( )g(X1 , )]| > g/4

(
# 'i'1 P |(1/n)'j'1sup0 ( )g(Xj , ) & E[sup0 ( )g(X1 , )]| > g/8

N() n
( (
% 'i'1 P |(1/n)'j'1inf0 ( )g(Xj , ) & E[inf0 ( )g(X1 , )]| > g/8 6 0 as n 6 4 .

N() n
( (
6.B.
Almost sure convergence and strong laws of large numbers
6.B.1. Preliminary results First, I will show the equivalence of (6.6) and (6.7) in Definition 6.3:
Theorem 6.B.1: Let Xn and X be random variables defined on a common probability space {S,,P}. Then limn64P(*Xm & X* # g for all m $ n) = 1 for arbitrary g > 0 if and only if P(limn64Xn ' X) = 1. This result carries over to random vectors, by replacing |.| with the
243 Euclidean norm ||.||.
Proof: Note that the statement P(limn64Xn ' X) = 1 reads: There exists a set N 0 with P(N) = 0 such that limn64Xn() ' X() pointwise in T 0 S\N. Such a set N is called a null set. Denote An(g) ' _m'n{ 0 : *Xm() & X()* # g} .
4
(6.67)
First, assume that for arbitrary g > 0, limn64P(An(g)) ' 1. Since An(g) d An%1(g) it follows that
4 4 4 countable union N ' ^k'1N(1/k) . Now let T 0 S\N. Then 0 \ ^k'1N(1/k) ' _k'1N (1/k) = 4 4 4
P[^n'1An(g)] ' limn64P(An(g)) ' 1, hence N(g) ' \ ^n'1An(g) is a null set, and so is the
4 4
_k'1^n'1An(1/k) , hence for each positive integer k, 0 ^n'1An(1/k) . Since An(1/k) d An%1(1/k)
it follows now that for each positive integer k there exists a positive integer nk() such that 0 An(1/k) for all n $ nk(). Let k(g) be the smallest integer $ 1/g, and let n0( , g) ' nk(g)() . Then for arbitrary g > 0, *Xn() & X()* # g if n $ n0( , g) . Therefore, limn64Xn() ' X() pointwise in T 0 S\N, hence P(limn64Xn ' X) = 1. Next, assume that the latter holds, i.e., the exists a null set N such that limn64Xn() = X() pointwise in T 0 S\N. Then for arbitrary g > 0 and T 0 S\N there exists a positive integer consequently, 1 = P( \ N) # P[^n'1An(g) ] . Since An(g) d An%1(g) it follows now that
4
n0( , g) such that 0 An ( , g)(g) and therefore also 0 ^n'1An(g) . Thus, \ N d ^n'1An(g) and
4 4
0
limn64P(An(g)) = P[^n'1An(g)] = 1. Q.E.D.

4
The following theorem, known as the Borel-Cantelli lemma, provides a convenient condition for almost sure convergence.
Theorem 6.B.2: (Borel-Cantelli). If for arbitrary g > 0, 'n'1P(*Xn & X* > g) < 4 , then
4
244
Xn 6 X a.s.
(g) be the complement of the set A (g) in (6.67). Then Proof: Let A n n where the latter conclusion follows from the condition that 'n'1P(*Xn & X* > g) < 4 . 6 Thus,
4 4 (g)) ' P[^4 P(A m'n{ 0 : *Xm() & X()* > g} ] # 'm'nP[|X n & X| > g] 6 0 , n
(g)) ' 0 , hence lim P(A (g)) ' 1. Q.E.D. limn64P(A n n64 n The following theorem establishes the relationship between convergence in probability and almost sure convergence:
Theorem 6.B.3: Xn 6p X if and only if every subsequence nm of n = 1,2,3,... contains a further subsequence nm(k) such that for k 6 4, Xn
m(k)
6 X a.s.
Proof: Suppose that Xn 6p X is not true, but every subsequence nm of n = 1,2,3,... contains a further subsequence nm(k) such that for k 6 4, Xn
m(k)
6 X a.s. Then there exist numbers g > 0,
0 (0,1) and a subsequence nm such that supm$1P[|Xn & X| # g] # 1& . Clearly, the same holds
m
for every further subsequence nm(k) , which contradicts the assumption that there exists a further subsequence nm(k) such that for k 6 4, Xn
m(k)
6 X a.s. This proves the only if part.

m
Next, suppose that Xn 6p X . Then for every subsequence nm, Xn 6p X . Consequently, for each positive integer k, limm64P[|Xn & X| > k &2] = 0, hence for each k we can find a
m
positive integer nm(k) such that P[|Xn

4
'k'1k &2 < 4 . The latter implies that 'k'1P[|Xn

4
m(k)
& X| > k &2] # k &2 . Thus, 'k'1P[|Xn

4
m(k)
m(k)
& X| > k &2] #
& X| > g] < 4 for each g > 0, hence by
245 Theorem 6.B.2, Xn

m(k)
6 X a.s. Q.E.D.
6.B.2. Slutskys theorem Theorem 6.B.1 can be used to prove Theorem 6.7. Theorem 6.3 was only proved for the special case that the probability limit X is constant. However, the general result of Theorem 6.3 follows straightforwardly from Theorems 6.7 and 6.B.3. Let us restate Theorems 6.3 and 6.7 together:
Theorem 6.B.4: (Slutsky's theorem). Let Xn a sequence of random vectors in k converging a.s. [in probability] to a (random or constant) vector X. Let Q(x) be an m -valued function on k which is continuous on an open (Borel) set B in k for which P(X 0 B) = 1). Then Q(Xn) converges a.s. [in probability] to Q(X).
Proof: Let Xn 6 X a.s. and let {S ,,P} be the probability space involved. According to Theorem 6.B.1 there exists a null set N1 such that limn64Xn() ' X() pointwise in T 0 S\N1. Moreover, let N2 ' { 0 : X() B} . Then also N2 is a null set, and so is N ' N1 ^ N2 .
Pick an arbitrary T 0 S\N. Since Q is continuous in X() it follows from standard calculus that limn64(Xn()) = (X()) . By Theorem 6.B.1 this result implies that (Xn) 6 (X) a.s. Since the latter convergence result holds along any subsequence, it follows from Theorem 6.B.3 that Xn 6p X implies (Xn) 6p (X) . Q.E.D.
246 6.B.3. Kolmogorovs strong law of large numbers I will now provide the proof of Kolmogorovs strong law of large numbers, based on the elegant and relatively simple approach of Etemadi (1981). This proof (and other versions of the proof as well) employs the notion of equivalent sequences:
Definition 6.B.1: Two sequences of random variables, Xn and Yn, n $1, are said to be equivalent if 'n'1P[Xn Y n] < 4.
4
The importance of this concept lies in the fact that if one of the equivalent sequences obeys a strong law of large numbers, then so does the other one:
Lemma 6.B.1: If Xn and Yn are equivalent and (1/n)'j'1Y j 6 a.s. then (1/n)'j'1Xj 6 a.s.
n n
Proof: Without loss of generality we may assume that = 0. Let {S ,,P} be the probability space involved, and let An ' ^m'n{ 0 : Xm() Y m()} .
4
Then P(An) # 'm'nP(Xm Y m) 6 0 , hence limn64P(An) ' 0 and thus P(_n'1An) ' 0 . The
4 4
latter implies that for each 0 \{_n'1An} there exists a natural number n(() such that
4
Xn() ' Y n() for all n $ n((), because if not there exists a countable infinite subsequence 0 _n'1An . Now let N1 be the null set on which (1/n)'j'1Y j 6 0 a.s. fails to hold, and let N =
4 n 4
nm() , m ' 1,2,3,.... , such that Xn ()() Yn ()() , hence 0 An for all n $1 and thus
k k
N1^ {_n'1An} . Since for each 0 \N , Xj() and Y j() differ for at most a finite number of js,
and limn64(1/n)'j'1Y j() ' 0 , it follows that also limn64(1/n)'j'1Xj() ' 0 . Q.E.D.
n n
247
The following construction of equivalent sequences plays a key-role in the proof of the strong law of large numbers.
Lemma 6.B.2: Let Xn , n $1, be i.i.d., with E[|Xn|] < 4 , and let Y n ' Xn . I(|Xj| # n) . Then Xn and Yn are equivalent.
Proof: The lemma follows from:

4 'n'1P[Xn
Y n] '
4 'n'1P[|Xn|
> n] '
4 'n'1P[|X1|
> n] #
|X1|
P[|X1| > t]dt ' E[I(|X1| > t)]dt m m

0 0
# E Q.E.D.
m0
I(|X1| > t)]dt ' E
m0
dt ' E[|X1|] < 4 .
6 E[max(0,X1)] a.s. and (1/n)'j'1max(0,&Xj) 6 E[max(0,&X1)] a.s. Then it is easy to verify from
n
Now let Xn , n $1, be the sequence in Lemma 6.B.2, and suppose that (1/n)'j'1max(0,Xj)
n
Theorem 6.B.1, by taking the union of the null sets involved, that
n max(0,Xj) 1 j n j'1 max(0,&Xj)
E[max(0,X1)] E[max(0,&X1)]
a.s.
n
Applying Slutskys theorem (Theorem 6.B.4) with (x,y) = x !y it follows that (1/n)'j'1Xj 6 E[X1] a.s. Therefore, the proof of Kolmogorovs strong law of large numbers is completed by Lemma 6.B.3 below.
248
1. Then (1/n)'j'1Xj 6 E[X1] a.s.

n
Lemma 6.B.3: Let the conditions of Lemma 6.B.2 hold, and assume in addition that P[Xn $ 0] =
Proof: Let Z(n) ' (1/n)'j'1Y j and observe that

n
Var(Z(n)) # (1/n 2)'j'1E[Y j ] ' (1/n 2)'j'1E[Xj I(Xj # j)]

n 2 n 2
(6.68) # n
&1 2 E[X1 I(X1
# n)] .
Next let > 1 and g > 0 be arbitrary. It follows from (6.68) and Chebishevs inequality that j P[|Z([ ]) & E[Z([ ])]| > g] # j Var(Z([ ]))/g # j
4 4 4 n n n 2 n'1 n'1 n'1
E[X1 I(X1 # [n])] g2[n] (6.69)
# g&2E X1 'n'1I(X1 # [n])/[n] ,

2 4
where [n] is the integer part of n . Let k be the smallest natural number such that X1 # [k] , and note that [n] > n/2 . Then the last sum in (6.69) satisfies 'n'1I(X1 # [n])/[n] # 2j &n ' 2. 'n'0&n &k #
4 4 4 n'k
2 , (&1)X1
hence E X1 'n'1I(X1 # [n])/[n] #

2 4
2 E[X1] < 4. &1
Consequently, it follows from the Borel-Cantelli lemma that Z([n]) & E[Z([n]) 6 0 a.s. Moreover, it is easy to verify that E[Z([n]) 6 E[X1] . Hence, Z([n]) 6 E[X1] a.s.
249 For each natural number k > there exists a natural number nk such that [ k] # k # [
n k%1 n
] , and since the Xjs are non-negative we have [ k] [

n k%1 n
Z([ k]) # Z(k) # ]
n k%1 n
Z([
n k%1
])
[ k]
(6.70)
The left-hand side expression in (6.70) converges a.s. to E[X1]/ as k 6 4 , and the right-hand side converges a.s. to E[X1] , hence we have with probability 1, 1 E[X1] # liminfk64Z(k) # limsupk64Z(k) # E[X1] In other words, denoting Z ' liminfk64Z(k), Z ' limsupk64Z(k) , there exists a null set N (depending on ) such that for all 0 \N , E[X1]/ # Z() # Z() # E[X1] . Taking the union N of N over all rational > 1, so that N is also a null set7, the same holds for all 0 \N and all rational > 1. Letting 91 along the rational values then yields limk64Z(k) = (1/n)'j'1Y j 6 E[X1] a.s., which by Lemmas 6.B.2 and 6.B.3 implies that
n
Z() ' Z() ' E[X1] for all 0 \N . Therefore, by Theorem 6.B.1,
(1/n)'j'1Xj 6 E[X1] a.s.. Q.E.D.

n
This completes the proof of Theorem 6.6.
6.B.5. The uniform strong law of large numbers and its applications Proof of Theorem 6.13: It follows from (6.63) , (6.64) and Theorem 6.6 that limsup sup *(1/n)'j'1g(Xj , ) & E[g(X1 , )*
n n64 0(()
# 2 E[ sup g(X1 , )] & E[ inf g(X1 , )]

0(() 0(()
< g/2 a.s. ,
250 hence (6.66) can now be replaced by limsup sup*(1/n)'j'1g(Xj , ) & E[g(X1 , )]*
n n64 0
# limsup max
n64
1#i#N() 0(i)
sup *(1/n)'j'1g(Xj , ) & E[g(X1 , )]* # g/2 a.s.

n
(6.71)
Replacing g/2 with 1/m, m $ 1, the last inequality in (6.71) reads: Let {S ,,P} be the probability space involved. For m = 1,2,3, ... there exist a null sets Nm such that for all 0 \Nm , limsup sup*(1/n)'j'1g(Xj() , ) & E[g(X1 , )]* # 1/m
n n64 0 4
(6.72)
and the same holds for all 0 \^k'1Nk , uniformly in m. Letting m 6 4 in (6.72), Theorem 6.13 follows. Note that this proof is based on a seminal paper by Jennrich (1969). An issue that has not yet been addressed is whether sup0*(1/n)'j'1g(Xj , )&E[g(X1 , )*
n
is a well-defined random variable. If so, we must have that for arbitrary y > 0, { 0 : sup0*(1/n)'j'1g(Xj() , )&E[g(X1 , )* # y}
n
' _0{ 0 : *(1/n)'j'1g(Xj() , )&E[g(X1 , )* # y} 0 .

n
However, this set is an uncountable intersection of sets in and therefore not necessarily a set in itself. The following lemma, which is due to Jennrich (1969), shows that in the case under review there is no problem.
Lemma 6.B.4: Let f(x,) be a real function on B , B d k , d m , where B is a Borel set
251 and is compact (hence is a Borel set) such that for each x in B, f(x,) is continuous in 0 , and for each 0 , f(x,) is Borel measurable. Then there exists a Borel measurable mapping (x): B 6 such that f(x,(x)) ' inf0f(x,) , hence the latter is Borel measurable itself. The same result holds for the sup case.
Proof: I will only prove this result for the special case k = m = 1, B = , = [0,1]. Denote n ' ^j'1{0 , 1/j , 2/j , .... ,(j&1)/j , 1} , and observe that
n 4
n d n%1, and that ( ' ^n'1n is the set of all rational numbers in [0,1]. Since n is finite,
for each positive integer n there exists a Borel measurable function n(x): 6 n such that f(x,n(x)) ' inf0 f(x,) . Let (x) = liminfn64n(x) . Note that (x) is Borel measurable. For each x
n
there exists a subsequence nj (which may depend on x) such that (x) ' limj64n (x) . Hence by
j
continuity, f(x,(x)) = limj64f(x,n (x)) ' limj64inf0 f(x,) . Now suppose that for some g > 0
j nj
the latter is greater or equal to g + inf0 f(x,) . Then, since for

(
m # nj , inf0 f(x,) # inf0 f(x,) , and the latter is monotonic non-increasing in m, it follows
nj m
that for all n $1, inf0 f(x,) $ g % inf0 f(x,) . It is not too hard to show, using the continuity
n (
of f(x,) in , that this is not possible. Therefore, f(x,(x)) ' inf0 f(x,) , hence by continuity,
(
f(x,(x)) ' inf0f(x,) . Q.E.D. Proof of Theorem 6.14: Let {S ,,P} be the probability space involved, and denote . . Now (6.9) becomes Q( ) 6 Q( ) a.s., i.e., there exists a null set N such that for all n ' n 0 0 \N , limn64Q(n()) ' Q(0) . Suppose that for some 0 \N there exists a subsequence nm() and an g > 0 such that (6.73)
252 infm$12n
m()
() & 02 > g . Then by the uniqueness condition there exists a () > 0 such that
m()
Q(0) & Q(n
()) $ () for all m $ 1 , which contradicts (6.73). Hence, for every
subsequence nm() we have limm64n m()() ' 0 , which implies that limn64n() ' 0 .
Proof of Theorem 6.15: The condition Xn 6 c a.s. translates as: There exists a null set N1 such that for all 0 \N1 , limn64Xn() ' c. By the continuity of M on B the latter implies that limn64|(Xn()) & (c)| ' 0 , and that for at most a finite number of indices n, Xn() B . Similarly, the uniform a.s. convergence condition involved translates as: There exists a null set N2 such that for all 0 \N2 , limn64supx0B*n(x,) & (x)* 6 0 . Take N ' N1^ N2 . Then for all 0 \N , limsupn64|n(Xn(),) & (c)| # limsupn64|n(Xn(),) & (Xn())| % limsupn64|(Xn()) & (c)| # limsupn64supx0B*n(x,) & (x)* % limsupn64|(Xn()) & (c)| ' 0 .
6.C.
Convergence of characteristic functions and distributions In this appendix I will provide the proof of the univariate version of Theorem 6.22. Let Fn
be a sequence of distribution functions on with corresponding characteristic functions nn(t), and let F be a distribution function on with characteristic function n(t) ' limn64nn(t) . Denote F(x) ' lim90liminfn64Fn(x%) , F(x) ' lim90limsupn64Fn(x%) . The function F(x) is right continuous and monotonic non-decreasing in x but not necessarily a distribution function itself, because limx84 F(x) may be less than one, or even zero. On the other
253 hand, it is easy to verify that limx9&4 F(x) = 0. Therefore, if limx64 F(x) = 1 then F is a distribution function. The same applies to F(x): If limx64 F(x) = 1 then F is a distribution function. I will first show that limx64 F(x) ' limx64 F(x) ' 1 , and then that F(x) ' F(x) .
Lemma 6.C.1: Let Fn be a sequence of distribution functions on with corresponding characteristic functions nn(t), and suppose that n(t) ' limn64nn(t) pointwise for each t in , where n is continuous in t = 0. Then F(x) ' lim90liminfn64Fn(x%) is a distribution function, and so is F(x) ' lim90limsupn64Fn(x%) .
Proof: For T > 0 and A > 0 we have 1 1 1 nn(t)dt ' exp(i.t.x)dFn(x)dt ' exp(i.t.x)dtdFn(x) m m m 2T 2T 2T m m
&T T T 4 4 T
'
2A
1 sin(Tx) cos(t.x)dtdFn(x) ' dFn(x) m m m 2T Tx

&4 &T &4 &2A 4
&T &4 4 T
&4 &T
(6.74)
'
&2A
sin(Tx) sin(Tx) sin(Tx) dFn(x) % dFn(x) % dFn(x) m Tx m Tx m Tx

&4 2A
Since |sin(x)/x| # 1 and |Tx|&1 # (2TA)&1 for |x| > 2A it follows from (6.74) that 1 /0 1 n (t)dt/0 # 2 dF (x) % 1 dFn(x) % dF (x) n 00 T m n 00 m AT m AT m n 00 &T 00 &2A &4 2A
T 2A &2A 4
1 1 1 1 ' 2 1 & dFn(x) % ' 2 1 & n([&2A,2A]) % , 2AT m AT 2AT AT

&2A
2A
(6.75)
where n is the probability measure on the Borel sets in corresponding to Fn. Hence, putting
254 T ' A &1 it follows from (6.75) that n([&2A,2A]) $ /00A nn(t)dt/00 & 1 , 00 m 00 0 &1/A 0
1/A
(6.76)
which can be rewritten as Fn(2A) $ /00A nn(t)dt/00 & 1 % Fn(&2A) & n({&2A}) 00 m 00 0 &1/A 0
1/A
(6.77)
Now let 2A and !2A be continuity points of F . Then it follows from (6.77) , the condition that n(t) ' limn64nn(t) pointwise for each t in , and the bounded8 convergence theorem that F(2A) $ /00A n(t)dt/00 & 1 % F(&2A) 00 m 00 0 &1/A 0
1/A
(6.78)
Since n(0) = 1 and n is continuous in 0 the integral in (6.78) converges to 2 for A 6 4 . Moreover, F(&2A) 9 0 if A 6 4 . Consequently, it follows from (6.78) that limA64 F(2A) ' 1 . By the same argument it follows that limA64 F(2A) ' 1 . Thus, F and F are distribution functions. Q.E.D.
Lemma 6.C.2: Let Fn be a sequence of distribution functions on such that F(x) = lim90liminfn64Fn(x%) and F(x) ' lim90limsupn64Fn(x%) are distribution functions. Then for every bounded continuous function on and every g > 0 there exist subsequences n k(g) and n k(g) such that
255 limsupk64 (x)dFn (g)(x) & (x)dF(x) < g , limsupk64 (x)dFn (g)(x) & (x)dF(x) < g . k k m /m m / /m /
Proof: Without loss of generality we may assume that (x) 0 [0,1] for all x. For any g > 0 we can choose continuity points a 1!g. Moreover, we can choose continuity points a = c1 < c2 <...< cm = b of F(x) such that for j = 1,..,m-1, sup (x) &
x0(c j,cj%1]
inf (x) # g .
x0(c j,cj%1]
(6.79)
Furthermore, there exists a subsequence nk (possibly depending g) on such that limk64Fn (cj) ' F(cj) for j '1,2.,...,m .
k
(6.80)
Now define (x) ' inf (x) for x 0 (cj,cj%1] , j ' 1,.., m&1 , (x) ' 0 elsewhere. (6.81)
x0(c j,cj%1]
Then by (6.79), 0 # (x) & (x) # g for x 0 (a,b] , and 0 # (x) & (x) # 1 for x (a,b] , hence limsup (x)dFn(x) & (x)dFn(x) m / n64 /m # limsup
n64
x0(a,b]
*(x)&(x)*dFn(x) %
x(a,b]
*(x)&(x)*dFn(x)
(6.82)
# g % 1 & limsup Fn(b) & Fn(a) # g % 1 & F(b) & F(a) # 2g .

n64
Moreover, if follows from (6.79) and (6.81) that (x)dF(x) & (x)dF(x) # 2g , /m m / (6.83)
256 and from (6.80) that limk64 (x)dFn (x) ' (x)dF(x) . k m m Combining (6.82), (6.83) and (6.84) it follows that limsupk64 (x)dFn (x) & (x)dF(x) # limsupk64 (x)dFn (x) & (x)dFn (x) k k k /m m / /m m / % limsupk64 (x)dFn (x) & (x)dF(x) % limsupk64 (x)dF(x) & (x)dF(x) < 4g k /m m / /m m / A similar result holds for the case F . Q.E.D. Let n((t) be the characteristic function of F . Since n(t) ' limn64nn(t) , it follows from Lemma 6.C.2 that for each t and arbitrary g > 0, |n(t) & n((t)| < g , hence n(t) ' n((t) . The same result holds for the characteristic function n((t) of F : n(t) ' n((t) . Consequently, n(t) = limn64nn(t) is the characteristic function of both F and F , which by Lemma 6.C.1 are distribution functions. By the uniqueness of characteristic functions (see Appendix 2.C in Chapter 2) it follows that both distributions are equal: F(x) ' F(x) ' F(x) , say. Thus, for each continuity point x of F, F(x) ' limn64Fn(x) . Note that we have not assumed from the outset that n(t) ' limn64nn(t) is a characteristic function, but only that this pointwise limit exists and is continuous in zero. Consequently, the univariate version of the if part of Theorem 6.22 can be restated more generally as follows. (6.84)
(6.85)
Theorem 6.C.1: Let Xn be a sequence of random variables with corresponding characteristic functions n(t) . If (t) ' limn64n(t) exists for all t 0 and (t) is continuous in t = 0, i.e., limt60(t) ' 1 , then:
257 (a) (b) (t) is a characteristic function itself. Xn 6d X, where X is a random variable with characteristic function (t) .
This result carries over to the multivariate case, but the proof is rather complicated, and therefore omitted. See Section 29 in Billingsley (1986).
Endnotes 1. Recall that n(Xn % 1)/2 = 'j'1(Y j % 1)/2 has a Binomial (n,1/2) distribution, so that the distribution function Fn(x) of Xn is
n
where [z] denotes the largest integer # z, and the sum 'k'0 is zero if m < 0.
m
Fn(x) ' P[Xn # x] ' P[n(Xn%1)/2 # n(x % 1)/2] ' 'k'0
min(n,[n(x % 1)/2]) n k
(1/2)n ,
2. The Yjs have been generated as Yj = 2.I(Uj > 0.5) !1, where the Uj s are random drawings from the uniform [0,1] distribution, and I(.) is the indicator function. 3. 4. 5. 6. Recall that open subsets of a Euclidean space are Borel sets. See Appendix II. Thus M is continuous in y on a little neighborhood of c. Let am, m $1, be a sequence of non-negative numbers such that 'm'1am = K < 4. Then
4 n&1 4 4 n&1 4 4 4
K ' 'm'1am ' limn64'm'1am % limn64'm'nam ' K % limn64'm'nam . Thus, limn64'm'nam = 0. 7. Note that ^0(1,4)N is an uncountable union and may therefore not be a null set. Note that |n(t)| # 1.
'm'1am is monotonic non-decreasing in n $ 2, with limit limn64'm'1am ' 'm'1am ' K , hence
n&1
Therefore, we need to confine the union to all rational " > 1, which is countable. 8.
258
Chapter 7 Dependent Laws of Large Numbers and Central Limit Theorems
In Chapter 6 I have focused on convergence of sums of i.i.d. random variables, in particular the law of large numbers and the central limit theorem. However, macroeconomic and financial data are time series data, for which the independence assumption does not apply. Therefore, in this chapter I will generalize the weak law of large numbers and the central limit theorem to certain classes of time series.
7.1.
Stationarity and the Wold decomposition In Chapter 3 I have introduced the concept of strict stationarity, which for convenience
will be restated here:
Definition 7.1: A time series process Xt is said to be strictly stationary if for arbitrary integers m1 < m2 <....< mn the joint distribution of Xt&m ,......,Xt&m does not depend on the time index t.
1 n
A weaker version of stationarity is covariance stationarity, which requires that the first and second moments of any set Xt&m ,......,Xt&m of time series variables do not depend on the
1 n
time index t.
259 Definition 7.2: A time series process Xt 0 k is covariance stationary (or weakly stationary) if E[||Xt||2] < 4, and for all integers t and m, E[Xt] ' and E[(Xt&)(Xt&m&)T] = (m) do not depend on the time index t.
Clearly, a strictly stationary time series process Xt is covariance stationary if E[||Xt||2] < 4. For zero-mean covariance stationary processes the famous Wold (1938) decomposition theorem holds. This theorem is the basis for linear time series analysis and forecasting, in particular the Box-Jenkins (1979) methodology, and vector autoregression innovation response analysis. See Sims (1980, 1982, 1986) and Bernanke (1986) for the latter.
Theorem 7.1: (Wold decomposition) Let Xt 0 be a zero-mean covariance stationary process. Then we can write Xt ' 'j'0jUt&j % Wt , where 0 ' 1, 'j'0j < 4 , the Uts are zero-mean
4 4 2
covariance stationary and uncorrelated random variables, and Wt is a deterministic process, i.e., there exist coefficients j such that P[Wt ' 'j'1jWt&1] ' 1 . Moreover, Ut = Xt & 'j'1jXt&1
4 4
and E[Ut%mWt] ' 0 for all integers m and t .
Intuitive proof: The exact proof employs Hilbert space theory, and will therefore be given in the Appendix to this chapter. However, the intuition behind the Wold decomposition is not too difficult. It is possible to find a sequence j , j = 1,2,3,....., of real numbers such that
4
E[(Xt & 'j'1jXt&j)2] is minimal. The random variable
' '4 X j'1jXt&j t is then called the linear projection of Xt on Xt!j, j $ 1. Denoting it follows from the first-order condition ME[(Xt & 'j'1jXt&j)2] / Mj ' 0 that
4
260 (7.1)
Ut ' Xt & 'j'1jXt&j

4
(7.2)
E[UtXt&m] ' 0 for m ' 1,2,3,..... Note that (7.2) and (7.3) imply E[Ut] ' 0 , E[UtUt&m] ' 0 for m ' 1,2,3,..... Moreover, note that by (7.2) and (7.3), E[Xt ] ' E[(Ut % 'j'1jXt&j)2] ' E[Ut ] % E[('j'1jXt&j)2] ,
2 4 2 4
(7.3)
(7.4)
so that by the covariance stationarity of Xt, E[Ut ] ' u # E[Xt ] and

2 2 2 2] ' E[('4 E[X j'1jXt&j) ] ' X # E[X t ] t 2 2 2
(7.5)
(7.6)
for all t. Thus it follows from (7.4) and (7.5) that Ut is a zero-mean covariance stationary time series process itself. Next, substitute Xt&1 ' Ut&1 % 'j'1jXt&1&j in (7.1). Then (7.1) becomes
4 4 4 ' (U % '4 X j'1jXt&1&j) % 'j'2jXt&j ' 1Ut&1 % 'j'2(j%1j&1)Xt&j t 1 t&1
' 1Ut&1 %
2 (2%1)Xt&2 4
4 'j'3(j%1j&1)Xt&j .
(7.7)
Now replace Xt&2 in (7.7) by Ut&2 % 'j'1jXt&2&j . Then (7.7) becomes:
261
4 4 ' U % ( %2 X 1)(Ut&2 % 'j'1jXt&2&j) % 'j'3(j%1j&1)Xt&j t 1 t&1 2
' 1Ut&1 % (2%1)Ut&2 % 'j'3[(2%1)j&2%(j%1j&1)]Xt&j

2 4 2
' 1Ut&1 % (2%1)Ut&2 % [(2%1)1%(3%12)]Xt&3 % 'j'4[(2%1)j&2%(j%1j&1)]Xt&j .

2 2 4 2
Repeating this substitution m times yields an expression of the type

4 ' 'm X j'1jUt&j % 'j'm%1m,jXt&j , t
(7.8)
say. It follows now from (7.3), (7.4), (7.5) and (7.8) that
m 2 4 2 2 ] ' 2 E[X u'j'1j % E[('j'm%1m,jXt&j) ] . Hence, letting m 6 4, we have t 4 2 4 2 2 2 ] ' 2 E[X u'j'1j % limm64E[('j'm%1m,jXt&j) ] ' X < 4. t
Therefore, we can write Xt as

4 2 'j'0j
Xt ' 'j'0jUt&j % Wt ,
4
where 0 ' 1 and satisfies
< 4 , with Wt '
4 plimm64'j'm%1m,jXt&j
(7.9) a remainder term which
E[Ut%mWt] ' 0 for all integers m and t . Finally, observe from (7.2) and (7.9) that Ut & Wt & 'j'1jWt&j ' (Xt&Wt) & 'j'1j(Xt&j&Wt&j)
4 4
(7.10)
'
4 4 'j'0j Ut&j&'m'1mUt&j&m
' Ut %
4 'j'1jUt&j ,
(7.11) say .
It follows now straightforwardly from (7.4), (7.5) and (7.10) that j ' 0 for all j $ 1 , hence Wt ' 'j'1jWt&j
4
(7.12)
with probability 1. Q.E.D.
262 Theorem 7.1 carries over to vector-valued covariance stationary processes:
Theorem 7.2: (Multivariate Wold decomposition) Let Xt 0 k be a zero-mean covariance stationary process. Then we can write Xt ' 'j'0AjUt&j % Wt , where A0 ' Ik, 'j'0AjAj is
4 4 T
finite, the Uts are zero-mean covariance stationary and uncorrelated random vectors, i.e., that P[Wt ' 'j'1BjWt&1] ' 1 . Moreover, Ut = Xt & 'j'1BjXt&1 and
4 4
E[UtUt&m] ' O for m $ 1, and Wt is a deterministic process, i.e., there exist matrices Bj such
E[Ut%mWt ] ' O for all integers m and t .
Although the process Wt is deterministic, in the sense that it is perfectly predictable from its past values, it still may be a random process. If so, let W ' (Wt , Wt&1 , Wt&2 , .......) be the F!algebra generated by Wt!m for m $ 0. Then all Wt s are measurable W for arbitrary natural numbers m, hence all Wt s are measurable W ' _t'0W . However, it follows from (7.2) and
&4 4 &t t&m t
(7.9) that each Wt can be constructed from Xt!j for j $ 0, hence X ' (Xt , Xt&1 , Xt&2 , .......) e W , and consequently, all Wt s are measurable X ' _t'0X . This implies that Wt ' E[Wt|X ] . See Chapter 3.
&4 4 &t &4 &4 t t
The F!algebra X represents the information contained in the remote past of Xt. Therefore, X is called the remote F!algebra, and the events therein are called the remote events. If X is the trivial F!algebra { ,i}, so that the remote past of Xt is uninformative, then E[Wt|X ] = E[Wt] , hence Wt = 0. However, the same result holds if all the remote events have either probability zero or one, as is easy to verify from the definition of conditional expectations with respect to a F!algebra. This condition follows automatically from
&4 &4 &4
263 Kolmogorovs zero-one law if the Xts are independent (see Theorem 7.5 below) , but for dependent processes this is not guaranteed. Nevertheless, for economic time series this is not too farfetched an assumption, as in reality they always start from scratch somewhere in the far past, say five hundred years ago for US time series.
Definition 7.3: A time series process Xt has a vanishing memory if the events in the remote F!algebra X ' _t'0(X&t , X&t&1 , X&t&2 , .......) have either probability zero or one.
&4 4
Thus, under the conditions of Theorems 7.1 and 7.2 and the additional assumption that the covariance stationary time series process involved has a vanishing memory, the deterministic term Wt in the Wold decomposition is zero or is a zero vector, respectively.
7.2.
Weak laws of large numbers for stationary processes I will show now that covariance stationary time series processes with a vanishing memory
obey a weak law of large numbers, and then specialize this result to strictly stationary processes. Let Xt 0 be a covariance stationary process, i.e., for all t, E[Xt] = , var[Xt] ' 2, and cov(Xt , Xt&m) ' (m) . If Xt has a vanishing memory then by Theorem 7.1 there exist uncorrelated random variables Ut 0 with zero expectations and common finite variance u such that Xt & ' 'm'0mUt&m , where 'm'0m < 4. Then
4 4 2 4 4 2
Since 'm'0m < 4, it follows that limk64'm'km ' 0. Hence it follows from (7.13) and
4 2 4 2
(k) ' E 'm'0m%kUt&m 'm'0mUt&m .
(7.13)
Schwarz inequality that
|(k)| # u 'm'km 'm'0m 6 0 as k 6 4

2 4 2 4 2
264 (7.14)
Consequently, var (1/n)'t'1Xt ' 2/n % 2(1/n 2)'t'1 'm'1(m) ' 2/n % 2(1/n 2)'m'1(n&m)(m)
n n&1 n&t n&1
# /n %
2
n 2(1/n)'m'1|(m)|
(7.15) 6 0 as n 6 4 .
Using Chebishevs inequality, it follows now from (7.15) that:
Theorem 7.3: If Xt is a covariance stationary time series process with vanishing memory then plimn64(1/n)'t'1Xt ' E[X1] .
n
This results requires that the second moment of Xt is finite. However, this condition can be relaxed by assuming strict stationarity:
Theorem 7.4: If Xt is a strictly stationary time series process with vanishing memory, and E[|X1|] < 4, then plimn64(1/n)'t'1Xt ' E[X1] .
n
Proof: Assume first that P[Xt $ 0] ' 1. For any positive real number M, Xt I(Xt # M) is a covariance stationary process with vanishing memory, hence by Theorem 7.3, plimn64(1/n)'t'1 Xt I(Xt # M) & E[X1 I(X1 # M)] ' 0
n
(7.16)
Next, observe that
265 |(1/n)'t'1(Xt & E[X1])| # |(1/n)'t'1(Xt I(Xt # M) & E[X1 I(X1 # M)])|
n n
n |(1/n)'t'1(Xt I(Xt
(7.17) > M) & E[X1 I(X1 > M)])|
Since for nonnegative random variables Y and Z, P[Y%Z > g] # P[Y > g/2] % P[Z > g/2] , it follows from (7.17) that for arbitrary g > 0, P[|(1/n)'t'1(Xt & E[X1])| > g]
n
# P[|(1/n)'t'1(Xt I(Xt # M) & E[X1 I(X1 # M)])| > g/2]

n
(7.18)
% P[|(1/n)'t'1(Xt I(Xt > M) & E[X1 I(X1 > M)])| > g/2].
n
For an arbitrary * 0 (0,1) we can choose M so large that E[X1 I(X1 > M)] < g/8 . Hence, using Chebishevs inequality for first moments, the last probability in (7.18) can be bounded by */2: P[|(1/n)'t'1(Xt I(Xt > M) & E[X1 I(X1 > M)])| > g/2] # 4E[X1 I(X1 > M)]/g < /2.
n
(7.19)
Moreover, it follows from (7.16) that there exists a natural number n0(g,) such that
n
P[|(1/n)'t'1(Xt I(Xt # M) & E[X1 I(X1 # M)])| > g/2] < /2 if n $ n0(g,) .
(7.20)
Combining (7.18), (7.19) and (7.20), the theorem follows for the case P[Xt $ 0] ' 1. The general case follows easily from Xt ' max(0,Xt) & max(0,&Xt) and Slutskys theorem. Q.E.D. Most stochastic dynamic macroeconomic models assume that the model variables are driven by independent random shocks, so that the model variables involved are functions of these independent random shocks and their past. These random shock are said to form a base for the model variables involved:
266 Definition 7.4: A time series process Ut is a base for a time series process Xt if for each t, Xt is measurable &4 = (Ut , Ut&1 , Ut&2 , ....).
t
If Xt has an independent base then it has a vanishing memory, due to Kolmogorovs zero-one law:
Theorem 7.5: (Kolmogorovs zero-one law). Let Xt be a sequence of independent random variables or vectors, and let &4 ' (Xt , Xt&1 , Xt&2 , ....). Then the sets in the remote F!algebra &4 ' _t'1&4 have either probability zero or one.
4 t t
Proof: Denote by t
t&1
t%k
the F!algebra generated by Xt , .... , Xt%k . Moreover, denote by

t%k
t&m the F!algebra generated by Xt&1 , ... , Xt&m . Each set A1 in t
takes the form
for some Borel set B1 0 k%1 . Similarly, each set A2 in ^m'1t&m takes the form
4 t&1
A1 ' { 0 : (Xt() , .... , Xt%k())T 0 B1}
A2 ' { 0 : (Xt&1() , .... , Xt&m())T 0 B2} for some m $ 1 and some Borel set B2 0 m. Clearly, A1 and A2 are independent.
t&1 4 t&1
F!algebra containing ^m'1t&m. Note that ^m'1t&m may not be a F!algebra itself, but it is
4 t&1 4 t&1
I will now show that the same holds for sets A2 in &4 ' (^m'1t&m) , the smallest
easy to verify that it is an algebra, because t&m d t&m&1. For a given set C in t
4 t&1
t&1
t&1
t%k
with
probability measure on the algebra ^m'1t&m , which has a unique extension to the smallest
4 t&1 4 t&1
positive probability, and for all sets A in ^m'1t&m we have P(A|C) ' P(A) . Thus P(@|C) is a F!algebra containing ^m'1t&m . See Chapter 1. Consequently, P(A|C) ' P(A) is true for all
sets A in &4 . Moreover, if C has probability zero then P(A_C) # P(C) ' 0 ' P(A)P(C) .
t&1
267
Thus for all sets C in t
t%k
therefore A 0 &4 , hence P(A_C) ' P(A)P(C) . By a similar argument as before it can be
t&k&1
C 0 ^k'1t&k . Then for some k, C is a set in t&k , and A is a set in &4 for all m, and
4 t t m
Next, let A 0 _t&4 , where the intersection is taken over all integers t, and let
t&1
and all sets A in &4 , P(A_C) ' P(A)P(C) .

t&1
shown that P(A_C) ' P(A)P(C) for all sets A 0 _t&4 and C 0 (^k'1t&k) . But
t&1 4 t t&1 4 t
&4 ' _t&4 d (^k'1t&k) , so that we may choose C = A. Thus for all sets A 0 _t&4 ,
t&1
P(A) ' P(A)2 , which implies that P(A) is either zero or one. Q.E.D.
7.3.
Mixing conditions Inspection of the proof of Theorem 7.5 reveals that the independence assumption can be
relaxed. We only need independence of an arbitrary set A in &4 and an arbitrary set C in t&k ' (Xt , Xt&1 , Xt&2 , ... , Xt&k) for k $ 1. A sufficient condition for this is that the process Xt is "!mixing or n!mixing:
t
Definition 7.5: Denote &4 ' (Xt , Xt&1 , Xt&2 , ... ) , t ' (Xt , Xt%1 , Xt%2 , ... ) , and let (m) ' sup n(m) ' sup sup sup |P(A_B) & P(A).P(B)|, |P(A|B) & P(A)|.
t A 0 4 , B 0 t&m t &4
t A 0 4 , B 0 t&m t &4
If limm64(m) ' 0 then the time series process Xt involved is said to be "!mixing, and if limm64n(m) ' 0 then Xt is said to be n!mixing.
268 Note that in the "!mixing case sup

A 0
t t&k ,
|P(A_B) & P(A).P(B)| # limsup sup

m64
sup
|P(A_B) & P(A).P(B)|
B 0 &4
t A 0 4 , B 0 t&k&m t&k &4
' limsup (m) ' 0 ,

m64
hence the sets A 0 t&k , B 0 &4 are independent. Moreover, note that (m) # n(m) , so that n!mixing implies "!mixing. Thus the latter is the weaker condition, which is sufficient for a zero-one law:
Theorem 7.6: Theorem 7.5 carries over for "!mixing processes.
Therefore, another version of the weak law of large numbers is:
Theorem 7.7: If Xt is a strictly stationary time series process with an "!mixing base, and E[|X1|] < 4, then plimn64(1/n)'t'1Xt ' E[X1] .
n
7.4. 7.4.1
Uniform weak laws of large numbers Random functions depending on finite-dimensional random vectors Due to Theorem 7.7, all the convergence in probability results in Chapter 6 for i.i.d.
random variables or vectors carry over to strictly stationary time series processes with an "!mixing base. In particular, the uniform weak law of large numbers can now be restated as follows.
269 Theorem 7.8(a): (UWLLN). Let Xt a strictly stationary k-variate time series process with an "!mixing base, and let 0 be non-random vectors in a compact subset d m . Moreover, let g(x,2) be a Borel measurable function on k such that for each x, g(x,2) is a continuous function on 1. Finally, assume that E[sup0*g(Xj , )*] < 4. Then plimn64sup0*(1/n)'j'1g(Xj , ) & E[g(X1 , )]* ' 0 .
n
Theorem 7.8(a) can be proved along the lines as in the proof of the uniform weak law of large numbers for the i.i.d. case in Appendix 6.A of Chapter 6, simply by replacing the reference to the weak law of large numbers for i.i.d random variables by a reference to Theorem 7.7.
7.4.2
Random functions depending on infinite-dimensional random vectors In time series econometrics we quite often have to deal with random functions that
depend on a countable infinite sequence of random variables or vectors. As an example, consider the time series process: Y t ' 0Yt&1 % Xt , with Xt ' Vt & 0Vt&1 , (7.21)
where the Vts are i.i.d. with zero expectation and finite variance F2, and the parameters involved satisfy |$0| < 1 and |(0| < 1. The part Y t ' 0Yt&1 % Xt is an Auto-Regression of order 1, denoted by AR(1), and the part Xt ' Vt & 0Vt&1 (7.23) (7.22)
is a Moving Average process or order 1, denoted by MA(1). Therefore, model (7.21) is called an ARMA(1,1) model. See Box and Jenkins (1976). The condition |$0| < 1 is necessary for the strict
270 stationarity of Yt because then by backwards substitution of (7.21) we can write model (7.21) as Y t ' 'j'00(Vt&j & 0Vt&1&j) ' Vt % (0 & 0)'j'10 Vt&j .
4 j 4 j&1
(7.24)
This is the Wold decomposition of Yt . The MA(1) model (7.23) can be written as an AR(1) model in Vt : Vt ' 0Vt&1 % Ut . If |(0| < 1 then by backwards substitution of (7.25) we can write (7.23) as Xt ' &'j'10Xt&j % Vt.
4 j
(7.25)
(7.26)
Substituting Xt ' Y t & 0Yt&1 in (7.26), the ARMA(1,1) model (7.21) can now be written as an infinite-order AR model: Y t ' 0Yt&1 & 'j'10(Yt&j & 0Yt&1&j) % Vt ' (0 & 0)'j'10 Yt&j % Vt.
4 j 4 j&1
(7.27)
Note that if 0 ' 0 then (7.27) and (7.24) reduce to Y t ' Vt , so that then there is no way to identify the parameters. Thus, we need to assume that 0 0 . Moreover, observe from (7.24) that Yt is strictly stationary, with an independent (hence "-mixing) base. There are different ways to estimate the parameters 0 , 0 in model (7.21) on the basis of observations on Yt for t = 0,1,....,n only. If we assume that the Vts are normally distributed we can use maximum likelihood. See Chapter 8. But it is also possible to estimate the model by nonlinear least squares (NLLS). If we would observe all the Yts for t < n then the nonlinear least squares estimator of 0 ' (0 , 0)T is
2 ' argmin (1/n)'n t'1(Y t & f t()) , 0
(7.28)
where f t() ' ( & )'j'1 j&1Yt&j , with ' ( , )T ,

4
(7.29)
271 and ' [&1%g , 1&g][&1%g , 1&g] , g 0 (0,1) , (7.30)
say, where g is a small number. If we only observe the Yts for t = 0,1,....,n, which is the usual case, then we still can use NLLS by setting the Yts for t < 0 to zero. This yields the feasible NLLS estimator
2 ' argmin (1/n)'n t'1(Y t & f t()) , 0
(7.31)
where
t ft() ' ( & )'j'1 j&1Yt&j .
(7.32)
For proving the consistency of (7.31) we need to show first that

n plimn64 sup0 * (1/n)'t'1 (Y t & ft())2 & (Y t & f t())2] * ' 0
(7.33)
(Exercise), and plimn64 sup0 * (1/n)'t'1 (Y t & ft())2 & E[(Y1 & f1())2] * ' 0 .
n
(7.34)
(Exercise) However, the random functions gt() ' (Y t & ft())2 depend on infinite-dimensional random vectors (Y t , Yt&1 , Yt&2 , Yt&2 , .......)T , so that Theorem 7.8(a) is not applicable to (7.34). Therefore, we need to generalize Theorem 7.8(a) in order to prove (7.34):
Theorem 7.8(b): (UWLLN). Let t ' (Vt , Vt&1 , Vt&2 , ....), where Vt is a time series process with an "!mixing base. Let gt() be a sequence of random function on a compact subset 1 of a Euclidean space. Denote for ( 0 and * $ 0, N(() ' { 0 : ||&(|| # } . If for each ( 0 and each * $ 0, (a) (b) sup0N ( )gt() and inf0N ( )gt() are measurable t and strictly stationary,
( (
E[sup0N ( )gt()] < 4 and E[inf0N ( )gt()] > &4 ,

( (
272 (c) then plimn64sup0|(1/n)'t'1gt() & E[g1()]| ' 0.

n
lim90E[sup0N ( )gt()] ' lim90E[inf0N ( )g t()] ' E[gt(()] ,

( (
Theorem 7.8(b) can also be proved easily along the lines of the proof of the uniform weak law of large numbers in Appendix 6.A of Chapter 6. Note that it is possible to strengthen the (uniform) weak laws of large numbers to corresponding strong laws or large numbers by imposing conditions on the speed of convergence to zero of "(m). See McLeish (1975). It is not too hard (but rather tedious) to verify that the conditions of Theorem 7.8(b) apply to the random functions gt() = (Y t & f t())2 with Yt defined by (7.21) and ft() by (7.29).
7.4.3
Consistency of M-estimators Further conditions for the consistency of M-estimators are stated in the next theorem,
which is a straightforward generalization of a corresponding result in Chapter 6 for the i.i.d. case:
Theorem 7.9: Let the conditions of Theorem 7.8(b) hold, and let 0 ' argmax0E[g1()], ' argmin (1/n)'n 0. Similarly, if 0 ' argmin0E[g1()], t'1gt() , and for * > 0, 0 ' . inf0\N ( )E[g1()] > E[g1(0)], then plimn64 0
0 0
' argmax (1/n)'n t'1gt() . If for * > 0, sup0\N ( )E[g1()] < E[g1(0)] then plimn64 = 0
Again, it is not too hard (but also rather tedious) to verify that the conditions of Theorem 7.9 apply to (7.28), with Yt defined by (7.21) and ft() by (7.29). Thus the feasible NLLS estimator
273 (7.31) is consistent..
7.5. 7.5.1
Dependent central limit theorems Introduction Similarly to the conditions for asymptotic normality of M-estimators in the i.i.d. case (see
Chapter 6), the crucial condition for asymptotic normality of the NLLS estimator (7.28) is that 1 j Vt Mft(0) / M0 6d N2[0 , B] ,
n T
n t'1 where
(7.35)
B ' E V1 Mf1(0) / M0 Mf1(0) / M0 . It follows from (7.24) and (7.29) that ft(0) ' (0 & 0)'j'10 Vt&j ,
4 j&1
(7.36)
(7.37)
which is measurable t&1 ' (Vt&1 , Vt&2 , Vt&3 , ....), and so is 'j'1 0 % (0 & 0)(j&1) 0 Vt&j
4 j&2
T Mf t(0)/M0
'
&'j'10 Vt&j
4 j&1
(7.38)
Therefore, it follows from the law of iterated expectations (see Chapter 3) that B ' 2E Mf1(0) / M0 Mf1(0) / M0 '
4 T
&'j'1 0 % (0 & 0)(j&1) 0

4
'j'1 0 % (0 & 0)(j&1) 20

4
2(j&2) 2(j&2)
&'j'1 0 % (0 & 0)(j&1) 0

4
2(j&2)
(7.39)
'j'10
4
2(j&1)
and
274 P E[Vt ( Mft(0) / M0 ) | t&1] ' 0 ' 1.

T T
(7.40)
The result (7.40) makes Vt ( Mf t(0) / M0 ) a bivariate martingale difference process, and for an arbitrary nonrandom 0 2 , 0 , the process Ut ' Vt T( Mft(0) / M0 ) is then a univariate martingale difference process:
T
Definition 7.4: Let Ut be a time series process defined on a common probability space {S,,P}, and let t be a sequence of sub- F-algebra of . If for each t, (a) (b) (c) (d) Ut is measurable t, t-1 d t, E[|Ut|] < 4, P(E[Ut|t-1] = 0) = 1,
then {Ut, t } is called a martingale difference process.
If condition (d) is replace by P(E[Ut|t&1] ' Ut&1) ' 1 then {Ut, t } is called a martingale. In that case Ut ' Ut & Ut&1 ' Ut & E[Ut|t&1] satisfies P(E[Ut|t&1] ' 0) ' 1 . This is the reason for calling the process in Definition 7.4 a martingale difference process. Thus, what we need for proving (7.35) is a martingale difference central limit theorem.
7.5.2
A generic central limit theorem In this section I will explain McLeish (1974) central limit theorems for dependent random
variables, with a specialization to stationary martingale difference processes. The following approximation of exp(i.x) plays a key role in proving central limit
275 theorems for dependent random variables.
Lemma 7.1. For x 0 with |x| < 1, exp(i.x) ' (1%i.x)exp(&x 2/2 % r(x)) , where |r(x)| # |x|3.
Proof: It follows from the definition of the complex logarithm and the series expansion of log(1+i.x) for |x| < 1 (see Appendix III) that log(1%i.x) ' i.x % x 2/2 % 'k'3(&1)k&1i kx k/k % i.m. ' i.x % x 2 / 2 & r(x) % i.m.,
4
where r(x) ' &'k'3(&1)k&1i kx k/k . Taking the exp of both sides of the equation for log(1+i.x)
4
yields exp(i.x) ' (1%i.x)exp(&x 2/2 % r(x)) . In order to prove the inequality |r(x)| # |x|3 , observe that r(x) ' &'k'3(&1)k&1i kx k/k ' x 3'k'0(&1)k i k%1x k/(k%3)
4 4
' x 3'k'0(&1)2k i 2k%1x 2k/(2k%3) % x 3'k'0(&1)2k%1 i 2k%2x 2k%1/(2k%4)

4 4
' x 3'k'0(&1)kx 2k%1/(2k%4) % i.x 3'k'0 (&1)kx 2k/(2k%3)

4 4
'
4 'k'0(&1)kx 2k%4/(2k%4)
4 i.'k'0 (&1)kx 2k%3/(2k%3)
(7.41)
'
m 1%y 2
0
y3
dy % i.
m 1%y 2
0
y2
dy
where the last equality follows from d 4 x3 4 4 'k'0(&1)kx 2k%4/(2k%4) ' 'k'0(&1)kx 2k%3 ' x 3'k'0(&x 2)k ' dx 1%x 2 for |x| < 1, and similarly
(7.42)
276 d 4 x2 . 'k'0 (&1)kx 2k%3/(2k%3) ' dx 1%x 2 The theorem now follows from the easy inequalities
3 /0 y dy/0 # y 3dy ' 1 |x|4 # |x|3 / 2 00m 00 2 m 4 00 0 1%y 00 0 x |x|
(7.43)
and
2 /0 y dy/0 # y 2dy ' 1 |x|3 # |x|3 / 2 00m 00 2 m 3 00 0 1%y 00 0 x |x|
which hold for |x| < 1. Q.E.D. The result of Lemma 7.1 plays a key-role in the proof of the following generic central limit theorem:
Lemma 7.2: Let Xt, t = 1,2,...,n,..., be a sequence of random variables satisfying the following four conditions: plimn64max1#t#n|Xt|/ n ' 0 ,
n 2
(7.44) (7.45) (7.46)
limn64E (t'1(1%i..Xt/ n) ' 1, 0 ,

n
plimn64(1/n)'t'1Xt ' 2 0 (0,4) ,
and supn$1E (t'1(1%2Xt /n) < 4, 0 .

n 2
(7.47)
Then
277 1 n t'1
2 j Xt 6d N(0, ) . n
(7.48)
Proof: Without loss of generality we may assume that F2 = 1, because if not we may replace Xt by Xt /F. It follows from the first part of Lemma 7.1 that
n n)'t'1Xt
exp i(1/
' k 1%iXt/ n exp &(2 / 2)(1/n)'t'1Xt exp 't'1r Xt / n

n n 2 n t'1
(7.49)
Condition (7.45) implies that plimn64exp &(2 / 2)(1/n)'t'1Xt ' exp(&2 / 2) .

n 2
(7.50)
Moreover, it follows from (7.44), (7.45) and the inequality |r(x)| # |x|3 for |x| <1 that
n /'t'1r
Xt / n I |Xt / n| < 1 / # # ||3 max1#t#n|Xt| n
||3 n n
3 j |Xj| I |Xt / n | < 1 n t'1
n 2 (1/n)'t'1Xt
(7.51) 6p 0 ,
Next, observe that /'t'1r Xt / n I |Xt / n| $ 1 / # 't'1/r Xt / n / I |Xt / n| $ 1

n n
# I ||.max1#t#n|Xt|/ n| $ 1 The latter and condition (7.44) imply that
n 't'1/r
X t / n /
(7.52)
P 't'1r Xt / n I |Xt / n| $ 1 ' 0 $ P ||.max1#t#n|Xt|/ n < 1 6 1 .

n
(7.53)
Therefore, it follows from (7.44) and (7.45) that
278 plimn64exp 't'1r Xt / n ' 1 .

n
(7.54)
Thus we can write exp i(1/ n)'t'1Xt ' k 1%iXt/ n exp(&2 / 2) % k 1%iXt/ n Zn()
n n n t'1 t'1
(7.55)
where Zn() ' exp(&2 / 2) & exp &(2 / 2)(1/n)'t'1Xt exp 't'1r Xt / n 6p 0
n 2 n
(7.56)
Since |Zn()| # 2 with probability 1 because |exp(&x 2/2 % r(x))| # 1 , it follows from (7.56) and the dominated convergence theorem that limn64E[|Zn()|2] ' 0 . .w and |z| ' z z ) that Moreover, condition (7.47) implies (using z w ' z supn$1 E /(t'1(1%iXt / n)/ ' supn$1E (t'1(1%iXt / n)(1&iXt / n)
n 2 n
(7.57)
(7.58)
' supn$1E
n 2 (t'1(1%2Xt /n)
(7.59)
< 4.
Therefore, it follows from the Cauchy-Schwarz inequality that /0 limn64 E Zn() k 1%iXt / n /0 # 00 00 t'1
n
limn64E[|Zn()|2] supn$1E [ (t'1(1%2Xt /n)]

n 2
(7.60)
' 0 Finally, it follows now from (7.46), (7.55) and (7.60) that limn64E[exp(i(1/ n)'t'1Xt)] ' exp(&2 / 2) .
n
(7.61)
Since the right-hand side of (7.61) is the characteristic function of the N(0,1) distribution, the
279 theorem follows for the case F2 = 1. Q.E.D. Lemma 7.2 is the basis for various central limit theorems for dependent processes. See for example Davidsons (1994) textbook. In the next section I will specialize Lemma 7.2 to martingale difference processes.
7.5.3
Martingale difference central limit theorems Note that Lemma 7.2 carries over if we replace the Xts by a double array Xn,t, t =1,2,...,n,
n = 1,2,3,..... In particular, let Yn,1 ' X1 , Yn,t ' Xt I (1/n)'k'1Xk # 2%1 for t $2.
t&1 2
(7.62)
Then by condition (7.45), P[Yn,t Xt for some t # n] # P[(1/n)'t'1Xt > 2%1] 6 0

n 2
(7.63)
hence (7.48) holds if 1

2 j Yn,t 6d N(0, ) . n
n t'1
(7.64)
Therefore, it suffices to verify the conditions of Lemma 7.2 for (7.62). First, it follows straightforwardly from (7.63) that condition (7.45) implies plimn64(1/n)'t'1Yn,t ' 2 .
n 2 2
(7.65)
Moreover, if Xt is strictly stationary with an "!mixing base, and E[X1 ] ' 2 0 (0,4) then it follows from Theorem 7.7 that (7.45) holds, and so does (7.65).
280 Next, let us have a closer look at condition (7.44). It is not hard to verify that for arbitrary g > 0, P[max1#t#n|Xt|/ n > g] ' P[(1/n)'t'1Xt I(|Xt|/ n > g) > g2]
n 2
(7.66)
hence, (7.44) is equivalent to the condition that for arbitrary g > 0, (1/n)'t'1Xt I(|Xt| > g n) 6p 0 .
n 2
(7.67)
Note that (7.67) is true if Xt is strictly stationary, because then E[(1/n)'t'1Xt I(|Xt| > g n)] ' E[X1 I(|X1| > g n)] 6 0
n 2 2
(7.68)
Now consider condition (7.47) for the Yn,ts. Observe that

n 2 (t'1(1%2Yn,t / n) n (t'1 t&1 2 (1/n)'k'1Xk
'
1%
2 Xt I
# %1 / n ' k 1%2Xt / n ,
2 2 t'1
Jn
(7.69)
where Jn ' 1 % j I (1/n)'k'1Xk # 2%1 .

n t&1 2 t'2
(7.70)
Hence
n 2 (t'1(1%2Yn,t / n) J n&1
ln
' j ln 1%2Xt / n % ln 1%2XJ n / n

2 2 t'1
J n&1
(7.71)
1 2 2 2 # 2 j Xt % ln 1%2XJ n / n # (2%1)2% ln 1%2XJ n / n n t'1 where the last inequality follows (7.70). Therefore,
281 supE (t'1(1%2Yn,t / n) # exp((2%1)2) 1%2supE[XJ n] / n

n 2 2 n$1 n$1
# exp(( %1) ) 1%
2 2
n 2 supn$1 (1/n)'t'1E[Xt ]
(7.72) .
Thus (7.72) is finite if supn$1(1/n)'t'1E[Xt ] < 4 ,

n 2
(7.73)
which in its turn is true if Xt is covariance stationary. Finally, it follows from the law of iterated expectations that for a martingale difference process Xt, E (t'1(1%iXt / n) ' E (t'1(1%iE[Xt|t&1] / n) ' 1, 0 ,
n n
(7.74)
and therefore also E (t'1(1%iYn,t / n) ' E (t'1(1%iE[Yn,t|t&1] / n) ' 1, 0

n n
(7.75)
We can now specialize Lemma 7.2 for martingale difference processes:
Theorem 7.10: Let Xt 0 be a martingale difference process satisfying the following three conditions: (a) (b) (c) (1/n)'t'1Xt 6p 2 0 (0,4) ;
n 2 n
For arbitrary g > 0, (1/n)'t'1Xt I(|Xt| > g n) 6p 0 ;

2
Then (1/ n) 't'1Xt 6d N(0,2) .

n
supn$1(1/n)'t'1E[Xt ] < 4 .
n 2
282 Moreover, it is not hard to verify that the conditions of Theorem 7.10 hold if the martingale difference process Xt is strictly stationary with an "!mixing base, and E[X1 ] ' 2 0 (0,4):
2
Theorem 7.11: Let Xt, 0 be a strictly stationary martingale difference process with an "!mixing base, satisfying E[X1 ] ' 2 0 (0,4) . Then (1/ n) 't'1Xt 6d N(0,2) .
2 n
7.6. 1.
Exercises Let U and V be independent standard normal random variables, and let for all integers t
and some nonrandom number 8 0 (0,B), Xt ' U.cos(t) % V.sin(t). Prove that Xt is covariance stationary and deterministic. 2. Show that the process Xt in problem 1 does not have a vanishing memory, but that
n
nevertheless plimn64(1/n)'t'1Xt ' 0 . 3. remote F!algebra &4 ' _t'0(X&t , X&t&1 , X&t&2 , .......) have either probability zero or one. Show
4
Let Xt be a time series process satisfying E[|Xt|] < 4 , and suppose that the events in the
that P(E[Xt|&4] ' E[Xt]) ' 1 . 4. 5. Prove (7.33) . Prove (7.34) by verifying the conditions on Theorem 7.8(b) for gt() = (Y t & f t())2 ,
with Yt defined by (7.21) and ft() by (7.29). 6. Verify the conditions of Theorem 7.9 for gt() = (Y t & f t())2 , with Yt defined by (7.21)
and ft() by (7.29). 7. Prove (7.57).
283 8. Prove (7.66).
Appendix
7.A. Hilbert spaces
7.A.1 Introduction Loosely speaking, a Hilbert space is a space of elements for which similar properties hold as for Euclidean spaces. We have seen in Appendix I that the Euclidean space n is a special case of a vector space, i.e., a space of elements endowed with two arithmetic operations: addition, denoted by "+", and scalar multiplication, denoted by a dot. In particular, a space V is a vector space if for all x, y and z in V, and all scalars c, c1 and c2, (a) (b) (c) (d) (e) (f) (g) (h) x + y = y + x; x + (y + z) = (x + y) + z; There is a unique zero vector 0 in V such that x + 0 = x; For each x there exists a unique vector !x in V such that x + (!x) = 0; 1.x = x; (c1c2).x = c1.(c2.x); c.(x + y) = c.x + c.y; (c1 + c2).x = c1.x + c2.x. Scalars are real or complex numbers. If the scalar multiplication rules are confined to real
284 numbers, the vector space V is a real vector space. In the sequel I will only consider real vector spaces. The inner product of two vectors x and y in n is defined by xTy. Denoting <x,y> = xTy, it is trivial that this inner product obeys the rules in the more general definition of inner product:
Definition 7.A.1: An inner product on a real vector space V is a real function <x,y> on VV such that for all x, y, z in V and all c in , (1) (2) (3) (4) <x,y> = <y,x> <cx,y> = c<x,y> <x+y,z> = <x,z> + <y,z> <x,x> > 0 when x 0.
A vector space endowed with an inner product is called an inner product space. Thus, n is an inner product space. In n the norm of a vector x is defined by ||x|| ' x Tx . Therefore, the norm on a real inner product space is defined similarly as ||x|| ' < x , x > . Moreover, in n the distance between two vectors x and y is defined by ||x&y|| ' (x&y)T(x&y) . Therefore, the distance between two vectors x and y in a real inner product space is defined similarly as ||x&y|| ' < x&y , x&y > . The latter is called a metric. An inner product space with associated norm and metric is called a pre-Hilbert space. The reason for the "pre" is that still one crucial property of n is missing, namely that every Cauchy sequence in n has a limit in n .
285 Definition 7.A.2: A sequence of elements xn of a inner product space with associated norm and metric is called a Cauchy sequence if for every g > 0 there exists an n0 such that for all k,m $ n0, ||xk!xm|| < g.
Theorem 7.A.1: Every Cauchy sequence in R, R < 4 , has a limit in the space involved.
Proof: Consider first the case . Let x ' limsupn64xn , where xn is a Cauchy sequence. I will show first that x < 4 . There exists a subsequence nk such that x ' limk64xn . Note that xn is also a Cauchy
k k
sequence. For arbitrary g > 0 there exists an index k0 such that |xn & xn | < g if k,m $ k0.
k m
Keeping k fixed and letting m 6 4 it follows that |xn & x| < g, hence x < 4 . Similarly,
k
x ' liminfn64xn > &4 . Now we can find an index k0 and sub-sequences nk and nm such that for k,m $ k0, |xn & x| < g, |xn & x| < g, and |xn & xn | < g, hence |x & x| < 3g. Since g is
k m k m
arbitrary, we must have x ' x ' limn64xn . Applying this argument to each component of a vector-valued Cauchy sequence the result for the case R follows. Q.E.D. In order for an inner product space to be a Hilbert space we have to require that the result in Theorem 7.A1 carries over to the inner product space involved:
Definition 7.A.3: A Hilbert space H is a vector space endowed with an inner product and associated norm and metric, such that every Cauchy sequence in H has a limit in H.
286 7.A.2 A Hilbert space of random variables Let U0 be the vector space of zero-mean random variables with finite second moments defined on a common probability space {,,P}, endowed with the inner product <X,Y> = E[X.Y], norm ||X|| ' E[X 2] and metric ||X-Y||.
Theorem 7.A.2: The space U0 defined above is a Hilbert space.
Proof: In order to show that U0 is a Hilbert space, we need to show that every Cauchy sequence Xn ,n $ 1, has a limit in U0. Since by Chebishevs inequality, P[|Xn&Xm| > g] # E[(Xn&Xm)2]/g2 ' ||Xn&Xm||2/g2 6 0 as n,m 6 4
for every g > 0, it follows that |Xn&Xm| 6p 0 as n,m 6 4. In Appendix 6.B of Chapter 6 we have seen that convergence in probability implies convergence a.s. along a subsequence. Therefore there exists a subsequence nk such that |Xn &Xn | 6 0 a.s. as n,m 6 4. The latter implies that
k m
there exists a null set N such that for every 0 \N , Xn () is a Cauchy sequence in , hence
k
limk64Xn () ' X() exists for every 0 \N . Now for every fixed m,
k
(Xn &Xm)2 6 (X&Xm)2 a.s. as k 6 4.

k
By Fatous lemma (see below) and the Cauchy property the latter implies that ||X&Xm||2 ' E[(X&Xm)2] # liminfk64E[(Xn &Xm)2] 6 0 as m 6 4 .
k
Moreover, it is easy to verify that E[X] ' 0 and E[X 2] < 4 . Thus, every Cauchy sequence in U0 has a limit in U0 , hence U0 is a Hilbert space. Q.E.D.
287 Lemma 7.A.1: (Fatous lemma). Let Xn ,n $ 1, be a sequence of non-negative random variables. Then E[liminfn64Xn] # liminfn64E[Xn] .
Proof: Put X ' liminfn64Xn and let n be a simple function satisfying 0 # n(x) # x. Moreover, put Y n ' min(n(X) , Xn). Then Y n 6p n(X) because for arbitrary g > 0, P[|Y n & n(X)| > g] ' P[Xn < n(X)&g] # P[Xn < X&g] 6 0 .
Since E[n(X)] < 4 because n is a simple function, and Y n # n(X) , it follows from Y n 6p n(X) and the dominated convergence theorem that E[n(X)] ' limn64E[Y n] ' liminfn64E[Y n] # liminfn64E[Xn] . (7.76)
Taking the supremum over all simple functions n satisfying 0 # n(x) # x it follows now from (7.76) and the definition of E[X] that E[X] # liminfn64E[Xn] . Q.E.D.
7.A.3 Projections Similarly to the Hilbert space n , two elements x and y in a Hilbert space H are said to be orthogonal if <x,y> = 0, and orthonormal is in addition ||x|| = 1 and ||y|| = 1. Thus, in the Hilbert space U0 two random variables are orthogonal if they are uncorrelated.
Definition 7.A.4: A linear manifold of a real Hilbert space H is a non-empty subset M of H such that for each pair x, y in M and all real numbers " and $, .x % .y 0 M . The closure M of M is called a subspace of H. The subspace spanned by a subset C of H is the closure of the
288 intersection of all linear manifolds containing C.
In particular, if S is the subspace spanned by a countable infinite sequence x1 , x2 , x3 , ..... of

4
vectors in H then each vector x in S takes the form x ' 'n cn.xn , where the coefficients cn are
such that ||x|| < 4. It is not hard to verify that a subspace of a Hilbert space is a Hilbert space itself.
Definition 7.A.5: The projection of an element y in a Hilbert space H on a subspace S of H is an element x of S such that ||y&x|| ' minz0S ||y&z||.
For example, if S is a subspace spanned by vectors x1 , ... , xk in H and y 0 H\S then the projection of y on S is a vector x ' c1.x1 % ... % ck.xk 0 S where the coefficients cj are chosen such that ||y & c1.x1 & ... & ck.xk|| is minimal. Of course, if y 0 S then the projection of y on S is y itself. Projections always exist and are unique:
Theorem 7.A.3: (Projection theorem) If S is a subspace of a Hilbert space H and y is a vector in H then there exists a unique vector x in S such that ||y!x|| = minz0S ||y&z||. Moreover, the residual vector u = y!x is orthogonal to any z in S.
Proof: Let y 0 H\S and infz0S ||y&z|| ' . By the definition of infimum it is possible to select vectors xn in S such that ||y&xn|| # % 1/n . The existence of the projection x of y on S
289 then follows by showing that xn is a Cauchy sequence, as follows. Observe that ||xn&xm||2 ' ||(xn&y)&(xm&y)||2 ' ||xn&y||2 % ||xm&y||2 & 2<xn&y,xm&y>
and 4||(xn%xm)/2&y||2 ' ||(xn&y)%(xm&y)||2 ' ||xn&y||2 % ||xm&y||2 % 2<xn&y,xm&y>.
Adding these two equations up yields ||xn&xm||2 ' 2||xn&y||2 % 2||xm&y||2 & 4||(xn%xm)/2&y||2 (7.77)
Since (xn%xm)/2 0 S it follows that ||(xn%xm)/2&y||2 $ 2 , hence it follows from (7.77) that ||xn&xm||2 # 2||xn&y||2 % 2||xm&y||2 & 42 # 4/n % 1/n 2 % 4/m % 1/m 2 .
Thus xn is a Cauchy sequence in S, and since S is a Hilbert space itself, xn has a limit x in S. As to the orthogonality of u = y!x with any vector z in S, note that for every real number c and every z in S, x+c.z is a vector in S, so that 2 # ||y&x&c.z||2 ' ||u&c.z||2 ' ||y&x||2% ||c.z||2&2<u,c.z> ' 2%c 2||z||2&2c<u,z>. (7.78)
Minimizing the right-hand side of (7.78) to c yields the solution c0 ' <u,z>/||z||2, and substituting this solution in (7.78) yields the inequality (<u,z>)2/||z||2 # 0 . Thus <u,z> = 0. Finally, suppose that there exists another vector p in S such that ||y&p|| ' . Then y!p is orthogonal to any vector z in S: <y&p , z> ' 0 . But x!p is a vector in S, so that <y&p , x&p> ' 0 and <y&x , x&p> ' 0 , hence 0 ' <y&p , x&p> & <y&x , x&p> ' <x&p , x&p> ' ||x&p||2 . Thus,
290 p ' x . Q.E.D.
7.A.5 Proof of the Wold decomposition Let Xt be a zero-mean covariance stationary process, and denote E[Xt ] ' 2 . Then the Xts are members of the Hilbert space U0 defined in Section 7.A.2. Let S&4 be the subspace
t&1 be the projection of Xt on S& spanned by Xt!j, j $1, and let X 4 . Then Ut ' X t & Xt is t t&1 2
orthogonal to all Xt!j, j $1, i.e., E[Ut Xt&j] = 0 for j $1. Since Ut&j 0 S&4 for j $1, the Ut s are also orthogonal to each other: E[Ut Ut&j] = 0 for j $1. takes the form X ' '4 Note that in general X j'1t,jXt&j , where the coefficients $t,j are t t
2
t&1
such that ||Y t||2 ' E[Y t ] < 4 . However, since Xt is covariance stationary the coefficients $t,j do not depend on the time index t, because they are the solutions of the normal equations (m) ' E[XtXt&m] ' 'j'1jE[Xt&jXt&m] ' 'j'1j(|j&m|) , m ' 1,2,3,.......
4 4
' '4 Thus the projections X j'1jXt&j are covariance stationary, and so are the Ut s because t ||2 ' ||U ||2 % ||X ||2 % 2 ' ||U ||2 % ||X ||2 ' E[Ut2] % E[X 2] , 2 ' ||Xt||2 ' ||Ut % X t t t t t t t t
2 2
so that E[Ut ] ' u # 2 .
Next, let Zt,m ' 'j'1jUt&j , where j ' <Xt , Ut> ' E[XtUt&j]. Then
m
||Xt&Zt,m||2 ' ||Xt&'j'1jUt&j||2 ' E[Xt ] & 2'j'1jE[XtUt&j] % 'i'1'j'1ijE[UiUj]

m 2 m m m
' E[Xt ] & 'j'1j $ 0 ,

2 m 2
for all m $ 1, hence 'j'1j < 4 . The latter implies that 'j'mj 6 0 for m 6 4 , so that for
4 2 4 2
291 fixed t, Zt,m is a Cauchy sequence in S&4 , and Xt&Zt,m is a Cauchy sequence in S&4 . Consequently, Zt ' 'j'1jUt&j 0 S&4 and Wt ' Xt ! 'j'1jUt&j 0 S&4 exist.
4 t&1 4 t t&m t&1 t
As to the latter, it follows easily from (7.8) that Wt 0 S&4 for every m, hence Wt 0 _&4<t<4 S&4 .
t
(7.79)
Consequently, E[Ut%mWt] ' 0 for all integers t and m. Moreover, it follows from (7.79) that the projection of Wt on any S&4 is Wt itself, hence Wt is perfectly predictable from any set {Xt&j , j $ 1} of past values of Xt, as well as from any set {Wt&j , j $ 1} of past values of Wt.
t&m
292
Chapter 8 Maximum Likelihood Theory
8.1.
Introduction Consider a random sample Z1,...,Zn from a k-variate distribution with density f(z*0) ,
where 0 0 d m is an unknown parameter vector, with 1 a given parameter space. As is well known, due to the independence of the Zj's, the joint density function of the random vector Z ' (Z1 , .... , Z n )T is the product of the marginal densities: (j'1f(zj*0) . The likelihood function
T T n
in this case is defined as this joint density, with the non-random arguments zj replaced by the corresponding random vectors Zj, and 0 by : () = (n L j'1f(Z j*) . n
(8.1)
= argmax L The maximum likelihood (ML) estimator of 0 is now 0 n() , or equivalently () , = argmax ln L 0 n
(8.2)
where "argmax" stands for the argument for which the function involved takes its maximum value. The ML estimation method is motivated by the fact that in this case ( ))] . ())] # E[ln(L E[ln(L n n 0 (8.3)
To see this, note that ln(u) = u!1 for u = 1, and ln(u) < u!1 for 0 1. Therefore,
293 taking u ' f(Zj|)/f(Z j|0) it follows that for all 2, ln(f(Zj|)/f(Z j|0)) # f(Zj|)/f(Z j|0) ! 1, and taking expectations it follows now that E[ln(f(Z j|)/f(Zj|0))] # E[f(Z j|)/f(Zj|0)] & 1 ' f(z*) f(z*0)dz & 1 ' f(z*)dz & 1 # 0 m f(z*0) m{z0k: f(z|0)>0} k
Summing up, (8.3) follows. This argument reveals that neither the independence assumption of the data Z = (Z1 , .... , Z n )T nor the absolute continuity assumption are necessary for (8.3). The only thing that matters is that ()/L ( )] # 1 E[L n n 0 (8.4)
T T
for all 2 0 1 and n $ 1. Moreover, if the support of Zj is not affected by the parameters in 20, i.e., if in the above case the set {z 0 m: f(z|) > 0} is the same for all 2 0 1, then the inequality in (8.4) becomes an equality: ()/L ( )] ' 1 E[L n n 0 (8.5)
for all 2 0 1 and n $ 1. Equality (8.5) is the most common case in econometrics . In order to show that absolute continuity is not essential for (8.3), suppose that the Zjs are independent and identically discrete distributed with support =, i.e., for all z 0 , the probability model involved. Of course, f(z|) should be specified such that 'z0f(z|) ' 1 for all 2 0 1. For example, suppose that the Zjs are independent Poisson (20) distributed, so P[Zj ' z] > 0 and 'z0P[Zj ' z] = 1. Moreover, let now f(z|0) ' P[Zj ' z], where f(z|) is
294 that f(z|) ' e &z/z! and = = {0,1,2,.....}. Then the likelihood function involved also takes the form (8.1), and E[f(Zj|)/f(Z j|0)] ' j
z0
f(z*) f(z*0) ' j f(z*) ' 1, f(z*0) z0
hence (8.5) holds in this case as well, and therefore so does (8.3). In this and the previous case the likelihood function takes the form of a product. However, also in the dependent case we can write the likelihood function as a product. For example, let Z = (Z1 , .... , Z n )T be absolutely continuously distributed with joint density fn(z n,......,z1|0), where the Zjs are no longer independent. It is always possible to decompose a joint density as a product of conditional densities and an initial marginal density. In particular, denoting for t $ 2, ft(z t|zt&1,...,z1,) ' ft(z t,...,z1|)/ft&1(zt&1,...,z1|), we can write fn(z n,...,z1|) ' f1(z1|)(t'2 ft(z t|zt&1,...,z1,) .
n T T
Therefore, the likelihood function in this case can be written as () ' f (Z ,...,Z |) ' f (Z |)(n L t'2 f t(Z t|Zt&1,...,Z1,) . n n n 1 1 1 (8.6)
It is easy to verify that in this case (8.5) holds also, and therefore so does (8.3). Moreover, it follows straightforwardly from (8.6) and the above argument that in the time series case involved
295 /0 Z ,...,Z # 1 1 0 t&1 Lt(0)/Lt&1(0) 000 ()/L () L t t&1
P E
' 1 for t ' 2,3,...,n,
(8.7)
hence ()/L ()) & ln(L ( )/L ( )) Z ,...,Z # 0 P E ln(L t t&1 t 0 t&1 0 / t&1 1 Of course, these results hold in the independent case as well.
' 1 for t ' 2,3,...,n.
(8.8)
8.2.
Likelihood functions There are quite a few cases in econometrics where the distribution of the data is neither
absolute continuous nor discrete. The Tobit model discussed below is such a case. In these cases we cannot construct a likelihood function in the way I have done here, but still we can define a likelihood function indirectly, using the properties (8.4) and (8.7):
() , n $ 1, of non-negative random functions on a parameter Definition 8.1: A sequence L n space 1 is a sequence of likelihood functions if the following conditions hold: (a) There exists an increasing sequence n , n $ 0, of F-algebras such that for each 2 0 1
() is measurable . and n $ 1, L n n (b) There exists a 0 0 such that for all 2 0 1, P(E [L1()/L1(0) * 0] # 1) ' 1, and /0 0 n&1 # 1 Ln(0)/Ln&1(0) 000 () ()/L L n n&1
for n $ 2, P E
' 1.
(c)
( ) ' L ( )* ] < 1, and for n $ 2, For all 1 2 in , P[L 1 1 1 2 0
296 ( )/L ( ) ' L ( )/L ( )* ] < 1. P[L n 1 n&1 1 n 2 n&1 2 n&1

1
() is constant on 1. Moreover, these conditions The conditions in (c) exclude the case that L n also guarantee that 0 0 is unique:
()/L ( ))] < 0. Theorem 8.1: For all 2 0 1\{ 0 } and n $ 1 , E[ln(L n n 0
()/L ( )) < L ()/L ( ) ! 1 Proof: First, let n = 1. I have already established that ln(L 1 1 0 1 1 0 ()/L ( ) 1 . Thus, denoting Y() ' L ()/L ( ) & ln(L ()/L ( )) & 1 and X() = if L n n 0 n n 0 n n 0 ()/L ( ) we have Y() $ 0, and Y() > 0 if and only if X() 1. Now suppose that L n n 0 P(E[Y()|0] ' 0) ' 1. Then P[Y() ' 0 |0] ' 1 a.s. because Y() $ 0, hence P[X() ' 1|0] ' 1 a.s. Condition (c) in Definition 8.1 now excludes the possibility that ()/L ( ))| ] < 0) ' 1 if and only if . In turn this result 0 , hence P(E[ln(L 1 1 0 0 0 implies that ()/L ( ))] < 0 if . E[ln(L 1 1 0 0 By a similar argument it follows that for n $ 2, ()/L ()) & ln(L ( )/L ( ))] < 0 if . E [ln(L n n&1 n 0 n&1 0 0 The theorem now follows from (8.9) and (8.10). Q.E.D. As we have seen for the case (8.1), if the support {z: f(z|2) > 0}of f(z|) does not depend on 2 then the inequalities in condition (b) becomes equalities, with n ' (Zn,.....,Z1) for n $ 1, and 0 the trivial F-algebra. Therefore, (8.10) (8.9)
297 () , n $ 1, of likelihood functions has invariant support if for Definition 8.2: The sequence L n ()/L ( ) * ] ' 1) ' 1, and for n $ 2, all 0 , P(E [L 1 1 0 0 /0 ' 1 ( )/L ( ) 000 n&1 L n 0 n&1 0 0 ()/L () L n n&1
P E
' 1.
As said before, this is the most common case in econometrics.
8.3. 8.3.1
Examples The uniform distribution Let Zj , j ' 1,...,n , be independent random drawings from the uniform [0,20] distribution,
where 20 > 0. The density function of Zj is f(z|0) ' 0 I(0 # z # 0) , so that the likelihood function involved is: () ' 1 L k I(0 # Z j # ) . n n j'1
n
&1
(8.11)
In this case n ' (Zn,.....,Z1) for n $ 1, and 0 is the trivial F-algebra {S,i}. The conditions (b) in Definition 8.1 now read as ()/L ( )| ] ' E[L ()/L ( )|] ' min(, )/ # 1, E[L 1 1 0 0 1 1 0 0 E ()/L ( )|] ' min(, )/ # 1 for n $ 2. /0 ' E[L 1 1 0 0 ( )/L ( ) 000 n&1 L n 0 n&1 0 0 ()/L () L n n&1
Moreover, the conditions (c) in Definition 8.1 read as
298 P[1 I(0 # Z1 # 1) ' 2 I(0 # Z1 # 2)] ' P(Z1 > max(1 , 2)) < 1 if 1 2.
&1 &1
Hence, Theorem 8.1 applies. Indeed, ()/L ( ))] ' nln( /) % nE[ln(I(0 # Z # ))] & E[ln(I(0 # Z # ))] E[ln(L n n 0 0 1 1 0 &4 if < ,0 ' nln(0/) % nE[ln(I(0 # Z1 # ))] ' nln(0/) < 0 if > 0, 0 if ' 0.
8.3.2
Linear regression with normal errors Let Zj ' (Y j , Xj )T , j ' 1,...,n , be independent random vectors such that Y j ' 0 % 0 Xj % Uj , Uj|Xj - N(0 , 0) ,
2 T 2 T
where the latter means that the conditional distribution of Uj given Xj is a normal N(0 , 0) distribution. The conditional density of Yj given Xj is exp[&(y&0&0 Xj)2/0] 0 2
T 2
f(y|0 , Xj) '
, where 0 ' (0,0 ,0)T .
Next, suppose that the Xjs are absolutely continuously distributed with density g(x). Then the likelihood function is
299
n (j'1f(Y j| , Xj) n (j'1g(Xj)
() ' L n
'
exp[&'j'1(Y j&&TXj)2/2]
n
( 2)
n
(j'1g(Xj)
n
(8.12)
where ' (,T,2)T . However, note that in this case the marginal distribution of Xj does not , because this distribution does not depend on the parameter matter for the ML estimator as function of the data is vector 0 . More precisely, the functional form of the ML estimator invariant to the marginal distributions of the Xjs, although the asymptotic properties of the ML estimator (implicitly) depend on the distributions of the Xjs. Therefore, without loss of generality we may ignore the distribution of the Xjs in (8.12) and work with the conditional likelihood function: c() ' L k f(Y j| , Xj) ' n
n j'1
exp[&'j'1(Y j&&TXj)2/2]
n
( 2)
n
, where ' (,T,2)T .
(8.13)
As to the F-algebras involved, we may take 0 ' ({Xj}j'1) and for n $1, n ' ({Y j}j'1) w0 , where w denotes the operation "take the smallest F-algebra containing the two F-algebras involved". 2 The conditions (b) in Definition 8.1 then read ()/L ( )| ] ' E[f(Y |, X )/f(Y | , X )|X ] ' 1, E[L 1 1 0 0 1 1 1 0 1 1 E /0 n&1 ' E[f(Y n|, Xn)/f(Y n|0, Xn)|Xn] ' 1 for n $ 2. c c 0 L n (0)/Ln&1(0) 000 c () c()/L L n n&1
c c n
Thus, Definition 8.2 applies. Moreover, it is easy to verify that the conditions (c) of Definition 8.1 now read as P[f(Y n|1, Xn) ' f(Y n|2, Xn)|Xn] < 1 if 1 2. This is true, but tedious to verify.
300 8.3.3 Probit and Logit models Again, let Zj ' (Y j , Xj )T , j ' 1,...,n , be independent random vectors, but now Yj takes only two values, 0 and 1, with conditional Bernoulli probabilities P(Y j'1|0 , Xj) ' F(0%0 Xj), P(Y j'0|0 , Xj) ' 1 & F(0%0 Xj),
T T T T
(8.14)
where F is a given distribution function and 0 ' (0,0 )T . For example, let the sample be a survey of households, where Yj indicates home ownership, and Xj is a vector of household characteristics such as marital status, number of children living at home, and income. If F is the logistic distribution function, F(x) ' 1/[1%exp(&x)], then model (8.14) is called the Logit model, and if F is the distribution function of the standard normal distribution then model (8.14) is called the Probit model. In this case the conditional likelihood function is
T T T T c() ' L k [Y jF(% Xj) % (1&Y j)(1 & F(% Xj))] , where ' (, ) . n n j'1
(8.15)
Also in this case the marginal distribution of Xj does not affect the functional form of the ML estimator as function of the data. The F-algebras involved are the same as in the regression case, namely 0 ' ({Xj}j'1) and for n $1, n ' ({Y j}j'1) w0 . Moreover, note that
c c 1 n 4
()/L ( )| ] ' 'y'0[yF(%TX ) % (1&y)(1 & F(%TX ))] ' 1 , E[L 1 1 0 0 1 1
and similarly c()/L c () L n n&1 c( )/L c ( ) L n n&1 0 0

T T /0 n&1 ' '1 y'0[yF(% Xn) % (1&y)(1 & F(% X n))] ' 1 , 00 00
301 hence the conditions (b) of Definition 8.1 and the conditions of Definition 8.2 apply. Also the conditions (c) in Definition 8.1 apply, but again it is rather tedious to verify this.
8.3.4
The Tobit model Let Zj ' (Y j , Xj )T , j ' 1,...,n , be independent random vectors such that Y j ' max(Y j ,0) , where Y j ' 0 % 0 Xj % Uj with Uj|Xj - N(0 , 0).
( ( T 2 T
(8.16)
The random variables Y j are only observed if they are positive. Note that P[Y j'0|Xj] ' P[0 % 0 Xj % Uj # 0|Xj] ' P[Uj > 0 % 0 Xj|Xj] ' 1 & (0 % 0 Xj)/0 , where (x) '
T T T
exp(&u 2/2)/ 2 du . m&4

x
This is a Probit model. Since model (8.16) was proposed by Tobin (1958) and involves a Probit model for the case Yj = 0 it is called the Tobit model. For example, let the sample be a survey of households, where Yj is the amount of money household j spends on tobacco products, and Xj is a vector of household characteristics. But there are households where nobody smokes, so that for these households Yj = 0. In this case the setup of the conditional likelihood function is not as straightforward as in the previous examples, because the conditional distribution of Yj given Xj is neither absolutely continuous nor discrete. Therefore, in this case it is easier to derive the likelihood function indirectly from Definition 8.1, as follows. First note that the conditional distribution function of Yj given Xj and Yj > 0 is
302 P[0 < Y j # y |Xj] P[Y j > 0|Xj]

T
P[Y j # y|Xj , Y j>0] '
'
P[&0 & 0 Xj < Uj # y & 0 & 0 Xj|Xj] P[Y j > 0|Xj]

T
'
(y & 0 & 0 Xj)/0 & (&0 & 0 Xj)/0 (0 % 0 Xj)/0

T
I(y > 0) ,
hence the conditional density function of Yj given Xj and Yj > 0 is n (y & 0 & 0 Xj)/0 0 (0 % 0 Xj)/0
T T
h(y|0 , Xj , Y j>0) '
I(y > 0) , where n(x) '
exp(&x 2/2) 2
Next, observe that for any Borel measurable function g of (Yj,Xj) such that E[|g(Yj,Xj)|] < 4 we have E[g(Y j,Xj)|Xj] ' g(0,Xj)P[Y j ' 0|Xj] % E[g(Y j,Xj)I(Y j > 0)|Xj] ' g(0,Xj)P[Y j ' 0|Xj] % E E[g(Y j,Xj)|Xj ,Y j > 0)|Xj]I(Y j > 0)|Xj ' g(0,Xj) 1 & (0 % 0 Xj)/0 % E ' g(0,Xj) 1 & (0 % 0 Xj)/0 %
T T T
m0
g(y,Xj)h(y|0 , Xj , Y j>0)dy.I(Y j > 0)|Xj g(y,Xj)h(y|0 , Xj , Y j>0)dy. (0 % 0 Xj)/0 1 4 T g(y,Xj)n (y & 0 & 0 Xj)/0 dy . m 0 0
T
(8.17)
m0
' g(0,Xj) 1 & (0 % 0 Xj)/0 % Hence, choosing
g(Y j ,Xj) '
1 & ( % TXj)/ I(Y j ' 0) % &1n (Y j & & TXj)/ I(Y j > 0) 1 & (0 % 0 Xj)/0 I(Y j ' 0) % 0 n Y j & 0 & 0 Xj)/0 I(Y j > 0)
T &1 T
(8.18)
it follows from (8.17) that
303 E[g(Y j,Xj)|Xj] ' 1 & ( % TXj)/ % 1 4 n (y & & TXj)/ dy m0 ' 1 & ( % TXj)/ % 1 & (& & TXj)/ ' 1 .
(8.19)
In view of Definition 8.1, (8.18) and (8.19) suggest to define the conditional likelihood function of the Tobit model as c() ' L k n
n j'1
1 & ( % TXj)/ I(Y j ' 0) % &1n (Y j & & TXj)/ I(Y j > 0) .
The conditions (b) in Definition 8.1 now follow from (8.19), with the F-algebras involved defined similar as in the regression case. Moreover, also the conditions (c) apply. Note that 0n (0 % 0 Xj)/0 (0 %
T 0 Xj)/0 T
E[Y j|Xj ,Y j>0] ' 0 %
T 0 X j
(8.20)
Therefore, if one would estimate a linear regression model using the observations with Yj > 0 only, the OLS estimates will be inconsistent, due to the last term in (8.20).
8.4. 8.4.1
Asymptotic properties of ML estimators Introduction () )] Without the conditions (c) in Definition 8.1 the solution 0 = argmax0 E[ln (L n
may not be unique. For example, if Zj = cos(Xj+20) with the Xj s independent absolutely continuously distributed random variables with common density, then the density function f(z|20) of Zj satisfies f(z|20) = f(z|20+2sB) for all integers s. Therefore, the parameter space 1 has to be chosen small enough to make 20 unique.
304 () )] at ' Also, the first and second-order conditions for a maximum of E[ln (L n 0 may not be satisfied. The latter is for example the case for the likelihood function (8.11): if () )] ' &4 , and if $ then E[ln (L () )] ' &n.ln() , so that the left < 0 then E[ln (L n 0 n () )] in ' is lim (E[ln (L ( ) )] & E[ln (L ( &) )]) / = 4, and derivative of E[ln (L n 0 90 n 0 n 0 ( %) )] & E[ln (L ( ) )]) / = &n/ . Since the first and the right-derivative is lim90(E[ln (L n 0 n 0 0 second-order conditions play a crucial role in deriving the asymptotic normality and efficiency of the ML estimator (see below), the rest of this chapter does not apply to the case (8.11).
8.4.2
First and second-order conditions The following conditions guarantee that the first and second-order conditions for a
maximum hold.
Assumption 8.1: The parameter space 1 is convex, and 20 is an interior point of 1. The () is, with probability 1, twice continuously differentiable in an open likelihood function L n neighborhood 0 of 0 , and for i1 , i2 ' 1,2,3,...,m , () M2 L n / /0 < 4 E sup 00 0 00 00 Mi Mi 00 0 1 20 and () M2ln L n / E sup 00 00 00 Mi Mi 1 2 0 /0 < 4 . 00 00
(8.21)
(8.22)
305
Theorem 8.2: Under Assumption 8.1, ()) Mln (L n MT *' () ) M2ln (L n MMT *' () ) Mln(L n MT *= .
0
= 0 and E
= &Var
Proof: For notational convenience I will prove this theorem for the univariate parameter case m = 1 only. Moreover, I will focus on the case that Z ' (Z1 , .... , Z n )T is a random sample from an absolutely continuous distribution with density f(z|20). Observe that () ) / n = 1 E ln (f(Z *) ) = ln (f(z*) )f(z* )dz , E ln (L j n j 0 m n j'1
n T T
(8.23)
It follows from Taylors theorem that for 0 0 and every * 0 for which % 0 0 there exists a 8(z,*) 0 [0,1] such that ln (f(z*%) ) & ln (f(z*) ) ' dln (f(z*) ) 1 d 2ln (f(z*%(z,)) ) % 2 . d 2 (d(%(z,)))2
(8.24)
Note that by the convexity of , 0 % (z,) 0 . Therefore, it follows from condition (8.22), the definition of a derivative, and the dominated convergence theorem that d dln (f(z*) ) ln (f(z*) )f(z*0)dz ' f(z*0)dz m d m d
(8.25)
Similarly, it follows from condition (8.21) , Taylors theorem and the dominated convergence theorem that
306 df(z*) d d dz ' f(z*) dz ' 1 ' 0. m d d m d Moreover, dln (f(z*) ) df(z*) / d df(z*) f(z*0)dz *' = f(z*0)dz *= ' dz *= 0 0 0 m m f(z*) m d d
(8.26)
(8.27)
The first part of theorem now follows from (8.23) through (8.27). Similarly to (8.25) and (8.26) it follows from the mean value theorem and the conditions (8.21) and (8.22) that ln (f(z*) )f(z*0)dz ' m (d)2 m and m (d)2 d 2f(z*) (d)2 m d f(z*) dz ' 0 . d2 d 2ln (f(z*) ) (d)2 f(z*0)dz
(8.28)
dz '
(8.29)
The second part of the theorem follows now from (8.28), (8.29) and m d 2ln f(z*) (d)2 d 2f(z*) f(z*0) df(z*) / d dz *= & 0 2 m (d) f(z*) m f(z*)
2
f(z*0)dz *' =
0
f(z*0)dz *=
'
m (d)2
d 2f(z*)
dz *= &
0
dln f(z*) /d 2f(z*0)dz *= .

0
The adaptation of the proof to the general case is pretty straightforward and is therefore left as an exercise. Q.E.D. The matrix
307 () )/MT * Hn = Var Mln(L n = .

0
(8.30)
is called the Fisher information matrix. As we have seen in Chapter 5, the inverse of the Fisher information matrix is just the Cramer-Rao lower bound of the variance matrix of an unbiased estimator of 0 .
8.4.3
Generic conditions for consistency and asymptotic normality The ML estimator is a special case of an M-estimator. In Chapter 6 I have derived generic
conditions for consistency and asymptotic normality of M-estimators, which in most cases apply to ML estimators as well. The case (8.11) is one of the exceptions, though. In particular, if
() /L ( )) & E[ln (L () /L ( ))] | ' 0 and Assumption 8.2: plimn64sup0 |ln (L n n 0 n n 0 () /L ( ))] & R( | )| ' 0 , where R( | ) is a continuous function in 0 limn64sup0 |E[ln (L n n 0 0 0 such that for arbitrary * > 0, sup0: ||& ||$ R( | 0) < 0 ,
0
then the ML estimator is consistent:
= . Theorem 8.3: Under Assumption 8.2, plimn64 0
The conditions in Assumption 8.2 need to be verified on a case-by-case basis. In particular, the uniform convergence in probability condition has to be verified from the
308 conditions of the uniform weak law of large numbers. The last condition in Assumption 8.2, i.e., sup0: ||& ||$ R( | 0) < 0 , follows easily from Theorem 8.1 and the continuity of R( | 0) .
0
Some of the conditions for asymptotic normality of the ML estimator are already listed in Assumption 8.1, in particular the convexity of the parameter space 1, and the condition that 20 is an interior point of 1. The other (high-level) conditions are:
Assumption 8.3: For i1 , i2 ' 1,2,3,...,m , plimn64 sup /00 0 00 0 and limn64 sup /00 E 0 00 0 where hi
1 , i2
() / n M2ln L n Mi Mi
1 2
& E
1 2
/0 ' 0 00 00
(8.31)
1 2
% hi
1 , i2
() /00 ' 0, 00 0
1 , i2
(8.32) (0) is non-
() is continuous in 20. Moreover, the mm matrix H with elements hi
singular. Furthermore, ( ) ) / n Mln(L n 0 M0

T
6d Nm[0 , H ] .
(8.33)
Note that the matrix H is just the limit of Hn /n, with Hn the Fisher information matrix (8.30). Condition (8.31) can be verified from the uniform weak law of large numbers. Condition (8.32) is a regularity condition which accommodates data-heterogeneity. In quite a few cases we may take hi
1 , i2
()) /(M M )]. Finally, condition (8.33) can be verified from () = &n &1E[M2ln(L n i i
1 2
309 the central limit theorem.
- ) 6 N [0 , H &1] . Theorem 8.4: Under Assumptions 8.1-8.3, n( 0 d m
Proof: It follows from the mean value theorem (see Appendix II) that for each i 0 0 [0,1] such that {1,...,m} there exists a i () ) / n Mln (L n Mi
2
*' =
() ) / n Mln (L n Mi
*'
(8.34) % ()) / n M ln(L *' % 0 i(&0) MMi - ), n( 0
The first-order condition for (8.2) and the condition that 20 is an interior point of 1 imply ()) / M | ' 0 . plimn64n &1/2Mln(L n i ' (8.35)
( Moreover, the convexity of 1 guarantees that the mean value 0% i &0) is contained in 1. It and the conditions (8.31) and (8.32) that follows now from the consistency of ()) / n M2ln(L n MM1 ' H MMm ! ()) / n M2ln(L n *' %
0 m(&0)
*' % ( & )
0 1 0
6p H .
(8.36)
The condition that H is nonsingular allows us to conclude from (8.36) and Slutskys theorem that plimn64H
&1
' H &1 ,
(8.37)
310 hence it follows from (8.34) and (8.35) that & ) ' &H &1(Mln (L ( ) ) /MT n( 0 ) / n + o p(1) . 0 n 0 (8.38)
Theorem 8.4 follows now from condition (8.33) and the results (8.37) and (8.38). Q.E.D. In the case of a random sample Z1,...,Zn the asymptotic normality condition (8.33) can easily be derived from the central limit theorem for i.i.d. random variables. For example, let again the Zjs be k-variate distributed with density f(z*0) . Then it follows from Theorem 8.2 that under Assumption 8.1,
T ( ))/MT E[Mln(f(Zj*0)) / M0 ] ' n &1E[Mln (L 0] ' 0 n 0
and ( ))/M0 ] ' H , Var[Mln(f(Zj*0)) / M0 ] ' n &1Var[Mln (L n 0 say, so that (8.33) straightforwardly follows from the central limit theorem for i.i.d. random vectors.
T T
8.4.4
Asymptotic normality in the time series case In the time series case (8.6) we have ( ))/M0 Mln(L n 0 n
T
'
j Ut , n t'1
n
(8.39)
where U1 ' Mln( f1(Z1|0))/M0 , Ut ' Mln( f t(Zt|Zt&1,...,Z1,0))/M0 for t $ 2 . The process Ut is a martingale difference process (see Chapter 7): Denoting for t $ 1,
T T
(8.40)
311 t ' (Z1,...,Zt), and letting 0 be the trivial F-algebra {S,i}, it is easy to verify that for t $ 1, E[Ut|t&1] ' 0 a.s. Therefore, condition (8.33) can in principle be derived from the conditions of the martingale difference central limit theorems (Theorems 7.10-7.11) in Chapter 7. Note that even if Zt is a strictly stationary process, the Uts may not be strictly stationary. In that case condition (8.33) can be proved by specializing Theorem 7.10 in Chapter 7. An example where condition (8.33) follows from Theorem 7.11 in Chapter 7 is the AutoRegressive (AR) model of order 1: Zt ' % Zt&1 % gt , where gt is i.i.d. N(0,2) and || < 1. (8.41)
The condition |$| < 1 is necessary for strict stationarity of Zt. Then for t $ 2 the conditional distribution of Zt given t&1 ' (Z1,...,Zt&1) is N( % Zt&1 , 2) , so that, with 0 = ( , , 2)T , (8.40) becomes gt ' 1
2
Ut '
M &(Z t&&Zt&1) / & ln( ) & ln( 2)

2 2 2
M(,, )
2
gtZt&1 (gt /2 & 1)

2
(8.42)
Since the gt s are i.i.d. N(0,2) and gt and Zt&1 are mutually independent it follows that (8.42) is a martingale difference process, not only with respect to t ' (Z1,...,Zt) but also with respect to &4 ' ({Zt&j}j'0) , i.e., E[Ut|&4 ] = 0 a.s.
t 4 t&1
By backwards substitution of (8.41) it follows that Zt ' 'j'0j (%gt&j) so that the
4
marginal distribution of Z1 is N[/(1&) , 2/(1&2)]. However, there is no need to derive U1 in this case, because this term is irrelevant for the asymptotic normality of (8.39). Therefore, the asymptotic normality of (8.39) in this case follows straightforwardly from the stationary martingale difference central limit theorem, with asymptotic variance matrix
312 1& 2 (1&)2 1&2 0 % 2
1 H ' Var(Ut) ' 1

2
0 0 1 22 .
1& 0
8.4.5
Asymptotic efficiency of the ML estimator As said before, the ML estimation approach is a special case of the M-estimation
approach discussed in Chapter 6. However, the position of the ML estimator among the Mestimators is a special one, namely the ML estimator is (under some conditions) asymptotically efficient. In order to explain and prove asymptotic efficiency, let ' argmax (1/n)'n j'1g(Z j,) 0 be an M-estimator of 0 ' argmax0E[g(Z1,)] , (8.44) (8.43)
where again Z1,...,Zn is a random sample from a k-variate absolutely continuous distribution with density f(z*0) , and d m is the parameter space. In Chapter 6 I have set forth conditions such that & ) 6 N [0 , A &1BA &1] , n( 0 d m where (8.45)
313 M2g(Z1 , 0) M0 M0
T
A ' E
'
mk M MT 0 0
M2g(z , 0)
f(z|0)dz
(8.46)
and B ' E Mg(Z1 , 0)/M0 Mg(Z1 , 0)/M0

T
'
Mg(z , 0)/M0 Mg(z , 0)/M0 f(z|0)dz
(8.47)
As will be shown below, the matrix A &1BA &1 & H &1 is positive semi-definite, hence the is "larger" (or at least not smaller) than the asymptotic variance asymptotic variance matrix of . In other words, the ML estimator is an asymptotically matrix H &1 of the ML estimator efficient M-estimator. This proposition can be motivated as follows. Under some regularity conditions, similarly to Assumption 8.1, it follows from the first-order condition for (8.44) that m Mg(z,0)/M0 f(z|0)dz ' k
T
mk
ME[g(z,)]/MT f(z|0)dz|' ' 0

0
(8.48)
Since equality (8.48) does not depend on the value of 20 it follows that for all 2, mk Mg(z,)/MT f(z|)dz ' 0 . (8.49)
Taking derivatives inside and outside the integral (8.49) again yield: mk MMT ' mk MMT M2g(z,) M2g(z,) f(z|)dz % mk mk Mg(z,)/MT Mf(z|)/M dz (8.50) f(z|)dz % Mg(z,)/MT Mln(f(z|))/M f(z|)dz ' O .
Replacing 2 by 20 it follows from (8.46) and (8.50) that Mg(Z1,0) M0

T
Mln(f(Z1|0)) M0
' &A.
(8.51)
Since the two vectors in (8.51) have zero expectation, (8.51) also reads as
314 Mg(Z1,0) Mln(f(Z1|0)) , T T M0 M0 ' &A.
Cov
(8.52)
It follows now from (8.47), (8.52) and Assumption 8.3 that Mg(Z1,0)/M0
T T
Var
Mln(f(Z1|0))/M0
'
B &A &A H
which of course is positive semi-definite, and therefore so is B &A &A H A &1 H

&1
A ,H
&1
&1
' A &1BA &1 & H &1 .
Note that this argument does not hinge on the independence and absolute continuity assumptions made here. We only need that (8.45) holds for some positive definite matrices A and B, and that 'j'1Mg(Zj,0)/M0
n T
1 n
( ))/MT Mln(L 0 n 0
6d N2m
0 0
B &A &A H .
8.5. 8.5.1
Testing parameter restrictions The pseudo t test and the Wald test In view of Theorem 8.2 and Assumption 8.3 the matrix H can be estimated consistently
in (8.53): by the matrix H
315 ())/n M2ln(L n MM

T
' & H
/0 6p H. 00 0'
(8.53)
Denoting by ei the i-th column of the unit matrix Im, it follows now from (8.53), Theorem 8.4 and the results in Chapter 6 that:
T T &1 / ei H ei 6d N(0,1) if Theorem 8.5: (Pseudo t-test) Under Assumptions 8.1-8.3, ti = n ei
e i 0 = 0 .
Thus the null hypothesis H0: ei 0 = 0 , which amounts to the hypothesis that the i-th component of 0 is zero, can now be tested by the pseudo t-value ti in the same way as for Mestimators. Next, consider the partition 1,0 2,0
0 =
, 1,0 0 m&r , 2,0 0 r ,
(8.54)
and suppose that we want to test the null hypothesis 2,0 ' 0 . This hypothesis corresponds to the linear restriction R0 = 0 , where R = (O,Ir). It follows from Theorem 8.4 that under this null hypothesis 6 N (0 , RH R T ). nR d r , H &1 and H &1 conformably to (8.54) as Partitioning
&1
(8.55)
1 2
, H
&1
(1,1) H (1,2) H (2,1) H (2,2) H
, H
&1
(1,1) H (1,2) H (2,1) H (2,2) H
(8.56)
316
= R , H (2,2) = RH &1R T , and H (2,2) = RH &1R T , hence it follows from (8.55) it follows that 2 (2,2) &1/2 n 6 N (0 , I ) and consequently: that H 2 d r r
T H (2,2) &1 6 2 Theorem 8.6: (Wald test) Under Assumptions 8.1-8.3, n 2 2 d r if 2,0 = 0.
8.5.2
The Likelihood Ratio test An alternative to the Wald test is the Likelihood Ratio (LR) test, which is based on the
ratio () max0: '0L n

2
() max0L n
( L n ) ( L n )
where is partitioned conformably to (8.54) as 1 2
and 1 2 1 0
() , = argmaxL n
0: 2'0
(8.57)
is always between zero and one. The intuition behind is the restricted ML estimator. Note that will approach 1 (in probability) as n 64 because then both the LR test is that if 2,0 = 0 then and the restricted ML estimator are consistent . In particular: the unrestricted ML estimator
317 ) 6 2 Theorem 8.7: (LR test) Under Assumptions 8.1-8.3, &2 ln( d r if 2,0 = 0.
Proof: Similarly to (8.38) we have () ) / n Mln (L n M1

T
- ) = &H &1 n( 1,1 1 1,0
*'
+ o p(1) ,
where H 1,1 is the upper-left (m&r)(m&r) block of H:
H '
H1,1 H1,2 H2,1 H2,2
and consequently,
1 H& 1,1 O
- ) = & n( 0
( ) ) / n Mln (L n 0
T M0
+ o p(1).
(8.58)
Subtracting (8.58) from (8.34) and using condition (8.33) yield &1 O H 1,1 O where &1 O H 1,1 O O &1 O H 1,1 O O &1 O H 1,1 O O O ( ) ) / n Mln (L n 0 M0
T
- ) = & H &1 n(
+ op(1) 6d N m(0 , ) ,
(8.59)
&1 = H
H &1 H
&1 = H
(8.60)
The last equality in (8.60) follows straightforwardly from the partition (8.56). Next, it follows from the second-order Taylor expansion around the unrestricted ML
318 that for some estimator 0 [0,1], () Mln L n MT
) = ln L T ( ln( n ) - ln (L n() ) = (-) () ) / n M2ln (L 1 n T + n(-) *= % & ) ( 2 MMT
*= (8.61)
- ) = - 1 n( - )T H n( - ) + o (1) , n( p 2
where the last equality in (8.61) follows from the fact that similarly to (8.36), () ) / n M2ln (L MMT 6p &H . /0 00' % & ) (
(8.62)
Thus we have ) = &1/2 n( - ) 1/2H 1/2 &1/2 n( - ) + o (1) . &2ln( p & ) 6 Nm(0,Im) is distr., and by (8.60) the matrix 1/2H1/2 is Since by (8.59), &1/2 n( idempotent with rank( 1/2H1/2 ) = trace( 1/2H1/2 ) = r, the theorem follows from the results in Chapters 5 and 6. Q.E.D.
T
(8.63)
8.5.3 The Lagrange Multiplier test can also be obtained from the first-order conditions of the The restricted ML estimator ()) - 2 , where 0 r is a vector of Lagrange Lagrange function ( , ) = ln(L n multipliers. These first-order conditions are:
T
319
T T M( , )/M1 *' , ' ' 0, ' Mln L() /M1 *' T T M( , )/M2 *' ' 0, , ' & ' Mln L() /M2 *'
M( , )/M T *' , ' ' 2 ' 0 .
Hence 1 0 n () / n Mln L MT
'
*' .
(8.64)
Again, using the mean value theorem we can expand this expression around the unrestricted ML , which then yields estimator
1 0 n hence TH
(2,2,)
& ) % o (1) , ' &H n( p
(8.65)
'
1 T T &1 0 )H (0 , n
'
& )TH n( & ) % o (1) 6 r . n( p d
(8.66)
Replacing H in this expression by a consistent estimator on the basis of the restricted ML , say: estimator /0 . 00 0'
' & H
())/n M2ln(L n MMT
(8.67)
&1 similarly to (8.56) as and partitioning H
320 (1,1) H (1,2) H (2,1) H (2,2) H
&1
'
(8.68)
we have
2 (2,2) Theorem 8.8: (LM test) Under Assumptions 8.1-8.3, TH / n 6d r if 2,0 = 0.
8.5.4
Which test to use? The Wald, LR and LM tests basically test the same null hypothesis against the same
alternative, so which one should we use? The Wald test employs only the unrestricted ML , so that this test is the most convenient if we have to conduct unrestricted ML estimator , and there are estimation anyhow. The LM test is entirely based on the restricted ML estimator situations where we start with restricted ML estimation, or where restricted ML estimation is much easier to do than unrestricted ML estimation, or even where unrestricted ML estimation is not feasible because without the restriction imposed the model is incompletely specified. Then the LM test is the most convenient test. Both the Wald and the LM tests require the estimation of the matrix H. That may be a problem for complicated models because of the partial derivatives involved. In that case use the LR test. Although I have derived the Wald, LR and LM tests for the special case of a null hypothesis of the type 2,0 ' 0 , the results involved can be modified to general linear hypotheses
321 of the form R0 ' q , where R is a r m matrix of rank r, by reparametrizing the likelihood function, as follows. Specify a (m-r) m matrix R* such that the matrix R( R is nonsingular. Then define new parameters by 1 2 R( R 0 q 0 q
Q '
'
'
&
' Q &
Substituting 0 q
' Q &1 % Q &1
in the likelihood function, the null hypothesis involved is equivalent to 2 ' 0 .
8.6. 1.
Exercises ' argmax L Derive n() for the case (8.11), and show that if Z1,...,Zn is a random
sample then the ML estimator involved is consistent. 2. 3. ' argmax L Derive n() for the case (8.13). Show that the log-likelihood function of the Logit model is unimodal, i.e., the matrix
()]/(MMT) is negative-definite for all 2. M2ln[L n 4. 5. Prove (8.20). Extend the proof of Theorem 8.2 to the multivariate parameter case.
322 6. Let (Y1,X1),...,(Yn,Xn) be a random sample from a bivariate continuous distribution with
conditional density f(y*x , 0) ' x / 0 exp &y . x / 0 if x > 0 and y > 0 ; f(y*x , 0) ' 0 elsewhere ,
where 20 > 0 is an unknown parameter. The marginal density h(x) of Xj is unknown, but we do know that h does not depend on 20, and h(x) = 0 for x # 0. (a) (b) (c) (d) (e) (f) c() . Specify the conditional likelihood function L n of 20. Derive the maximum likelihood estimator is unbiased. Show that is equal to 2 Show that the variance of 0/n. Verify that this variance is equal to the Cramer-Rao lower bound. Derive the test statistic of the LR test of the null hypothesis 20 = 1, in the form for which
2
it has an asymptotic 1 null distribution. (g) (h) (i) Derive the test statistic of the Wald test of the null hypothesis 20 = 1. Derive the test statistic of the LM test of the null hypothesis 20 = 1. Show that under the null hypothesis 20 = 1 the LR test in part (f) has a limiting 1
2
distribution. 7. Let Z1,....,Zn be a random sample from the (nonsingular) Nk[,E] distribution. Determine
the maximum likelihood estimators of and E. 8. In the case where the dependent variable Y is a duration, for example an unemployment
duration spell, the conditional distribution of Y given a vector X of explanatory variables is often modeled by the proportional hazard model
where 8(t) is a positive function on (0,4) such that *0 (t)dt ' 4 , and n is a positive function.
4
P[Y # y|X ' x] ' 1 & exp &n(x)*0 (t)dt , y > 0 ,

y
323 (8.70)
The reason for calling this model a proportional hazard model is the following. Let f(y|x) be the conditional density of Y given X = x, and let G(y|x) ' exp &n(x)*0 (t)dt , y > 0 . The
y
latter function is called the conditional survival function. Then f(y|x)/G(y|x) ' n(x)(y) is called the hazard function, because for a small * > 0, f(y|x)/G(y|x) is approximately the conditional probability (hazard) that Y 0 (y , y%], given that Y > y and X = x. Convenient specifications of 8(t) and n(x) are: (t) ' t &1 , > 0 (Weibull specification) n(x) ' exp( % x)
T
(8.71)
Now consider a random sample of size n of unemployed workers. Each unemployed worker j is interviewed twice. The first time worker j tells the interviewer how long he or she has been unemployed, and reveals his or her vector Xj of characteristics. Call this time Y1,j . A fixed period of length T later the interviewer asks worker j whether he or she is still (uninterruptedly) unemployed, and if not how long it took during this period to find employment for the first time. Call this duration Y2,j. In the latter case the observed unemployment duration is Y j ' Y1,j % Y2,j, but if the worker is still unemployed we only know that Y j > Y1,j % T . The latter is called censoring. Assuming that the Xjs do not change over time, setup the conditional likelihood function for this case, using the specifications (8.70) and (8.71).
324 Endnotes 1. 2. See Chapter 3 for the definition of these conditional probabilities. Recall from Chapter 1 that the union of F-algebras is not necessarily a F-algebra.
325
Appendix I Review of Linear Algebra

I.1. Vectors in a Euclidean space A vector is a set of coordinates which locates a point in a Euclidean space. For example, in the two-dimensional Euclidean space 2 the vector a1 a2 6 4
a '
'
(I.1)
is the point which location in a plane is determined by moving a1 ' 6 units away from the origin along the horizontal axis (axis 1), and then moving a2 ' 4 units away parallel to the vertical axis (axis 2), as displayed in Figure I.1.
Figure I.1: A vector in 2 The distances a1 and a2 are called the components of the vector a involved. An alternative interpretation of the vector a is a force pulling from the origin (the intersection of the two axes). This force is characterized by its direction (the angle of the line in Figure I.1) and its strength (the length of the line piece between point a and the origin). As to the
326 latter, it follows from Pythagoras Theorem that this length is the square root of the sum of the squared distances of point a from the vertical and horizontal axes: a1 %a2 ' and is denoted by 2a2 . More generally, the length of a vector x1 x ' x2 ! xn
n in is defined by 2 2
62%42 ' 3 6 ,
(I.2)
2x2 '
def.
'j'1xj .
n 2
(I.3)
There are two basic operations that apply to vectors in n . The first basic operation is scalar multiplication: c.x1
def.
c.x '
c.x2 ! c.xn
(I.4)
where c 0 is a scalar. Thus, vectors in n are multiplied by a scalar by multiplying each of the components by this scalar. The effect of scalar multiplication is that the point x is moved a factor c along the line through the origin and the original point x. For example, if we multiply the vector a in Figure I.1 by c = 1.5, the effect is the following:
327
Figure I.2: Scalar multiplication
The second operation is addition: Let x be the vector (I.2), and let y1 y ' y2 ! yn Then x1%y1
def.
(I.5)
x % y '
x2%y2 ! xn%yn
(I.6)
Thus, vectors are added by adding up the corresponding components. Of course, this operation is only defined for conformable vectors, i.e., vectors with the same number of components. As an example of the addition operation, let a be the vector (I.1), and let
328 b1 b2 3 7
b ' Then 6 4
'
(I.7)
a % b '
3 7
'
9 11
' c,
(I.8)
say. This result is displayed in Figure I.3 below. We see from Figure I.3 that the origin together with the points a, b and c = a + b form a parallelogram (which is easy to prove). In terms of forces, the combined forces represented by the vectors a and b result in the force represented by the vector c = a + b.
Figure I.3: c = a + b The distance between the vectors a and b in Figure I.3 is 2a & b2 . To see this, observe that the length of the horizontal line piece between the vertical line through b and point a is a1&b1 , and similarly the vertical line piece between b and the horizontal line through a has length b2&a2 . These two line pieces, together with the line piece connecting the points a and b, form a triangle for which Pythagoras Theorem applies: The squared distance between a and b is
329 equal to (a1&b1)2 % (a2&b2)2 ' 2a&b22 . More generally, The distance between the vector x in (I.2) and the vector y in (I.5) is 2x & y2 ' 'j'1(xj & yj)2 .
n
(I.9)
Moreover, it follows from (I.9) and the Law of Cosines1 that: The angle N between the vector x in (I.2) and the vector y in (I.5) satisfies 'j'1xj yj 2x22 % 2y22 & 2x&y22 cos() ' ' . 22x2.2y2 2x2.2y2
n
(I.10)
I.2.
Vector spaces The two basic operations, addition and scalar multiplication, make a Euclidean space n
a special case of a vector space:
Definition I.1: Let V be a set endowed with two operations, the operation "addition", denoted by "+", which maps each pair (x,y) in V V into V, and the operation "scalar multiplication", denoted by a dot (.), which maps each pair (c,x) in V into V. The set V is called a vector space if the addition and multiplication operations involved satisfy the following rules, for all x, y and z in V, and all scalars c, c1 and c2 in : (a) (b) (c) (d) x + y = y + x; x + (y + z) = (x + y) + z; There is a unique zero vector 0 in V such that x + 0 = x; For each x there exists a unique vector !x in V such that x + (!x) = 0;
330 (e) (f) (g) (h) 1.x = x; (c1c2).x = c1.(c2.x); c.(x + y) = c.x + c.y; (c1 + c2).x = c1.x + c2.x.
It is trivial to verify that with addition "+" defined by (I.6) and scalar multiplication c.x defined by (I.4) the Euclidean space n is a vector space. However, the notion of a vector space is much more general. For example, let V be the space of all continuous functions on , with pointwise addition and scalar multiplication defined the same way as for real numbers. Then it is easy to verify that this space is a vector space. Another (but weird) example of a vector space is the space V of positive real numbers endowed with the "addition" operation x + y = x.y and the "scalar multiplication" c.x = xc. In this case the null vector 0 is the number 1, and !x = 1/x.
Definition I.2: A subspace V0 of a vector space V is a non-empty subset of V which satisfies the following two requirements: (a) (b) For any pair x, y in V0, x + y is in V0; For any x in V0 and any scalar c, c.x is in V0.
It is not hard to verify that a subspace of a vector space is a vector space itself, because the rules (a) through (h) in Definition I.1 are inherited from the "host" vector space V. In particular, any subspace contains the null vector 0, as follows from part (b) of Definition I.2 with
331 c = 0. For example, the line through the origin and point a in Figure I.1, extended indefinitely in both directions, is a subspace of 2 . This subspace is said to be spanned by the vector a. More generally,
Definition I.3: Let x1,x2,....,xn be vectors in a vector space V. The space V0 spanned by x1,x2,....,xn is the space of all linear combinations of x1,x2,....,xn , i.e., each y in V0 can be written as y ' 'j'1cjxj for some coefficients cj in .
n
Clearly, this space V0 is a subspace of V. For example, the two vectors a and b in Figure I.3 span the whole Euclidean space 2 , because any vector x in 2 can be written as, x1 x2 6 4 3 7 6c1%3c2 4c1%7c2
x '
' c1
% c2
'
where c1 ' 7 1 2 1 x1 & x2 , c2 ' & x1 % x2 . 30 10 15 5
The same applies to the vectors a, b and c in Figure I.3: They also span the whole Euclidean space 2 . However, in this case any pair of a, b and c does the same, so one of these three vectors is redundant, because each of the vectors a, b and c can already be written as a linear combination of the other two. Such vectors are called linear dependent:
332 Definition I.4: A set of vectors x1,x2,....,xn in a vector space V is linear dependent if one or more of these vectors can be written as a linear combination of the other vectors, and the set is called linear independent if none of them can be written as a linear combination of the other vectors. In particular, x1,x2,....,xn are linear independent if and only if 'j'1cjxj = 0 implies that
n
c1 ' c2 ' ' cn ' 0 .
For example, the vectors a and b in Figure I.3 are linear independent, because if not then there would exists a scalar c such that b = c.a, hence 6 = 3c and 4 = 7c, which is impossible. A set of such linear independent vectors is called a basis for the vector space they span:
Definition I.5: A basis for a vector space is a set of vectors having the following two properties: (a) (b) It is linear independent; The vectors span the vector space involved.
We have seen that each of the subsets {a,b}, {a,c} and {b,c} of the set {a, b, c} of vectors in Figure I.3 is linear independent, and span the vector space 2 . Thus, there are in general many bases for the same vector space, but what they have in common is their number: This number is called the dimension of V.
Definition I.6: The number of vectors that form a basis of a vector space is called the dimension of this vector space.
333 In order to show that this definition is unambiguous, let {x1,x2,....,xn} and {y1,y2,....,ym} be two different bases for the same vector space, and let m = n +1. Each of the yis can be written as a linear combination of x1,x2,....,xn : yi ' 'j'1ci,jxj . If {y1,y2,....,yn+1} is linear independent then
n n%1 n n%1
linear independent we must also have that 'i'1 zici,j ' 0 for j = 1,...,n. The latter is a system of
n%1
'i'1 ziyi ' 'j'1'i'1 zici,jxj = 0 if and only if z1 ' ...... ' zn%1 ' 0 . But since {x1,x2,....,xn} is
n linear equations in n+1 unknown variables zi and therefore has a non-trivial solution, in the sense that there exists a solution z1,...,zn%1 such that least one of the zs is non-zero. Consequently, {y1,y2,....,yn+1} cannot be linear independent. Note that in principle the dimension of a vector space can be infinite. For example, consider the space 4 of all countable infinite sequences x ' (x1 , x2 , x3 , ............) of real numbers, endowed with the addition operation x % y ' (x1 , x2 , x3 , ............) % (y1 , y2 , y3 , ............) ' (x1%y1 , x2%y2 , x3%y3 , ............) and the scalar multiplication operation c.x ' (c.x1 , c.x2 , c.x3 , ............) . Let yi be a countable infinite sequence of zeros, except for the i-th element in this sequence, which is equal to 1. Thus, y1 ' (1 , 0 , 0 , ) , y2 ' (0 , 1 , 0 , ) , etc. Then {y1,y2,y3,...} is a basis for 4 , with dimension 4. Also in this case there are many different bases; for example, another basis for 4 is y1 ' (1 , 0 , 0 , 0 ,) , y2 ' (1 , 1 , 0 , 0 ,) , y3 ' (1 , 1 , 1 , 0 ,) , etc.
334 I.3. Matrices In Figure I.3 the location of point c can be determined by moving nine units away from the origin along the horizontal axis 1, and then moving eleven units away from axis 1 parallel to the vertical axis 2. However, given the vectors a and b an alternative way of determining the location of point c is: Move 2a2 units away from the origin along the line through the origin and point a (the subspace spanned by a), and then move 2b2 units away parallel to the line through the origin and point b (the subspace spanned by b). Moreover, if we take 2a2 as the new distance unit along the subspace spanned by a, and 2b2 as the new distance unit along the subspace spanned by b, then point c can be located by moving one (new) unit away from the origin along the new axis 1 formed by the subspace spanned by a, and then move one (new) unit away from this new axis 1 parallel to the subspace spanned by b (which is now the new axis 2). We may interpret this as moving the point
1 1
to a new location: point c. This is precisely what a matrix
does: moving points to a new location by changing the coordinate system. In particular, the matrix 6 3 4 7
A ' a,b ' moves any point x1 x2
(I.11)
x '
(I.12)
to a new location, by changing the original perpendicular coordinate system to a new coordinate system, where the new axis 1 is the subspace spanned by the first column, a, of the matrix A, with new unit distance the length of a, and the new axis 2 is the subspace spanned by the second
335 column, b, of A, with new unit distance the length of b. Thus, this matrix A moves point x to point 6 4 3 7 6x1%3x2 4x1%7x2
y ' Ax ' x1.a % x2.b ' x1. In general, an m n matrix
% x2.
'
(I.13)
a1,1 a1,n A ' ! " ! (I.14) am,1 am,n moves the point in n corresponding to the vector x in (I.2) to a point in the subspace of m spanned by the columns of A, namely to point 'j'1a1,jxj
n
y ' Ax ' j xj
n j'1
a1,j ! am,j '
y1 ' ! . ym (I.15)
'j'1am,jxj
n
Next, consider the k m matrix b1,1 b1,m B ' ! " ! . (I.16) bk,1 bk,m and let y be given by (I.15). Then b1,1 b1,m By ' B(Ax) ' ! " ! bk,1 bk,m where 'j'1a1,jxj
n
'j'1 's'1b1,sas,j xj
n m
'j'1am,jxj
n
'
'j'1 's'1bk,sas,j xj
n m
' Cx,
(I.17)
336 c1,1 c1,n C ' ! " ! ck,1 ck,n This matrix C is called the product of the matrices B and A, and is denoted by BA. Thus, with A given by (I.14) and B given by (I.16), b1,1 b1,m BA ' ! " ! bk,1 bk,m a1,1 a1,n ! " ! ' am,1 am,n 's'1b1,sas,1 's'1b1,sas,n
m m
with ci,j ' 's'1bi,sas,j .

m
(I.18)
's'1bk,sas,1 's'1bk,sas,n
m m
"
(I.19)
which is a k n matrix. Note that the matrix BA only exists if the number of columns of B is equal to the number of rows of A. Such matrices are called conformable. Moreover, note that if A and B are also conformable, so that AB is defined2, then the commutative law does not hold, i.e., in general AB BA. However, the associative law (AB)C = A(BC) does hold, as is easy to verify. Let A be the m n matrix (I.14), and let now B be another m n matrix: b1,1 b1,n B ' ! " ! . (I.20) bm,1 bm,n As argued before, A maps a point x 0 n to a point y = Ax 0 m, and B maps x to a point z = Bx 0 m. It is easy to verify that y+z = Ax + Bx = (A+B)x = Cx, say, where C = A + B is the m n formed by adding up the corresponding elements of A and B:
337 a1,1 a1,n A % B ' ! " ! % am,1 am,n b1,1 b1,n ! " ! ' bm,1 bm,n a1,1%b1,1 a1,n%b1,n ! " ! . (I.21) am,1%bm,1 am,n%bm,n
Thus, conformable matrices are added up by adding up the corresponding elements. Moreover, for any scalar c we have A(c.x) = c.(Ax) = (c.A)x, where c.A is the matrix formed by multiplying each element of A by the scalar c: a1,1 a1,n c.A ' c. ! " ! ' am,1 am,n c.a1,1 c.a1,n ! " ! . (I.22) c.am,1 c.am,n
Now with addition and scalar multiplication defined in this way, it is easy to verify that all the conditions in Definition I.1 hold for matrices as well, i.e., the set of all m n matrices is a vector space. In particular, the "zero" element involved is the m n matrix with all elements equal to zero: 0 0 Om,n ' ! " ! . 0 0 I.4. The inverse and transpose of a matrix The question I now will address is whether for a given m n matrix A there exists a n m matrix B such that, with y = Ax, By = x. If so, the action of A is undone by B, i.e., B moves y back to the original position x. If m < n there is no way to undo the mapping y = Ax, i.e., there does not exists an n m matrix B such that By = x. To see this, consider the 1 2 matrix A = ( 2,1). Then with x as in (I.23)
338 (I.12), Ax = 2x1 + x2 = y, but if we know y and A, then we only know that x is located on the line x2 = y - 2x1, but there is no way to determine where on this line. If m = n in (I.14), so that the matrix A involved is a square matrix, we can undo the mapping A if the columns3 of the matrix A are linear independent. Take for example the matrix A in (I.11) and the vector y in (I.13), and let 7 30 2 & 15 1 10 1 5
&
B '
(I.24)
Then 7 30 2 & 15 1 10 1 5
&
By '
6x1%3x2 4x1%7x2
'
x1 x2
' x,
(I.25)
so that this matrix B moves the point y = Ax back to x. Such a matrix is called the inverse of A, and is denoted by A &1 . Note that for an invertible n n matrix A, A &1A ' In , where In is the n n unit matrix: 1 0 0 0 0 1 0 0 In ' 0 0 1 0 . ! ! ! " ! 0 0 0 1 Note that a unit matrix is a special case of a diagonal matrix, i.e., a square matrix with all offdiagonal elements equal to zero. We have seen that the inverse of A is a matrix A &1 such that A &1A ' I . 4 But what about (I.26)
339 AA &1 ? Does the order of multiplication matter? The answer is no:
Theorem I.1: If A is invertible, then A A!1 = I, i.e., A is the inverse of A!1,
because it is trivial that
Theorem I.2: If A and B are invertible matrices then (AB)!1 = B!1A!1.
Now let us give a formal proof of our conjecture that:
Theorem I.3: A square matrix is invertible if and only if its columns are linear independent.
Proof: Let A be n n the matrix involved. I will show first that: (a) The columns a1,....,an of A are linear independent if and only if for every b 0 n the
system of n linear equations Ax = b has a unique solution. To see this, suppose that there exists another solution y: Ay = b. Then A(x!y) = 0 and x!y 0, which imply that the columns a1,....,an of A are linear dependent. Similarly, if for every b 0 n the system Ax = b has a unique solution, then the columns a1,....,an of A must be linear independent, because if not then there exists a vector c 0 in n such that Ac = 0, hence if x is a solution of Ax = b then so is x + c. Next, I will show that:
340 (b) A is invertible if and only if for every b 0 n the system of n linear equations Ax = b
has a unique solution. First, if A is invertible then the solution of Ax = b is x = A!1b, which for each b 0 n is unique. Second, let b = ei be the i-th column of the unit matrix In, and let xi be the unique solution of Axi = ei. Then the matrix X with columns x1,...,xn satisfies AX ' A(x1 , ,xn) ' (Ax1 , , Axn) ' (e1 , , en) ' In , hence A is the inverse of X: A = X!1. It follows now from Theorem I.1 that X is the inverse of A: X = A!1. Q.E.D. If the columns of a square matrix A are linear dependent, then Ax maps point x into a lower-dimensional space, namely the subspace spanned by the columns of A. Such a mapping is called a singular mapping, and the corresponding matrix A is therefore called singular. Consequently, a square matrix with linear independent columns is called non-singular. It follows from Theorem I.3 that a non-singularity is equivalent to invertibility, and singularity is equivalent to absence of invertibility. If m > n in (I.14), so that the matrix A involved has more rows than columns, we can also undo the action of A if the columns of the matrix A are linear independent, as follows. First, consider the transpose5 AT of the matrix A in (I.14): a1,1 am,1 AT ' ! " ! , (I.27) a1,n am,n i.e., AT is formed by filling its columns with the elements of the corresponding rows of A. Note that
341 Theorem I.4: (AB)T = BTAT. Moreover, if A and B are square and invertible then (A T )&1 ' (A &1)T , (AB)T
&1
(AB)&1
&1
' B &1A &1 BT

&1
' A &1
T
B &1
' AT
&1
BT
&1
, and similarly ,
' B TA T
' AT
&1
' A &1
B &1 T .
Proof: Exercise. Since a vector can also be interpreted as a matrix with only one column, the transpose operation also applies to vectors. In particular, the transpose of the vector x in (I.2) is: x T ' (x1 , x2 . , xn) , which may be interpreted as a 1n matrix. Now if y = Ax then ATy = ATAx, where ATA is an n n matrix. If ATA is invertible, then (ATA)!1ATy = x, so that then the action of the matrix A is undone by the n m matrix (ATA)!1AT. Thus, it remains to be shown that: (I.28)
Theorem I.5: ATA is invertible if and only if the columns of the matrix A are linear independent.
Proof: Let a1,....,an be the columns of A. Then ATa1,....,ATan are the columns of ATA. Thus, the columns of ATA are linear combinations of the columns of A. Suppose that the columns of ATA are linear dependent. Then there exists coefficients cj not all equal to zero such that c1A Ta1 % % cnA Ta n ' 0 . This equation can be rewritten as A T(c1a1 % % cnan) ' 0. Since a1,....,an are linear independent, we have c1a1 % % cna n 0, hence the columns of AT are linear dependent. However, this is impossible, because of the next theorem. Therefore, if the columns of A are linear independent, then so are the columns of ATA. Thus, the theorem under
342 review follows from Theorem I.3 and Theorem I.6 below.
Theorem I.6: The dimension of the subspace spanned by the columns of a matrix A is equal to the dimension of the subspace spanned by the columns of its transpose AT.
The proof of Theorem I.6 has to be postponed, because we need for it the results in the next sections. In particular, Theorem I.6 follows from Theorems I.11, I.12 and I.13 below.
Definition I.7: The dimension of the subspace spanned by the columns of a matrix A is called the rank of A.
Thus, a square matrix is invertible if and only if its rank equals its size, and if a matrix is invertible then so is its transpose.
I.5.
Elementary matrices and permutation matrices Let A be the m n matrix in (I.14). An elementary m m matrix E is a matrix such
that the effect of EA is that a multiple of one row of A is added to another row of A. For example, let Ei,j(c) be an elementary matrix such that the effect of Ei,j(c)A is that c times row j is added to row i < j:
343 a1,1 ! ai&1,1 Ei,j(c)A ' " a1,n ! ai&1,n
ai,1%caj,1 ai,n%caj,n ai%1,1 ! aj,1 ! am,1 " " ai%1,n ! aj,n ! am,n . (I.29)
Then Ei,j(c)6 is equal to the unit matrix Im (compare (I.26)), except that the zero in the (i,j)'s position is replaced with a nonzero constant c. In particular, if i=1 and j = 2 in (I.29), so that E1,2(c)A adds c times row 2 of A to row 1 of A, then 1 c 0 0 0 1 0 0 E1,2(c) ' 0 0 1 0 . ! ! ! " ! 0 0 0 1 This matrix is a special case of an upper-triangular matrix; that is a square matrix with all the elements below the diagonal equal to zero. Moreover, E2,1(c)A adds c times row 1 of A to row 2 of A: 1 0 0 0 c 1 0 0 E2,1(c) ' 0 0 1 0 , ! ! ! " ! 0 0 0 1 (I.30)
344 which is a special case of a lower-triangular matrix, i.e., a square matrix with all the elements above the diagonal equal to zero. Similarly, if E is an elementary n n matrix, then the effect of AE is that one of the columns of A, times a nonzero constant, is added to another column of A. Thus,
Definition I.8: An elementary matrix is a unit matrix with one off-diagonal element replaced with a nonzero constant.
Note that the columns of an elementary matrix are linear independent, hence an elementary matrix is invertible. The inverse of an elementary matrix is easy to determine: If the effect of EA is that c times row j of A is added to row i of A, then E!1 is an elementary matrix such that the effect of E!1EA is that -c times row j of EA is added to row i of A, so that then E!1EA restores A. For example, the inverse of the elementary matrix (I.30) is: 1 0 0 0 c 1 0 0 E2,1(c)&1 ' 0 0 1 0 ! ! ! " ! 0 0 0 1 We now turn to permutation matrices: '
&1
1 0 0
0 0 0 0 1 0 0 0 1 ' E2,1(&c) .
&c 1 0 0 ! ! ! " !
Definition I.9: An elementary permutation matrix is a unit matrix with two columns or rows swapped. A permutation matrix is a matrix whose columns or rows are permutations of the columns or rows of a unit matrix.
345 In particular, the elementary permutation matrix that is formed by swapping the columns i and j of a unit matrix will be denoted by Pi,j. The effect of an (elementary) permutation matrix on A is that PA swaps two rows, or permutates the rows, of A. Similarly, AP swaps or permutates the columns of A. Whether you swap or permutate columns or rows of a unit matrix does not matter, because the resulting (elementary) permutation matrix is the same. An example of an elementary permutation matrix is 0 1 0 0 1 0 0 0 P1,2 ' 0 0 1 0 . ! ! ! " ! 0 0 0 1 Note that a permutation matrix P can be formed as a product of elementary permutation matrices, say P ' Pi ,j .......Pi ,j . Moreover, note that if an elementary permutation matrix Pi,j is applied to
1 1 k k
itself, i.e., Pi,jPi,j, then the swap is undone, and the result is the unit matrix: Thus, the inverse of an elementary permutation matrix Pi,j is Pi,j itself. This result holds only for elementary permutation matrices, though. In the case of the permutation matrix P ' Pi ,j .......Pi ,j we have
1 1 k k P
&1
' Pi ,j .......Pi ,j . Since elementary permutation matrices are symmetric: Pi,j ' Pi,j , it
k k 1 1
follows that P &1 ' Pi k,j k.......Pi1,j1 ' P T . Moreover, if E is an elementary matrix and Pi,j an elementary permutation matrix then Pi,jE ' EPi,j . Combining these results, it follows:
Theorem I.7: If E is an elementary matrix and P is a permutation matrix, then PE ' EP T . Moreover, P &1 ' P T .
346 I.6. Gaussian elimination of a square matrix, and the Gauss-Jordan iteration for inverting a matrix I.6.1 Gaussian elimination of a square matrix The results in the previous section are the tools we need to derive the following result:
Theorem I.8: Let A be a square matrix. (a) There exists a permutation matrix P, possibly equal to the unit matrix I, a lower-
triangular matrix L with diagonal elements all equal to 1, a diagonal matrix D, and an uppertriangular matrix U with diagonal elements all equal to 1, such that PA = LDU. (b) If A is non-singular and P = I this decomposition is unique, i.e., if A = LDU = L*D*U*,
then L( ' L , D( ' D , and U( ' U .
The proof of part (b) is as follows: LDU = L*D*U* implies L &1L(D( ' DUU(
&1
(I.31)
It is easy to verify that the inverse of a lower triangular matrix is lower triangular, and that the product of lower triangular matrices is lower triangular. Thus the left-hand side of (I.31) is lower triangular. Similarly, the right-hand side of (I.31) is upper triangular. Consequently, the offdiagonal elements in both sides are zero: Both matrices in (I.31) are diagonal. Since D( is diagonal and non-singular, it follows from (I.31) that L &1L( ' DUU( D( is diagonal. Moreover, since the diagonal elements of L &1 and L( are all equal to one, the same applies to L &1L( , i.e., L &1L( ' I , hence L ' L( . Similarly we have U ' U( . Then D ' L &1AU &1 and D( ' L &1AU &1 .
&1 &1
347 Rather than giving a formal proof of part (a) of Theorem I.8, I will demonstrate the result involved by two examples, one for the case that A is non-singular, and one for the case that A is singular. Example 1: A is nonsingular. Let 2 4 2 A ' 1 2 3 &1 1 &1 We are going to multiply A by elementary matrices and elementary permutation matrices such that the end-result will be a upper-triangular matrix. This is called Gaussian elimination. First, add ! times row 1 to row 2 in (I.32). This is equivalent to multiplying A by the elementary matrix E2,1(!). (Compare (I.30), with c = !.). Then 1 0 0 0 0 1 2 4 2 1 2 3 &1 1 &1 ' 2 4 2 0 0 2 &1 1 &1 . (I.33) . (I.32)
E2,1(!)A ' !0.5 1 0
Next, add times row 1 to row 3, which is equivalent to multiplying (I.33) by the elementary matrix E3,1(): 1 E3,1()E2,1(!)A ' 0 0 0 1 0 2 4 2 0 0 2 &1 1 &1 ' 2 4 2 0 0 2 . 0 3 0 (I.34)
0.5 0 1
Now swap the rows 2 and 3 of the right-hand side matrix in (I.34). This is equivalent to multiplying (I.34) by the elementary permutation matrix P2,3 formed by swapping the columns 2
348 and 3 of the unit matrix I3. Then 1 0 0 P2,3E3,1()E2,1(!)A ' 0 0 1 0 1 0 2 0 0 ' 0 3 0 0 0 2 1 2 1 0 1 0 0 0 1
&1
2 4 2 0 0 2 0 3 0
2 4 2 ' 0 3 0 0 0 2 (I.35)
' DU ,
say. Moreover, since P2,3 is an elementary permutation matrix, we have that P2,3 ' P2,3 , hence it follows from Theorem I.7 and (I.35) that P2,3E3,1()E2,1(!)A ' E3,1()P2,3E2,1(!)A ' E3,1()E2,1(!)P2,3A ' DU . Furthermore, observe that 1 0 hence 1 E3,1()E2,1(!)
&1
(I.36)
0 0 0 1
1 0
0 0 1 0
0 0 (I.37)
E3,1()E2,1(!) ' !0.5 1 0
' &0.5 1 0 0.5 0 1
0.5 0 1
0 0
&1
1 '
0 0 ' L, (I.38)
' &0.5 1 0 0.5 0 1
0.5 1 0 &0.5 0 1
say. Combining (I.36) and (I.38), it follows now that P2,3A ' LDU . Example 2: A is singular. Theorem I.8 also holds for singular matrices. The only difference with the non-singular case is that if A is singular then the diagonal matrix D will have zeros on the diagonal. To demonstrate this, let now
349 2 4 2 A ' 1 2 1 &1 1 &1 Since the first and last column of this matrix A are equal, the columns are linear dependent, hence A is singular. Now (I.33) becomes 1 0 (I.34) becomes 1 E3,1()E2,1(!)A ' 0 0 0 1 0 2 4 2 0 0 0 &1 1 &1 ' 2 4 2 0 0 0 , 0 3 0 (I.41) 0 0 0 1 2 4 2 1 2 1 &1 1 &1 ' 2 4 2 0 0 0 &1 1 &1 , (I.40) . (I.39)
E2,1(!)A ' !0.5 1 0
0.5 0 1 and (I.35) becomes 1 0 0 P2,3E3,1()E2,1(!)A ' 0 0 0 0 1 0 2 0 0 ' 0 3 0 0 0 0
2 4 2 0 0 0 0 3 0
2 4 2 ' 0 3 0 0 0 0 (I.42)
1 2 1 0 1 0 0 0 1 ' DU .
The formal proof of part (a) of Theorem I.8 is similar to the argument in these two examples, and is therefore omitted. Note that the result (I.42) demonstrates that:
350 Theorem I.9: The dimension of the subspace spanned by the columns of a square matrix A is equal to the number of non-zero diagonal elements of the matrix D in Theorem I.8.
Example 3: A is symmetric and nonsingular Next, consider the case that A is symmetric, that is, AT = A. For example, let 2 4 2 A ' 4 0 1 2 1 &1 Then E3,2(&3/8)E3,1(&1)E2,1(&2)AE1,2(&2)E1,3(&1)E2,3(&3/8) 2 0 ' 0 &8 0 0 ' D, (I.44) . (I.43)
0 0 &15/8 hence A ' E3,2(&3/8)E3,1(&1)E2,1(&2) &1D E1,2(&2)E1,3(&1)E2,3(&3/8) &1 ' LDL T . (I.45)
Thus, in the symmetric case we can eliminate each pair of non-zero elements opposite of the diagonal jointly by multiplying A from the left by an appropriate elementary matrix and multiplying A from the right by the transpose of the same elementary matrix. Example 4: A is symmetric and singular Although I have demonstrated this result for a non-singular symmetric matrix, it holds for the singular case as well. For example, let now 2 4 2 A ' 4 0 4 . 2 4 2 (I.46)
351
Then 2 0 0 E3,1(&1)E2,1(&2)AE1,2(&2)E1,3(&1) ' 0 &8 0 0 0 0 Example 5: A is symmetric and has a zero in a pivot position If there is a zero in a pivot position7, then we need a row exchange. In that case the result A = LDLT will no longer be valid. For example, let 0 4 2 A ' 4 0 4 . 2 4 2 Then 4 0 4 E3,2(&1)E3,1(&1/2)P1,2A ' 0 4 2 0 0 &2 4 0 0 ' 0 4 0 0 0 &2 1 L ' E3,2(&1)E3,1(&1/2)
&1
' D.
(I.47)
(I.48)
1 0 0 0 0 0 1 0
1 ' DU (I.49) 1
0 1 1/2
' E3,1(1/2)E3,2(1) '
U T.
1/2 1 1 Thus, examples 3, 4 and 5 demonstrate that:
Theorem I.10: If A is symmetric and the Gaussian elimination can be conducted without need for row exchanges, then there exists a lower triangular matrix L with diagonal elements all equal to one, and a diagonal matrix D, such that A = LDLT.
352 I.6.2 The Gauss-Jordan iteration for inverting a matrix The Gaussian elimination of the matrix A in the first example in the previous section suggests that this method can also be used to compute the inverse of A, as follows. Augment the matrix A in (I.32) to a 3 6 matrix, by augmenting the columns of A with the columns of the unit matrix I3: 2 4 2 B ' (A, I3) ' 1 2 3 &1 1 &1 1 0 0 0 1 0 . 0 0 1 (I.50)
Now follow the same procedure as in Example 1, up to (I.35), with A replaced by B. Then (I.35) becomes: P2,3E3,1()E2,1(!)B ' P2,3E3,1()E2,1(!)A , P2,3E3,1()E2,1(!) 2 4 2 ' 0 3 0 0 0 2 1 0 0 ' U( , C , (I.51)
0.5 0 1 &0.5 1 0
say, where U* in (I.51) follows from (I.35) and 1 C ' P2,3E3,1()E2,1(!) ' 0 0 (I.52)
0.5 0 1 . &0.5 1 0
Now multiply (I.51) by elementary matrix E13(-1), i.e., subtract row 3 from row 1:
353 E1,3(&1)P2,3E3,1()E2,1(!)A , E1,3(&1)P2,3E3,1()E2,1(!) 2 4 0 ' 0 3 0 0 0 2 1.5 &1 0 0.5 0 1 , &0.5 1 0 (I.53)
multiply (I.53) by elementary matrix E12(-4/3), i.e., subtract 4/3 times row 3 from row 1: E1,2(&4/3)E1,3(&1)P2,3E3,1()E2,1(!)A , E1,2(&4/3)E1,3(&1)P2,3E3,1()E2,1(!) 2 0 0 ' 0 3 0 0 0 2 5/6 &1 &4/3 0.5 0 1 0 , &0.5 1 (I.54)
and finally, divide row 1 by 2, row 2 by 3, and row 3 by 2, or equivalently, multiply (I.54) by the diagonal matrix D* with diagonal elements 1/2, 1/3 and 1/2: D(E1,2(&4/3)E1,3(&1)P2,3E3,1()E2,1(!)A , D(E1,2(&4/3)E1,3(&1)P2,3E3,1()E2,1(!) ' I3 , D(E1,2(&4/3)E1,3(&1)P2,3E3,1()E2,1(!) 1 0 0 ' 0 1 0 0 0 1 5/12 &1/2 &2/3 1/6 0 1/3 . 0 &1/4 1/2 (I.55)
Observe from (I.55) that the matrix (A,I3) has been transformed into a matrix of the type (I3,A*) = (A*A,A*) , where A ( ' D(E1,2(&4/3)E1,3(&1)P2,3E3,1()E2,1(!) is the matrix consisting of the last three columns of (I.55). Consequently, A ( ' A &1 . This way of computing the inverse of a matrix is called the Gauss-Jordan iteration. In practice the Gauss-Jordan iteration is done in a slightly different but equivalent way, using a sequence of tableaus. Take again the matrix A in (I.32). The Gauss-Jordan iteration then starts
354 from the initial tableau: Tableau 1 A 2 1 4 2 2 3 I 1 0 0 0 1 0 0 0 1
&1 1 &1
If there is a zero in a pivot position, you have to swap rows, as we will see below. In the case of Tableau 1 there is not yet a problem, because the first element of row 1 is non-zero. The first step is to make all the non-zero elements in the first column equal to one, by dividing all the rows by their first element, provided that they are non-zero. Then we get: Tableau 2 1 2 1 1 2 3 1 &1 1 1/2 0 0 0 0 1 0 0 &1
Next, wipe out the first elements of rows 2 and 3, by subtracting row 1 from them: Tableau 3 1 2 1 0 0 2 0 &3 0 1/2 0 0 &1/2 1 0 &1/2 0 &1
Now we have a zero in a pivot position, namely the second zero of row 2. Therefore, swap rows 2 and 3:
355 Tableau 4 1 2 1 0 &3 0 0 0 2 Divide row 2 by -3 and row 3 by 2: Tableau 5 1 2 1 0 1 0 0 0 1 The left 33 block is now upper-triangular. Next, we have to wipe out, one by one, the elements in this block above the diagonal. Thus, subtract row 3 from row 1: Tableau 6 1 2 0 0 1 0 0 0 1 Finally, subtract 2 times row 2 from row 1: Tableau 7 I 1 0 0 0 1 0 0 0 1 1/6 A &1 5/12 &1/2 &2/3 0 1/3 0 &1/4 1/2 3/4 &1/2 1/6 0 &1/4 1/2 0 1/3 0 1/2 1/6 0 0 0 1/3 0 1/2 0 0 &1/2 0 &1 &1/2 1 0
&1/4 1/2
This is the final tableau. The last three columns now form A!1. Once you have calculated A!1, you can solve the linear system Ax = b by computing x = A!1b. However, you can also incorporate the latter in the Gauss-Jordan iteration, as follows. Let again A be the matrix in (I.32), and let for example
356 1 b ' 1 . 1 Insert this vector in Tableau 1: Tableau 1( A 2 1 4 2 2 3 b 1 1 1 I 1 0 0 0 1 0 0 0 1
&1 1 &1
and perform the same row operations as before. Then Tableau 7 becomes: Tableau 7( I 1 0 0 0 1 0 0 0 1 A &1b &5/12 1/2 1/4 1/6 A &1 5/12 &1/2 &2/3 0 1/3 0 &1/4 1/2
This is how matrices were inverted and system of linear equations were solved fifty and more years ago, using only a mechanical calculator. Nowadays of course you would use a computer, but the Gauss-Jordan method is still handy and not too time consuming for small matrices like the one in this example.
I.7.
Gaussian elimination of a non-square matrix The Gaussian elimination of a non-square matrix is similar to the square case, except that
in the final result the upper-triangular matrix now becomes an echelon matrix:
357 Definition I.10: An m n matrix U is an echelon matrix if for i = 2,...,m the first non-zero element of row i is farther to the right than the first non-zero element of the previous row i!1.
Theorem I.8 can now be generalized to:
Theorem I.11: For each matrix A there exists a permutation matrix P, possibly equal to the unit matrix I, a lower-triangular matrix L with diagonal elements all equal to 1, and an echelon matrix U, such that PA = LU. If A is a square matrix then U is an upper-triangular matrix. Moreover, in that case PA = LDU, where now U is an upper-triangular matrix with diagonal elements all equal to 1, and D is a diagonal matrix.8
Again, I will only prove the general part this theorem by examples. The parts for square matrices follow trivially from the general case. First, let 2 4 2 1 A ' 1 2 3 1 . &1 1 &1 0 which is the matrix (I.32) augmented with an additional column. Then it follows from (I.52) that 1 P2,3E3,1()E2,1(!)A ' 0 0 2 4 2 1 1 2 3 1 &1 1 &1 0 (I.57) 2 4 2 ' 1 ' U, 0 3 0 1/2 0 0 2 1/2 (I.56)
0.5 0 1 &0.5 1 0
358 where U is now an echelon matrix. As another example, take the transpose of the matrix A in (I.56): 2 1 &1 AT ' 4 2 1 2 3 &1 1 1 0 Then 2 1 &1 P2,3E4,2(&1/6)E4,3(1/4)E2,1(&2)E3,1(&1)E4,1(&1/2)A T ' 0 2 0 0 0 3 0 0 0 where again U is an echelon matrix. ' U, (I.59) . (I.58)
I.8.
Subspaces spanned by the columns and rows of a matrix The result in Theorem I.9 also reads as: A = BU, where B = P!1L is a non-singular matrix.
Moreover, note that the size of U is the same as the size of A, i.e., if A is an m n matrix, then so is U. Denoting the columns of U by u1,...,un, it follows therefore that the columns a1,...,an, of A are equal to Bu1,...,Bun, respectively. This suggests that the subspace spanned by the columns of A has the same dimension as the subspace spanned by the columns of U. To prove this conjecture, let VA be the subspace spanned by the columns of A, and let VU be the subspace spanned by the columns of U. Without loss or generality we may reorder the columns of A such that the first k columns a1,...,ak of A form a basis for VA . Now suppose that u1,...,uk are linear also 'j'1cjBuj ' 'j'1cja j = 0, which by the linear independence of a1,...,ak implies that all the
k k
dependent, i.e., there exist constants c1,...,ck not all equal to zero such that 'j'1cjuj ' 0 . But then
k
359 cjs are equal to zero. Hence, u1,...,uk are linear independent, and therefore the dimension of VU is greater or equal to the dimension of VA . But since U = B!1A, the same argument applies the other way around: the dimension of VA is greater or equal to the dimension of VU . Thus we have:
Theorem I.12: The subspace spanned by the columns of A has the same dimension as the subspace spanned by the columns of the corresponding echelon matrix U in Theorem I.9.
Next, I will show that
Theorem I.13: The subspace spanned by the columns of AT is the same as the subspace spanned by the columns of the transpose UT of the corresponding echelon matrix U in Theorem I.9.
Proof: Let A be an m n matrix. The equality A = BU implies that A T ' U TB T . The subspace spanned by the columns of AT consists of all vectors x 0 m for which there exists a vector c1 0 n such that x ' A Tc1 , and similarly the subspace spanned by the columns of UT consists of all vectors x 0 m for which there exists a vector c2 0 n such that x ' U Tc2 . Letting c2 ' B Tc1 the theorem follows. Q.E.D. Now let us have a closer look at a typical echelon matrix:
360 0 0 ( ( ( ( ( ( ( ( 0 0 0 0 ( ( ( ( ( ( U ' 0 0 0 0 0 0 ( ( ( ( 0 0 0 0 0 0 0 0 ( ( ! " ! ! " ! ! " ! ! " ! ! ! 0 0 0 0 0 0 0 0 0 0 where each smiley face ( (called a pivot) indicates the first nonzero elements of the row involved, and the *s indicate possible nonzero elements. Since the elements below the pivot in each column with a smiley face ( are zero, the columns involved are linear independent. In particular, it is impossible to write the last column with a pivot as a linear combination of the other ones. Moreover, it is easy to see that all the columns without a pivot can be formed as linear combinations of the columns with a pivot. Consequently, the columns of U with a pivot form a basis for the subspace spanned by the columns of U. But the transpose UT of U is also an echelon matrix, and the number of rows of U with a pivot is the same as the number of columns with a pivot, hence: , (I.60)
Theorem I.14: The dimension of the subspace spanned by the columns of an echelon matrix U is the same as the dimension of the subspace spanned by the columns of its transpose UT.
Combining Theorems I.11, I.12 and I.13, it follow now that Theorem I.6 holds.. The subspace spanned by the columns of a matrix A is called the column space of A, and is denoted by R(A). The row space of A is the space spanned by the columns of AT, i.e., the row space of A is R(AT). Theorem I.14 implies that the dimension of R(A) is equal to the
361 dimension of R(AT). There is also another space associated with a matrix A, namely the null space of A, denoted by N(A). This the space of all vectors x for which Ax = 0, which is also a subspace of a vector space. If A is square and non-singular, then N(A) = {0}, but if not it follows from Theorem I.12 that N(A) = N(U), where U is the echelon matrix in Theorem I.12. In order to determine the dimension of N(U), suppose that A is an m n matrix with rank r, and thus U is an m n matrix with rank r. Let R be an n n permutation matrix such that the first r columns of UR are the r columns of U with a pivot. Clearly, the dimension of N(U) is the same as the dimension of N(UR). We can partition UR as (Ur, Un!r), where Ur is the m r matrix consisting of the columns of U with a pivot, and Un!r is the m (n!r) matrix consisting of the other columns of U. Partitioning a vector x in N(UR) accordingly, i.e., x ' (xr , xn&r)T , we have URx ' Urxr % Un&rxn&r ' 0 . It follows from Theorem I.5 that Ur Ur is invertible, hence it follows from (I.61) and the partition x ' (xr , xn&r)T that &(Ur Ur)&1Ur Un&r In&r
T T T T T T T
(I.61)
x '
xn&r .
(I.62)
Therefore, N(UR) is spanned by the columns of the matrix in (I.62), which has rank n!r, and thus the dimension of N(A) is n!r. By the same argument it follows that the dimension of
362 N(AT) is m!r. The subspace N(AT) is called the left null space of A, because it consists of all vectors y for which y TA ' 0T . Summarizing, it has been shown that the following results hold.
Theorem I.15: Let A be an m n matrix with rank r. Then R(A) and R(AT) have dimension r, N(A) has dimension n!r, and N(AT) has dimension m!r.
Note that in general the rank of a product AB is not determined by the ranks r and s of A and B, respectively. At first sight one might guess that the rank of AB is min(r,s), but that is in general not true. For example, let A = (1,0) and BT = (0,1). Then A and B have rank 1, but AB = 0, which has rank zero. The only thing we know for sure is that the rank of AB cannot exceed min(r,s). Of course, if A and B are conformable invertible matrices, then AB is invertible, hence the rank of AB is equal to the rank of A and the rank of B, but that is a special case. The same applies to the case in Theorem I.5.
I.9.
Projections, projection matrices, and idempotent matrices Consider the following problem: Which point on the line through the origin and point a in
Figure I.3 is the closest to point b? The answer is: point p in Figure I.4 below. The line through b and p is perpendicular to the subspace spanned by a, and therefore the distance between b and any other point in this subspace is larger than the distance between b and p. Point p is called the
363 projection of b on the subspace spanned by a. In order to find p, let p = c.a, where c is a scalar. The distance between b and p is now 2b & c.a2 , so the problem is to find the scalar c which minimizes this distance. Since 2b & c.a2 is minimal if and only if 2b & c.a22 ' (b&c.a)T(b&c.a) ' b Tb & 2c.a Tb % c 2a Ta is minimal, the answer is: c ' a Tb/a Ta , hence p ' (a Tb/a Ta).a .
Figure I.4: Projection of b on the subspace spanned by a
Similarly, we can project a vector y in n on the subspace of n spanned by a basis {x1,...,xk}, as follows. Let X be the n k matrix with columns x1,...,xk. Any point p in the column space R(X) of X can be written as p = Xb, where b 0 k . Then the squared distance between y and p = Xb is 2y & Xb22 ' (y & Xb)T(y & Xb) ' y Ty & b TX Ty & y TXb % b TX TXb (I.63) ' y Ty & 2b TX Ty % b TX TXb ,
where the last equality follows from the fact that y TXb is a scalar (or equivalently, a 11 matrix), hence y TXb ' (y TXb)T ' b TX Ty . Given X and y, (I.63) is a quadratic function of b.
364 The first-order condition for a minimum of (I.63) is given by M2y & Xb22 Mb which has solution b ' (X TX)&1X Ty . Thus, the vector p in R(X) closest to y is p ' X(X TX)&1X Ty , which is the projection of y on R(X) . Matrices of the type in (I.66) are called projection matrices: (I.66) (I.65)
T
' &2X Ty % 2X TXb ' 0 ,
(I.64)
Definition I.11: Let A be an n k matrix with rank k. Then the n n matrix P = A(A TA)&1A T is called a projection matrix: For each vector x in n , Px is the projection of x on the column space of A.
Note that this matrix P is such that PP ' A(A TA)&1A TA(A TA)&1A T) ' A(A TA)&1A T ' P . This is not surprising, though, because p = Px is already in R(A), hence the point in R(A) closest to p is p itself.
Definition I.12: An n n matrix M is called idempotent if MM = M.
Thus, projection matrices are idempotent.
365 I.10. Inner product, orthogonal bases, and orthogonal matrices It follows from (I.10) that the cosines of the angle N between the vectors x in (I.2) and y in (I.5) is 'j'1xj yj
n
cos() '
2x2.2y2
'
x Ty . 2x2.2y2
(I.67)
Definition I.13: The quantity x Ty is called the inner product of the vectors x and y.
If x Ty ' 0 then cos(N) = 0, hence N = B/2 or N = 3B/4. This corresponds to an angle of 90 degrees and 270 degrees, respectively, hence x and y are then perpendicular. Such vectors are said to be orthogonal.
Definition I.14: Conformable vectors x and y are orthogonal if their inner product x Ty is zero. Moreover, they are orthonormal if in addition their lengths are 1: 2x2 ' 2y2 ' 1 .
If we flip in Figure I.4 point p over to the other side of the origin along the line through the origin and point a, and add b to !p, then the resulting vector c = b ! p is perpendicular to the line through the origin and point a. This is illustrated in Figure I.5. More formally, a Tc ' a T(b&p) ' a T(b & (a Tb/2a22)a ' a Tb & (a Tb/2a22)2a22 ' 0 .
366
Figure I.5: Orthogonalization
This procedure can be generalized to convert any basis of a vector space into an orthonormal basis, as follows. Let a1,......,ak, k # n , be a basis for a subspace of n , and let q1 ' 2a12&1.a1 . The projection of a2 on q1 is now p ' (q1 a2).q1 , hence a2 ' a2 & (q1 a2).q1 is orthogonal to q1 . Thus, let q2 ' 2a2 2&1a2 . The next step is to erect a3 perpendicular to q1 and q2 , which can be done by subtracting from a3 its projections on q1 and q2 : a3 ' a3 & (a3 q1)q1 & (a3 q2)q2 . Using the facts that by construction,
( T T ( T ( ( T
q1 q1 ' 1, q2 q2 ' 1, q1 q2 ' 0 , q2 q1 ' 0 , we have indeed that q1 a3 ' q1 a3 & (a3 q1)q1 q1 & (a3 q2)q1 q2 ' q1 a3 & a3 q1 ' 0 and similarly, q2 a3 ' 0 . Thus, let now q3 ' 2a3 2&1a3 . Repeating this procedure yields :
T ( ( ( T ( T T T T T T T
Theorem I.16: Let a1,......,ak be a basis for a subspace of n , and construct q1,......,qk recursively by:
367 q1 ' 2a12&1.a1 and aj ' aj & j (aj qi)qi , q j ' 2aj 2&1aj for j ' 2,3,....,k.
( j&1 i'1 T ( (
(I.68)
Then q1,......,qk is an orthonormal basis for the subspace spanned by a1,......,ak .
The construction (I.68) is known as the Gram-Smidt process. The orthonormality of q1,......,qk has already been shown, but it still has to be shown that q1,......,qk spans the same subspace as a1,......,ak . To show the latter, observe from (I.68) that a1,......,ak is related to q1,......,qk by aj ' j ui,jqi , j ' 1,2,...,k ,
j i'1
(I.69)
where uj,j ' 2aj 2, ui,j ' qi a j for i < j , ui,j ' 0 for i > j, i,j ' 1,....,k ,.
( ( T
(I.70)
with a1 ' a1 . It follows now from (I.69) that a1,......,ak are linear combinations of q1,......,qk , and it follows from (I.68) that q1,......,qk are linear combinations of a1,......,ak , hence the two bases span the same subspace. Observe from (I.70) that the kk matrix U with elements ui,j is an upper triangular matrix with positive diagonal elements. Moreover, denoting by A the nk matrix with columns a1,......,ak and by Q the nk matrix with columns q1,......,qk it follows from (I.69) that A = QU. Thus it follows from Theorem I.16, (I.69) and (I.70):
Theorem I.17: Let A be an nk matrix with rank k. There exists an nk matrix Q with orthonormal columns, and an upper triangular kk matrix U with positive diagonal elements, such that A = QU.
In the case k = n the matrix Q in Theorem I.17 is called an orthogonal matrix:
368 Definition I.15: An orthogonal matrix Q is a square matrix with orthonormal columns: QTQ = I.
In particular, if Q is an orthogonal nn matrix with columns q1,......,qn then the elements of the matrix QTQ are qi qj ' I(i ' j) , where I(.) is the indicator function9, hence QTQ = In. Thus QT = Q!1. It follows now from Theorem I.1 that also QQT = In, i.e, the rows of an orthogonal matrix are also orthonormal. Orthogonal transformations of vectors leave the angles between the vectors, and their lengths, the same. In particular, let x and y be vectors in n , and let Q be an orthogonal nn matrix. Then (Qx)T(Qy) ' x TQ TQy ' x Ty , 2Qx2 ' (Qx)T(Qx) ' x Tx ' 2x2 , hence it
T
follows from (I.67) that the angle between Qx and Qy is the same as the angle between x and y. In the case n = 2 the effect of an orthogonal transformation is a rotation. A typical orthogonal 22 matrix takes the form cos() sin()
Q '
sin() &cos()
(I.71)
This matrix transforms the unit vector e1 = (1, 0)T into the vector q ' (cos() , sin())T , and it follows from (I.67) that 2 is the angle between the two. By moving 2 from 0 to 2B the vector q rotates anti-clockwise from the initial position e1 back to e1.
369 I.11. Determinants: Geometric interpretation and basic properties The area enclosed by the parallelogram in Figure I.3 has a special meaning, namely the determinant of the matrix a1 b1 a2 b2 6 3 4 7
A ' (a,b) '
'
(I.72)
The determinant is denoted by det(A). This area is two times the area enclosed by the triangle formed by the origin and the points a and b in Figure I.3, which in its turn is the sum of the areas enclosed by the triangle formed by the origin, point b, and the projection p ' (a Tb/a Ta).a ' (a Tb/2a22).a of b on a, and the triangle formed by the points p, a, and b, in Figure I.4. The first triangle has area 2b & p2 times the distance of p to the origin, and the second triangle has area equal to 2b & p2 times the distance between p and a, hence the determinant of A is: det(A) ' 2b & p2.2a2 ' 2b & (a Tb/2a22)2.2a2 ' 2a222b22 & (a Tb)2 ' (a1 %a2 )(b1 %b2 ) & (a1b1%a2b2)2 ' (a1b2 & b1a2)2 ' *a1b2 & b1a2* ' a1b2 & b1a2 .
2 2 2 2
(I.73)
The latter equality is a matter of normalization, as &(a1b2 & b1a2) would also fit (I.73), but the chosen normalization is appropriate for (I.72), because then det(A) ' a1b2 & b1a2 ' 67 & 34 ' 30 . However, as I will show below, a determinant can be negative or zero. Equation (I.73) reads in words: (I.74)
370 Definition I.16: The determinant of a 22 matrix is the product of the diagonal elements minus the product of the off-diagonal elements.
We can also express (I.73) in terms of the angles Na and Nb of the vectors a and b, respectively, with the right hand side of the horizontal axis: a1 ' 2a2cos(a) , a2 ' 2a2sin(a) , b1 ' 2b2cos(b) , b2 ' 2b2sin(b) , hence det(A) ' a1b2 & b1a2 ' 2a2.2b2. cos(a)sin(b) & sin(a)cos(b) ' 2a2.2b2.sin(b & a) Since in Figure I.3, 0 0 . As an example of a negative determinant, let us swap the columns of A, and call the result matrix B: b1 a1 b2 a2 3 6 7 4 (I.75)
B ' AP1,2 ' (b,a) ' where
'
(I.76)
P1,2 '
0 1 1 0
is the elementary permutation matrix involved. Then det(B) ' b1a2 & a1b2 ' &30 . (I.77)
At first sight this looks odd, because it seems that the area enclosed by the parallelogram in Figure I.3 has not been changed. However, it has! Recall the interpretation of a matrix as a
371 mapping: A matrix moves a point to a new location, by replacing the original perpendicular coordinate system by a new system formed by the columns space of the matrix involved, with new units of measurement the lengths of the columns. In the case of the matrix B in (I.76) we have: Unit vectors Axis 1: Original e1 ' e2 ' 1 0 0 1 6 b ' 6 a ' New 3 7 6 4
2:
Thus, b is now the first unit vector, and a is the second. If we adopt the convention that the natural position of unit vector 2 is above the line spanned by the first unit vector, as is the case for e1 and e2, then we are actually looking at the parallelogram in Figure I.3 from the backside, as in Figure I.6:
Figure I.6: Backside of Figure I.3 Thus, the effect of swapping the columns of the matrix A in (I.72) is that Figure I.3 is flipped
372 over vertically 180 degrees. Since we are now looking at Figure I.3 from the back, which is the negative side, the area enclosed by the parallelogram is negative too! Note that this corresponds to (I.75): If we swap the columns of A, then we swap the angles Na and Nb in (I.75), and consequently the determinant flips sign.
Figure I.7: det(a,b) > 0
Figure I.8: det(a,b) < 0
373 As another example, let a be as before, but now position b in the south-west quadrant, as in Figure I.7 and Figure I.8. The fundamental difference between these two cases is that in Figure I.7 point b is above the line through a and the origin, so that Nb ! Na < B, whereas in Figure I.8 point b is below that line: Nb ! Na > B. Therefore, the area enclosed by the parallelogram in Figure I.7 is positive, whereas the area enclosed by the paralelllogram in Figure I.8 is negative. Hence in the case of Figure I.7, det(a,b) > 0, and in the case of Figure I.8, det(a,b) < 0. Again, in Figure I.8 we are looking at the backside of the picture; you have to flip it vertically to see the front side. What I have demonstrated here for 22 matrices is that if the columns are interchanged then the determinant changes sign. It is easy to see that the same applies to the rows. This property holds for general nn matrices as well, in the following way.
Theorem I.18: If two adjacent columns or rows of a square matrix are swapped10 then the determinant changes sign only.
Next, let us consider determinants of special 22 matrices. The first special case is the orthogonal matrix. Recall that the columns of an orthogonal matrix are perpendicular, and have unit length. Moreover, recall that an orthogonal 22 matrix rotates a set of points around the origin, leaving angles and distances the same. In particular, consider the set of points in the unit square formed by the vectors (0,0)T, (0,1)T, (1,0)T and (1,1)T. Clearly, the area of this unit square equals 1, and since the unit square corresponds to the 22 unit matrix I2., the determinant of I2 equals 1. Now multiply I2 by an orthogonal matrix Q. The effect is that the unit square is rotated,
374 without affecting its shape or size. Therefore,
Theorem I.19: The determinant of an orthogonal matrix is either 1 or -1, and the determinant of a unit matrix is 1.
The "either-or" part follows from Theorem I.18: swapping adjacent columns of an orthogonal matrix preserves orthonormality of the columns of the new matrix, but switches the sign of the determinant. For example, consider the orthogonal matrix Q in (I.71). Then it follows from Definition I.16 that det(Q) ' &cos2() & sin2() ' &1 .
Now swap the columns of the matrix (I.71): sin() &cos() cos() sin()
Q ' Then it follows from Definition I.16 that
det(Q) ' sin2() % cos2() ' 1 . Note that Theorem I.19 is not confined to the 22 case: it is true for orthogonal and unit matrices of any size. Next, consider the lower-triangular matrix a 0 b c
L '
(I.78)
According to Definition I.16, det(L) = a.c - 0.c = a.c, so that in the 22 case the determinant of a
375 lower-triangular matrix is the product of the diagonal elements. This is illustrated in Figure I.9 below. The determinant of L is the area in the parallelogram, which is the same as the area in the rectangle formed by the vectors (a,0)T and (0,c)T . This area is a.c. Thus, you can move b freely along the vertical axis without affecting the determinant of L. If you would flip the picture over vertically, which corresponds to replacing a by -a, the parallelogram will be viewed from the backside, hence the determinant flips sign.
Figure I.9: det(L) The same result applies of course to upper-triangular and diagonal 22 matrices. Thus we have:
Theorem I.20: The determinant of a lower-triangular matrix is the product of the diagonal elements. The same applies to an upper-triangular matrix and a diagonal matrix.
Again, this result is not confined to the 22 case, but holds in general. Now consider the determinant of a transpose matrix. In the 22 case the transpose AT of A can be formed by first swapping the columns and then swapping the rows. Then it follows from Theorem I.18 that in each of the two steps only the sign flips, hence
376 Theorem I.21: det(A) = det(AT).
The same applies to the general case: the transpose of A can be formed by a sequence of column exchanges and a corresponding sequence of row exchanges, and the total number of column and row exchanges is an even number. It follows from Theorem I.11 that in the case of a square matrix A there exist a permutation matrix P, possibly equal to the unit matrix I, a lower-triangular matrix L with diagonal elements all equal to 1, a diagonal matrix D, and an upper-triangular matrix U with diagonal elements all equal to 1, such that PA = LDU. Moreover, recall that a permutation matrix is orthogonal, because it consists of permutations of the columns of the unit matrix. Thus we can write A = PTLDU. Now consider the parallelogram formed by the columns of U. Since the diagonal elements of U are 1, the area of this parallelogram is the same as the area of the unit square: det(U) = det(I). Therefore, the effect of the transformation PTLD on the area of the parallelogram formed by the columns of U is the same as the effect of PTLD on the area of the unit square, and consequently det(PTLDU) = det(PTLD). The effect of multiplying D by L is that the rectangle formed by the columns of D is tilted and squeezed, without affecting the area itself. Therefore, det(LD) = det(D), and consequently det(PTLDU) = det(PTD). Next, PT permutates the rows of D, so the effect on det(D) is a sequence of sign switches only. The number of sign switches involved is the same as the number of column exchanges of PT necessary to convert PT into the unit matrix. If this number of swaps is even, then det(P) = det( PT) = 1, else det(P) = det(PT) = -1. Thus, in the 22 case (as well as in the general case) we have.
377 Theorem I.22: det(A) = det(P).det(D), where P and D are the permutation matrix and the diagonal matrix, respectively, in the decomposition PA = LDU in Theorem I.11 for the case of a square matrix A.
This result yields two important corollaries. First:
Theorem I.23: The determinant of a singular matrix is zero.
To see this, observe from the decomposition PA = LDU that A is singular if and only if D is singular. If D is singular then at least one of the diagonal elements of D is zero, hence det(D) = 0. Second, for conformable square matrices A and B we have
Theorem I.24: det(AB) = det(A).det(B).
This result can be shown in the same way as Theorem I.22, i.e., by showing that det(A) = det(PTLDUB) = det(P).det( DB) and det(DB) = det(D).det(B). Moreover, Theorems I.20 and I.24 imply that
Theorem I.25: Adding or subtracting a constant times a row or column to another row or column, respectively, does not change the determinant.
The reason is that this operation is equivalent to multiplying a matrix by an elementary matrix,
378 and that an elementary matrix is triangular with diagonal elements equal to 1. Furthermore, we have:
Theorem I.26: Let A be an nn matrix and let c be a scalar. If one of the columns or rows is multiplied by c, then the determinant of the resulting matrix is c.det(A). Consequently, det(c.A) = cn.det(A).
This theorem follows straightforwardly from Theorems I.20 and I.24. For example, let B be a diagonal matrix with diagonal elements 1, except for one element, say diagonal element i, which equals c. Then BA is the matrix A with the i-th column multiplied by c. Since by Theorem I.20, det(B) = c, the first part of Theorem I.26 for the "column" case follows from Theorem I.24, and the "row" case follows from det(AB) = det(A).det(B) = c.det(A). The second part follows by choosing B = c.In.. The results in this section merely serve as a motivation for what a determinant is, and its geometric interpretation and basic properties. All the results so far can be derived from three fundamental properties, namely the results in Theorems I.18, I.20 and I.21. If we would assume the that the results in Theorems I.18, I.20 and I.21 hold, and treat these properties as axioms, all the other results follow from these properties and the decomposition PA = LDU. Moreover, the function involved is unique. As to the latter, suppose that *(A) is a function satisfying (a) (b) If two adjacent rows or columns are swapped then * switches sign only. If A is triangular then *(A) is the product of the diagonal elements of A.
379 (c) *(AB) = *(A). *(B)
Then it follows from the decomposition A = PTLDU and axiom (c) that *(A) = *(PT)*(L)*(D)*(U). Moreover, it follows from axiom (b) that *(L) = *(U) =1 and *(D) = det(D). Finally, it follows from axiom (b) that the functions *(.) and det(.) coincide for unit matrices, so that by axiom (a), *(PT) = *(P) = det(P). Thus, *(A) = det(A), hence, the determinant is uniquely defined by the axioms (a), (b) and (c). Therefore,
Definition I.17: The determinant of a square matrix is uniquely defined by three fundamental properties: (a) (b) (c) If two adjacent rows or columns are swapped then the determinant switches sign only. The determinant of a triangular matrix is the product of the diagonal elements. The determinant of AB is the product of the determinants of A and B.
These three axioms can be used to derive a general expression for the determinant, together with the results below regarding determinants of block-triangular matrices.
380 I.12. Determinants of block-triangular matrices Consider a square matrix A partitioned as A1,1 A1,2 A2,1 A2,2
A '
(I.79)
where A1,1 and A2,2 are sub-matrices of size kk and mm, respectively, A1,2 is a km matrix and A2,1 is an mk matrix. This matrix A is block-triangular if either A1,2 or A2,1 is a zero matrix, and it is block-diagonal if both A1,2 and A2,1 are zero matrices. In the latter case A1,1 O O A2,2 , (I.80)
A '
where the two O blocks represent zero elements. For each block A1,1 and A2,2 we can apply Theorem I.11, i.e. A1,1 ' P1 L1D1U1 , A2,2 ' P2 L2D2U2 , hence
T T
A '
P1 L1D1U1 O
T
O ' P2 L2D2U2
P1 O O P2
L1 O O L2
D1 O O D2
U1 O O U2 (I.81)
' P TLDU , say. Then det(A) = det(P).det(D) = det(P1).det(P2).det(D1).det(D2) = det(A1,1).det(A2,2) . More generally, we have that
Theorem I.27: The determinant of a block-diagonal matrix is the product of the determinants of the diagonal blocks.
Next, consider the lower block-diagonal matrix
381 A1,1 O . (I.82)
A '
A2,1 A2,2
where again A1,1 and A2,2 are kk and mm matrices, respectively, and A2,1 is an mk matrix. Then it follows from Theorem I.25 that for any km matrix C, A1,1 O . (I.83)
det(A) ' det
A2,1&CA1,1 A2,2
&1
If A1,1 is nonsingular, then we can choose C ' A1,1 A2,1 so that A2,1&CA1,1 ' O . In that case it follows from Theorem I.27 that det(A) = det(A1,1).det(A2,2) . If A1,1 is singular, then the rows of A1,1 are linear dependent, and so are the first k rows of A. Hence, if A1,1 is singular then A is singular, so that by Theorem I.23, det(A) = det(A1,1).det(A2,2) ' 0 . Thus
Theorem I.28: The determinant of a block-triangular matrix is the product of the determinants of the diagonal blocks.
I.13.
Determinants and co-factors Consider the n n matrix a1,1 a1,n A ' ! " ! (I.84) an,1 an,n
and define the following matrix-valued function of A:
Definition I.18: The transformation (A | i1 , i2 ,..... , in) is a matrix formed by replacing in rows k
382 = 1,..,n of matrix (I.84) all but the iks element ak,i by zeros. Similarly, the transformation
k
(A | i1 , i2 ,..... , in) is a matrix formed by replacing in columns k = 1,..,n of matrix (I.84) all but the iks element ai , k by zeros.
k
For example, in the 33 case, 0 (A |2 , 3 ,1) ' 0 a3,1 a1,2 0 0 0 a2,3 , (A |2 , 3 ,1) ' 0 0 a2,1 0 0 0 a3,2 a1,3 0 0 .
Recall that a permutation of the numbers 1,2,....,n is an ordered set of these n numbers, and that there are n! of these permutations, including the trivial permutation 1,2,...,n. Moreover, it is easy to verify that for each permutation i1 , i2 ,..... , in of 1,2,....,n there exists a unique permutation j1 , j2 ,..... , jn such that (A | i1 , i2 ,..... , in) = (A | j1 , j2 ,..... , jn) and vice versa. Now define the function (A) ' j det[(A | i1 , i2 ,..... , in)] ' j det[(A | i1 , i2 ,..... , in)] , where the summation is over all permutations i1 , i2 ,..... , in of 1,2,....,n. Note that det[(A | i1 , i2 ,..... , in)] ' a1,i a2,, i .....an, i , where the sign depends on how
1 2 n
(I.85)
many row or column exchanges are needed to convert (A | i1 , i2 ,..... , in) into a diagonal matrix. If the number of exchanges is even, the sign is + and the sign is ! if this number is odd. Clearly, this sign is the same as the sign of the determinant of the permutation matrix (n | i1 , i2 ,..... , in)] , where +n is the n n matrix with all elements equal to 1. I will show now that *(A) in (I.85) satisfies the axioms in Definition I.17, so that:
383 Theorem I.29: The function *(A) in (I.85) is the determinant of A: *(A) = det(A).
Proof: First, exchange rows of A, say rows 1 and 2. The new matrix is P12A, where P12 is the elementary permutation matrix involved, i.e., the unit matrix with the first two columns exchanged. Then (P12A | i1 , i2 ,..... , in) ' P12(A | i1 , i2 ,..... , i n) , hence (P12A) ' det(P1,2)(A) = &(A) . Thus, *(A) satisfies axiom (a) in Definition I.17. Second, let A be lower-triangular. Then (A | i1 , i2 ,..... , in) is lower-triangular, but has at least one zero diagonal element for all permutations i1 , i2 ,..... , in except for the trivial permutation 1,2,....,n. Thus in this case (A) ' det[(A |1, , 2 , .... ,n) ' det(A) . The same applies of course to upper-triangular and diagonal matrices. Consequently *(A) satisfies axiom (b) in Definition I.17. Finally, observe that (AB | i1 , i2 ,..... , in) is a matrix with elements 'k'1am,kbk, i in position
n
m
(m,im), m = 1,....,n, and zeros elsewhere. Hence (AB | i1 , i2 ,..... , in) = A.(B | i1 , i2 ,..... , in) , which implies that (AB) ' det(A).(B) . Now write B as B = PTLDU , and observe from (I.86) and axiom (b) that (B) ' (P TLD)U ' det(P TLD)(U) ' det(P TLD)det(U) ' det(B) . The same applies to A. Thus, (AB) ' det(A).det(B) ' (A).(B) . Q.E.D. Next, consider the transformation: (I.87) (I.86)
384 Definition I.19: The transformation (A |k,m) is a matrix formed by replacing all elements in row k and column m by zeros, except element ak,m itself .
For example, in the 33 case, a1,1 a1,2 (A |2 , 3) ' 0 0 0 a2,3 . 0 (I.88)
a3,1 a3,2 Then it follows from (I.85) and Theorem I.29 that
det[(A |k,m)] ' j det[(A | i1 , i2 ,..... , i n)] ' j det[(A | i1 , i2 ,..... , i n)]
i k'm i k'k
(I.89)
hence:
Theorem I.30: For nn matrices A, det(A) ' 'm'1det[(A|k,m)] for k ' 1,2,....,n , and
n
det(A) ' 'k'1det[(A|k,m)] for m ' 1,2,....,n .

n
Now let us evaluate the determinant of the matrix (I.88). Swap rows 1 and 2, and then swap recursively columns 2 and 3 and columns 1 and 2. The total number of row and column exchanges is 3, hence a2,3 det[(A |2 , 3)] ' (&1)3det 0 0 0 0 ' a2,3(&1)2%3det a1,1 a1,2 a3,1 a3,2 (I.90)
a1,1 a1,2 a3,1 a3,2
' a2,3cof2,3(A) , say, where cof2,3(A) is the co-factor of element a2,3 of A . Note that the second equality follows
385 from Theorem I.27. Similarly, we need k-1 row exchanges and m-1 column exchanges to convert (A |k,m) into a block-diagonal matrix. More generally:
Definition I.20: The co-factor cofk,m(A) of an nn matrix A is the determinant of the (n!1)(n!1) matrix formed by deleting row k and column m, times (&1)k%m .
Thus, Theorem I.30 now reads as:
Theorem I.31: For nn matrices A, det(A) ' 'm'1ak,mcofk,m(A) for k ' 1,2,....,n , and also
n
det(A) ' 'k'1ak,mcofk,m(A) for m ' 1,2,....,n .

n
I.14.
Inverse of a matrix in terms of co-factors Theorem I.31 now enables us to write the inverse of a matrix A in terms of co-factors and
the determinant, as follows. Define
Definition I.20: The matrix cof1,1(A) cofn,1(A) Aadjoint ' ! " ! (I.91) cof1,n(A) cofn,n(A) is called the adjoint matrix of A.
Note that the adjoint matrix is the transpose of the matrix of co-factors with typical (i,j)s
386 element cofi , j(A) . Now observe from Theorem I.31 that det(A) ' 'k'1ai,kcofi,k(A) is just diagonal element i
n
of A.Aadjoint . Moreover, suppose that row j of A is replaced by row i, and call this matrix B. This has no effect on cofj,k(A) , but 'k'1ai,kcofj,k(A) ' 'k'1ai,kcofi,k(B) is now the determinant of B,
n n
and since the rows of B are linear dependent, det(B) = 0. Thus we have: 'k'1ai,kcofj,k(A) ' det(A) if i ' j ,
n
(I.92) ' 0 if i j , hence:
Theorem I.32: If det(A) 0 then A &1 '
1 . A det(A) adjoint
Note that the co-factors cofj,k(A) do not depend on ai,j. It follows therefore from Theorem I.31 that Mdet(A) ' cofi,j(A) . M ai , j
(I.93)
Using the well-known fact that dln(x)/dx ' 1 / x it follows now from Theorem I.32 and (I.93) that
Theorem I.33: If det(A) > 0 then
387 Mln[det(A)] Mln[det(A)] Ma1,1 Man,1 ! " ! ' A &1 . (I.94) Mln[det(A)] Mln[det(A)] Ma1,n Man,n
Mln[det(A)] MA
def.
'
Note that Theorem I.33 generalizes the formula d ln(x) / dx ' 1 / x to matrices. This result will be useful in deriving the maximum likelihood estimator of the variance matrix of the multivariate normal distribution.
I.15.
Eigenvalues and eigenvectors
I.15.1 Eigenvalues Eigenvalues and eigenvectors play a key role in modern econometrics, in particular in cointegration analysis. These econometric applications are confined to eigenvalues and eigenvectors of symmetric matrices, i.e., square matrices A for which A = AT. Therefore, I will mainly focus on the symmetric case.
Definition I.21: The eigenvalues11 of an nn matrix A are the solutions for 8 of the equation det(A!8In) = 0. It follows from Theorem I.29 that det(A) ' ' a1,i a2,, i .....an, i , where the summation is
1 2 n
over all permutations i1 , i2 ,..... , in of 1,2,....,n. Therefore, replacing A by A ! 8In it is not hard to verify that det(A ! 8In) is a polynomial of order n in 8: det(A ! 8In) = 'k'0 ck k , where the
n
388 coefficients ck are functions of the elements of A. For example, in the 22 case a1,1 a1,2 a2,1 a2,2
A ' we have a1,1 & a2,1
det(A & I2) ' det
a1,2 a2,2&
' (a1,1 & )(a2,2&) & a1,2a2,1
' 2 & (a1,1 % a2,2) % a1,1a2,2 & a1,2a2,1 , which has two roots, i.e., the solutions of 2 & (a1,1 % a2,2) % a1,1a2,2 & a1,2a2,1 = 0: a1,1 % a2,2 % (a1,1 & a2,2)2 % 4a1,2a2,1 2 a1,1 % a2,2 & (a1,1 & a2,2)2 % 4a1,2a2,1 2
1 '
2 '
There are three cases to be distinguished. If (a1,1 & a2,2)2 % 4a1,2a2,1 > 0 then 1 and 2 are different and real valued. If (a1,1 & a2,2)2 + 4a1,2a2,1 = 0 then 1 ' 2 and real valued. However, if (a1,1 & a2,2)2 % 4a1,2a2,1 < 0 then 1 and 2 are different but complex valued: a1,1 % a2,2 % i . &(a1,1 & a2,2)2 & 4a1,2a2,1 2 a1,1 % a2,2 & i. &(a1,1 & a2,2)2 & 4a1,2a2,1 2
1 '
2 '
where i = &1 . In this case 1 and 2 are complex conjugate: 2 ' 1 . 12 Thus, eigenvalues can be complex-valued!
389 Note that if the matrix A involved is symmetric : a1,2 ' a2,1 , then a1,1 % a2,2 % (a1,1 & a2,2)2 % 4a1,2 2 a1,1 % a2,2 & (a1,1 & a2,2)2 % 4a1,2 2
2 2
1 '
2 '
so that in the symmetric 22 case the eigenvalues are always real valued. It will be shown below that this is true for all symmetric nn matrices.
I.15.2 Eigenvectors By Definition I.21 it follows that if 8 is an eigenvalue of an nn matrix A, then A !8In is a singular matrix (possibly complex-valued!). Suppose first that 8 is real valued. Since the rows of A !8In are linear dependent there exists a vector x 0 n such that (A !8In)x = 0 (0 n ), hence Ax = 8x. Such a vector x is called an eigenvector of A corresponding to the eigenvalue 8. Thus in the real eigenvalue case:
Definition I.22: An eigenvector13 of an nn matrix A corresponding to an eigenvalue 8 is a vector x such that Ax = 8x.
However, this definition also applies to the complex eigenvalue case, but then the eigenvector x has complex-valued components: x 0 n. To show the latter, consider the case that 8 is complex-valued: 8 = " + i.$, ",$ 0 , $ 0. Then A !8In = A !"In ! i.$In
390 is complex-valued with linear dependent rows, in the following sense. There exist a vector x = a+i.b with a,b 0 n and length14 2x2 ' a Ta % b Tb > 0, such that (A !"In ! i.$In)(a + i.b) = [(A !"In )a + $b] + i.[(A !"In )b !$a] = 0 (0 n ). Consequently, (A !"In )a + $b = 0 and (A !"In )b ! $a = 0, and thus, A&In &In I n A&In a b 0 0
'
0 2n .
(I.95)
Therefore, in order for the length of x to be positive, the matrix in (I.95) has to be singular, and then
a b
can be chosen from the null space of this matrix.
I.15.3 Eigenvalues and eigenvectors of symmetric matrices On the basis of (I.95) it is easy to show that in the case of a symmetric matrix A, $ = 0 and b = 0:
Theorem I.34: The eigenvalues of a symmetric nn matrix A are all real valued, and the corresponding eigenvectors are contained in n.
Proof: First, note that (I.95) implies that for arbitrary > 0 , b a
T
0 '
A&In &In
I n A&In
a b (I.96)
' a TAb % b TAa &b Ta&a Tb% b Tb & a Ta Next observe that b Ta ' a Tb and by symmetry, b TAa ' (b TAa)T ' a TA Tb ' a TAb , where the first equality follows from the fact that b TAa is a scalar (or 11 matrix). Then we have for
391 arbitrary > 0 , (%1)a TAb & (%1)a Tb % (b Tb & a Ta) ' 0 . (I.97)
If we choose > = !1 in (I.97) then (b Tb % a Ta) ' .2x22 ' 0 , so that $ = 0 and thus 8 = " 0 . It is now easy to see that b no longer matters, so that we may choose b = 0. Q.E.D. There is more to say about the eigenvectors of symmetric matrices, namely:
Theorem I.35: The eigenvectors of a symmetric nn matrix A can be chosen orthonormal.
Proof: First assume that all the eigenvalues 1 , 2 , ..... , n of A are different. Let x1 , x2 , ..... , xn be the corresponding eigenvectors. Then for i j , xi Axj ' jxi xj and xj Axi ' ixi xj , hence (i &j)xi xj ' 0, because by symmetry, xi Axj ' (xi Axj)T ' xj A Txi ' xj Axi . Since i j , it follows now that xi xj ' 0 . Upon normalizing the eigenvectors as qj ' 2xj2&1xj the result follows. The case where two or more eigenvalues are equal requires a completely different proof. First, normalize the eigenvectors as qj ' 2xj2&1xj . Using the approach in Section I.10 we can always construct vectors y2 , ... , yn 0 n such that q1 , y2 , ... , yn is an orthonormal basis of n. Then Q1 ' (q1 , y2 , ... , yn) is an orthogonal matrix. The first column of Q1 AQ1 is Q1 Aq1 ' Q1 q1 . But by the orthogonality of Q1 , q1 Q1 ' q1 (q1 , y2 , ... , yn) ' (q1 q1 , q1 y2 , ... , q1 yn) ' (1 ,0 ,0 ,.... ,0) hence the first column of Q1 AQ1 is (1 , 0 , 0 , ...... , 0)T and by symmetry of Q1 AQ1 the first row is (1 , 0 , 0 , ...... , 0) . Thus Q1 AQ1 takes the form
T T T T T T T T T T T T T T T T T T T T T
392 1 0T . (I.98)
T Q1 AQ1
'
0 An&1
Next, observe that det(Q1 AQ1 & In) ' det(Q1 AQ1 & Q1 Q1) ' det[Q1 (A & In)Q1] ' det(Q1 )det(A & In)det(Q1) ' det(A & In)
T T T T T T
so that the eigenvalues of Q1 AQ1 are the same as the eigenvalues of A, and consequently the eigenvalues of An!1 are 2 , ..... , n . Applying the same argument as above to An!1 , there exists an orthogonal (n!1)(n!1) matrix Q2 such that 2 0T . (I.99)
(
( T ( Q2 An&1Q2
'
0 An&2
Hence, denoting 1 0T 0 Q2
(
Q2 '
(I.100)
which an orthogonal nn matrix, we can write 2 O (I.101)
Q2 Q1 AQ1Q2 '
O An&2
where 72 is a diagonal matrix with diagonal elements 1 and 2 . Repeating this procedure n-3 more times yields Qn Q2 Q1 AQ1Q2Qn ' where 7 is the diagonal matrix with diagonal elements 1 , 2 , ..... , n . Note that Q ' Q1Q2Qn is an orthogonal matrix itself, and it is now easy to verify that
T T T
(I.102)
393 the columns of Q are the eigenvectors of A. Q.E.D. In view of this proof, we can now restate Theorem I.35 as:
Theorem I.36: A symmetric matrix A can be written as A = Q7QT, where 7 is a diagonal matrix with the eigenvalues of A on the diagonal, and Q is the orthogonal matrix with the corresponding eigenvectors as columns.
This theorem yields a number of useful corollaries. The first one is trivial:
Theorem I.37: The determinant of a symmetric matrix is the product of its eigenvalues.
The next corollary concerns idempotent matrices [see Definition I.12]:
Theorem I.38: The eigenvalues of a symmetric idempotent matrix are either 0 or 1. Consequently, the only nonsingular symmetric idempotent matrix is the unit matrix I.
Proof: Let the matrix A in Theorem I.36 be idempotent: A.A = A. Then A = Q7QT = A.A = Q7QTQ7QT = Q72QT, hence 7 = 72. Since 7 is diagonal, each diagonal element 8j satisfies j ' j , hence j(1 & j) ' 0 . Moreover, if A is non-singular and idempotent then none of the eigenvalues can be zero, hence they are all equal to 1: 7 = I. Then A = QIQT = A = QQT = I. Q.E.D.
2
394 I.16. Positive definite and semi-definite matrices Another set of corollaries of Theorem I.36 concern positive [semi-] definite matrices. Most of the symmetric matrices we will encounter in econometrics are positive [semi-] definite or negative [semi-] definite. Therefore, the results below are of utmost importance to econometrics.
Definition I.23: An nn matrix A is called positive definite if for arbitrary vectors x 0 n unequal to the zero vector, xTAx > 0, and it is called positive semi-definite if for such vectors x, xTAx $ 0. Moreover, A is called negative [semi-] definite if !A is positive [semi-] definite.
Note that symmetry is not required for positive [semi-] definiteness. However, xTAx can always be written as x TAx ' x T 1 1 A % A T x ' x TAs x , 2 2
(I.103)
say, where As is symmetric, so that A is positive or negative [semi-] definite if and only if As is positive or negative [semi-] definite.
Theorem I.39: A symmetric matrix is positive [semi-] definite if and only if all its eigenvalues are positive [non-negative].
Proof: This result follows easily from xTAx = xTQ7QTx = yT7y = 'jjyj , where y = QTx
2
with components yj. Q.E.D.
395 Due to Theorem I.39, we can now define arbitrary powers of positive definite matrices:
Definition I.24: If A is a symmetric positive [semi-]definite nn matrix, then for " 0 [" > 0] the matrix A to the power " is defined by A" =Q7"QT, where 7" is a diagonal matrix with diagonal elements the eigenvalues of A to the power ": 7" = diag(1 , ..... , n) , and Q is the orthogonal matrix of corresponding eigenvectors.

The following theorem is related to Theorem I.8.
Theorem I.40: If A is symmetric and positive semi-definite then the Gaussian elimination can be conducted without need for row exchanges. Consequently, there exist a lower triangular matrix L with diagonal elements all equal to one, and a diagonal matrix D, such that A = LDLT.
Proof: First note that by Definition I.24 with " = 1/2, A1/2 is symmetric and (A1/2)TA1/2 = A1/2 A1/2 = A. Second, recall that according to Theorem I.17 there exist an orthogonal matrix Q and an upper-triangular matrix U such that A1/2 = QU, hence A = (A1/2)TA1/2 =UTQTQU = UTU. The matrix UT is lower-triangular, and can be written as UT = LD*, where D* is a diagonal matrix and L is a lower-triangular matrix with diagonal elements all equal to 1. Thus, A = LD*D*LT = LDLT, where D = D*D*. Q.E.D.
396 I.17. Generalized eigenvalues and eigenvectors The concepts of generalized eigenvalues and eigenvectors play a key role in cointegration analysis. Cointegration analysis is an advanced econometric time series topic, and will therefore likely not be covered in an introductory Ph.D. level econometrics course for which this review of linear algebra is intended. Nevertheless, to conclude this review I will briefly discuss what generalized eigenvalues and eigenvectors are, and how they relate to the standard case. Given two nn matrices A and B, the generalized eigenvalue problem is: Find the values for 8 for which det(A & B) ' 0 . (I.104)
Given a solution 8, which is called the generalized eigenvalue of A and B, the corresponding generalized eigenvector (relative to B) is a vector x in n such that Ax = 8Bx. However, if B is singular then the generalized eigenvalue problem may not have n solutions as in the standard case, and may even have no solution at all. To demonstrate this, consider the 22 case: a1,1 a1,2 a2,1 a2,2 b1,1 b1,2 b2,1 b2,2
A ' Then
, B '
397 a1,1&b1,1 a1,2&b1,2 a2,1&b2,1 a2,2&b2,2
det(A & B) ' det
' (a1,1&b1,1)(a2,2&b2,2) & (a1,2&b1,2)(a2,1&b2,1) ' a1,1a2,2&a1,2a2,1 % (a2,1b1,2&a2,2b1,1&a1,1b2,2%b2,1a1,2) % (b1,1b2,2&b2,1b1,2)2 If B is singular then b1,1b2,2&b2,1b1,2 ' 0 , so that then the quadratic term vanishes. But things can even be worse! It is possible that also the coefficient of 8 vanishes, whereas the constant term a1,1a2,2&a1,2a2,1 remains nonzero. In that case the generalized eigenvalues do not exist at all. This is for example the case if 1 0 0 &1 1 1 1 1
A ' Then
, B '
det(A & B) ' det
1&
&
& &1&
' &(1&)(1%) & 2 ' &1 ,
so that the generalized eigenvalue problem involved has no solution. Therefore, in general we need to require that the matrix B is non-singular. In that case the solutions of (I.104) are the same as the solutions of the standard eigenvalue problems det(AB!1!8I) = 0 and det(B &1A&I) ' 0 . The generalized eigenvalue problems that we shall encounter in advanced econometrics always involve a pair of symmetric matrices A and B, with B positive definite. Then the solutions of (I.104) are the same as the solutions of the symmetric standard eigenvalue problem det(B &1/2AB &1/2 & I) ' 0 . (I.105)
The generalized eigenvectors relative to B corresponding to the solutions of (I.104) can be
398 derived from the eigenvectors corresponding to the solutions of (I.105): B &1/2AB &1/2x ' x ' B 1/2B &1/2x Y A(B &1/2x) ' B(B &1/2x) Thus if x is an eigenvector corresponding to a solution 8 of (I.105) then y = B!1/2x is the generalized eigenvector relative to B corresponding to the generalized eigenvalue 8. Finally, note that generalized eigenvectors are in general not orthogonal, even if the two matrices involved are symmetric. However, in the latter case the generalized eigenvectors are "orthogonal with respect to the matrix B", in the sense that for different generalized eigenvectors y1 and y2, y1 By2 = 0. This follows straightforwardly from the link y = B!1/2x between generalized eigenvectors y and standard eigenvectors x.
T
(I.106)
I.18. 1.
Exercises Consider the matrix 2 A ' 1 1
4 &6 0 . &2 7 2
(a)
Conduct the Gaussian elimination by finding a sequence Ej of elementary matrices such
that (Ek Ek-1 .... E2 . E1) A = U = upper triangular. (b) Then show that by undoing the elementary operations Ej involved one gets the LU
decomposition A = LU, with L a lower triangular matrix with all diagonal elements equal to 1. (c) 2. Finally, find the LDU factorization. Find the 33 permutation matrix that swaps rows 1 and 3 of a 33 matrix.
399 3. Let 1 v1 0 0 A ' 0 v2 0 0 0 v3 1 0 0 v4 0 1 where v2 0. (a) (b) 4. Factorize A into LU. Find A-1, which has the same form as A. Compute the inverse of the matrix 1 2 0 A ' 2 6 4 . by any method. 5. Consider the matrix 1 1 (a) (b) (c) (d) 2 0 2 1 1 0 . 0 4 11 .
A ' &1 &2 1
2 &3 &7 &2
Find the echelon matrix U in the factorization PA = LU. What is the rank of A? Find a basis for the null space of A. Find a basis for the column space of A.
400 6. (a) (b) (c) 7. Find a basis for the following subspaces of 4: The vectors (x1 , x2 , x3 , x4)T for which x1 = 2x4. The vectors (x1 , x2 , x3 , x4)T for which x1 + x2 + x3 = 0 and x3 + x4 = 0. The subspace spanned by (1,1,1,1)T, (1,2,3,4)T, and (2,3,4,5)T. Let b1 and b ' b2 . b3
1 2 0 3 A ' 0 0 0 0 2 4 0 (a) (b) (c) (d) (e) 8. 1
Under what conditions on b does Ax = b have a solution? Find a basis for the nullspace of A. Find the general solution of Ax = b when a solution exists. Find a basis for the column space of A. What is the rank of AT ? Apply the Gram-Smidt process to the vectors 0 a ' 0 , b ' 1 0 1 , c ' 1 1 1 1
and write the result in the form A = QU, where Q is an orthogonal matrix and U is upper triangular. 9. 10. With a, b and c as in problem 8, find the projection of c on the space spanned by a and b. Find the determinant of the matrix A in problem 1.
401 11. Consider the matrix 1 a &1 1
A ' For which values of a has this matrix (a) (b) (c) (d) 12. two different real valued eigenvalues? two complex valued eigenvalues? two equal real valued eigenvalues? at least one zero eigenvalue?
For the case a = -4, find the eigenvectors of the matrix A in problem 11 and standardized
them to unit length. 13. Let A be a matrix with eigenvalues 0 and 1 and corresponding eigenvectors (1,2)T and
(2,!1)T. (a) (b) (c) 14. How can you tell in advance that A is symmetric? What is the determinant of A? What is A? The trace of a square matrix is the sum of the diagonal elements. Let A be a positive
definite kk matrix. Prove that the maximum eigenvalue of A can be found as the limit of the ratio trace(An)/trace(An!1) for n 6 4.
402 Endnotes 1. Law of Cosines: Consider a triangle ABC, let N be the angle between the legs C6A and C6B, and denote the lengths of the legs opposite to the points A, B and C by ", $, and (, respectively. Then 2 ' 2 % 2 & 2cos() . 2. In writing a matrix product it is from now on implicitly assumed that the matrices involved are conformable. 3. 4. 5. Here and in the sequel the columns of a matrix are interpreted as vectors. Here and in the sequel I denotes a generic unit matrix. The transpose of a matrix A is also denoted in the literature by A ) .
6. The notation Ei,j(c) will be used for a specific elementary matrix, and a generic elementary matrix will be denoted by "E". 7. A pivot is an element on the diagonal to be used to wipe out the elements below that diagonal element. 8. Note that the diagonal elements of D are the diagonal elements of the former uppertriangular matrix U. 9. I(true) ' 1 , I(false) ' 0 .
10. The operation of swapping a pair of adjacent columns or rows is also called a column or row exchange, respectively. 11. Eigenvalues are also called characteristic roots. The name "eigen" comes from the German word "Eigen" , which means "inherent", or "characteristic". 12. Recall that the complex conjugate of x = a + i.b, a,b 0 , is x ' a & i.b . See Appendix III. 13. 14. Eigenvectors are also called characteristic vectors. Recall (see Appendix III) that the length (or norm) of a complex number x = a + i.b, a,b a Ta%b Tb .
0 , is defined as |x| ' (a %i.b ).(a & i.b ) ' a 2%b 2 . Similarly, in the vector case x = a + i.b, a,b 0 n, the length of x is defined as 2x2 = (a % i.b )T(a & i.b ) '
403
Appendix II Miscellaneous Mathematics
In this appendix I will review various mathematical concepts, topics and related results that are used throughout the main text.
II.1.
Sets and set operations
II.1.1 General set operations The union AcB of two sets A and B is the set of elements that belong to either A or B or to both. Thus, denoting "belongs to" or "is element of" by the symbol 0, x 0 AcB implies that x 0 A or x 0 B, or in both, and vice versa. A finite union ^j'1Aj of sets A1,...,An is the set with the
n n
property that for each x 0 ^j'1Aj there exists an index i, 1 # i # n, for which x 0 Ai , and vice versa: If x 0 Ai for some index i, 1 # i # n, then x 0 ^j'1Aj . Similarly, the countable union
n
finite index i $ 1 then x 0 ^j'1Aj .

4
x 0 ^j'1Aj there exists a finite index i $ 1 for which x 0 Ai , and vice versa: If x 0 Ai for some
4
^j'1Aj of an infinite sequence of sets Aj, j = 1,2,3,....., is a set with the property that for each
4
The intersection A1B of two sets A and B is the set of elements which belong to both A and B. Thus, x 0A1B implies that x 0 A and x 0 B, and vice versa. The finite intersection _j'1Aj
n
vice versa: If x 0 Ai for all i = 1, ..., n, then x 0 _j'1Aj . Similarly, the countable intersection
n
of sets A1,...,An is the set with the property that if x 0 _j'1Aj then for all i = 1,...,n, x 0 Ai , and
n
_j'1Aj of an infinite sequence of sets Aj, j = 1,2,..., is a set with the property that if x 0 _j'1Aj
4 4
then for all indices i $ 1, x 0 Ai , and vice versa: If x 0 Ai for all indices i $ 1 then x 0 _j'1Aj .
4
404
A set A is a subset of a set B, denoted by AdB, if all the elements of A are contained in B. If AdB and BdA then A = B. The difference A\B (also denoted by A-B) of sets A and B is the set of elements of A that are not contained in B. The symmetric difference of two sets A and B is denoted and defined by AB ' (A/B) ^ (B/A) . = B/A ( also denoted by ~A) is called the complement of A with If AdB then the set A and ~_ A ' ^ A respect to B. If Aj for j = 1,2,3,..... are subsets of B then ~^jAj ' _jA j j j j j , for finite as well as countable infinite unions and intersections. . Sets A and B are disjoint if they do not have elements in common: A1B = i, where i
denotes the empty set, i.e., a set without elements. Note that Aci = A and A1i = i. Thus the empty set i is a subset of any set, including i itself. Consequently, the empty set is disjoint with any other set, including i itself. In general, a finite or countable infinite sequence of sets is disjoint if their finite or countable intersection is the empty set i. For every sequence of sets Aj , j = 1,2,3,....., there exists a sequence Bj , j = 1,2,3,....., of disjoint sets such that for each j, BjdAj, and ^jAj ' ^jBj . In particular, let B1 = A1 and Bn = An \ ^j'1 Aj for n = 2,3,4,.....
n&1
The order in which unions are taken does not matter, and the same applies to intersections. However, if you take unions and intersections sequentially it matters what is done first. For example, (AcB)1C = (A1C)c(B1C), which is in general different from Ac(B1C), except if AdC. Similarly, (A1B)c C = (AcC)1(BcC), which is in general different from A1(BcC), except if AdB.
405 II.1.2 Sets in Euclidean spaces An open g-neighborhood of a point x in a Euclidean space k is a set of the form Ng(x) ' {y 0 k : ||y&x|| < g} , g > 0 , and a closed g-neighborhood is a set of the form Ng(x) ' {y 0 k : ||y&x|| # g} , g > 0 . A set A is called open if for every x 0 A there exists a small open g-neighborhood Ng(x) contained in A. In short-hand notation: x 0 A g > 0: Ng(x) d A , where stands for for all and stands for there exists. Note that the gs may be different for different x. A point x called a point of closure of a subset A of k if every open g-neighborhood of A. Note that points of Ng(x) contains a point in A as well as a point in the complement A closure may not exist, and if one exists it may not be contained in A. For example, the Euclidean space k itself has no points of closure because its complement is empty. Moreover, the interval (0,1) has two points of closure, 0 and 1, both not included in (0,1). The boundary of a set A, denoted by MA, is the set of points of closure of A. Again, MA may be empty. A set A is closed if it contains all its points of closure if they exist. In other words, A is closed if and only if . The closure of a set A, MA i and MA d A. Similarly, a set A is open if either MA = i or MA d A denoted by A , is the union of A and its boundary MA: A ' A ^ MA . The set A\MA is the interior of A. Finally, if for each pair x, y of points in a set A and an arbitrary 0 [0,1] the convex combination z ' x % (1&)y is also a point in A then the set A is called convex.
406 II.2. Supremum and infimum The supremum of a sequence of real numbers, or a real function, is akin to the notion of a maximum value. In the latter case the maximum value is taken at some element of the sequence, or in the function case some value of the argument. Take for example the sequence an = (!1)n/n for n = 1,2,......., i.e., a1 = -1, a2 = 1/2, a3 = -1/3, a4 = 1/4, ..... Then clearly the maximum value is , which is taken by a2. The latter is what distinguishes a maximum from a supremum. For example, the sequence an = 1!1/n for n = 1,2,....... is bounded by 1: an < 1 for all indices n $1, and the upper bound 1 is the lowest possible upper bound, but there does not exist a finite index n for which an = 1. More formally, the (finite) supremum of a sequence an (n = 1,2,3,.......) is a number b, denoted by supn$1an , such that an # b for all indices n $1, and for every arbitrary small positive number g there exists a finite index n such that an > b!g. Clearly, this definition fits a maximum as well: a maximum is a supremum, but a supremum is not always a maximum. If the sequence an is unbounded from above, in the sense that for every arbitrary large real number M there exists an index n $1 for which an > M, then we say that the supremum is infinite: supn$1an = 4. The notion of a supremum also applies to functions. For example the function f(x) = exp(!x2) takes its maximum 1 at x = 0, but the function f(x) = 1!exp(!x2) does not have a maximum; it has supremum 1 because f(x) # 1 for all x but there does not exists a finite x for which f(x) = 1. As another example, let f(x) = x on the interval [a,b]. Then b is the maximum of f(x) on [a,b] but b is only the supremum f(x) on [a,b) because b is not contained in [a,b). More generally, the finite supremum of a real function f(x) on a set A, denoted by supx0A f(x) , is a real number b such that f(x) # b for all x in A, and for every arbitrary small positive number g there
407 exists an x in A such that f(x) > b!g. If f(x) = b for some x in A then the supremum coincides with the maximum. Moreover, the supremum involved is infinite, supx0A f(x) ' 4 , if for every arbitrary large real number M there exists an x in A for which f(x) > M. The minimum versus infimum cases are similar: infn$1an = !supn$1(!an) and infx0A f(x) = &supx0A (&f(x)) . The concepts of supremum and infimum apply to any collection {c", " 0 A} of real numbers, where the index set A may be uncountable, as we may interpret c" as a real function on the index set A, say c" = f(").
II.3.
Limsup and liminf Let an (n = 1,2,.......) be a sequence of real numbers, and define the sequence bn as b n ' supm$n am . (II.1)
Then bn is a non-increasing sequence: bn $ bn+1 because if an is greater than the smallest upper bound of an%1 , an%2 , an%3 ,...... then an is the maximum of an , an%1 , an%2 , an%3 ,...... , hence bn ' a n > bn%1 , and if not then bn ' bn%1 . Non-increasing sequences always have a limit, although the limit may be !4. The limit of bn in (II.1) is called the limsup of an :
def.
limsup a n ' lim supm$n am .

n64 n64
(II.2)
Note that since bn is non-increasing, the limit of bn is equal to the infimum of bn . Therefore, the limsup of an may also be defined as
408
def.
limsup an ' inf supm$n am .

n64 n$1
(II.3)
Note that the limsup may be +4 or !4, for example in the cases an = n and an = !n, respectively. Similarly, the liminf of an is defined by
def.
liminf an ' lim infm$n a m

n64 n64
(II.4)
or equivalently by
def.
liminf an ' sup infm$n a m .

n64 n$1
(II.5)
Again, it is possible that the liminf is +4 or !4. Note that liminfn64 an # limsupn64 an , because infm$n am # supm$n am for all indices n $1, and therefore the inequality must hold for the limits as well.
Theorem II.1: (a) If liminfn64 an ' limsupn64 an then limn64 an ' limsup an , and if liminfn64 an <
n64
limsupn64 an then the limit of an does not exist. (b) Every sequence an contains a sub-sequence an such that limk64 an ' limsupn64 an , and
k k m m
an also contains a sub-sequence an such that limm64 an ' liminfn64 an .
Proof: The proof of (a) follows straightforwardly from (II.2), (II.4) and the definition of a limit. The construction of the sub-sequence an in part (b) can be done recursively, as follows.
k
409 Let b ' limsupn64 an < 4 . Choose n1 = 1, and suppose that we have already constructed an for
j
j=1,...,k $1. Then there exists an index nk+1 > nk such that an
k%1
> b & 1/(k%1) , because
otherwise am # b & 1/(k%1) for all m $ nk , which would imply that limsupn64 an # b !1/(k+1). Repeating this construction yields a sub-sequence an such that from large enough k onwards,
k
b & 1/k < an # b . Letting k64, the limsup case of part (b) follows. If limsupn64 an ' 4 then
k
for each index nk we can find an index nk+1 > nk such that an
k%1
> k%1 , hence then limk64 an =

k
4. The sub-sequence in the case limsupn64 an ' &4 and in the liminf case can be constructed similarly. Q.E.D. The concept of a supremum can be generalized to sets. In particular, the countable union ^j'1Aj may be interpreted as the supremum of the sequence of sets Aj, i.e., the smallest set
4 4
containing all the sets Aj. Similarly, we may interpret the countable intersection _j'1Aj as the infimum of the sets Aj, i.e., the largest set contained in each of the sets Aj. Now let for n = 1,2,3,..., Bn ' ^j'nAj . This is a non-increasing sequence of sets: Bn+1 d Bn , hence _j'1Bn ' Bn .
4 n
The limit of this sequence of sets is the limsup of An for n 64, i.e., similarly to (II.3) we have limsup An ' _ ^ Aj .
def. 4 4 n64 n'1 j'n
hence ^j'1Cn ' C n . The limit of this sequence of sets is the liminf of An for n 64, i.e. similarly
n
Next, let for n = 1,2,3,..., Cn ' _j'nAj . This is a non-decreasing sequence of sets: Cn d Cn+1,
4
to (II.5) we have liminf An ' ^ _ Aj .

def. 4 4 n64 n'1 j'n
410 II.4. Continuity of concave and convex functions A real function on a subset of a Euclidean space is convex if for each pair of points a,b and every 0 [0,1], (a%(1&)b) $ (a) % (1&)(b) . For example, (x) ' x 2 is a convex function on the real line, and so is (x) ' exp(x) . Similarly, is concave if for each pair of points a,b and every 0 [0,1], (a%(1&)b) # (a) % (1&)(b) . I will prove the continuity of convex and concave functions by contradiction. Suppose that is convex but not continuous in a point a. Then (a%) ' lim(b) (a)
b 9a
(II.6)
or (a&) ' lim(b) (a)

b 8a
(II.7)
or both. In the case (II.6) we have (a%) ' lim(a % 0.5(b&a)) ' lim(0.5a%0.5b)
b 9a b 9a
# 0.5(a) % 0.5lim(b) ' 0.5(a) % 0.5(a%) ,

b 9a
hence (a%) # (a) and therefore by (II.6), (a%) < (a) . Similarly, if (II.7) is true then (a&) < (a) . Now let > 0. By the convexity of it follows that (a) ' (0.5(a&) % 0.5(a%)) # 0.5(a&) % 0.5(a%) , and consequently, letting 9 0 , and using the fact that (a%) < (a) , or (a&) < (a) , or both, we have (a) # 0.5(a&) % 0.5(a%) < (a) . Since this result is impossible, it follows that (II.6) and (II.7) are impossible, hence is continuous. If is concave, then & is convex and thus continuous, hence concave functions are continuous.
411 II.5. Compactness An (open) covering of a subset of a Euclidean space k is a collection of (open) subsets U(), 0 A, of k , where A is a possibly uncountable index set, such that d ^0AU() . A set is called compact if every open covering has a finite sub-covering, i.e., if U("), " 0 A, is an open covering of 1 and 1 is compact then there exists a finite subset B of A such that d ^0BU() . The notion of compactness extends to more general spaces than only Euclidean spaces. However,
Theorem II.2: Closed and bounded subsets of Euclidean spaces are compact.
Proof: I will prove the result for sets in only. First note that boundedness is a necessary condition for compactness, because a compact set can always be covered by a finite number of bounded open sets. Next let is a closed and bounded subset of the real line. By boundedness, there exists points a and b in such that is contained in [a,b]. Since every open covering of can be extended to an open covering of [a,b], we may without loss of generality assume that = [a,b]. For notational convenience, let = [0,1]. There always exists an open covering of [0,1], because for arbitrary g > 0, [0,1] d ^0#x#1(x&g,x%g) . Let U(), 0 A, be an open covering of [0,1]. Without loss of generality we may assume that each of the open sets U() takes the form (a(),b()). Moreover, if for two different indices and , a() = a(), then either (a(),b()) d (a(),b()), so that (a(),b()) is superfluous, or (a(),b()) e (a(),b()), so that (a(),b()) is
412 superfluous. Thus, without loss of generality we may assume that the a()s are all distinct and can be arranged in increasing order. Consequently, we may assume that the index set A is the set of the a()s themselves, i.e., U(a) = (a,b(a)), a 0 A, where A is a subset of such that [0,1] d â0A(a,b(a)) . Furthermore, if a1 < a2 then b(a1) < b(a2) , as otherwise (a2,b(a2)) is superfluous. Now let 0 0 (a1 ,b(a1)) , and define for n = 2,3,4,..., an ' (an&1%b(an&1))/2 . Then [0,1] d ^n'1(an,b(an)) . This implies that 1 0 ^n'1(an,b(an)) , hence there exists an n such that
4 4
1 0 (an,b(a n)) . Consequently, [0,1] d ^j'1(aj,b(a j)) . Thus, [0,1] is compact. This argument
n
extends to arbitrary closed and bounded subsets of a Euclidean space. Q.E.D. A limit point of a sequence xn of real numbers is a point x* such that for every g > 0 there exists an index n for which |xn&x(| < g . Consequently, a limit point is a limit along a subsequence. Sequences xn confined to an interval [a,b] always have at least one limit point, and these limit points are contained in [a,b], because limsupn64xn and liminfn64xn are limit points contained in [a,b], and any other limit point must lie between liminfn64xn and limsupn64xn . This property carries over to general compact sets:
Theorem II.3: Every infinite sequence n of points in a compact set has at least one limit point, and all the limit points are contained in .
Proof: Let . be a compact subset of a Euclidean space and let k, k ' 1,2,...., be a decreasing sequence of compact subsets of each containing infinitely many n s , to be constructed as follows. Let 0 ' and k $ 0. There exist a finite number of points k,j , j ' 1,....,mk, such that, which Uk(() ' {: ||&(|| < 2&k}, k is contained in
(
^j'k1Uk(k,j). Then at least one of these open sets contains infinity many points n , say Uk(k,1).
m ( (
413
Next, let k%1 ' {: ||&k,1|| # 2&k}_k ,

(
which is compact, and contains infinity many points n . Repeating this construction it is easy to verify that _k'0k is a singleton, and that this singleton is a limit point contained in . Finally,
4
if a limit point ( is located outside then for some large k, Uk(()_ ' i, which contradicts the requirement that Uk(() contains infinitely many n s. Q.E.D.
Theorem II.4: Let n be a sequence of points in a compact set . If all the limit points of n are the same, then limn64n exists and is a point in .
Proof: Let ( 0 be the common limit point. If the limit does not exists, then there exists a > 0 and an infinite subsequence n such that |n &(| $ for all k. But n has also
k k k
limit point ( , so that there exists a further subsequence n (m) which converges to ( . Therefore,
k
the theorem follows by contradiction. Q.E.D.
Theorem II.5: For a continuous function g on a compact set , sup0g() = max0g() and inf0g() = min0g() . Consequently, argmax0g() 0 and argmin0g() 0 .
Proof: It follows from the definition of sup0g() that or each k $ 1 there exists a point k 0 such that g(k) > sup0g() & 2&k , hence limk64g(k) ' sup0g() . Since is compact the sequence k has a limit point ( 0 (see Theorem II.3), hence by the continuity
414 of g, g(() ' sup0g() . Consequently, sup0g() = max0g() ' g(() . Q.E.D.
II.6.
Uniform continuity A function g on k is called uniformly continuous if for every g > 0 there exists a > 0
such that |g(x) & g(y)| < g if 2x & y2 < . In particular,
Theorem II.6: If a function g is continuous on a compact subset of k then it is uniformly continuous on .
Proof: Let g > 0 be arbitrary, and observe from the continuity of g that for each x in there exists a (x) > 0 such that |g(x) & g(y)| < g/2 if 2x & y2 < 2(x) . Now let U(x) = {y 0 k: 2y & x2 < (x)} . Then the collection {U(x), x 0 } is an open covering of , hence by compactness of there exists a finite number of points 1 ,..... , n in such that
n
d ^j'1U(j) . Next, let ' min1#j#n(j) . Each point x 0 belongs to at least one of the
open sets U(j): x 0 U(j) for some j. Then 2x & j2 < (j) < 2(j) , hence |g(x) & g(j)| < g/2. Moreover, if 2x & y2 < then 2y & j2 ' 2y & x % x & j2 # 2x & y2 % 2x & j2 < % (j) # 2(j) , hence |g(y) & g(j)| < g/2 . Consequently, |g(x) & g(y)| # |g(x) & g(j)| % |g(y) & g(j)| < g if 2x & y2 < . Q.E.D.
415 II.7. Derivatives of functions of vectors and matrices Consider a real function f(x) ' f(x1 , ..... , xn) on n , where x ' (x1 , ..... , xn)T . Recall that the partial derivative of f to a component xi of x is denoted and defined by Mf(x1 , ..... , xn) def. f(x , .. , xi&1 , xi% , xi%1 , ... , xn) & f(x1 , .. , xi&1 , xi , xi%1 , ... , xn) Mf(x) ' ' lim 1 . Mx i Mx i 60 For example, let f(x) ' Tx ' x T ' 1x1 % ... nxn . Then Mf(x)/Mx1 ! Mf(x)/Mxn ' 1 ! n ' .
This result could also have been obtained by treating xT as a scalar and taking the derivative of f(x) ' x T to xT: M(x T)/Mx T ' . This motivates the convention to denote the column vector of partial derivative of f(x) by Mf(x)/Mx T . Similarly, if we treat x as a scalar and take the derivative of f(x) ' Tx to x, then the result is a row vector: M(Tx)/Mx ' T . Thus in general, Mf(x)/Mx1 ! Mf(x)/Mxn ,
Mf(x) Mx T
def.
'
Mf(x) def. ' Mf(x)/Mx1 , .... , Mf(x)/Mxn . Mx
If the function H is vector-valued, say H(x) ' (h1(x) , ..... , hm(x))T , x 0 n , then applying the operation M/Mx to each of the components yields an mn matrix: Mh1(x)/Mx ! Mhm(x)/Mx ' Mh1(x)/Mx1 Mh1(x)/Mxn ! ! . Mh m(x)/Mx1 Mhm(x)/Mxn
MH(x) ' Mx
def.
Moreover, applying the latter to a column vector of partial derivatives of a real function f yields
416 M2f(x) M2f(x) Mx1Mx1 Mx1Mxn ! " ! ' M2f(x) M2f(x) MxnMx1 Mx n Mx n
M(Mf(x)/Mx T) ' Mx
M2f(x) MxMx T
say. In the case of an mn matrix X with columns x1 , ..... ,xn 0 k , xj ' (x1,j , .... ,xm,j)T , and a differentiable function f(X) on the vector space of kn matrices, we may interpret X = (x1 , ... , xn) as a row of column vectors, so that Mf(X)/Mx1 ! Mf(X)/Mxn
def.
Mf(X) Mf(X) ' ' MX M(x1 ,.....,xn)
def.
def.
Mf(X)/Mx1,1 Mf(X)/Mxm,1 ! " Mf(X)/Mx1,n Mf(X)/Mxm,n
'
is an nm matrix. For the same reason, Mf(X)/MX T ' (Mf(X)/MX)T . An example of such a derivative to a matrix is given by Theorem I.33 in Appendix I, which states that if X is a square nonsingular matrix then Mln[det(X)]/MX ' X &1 . Next, consider the quadratic function f(x) = a + xTb + xTCx, where x1 x ' : , b ' xn b1 : , C ' bn c1,1 .... c1,n : .... : , with ci,j ' cj,i . cn,1 .... cn,n
Thus, C is a symmetric matrix. Then
417 M a % 'i'1bixi % 'i'1'j'1xici,jxj

n n n
Mf(x)/Mxk ' ' j bi

n i'1
Mx k Mxici,jxj Mx k
n
Mx i Mx k
% jj
n n i'1 j'1
' bk % 2ck,kxk % j xici,k % j ck,jxj

n n i'1 ik j'1 jk
' bk % 2j ck,jxj , k ' 1 , ... , n ,

j'1
hence, stacking these partial derivatives in a column vector yields Mf(x)/Mx T ' b % 2Cx . (II.8)
If C is not symmetric, we may without loss of generality replace C in the function f(x) by the symmetric matrix (C + CT)/2, because xTCx = (xTCx)T = xTCTx, so that then Mf(x)/Mx T ' b % Cx % C Tx . The result (II.8) for the case b = 0 can be used to give an interesting alternative interpretation of eigenvalues and eigenvectors of symmetric matrices, namely as the solutions of a quadratic optimization problem under quadratic restrictions. Consider the optimization problem max or min x TAx s.t. x Tx ' 1 , (II.9)
where A is a symmetric matrix, and max and min include local maxima and minima, and saddle-point solutions. The Lagrange function for solving this problem is (x , ) ' x TAx % (1 & x Tx) , with first-order conditions M(x , )/Mx T ' 2Ax & 2x ' 0 Y Ax ' x , M(x , )/M ' 1 & x Tx ' 0 Y 2x2 ' 1 . (II.10) (II.11)
Condition (II.10) defines the Lagrange multiplier . as the eigenvalue and the solution for x as the
418 corresponding eigenvector of A, and (II.11) is the normalization of the eigenvector to unit length. Combining (II.10) and (II.11) it follows that = xTAx.
II.8.
The mean value theorem Consider a differentiable real function f(x), displayed as the curved line in the following
figure:
Figure II.1. The mean value theorem
We can always find a point c in the interval [a,b] such that the slope of f(x) at x = c, which is equal to the derivative f )(c) , is the same as the slope of the straight line connecting the points (a,f(a)) and (b,f(b)), simply by shifting the latter line parallel to the point where it be comes tangent to f(x). The slope of this straight line through the points (a, f(a)) and (b,f(b)) is: (f(b)!f(a))/(b!a). Thus, at x = c we have f )(c) ' (f(b) & f(a)) / (b & a) , or equivalently
419 f(b) ' f(a) % (b & a)f )(c) . This easy result is called the mean value theorem. Since this point c can also be expressed as c ' a % (b & a) , with 0 # ' (c & a) / (b & a) # 1 , we can now state the mean value theorem as:
Theorem II.7(a): Let f(x) be a differentiable real function on an interval [a,b], with derivative f )(x) . For any pair of points x , x0 0 [a,b] there exists a 0 [0,1] such that f(x) = f(x0) % (x & x0)f )(x0 % (x&x0)) .
This result carries over to real functions of more than one variable:
Theorem II.7(b): Let f(x) be a differentiable real function on a convex subset C of k . For any pair of points x , x0 0 C there exists a 0 [0,1] such that f(x) ' f(x0) % (x & x0)T(M/My T)f(y)*y'x %(x&x ) .
0 0
II.9.
Taylors theorem The mean value theorem implies that if for two points a < b, f(a) = f(b), then there
exists a point c 0 [a,b] such that f )(c) ' 0 . This fact is the core of the proof of Taylors theorem:
Theorem II.8(a): Let f(x) be an n-times continuously differentiable real function on an interval [a,b], with the n-th derivative denoted by f (n)(x) . For any pair of points x , x0 0 [a,b] there exists
420 a 0 [0,1] such that f(x) ' f(x0) % j

n&1 k'1
(x & x0)k k!
f (k)(x0) %
(x & xn)n n!
f (n)(x0 % (x&x0)) .
Proof: Let a # x0 < x # b be fixed. We can always write f(x) ' f(x0) % j
n&1 k'1
(x & x0)k k!
n&1 k'1
f (k)(x0) % Rn ,
(II.12)
where Rn is the remainder term. Now let a # x0 < x # b be fixed, and consider the function g(u) ' f(x) & f(u) & j with derivative
n&1 nRn(x & u)n&1 (x & u)k&1 (k) (x & u)k (k%1) f (u) & j f (u) % g (u) ' &f (u) % j (k&1)! k! k'1 k'1 (x & x0)n n&1 ) )
R (x & u)n (x & u)k (k) f (u) & n k! (x & x0)n
' &f )(u) % j

n&2 k'0
n&1 nRn(x & u)n&1 (x & u)k (k%1) (x & u)k (k%1) f (u) & j f (u) % k! k! k'1 (x & x0)n
' &
nRn(x & u)n&1 (x & u)n&1 (n) f (u) % . (n&1)! (x & x0)n
Then g(x) ' g(x0) ' 0 , hence there exists a point c 0 [x0,x] such that g )(c) ' 0 : nRn(x & c)n&1 (x & c)n&1 (n) f (c) % . (n&1)! (x & x0)n
0 ' & Therefore, (x & xn)n n!
Rn '
f (c) '
(n)
(x & xn)n n!
f (n) x0 % (x&x0) ,
(II.13)
where c ' x0 % (x&x0). Combining (II.12) and (II.13) the theorem follows. Q.E.D.
421 Also Taylors theorem carries over to real functions of more than one variable, but the result involved is awkward to display for n > 2. Therefore, we only state the second-order Taylor expansion theorem involved:
Theorem II.8(b): Let f(x) be a twice continuously differentiable real function on a convex subset of n . For any pair of points x , x0 0 there exists a 0 [0,1] such that Mf(y) / My T 000y'x 0 1 M2f(y) (x & x0)T (x & x0) . / 2 MyMy T 000y'x %(x&x )
0 0
f(x) ' f(x0) % (x & x0)T
(II.14)
II.10. Optimization Theorem II.8(b) shows that the function f(x) involved is locally quadratic. Therefore, the conditions for a maximum or a minimum of f(x) in a point x0 0 can be derived from (II.14) and the following theorem.
Theorem II.9: Let C be a symmetric nn matrix, and let f(x) = a + xTb + xTCx, x 0 n, where a is a given scalar and b is a given vector in n. If C is positive [negative] definite then f(x) takes a unique minimum [maximum], at x ' &C &1b .
Proof: The first-order condition for a maximum or minimum is Mf(x)/Mx T ' 0 (0 n) , hence x ' &C &1b . As to the uniqueness issue, and the question whether the optimum is a minimum or a maximum, recall that C = QQT, where is the diagonal matrix of the eigenvalues of C and Q is the corresponding matrix of eigenvectors. Thus we can write f(x) as
422 f(x) = a + xT QQT b + xT QQT x. Let y = QT x = (y1,...,yn)T and = QT b = (1,...,n)T . Then f(Qy) = a + yT + yT y = a % 'j'1(jyj % jyj ) . The latter is a sum of quadratic functions in
n 2
one variable which each has a unique minimum if j > 0 and a unique maximum if j < 0. Q.E.D. It follows now from (II.14) and Theorem II.9 that:
Theorem II.10: The function f(x) in Theorem II.8(b) takes a local maximum [minimum] in a point x0 0 , i.e., x0 is contained in an open subset 0 of such that for all x 0 0\{x0}, f(x) < f(x0) [f(x) > f(x0)], if and only if Mf(x0)/Mx0 ' 0 (0 n) , and the matrix M2f(x0)/(Mx0Mx0 ) is negative [positive] definite.
T T
A practical application of the Theorems II.8(a), II.9 and II.10 is the Newton iteration for finding a minimum or a maximum of a function. Suppose that the function f(x) in Theorem II.8(b) takes a unique global maximum at x( 0 . Starting from an initial guess x0 of x*, let for k $ 0, M2f(xk)
T Mx k Mx k &1
xk%1 ' xk &
Mf(xk)
T Mx k
Thus, the Newton iteration maximizes or minimized the local quadratic approximation of f in xk. The iteration is stopped if for some small threshold g > 0 , ||xk%1 & xk|| < g .
423
Appendix III A Brief Review of Complex Analysis
III.1. The complex number system Complex numbers have many applications. The complex number system allows to conduct computations that would be impossible to perform in the real world. In probability and statistics we mainly use complex numbers in dealing with characteristic functions, but in time series analysis complex analysis plays a key-role. See for example Fuller (1996). Therefore, in this appendix I will review the basics of complex analysis. Complex numbers are actually two-dimensional vectors endowed with arithmetic operations that make them act as numbers. Therefore, I will introduce the complex numbers in their "real" form, as vectors in 2. Next to the usual addition and scalar multiplication operators on the elements of 2 (see Appendix I), define the vector multiplication operator "" by: a b Observe that a b c d c d a b c d
def.
'
a.c&b.d b.c%a.d
(III.1)
'
(III.2)
Moreover, define the inverse operator !1" by
424 a b so that a b
&1 &1 def.
'
1 a %b
2 2
a &b
, provided that a 2%b 2 > 0 ,
(III.3)
a b
'
a b
a b
&1
'
a 2%b 2 &b
a b
'
1 0
(III.4)
The latter vector plays the same role as the number 1 in the real number system. Furthermore, we can now define the division operator / by a b c d
def.
'
a b
c d
&1
'
c 2%d 2 b
c &d
'
a.c%b.d
c 2%d 2 b.c&a.d
(III.5)
provided that c 2%d 2 > 0. Note that 1 0

2
c d
'
c 2%d 2 &d
'
c d
&1
(III.6)
In the subspace 1 ' { (x,0)T , x 0 } these operators work the same as for real numbers: a 0 c 0 a.c 0 c 0
&1
'
'
1/c 0
a 0
c 0
'
a/c 0
(III.7)
provided that c 0. Therefore, all the basic arithmetic operations (addition, subtraction, multiplication, division) of the real number system apply to 1 , and vice versa. In the subspace 2 ' { (0,x)T , x 0 } the multiplication operator yields 0 b In particular, note that 0 d &b.d 0
2 2
'
(III.8)
425 0 1 Now denote

def.
0 1
'
&1 0
(III.9)
a % i.b '
1 0
a %
0 1
b , where i '
0 1
(III.10)
and interpret a + i.0 as the mapping a 0
a % i.0 : Then it follows from (III.1) and (III.10) that a b c d
6 a.
(III.11)
(a%i.b)(c%i.d) '
'
a.c&b.d b.c%a.d
' (a.c&b.d) % i.(b.c%a.d) .
(III.12)
However, the same result can be obtained by using standard arithmetic operations, treating the identifier i as &1 : (a%i.b)(c%i.d) ' a.s%i 2.b.d%i.b.c%i.a.d ' (a.s&b.d)%i.(b.c%a.d) In particular, it follows from (III.9), (III.10) and (III.11) that 0 1 0 1 &1 0 (III.13)
ii '
'
' &1%i.0 6 &1
(III.14)
which can also be obtained by standard arithmetic operations, treating i as &1 and i.0 as 0. Similarly, we have a b c d 1 a.c%b.d a.c%b.d c 2%d 2 b.c&a.d c 2%d 2
(a%i.b)/(c%i.d) '
'
c 2%d 2 b.c&a.d
'
%i.
(III.15)
provided that c 2%d 2 > 0. Again, this result can also be obtained by standard arithmetic
426 operations, treating i as &1 : (a%i.b)/(c%i.d) ' a%i.b c&i.d (a%i.b)(c&i.d) a.c%b.d b.c&a.d ' ' %i. . c%i.d c&i.d (c%i.d)(c&i.d) c 2%d 2 c 2%d 2
(III.16)
The Euclidean space 2 endowed with the arithmetic operations (III.1), (III.3) and (III.5) resembles a number system, except that the numbers involved cannot be ordered. However, it is possible to measure the distance between these numbers, using the Euclidean norm:
def. a |a%i.b| ' 4 ' a 2%b 2 ' 4 5 5 5 5 b 5 5
(a%i.b)(a&i.b).
(III.17)
If the numbers in this system are denoted by (III.10), and standard arithmetic operations are applied, treating i as &1 and i.0 as 0, the results are the same as for the arithmetic operations (III.1), (III.3) and (III.5) on the elements of 2. Therefore, we may interpret (III.10) as a number, bearing in mind that this number has two dimensions if b 0. From now on I will use the standard notation for multiplication, i.e., (a%i.b)(c%i.d) instead of (III.13). The part a of a%i.b is called the real part of the complex number involved, denoted by Re(a%i.b) ' a , and b is called the imaginary part, denoted by Im(a%i.b) ' b . Moreover, a&i.b is called the complex conjugate of a%i.b and vice versa. The complex conjugate of z ' a%i.b is denoted by a bar: z ' a&i.b. It follows from (III.12) that for z ' a%i.b and w = c + i.d, zw ' z .w . Moreover, |z| ' z z . Finally, the complex number system itself is denoted by .
427 III.2. The complex exponential function Recall that for real-valued x the exponential function e x , also denoted by exp(x) , has the series representation e ' j
4 x k'0
xk . k!
(III.18)
x%y ' e xe y corresponds to the equality The property e
(x%y)k 1 ' j j j k! k'0 k'0 k! m'0

4 4 k
k k&m m x k&m y m x y ' jj m k'0 m'0 (k&m)! m!

4 k
'
j
4 k'0
x k!
m'0
j
4
(III.19) y m!
m
The first equality in (III.19) is due to the binomial expansion, and the last equality follows easily by rearranging the summation. It is easy to see that (III.19) also holds for complex valued x and y. Therefore, we can define the complex exponential function by the series expansion (III.18): e a%i.b ' j
def. 4 k'0
(a%i.b)k ak (i.b)m i m.b m a ' j ' e j j k! m! m! k'0 k! m'0 m'0

4 4 4 4
' ea j
m'0
(&1) .b (2m)!
2m
% i. j
4
(III.20) (&1) .b (2m%1)!

m 2m%1
m'0
Moreover, it follows from Taylors theorem that cos(b) ' j

4
m'0
(&1)m.b 2m (&1)m.b 2m%1 , sin(b) ' j , (2m)! (2m%1)! m'0

4
(III.21)
so that e a%i.b ' e a [cos(b) % i.sin(b)] . (III.22)
Setting a = 0, the latter equality yields the following expressions for the cosines and sinus in
428 terms of the complex exponential function: e i.b%e &i.b e i.b&e &i.b , sin(b) ' . 2 2.i
cos(b) '
(III.23)
These expressions are handy in recovering the sinus-cosines formulas: sin(a)sin(b) ' [cos(a & b) & cos(a % b)]/2 sin(a)cos(b) ' [sin(a % b) % sin(a & b)]/2 cos(a)sin(b) ' [sin(a % b) & sin(a & b)]/2 cos(a)cos(b) ' [cos(a % b) % cos(a & b)]/2 sin(a % b) ' sin(a)cos(b) % cos(a)sin(b) cos(a % b) ' cos(a)cos(b) & sin(a)sin(b) sin(a & b) ' sin(a)cos(b) & cos(a)sin(b) cos(a & b) ' cos(a)cos(b) % sin(a)sin(b
(III.24)
Moreover, it follows from (III.22) that for natural numbers n, e i.n.b ' [cos(b) % i.sin(b)]n ' cos(n.b) % i.sin(n.b) . (III.25)
This result is known as DeMoivres formula. It also holds for real numbers n, as we will see below. Finally, note that any complex number z = a + i.b can be expressed as z ' a %i .b ' |z|. a a %b
2 2
% i.
b a %b
2 2
' |z|.[cos(2) % i.sin(2)] (III.26)
' |z|.exp(i.2) , where 0 [0,1] is such that 2 ' arccos(a/ a 2%b 2) ' arcsin(b/ a 2%b 2) .
III.3. The complex logarithm Similarly to the natural logarithm ln(.), the complex logarithm log(z), z 0 , is a complex number a+i.b = log(z) such that exp(a+i.b) = z, hence it follows from (III.25) that z =
429 exp(a)[cos(b) + i.sin(b)] and consequently, since |exp(-a).z| = |cos(b)+ i.sin(b)| = cos2(b)%sin2(b) = 1,
we have that exp(a) = |z| and exp(i.b) = z/|z|. The first equation has a unique solution, a = ln(|z|), as long as z 0. The second equation reads as cos(b) % i.sin(b) ' (Re(z) % i.Im(z))/|z|, (III.27)
hence cos(b) = Re(z)/|z|, sin(b) = Im(z)/|z|, so that b = arctan(Im(z)/Re(z)). However, equation (III.27) also holds if we add or subtract multiples of B to or from b, because tan(b) = tan(b+m.B) for arbitrary integers m, hence log(z) ' ln(|z|) % i.[arctan(Im(z)/Re(z)) % m] , m ' 0,1,2,3,..... Therefore, the complex logarithm is not uniquely defined. The imaginary part of (III.28) is usually denoted by arg(z) ' arctan(Im(z)/Re(z)) % m , m ' 0,1,2,3,..... (III.29) (III.28)
It is the angle in radians indicated in Figure 1, eventually rotated multiples of 180 degrees clockwise or anticlockwise:
Figure III.1: arg(z)
430 Note that Im(z)/Re(z) is the tangents of the angle arg(z), hence arctan(Im(z)/Re(z)) is the angle itself. With the complex exponential function and logarithm defined, we can now define the power zw as the complex number a+i.b such that a+i.b = exp(w.log(z)), which exists if |z| > 0. Consequently, DeMoivres formula carries over to all real-valued powers n: [cos(b) % i.sin(b)]n ' e i.b n ' e i.n.b ' cos(n.b) % i.sin(n.b) . (III.30)
III.4
Series expansion of the complex logarithm For the case x 0 , |x| < 1, it follows from Taylors theorem that ln(1+x) has the series
representation ln(1%x) ' j (&1)k&1x k/k .

4 k'1
(III.31)
The question I will address now is whether this series representation carries over if we replace x by i.x, because this will yield a useful approximation of exp(i.x) which plays a key role in proving central limit theorems for dependent random variables.1 See Chapter 7. If (III.31) carries over we can write, for arbitrary integers m, log(1%i.x) ' j (&1)k&1i kx k/k % i.m
4 k'1 4
' j (&1)2k&1i 2kx 2k/(2k) % j (&1)2k&1&1i 2k&1x 2k&1/(2k&1) % i.m

4 k'1 k'1
(III.32)
' j (&1)k&1x 2k/(2k) % i.j (&1)k&1x 2k&1/(2k&1)% i.m

4 4 k'1 k'1
On the other hand, it follows from (III.28) that log(1%i.x) '

1 2
ln(1%x 2) % i.[arctan(x) % m] .
(III.33)
431 Therefore, we need to verify that for x 0 , |x| < 1, 1 ln(1%x 2) ' j (&1)k&1x 2k/(2k) 2 k'1
4
(III.34)
and arctan(x) ' j (&1)k&1x 2k&1/(2k&1).

4 k'1
(III.35)
Equation (III.34) follows from (III.31) by replacing x with x2. Equation (III.35) follows from d 1 k&1 2k&1 /(2k&1) ' j (&1)k&1x 2k&2 ' j &x 2 k ' j (&1) x dx k'1 k'1 k'0 1%x 2
4 4 4
(III.36)
and the facts that arctan(0) = 0 and darctan(x) 1 ' . dx 1%x 2 Therefore, the series representation (III.32) is true.
(III.37)
III.5. Complex integration In probability and statistics we encounter complex integrals mainly in the form of characteristic functions, which for absolutely continuous random variables are integrals over complex-valued functions with real-valued arguments. Such functions take the form f(x) ' (x) % i.(x) , x 0 , (III.38)
where N and R are real-valued functions on . Therefore, we may define the (Lebesgue) integral of f over an interval [a,b] simply as
432 f(x)dx ' (x)dx % i. (x)dx , m m m

a a a b b b
(III.39)
provided of course that the latter two integrals are defined. Similarly, if is a probability measure on the Borel sets in k and Re(f(x)) and Im(f(x)) are Borel measurable real functions on k, then f(x)d(x) ' Re(f(x))d(x) % i. Im(f(x))d(x) , m m m provided that the latter two integrals are defined. Integrals of complex-valued functions of complex variables are much trickier, though. See for example Ahlfors (1966). However, these types of integrals have limited applicability in econometrics, and are therefore not discussed here. (III.40)
Endnote 1. For x 0 with |x| < 1, exp(i.x) ' (1%i.x)exp(&x 2/2 % r(x)) , where |r(x)| # |x|3.
433
References
Ahlfors, L. V. (1966): Complex Analysis. New York: McGraw-Hill. Bernanke, B. S. (1986): "Alternative Explanations of the Money-Income Correlation", CarnegieRochester Conference Series on Public Policy 25, 49-100. Bierens, H. J. (1994): Topics in Advanced Econometrics: Estimation, Testing, and Specification of Cross-Section and Time Series Models, Cambridge, UK: Cambridge University Press. Billingsley, P. (1986): Probability and Measure. New York: John Wiley. Box, G. E. P., and G. M. Jenkins (1976): Time Series Analysis: Forecasting and Control. San Francisco: Holden-Day. Chung, K. L. (1974): A Course in Probability Theory (Second edition). New York: Academic Press. Davidson, J. (1994): Stochastic Limit Theory. Oxford, UK: Oxford University Press. Etemadi, N. (1981): "An Elementary Proof of the Strong Law of Large Numbers", Zeitschrift fur Wahrscheinlichkeitstheorie und Verwandte Gebiete 55, 119-122. Fuller, W. A. (1996): Introduction to Statistical Time Series. New York: John Wiley. Gallant, A. R. (1997): An Introduction to Econometric Theory. Princeton, NJ.: Princeton University Press. Jennrich, R. I. (1969): Asymptotic Properties of Non-Linear Least Squares Estimators, Annals of Mathematical Statistics 40, 633-643. McLeish, D. L. (1974): Dependent Central Limit Theorems and Invariance Principles, Annals of Probability 2, 620-628.
434 McLeish, D. L. (1975): A Maximal Inequality and Dependent Strong Laws, Annals of Probability 3, 329-339. Press, W. H., B. P. Flannery, S. A Teukolsky, and W. T. Vetterling (1989): Numerical Recipes (FORTRAN Version), Cambridge, UK: Cambridge University Press. Royden, H.L. (1968): Real Analysis. London: Macmillan. Sims, C.A. (1980): "Macroeconomics and Reality", Econometrica 48, 1-48. Sims, C.A. (1982): "Policy Analysis with Econometric Models", Brookings Papers on Economics Activity 1, 107-152. Sims, C.A. (1986): "Are Forecasting Models Usable for Policy Analysis?", Federal Reserve Bank of Minneapolis Quarterly Review, 1-16. Stokey, N.L., R.E. Lucas, and E. Prescott (1989): Recursive Methods in Economic Dynamics. Cambridge, MA: Harvard University Press. Tobin, J. (1958): ''Estimation of Relationships for Limited Dependent Variables'', Econometrica 26, 24-36. Wold, H. (1938): A Study in the Analysis of Stationary Time Series, Upsala, Sweden: Almqvist and Wiksell.

Bierens - Statistical Foundations of Econometrics

Uploaded by

Copyright:

Available Formats

Bierens - Statistical Foundations of Econometrics

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bierens - Statistical Foundations of Econometrics

Uploaded by

Copyright:

Available Formats

1

Introduction to the Mathematical and Statistical Foundations of Econometrics

4 2.8.2 Characteristic functions 2.9. Exercises

Appendix: 2.A. Uniqueness of characteristic functions

Appendix: 3.A. Proof of Theorem 3.12

Appendices: 4.A: 4.B: Tedious derivations Proof of Theorem 4.4

6 5.7.3 Hypotheses testing 5.8. Exercises

Appendix: 5.A. Proof of Theorem 5.8

Appendix: 7.A. Hilbert spaces

10 III.5. Complex integration

Chapter 1 Probability and Measure

k (50 & j) ' k k '

50! . (50 & 6)!

50! ' 15,890,700. 6!(50&6)!

objects without replacement is: n k

17 n 0 ' n n ' n! 1 ' ' 1. 0!n! 0!

for n $ 2 , k ' 1,....,n&1 .

(1.8) (1.9) (1.10)

if 0 # k # min(n,K) , P({k}) ' 0 elsewhere ,

p1(r,n) ' 1 & min pr(n,K),

(j'1(K&k%j) (j'1 (N&K&n%k%j)

K k j K n k j n&k & % (j'1 1& & % % N N N N N N N n j n (j'1 1& % N N

p k(1&p)n&k if N 6 4 and K/N 6 p .

k n! 1 & /n k k! n (n&k)! 1 & /n

because n! n k(n&k)! (j'1(n&k%j)

k j k ' (j'1 1& % n n

6 (j'11 ' 1 for n 6 4 ,

1 & /n and 1 & /n

satisfies the conditions (1.8), (1.9), and (1.10).

if _q'1^n'1_m'nBq,m 0 . In order to guarantee this, we need to require that is a -algebra.

and probability measures. Our first result is trivial:

30 is not a & algebra.

Proof: Exercise. Similarly, we have

_j'1Aj 0 . A collection of subsets of a nonempty set is a & algebra if it satisfies

where is a possibly uncountable index set. Then ' _0 is a & algebra.

Theorem 1.5: Let , 0 , be a collection of & algebras of subsets of a given set ,

containing . Then = _0 is the smallest & algebra containing .

However, it is always possible to extend ^0 to a & algebra, often in various ways,

Theorem 1.6: B = ({(&4,a] : a 0 }) .

Proof: Let ( ' {(&4 , a] : a 0 } . (a) (1.19)

0 ( ) . ~(a,b] ' (&4 , a]^(b , 4) ' A^B (

0 () , and consequently (&4 , a] ' _4 countable union of sets in , hence A m'1Am = m

measures. Here we list only the most important ones.

(f) If An e An%1 for n ' 1,2,..., then P(An) 9 P(_n'1An) ,

(g) P(^n'1An) # 'n'1P(An) .

Since the Bj s are disjoint, it follows from axiom (1.10) that

Part (e) follows now from the fact that 'j'n%1P(Bj) 9 0 .

P(^j'1Aj) ' 'j'1P(Bj) ' 'j'1P(Bj) % 'j'n%1P(Bj) ' P(An) % 'j'n%1P(Bj) .

union of sets in the algebra 0 : A d ^j'1Aj , where Aj 0 0 , hence the probability of A is

Note that it is not required in (1.21) that ^j'1Aj 0 0 .

'j'1((a j , bj)) '

'j'1(bj & aj) .

42 (B) ' inf

'j'1 i'1(ai,j , bi,j) '

'j'1 (i'1(bi,j & ai,j) .

inf f(x) (Bm) ,

P(X ' 1) P(X ' 0)

' P({H , T}) '

_j'1{0 : Xj() # xj} ' {0 : X() 0 j'1(&4 , xj]}

g(x) & g(x() 6 0