Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
STAT 6710 Mathematical Statistics I Fall Semester 1999 Dr. Jurgen Symanzik Utah State University Department of Mathematics and Statistics 3900 Old Main Hill Logan, UT 84322{3900 Tel.: (435) 797{0696 FAX: (435) 797{1822 e-mail: [email protected] Contents Acknowledgements 1 1 Axioms of Probability 1 1.1 1.2 1.3 1.4 {Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Manipulating Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Combinatorics and Counting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Conditional Probability and Independence . . . . . . . . . . . . . . . . . . . . . . . . 16 2 Random Variables 2.1 2.2 2.3 2.4 23 Measurable Functions . . . . . . . . . . . . . . Probability Distribution of a Random Variable Discrete and Continuous Random Variables . . Transformations of Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Moments and Generating Functions 3.1 3.2 3.3 3.4 3.5 42 Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Generating Functions . . . . . . . . . . . . . . . . . . . . . . . . . Complex{Valued Random Variables and Characteristic Functions Probability Generating Functions . . . . . . . . . . . . . . . . . . Moment Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Random Vectors 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 23 27 31 36 42 51 57 68 69 72 Joint, Marginal, and Conditional Distributions Independent Random Variables . . . . . . . . . Functions of Random Vectors . . . . . . . . . . Order Statistics . . . . . . . . . . . . . . . . . . Multivariate Expectation . . . . . . . . . . . . Multivariate Generating Functions . . . . . . . Conditional Expectation . . . . . . . . . . . . . Inequalities and Identities . . . . . . . . . . . . 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 77 82 91 93 98 104 106 5 Particular Distributions 113 5.1 Multivariate Normal Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 5.2 Exponential Familty of Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 6 Limit Theorems 122 6.1 Modes of Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 6.2 Weak Laws of Large Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 6.3 Strong Laws of Large Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 2 Acknowledgements I would like to thank my students, Hanadi B. Eltahir, Rich Madsen, and Bill Morphet, who helped in typesetting these lecture notes using LATEX and for their suggestions how to improve some of the material presented in class. In addition, I particularly would like to thank Mike Minnotte and Dan Coster, who previously taught this course at Utah State University, for providing me with their lecture notes and other materials related to this course. Their lecture notes, combined with additional material from Rohatgi (1976) and other sources listed below, form the basis of the script presented here. The textbook required for this class is: Rohatgi, V. K. (1976): An Introduction to Probability Theory and Mathematical Statistics, John Wiley and Sons, New York. A Web page dedicated to this class is accessible at: http://www.math.usu.edu/~symanzik/teaching/1999_stat6710/stat6710.html This course closely follows Rohatgi (1976) as described in the syllabus. Additional material originates from the lectures from Professors Hering, Trenkler, and Gather I have attended while studying at the Universitat Dortmund, Germany, the collection of Masters and PhD Preliminary Exam questions from Iowa State University, Ames, Iowa, and the following textbooks: Bandelow, C. (1981): Einfuhrung in die Wahrscheinlichkeitstheorie, Bibliographisches Institut, Mannheim, Germany. Casella, G., and Berger, R. L. (1990): Statistical Inference, Wadsworth & Brooks/Cole, Pacic Grove, CA. Fisz, M. (1989): Wahrscheinlichkeitsrechnung und mathematische Statistik, VEB Deutscher Verlag der Wissenschaften, Berlin, German Democratic Republic. Kelly, D. G. (1994): Introduction to Probability, Macmillan, New York, NY. Mood, A. M., and Graybill, F. A., and Boes, D. C. (1974): Introduction to the Theory of Statistics (Third Edition), McGraw-Hill, Singapore. Parzen, E. (1960): Modern Probability Theory and Its Applications, Wiley, New York, NY. Searle, S. R. (1971): Linear Models, Wiley, New York, NY. 3 Additional denitions, integrals, sums, etc. originate from the following formula collections: Bronstein, I. N. and Semendjajew, K. A. (1985): Taschenbuch der Mathematik (22. Auage), Verlag Harri Deutsch, Thun, German Democratic Republic. Bronstein, I. N. and Semendjajew, K. A. (1986): Erganzende Kapitel zu Taschenbuch der Mathematik (4. Auage), Verlag Harri Deutsch, Thun, German Democratic Republic. Sieber, H. (1980): Mathematische Formeln | Erweiterte Ausgabe E, Ernst Klett, Stuttgart, Germany. Jurgen Symanzik, January 18, 2000 4 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 2: Wednesday 9/1/1999 Scribe: Jurgen Symanzik 1 Axioms of Probability 1.1 {Fields Let be the sample space of all possible outcomes of a chance experiment. Let ! 2 (or x 2 ) be any outcome. Example: Count # of heads in n coin tosses. = f0; 1; 2; : : : ; ng. Any subset A of is called an event. For each event A , we would like to assign a number (i.e., a probability). Unfortunately, we cannot always do this for every subset of . Instead, we consider classes of subsets of called elds and {elds. Denition 1.1.1: A class L of subsets of is called a eld if 2 L and L is closed under complements and nite unions, i.e., L satises (i) 2 L (ii) A 2 L =) AC 2 L (iii) A; B 2 L =) A [ B 2 L Since C = , (i) and (ii) imply 2 L. Therefore, (i)': 2 L [can replace (i)]. Recall De Morgan's Laws: [ A2A \ A=( A2A AC )C and 1 \ A2A A=( [ A2A AC )C : Note: So (ii), (iii) imply (iii)': A; B 2 L =) A \ B 2 L [can replace (iii)]. Proof: A; B 2 L =(ii)) AC ; B C 2 L =(iii)) (AC [ B C ) 2 L =(ii)) (AC [ B C )C 2 L =DM )A\B 2L Denition 1.1.2: A class L of subsets of is called a {eld (Borel eld, {algebra) if it is a eld and closed under countable unions, i.e., 1 [ (iv) fAn g1 2 L = ) An 2 L. n=1 n=1 Note: (iv) implies (iii) by taking An = for n 3. Example 1.1.3: For some , let L contain all nite and all conite sets (A is conite if AC is nite). Then L is a eld. But L is a {eld i (if and only if) is nite. Example: 1 [ = Z . Take An = fng, each nite, so An 2 L. But An = Z + 62 L, since the set is not nite (it is innite) and also not conite (( 1 [ n=1 An )C n=1 ; = Z0 is innite, too). Question: Does this construction work for = Z + ?? The largest {eld in is the power set P( ) of all subsets of . The smallest {eld is L = f; g. Terminology: A set A 2 L is said to be \measurable L". 2 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 3: Friday 9/3/1999 Scribe: Jurgen Symanzik We often begin with a class of sets, say a, which may not be a eld or a {eld. Denition 1.1.4: The {eld generated by a, (a), is the smallest {eld containing a, or the intersection of all {elds containing a. Note: (i) Such {elds containing a always exist (e.g., P( )), and (ii) the intersection of an arbitrary # of {elds is always a {eld. Proof: \ (ii) Suppose L = L . We have to show that conditions (i) and (ii) of Def. 1.1.1 and (iv) of Def. 1.1.2 are fullled: (i) 2 L 8 =) 2 L (ii) Let A 2 L =) A 2 L 8 =) AC 2 L 8[ =) AC 2 L [ (iv) Let An 2 L 8n =) An 2 L 8 8n =) An 2 L 8 =) An 2 L n n Example 1.1.5: = f0; 1; 2; 3g; a = ff0gg; b = ff0g; f0; 1gg. What is (a)? (a): must include ; ; f0g also: f1; 2; 3g by 1.1.1 (ii) Since all unions are included, we have (a) = f ; ; f0g; f1; 2; 3gg What is (b)? (b): must include ; ; f0g; f0; 1g also: f1; 2; 3g; f2; 3g by 1.1.1 (ii) f0; 2; 3g by 1.1.1 (iii) 3 f1g by 1.1.1 (ii) Since all unions are included, we have (b) = f ; ; f0g; f1g; f0; 1g; f2; 3g; f0; 2; 3g; f1; 2; 3gg If is nite or countable, we will usually use L = P( ). If j j= n < 1, then j L j= 2n . If is uncountable, P( ) may be too large to be useful and we may have to use some smaller {eld. Denition 1.1.6: If = IR, an important special case is the Borel {eld, i.e., the {eld generated from all half{open intervals of the form (a; b], denoted B or B1 . The sets of B are called Borel sets. The Borel {eld on IRd (Bd ) is the {eld generated by d{dimensional rectangles of the form f(x1 ; x2 ; : : : ; xd) j ai < xi bi ; i = 1; 2; : : : ; dg. Note: 1 \ B contains all points: fxg = (x ; n1 ; x] n=1 closed intervals: [x; y] = (x; y] + fxg = (x; y] [ fxg open intervals: (x; y) = (x; y] ; fyg = (x; y] \ fygC 1 [ and semi{innite intervals: (x; 1) = (x; x + n] n=1 We now have a measurable space ( ; L). We next dene a probability measure P () on ( ; L) to obtain a probability space ( ; L; P ). Denition 1.1.7: Kolmogorov Axioms of Probability A probability measure (pm), P , on ( ; L) is a set function P : L ! IR satisfying (i) 0 P (A) 8A 2 L (ii) P ( ) = 1 (iii) If fAn g1 n=1 are disjoint sets in L and 1 [ n=1 An 2 L, then P ( Note: 1 [ n=1 1 [ n=1 An ) = 1 X n=1 P (An ). An 2 L holds automatically if L is a {eld but it is needed as a precondition in the case that L is just a eld. Property (iii) is called countable additivity. 4 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 4: Wednesday 9/8/1999 Scribe: Jurgen Symanzik, Bill Morphet 1.2 Manipulating Probability Theorem 1.2.1: For P a pm on ( ; L), it holds: (i) P () = 0 (ii) P (AC ) = 1 ; P (A) 8A 2 L (iii) P (A) 1 8A 2 L (iv) P (A [ B ) = P (A) + P (B ) ; P (A \ B ) 8A; B 2 L (v) If A B , then P (A) P (B ). Proof: (i) An ; 8n ) 1 [ n=1 A n = ; 2 L. Ai \ Aj = ; \ ; = ; 8i; j ) An are disjoint 8n. P (;) = P ( 1 [ n=1 :1:7(iii) An ) Def 1= 1 X n=1 P (A n ) = This can only hold if P (;) = 0. 1 X n=1 P (;) (ii) A1 A; A2 AC ; An ; 8n 3. 1 [ An = A1 [ A2 [ 1 [ An = A1 [ A2 [ ;. n=1 n=3 A1 \ A2 = A1 \ ; = A2 \ ; = ; ) A1 ; A2 ; ; are disjoint. = 1 = P ( ) 1 [ = P( Def 1:1:7(iii) 1 X = n=1 n=1 5 An ) P (An ) P (A1 ) + P (A2 ) + = Th1:2:1(i) P (A1 ) + P (A2 ) P (A) + P (AC ) = = 1 X n=3 P (An) =) P (AC ) = 1 ; P (A) 8A 2 L. (iii) By Th. 1.2.1 (ii) P (A) = 1 ; P (AC ) =) P (A) 1 8A 2 L since P (AC ) 0 by Def 1.1.7 (i) (iv) A [ B = (A \ B C ) [ (A \ B ) [ (B \ AC ). So, (A [ B ) can be written as a union of disjoint sets (A \ B C ); (A \ B ); (B \ AC ): =) P (A [ B ) = Def:1:1:7(iii) = = = = P ((A \ B C ) [ (A \ B ) [ (B \ AC )) P (A \ B C ) + P (A \ B ) + P (B \ AC ) P (A \ B C ) + P (A \ B ) + P (B \ AC ) + P (A \ B ) ; P (A \ B ) (P (A \ B C ) + P (A \ B )) + (P (B \ AC ) + P (A \ B )) ; P (A \ B ) P (A) + P (B ) ; P (A \ B ) (v) B = (B \ AC ) [ A where (B \ AC ) and A are disjoint sets. :1:7(iii) P (B ) = P ((B \ AC ) [ A) Def 1= P (B \ AC ) + P (A) =) P (A) = P (B ) ; P (B \ AC ) =) P (A) P (B ) since P (B \ AC ) 0 by Def 1.1.7 (i) Theorem 1.2.2: Principle of Inclusion{Exclusion Let A1 ; A2 ; : : : ; An 2 L. Then P( [n k=1 Ak ) = n X k=1 P (Ak ) ; n X k1 <k2 P (Ak1 \ Ak2 )+ n X k1 <k2 <k3 P (Ak1 \ Ak2 \ Ak3 ) ; : : : +(;1)n+1 P ( k=1 Proof: n = 1 is trivial n = 2 is Theorem 1.2.1 (iv) use induction for higher n (Homework) Theorem 1.2.3: Bonferroni's Inequality Let A1 ; A2 ; : : : ; An 2 L. Then n X i=1 P (Ai ) ; X i<j [n n X i=1 i=1 P (Ai \ Aj ) P ( Ai ) 6 \n P (Ai ) Ak ) Proof: Right side induction base: For n = 1, Th. 1.2.3 right side evaluates to P (A1 ) P (A1 ), which is true. For n = 2, Th. 1.2.3 right side evaluates to P (A1 [ A2 ) P (A1 ) + P (A2 ). P (A1 [ A2 ) Th:1:=2:1(iv) P (A1 ) + P (A2 ) ; P (A1 \ A2 ) P (A1) + P (A2 ) since P (A1 \ A2 ) 0 by Def. 1.1.7 (i). This establishes the induction base for the right side of Th. 1.2.3. Right side induction step assumes Th. 1.2.3 right side is true for n and shows that it is true for n + 1: P( n[ +1 i=1 Ai) = Th:1:2:1(iv) = Def:1:1:7(i) I:B: = [n P (( Ai ) [ An+1 ) i=1 [n [n i=1 [n i=1 P ( Ai ) + P (An+1) ; P (( Ai) \ An+1) P ( Ai ) + P (An+1 ) i=1 n X i=1 nX +1 i=1 P (Ai ) + P (An+1) P (Ai ) Left side induction base: For n = 1, Th. 1.2.3 left side evaluates to P (A1 ) P (A1 ), which is true. For n = 2, Th. 1.2.3 left side evaluates to P (A1 ) + P (A2 ) ; P (A1 \ A2 ) P (A1 [ A2 ), which is true by Th. 1.2.1 (iv). For n = 3, Th. 1.2.3 left side evaluates to P (A1 ) + P (A2 ) + P (A3 ) ; P (A1 \ A2 ) ; P (A1 \ A3 ) ; P (A2 \ A3 ) P (A1 [ A2 [ A3 ). 7 This holds since P (A1 [ A2 [ A3 ) = P ((A1 [ A2 ) [ A3 ) Th:1:2:1(iv) = P (A1 [ A2 ) + P (A3 ) ; P ((A1 [ A2 ) \ A3 ) = P (A1 [ A2 ) + P (A3 ) ; P ((A1 \ A3 ) [ (A2 \ A3 )) Th:1:2:1(iv) P (A1 ) + P (A2 ) ; P (A1 \ A2 ) + P (A3 ) ; P (A1 \ A3 ) ; P (A2 \ A3 ) +P ((A1 \ A3 ) \ (A2 \ A3 )) = P (A1 ) + P (A2 ) + P (A3 ) ; P (A1 \ A2 ) ; P (A1 \ A3 ) ; P (A2 \ A3 ) + P (A1 \ A2 \ A3 ) = Def 1:1:7(i) P (A1 ) + P (A2 ) + P (A3 ) ; P (A1 \ A2) ; P (A1 \ A3 ) ; P (A2 \ A3 ) This establishes the induction base for the left side of Th. 1.2.3. Left side induction step assumes Th. 1.2.3 left side is true for n and shows that it is true for n + 1: n[ +1 P( i=1 Ai) = = left I:B: = [n P (( Ai ) [ An+1 ) i=1 [n i=1 n X i=1 nX +1 i=1 +1 Th:1:2:3 right side nX = [n P ( Ai) + P (An+1 ) ; P (( Ai ) \ An+1 ) i=1 nX +1 i=1 P (Ai ) ; P (Ai ) ; P (Ai ) ; P (Ai ) ; n X i<j n X i=1 [n P (Ai \ Aj ) + P (An+1 ) ; P (( Ai ) \ An+1) i<j n X i<j nX +1 [n i=1 P (Ai \ Aj ) ; P ( (Ai \ An+1 )) i=1 P (Ai \ Aj ) ; i<j 8 P (Ai \ Aj ) n X i=1 P (Ai \ An+1 ) Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 5: Friday 9/10/1999 Scribe: Jurgen Symanzik, Rich Madsen Theorem 1.2.4: Boole's Inequality Let A; B 2 L. Then (i) P (A \ B ) P (A) + P (B ) ; 1 (ii) P (A \ B ) 1 ; P (AC ) ; P (B C ) Proof: Homework Denition 1.2.5: Continuity of sets For a sequence of sets fAn g1 n=1 ; An 2 L and A 2 L, we say (i) An " A if A1 A2 A3 : : : and A = (ii) An # A if A1 A2 A3 : : : and A = 1 [ n=1 1 \ n=1 An . An . Theorem 1.2.6: If fAn g1 n=1 ; An 2 L and A 2 L, then nlim !1 P (An ) = P (A) if 1.2.5 (i) or 1.2.5 (ii) holds. Proof: Part (i): Assume that 1.2.5 (i) holds. Let B1 = A1 and Bk = Ak ; Ak;1 = Ak \ ACk;1 8k 2 By construction, Bi \ Bj = ; for i = 6 j 1 1 [ [ It is A = n=1 An = and also An = [n i=1 n=1 Bn Ai = [n i=1 Bi 9 P(A) = P( 1 [ k=1 Bk ) By Def.=1.1.7 (iii) again by Def. 1.1.7 (iii) k=1 n [ P (Bk ) = nlim !1[ n X k=1 n [ P (Bk )] nlim !1[P (k=1 Bk )] = nlim !1[P (k=1 Ak )] = nlim !1 P (An ) [n is possible since An = Ak k=1 = The last step 1 X Part (ii): Assume that 1.2.5 (ii) holds. 1 1 \ [ Then, AC1 AC2 AC3 : : : and AC = ( An )C DeMorgan = ACn P (AC ) By Part (i) = C nlim !1 P (An ) n=1 n=1 C So 1 ; P (AC ) = 1 ; nlim !1 P (An ) C =) P (A) = nlim !1(1 ; P (An )) = nlim !1 P (An ) Theorem 1.2.7: (i) Countable unions of probability 0 sets have probability 0. (ii) Countable intersections of probability 1 sets have probability 1. Proof: Part (i): Let fAn g1 n=1 2 L, P (An ) = 0 8n 1 1 1 [ X By K.A.P. (i) By Bonferroni's Inequality X 0 P ( An ) P (An ) = 0 = 0 Therefore P ( n=1 1 [ n=1 n=1 An ) = 0 Part (ii): Let fAn g1 n=1 2 L, P (An ) = 1 8n =) by Th. 1.2.1 (ii) n=1 1.2.7 (i) P (ACn ) = 0 8n by Th.=) P( 1 [ n=1 ACn ) = 0 De=Morgan ) P( 1 \ n=1 10 An ) = 1 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 6: Monday 9/13/1999 Scribe: Jurgen Symanzik, Hanadi B. Eltahir 1.3 Combinatorics and Counting For now, we restrict ourselves to sample spaces containing a nite number of points. X Let = f!1 ; : : : ; !n g and L = P( ). For any A 2 L; P (A) = P (!j ). !j 2A Denition 1.3.1: We say the elements of are equally likely (or occur with uniform probability) if P (!j ) = n1 8j = 1; : : : ; n. Note: !j in A If this is true, P (A) = number number !j in . Therefore, to calculate such probabilities, we just need to be able to count elements accurately. Theorem 1.3.2: Fundamental Theorem of Counting If we wish to select one element (a1 ) out of n1 choices, a second element (a2 ) out of n2 choices, and so on for a total of k elements, there are n1 n2 n3 : : : nk ways to do it. Proof: (By Induction) Induction Base: k = 1: trivial k = 2: n1 ways to choose a1 . For each, n2 ways to choose a2 . Total # of ways = n| 2 + n2 + {z : : : + n2} = n1 n2. n1 times Induction Step: Suppose it is true for (k ; 1). We show that it is true for k = (k ; 1) + 1. There are n1 n2 n3 : : : nk;1 ways to select one element (a1 ) out of n1 choices, a second element (a2 ) out of n2 choices, and so on, up to the (k ; 1)th element (ak;1 ) out of nk;1 choices. For each of these n1 n2 n3 : : : nk;1 possible ways, we can select the kth element (ak ) out of nk choices. Thus, the total # of ways = (n1 n2 n3 : : : nk;1 ) nk . 11 Denition 1.3.3: For positive integer n, we dene n factorial as n! = n (n ; 1) (n ; 2) : : : 2 1 = n (n ; 1)! and 0! = 1. Denition 1.3.4: For nonnegative integers n r, we dene the binomial coecient (read as n choose r) as n r = r!(nn;! r)! = n (n ; 1) 1(n 2; 31) : :: :: : r(n ; r + 1) : Note: Most counting problems consist of drawing a xed number of times from a set of elements (e.g., f1; 2; 3; 4; 5; 6g). To solve such problems, we need to know (i) the size of the set, n; (ii) the size of the sample, r; (iii) whether the result will be ordered (i.e., is f1; 2g dierent from f2; 1g); and (iv) whether the draws are with replacement (i.e, can results like f1; 1g occur?). Theorem 1.3.5: The number of ways to draw r elements from a set of n, if (i) ordered, without replacement, is (n;n!r)! ; (ii) ordered, with replacement, is nr ; (iii) unordered, without replacement, is r!(nn;! r)! = rn ; 1)! = n+rr;1 . (iv) unordered, with replacement, is (rn!(+nr;;1)! Proof: (i) n choices to select 1st n ; 1 choices to select 2nd .. . n ; r + 1 choices to select rth By Theorem 1.3.2, there are n (n ; 1) : : : (n ; r +1) = n(n;1):::(n(;nr;)!r+1)(n;r)! = (n;n!r)! ways to do so. 12 Corollary: The number of permutations of n objects is n!. (ii) n choices to select 1st n choices to select 2nd .. . n choices to select rth By Theorem 1.3.2, there are n| n {z: : : n} = nr ways to do so. r times (iii) We know from (i) above that there are (n;n!r)! ways to draw r elements out of n elements without replacement in the ordered case. However, for each unordered set of size r, there are r! related ordered sets that consist of the same elements. Thus, there are (n;n!r)! r1! = rn ways to draw r elements out of n elements without replacement in the unordered case. (iv) There is no immediate direct way to show this part. We have to come up with some extra motivation. We assume that there are (n ; 1) walls that separate the n bins of possible outcomes and there are r markers. If we shake everything, there are (n ; 1+ r)! permutations to arrange these (n ; 1) walls and r markers according to the Corollary. Since the r markers are indistinguishable and the (n ; 1) markers are also indistinguishable, we have to divide the number of permutations by r! to get rid of identical permutations where only the markers are changed and by (n ; 1)! to get rid of identical permutations where only the walls are r)! ;n+r;1 changed. Thus, there are (rn!(;n1+ ;1)! = r ways to draw r elements out of n elements with replacement in the unordered case. Theorem 1.3.6: The Binomial Theorem If n is a non{negative integer, then n X (1 + x)n = r=0 Proof: (By Induction) Induction Base: ! 0 0! X 0 r 0 n = 0: 1 = (1 + x) = x = 0 x0 = 1 r r=0 ! n xr r ! ! 1 1! X 1 1 r 0 1 n = 1: (1 + x) = x = 0 x + 1 x1 = 1 + x r r=0 13 Induction Step: Suppose it is true for k. We show that it is true for k + 1. (1 + x)k+1 = (1 + x)k (1 + x) IB = = = = () = = k k! ! X r r=0 r k k! X x (1 + x) k k! X xr + xr+1 r r r=0 r=0 ! X " ! !# ! k k x0 + k + k k xr + k xk+1 0 r r ; 1 ! r=1 X " ! !# ! k + 1 x0 + k k + k k + 1 xr + k + 1 xk+1 0 r r ; 1 =1 ! rX ! ! k + 1 x0 + k k + 1 xr + k + 1 xk+1 0 r k+1 r=1 ! kX +1 k + 1 xr r r=0 (*) Here we use Theorem 1.3.8 (i). Since the proof of Theorem 1.3.8 (i) only needs algebraic transformations without using the Binomial Theorem, part (i) of Theorem 1.3.8 can be applied here. Corollary 1.3.7: For a non{negative integer n, it holds: n n (i) 0 + 1 + : : : + nn = 2n n n n n (ii) 0 ; 1 + 2 ; 3 + : : : + (;1)n nn = 0 n n n (iii) 1 1 + 2 2 + 3 3 + : : : + n nn = n2n;1 n n n (iv) 1 1 ; 2 2 + 3 3 + : : : + (;1)n;1 n nn = 0 Proof: Use the Binomial Theorem: (i) Let x = 1. Then n X 2n = (1 + 1)n Bin:Th: = r=0 14 ! ! n n n 1r = X r r=0 r (ii) Let x = ;1. Then (iii) n d (1 + x)n = d X dx dx r=0 ! n X Bin:Th: n 0 = (1 + (;1)) = r=0 n xr r n X ! n (;1)r r ! =) n(1 + x)n;1 = r nr xr;1 r=1 Substitute x = 1, then ! n X n n2n;1 = n(1 + 1)n;1 = r r r=1 (iv) Substitute x = ;1 in (iii) above, then 0 = n(1 + (;1))n;1 since for P a = 0 also P(;a ) = 0. i i ! Note: A useful extension for the binomial coecient for n < r is n n (n ; 1) : : : 0 : : : (n ; r + 1) r = = 0: 1 2 ::: r Theorem 1.3.8: For non{negative integers, n; m; r, it holds: n;1 (i) n;r 1 + r ; 1 = rn n m n (ii) 0 r + 1 m ! n X = r nr (;1)r = r nr (;1)r;1 r=1 r=1 n X n m m+n r ; 1 + ::: + r 0 = r n+1 (iii) r0 + r1 + r2 + : : : + rn = r + 1 Proof: Homework 15 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 7: Wednesday 9/15/1999 Scribe: Jurgen Symanzik, Bill Morphet 1.4 Conditional Probability and Independence So far, we have computed probability based only on the information that is used for a probability space ( ; L; P ). Suppose, instead, we know that event H 2 L has happened. What statement should we then make about the chance of an event A 2 L ? Denition 1.4.1: Given ( ; L; P ) and H 2 L; P (H ) > 0, and A 2 L, we dene \ H ) = P (A) P (AjH ) = P (PA(H H ) and call this the conditional probability of A given H . Note: This is undened if P (H ) = 0. Theorem 1.4.2: In the situation of Denition 1.4.1, ( ; L; PH ) is a probability space. Proof: If PH is a probability measure, it must satisfy Def. 1.1.7. (i) P (H ) > 0 and by Def. 1.1.7 (i) P (A \ H ) 0 =) PH (A) = P P(A(H\H) ) 0 8A 2 L (ii) PH ( ) = P P( (H\H) ) = PP ((HH )) = 1 (iii) Let fAn g1 n=1 be a sequence of disjoint sets. Then, PH ( 1 [ n=1 An ) Def:1:4:1 = = 16 P (( 1 [ n=1 An ) \ H ) P (H ) P( 1 [ n=1 (An \ H )) P (H ) 1 X Def:1:1:7(iii) n=1 = = Def 1:4:1 = P (An \ H ) P (H ) 1 P (A \ H ) X ( n ) P (H ) n=1 1 X n=1 PH (An ) Note: What we have done is to move to a new sample space H and a new {eld LH = L \ H of subsets A \ H for A 2 L. We thus have a new measurable space (H; LH ) and a new probability space (H; LH ; PH ). Note: From Denition 1.4.1, if A; B 2 L; P (A) > 0, and P (B ) > 0, then P (A \ B ) = P (A)P (B jA) = P (B )P (AjB ); which generalizes to Theorem 1.4.3: Multiplication Rule n\ ;1 If A1 ; : : : ; An 2 L and P ( Aj ) > 0, then j =1 \n P( j =1 Aj ) = P (A1 ) P (A2 jA1 ) P (A3 jA1 \ A2 ) : : : P (An j Proof: Homework Denition 1.4.4: A collection of subsets fAn g1 n=1 of form a partition of if (i) 1 [ n=1 An = , and (ii) Ai \ Aj = 8i 6= j , i.e., elements are pairwise disjoint. 17 n\ ;1 j =1 Aj ): Theorem 1.4.5: Law of Total Probability If fHj g1 j =1 is a partition of , and P (Hj ) > 0 8j , then, for A 2 L, P (A) = 1 X j =1 P (A \ Hj ) = 1 X j =1 P (Hj )P (AjHj ): Proof: By the Note preceding Theorem 1.4.3, the summands on both sides are equal =) the right side of Th. 1.4.5 is true. The left side proof: Hj are disjoint ) A \ Hj are disjoint A=A\ Def 1:4:4 = 1 [ A\( 1 [ j =1 Hj ) = 1 [ (A \ Hj ) j =1 :1:7(iii) =) P (A) = P ( (A \ Hj )) Def 1= j =1 1 X j =1 P (A \ Hj ) Theorem 1.4.6: Bayes' Rule Let fHj g1 j =1 be a partition of , and P (Hj ) > 0 8j . Let A 2 L and P (A) > 0. Then P (Hj )P (AjHj ) 8j: P (Hj jA) = X 1 P (Hn)P (AjHn ) n=1 Proof: P (Hj \ A) Def=1:4:1 P (A) P (Hj jA) = P (Hj ) P (AjHj ) P (Hj )P (AjHj ) . =) P (Hj jA) = P (HjP)P(A()AjHj ) Th:=1:4:5 X 1 P (Hn)P (AjHn) n=1 Denition 1.4.7: For A; B 2 L, A and B are independent i P (A \ B ) = P (A)P (B ). Note: There are no restrictions on P (A) or P (B ). If A and B are independent, then P (AjB ) = P (A) (given that P (B ) > 0) and P (B jA) = P (B ) (given that P (A) > 0). 18 If A and B are independent, then the following events are independent as well: A and B C ; AC and B ; AC and B C . Denition 1.4.8: Let A be a collection of L{sets. The events of A are pairwise independent i for every distinct A1 ; A2 2 A it holds P (A1 \ A2 ) = P (A1 )P (A2 ). Denition 1.4.9: Let A be a collection of L{sets. The events of A are mutually independent (or completely inde- \k pendent) i for every nite subcollection fAi1 ; : : : ; Aik g; Aij 2 A, it holds P ( j =1 Ai ) = j Yk j =1 P (Ai ). j Note: To check for mutually independence of n events fA1 ; : : : ; An g 2 L, there are 2n ; n ; 1 relations (i.e., all subcollections of size 2 or more) to check. Example 1.4.10: Flip a fair coin twice. = fHH; HT; TH; TT g. A1 = \H on 1st toss" A2 = \H on 2nd toss" A3 = \Exactly one H " Obviously, P (A1 ) = P (A2 ) = P (A3 ) = 12 . Question: Are A1; A2 and A3 pairwise independent and also mutually independent? P (A1 \ A2 ) = :25 = :5 :5 = P (A1 ) P (A2 ) ) A1; A2 are independent. P (A1 \ A3 ) = :25 = :5 :5 = P (A1 ) P (A3 ) ) A1; A3 are independent. P (A2 \ A3 ) = :25 = :5 :5 = P (A2 ) P (A3 ) ) A2; A3 are independent. Thus, A1 ; A2 ; A3 are pairwise independent. P (A1 \ A2 \ A3) = 0 6= :5 :5 :5 = P (A1 ) P (A2 ) P (A3 ) ) A1 ; A2 ; A3 are not mutually independent. 19 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 8: Friday 9/17/1999 Scribe: Jurgen Symanzik, Rich Madsen Example 1.4.11: (from Rohatgi, page, Example 5) r students. 365 possible birthdays for each student that are equally likely. One student at a time is asked for his/her birthday. If one of the other students hears this birthday and it matches his/her birthday, this other student has to raise his/her hand | if at least one other student raises his/her hand, the procedure is over. We are interested in pk = P (procedure terminates at the kth student) = P (a hand is rst risen when the kth student is asked for his/her birthday) It is p1 = P (at least 1 other (from r-1) students has a birthday on this particular day.) = 1 ; P (all (r-1) students have a birthday on the remaining 364 out of 365 days) 364 r;1 = 1 ; 365 p2 = P (no student has a birthday matching the rst student and at least one of the other (r-2) students has a b-day matching the second student) Let A No student has a b-day matching the 1st student Let B At least one of the other (r-2) has b-day matching 2nd So p2 = P (A \ B ) = P (A) P (B jA) = P (no student has a matching b-day with the 1st student ) P (at least one of the remaining students has a mathching b-day with the second, given that no one matched the rst.) 20 = (1 ; p1 )[1 ;" P (all (r-2) students have a b-day on the remaining 363 out of 364 days) 364 r;1 363 r;2# 1 ; 364 = 365 365 ; 1 r;1 " 363 r;2# = 1 ; 364 365 Working backwards from the book " 365P2;1 r;2+1 365 ; 2 r;2 2 ; 1 1 ; 365 ; 2 + 1 p2 = (365)2;1 1 ; 365 1 r;1 " 363 r;2# 365 = 365 1 ; 365 1 ; 364 365 ; 1 r;1 " 363 r;2# = 1 ; 364 365 # (Same as what we found before) p3 = P (No one has same b-day as rst and no one same as second, and at least one of the remaining (r ; 3) has a matching b-day with the 3rd student) Let A No one has the same b-day as the rst student Let B No one has the same b-day as the second student Let C At least one of the other (r ; 3) has the same b-day as the third students Now: p3 = P (A \ B \ C ) = P (A) P (B jA) P (C jA \ B ) 364 r;1 363 r;2 [1 ; P (all (r ; 3) students have a b-day on the remaining 362 out of 363 days] = 365 364 364 r;1 363 r;2 " 362 r;3# = 365 1 ; 363 364 362 r;3# r;1 (363)r;2 " (364) = (365)r;1 (364)r;2 1 ; 363 362 r;3# r;1 ! 363r;2 ! " 364 = 364r;2 365r;1 1 ; 363 21 364 363r;2 ! " 362 r;3# = 365 365r;2 1 ; 363 Working backwards from the book p3 = = = = " 365P3;1 3 ; 1 r;3+1 1 ; 365 ; 3 r;3 1 ; (365)3;1 365 365 ; 3 + 1 365P2 2 r;2 " 362 r;3 # 1 ; 363 (365)2 1 ; 365 (365)(364) 363 r;2 " 362 r;3 # 1 ; 363 (365)2 365 364 363 r;2 " 362 r;3# 1 ; 363 365 365 (Same as what we found before) For general pk and restrictions on r and k see Homework. 22 # 2 Random Variables 2.1 Measurable Functions Denition 2.1.1: A random variable (rv) is a set function from to IR. More formally: Let ( ; L; P ) be any probability space. Suppose X : ! IR and that X is a measurable function, then we call X a random variable. More generally: If X : ! IRk , we call X a random vector, X = (X1 (!); X2 (!); : : : ; Xk (!)). What does it mean to say that a function is measurable? Denition 2.1.2: Suppose ( ; L) and (S; B) are two measurable spaces and X : ! S is a mapping from to S. We say that X is measurable L ; B if X ;1 (B ) 2 L for every set B 2 B, where X ;1 (B ) = f! 2 : X (!) 2 B g. Example 2.1.3: Record the opinion of 50 people: \yes" (y) or \no" (n). = fAll 250 possible sequences of y/ng | HUGE ! L = P( ) X : ! S = fAll 250 possible sequences of 1 (=y) and 0 (=n)g B = P(S) X is a random vector since each element in S has a corresponding element in , for B 2 B; X ;1 (B ) 2 L = P( ). Consider X : ! S = f0; 1; 2; : : : ; 50g, where X (!) = \# of y's in !" is a more manageable random variable. A simple function, which takes only nite many values x1 ; : : : ; xk is measurable i X ;1 (xi ) 2 L 8xi. Here, X ;1 (k) = f! 2 : # 1's in sequence ! = kg is a subset of , so it is in L = P( ) 23 Example 2.1.4: Let = \innite fair coin tossing space", i.e., innite sequence of H's and T's. Let Ln be a {eld for the 1st n tosses. 1 [ Dene L = ( Ln ). n=1 Let Xn : ! IR be Xn (!) = \proportion of H's in 1st n tosses". For each n, Xn () is simple (values f0; n1 ; n2 ; : : : ; ng) and Xn;1 ( nk ) 2 Ln 8k = 0; 1; : : : ; n. Therefore, Xn;1 ( nk ) 2 L. So every random variable Xn () is measurable L ; B. Now we have a sequence of rv's fXn g1 n=1 . We 1 will show later that P (f! : Xn (!) ! 2 g) = 1, i.e., the Strong Law of Large Numbers (SLLN). 24 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 9: Monday 9/20/1999 Scribe: Jurgen Symanzik Some Technical Points about Measurable Functions 2.1.5: Suppose ( ; L) and (S; B) are measure spaces and that a collection of sets A generates B, i.e., (A) = B. Let X : ! S. If X ;1 (A) 2 L 8A 2 A, then X is measurable L ; B. This means we only have to check measurability on a basis collection A. The usage is: B on IR is generated by f(;1; x] : x 2 IRg. 2.1.6: If ( ; L); ( 0 ; L0 ), and ( 00 ; L00 ) are measure spaces and X : ! 0 and Y : 0 ! 00 are measurable, then the composition (Y X ) : ! 00 is measurable L ; L00 . 2.1.7: If f : IRi ! IRk is a continuous function, then f is measurable Bi ; Bk . 2.1.8: If fj : ! IR; j = 1; : : : k and g : IRk ! IR are measurable, then g(f1 (); : : : ; fk ()) is measurable. The usage is: g could be sum, average, dierence, product, (nite) maximums and minimums of x1 ; : : : ; xk , etc. 2.1.9: Limits: Extend the real line to [;1; 1] = IR [ f;1; 1g. We say f : ! IR is measurable L ; B if (i) f ;1 (B ) 2 L 8B 2 B, and (ii) f ;1 (;1); f ;1 (1) 2 L also. 25 2.1.10: Suppose f1 ; f2 ; : : : is a sequence of real{valued measurable functions ( ; L) ! (IR; B). Then it holds: (i) sup fn ; inf n fn; lim sup fn ; lim inf n fn , are measurable. n n (ii) If f = lim n fn exists, then f is measurable. (iii) The set f! : fn (!) convergesg 2 L. (iv) If f is any measurable function, the set f! : fn (!) ! f (!)g 2 L. 26 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 10: Wednesday 9/22/1999 Scribe: Jurgen Symanzik, Bill Morphet 2.2 Probability Distribution of a Random Variable The denition of a random variable X : ( ; L) ! (S; B) makes no mention of P . We now introduce a probability measure on (S; B). Theorem 2.2.1: A random variable X on ( ; L; P ) induces a probability measure on a space (IR; B; Q) with the probability distribution Q of X dened by Q(B ) = P (X ;1 (B )) = P (f! : X (!) 2 B g) 8B 2 B: Note: By the denition of a random variable, X ;1 (B ) 2 L 8B 2 B. Proof: If X induces a probability measure Q on (IR; B), then Q must satisfy the Kolmogorov Axioms of probability. X : ( ; L) ! (S; B). X is a rv ) X ;1 (B ) = f! : X (!) 2 B g = A 2 L 8B 2 B. (i) Q(B ) = P (X ;1 (B )) = P (f! : X (!) 2 B g) = P (A) Def:1:1:7(i) 0 8B 2 B (ii) Q(IR) = P (X ;1 (IR)) X=rv P ( ) Def:1=:1:7(ii) 1 (iii) Let fBn g1 n=1 2 B; Bi \ Bj = ; 8i 6= j . Then, 1 [ Q( n=1 1 [ Bn ) = P (X ;1 ( n=1 1 [ 1 X n=1 n=1 Bn )) (=) P ( (X ;1 (Bn))) Def:1=:1:7(iii) P (X ;1 (Bn)) = 1 X n=1 Q(Bn ) () holds since X ;1 () commutes with unions/intersections and preserves disjointedness. 27 Denition 2.2.2: A real{valued function F on (;1; 1) that is non{decreasing, right{continuous, and satises F (;1) = 0; F (1) = 1 is called a cumulative distribution function (cdf) on IR. Note: No mention of probability space or measure P in Denition 2.2.2 above. Denition 2.2.3: Let P be a probability measure on (IR; B). The cdf associated with P is F (x) = FP (x) = P ((;1; x]) = P (f! : X (!) xg) = P (X x) for a random variable X dened on (IR; B; P ). Note: F () dened as in Denition 2.2.3 above indeed is a cdf. Proof: (i) Let x1 < x2 =) (;1; x1 ] (;1; x2 ] Th:1:2:1(v) =) F (x1 ) = P (f! : X (!) x1 g) P (f! : X (!) x2 g) = F (x2 ) Thus, since x1 < x2 and F (x1 ) F (x2 ), F (:) is non{decreasing. (ii) Since F is non-decreasing, it is sucient to show that F (:) is right{continuous if for any sequence of numbers xn ! x+ (which means that xn is approaching x from the right) with x1 > x2 > : : : > xn > : : : > x : F (Xn ) ! F (X ): Let An = f! : X (!) 2 (x; xn ]g 2 L and An # ;. None of the intervals (x; xn ] contains x. As xn ! x+, the number of points ! in An diminishes until the set is empty. Formally, \n 1 \ nlim !1 An = nlim !1 i=1 Ai = n=1 An = ;. By Theorem 1.2.6 it follows that nlim !1 P (An ) = P (nlim !1 An ) = P (;) = 0. 28 It is P (An ) = P (f! : X (!) xn g) ; P (f! : X (!) xg) = F (xn ) ; F (x). =) (nlim !1 F (xn )) ; F (x) = nlim !1(F (xn ) ; F (x)) = nlim !1 P (An ) = 0 =) nlim !1 F (xn ) = F (x) =) F (x) is right{continuous. (iii) F (;n) Def:=2:2:3 P (f! : X (!) ;ng) =) F (;1) = = = = = nlim !1 F (;n) nlim !1 P (f! : X (!) ;ng) P (nlim !1f! : X (!) ;ng) P (;) 0 (iv) F (n) Def:=2:2:3 P (f! : X (!) ng) =) F (1) = nlim !1 F (n) = nlim !1 P (f! : X (!) ng) = P (nlim !1f! : X (!) ng) = P ( ) = 1 Note that (iii) and (iv) implicitly use Theorem 1.2.6. In (iii), we use An = (;1; ;n) where An An+1 and An # ;. In (iv), we use An = (;1; n) where An An+1 and An " IR. Denition 2.2.4: If a random variable X : ! IR has induced a probability measure PX on (IR; B) with cdf F (x), we say (i) rv X is continuous if F (x) is continuous in x. (ii) rv X is discrete if F (x) is a step function in x. 29 Note: There are rvs that are mixtures of continuous and discrete rvs. One such example is a truncated failure time distribution. We assume a continuous distribution (e.g., exponential) up to a given truncation point x and assign the \remaining" probability to the truncation point. Thus, a single point has a probability > 0 and F (x) jumps at the truncation point x. Denition 2.2.5: Two random variables X and Y are identically distributed i PX (X 2 A) = PY (Y 2 A) 8A 2 L. Note: Def. 2.2.5 does not mean that X (!) = Y (!) 8! 2 . For example, X = # H in 3 coin tosses Y = # T in 3 coin tosses X; Y are both Bin(3; 0:5), i.e., identically distributed, but for ! = (H; H; T ); X (!) = 2 6= 1 = Y (!), i.e., X 6= Y . Theorem 2.2.6: The following two statements are equivalent: (i) X; Y are identically distributed. (ii) FX (x) = FY (x) 8x 2 IR. Proof: (i) ) (ii): FX (x) = = byDef:2:2:5 = = = PX ((;1; x]) P (f! : X (!) 2 (;1; x]g) P (f! : Y (!) 2 (;1; x]g) PY ((;1; x]) FY (X ) (ii) ) (i): Requires extra knowledge from measure theory. 30 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 11: Friday 9/24/1999 Scribe: Jurgen Symanzik 2.3 Discrete and Continuous Random Variables We now extend Denition 2.2.4 to make our denitions a little bit more formal. Denition 2.3.1: Let X be a real{valued random variable with cdf F on ( ; L; P ). X is discrete if there exists a countable set E IR such that P (X 2 E ) = 1, i.e., P (f! : X (!) 2 E g) = 1. The points of E which have positive probability are the jump points of the step function F , i.e., the cdf of X . Dene pi = P (f! : X (!) = xi ; xi 2 E g) = PX (X = xi ) 8i 1. Then, pi 0; 1 X i=1 pi = 1. We call fpi : pi 0g the probability mass function (pmf) (also: probability frequency function) of X . Note: 1 X Given any set of numbers fpn g1 ; p 0 8 n 1 ; pn = 1, fpn g1 n n=1 n=1 is the pmf of some rv X . n=1 Note: The issue of continuous rv's and probability density functions (pdfs) is more complicated. A rv X : ! IR always has a cdf F . Whether there exists a function f such that f integrates to F and F 0 exists and equals f (almost everywhere) depends on something stronger than just continuity. Denition 2.3.2: A real{valued function F is continuous in x0 2 IR i 8 > 0 9 > 0 8x : j x ; x0 j< )j F (x) ; F (x0) j< : F is continuous i F is continuous in all x 2 R. 31 Denition 2.3.3: A real{valued function F dened on [a; b] is absolutely continuous on [a; b] i 8 > 0 9 > 0 8 nite subcollection of disjoint subintervals [ai; bi ]; i = 1; : : : ; n : n n X X (bi ; ai ) < ) j F (bi ) ; F (ai ) j< : i=1 i=1 Note: Absolute continuity implies continuity. Theorem 2.3.4: (i) If F is absolutely continuous, then F 0 exists almost everywhere. (ii) A function F is an indenite integral i it is absolutely continuous. Thus, every absolutely continuous function F is the indenite integral of its derivative F 0 . Denition 2.3.5: Let X be a random variable on ( ; L; P ) with cdf F . We say X is a continuous rv i F is absolutely continuous. In this case, there exists a non{negative integrable function f , the probability density function (pdf) of X , such that F (x) = Zx ;1 f (t)dt = P (X x): From this it follows that, if a; b 2 IR; a < b, then PX (a < X b) = F (b) ; F (a) = Zb a f (t)dt exists and is well dened. Theorem 2.3.6: Let X be a continuous random variable with pdf f . Then it holds: (i) For every Borel set B 2 B; P (B ) = Z B f (t)dt. (ii) If F is absolutely continuous and f is continuous at x, then F 0 (x) = dFdx(x) = f (x). 32 Proof: Part (i): From Denition 2.3.5 above. Part (ii): By Fundamental Theorem of Calculus. Note: As already stated in the Note following Denition 2.2.4, not every rv will fall into one of these two (or if you prefer { three {, i.e., discrete, continuous/absolutely continuous) classes. However, most rv which arise in practice will. We look at one example that is unlikely to occur in practice in the next Homework assignment. However, note that every cdf F can be written as F (x) = aFd (x) + (1 ; a)Fc(x); 0 a 1; where Fd is the cdf of a discrete rv and Fc is a continuous (but not necessarily absolute continuous) cdf. Some authors, such as Marek Fisz Wahrscheinlichkeitsrechnung und mathematische Statistik, VEB Deutscher Verlag der Wissenschaften, Berlin, 1989, are even more specic. There it is stated that every cdf F can be written as F (x) = a1 Fd (x) + a2Fc (x) + a3 Fs(x); a1 ; a2 ; a3 0; a1 + a2 + a3 = 1: Here, Fd (x) and Fc (x) are discrete and continuous cdfs (as above). Fs (x) is called a singular cdf. Singular means that Fs (x) is continuous and its derivative F 0 (x) equals 0 almost everywhere (i.e., everywhere but in those points that belong to a Borel{measurable set of probability 0). Question: Does \continuous" but \not absolutely continuous" mean \singular"? | We will (hopefully) see later: : : 33 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 12: Monday 9/27/1999 Scribe: Jurgen Symanzik, Hanadi B. Eltahir Example 2.3.7: Consider 8 > 0; x<0 > > < 1=2; x=0 F (x) = > 1=2 + x=2; 0 < x < 1 > > : 1; x1 We can write F (x) as aFd (x) + (1 ; a)Fc (x); 0 a 1. How? Since F (x) has only one jump at x = 0, it is reasonable to get started with a pmf p0 = 1 and corresponding cdf ( Fd (x) = 0; x < 0 1; x 0 Since F (x) = 0 for x < 0 and F (x) = 1 for x 1, it must clearly hold that Fc (x) = 0 for x < 0 and Fc (x) = 1 for x 1. In addition F (x) increases linearly in 0 < x < 1. A good guess would be a pdf fc (x) = 1 I(0;1) (x) and corresponding cdf 8 > < 0; x 0 Fc (x) = > x; 0 < x < 1 : 1; x 1 Knowing that F (0) = 1=2, we have at least to multiply Fd (x) by 1=2. And, indeed, F (x) can be written as F (x) = 21 Fd (x) + 12 Fc (x): Denition 2.3.8: The two{valued function IA (x) is called indicator function and it is dened as follows: IA(x) = 1 if x 2 A and IA(x) = 0 if x 62 A for any set A. 34 An Excursion into Logic When proving theorems we only used direct methods so far. We used induction proofs to show that something holds for arbitrary n. To show that a statement A implies a statement B , i.e., A ) B , we used proofs of the type A ) A1 ) A2 ) : : : ) An;1 ) An ) B where one step directly follows from the previous step. However, there are dierent approaches to obtain the same result. A ) B is equivalent to :B ) :A is equivalent to :A _ B : A B A ) B :A :B :B ) :A :A _ B 1 1 0 0 1 0 1 0 1 0 1 1 0 0 1 1 0 1 0 1 1 0 1 1 1 0 1 1 A , B is equivalent to (A ) B ) ^ (B ) A) is equivalent to (:A _ B ) ^ (A _ :B ): A B A , B A ) B B ) A (A ) B ) ^ (B ) A) :A _ B A _ :B (:A _ B ) ^ (A _ :B ) 1 1 0 0 1 0 1 0 1 0 0 1 1 0 1 1 1 1 0 1 1 0 0 1 Negations of Quantiers: :8x 2 X : B (x) is equivalent to 9x 2 X : :B (x) :9x 2 X : B (x) is equivalent to 8x 2 X : :B (x) 9x 2 X 8y 2 Y : B (x; y) implies 8y 2 Y 9x 2 X : B (x; y) 35 1 0 1 1 1 1 0 1 1 0 0 1 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 13: Wednesday 9/29/1999 Scribe: Jurgen Symanzik, Rich Madsen 2.4 Transformations of Random Variables Let X be a real{valued random variable on ( ; L; P ), i.e., X : ( ; L) ! (IR; B). Let g be any Borel{measurable real{valued function on IR. Then, by statement 2.1.6, Y = g(X ) is a random variable. Theorem 2.4.1: Given a random rariable X with known induced distribution and a Borel{measurable function g, then the distribution of the random variable Y = g(X ) is determined. Proof: FY (y) = PY (Y y) = P (f! : g(X (!)) yg) = P (f! : X (!) 2 By g) where By = g;1 (;1; y] 2 B since g is Borel-measureable. = P (X ;1 (By )) Note: From now on, we restrict ourselves to real{valued (vector{valued) functions that are Borel{measurable, i.e., measurable with respect to (IR; B) or (IRk ; Bk ). More generally, PY (Y 2 C ) = PX (X 2 g;1 (C )) 8C 2 B. Example 2.4.2: Suppose X is a discrete random variable. Let A be a countable set such that P (X 2 A) = 1 and P (X = x) > 0 8x 2 A. Let Y = g(X ). Obviously, the sample space of Y is also countable. Then, PY (Y = y) = X x2g;1 (fyg) PX (X = x) = 36 X fx:g(x)=yg PX (X = x) 8y 2 g(A) Example 2.4.3: X U (;1; 1) so the pdf of X is fX (x) = 1=2I[;1;1] (x), which, according to Denition 2.3.8, reads as fX (x) = 1=2 for ;1 x 1 and 0 otherwise. ( x; x 0 + Let Y = X = 0; otherwise Then, 8 > 0; y<0 > > < 1=2; y=0 FY (y) = PY (Y y) = > 1=2 + y=2; 0 < y < 1 > > : 1; y1 This is the mixed discrete/continuous distribution from Example 2.3.7. Note: We need to put some conditions on g to ensure g(X ) is continuous if X is continuous and avoid cases as in Example 2.4.3 above. Denition 2.4.4: For a random variable X from ( ; L; P ) to (IR; B), the support of X (or P ) is any set A 2 L for which P (A) = 1. For a continuous random variable X with pdf f , we can think of the support of X as X = X ;1 (fx : fX (x) > 0g). Denition 2.4.5: Let f be a real{valued function dened on D IR; D 2 B. We say: f is (strictly) non{decreasing if x < y ) f (x) (<) f (y) 8x; y 2 D f is (strictly) non{increasing if x < y ) f (x) (>) f (y) 8x; y 2 D f is monotonic on D if f is either increasing or decreasing and write f " or f #. Theorem 2.4.6: Let X be a continuous rv with pdf fX and support X. Let y = g(x) be dierentiable for all x and either (i) g0 (x) > 0 or (ii) g0 (x) < 0 for all x. Then, Y = g(X ) is also a continuous rv with pdf d g;1 (y) j I (y): fY (y) = fX (g;1 (y)) j dy g(X) 37 Proof: Part (i): g0 (x) > 0 8x 2 X So g is strictly increasing and continuous. Therefore, x = g;1 (y) exists and it is also strictly increasing and also dierentiable. Then, from Rohatgi, page 9, Theorem 15: d g;1 (y) = d g(x) j ;1 ;1 > 0 x=g (y) dy dx We get FY (y) = PY (Y y) = PY (g(X ) y) = PX (X g;1 (y)) = FX (g;1 (y)) for y 2 g(X) and, by dierentiation, d (F (g;1 (y))) By Chain d g;1 (y) fY (y) = FY0 (y) = dy = Rule fX (g;1 (y)) dy X Part (ii): g0 (x) < 0 8x 2 X So g is strictly decreasing and continuous. Therefore, x = g;1 (y) exists and it is also strictly decreasing and also dierentiable. Then, from Rohatgi, page 9, Theorem 15: d g;1 (y) = d g(x) j ;1 ;1 < 0 x=g (y) dy dx We get FY (y) = PY (Y y) = PY (g(X ) y) = PX (X g;1 (y)) = 1 ; PX (X g;1 (y)) = 1 ; FX (g;1 (y)) for y 2 g(X) and, by dierentiation, d d d By Chain Rule 0 ; 1 ; 1 ; 1 ; 1 ; 1 fY (y) = FY (y) = dy (1;FX (g (y))) = ;fX (g (y)) dy g (y) = fX (g (y)) ; dy g (y) Since dyd g;1 (y) < 0, the negative sign will cancel out, always giving us a positive value. Hence the need for the absolute value signs. Combining parts (i) and (ii), we can therefore write d g;1 (y) j I (y): fY (y) = fX (g;1 (y)) j dy g(X) 38 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lectures 14 & 15: Friday 10/1/1999 & Monday 10/4/1999 Scribe: Jurgen Symanzik, Hanadi B. Eltahir Note: In Theorem 2.4.6, we can also write fY (y) = fdg((xx)) j dx ; y 2 g (X ) j x=g;1(y) If g is monotonic over disjoint intervals, we can also get an expression for the pdf/cdf of Y = g(X ) as stated in the following theorem: Theorem 2.4.7: Let Y = g(X ) where X is a rv with pdf fX (x) on support X. Suppose there exists a partition A0 ; A1 ; : : : ; Ak of X such that P (X 2 A0 ) = 0 and fX (x) is continuous on each Ai . Suppose there exist functions g1 (x); : : : ; gk (x) dened on A1 through Ak , respectively, satisfying (i) g(x) = gi (x) 8x 2 Ai , (ii) gi (x) is monotonic on Ai , (iii) the set Y = gi (Ai ) = fy : y = gi (x) for some x 2 Ai g is the same for each i = 1; : : : ; k, and (iv) gi;1 (y) has a continuous derivative on Y for each i = 1; : : : ; k. Then, fY (y) = k X i=1 d g;1 (y) j I (y) fX (gi;1 (y)) j dy Y i Note: Rohatgi, page 73, Theorem 4, removes condition (iii) by dening n = n(y) and x1 (y); : : : ; xn (y). 39 Example 2.4.8: Let X be a rv with pdf fX (x) = 2x2 I(0;) (x). Let Y = sin(X ). What is fY (y)? Since sin is not monotonic on (0; ), Theorem 2.4.6 cannot be used to determine the pdf of Y . Two possible approaches: Method 1: cdfs For 0 < y < 1 we have FY (y) = PY (Y y) = PX (sin X y) = PX ([0 X sin;1 (y)] or [ ; sin;1 (y) X ]) = FX (sin;1 (y)) + (1 ; FX ( ; sin;1 (y))) since [0 X sin;1 (y)] and [ ; sin;1 (y) X ] are disjoint sets. Then, fY (y) = FY0 (y) = fX (sin;1 (y)) p 1 2 + (;1)fX ( ; sin;1 (y)) p ;1 2 1;y 1;y = p 1 2 fX (sin;1 (y)) + fX ( ; sin;1 (y)) 1;y ;1 (y)) 2( ; sin;1 (y)) ! 1 2(sin = p + 2 2 1 ; y2 = 2 p 1 2 2 1;y = p 2 2 I(0;1) (y) 1;y Method 2: Use of Theorem 2.4.7 Let A1 = (0; 2 ), A2 = ( 2 ; ), and A0 = f 2 g. Let g1;1 (y) = sin;1 (y) and g2;1 (y) = ; sin;1 (y). It is dyd g1;1 (y) = p11;y2 = ; dyd g2;1 (y) and Y = (0; 1). Thus, by use of Theorem 2.4.7, we get fY (y) = 2 X i=1 d g;1 (y) j I (y) fX (gi;1 (y)) j dy i Y 40 ;1 (y)) 1 ;1 p 2 I(0;1) (y) = 2 sin2 (y) p 1 2 I(0;1) (y) + 2( ; sin 2 1;y 1;y = 22 p 1 2 I(0;1) (y) 1;y = p 2 2 I(0;1) (y) 1;y Obviously, both results are identical. Theorem 2.4.9: Let X be a rv with a continuous cdf FX (x) and let Y = FX (X ). Then, Y U (0; 1). Proof: We have to consider two possible cases: (a) FX is strictly increasing, i.e., FX (x1 ) < FX (x2 ) for x1 < x2 , and (b) FX is non{decreasing, i.e., there exists x1 < x2 and FX (x1 ) = FX (x2 ). Assume that x1 is the inmum and x2 the supremum of those values for which FX (x1 ) = FX (x2 ) holds. In (a), FX;1 (y) is uniquely dened. In (b), we dene FX;1 (y) = inf fx : FX (x) yg Without loss of generality: FX;1 (1) = +1 if FX (x) < 1 8x 2 IR and FX;1 (0) = ;1 if FX (x) > 0 8x 2 IR. For Y = FX (X ) and 0 < y < 1, we have P (Y y) = F ;1 " = = = = X () P (FX (X ) y) P (FX;1 (FX (X )) FX;1(y)) P (X FX;1 (y)) FX (FX;1 (y)) y At the endpoints, we have P (Y y) = 1 if y 1 and P (Y y) = 0 if y 0. But why is () true? | In (a), if FX is strictly increasing and continuous, it is certainly x = FX;1(FX (x)). In (b), if FX (x1 ) = FX (x2 ) for x1 < x < x2 , it may be that FX;1 (FX (x)) 6= x. But by denition, FX;1(FX (x)) = x1 8x 2 [x1; x2 ]. () holds since on [x1 ; x2 ], it is P (X x) = P (X x1 ) 8x 2 [x1 ; x2 ]. The at cdf denotes FX (x2 ) ; FX (x1 ) = P (x1 < X x2 ) = 0 by denition. 41 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 16: Wednesday 10/6/1999 Scribe: Jurgen Symanzik 3 Moments and Generating Functions 3.1 Expectation Denition 3.1.1: Let X be a real-valued rv with cdf FX and pdf fX if X is continuous (or pmf fX and support X if X is discrete). The expected value (mean) of a measurable function g() of X is 8Z1 > > < ;1 g(x)fX (x)dx; if X is continuous E (g(X )) = > X > : g(x)fX (x); if X is discrete x2X if E (j g(X ) j) < 1; otherwise E (g(X )) is undened, i.e., it does not exist. Example: X Cauchy; fX (x) = (1+1 x2 ) ; ;1 < x < 1: Z1 x 1 [log(1 + x2 )]1 = 1 2 E (j X j) = dx = 0 2 0 1+x So, E (X ) does not exist for the Cauchy distribution. Theorem 3.1.2: If E (X ) exists and a and b are nite constants, then E (aX + b) exists and equals aE (X ) + b. Proof: Continuous case only: Existence: E (j aX + b j) = Z1 j ax + b j fX (x)dx Z;1 1 (j a j j x j + j b j)fX (x)dx ;1Z Z1 1 fX (x)dx j x j fX (x)dx+ j b j = jaj ;1 ;1 42 = j a j E (j X j)+ j b j < 1 Numerical Result: E (aX + b) = Z1 (ax + b)fX (x)dx ;1 Z1 Z1 = a xfX (x)dx + b fX (x)dx ;1 ;1 = aE (X ) + b Theorem 3.1.3: If X is bounded (i.e., there exists a M; 0 < M < 1, such that P (j X j< M ) = 1), then E (X ) exists. Denition 3.1.4: The kth moment of X , if it exists, is mk = E (X k ). The kth central moment of X , if it exists, is k = E ((X ; E (X ))k ). Denition 3.1.5: The variance of X , if it exists, is the second central moment of X , i.e., V ar(X ) = E ((X ; E (X ))2 ): Theorem 3.1.6: V ar(X ) = E (X 2 ) ; (E (X ))2 . Proof: V ar(X ) = E ((X ; E (X ))2 ) = E (X 2 ; 2XE (X ) + (E (X ))2 ) = E (X 2 ) ; 2E (X )E (X ) + (E (X ))2 = E (X 2 ) ; (E (X ))2 43 Theorem 3.1.7: If V ar(X ) exists and a and b are nite constants, then V ar(aX + b) exists and equals a2 V ar(X ). Proof: Existence & Numerical Result: ; ; V ar(aX + b) = E ((aX + b) ; E (aX + b))2 exists if E j ((aX + b) ; E (aX + b))2 j exists. It holds that = = Th:= 3:1:6 Th:= 3:1:2 Th:= 3:1:2 = Th:= 3:1:6 < E j ((aX + b) ; E (aX + b))2 j E ((aX + b) ; E (aX + b))2 V ar(aX + b) E ((aX + b)2 ) ; (E (aX + b))2 E (a2 X 2 + 2abX + b2 ) ; (aE (X ) + b)2 a2 E (X 2 ) + 2abE (X ) + b2 ; a2 (E (X ))2 ; 2abE (X ) ; b2 a2 (E (X 2 ) ; (E (X ))2 ) a2 V ar(X ) 1 since V ar(X ) exists 44 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 17: Friday 10/8/1999 Scribe: Jurgen Symanzik Theorem 3.1.8: If the tth moment of a rv X exists, then all moments of order 0 < s < t exist. Proof: Continuous case only: Z j x js fX (x)dx + j x js fX (x)dx j x j 1 j x j > 1 Z Z 1 fX (x)dx + j x jt fX (x)dx jxj1 jxj>1 P (j X j 1) + E (j X jt) < 1 E (j X js) = Z Theorem 3.1.9: If the tth moment of a rv X exists, then t nlim !1 n P (j X Proof: Continuous case only: Z Z t 1 > j x j fX (x)dx = nlim !1 IR =) nlim !1 But, nlim !1 Z Z jxj>n jxj>n jxjn j> n) = 0: j x jt fX (x)dx j x jt fX (x)dx = 0 t j x jt fX (x)dx nlim !1 n Z jxj>n t fX (x)dx = nlim !1 n P (j X j> n) = 0 Note: t th The inverse is not necessarily true, i.e., if nlim !1 n P (j X j> n) = 0, then the t moment of a rv X does not necessarily exist. We can only approach t up to some > 0 as the following Theorem 3.1.10 indicates. 45 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 18: Monday 10/11/1999 Scribe: Jurgen Symanzik, Bill Morphet Theorem 3.1.10: t Let X be a rv with a distribution such that nlim !1 n P (j X j> n) = 0 for some t > 0. Then, E (j X js ) < 1 8 0 < s < t: Note: To prove this Theorem, we need Lemma 3.1.11 and Corollary 3.1.12. Lemma 3.1.11: Let X be a non{negative rv with cdf F . Then, E (X ) = Z1 0 (1 ; FX (x))dx (if either side exists). Proof: Continuous case only: To prove that the left side implies that the right side is nite and both sides are identical, we assume that E (X ) exists. It is Z1 Zn E (X ) = xfX (x)dx = nlim !1 xfX (x)dx 0 0 Replace the expression for the right side integral using integration by parts. Let u = x and dv = fX (x)dx, then Zn Zn n xfX (x)dx = (xF (x)) j0 ; FX (x)dx 0 0 Zn = nFX (n) ; 0FX (0) ; FX (x)dx 0 Zn = nFX (n) ; n + n ; FX (x)dx 0 Zn = nFX (n) ; n + [1 ; FX (x)]dx 0 46 = n[FX (n) ; 1] + Zn 0 = ;n[1 ; FX (n)] + = ;nP (X > n) + [1 ; FX (x)]dx Zn 0 n Z 0 X 0 [1 ; FX (x)]dx [1 ; FX (x)]dx = ;n1 P (jX j > n) + Zn 0 [1 ; FX (x)]dx 1 =) E (X 1 ) = nlim !1[;n P (jX j > n) + Th:=3:1:9 0 + = Z1 0 nlim !1 Zn 0 Zn 0 [1 ; FX (x)]dx] [1 ; FX (x)]dx [1 ; FX (x)]dx Thus, the existence of E (X ) implies that Z1 0 [1 ; FX (x)]dx is nite and that both sides are identical. We still have to show the converse implication: Z1 If [1 ; FX (x)]dx is nite, then E (X ) exists, i.e., E (j X j) = E (X ) < 1, and both sides are 0 identical. It is Zn 0 xfX (x)dx X=0 Zn 0 as seen above. Since ;n[1 ; FX (n)] 0, we get Zn 0 Thus, j x j fX (x)dx j x j fX (x)dx = ;n[1 ; FX (n)] + Zn 0 Zn [1 ; FX (x)]dx Z1 Z1 0 Zn 0 [1 ; FX (x)]dx [1 ; FX (x)]dx < 1 8n Z1 j x j fX (x) = j x j fX (x)dx [1 ; FX (x)]dx < 1 0 0 Z1 =) E (X ) exists and is identical to [1 ; FX (x)]dx as seen above. nlim !1 0 0 47 Corollary 3.1.12: E (j X js) = s Z1 0 ys;1 P (j X j> y)dy Proof: Z1 Z1 E (j X js ) Lemma= 3:1:11 [1 ; FjX js (z )]dz = P (j X js > z )dz . 0 0 dz = sys;1 and dz = sys;1dy. Therefore, Let z = ys . Then dy Z1 0 P (j X js> z)dz Z1 = = s monotonic " s = 0Z P (j X js > ys)sys;1 dy 1 s;1 y P (j X js > ys )dy Z01 0 ys;1P (j X j> y)dy Proof (of Theorem 3.1.10): For any given > 0, choose N such that the tail probability P (j X j> n) < nt 8n N . E (j X js) Cor:=3:1:12 s =s ZN 0 N Z 0 N = ycdy = ( ys;1 P (j X j> y)dy sys;1 1 dy + s = N s + s Z1 0 ys;1P (j X j> y)dy + s = ys jN0 + s It is Z1 Z1 N Z1 N ( Z1 N Z1 N ys;1P (j X j> y)dy ys;1 yt dy ys;1 y1t dy ys;1;t dy 1 c+1 1 c+1 y jN ; ln y j1 N; c 6= ;1 c=1 1; c ;1 1 c +1 ; c+1 N < 1; c < ;1 Thus, for E (j X js ) < 1, it must hold that s ; 1 ; t < ;1, or equivalently, s < t. So E (j X js ) < 1, i.e., it exists, for every s with 0 < s < t for a rv X with a distribution such t that nlim !1 n P (j X j> n) = 0 for some t > 0. 48 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 19: Wednesday 10/13/1999 Scribe: Jurgen Symanzik, Rich Madsen Theorem 3.1.13: Let X be a rv such that P (j X j> k) = 0 8 > 1: lim k!1 P (j X j> k) Then, all moments of X exist. Proof: For > 0; we select some k0 such that P (j X j> k) < 8k k : 0 P (j X j> k) Select k1 such that P (j X j> k) < 8k k1 : Select N = max(k0 ; k1 ). If we have some xed positive integer r: P (j X j> r k) = P (j X j> k) P (j X j> 2 k) P (j X j> 3 k) : : : P (j X j> r k) P (j X j> k) P (j X j> k) P (j X j> k) P (j X j> 2 k) P (j X j> r;1k) r k) P (j X j> (k)) P (j X j> (2 k)) : : : P (j X j> (r;1 k)) =) PP(j(jXXj>j>kk) ) = PP((j jXXj> j> k) P (j X j> 1 (k)) P (j X j> 1 (2 k)) P (j X j> 1 (r;1 k)) Note: Each of these r terms on the right side is > by our original statement of selecting some ) n k0 such that PP((jXjXj>k j>k) < 8k k0 and since > 1 and therefore k k0 . k) r Now we get for our entire expression that PP(j(XjXj> j>k) for k N (since in this case also k k0 ) and > 1: Overall, we have P (j X j> r k) r P (j X j> k) r+1 for k N (since in this case also k k1 ). For a xed positive integer n: Z1 ZN Z1 Cor: 3 : 1 : 12 n n ; 1 n ; 1 E(j X j ) = n x P(j X j> x)dx = n x P(j X j> x)dx + n xn;1 P(j X j> x)dx r 0 We know that: n ZN 0 0 xn;1 P (j X ZN N j> x)dx nxn;1dx = xn jN0 = N n < 1 0 49 but is Z1 n xn;1P (j X j> x)dx < 1 ? N To check the second part, we use: Z1 1 Z N X n ; 1 x P (j X j> x)dx = xn;1 P (j X j> x)dx r r=1r;1 N N We know that: Zr N r;1 N since r xn;1 P (j X j> x)dx r Zr N r;1 N xn;1 dx This step is possible P (j X j r;1N ) P (j X j> x) P (j X j r N ) 8x 2 (r;1 N; r N ) and N = max(k0 ; k1 ). Since (r;1 N )n;1 xn;1 (r N )n;1 8x 2 (r;1N; r N ), we get: Z N Z N r n ; 1 r r n ; 1 x dx ( N ) 1dx r (r N )n;1 (r N ) r (r N )n r r r;1 N r;1 N Now we go back to our original inequality: Z1 N xn;1 P (j X j> x)dx 1 X r Zr N r=1 r;1 N xn;1dx n n r=1 r (r N )n = N n n n n < 1 or, equivalently, if < 1 if = 1N; n n Since N1; is nite, all moments E (j X jn ) exist. n 1 X 50 1 X ( n )r r=1 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 20: Friday 10/15/1999 Scribe: Jurgen Symanzik, Hanadi B. Eltahir 3.2 Generating Functions Denition 3.2.1: Let X be a rv with cdf FX . The moment generating function (mgf) of X is dened as MX (t) = E (etX ) provided that this expectation exists in an (open) interval around 0, i.e., for ;h < t < h for some h > 0. Theorem 3.2.2: If a rv X has a mgf MX (t) that exists for ;h < t < h for some h > 0, then n E (X n ) = MX(n) (0) = dtd n MX (t) jt=0 : Proof: We assume that we can dierentiate under the integral sign. If, and when, this really is true will be discussed later in this section. d M (t) = d Z 1 etx f (x)dx X dt X dt ;1 Z1 @ = ( @t etx fX (x))dx Z;1 1 xetxfX (x)dx = E (XetX ) = ;1 Evaluating this at t = 0, we get: dtd MX (t) jt=0 = E (X ) By iteration, we get for n 2: ! dn M (t) = d dn;1 M (t) dtn X dt dtn;1 X Z 1 d n ; 1 tx = dt x e fX (x)dx ;1 Z1 @ = ( @t xn;1 etx fX (x))dx ;1 51 Z1 xn etx fX (x)dx = E (X n etX ) = ;1 Evaluating this at t = 0, we get: dtdnn MX (t) jt=0 = E (X n ) Example 3.2.3: X U (a; b); fX (x) = b;1 a I[a;b](x). Then, MX (t) = Z b etx etb ; eta dx = t (b ; a ) a b;a MX (0) = 0 0 L'Hospital = = betb ; aeta b;a 1 So MX (0) = 1 and since ettb(b;;eata) is continuous, it also exists in an open interval around 0 (in fact, it exists for every t 2 IR). MX0 (t) = = =) E (X ) = (betb ; aeta )t(b ; a) ; (etb ; eta )(b ; a) t2 (b ; a)2 t(betb ; aeta ) ; (etb ; eta ) t2 (b ; a) MX0 (0) = 00 = betb ; aeta + tb2 etb ; ta2 eta ; betb + aeta 2t(b ; a) = tb2 etb ; ta2eta 2t(b ; a) L'Hospital L'Hospital = = = = 0 t=0 0 b2 etb ; a2 eta + tb3 etb ; ta3eta 2(b ; a) b2 ; a2 2(b ; a) b+a 2 52 t=0 t=0 Note: In the previous example, we made use of L'Hospital's rule. This rule gives conditions under 1 ". which we can resolve indenite expressions of the type \ 00 " and \ 1 (i) Let f and g be functions that are dierentiable in an open interval around x0 , say in (x0 ; ; x0 + ); but not necessarily dierentiable in x0 . Let f (x0 ) = g(x0 ) = 0 and g0 (x) 6= 0 f 0(x) = A implies that also lim f (x) = A. The 8x 2 (x0 ; ; x0 + ) ; fx0g. Then, xlim x!x0 g(x) !x0 g0 (x) + ; same holds for the cases xlim !x0 f (x) = xlim !x0 g(x) = 1 and x ! x0 or x ! x0 . (ii) Let f and g be functions that are dierentiable for x > a (a > 0). Let xlim !1 f (x) = xlim !1 g(x) = 0 0 (x) 6= 0. Then, lim f (x) = A implies that also lim f (x) = A. 0 and xlim g !1 x!1 g0 (x) x!1 g(x) (iii) We can iterate this process as long as the required conditions are met and derivatives exist, e.g., if the rst derivatives still result in an indenite expression, we can look at the second derivatives, then at the third derivatives, and so on. (iv) It is recommended to keep expressions as simple as possible. If we have identical factors in the numerator and denominator, we can exclude them from both and continue with the simpler functions. (v) Indenite expressions of the form \0 1" can be handled by rearranging them to \ 1=01 " and f (x) can be handled by use of the rules for lim f (;x) . lim x!1 g(;x) x!;1 g(x) Note: The following Theorems provide us with rules that tell us when we can dierentiate under the integral sign. Theorem 3.2.4 relates to nite integral bounds a() and b() and Theorems 3.2.5 and 3.2.6 to innite bounds. Theorem 3.2.4: Leibnitz's Rule If f (x; ); a(), and b() are dierentiable with respect to (for all x) and ;1 < a() < b() < 1, then d Z b() f (x; )dx = f (b(); ) d b() ; f (a(); ) d a() + Z b() @ f (x; )dx: d d d @ a() The rst 2 terms are vanishing if a() and b() are constant in . Proof: Uses the Fundamental Theorem of Calculus and the chain rule. 53 a() Theorem 3.2.5: Lebesque's Dominated Convergence Theorem Z1 Let g be an integrable function such that g(x)dx < 1. If j fn j g almost everywhere (i.e., ;1 except for a set of Borel{measure 0) and if fn ! f almost everywhere, then fn and f are integrable and Z Z fn(x)dx ! f (x)dx: Note: If f is dierentiable with respect to , then @ f (x; ) = lim f (x; + ) ; f (x; ) !0 @ and Z1 @ Z 1 f (x; + ) ; f (x; ) lim dx !0 @ f (x; )dx = ;1 while ;1 @ Z 1 f (x; )dx = lim Z 1 f (x; + ) ; f (x; ) dx !0 ;1 @ ;1 Theorem 3.2.6: Let fZn(x; 0 ) = f (x;0 +nn);f (x;0 ) for some 0 . Suppose there exists an integrable function g(x) such 1 that g(x)dx < 1 and j fn (x; ) j g(x) 8x, then ;1 d Z1 d ;1 f (x; )dx =0 = Z1@ ;1 @ f (x; ) j=0 dx: Usually, if f is dierentiable for all , we write d Z 1 f (x; )dx = Z 1 @ f (x; )dx: d ;1 ;1 @ Corollary 3.2.7: Let there exists an integrable function g(x; ) such that Z 1 f (x; ) be dierentiable @ for all . Suppose g(x; )dx < 1 and @ f (x; ) j=0 g(x; ) 8x 80 in some {neighborhood of , then ;1 d Z 1 f (x; )dx = Z 1 @ f (x; ) j dx: =0 d ;1 ;1 @ 54 More on Moment Generating Functions @ etxfX (x) jt=t0 =j x j et0 xfX (x) for j t0 ; t j 0 : @t Choose t; 0 small enough such that t + 0 2 (;h; h) and t ; 0 2 (;h; h). Then, @ etxfX (x) jt=t0 g(x; t) @t Consider where ( (t+0 )x g(x; t) = j x j e(t;0 )x fX (x); x 0 jxje fX (x); x < 0 R To verify g(x; t)dx < 1, we need to know fX (x). Suppose mgf MX (t) exists for j t j h for some h > 1. Then j t + 0 + 1 j< h and j t ; 0 ; 1 j< h. Since j x j ejxj 8x, we get ( (t+0 +1)x g(x; t) e(t;0 ;1)x fX (x); x 0 e fX (x); x < 0 Then, Z1 ;1 Z1 0 g(x; t)dx MX (t + 0 +1) < 1 and g(x)dx < 1. Z0 ;1 g(x; t)dx MX (t ; 0 ; 1) < 1 and, therefore, Together with Corollary 3.2.7, this establishes that we can dierentiate under the integral in the Proof of Theorem 3.2.2. If h 1, we may need to check more carefully to see if the condition holds. Note: If MX (t) exists for t 2 (;h; h), then we have an innite collection of moments. Does a collection of integer moments fmk : k = 1; 2; 3; : : : g completely characterize the distribution, i.e., cdf, of X ? | Unfortunately not, as Example 3.2.8 shows. 55 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 21: Monday 10/18/1999 Scribe: Jurgen Symanzik Example 3.2.8: Let X1 and X2 be rv's with pdfs fX1 (x) = p1 x1 exp(; 12 (log x)2 ) I(0;1)(x) 2 and fX2 (x) = fX1 (x) (1 + sin(2 log x)) I(0;1) (x) It is E (X1r ) = E (X2r ) = er2 =2 for r = 0; 1; 2; : : : as you have to show in the Homeworks. Two dierent pdfs/cdfs have the same moment sequence! What went wrong? In this example, MX1 (t) does not exist as shown in the Homeworks! Theorem 3.2.9: Let X and Y be 2 rv's with cdf's FX and FY for which all moments exist. (i) If FX and FY have bounded support, then FX (u) = FY (u) 8u i E (X r ) = E (Y r ) for r = 0; 1; 2; : : : . (ii) If both mgf's exist, i.e., MX (t) = MY (t) for t in some neighborhood of 0, then FX (u) = FY (u) 8u. Note: The existence of moments is not equivalent to the existence of a mgf as seen in Example 3.2.8 above and some of the Homework assignments. Theorem 3.2.10: Suppose rv's fXi g1 M (t) = MX (t) 8t 2 (;h; h) for some i=1 have mgf's MXi (t) and that ilim !1 Xi h > 0 and that MX (t) itself is a mgf. Then, there exists a cdf FX whose moments are determined by MX (t) and for all continuity points x of FX (x) it holds that ilim F (x) = FX (x), i.e., the !1 Xi convergence of mgf's implies the convergence of cdf's. Proof: Uniqueness of Laplace transformations, etc. 56 Theorem 3.2.11: For constants a and b, the mgf of Y = aX + b is MY (t) = ebt MX (at); given that MX (t) exists. Proof: MY (t) = E (e(aX +b)t ) = E (eaXt ebt ) = ebt E (eXat ) = ebt MX (at) 3.3 Complex{Valued Random Variables and Characteristic Functions Recall the following facts regarding complexd numbers: p i0 = +1; i = ;1; i2 = ;1; i3 = ;i; i4 = +1; etc. in the planar Gauss'ian number plane it holds that i = (0; 1) z = a + ib = r(cos + i sin ) p r =j z j= a2 + b2 tan = ab Euler's Relation: z = r(cos + i sin ) = rei Mathematical Operations on Complex Numbers: z1 z2 = (a1 a2) + i(b1 b2 ) z1 z2 = r1 r2ei(1 +2) = r1 r2 (cos(1 + 2 ) + i sin(1 + 2 )) z1 r1 i(1 ;2 ) = r1 (cos(1 ; 2 ) + i sin(1 ; 2 )) z2 = r2 e r2 Moivre's Theorem: z n = (r(cos + i sin ))n = rn (cos(n) + i sin(n)) pn z = pn a + ib = pn r cos( +k3600 ) + i sin( +k3600 ) for k = 0; 1; : : : ; (n ; 1) and the main value n n for k = 0 ln z = ln(a + ib) = ln(j z j) + i i2n where = arctan ab and the main value for n = 0 57 Conjugate Complex Numbers: For z = a + ib, we dene the conjugate complex number z = a ; ib. It holds: z=z z = z i z 2 IR z1 z2 = z1 z2 z1 z2 = z1 z2 z1 z1 z2 = z2 z z = a2 + b2 Re(z ) = a = 12 (z + z ) Im(z ) = b = 21i (z ; z ) p p j z j= a2 + b2 = z z Denition 3.3.1: Let ( ; L; P ) be a probability space and X and Y real{valued rv's, i.e., X; Y : ( ; L) ! (IR; B) (i) Z = X + iY : ( ; L) ! (CI; BCI) is called a complex{valued random variable (CI-rv). (ii) If E (X ) and E (Y ) exist, then E (Z ) is dened as E (Z ) = E (X ) + iE (Y ) 2 CI. Note: E (Z ) exists i E (j X j) and E (j Y j) exist. It also holds that if E (Z ) exists, then j E (Z ) j E (j Z j) (see Homework). Denition 3.3.2: Let X be a real{valued rv on ( ; L; P ). Then, X (t) : IR ! CI with X (t) = E (eitX ) is called the characteristic function of X . Note: Z1 Z1 Z1 (i) X (t) = eitx fX (x)dx = cos(tx)fX (x)dx + i sin(tx)fX (x)dx if X is continuous. ;1 ;1 ;1 X X X (ii) X (t) = eitx P (X = x) = cos(tx)P (X = x)+ i sin(tx)P (X = x)(x) if X is discrete x2X x2X x2X and X is the support of X . (iii) X (t) exists for all real{valued rv's X since j eitx j= 1. 58 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 22: Wednesday 10/20/1999 Scribe: Jurgen Symanzik, Bill Morphet Theorem 3.3.3: Let X be the characteristic function of a real{valued rv X . Then it holds: (i) X (0) = 1. (ii) j X (t) j 1 8t 2 IR. (iii) X is uniformly continuous, i.e., 8 > 0 9 > 0 8t1 ; t2 2 IR :j t1 ; t2 j< )j (t1 ) ; (t2 ) j< . (iv) X is a positive denite function, i.e., 8n 2 IN 81 ; : : : ; n 2 CI 8t1 ; : : : ; tn 2 IR : n X n X l=1 j =1 l j X (tl ; tj ) 0. (v) X (t) = X (;t). (vi) If X is symmetric around 0, i.e., if X has a pdf that is symmetric around 0, then X (t) 2 IR 8t 2 IR. (vii) aX +b (t) = eitb X (at). Proof: See Homework for parts (i), (ii), (iv), (v), (vi), and (vii). Part (iii): Known conditions: (i) Let > 0. (ii) 9 a > 0 : P (;a < X < +a) > 1 ; 4 and P (j X j> a) < 4 (iii) 9 > 0 : j e{(t0 ;t)x ; 1 j< 2 8x s.t. j x j< a and 8(t0 ; t) s.t. 0 < (t0 ; t) < . This third condition holds since j e{0 ; 1 j= 0 and the exponential function is continuous. Therefore, if we select (t0 ; t) and x small enough, j e{(t0 ;t)x ; 1 j will be < 2 for a given . 59 Let t; t0 2 IR, t < t0 , and t0 ; t < . Then, j X (t0) ; X (t) j = j Z +1 = j ;1 Z ;a = j ;1 Z ;a j ;1 Z +1 ;1 Z +1 e{t0 xfX (x)dx ; ;1 e{tx fX (x)dx j (e{t0 x ; e{tx )fX (x)dx j (e{t0 x ; e{tx )f X (x)dx + Z +a ;a (e{t0 x ; e{tx )fX (x)dx j + j (e{t0 x ; e{tx )f Z +a ;a X (x)dx + Z +1 +a (e{t0 x ; e{tx )fX (x)dx j + j (e{t0 x ; e{tx )fX (x)dx j Z +1 +a (e{t0 x ; e{tx )fX (x)dx j We now take a closer look at the rst and third of these absolute integrals. It is: j Z ;a ;1 Z ;a (e{t0 x ; e{tx )fX (x)dx j = j Z ;a j Z ;a j e{t0 x j fX (x)dx + ;1 (A) = = ;1 e{t0 xfX (x)dx j + j Z ;a ;1 Z ;a ;1 1fX (x)dx + Z ;a ;1 ;1 Z ;a ;1 Z ;a ;1 e{t0 xfX (x)dx ; Z ;a ;1 e{tx fX (x)dx j e{tx fX (x)dx j j e{tx j fX (x)dx 1fX (x)dx 2fX (x)dx. (A) holds due to Note (iii) that follows Denition 3.3.2. Similarly, j Z +1 +a (e{t0 x ; e{tx )fX (x)dx j Z +1 +a 2fX (x)dx Returning to the main part of the proof, we get j X (t0) ; X (t) j =2 Z ;a ;1 Z ;a ;1 fX (x)dx + = 2P (j X j> a) + j 2fX (x)dx + j Z +1 +a Z +a ;a Z +a ;a (e{t0 x ; e{tx )fX (x)dx j + fX (x)dx + j Z +a ;a (e{t0 x ; e{tx )fX (x)dx j 60 Z +1 +a (e{t0 x ; e{tx )fX (x)dx j 2fX (x)dx Z +a 0 (e{t x ; e{tx )fX (x)dx j 24 + j Condition (ii) ;a Z +a = 2 +j e{tx (e{(t0 ;t)x ; 1)fX (x)dx j ;a 2 + Z +a ;a j e{tx (e{(t0 ;t)x ; 1) j fX (x)dx Z +a j e{tx j j (e{(t0 ;t)x ; 1) j fX (x)dx 2 + ;a Z +a 1 2 fX (x)dx 2+ (B ) ;a Z +1 2+ 2 fX (x)dx ;1 = 2 + 2 = (B) holds due to Note (iii) that follows Denition 3.3.2 and due to condition (iii). Theorem 3.3.4: Bochner's Theorem Let : IR ! CI be any function with properties (i), (ii), (iii), and (iv) from Theorem 3.3.3. Then there exists a real{valued rv X with X = . Theorem 3.3.5: Let X be a real{valued rv and E (X k ) exists for an integer k. Then, X is k times dierentiable and (Xk) (t) = ik E (X k eitX ). In particular for t = 0, it is (Xk) (0) = ik mk . Theorem 3.3.6: Let X be a real{valued rv with characteristic function X and let X be k times dierentiable, where k is an even integer. Then the kth moment of X , mk , exists and it is (Xk) (0) = ik mk . 61 Theorem 3.3.7: Levy's Theorem Let X be a real{valued rv with cdf FX and characteristic function X . Let a; b 2 IR, a < b. If P (X = a) = P (X = b) = 0, i.e., FX is continuous in a and b, then F (b) ; F (a) = 21 Z 1 e;ita ; e;itb X (t)dt: it ;1 Theorem 3.3.8: Let X and Y be a real{valued rv with characteristic functions X and Y . If X = Y , then X and Y are identically distributed. Theorem 3.3.9: Z1 Let X be a real{valued rv with characteristic function X such that j X (t) j dt < 1. Then ;1 X has pdf Z1 fX (x) = 21 e;itxX (t)dt: ;1 Theorem 3.3.10: Let X be a real{valued rv with mgf MX (t), i.e., the mgf exists. Then X (t) = MX (it). Theorem 3.3.11: 1 1 Suppose real{valued rv's fXi g1 i=1 have cdf's fFXi gi=1 and characteristic functions fXi (t)gi=1 . If lim (t) = X (t) 8t 2 (;h; h) for some h > 0 and X (t) is itself a characteristic function i!1 Xi (of a rv X with cdf FX ), then ilim F (x) = FX (x) for all continuity points x of FX (x), i.e., the !1 Xi convergence of characteristic functions implies the convergence of cdf's. 62 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 23: Friday 10/22/1999 Scribe: Jurgen Symanzik, Rich Madsen Theorem 3.3.12: Characteristic functions for some well{known distributions: Distribution X (t) (i) X Dirac(c) eitc (ii) X Bin(1; p) 1 + p(eit ; 1) (iii) X Poisson(c) exp(c(eit ; 1)) eitb ;eita (iv) X U (a; b) (b;a)it (v) X N (0; 1) exp(;t2 =2) (vi) X N (; 2 ) eit exp(;2 t2 =2) (vii) X ;(p; q) (1 ; itq );p (viii) X Exp(c) (1 ; itc );1 (1 ; 2it);n=2 (ix) X 2n Proof: (i) X (t) = E (eitX ) = eitc P (X = c) = eitc (ii) X (t) = (iii) X (t) = 1 X k=0 eitk P (X = k) = eit0 (1 ; p) + eit1 p = 1 + p(eit ; 1) X 1 1 xn n X X eitn cn! e;c = e;c n1! (c eit )n = e;c ece = ec(e ;1) since = ex n ! n=0 n=0 n2IN 0 Z it b (iv) X (t) = b;1 a eitx dx = a " # eitx b = eitb ; eita b ; a it a (b ; a)it 1 it (v) X N (0; 1) is symmetric around 0 ) X (t) is real since there is no imaginary part according to Theorem 3.3.3 (vi) ) X (t) = p12 Z1 ;1 cos(tx)e ;x2 2 dx 63 Since E (X ) exists, X (t) is dierentiable according to Theorem 3.3.5 and the following holds: 0X (t) = Re(0X (t)) 0 Z1 = Re B @ ;1 ix = = = = 1 p1 e dxC eitx A |{z} 2 cos(tx)+i sin(tx) Z 1 Z1 ;2 ;2 1 1 2 2 ;x sin(tx) p e dx Re ix cos(tx) p e dx + 2 2 ;1 ;1 Z 1 ;2 0 = ;t cos(tx) and v = ;e ;2 2 2 dx p1 (|; sin( tx )) xe j u | {z } {z } v0 2 ;1 u Z1 2 ;2 p1 sin(tx)e ;2 j1;1 ; p1 (;t cos(tx))(;e 2 )dx | 2 {z } 2 ;1 =0 since sin is odd Z1 ;2 ;t p1 cos(tx)e 2 dx ;x2 2 x x x x 2 ;1 = ;tX (t) x x x 0 Thus, 0X (t) = ;tX (t). It follows that XX ((tt)) = ;t and by integrating both sides, we get ln j X (t) j= ; 21 t2 + c with c 2 IR. For t = 0, we know that X (0) = 1 by Theorem 3.3.3 (i) and ln j X (0) j= 0. It follows that 0 = 0 + c. Therefore, c = 0 and j X (t) j= e; 12 t2 . If we take t = 0, then X (0) = 1 by Theorem 3.3.3 (i). Since X is continuous, X must take the value 0 before it can eventually take a negative value. However, since e; 12 t2 > 0 8t 2 IR, X cannot take 0 as a possible value and therefore cannot pass into the negative numbers. So, it must hold that X (t) = e; 21 t2 8t 2 IR: (vi) For > 0; 2 IR, we know that if X N (0; 1), then X + N (; 2 ). By Theorem 3.3.3 (vii) we have 1 X + (t) = eit X (t) = eit e; 2 2 t2 : (vii) X (t) = Z1 0 eitx (p; q; x)dx Z 1 qp xp;1e;(q;it)x dx 0 ;(p) Z1 p q ; p = ;(p) (q ; it) ((q ; it)x)p;1 e;(q;it)x (q ; it)dx j u = (q ; it)x; du = (q ; it)dx = 0 64 Z1 p = ;(q p) (q ; it);p (u)p;1 e;u du | 0 {z } =;(p) = qp(q ; it);p ;p = (q ; it) q;p q ; it ;p = q it ;p = 1; q (viii) Since an Exp(c) distribution is a ;(1; c) distribution, we get for X Exp(c) = ;(1; c): X (t) = (1 ; itc );1 (ix) Since a 2n distribution (for n 2 IN ) is a ;( n2 ; 12 ) distribution, we get for X 2n = ;( n2 ; 12 ): X (t) = (1 ; 1it=2 );n=2 = (1 ; 2it);n=2 Example 3.3.13: Since we know that m1 = E (X ) and m2 = E (X 2 ) exist for X Bin(1; p), we can determine these moments according to Theorem 3.3.5 using the characteristic function. It is X (t) = 1 + p(eit ; 1) 0X (t) = pieit 0X (0) = pi 0 ) m1 = Xi(0) = pii = p = E (X ) 00X (t) = pi2 eit 00X (0) = pi2 00 2 ) m2 = Xi2(0) = pii2 = p = E (X 2 ) ) V ar(X ) = E (X 2 ) ; (E (X ))2 = p ; p2 = p(1 ; p) 65 Note: Z1 The restriction j X (t) j dt < 1 in Theorem 3.3.9 works in such a way that we don't end up ;1 with a (non{existing) pdf if X is a discrete rv. For example, X Dirac(c): Z1 ;1 j X (t) j dt = Z1 Z;1 1 j eitc j dt = 1dt ;1 = x j1 ;1 which is undened. Also for X Bin(1; p): Z1 ;1 j X (t) j dt = = = = = Z1 which is undened for p 6= 1=2. If p = 1=2, we have Z1 ;1 j 1 + p(eit ; 1) j dt Z;1 1 j peit ; (p ; 1) j dt Z;1 1 j j peit j ; j (p ; 1) j j dt Z;1 Z1 1 j peit j dt ; j (p ; 1) j dt ;1 ;1 Z1 Z1 p 1dt ; (1 ; p) 1 dt ;1 Z ;1 1 (2p ; 1) 1 dt ;1 (2p ; 1)x j1 ;1 j peit ; (p ; 1) j dt = 1=2 = 1=2 Z1 Z;1 1 j eit + 1 j dt j cos t + i sin t + 1 j dt Z;1 1q = 1=2 (cos t + 1)2 + (sin t)2 dt ;1 Z1q = 1=2 cos2 t + 2 cos t + 1 + sin2 t dt Z;1 1p = 1=2 2 + 2 cos t dt ;1 which also does not exist. 66 Otherwise, X N (0; 1): Z1 ;1 Z1 exp(;t2 =2)dt ;1 p Z1 1 p exp(;t2 =2)dt = 2 p ;1 2 = 2 j X (t) j dt = < 1 67 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 24: Monday 10/25/1999 Scribe: Jurgen Symanzik, Hanadi B. Eltahir 3.4 Probability Generating Functions Denition 3.4.1: Let X be a discrete rv which only takes non{negative integer values, i.e., pk = P (X = k), and 1 X pk = 1. Then, the probability generating function (pgf) of X is dened as k=0 G(s) = 1 X k=0 pk sk : Theorem 3.4.2: G(s) converges for j s j< 1. Proof: 1 1 X X j G(s) j j pk sk j j pk j= 1 k=0 k=0 Theorem 3.4.3: Let X be a discrete rv which only takes non{negative integer values and has pgf G(s). Then it holds: dk G(s) j P (X = k) = k1! ds s=0 k Theorem 3.4.4: Let X be a discrete rv which only takes non{negative integer values and has pgf G(s). If E (X ) exists, then it holds: d G(s) j E (X ) = ds s=1 68 Denition 3.4.5: The kth factorial moment of X is dened as E [X (X ; 1)(X ; 2) : : : (X ; k + 1)] if this expectation exists. Theorem 3.4.6: Let X be a discrete rv which only takes non{negative integer values and has pgf G(s). If E [X (X ; 1)(X ; 2) : : : (X ; k + 1)] exists, then it holds: dk G(s) j E [X (X ; 1)(X ; 2) : : : (X ; k + 1)] = ds s=1 k Note: Similar to the Cauchy distribution for the continuous case, there exist discrete distributions where the mean (or higher moments) do not exist. See Homework. 3.5 Moment Inequalities Theorem 3.5.1: Let h(X ) be a non{negative Borel{measurable function of a rv X . If E (h(X )) exists, then it holds: P (h(X ) ) E (h(X )) 8 > 0 Proof: Continuous case only: E (h(X )) = = Z1 Z;1 ZA ZA h(x)fX (x)dx h(x)fX (x)dx + Z h(x)fX (x)dx AC h(x)fX (x)dx j where A = fx : h(x) g fX (x)dx = P (h(X ) ) 8 > 0 A Therefore, P (h(X ) ) E (h(X )) 8 > 0. 69 Corollary 3.5.2: Markov's Inequality Let h(X ) =j X jr and = kr where r > 0 and k > 0. If E (j X jr ) exists, then it holds: P (j X j k) E (j kXr j ) r Proof: Since P (j X j k) = P (j X jr kr ) for k > 0, it follows using Theorem 3.5.1: r Th:3:5:1 P (j X j k) = P (j X jr kr ) E (j X j ) kr Corollary 3.5.3: Chebychev's Inequality Let h(X ) = (X ; )2 and = k2 2 where E (X ) = , V ar(X ) = 2 < 1, and k > 0. Then it holds: P (j X ; j> k) k12 Proof: Since P (j X ; j> k) = P (j X ; j2 > k2 2 ) for k > 0, it follows using Theorem 3.5.1: Th:3:5:1 j2) = V ar(X ) = 2 = 1 P (j X ; j> k) = P (j X ; j2 > k22 ) E (j Xk2; 2 k2 2 k2 2 k2 Theorem 3.5.4: Lyapunov Inequality Let 0 < n = E (j X jn ) < 1. For arbitrary k such that 2 k n, it holds that 1 1 (k;1 ) k;1 (k ) k ; 1 i.e., (E (j X jk;1 )) k;1 (E (j X jk )) k1 . Proof: Continuous case only: k;1 k+1 Let Q(u; v) = E (u j X j 2 +v j X j 2 )2 . Obviously, Q(u; v) 0 8u; v 2 IR. Also, Q(u; v) = Z1 ;1 (u j x j k ;1 2 +v j x j k +1 2 2 ) fX (x)dx Z1 Z1 Z1 2 k ; 1 k 2 = u j x j fX (x)dx + 2uv j x j fX (x)dx + v j x jk+1 fX (x)dx ;1 ;1 ;1 = u2 k;1 + 2uvk + v2 k+1 0 8u; v 2 IR 70 Using the fact that Ax2 + 2Bxy + Cy2 0 8x; y 2 IR i A > 0 and AC ; B 2 > 0 (see Rohatgi, page 6, Section P2.4), we get with A = k;1 ; B = k , and C = k+1 : k;1 k+1 ; k2 0 =) k2 k;1 k+1 =) k2k kk;1 kk+1 This means that 12 0 2 , 24 12 32 , 36 23 43 , and so on. Multiplying these, we get: kY ;1 j =1 j2j kY ;1 j =1 jj;1 jj+1 = (0 2 )(12 32 )(23 43 )(34 54 ) : : : (kk;;32 kk;;12 )(kk;;21 kk;1 ) = Dividing both sides by kY ;2 j =1 kY ;2 k ; 2 k ; 1 0 k;1 k j2j j =1 j2j , we get: k2;k;1 2 0 kk;1kk;;12 0 =1 k =) k;1 kk;1 ;1 =) k1;1 k k k 1 1 =) kk;;11 kk 71 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 25: Wednesday 10/27/1999 Scribe: Jurgen Symanzik 4 Random Vectors 4.1 Joint, Marginal, and Conditional Distributions Denition 4.1.1: The vector X = (X1 ; : : : ; Xn )0 on ( ; L; P ) ! IRn dened by X (!) = (X1 (!); : : : ; Xn (!))0 ; ! 2 , is an n{dimensional random vector (n{rv) if X ;1 (I ) = f! : X1 (!) a1 ; : : : ; Xn (!) an g 2 L for all n{dimensional intervals I = f(x1 ; : : : ; xn ) : ;1 < xi ai ; ai 2 IR 8i = 1; : : : ; ng. Note: It follows that if X1 ; : : : ; Xn are any n rv's on ( ; L; P ), then X = (X1 ; : : : ; Xn )0 is an n{rv on ( ; L; P ) since for any I , it holds: X ;1 (I ) = f! : (X1 (!); : : : ; Xn (!)) 2 I g = f! : X1 (!) a1 ; : : : ; Xn (!) an g \n = f| ! : Xk ({z!) ak g} k=1 | {z2L 2L } Denition 4.1.2: For an n{rv X , a function F dened by F (x) = P (X x) = P (X1 x1 ; : : : ; Xn xn ) 8x 2 IRn is the joint cumulative distribution function of X . Note: (i) F is non{decreasing and right{continuous in each of its arguments xi . (ii) xlim !1 F (x) = x1 !1lim ;:::xn !1 F (x) = 1 and xklim !;1 F (x) = 0 8x1 ; : : : ; xk;1 ; xk+1 ; : : : ; xn 2 IR. 72 However, conditions (i) and (ii) together are not sucient for F to be a joint cdf. Instead we need the conditions from the next Theorem. Theorem 4.1.3: A function F (x) = F (x1 ; : : : ; xn ) is the joint cdf of some n{rv X i (i) F is non{decreasing and right{continuous with respect to each xi , (ii) F (;1; x2 ; : : : ; xn ) = F (x1 ; ;1; x3 ; : : : ; xn ) = : : : = F (x1 ; : : : ; xn;1 ; ;1) = 0 and F (1; : : : ; 1) = 1, and (iii) 8x 2 IRn 8i > 0; i = 1; : : : n, the following inequality holds: F (x + ) ; + n X F (x1 + 1 ; : : : ; xi;1 + i;1 ; xi ; xi+1 + i+1 ; : : : ; xn + n) i=1X 1i<j n F (x1 + 1 ; : : : ; xi;1 + i;1 ; xi ; xi+1 + i+1 ; : : : ; ::: + (;1)n F (x) 0 xj ;1 + j ;1 ; xj ; xj+1 + j+1; : : : ; xn + n) Note: We won't prove this Theorem but just see why we need condition (iii) for n = 2: P (x1 < X x2; y1 < Y y2 ) = P (X x2 ; Y y2 ) ; P (X x1 ; Y y2 ) ; P (X x2 ; Y y1) + P (X x1 ; Y y1 ) 0 We will restrict ourselves to n = 2 for most of the next Denitions and Theorems but those can be easily generalized to n > 2. The term bivariate rv is often used to refer to a 2{rv and multivariate rv is used to refer to an n{rv, n 2. Denition 4.1.4: A 2{rv (X; Y ) is discrete if there exists a countable collection XXof pairs (xi ; yi ) that has probability 1. Let pij = P (X = xi ; Y = yj ) > 0 8(xi ; yj ) 2 X. Then, pij = 1 and fpij g is the joint i;j probabiliy mass function of (X; Y ). 73 Denition 4.1.5: Let (X; Y ) be a discrete 2{rv with joint pmf fpij g. Dene pi = and pj = 1 X pij = j =1 1 X i=1 pij = 1 X j =1 1 X i=1 P (X = xi ; Y = yj ) = P (X = xi ) P (X = xi ; Y = yj ) = P (Y = yj ): Then fpi g is called the marginal probability mass function of X and fpj g is called the marginal probability mass function of Y . Denition 4.1.6: A 2{rv (X; Y ) is continuous if there exists a non{negative function f such that F (x; y) = Zx Zy ;1 ;1 f (u; v) dv du 8(x; y) 2 IR2 where F is the joint cdf of (X; Y ). We call f the joint probability density function of (X; Y ). Note: If F is continuous at (x; y), then d2 F (x; y) = f (x; y): dx dy Denition 4.1.7: Z1 Let (X; Y ) be a continuous 2{rv with joint pdf f . Then fX (x) = f (x; y)dy is called the ;1 Z1 marginal probability density function of X and fY (y) = f (x; y)dx is called the marginal ;1 probability density function of Y . Note: (i) Z1 ;1 fX (x)dx = Z 1 Z 1 ;1 ;1 f (x; y)dy dx = F (1; 1) = 1 = and fX (x) 0 8x 2 IR and fY (y) 0 8y 2 IR. Z 1 Z 1 ;1 ;1 f (x; y)dx dy = Z1 ;1 (ii) Given a 2{rv (X; Y ) with joint cdf F (x; y), how do we generate a marginal cdf FX (x) = P (X x) ? | The answer is P (X x) = P (X x; ;1 < Y < 1) = F (x; 1). 74 fY (y)dy Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 26: Friday 10/29/1999 Scribe: Jurgen Symanzik Denition 4.1.8: If FX (x1 ; : : : ; xn ) = FX (x) is the joint cdf of an n{rv X = (X1 ; : : : ; Xn ), then the marginal cumulative distribution function of (Xi1 ; : : : ; Xik ); 1 k n ; 1; 1 i1 < i2 < : : : < ik n, is given by lim F (x) = FX (1; : : : ; 1; xi1 ; 1; : : : ; 1; xi2 ; 1; : : : ; 1; xik ; 1; : : : ; 1): xi !1;i6=i1;:::;ik X Note: In Denition 1.4.1, we dened conditional probability distributions in some probability space ( ; L; P ). This denition extends to conditional distributions of 2{rv's (X; Y ). Denition 4.1.9: Let (X; Y ) be a discrete 2{rv. If P (Y = yj ) = pj > 0, then the conditional probability mass function of X given Y = yj (for xed j ) is dened as pijj = P (X = xi j Y = yj ) = P (XP=(Yxi=; Yy =) yj ) = ppij : j j Note: For a continuous 2{rv (X; Y ) with pdf f , P (X x j Y = y) is not dened. Let > 0 and suppose that P (y ; < Y y + ) > 0. For every x and every interval (y ; ; y + ], consider the conditional probability of X x given Y 2 (y ; ; y + ]. We have P (X x j y ; < Y y + ) = P (XP (y x;; y ;< Y<Yy + y)+ ) which is well{dened if P (y ; < Y y + ) > 0 holds. So, when does exist? See the next denition. lim P (X x j Y 2 (y ; ; y + ]) !0+ 75 Denition 4.1.10: The conditional cumulative distribution function of a rv X given that Y = y is dened to be FX jY (x j y) = lim P (X x j Y 2 (y ; ; y + ]) !0+ provided that this limit exists. If it does exist, the conditional probability density function of X given that Y = y is any non{negative function fX jY (x j y) satisfying FX jY (x j y) = Zx f (t j y)dt ;1 X jY 8x 2 IR: Note: Z1 For xed y, fX jY (x j y) 0 and fX jY (x j y)dx = 1. So it is really a pdf. ;1 Theorem 4.1.11: Let (X; Y ) be a continuous 2{rv with joint pdf fX;Y . It holds that at every point (x; y) where f is continuous and the marginal pdf fY (y) > 0, we have x; Y 2 (y ; ; y + ]) FX jY (x j y) = ! lim0+ P (XP (Y 2 (y ; ; y + ]) 1 0 Z x Z y+ 1 f ( u; v ) dv du B 2 ;1 y; X;Y CC = ! lim0+ B B@ CA Z y+ 1 f (v)dv Zx ;1 = = 2 y; Y fX;Y (u; y)du fY (y) Z x fX;Y (u; y) du: ;1 fY (y) (x;y) Thus, fX jY (x j y) exists and equals fX;Y fY (y) , provided that fY (y) > 0. Furthermore, since Zx ;1 fX;Y (u; y)du = fY (y)FX jY (x j y); we get the following marginal cdf of X : FX (x) = Z 1 Z x ;1 ;1 fX;Y (u; y)du dy = 76 Z1 ;1 fY (y)FX jY (x j y)dy Example 4.1.12: Consider fX;Y (x; y) = ( 2; 0 < x < y < 1 0; otherwise We calculate the marginal pdf's fX (x) and fY (y) rst: fX (x) = and Z1 ;1 fY (y) = fX;Y (x; y)dy = Z1 ;1 Z1 x fX;Y (x; y)dx = 2dy = 2(1 ; x) for 0 < x < 1 Zy 0 2dx = 2y for 0 < y < 1 The conditional pdf's fY jX (y j x) and fX jY (x j y) are calculated as follows: (x; y) = 2 = 1 for x < y < 1 (where 0 < x < 1) fY jX (y j x) = fX;Y f (x) 2(1 ; x) 1 ; x X and (x; y) = 2 = 1 for 0 < x < y (where 0 < y < 1) fX jY (x j y) = fX;Y f (y ) 2y y Y Thus, it holds that Y j X = x U (x; 1) and X j Y = y U (0; y), i.e., both conditional pdf's are related to uniform distributions. 4.2 Independent Random Variables Example 4.2.1: (from Rohatgi, page 119, Example 1) Let f1 ; f2 ; f3 be 3 pdf's with cdf's F1 ; F2 ; F3 and let j j 1. Dene f (x1 ; x2 ; x3 ) = f1 (x1 )f2(x2 )f3(x3 ) (1 + (2F1 (x1 ) ; 1)(2F2 (x2) ; 1)(2F3 (x3 ) ; 1)): We can show (i) f is a pdf for all 2 [;1; 1]. (ii) ff : ;1 1g all have marginal pdf's f1 ; f2 ; f3 . See book for proof and further discussion | but when do the marginal distributions uniquely determine the joint distribution? 77 Denition 4.2.2: Let FX;Y (x; y) be the joint cdf and FX (x) and FY (y) be the marginal cdf's of a 2{rv (X; Y ). X and Y are independent i FX;Y (x; y) = FX (x)FY (y) 8(x; y) 2 IR2: Lemma 4.2.3: If X and Y are independent, a; b; c; d 2 IR, and a < b and c < d, then P (a < X b; c < Y d) = P (a < X b)P (c < Y d): Proof: P (a < X b; c < Y d) = FX;Y (b; d) ; FX;Y (a; d) ; FX;Y (b; c) + FX;Y (a; c) = FX (b)FY (d) ; FX (a)FY (d) ; FX (b)FY (c) + FX (a)FY (c) = (FX (b) ; FX (a))(FY (d) ; FY (c)) = P (a < X b)P (c < Y d) Denition 4.2.4: A collection of rv's X1 ; : : : ; Xn with joint cdf FX (x) and marginal cdf's FXi (xi ) are mutually (or completely) independent i FX (x) = n Y i=1 FX (xi ) 8x 2 IRn: i Note: We often simply say that the rv's X1 ; : : : ; Xn are independent when we really mean that they are mutually independent. 78 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 27: Monday 11/1/1999 Scribe: Jurgen Symanzik, Bill Morphet Theorem 4.2.5: Factorization Theorem (i) A necessary and sucient condition for discrete rv's X1 ; : : : ; Xn to be independent is that P (X = x) = P (X1 = x1 ; : : : ; Xn = xn ) = where X IRn is the countable support of X . n Y i=1 P (Xi = xi) 8x 2 X (ii) For an absolutely continuous n{rv X = (X1 ; : : : ; Xn ), X1 ; : : : ; Xn are independent i fX (x) = fX1;:::;X (x1 ; : : : ; xn ) = n n Y i=1 fX (xi ); i where fX is the joint pdf and fX1 ; : : : ; fXn are the marginal pdfs of X . Proof: (i) Discrete case: This is a generalization of Rohatgi, page 120, Theorem 1 to n-dimensional random vectors. The Theorem is based on Lemma 4.2.3 (see also Rohatgi, page 120, Lemma 1) which gives the cumulative probability of a completely bounded region in a two-dimensional support, a < X b; c < Y d. The extension of the Lemma to n-dimensional space gives the cumulative probability in a completely bounded region in an n-dimensional support (hyperspace). Let X be a random vector whose components are independent random variables of the discrete type with P (X = b) > 0. Lemma 4.2.3 extends to: lim P (a < X b) a"b = = = indep: = lim P (a1 < X1 b1 ; a2 < X2 b2 ; ai "bi 8i2f1;:::;ng P (X1 = b1 ; X2 = b2 ; : : : ; Xn = bn) : : : ; an < Xn bn ) P (X = b) n Y P (Xi = bi ) i=1 Before considering the converse factorization of the joint cdf of an n-rv, recall that independence allows each value in the support of a component to combine with each of all possible 79 combinations of the other component values: a 3-dimensional vector where the rst component has support fx1a ; x1b ; x1c g, the second component has support fx2a ; x2b ; x2c g, and the third component has support fx3a ; x3b ; x3c g, has 3 3 3 = 27 points in its support. Due to independence, all these vectors can be arranged into three sets: the rst set having x1a , the second set having x1b , and the third set having x1c , with each set having 9 combinations of x2 and x3. =) F (x1 :; x2 :; x3 :) = P (X1 = x1a )F (x2 :; x3 :) + P (X1 = x1b )F (x2 :; x3 :) + P (X1 = x1c )F (x2 :; x3 :) = F (x1 :)F (x2 :; x3 :) = F (x1 :)[P (X2 = x2a )F (x3 :) + P (X2 = x2b )F (x3 :) + P (X2 = x2c )F (x3 :)] = F (x1 :)[P (X2 = x2a ) + P (X2 = x2b ) + P (X2 = x2c )]F (x3 :) = F (x1 :)F (x2 :)F (x3 :) More generally, for n dimensions, let xi = (xi1 ; xi2 ; : : : ; xin ); B = fxi : xi1 x1 ; xi2 x2 ; : : : ; xin xng; B 2 X. Then it holds: X FX (x) = P (X = xi) xi 2B = indep: = = = X xi 2B X xi 2B P (X1 = xi1 ; X2 = xi2 ; : : : ; Xn = xin ) P (X1 = xi1 )P (X2 = xi2; : : : ; Xn = xin) X xi1 x1 ; :::; xin xn X xi1 x1 P (X1 = xi1 )P (X2 = xi2 ; : : : ; Xn = xin) P (X1 = xi1 ) X X xi2 x2 ; :::; xin xn = FX1 (x1 ) indep: = FX1 (x1 ) = FX1 (x1 )FX2 (x2 ) = = ::: FX1 (x1 )FX2 (x2 ) : : : FX (xn) = n Y i=1 xi2 x2 ; :::; xin xn X xi2 x2 P (X2 = xi2; : : : Xn = xin) P (X2 = xi2 ; : : : ; Xn = xin ) P (X2 = xi2 ) X X xi3 x3 ; :::; xin xn xi3 x3 ; :::; xin xn n FX (xi ) i (ii) Continuous case: Homework 80 P (X3 = xi3 ; : : : ; Xn = xin ) P (X3 = xi3; : : : ; Xn = xin ) Theorem 4.2.6: n Y X1 ; : : : ; Xn are independent i P (Xi 2 Ai ; i = 1; : : : ; n) = P (Xi 2 Ai ) 8 Borel sets Ai 2 B i=1 (i.e., rv's are independent i all events involving these rv's are independent). Proof: Lemma 4.2.3 and denition of Borel sets. Theorem 4.2.7: Let X1 ; : : : ; Xn be independent rv's and g1 ; : : : ; gn be Borel{measurable functions. g1 (X1 ); g2 (X2 ); : : : ; gn(Xn ) are independent. Then Proof: Fg(X1 );g(X2 );:::;g(X )(h1 ; h2 ; : : : ; hn ) n = P (g(X1 ) h1; g(X2 ) h2 ; : : : ; g(Xn ) hn) () P (X1 2 g1;1 (;1; h1 ]; : : : ; Xn 2 gn;1 (;1; hn ]) = Th:=4:2:6 = = n Y i=1 n Y i=1 n Y i=1 P (Xi 2 gi;1 (;1; hi ]) P (gi (Xi ) hi) Fg (X ) (hi ) i i () holds since g1;1 (;1; h1 ] 2 B; : : : ; gn;1 (;1; hn ] 2 B Theorem 4.2.8: If X1 ; : : : ; Xn are independent, then also every subcollection Xi1 ; : : : ; Xik ; k = 2; : : : ; n ; 1, 1 i1 < i2 : : : < ik n, is independent. Denition 4.2.9: A set (or a sequence) of rv's fXn g1 n=1 is independent i every nite subcollection is independent. Note: Recall that X and Y are identically distributed i FX (x) = FY (x) 8x 2 IR according to Denition 2.2.5 and Theorem 2.2.6. 81 Denition 4.2.10: We say that fXn g1 n=1 is a set (or a sequence) of independent identically distributed (iid) rv's 1 if fXn gn=1 is independent and all Xn are identically distributed. Note: Recall that X and Y being identically distributed does not say that X = Y with probability 1. If this happens, we say that X and Y are equivalent rv's. Note: We can also extend the dention of independence to 2 random vectors X n1 and Y n1 : X and Y are independent i FX;Y (x; y) = FX (x)FY (y ) 8x; y 2 IRn . This does not mean that the components Xi of X or the components Yi of Y are independent. However, it does mean that each pair of components (Xi ; Yi ) are independent, any subcollections (Xi1 ; : : : ; Xik ) and (Yj1 ; : : : ; Yjl ) are independent, and any Borel{measurable functions f (X ) and g(Y ) are independent. Corollary 4.2.11: (to Factorization Theorem 4.2.5) If X and Y are independent rv's, then FX jY (x j y) = FX (x) 8x; and FY jX (y j x) = FY (y) 8y: 4.3 Functions of Random Vectors Theorem 4.3.1: If X and Y are rv's on ( ; L; P ) ! IR, then (i) X Y is a rv. (ii) XY is a rv. (iii) If f! : Y (!) = 0g = , then XY is a rv. 82 Theorem 4.3.2: Let X1 ; : : : ; Xn be rv's on ( ; L; P ) ! IR. Dene MAXn = maxfX1 ; : : : ; Xn g = X(n) by MAXn(!) = maxfX1 (!); : : : ; Xn (!)g 8! 2 and MINn = minfX1 ; : : : ; Xng = X(1) = ; maxf;X1 ; : : : ; ;Xn g by MINn (!) = minfX1 (!); : : : ; Xn (!)g 8! 2 : Then, (i) MINn and MAXn are rv's. (ii) If X1 ; : : : ; Xn are independent, then FMAX (z ) = P (MAXn z) = P (Xi z 8i = 1; : : : ; n) = n and FMIN (z ) = P (MINn z ) = 1 ; P (Xi > z 8i = 1; : : : ; n) = 1 ; n n Y i=1 n Y FX (z ) i (1 ; FXi (z )): i=1 (iii) If fXi gni=1 are iid rv's with common cdf FX , then FMAX (z ) = FXn (z) n and FMIN (z ) = 1 ; (1 ; FX (z ))n : If FX is absolutely continuous with pdf fX , then the pdfs of MAXn and MINn are n fMAX (z) = n FXn;1 (z ) fX (z) n and fMIN (z) = n (1 ; FX (z))n;1 fX (z) for all continuity points of FX . n 83 Note: Using Theorem 4.3.2, it is easy to derive the joint cdf and pdf of MAXn and MINn for iid rv's fX1 ; : : : ; Xn g. For example, if the Xi's are iid with cdf FX and pdf fX , then the joint pdf of MAXn and MINn is fMAX n ;MINn (x; y) = ( 0; xy n ; 2 n(n ; 1) (FX (x) ; FX (y)) fX (x)fX (y); x > y However, note that MAXn and MINn are not independent. See Rohatgi, page 129, Corollary, for more details. 84 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 28: Wednesday 11/3/1999 Scribe: Jurgen Symanzik, Rich Madsen Note: The previous transformations are special cases of the following Theorem: Theorem 4.3.3: If g : IRn ! IRm is a Borel{measurable function (i.e., 8B 2 Bm : g;1 (B ) 2 Bn ) and if X = (X1 ; : : : ; Xn ) is an n{rv, then g(X ) is an m{rv. Proof: If B 2 Bm , then f! : g(X (!)) 2 B g = f! : X (!) 2 g;1 (B )g 2 Bn . Question: How do we handle more general transformations of X ? Discrete Case: Let X = (X1 ; : : : ; Xn ) be a discrete n{rv and X IRn be the countable support of X , i.e., P (X 2 X) = 1 and P (X = x) > 0 8x 2 X. Dene ui = gi (x1 ; : : : ; xn ); i = 1; : : : ; n to be 1{to{1-mappings of X onto B . Let u = (u1 ; : : : ; un )0 . Then P (U = u) = P (g1 (X ) = u1 ; : : : ; gn (X ) = un) = P (X1 = h1 (u); : : : ; Xn = hn (u)) 8u 2 B where xi = hi (u); i = 1; : : : ; n, is the inverse transformation (and P (U = u) = 0 8u 62 B ). The joint marginal pmf of any subcollection of ui 's is now obtained by summing over the other remaining uj 's. 85 Example 4.3.4: Let X; Y be iid Bin(n; p); 0 < p < 1. Let U = YX+1 and V = Y + 1. Then X = UV and Y = V ; 1. So the joint pmf of U; V is P (U = u; V = v) = ! ! n puv (1 ; p)n;uv n pv;1 (1 ; p)n+1;v uv v;1 ! ! n n puv+v;1 (1 ; p)2n+1;uv;v = uv v;1 for v 2 f1; 2; : : : ; n + 1g and uv 2 f0; 1; : : : ; ng. Continuous Case: Let X = (X1 ; : : : ; Xn ) be a continuous n{rv with joint cdf FX and joint pdf fX . Let 0 B U =B @ 1 0 g (X ) BB 1 .. .. C C = g ( X ) = . A @ . U1 gn (X ) Un i.e., Ui = gi (X ), be a mapping from IRn into IRn . If B 2 Bn , then R R 1 CC ; A R R n Y P (U 2 B ) = P (X 2 g;1 (B )) = g;:1:(:B ) fX (x)d(x) = g;:1:(:B ) fX (x) dxi i=1 where g;1 (B ) = fx = (x1 ; : : : ; xn ) 2 IRn : g(x) 2 B g. Suppose we dene B as the half{innite n{dimensional interval Bu = f(u01 ; : : : ; u0n ) : ;1 < u0i < ui 8i = 1; : : : ; ng for any u 2 IRn . Then the joint cdf of U is R :::R G(u) = P (U 2 Bu ) = P (g1 (X ) u1 ; : : : ; gn (X ) un) = g;1 (B ) fX (x)d(x): u u) If G happens to be absolutely continuous, the joint pdf of U will be given by fU (u) = @u1@@uG2(:::@u n at every continuity point of fU . n Under certain conditions, we can write fU in terms of the original pdf fX of X as stated in the next Theorem: 86 Theorem 4.3.5: Multivariate Transformation Let X = (X1 ; : : : ; Xn ) be a continuous n{rv with joint pdf fX . (i) Let 0 B U =B @ 1 0 g (X ) 1 B . .. C C B . A = g(X ) = @ .. U1 1 CC ; A gn (X ) (i.e., Ui = gi (X )) be a 1{to{1{mapping from IRn into IRn , i.e., there exist inverses hi , i = 1; : : : ; n, such that xi = hi (u) = hi (u1 ; : : : ; un ); i = 1; : : : ; n, over the range of the transformation g. Un (ii) Assume both g and h are continuous. @xi = @hi (u) ; i; j = 1; : : : ; n, exist and are continuous. (iii) Assume partial derivatives @u @uj j (iv) Assume that the Jacobian of the inverse transformation @x1 @x1 @ (x ; : : : ; x ) @u 1 : : : @u .. J = @ (u1 ; : : : ; un ) = ... . 1 n @x @x @u1 : : : @u n n n n is dierent from 0 for all u in the range of g. Then the n{rv U = g(X ) has a joint absolutely continuous cdf with corresponding joint pdf fU (u) =j J j fX (h1 (u); : : : ; hn (u)): Proof: Let u 2 IRn and Bu = f(u01 ; : : : ; u0n ) : ;1 < u0i < ui 8i = 1; : : : ; ng: Then, R :::R GU (u) = g;1 (B ) fX (x)d(x) u R :::R Bu fX (h1 (u); : : : ; hn (u)) j J j d(u) The result follows from dierentiation of GU . = For additional steps of the proof see Rohatgi (page 135 and Theorem 17 on page 10) or a book on multivariate calculus. 87 Theorem 4.3.6: Let X = (X1 ; : : : ; Xn ) be a continuous n{rv with joint pdf fX . (i) Let 0 B U =B @ 1 0 g (X ) 1 B .. C . C B . A = g(X ) = @ .. U1 Un (i.e., Ui = gi (X )) be a mapping from IRn into IRn . gn (X ) 1 CC ; A (ii) Let X = fx : fX (x) > 0g be the support of X . (iii) Suppose that for each u 2 B = fu 2 IRn : u = g(x) for some x 2 Xg there is a nite number k = k(u) of inverses. (iv) Suppose we can partition X into X0 ; X1 ; : : : ; Xk s.t. (a) P (X 2 X0 ) = 0. (b) U = g(X ) is a 1{to{1{mapping 0 h (u) 1 from Xl onto B for all l = 1; : : : ; k, with inverse transl1 B . C formation hl (u) = B @ .. CA ; u 2 B , i.e., for each u 2 B , hl (u) is the unique x 2 Xl hln (u) such that u = g(x). @xi = @hli (u) ; l = 1; : : : ; k; i; j = 1; : : : ; n, exist and are continuous. (v) Assume partial derivatives @u @uj j (vi) Assume the Jacobian of each of the inverse transformations @x1 @x1 @h 1 : : : @u1 : : : @u @u1 . . .. = ... Jl = .. @x @x @h : : : @u1 @u @u1 : : : l n n n n ln .. ; l = 1; : : : ; k; . @h @u @hl1 @un ln n is dierent from 0 for all u in the range of g. Then the joint pdf of U is given by fU (u) = k X l=1 j Jl j fX (hl1 (u); : : : ; hln(u)): 88 Example 4.3.7: Let X; Y be iid N (0; 1). Dene U = g1 (X; Y ) = ( Y= 6 0 0; Y = 0 X; Y and V = g2 (X; Y ) =j Y j : X = IR2 , but U; V are not 1{to{1 mappings since (U; V )(x; y) = (U; V )(;x; ;y), i.e., conditions do not apply for the use of Theorem 4.3.5. Let X0 = f(x; y) : y = 0g X1 = f(x; y) : y > 0g X2 = f(x; y) : y < 0g Then P ((X; Y ) 2 X0 ) = 0. Let B = f(u; v) : v > 0g = g(X1 ) = g(X2 ). Inverses: B ! X1 : x y B ! X2 : x y h11 (u; v) = uv h12 (u; v) = v h21 (u; v) = ;uv h22 (u; v) = ;v v u ) jJ1 j =j v j J1 = = = = = 0 1 ; v ; u ) jJ2 j =j v j J2 = 0 ;1 fX;Y (x; y) = 21 e;x2 =2 e;y2 =2 fU;V (u; v) = j v j 21 e;(uv)2 =2 e;v2 =2 + j v j 21 e;(;uv)2 =2 e;(;v)2 =2 ;( 2 +1) 2 = v e 2 ; ;1 < u < 1; 0 < v < 1 u Marginal: fU (u) = = Z1v 0 Z1 0 e ;(u2 +1)v2 2 v 2 2 dz dv j z = (u +2 1)v ; dv = (u2 + 1)v 1 ;z (u2 + 1) e dz 89 1 = (u21+ 1) (;e;z ) 0 = (1 +1 u2 ) ; ;1 < u < 1 Thus, the quotient of two iid N (0; 1) rv's is a rv that has a Cauchy distribution. 90 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 29: Friday 11/5/1999 Scribe: Jurgen Symanzik, Hanadi B. Eltahir 4.4 Order Statistics Denition 4.4.1: Let (X1 ; : : : ; Xn ) be an n{rv. The kth order statistic X(k) is the kth smallest of the Xi0 s, i.e., X(1) = minfX1 ; : : : ; Xn g, X(2) = minffX1 ; : : : ; Xn g ; X(1) g, : : :, X(n) = maxfX1 ; : : : ; Xn g. It is X(1) X(2) : : : X(n) and fX(1) ; X(2) ; : : : ; X(n) g is the set of order statistics for (X1 ; : : : ; Xn ). Note: As shown in Theorem 4.3.2, X(1) and X(n) are rv's. This result will be extended in the following Theorem: Theorem 4.4.2: Let (X1 ; : : : ; Xn ) be an n{rv. Then the kth order statistic X(k) , k = 1; : : : ; n, is also an rv. Theorem 4.4.3: Let X1 ; : : : ; Xn be continuous iid rv's with pdf fX . The joint pdf of X(1) ; : : : ; X(n) is 8 Y n > < n! fX (xi); x1 x2 : : : xn fX(1);:::;X( ) (x1 ; : : : ; xn ) = > i=1 : 0; otherwise n Proof: For the case n = 3, look at the following scenario how X1 ; X2 , and X3 can be possibly ordered to yield X(1) < X(2) < X(3) . Columns represent X(1) ; X(2) , and X(3) . Rows represent X1 ; X2 , and X3 : 1 X1 < X2 < X3 : 0 0 1 X1 < X3 < X2 : 0 0 91 0 0 1 0 0 1 0 0 0 1 1 0 X2 < X1 < X3 : X2 < X3 < X1 : X3 < X1 < X2 : 0 1 0 0 1 0 0 0 1 0 0 1 1 0 0 0 0 1 0 1 0 0 1 0 1 0 0 1 0 0 0 1 X3 < X2 < X1 : 1 0 0 0 For n = 3, there are 3! = 6 possible arrangements. In general, there are n! arrangements of X1 ; : : : ; Xn for each (X(1) ; : : : ; X(n) ). This mapping is not 1{to{1. For each mapping, we have a n n matrix J that results from an n n identity matrix through the rearrangement or rows. Therefore, j J j= 1. By Theorem 4.3.6, we get fX(1) ;:::;X( ) (x(1) ; : : : ; x(n) ) = n!fX1;:::;X (x(k1 ) ; x(k2 ) ; : : : ; x(k ) ) n = n! = n! n Y i=1 n Y i=1 n n fX (x(k ) ) i i fX (xi ) Theorem 4.4.4: Let X1 ; : : : ; Xn be continuous iid rv's with pdf fX and cdf FX . Then the following holds: (i) The marginal pdf of X(k) , k = 1; : : : ; n, is n! k;1 n;k fX(k) (x) = (k ; 1)!( n ; k)! (FX (x)) (1 ; FX (x)) fX (x): (ii) The joint pdf of X(j ) and X(k) , 1 j < k n, is fX(j);X(k) (xj ; xk ) = (j ; 1)!(k ; jn!; 1)!(n ; k)! (FX (xj ))j ;1 (FX (xk ) ; FX (xj ))k;j ;1 (1 ; FX (xk ))n;k fX (xj )fX (xk ) if xj < xk and 0 otherwise. 92 4.5 Multivariate Expectation In this section, we assume that X = (X1 ; : : : ; Xn ) is an n{rv and g : IRn ! IRn is a Borel{ measurable function. Denition 4.5.1: If n = 1, i.e., g is univariate, we dene the following: (i) Let xi1 ; : : : ; Xn = xin ). If X X be discrete with joint pmf pi1;:::;in = P (X1 = X pi1;:::;in g(xi1 ; : : : ; xin ) and pi1 ;:::;in j g(xi1 ; : : : ; xin ) j< 1, we dene E (g(X )) = i1 ;:::;in i1 ;:::;in this value exists. (ii) Let X be continuous with joint pdf fX (x). If E (g(X )) = Z IRn Z g(x)fX (x)dx and this value exists. IRn j g(x) j fX (x) dx < 1, we dene Note: The above can be extended to vector{valued functions g (n > 1) in the obvious way. For example, if g is the identity mapping from IRn ! IRn , then 0 E (X ) 1 B . E (X ) = B @ .. E (Xn ) 1 0 CC = BB A @ 1 .. C . C A 1 n provided that E (j Xi j) < 1 8i = 1; : : : ; n. Similarly, provided that all expectations exist, we get for the variance{covariance matrix: V ar(X ) = X = E ((X ; E (X )) (X ; E (X ))0 ) with (i; j )th component E ((Xi ; E (Xi )) (Xj ; E (Xj ))) = Cov(Xi ; Xj ) and with (i; i)th component E ((Xi ; E (Xi )) (Xi ; E (Xi ))) = V ar(Xi ) = i2 : Joint higher{order moments can be dened similarly when needed. 93 Note: We are often interested in (weighted) sums of rv's or products of rv's and their expectations. This will be addressed in the next two Theorems: Theorem 4.5.2: n X Let Xi ; i = 1; : : : ; n, be rv's such that E (j Xi j) < 1. Let a1 ; : : : ; an 2 IR and dene S = ai Xi . i=1 Then it holds that E (j S j) < 1 and E (S ) = Proof: Continuous case only: E (j S j) = = = = Z j n X IRn i=1 Z X n n X i=1 aiE (Xi ): ai xi j fX (x)dx j ai j j xi j fX (x)dx IRn i=1 Z Z n X j ai j j xi j n;1 fX (x)dx1 : : : dxi;1 dxi+1 : : : dxn dxi IR IR i=1 Z n X j ai j j xi j fXi (xi )dxi IR i=1 n X j ai j E (j Xi j) i=1 < 1 It follows that E (S ) = n X i=1 ai E (Xi ) by the same argument without using the absolute values j Note: If Xi ; i = 1; : : : ; n, are iid with E (Xi ) = , then E (X ) = E ( n1 n X i=1 Xi) = 94 n 1 X i=1 n E (Xi ) = : j. Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 30: Monday 11/8/1999 Scribe: Jurgen Symanzik, Bill Morphet Theorem 4.5.3: Let Xi ; i = 1; : : : ; n, be independent rv's such that E (j Xi j) < 1. Let gi ; i = 1; : : : ; n, be Borel{ measurable functions. Then n n Y Y E ( gi (Xi )) = E (gi (Xi )) i=1 i=1 if all expectations exist. Proof: n Y By Theorem 4.2.5, fX (x) = fXi (xi ), and by Theorem 4.2.7, gi (Xi ); i = 1; : : : ; n, are also indei=1 pendent. Therefore, Yn E ( gi (xi )) i=1 = Th:= 4:2:5 = Th:=4:2:7 = = Z Y n IRn i=1 Z Y n gi (xi )fX (x)dx (gi (xi )fXi (xi )dxi ) IRn i=1 Z IR Z ::: Z Y n IR i=1 gi (xi ) n Y i=1 Z fX (xi ) i n Y i=1 dxi Z g1 (x1 )fX1 (x1 )dx1 g2 (x2 )fX2 (x2 )dx2 : : : gn (xn )fXn (xn)dxn IR IR IR n Z Y gi (xi )fXi (xi)dxi i=1 IR n Y E (gi (Xi )) i=1 Corollary 4.5.4: If X; Y are independent, then Cov(X; Y ) = 0. Theorem 4.5.5: Two rv's X; Y are independent i for all pairs of Borel{measurable functions g1 and g2 it holds that E (g1 (X ) g2 (Y )) = E (g1 (X )) E (g2 (Y )) if all expectations exist. Proof: \=)": It follows from Theorem 4.5.3 and the independence of X and Y that E (g1 (X )g2 (Y )) = E (g1 (X )) E (g2 (Y )): 95 \(=": From Theorem 4.2.6, we know that X and Y are independent i P (X 2 A1 ; Y 2 A2 ) = P (X 2 A1 ) P (Y 2 A2 ) 8 Borel sets A1 and A2. How do we relate Theorem 4.2.6 to g1 and g2 ? Let us dene two Borel{measurable functions g1 and g2 as: ( g1 (x) = IA1 (x) = 1; x 2 A1 0; otherwise ( g2(y) = IA2 (y) = 1; y 2 A2 0; otherwise Then, E (g1 (X )) = 0 P (X 2 Ac1 ) + 1 P (X 2 A1 ) = P (X 2 A1 ); E (g2 (Y )) = 0 P (Y 2 Ac2 ) + 1 P (Y 2 A2 ) = P (Y 2 A2 ) and E (g1 (X ) g2 (Y )) = P (X 2 A1; Y 2 A2 ): =) P (X 2 A1 ; Y 2 A2 ) = E (g1 (X ) g2 (Y )) given = E (g1 (X )) E (g2 (Y )) = P (X 2 A1 )P (Y 2 A2 ) =) P (X 2 A1 ; Y 2 A2 ) = P (X 2 A1 ) P (Y 2 A2 ) =) X; Y independent by Theorem 4.2.6. Denition 4.5.6: th th The (ith 1 ; i2 ; : : : ; in ) multi{way moment of X = (X1 ; : : : ; Xn ) is dened as mi1 i2 :::i = E (X1i1 X2i2 : : : Xni ) n n if it exists. th th The (ith 1 ; i2 ; : : : ; in ) multi{way central moment of X = (X1 ; : : : ; Xn ) is dened as n Y i1 i2:::i = E ( (Xj ; E (Xj ))i ) n j j =1 if it exists. Note: If we set ir = is = 1 and ij = 0 8j 6= r; s in Denition 4.5.6, we get 0 : : : 0 1 0 : : : 0 1 0 : : : 0 = rs = Cov(Xr ; Xs ): " r " s 96 Theorem 4.5.7: Cauchy{Schwarz{Inequality Let X; Y be 2 rv's with nite variance. Then it holds: (i) Cov(X; Y ) exists. (ii) (E (XY ))2 E (X 2 )E (Y 2 ). (iii) (E (XY ))2 = E (X 2 )E (Y 2 ) i there exists an (; ) 2 IR2 ; f(0; 0)g such that P (X + Y = 0) = 1. Proof: Assumptions: V ar(X ); V ar(Y ) < 1. Then also E (X 2 ); E (X ); E (Y 2 ); E (Y ) < 1. Result used in proof: 0 (a ; b)2 = a2 ; 2ab + b2 =) ab a2 +2 b2 0 (a + b)2 = a2 + 2ab + b2 =) ;ab a2 +2 b2 =)j ab j a2 +2 b2 8 a; b 2 IR () (i) E (j XY j) = Z IR2 j xy j fX;Y (x; y)dx dy Z x2 + y2 2 fX;Y (x; y)dx dy IR2 Z y2 Z x2 f ( x; y ) dx dy + fX;Y (x; y)dy dx = X;Y IR2 2 IR2 2 Z x2 Z y2 () 2 fX (x)dx + IR 2 fY (y)dy 2 2 = E (X ) +2 E (Y ) = IR < 1 =) E (XY ) exists =) Cov(X; Y ) = E (XY ) ; E (X )E (Y ) exists (ii) 0 E ((X + Y )2 ) = 2 E (X 2 ) + 2E (XY ) + 2 E (Y 2 ) 8 ; 2 IR (A) If E (X 2 ) = 0, then X has a degenerate 1{point Dirac distribution and the inequality trivially is true. Therefore, we can assume that E (X 2 ) > 0. As (A) is true for all ; 2 IR, we can choose = ;EE((XXY2 ) ) ; = 1. 97 2 2 )) (E (XY )) 2 =) (EE(XY (X 2 ) ; 2 E (X 2 ) + E (Y ) 0 =) ;(E (XY ))2 + E (Y 2 )E (X 2 ) 0 =) (E (XY ))2 E (X 2 ) E (Y 2 ) (iii) When are the left and right sides of the inequality in (ii) equal? Assume that E (X 2 ) > 0. (E (XY ))2 = E (X 2 )E (Y 2 ) holds i E ((X + Y )2 ) = 0. This can only happen if P (X + Y = 0) = 1. Otherwise, if P (X + Y = 0) = P (Y = ; X ) = 1 for some (; ) 2 IR2 ; f(0; 0)g, i.e., Y is linearly dependent on X with probability 1, this implies: 2 = ( )2 (E (X 2 ))2 = E (X 2 )( )2 E (X 2 ) = E (X 2 )E (Y 2 ) (E (XY ))2 = (E (X ;X )) 4.6 Multivariate Generating Functions Denition 4.6.1: Let X = (X1 ; : : : ; Xn ) be an n{rv. We dene the multivariate moment generating function (mmgf) of X as ! n X 0X t MX (t) = E (e ) = E (exp ti Xi ) i=1 v u n uX if this expectation exists for j t j= t t2i < h for some h > 0. i=1 Denition 4.6.2: Let X = (X1 ; : : : ; Xn ) be an n{rv. We dene the n{dimensional characteristic function X : IRn ! CI of X as 0 n 1 X 0 X (t) = E (eit X ) = E (exp @i tj Xj A): j =1 Note: (i) X (t) exists for any real{valued n{rv. (ii) If MX (t) exists, then X (t) = MX (it). 98 Theorem 4.6.3: (i) If MX (t) exists, it is unique and uniquely determines the joint distribution of X . X (t) is also unique and uniquely determines the joint distribution of X . (ii) MX (t) (if it exists) and X (t) uniquely determine all marginal distributions of X , i.e., MXi (ti ) = MX (0; ti ; 0) and and Xi (ti ) = X (0; ti ; 0). (iii) Joint moments of all orders (if they exist) can be obtained as mi1 :::i if the mmgf exists and i1 +i2 +:::+i = @i1 i2 M ( t ) = E (X1i1 X2i2 : : : Xni ) @t1 @t2 : : : @tin X t=0 n n n n i1 +i2 +:::+i X (0) = E (X1i1 X2i2 : : : Xni ): mi1:::i = ii1 +i21+:::+i @i1 i2 i @t1 @t2 : : : @tn n n n n n (iv) X1 ; : : : ; Xn are independent rv's i MX (t1 ; : : : ; tn ) = MX (t1 ; 0) MX (0; t2 ; 0) : : : MX (0; tn ) 8t1 ; : : : ; tn 2 IR; given that MX (t) exists. Similarly, X1 ; : : : ; Xn are independent rv's i X (t1 ; : : : ; tn ) = X (t1 ; 0) X (0; t2 ; 0) : : : X (0; tn ) 8t1; : : : ; tn 2 IR: Proof: Rohatgi, page 162: Theorem 7, Corollary, Theorem 8, and Theorem 9 (for mmgf and the case n = 2). Theorem 4.6.4: Let X1 ; : : : ; Xn be independent rv's. (i) If mgf's MX1 (t); : : : ; MXn (t) exist, then the mgf of Y = MY (t) = n Y i=1 n X i=1 MX (ait) i on the common interval where all individual mgf's exist. (ii) The characteristic function of Y = n X j =1 aj Xj is Y (t) = 99 n Y j =1 Xj (aj t) aiXi is (iii) If mgf's MX1 (t); : : : ; MXn (t) exist, then the mmgf of X is n Y MX (t) = i=1 MX (ti) i on the common interval where all individual mgf's exist. (iv) The n{dimensional characteristic function of X is X (t) = n Y j =1 Proof: Homework (parts (ii) and (iv) only) 100 Xj (tj ): Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 31: Wednesday 11/10/1999 Scribe: Jurgen Symanzik, Rich Madsen Theorem 4.6.5: Let X1 ; : : : ; Xn be independent discrete rv's on the non{negative integers with pgf's GX1 (s); : : : ; GXn (s). n X The pgf of Y = Xi is i=1 GY (s) = n Y i=1 GX (s): i Proof: Version 1: GX (s) i = E (sX ) GY (s) = E (sY ) = E (s indep: = = i P n Y =1 Xi ) n i E (sX ) i i=1 n Y GX (s) i=1 i Version 2: (case n = 2 only) GY (s) = = indep: = = = P (Y = 0) + P (Y = 1)s + P (Y = 2)s2 + : : : P (X1 = 0; X2 = 0) + (P (X1 = 1; X2 = 0) + P (X1 = 0; X2 = 1)) s + (P (X1 = 2; X2 = 0) + P (X1 = 1; X2 = 1) + P (X1 = 0; X2 = 2)) s2 + : : : P (X1 = 0)P (X2 = 0) + (P (X1 = 1)P (X2 = 0) + P (X1 = 0)P (X2 = 1)) s + 2 (P (X1 = 2)P (X2 = 0) + P (X1 = 1)2P (X2 = 1) + P (X1 = 0)P (X2 = 2)) s + : : : P (X1 = 0) + P (X1 = 1)s + P (X1 = 2)s + : : : P (X2 = 0) + P (X2 = 1)s + P (X2 = 2)s2 + : : : GX1 (s) GX2 (s) 101 A generalized proof for n 3 needs to be done by induction on n. Theorem 4.6.6: Let X1 ; : : : ; XN be iid discrete rv's on the non{negative integers with common pgf GX (s). Let N be a discrete rv on the non{negative integers with pgf GN (s). Let N be independent of the Xi 's. Dene SN = N X i=1 Xi . The pgf of SN is GS (s) = GN (GX (s)): N Proof: P (SN = k) = =) GSN (s) = = = = Th:=4:6:5 iid = = 1 X n=0 P (SN = kjN = n) P (N = n) 1 X 1 X k=0 n=0 1 X n=0 1 X n=0 1 X n=0 1 X n=0 1 X P (SN = kjN = n) P (N = n) sk 1 X P (N = n) P (SN = kjN = n) sk k=0 1 X P (N = n) P (Sn = k) sk k=0 1 X n X P (N = n) P ( Xi = k) sk k=0 n Y i=1 i=1 i P (N = n) GX (s) P (N = n) (GX (s))n n=0 GN (GX (s)) Example 4.6.7: Starting with a single cell at time 0, after one time unit there is probability p that the cell will have split (2 cells), probability q that it will survive without splitting (1 cell), and probability r that it will have died (0 cells). It holds that p; q; r 0 and p + q + r = 1. Any surviving cells have the same probabilities of splitting or dying. What is the pgf for the # of cells at time 2? GX (s) = GN (s) = ps2 + qs + r GS (s) = p(ps2 + ps + r)2 + q(ps2 + ps + r) + r N 102 Theorem 4.6.8: Let X1 ; : : : ; XN be iid rv's with common mgf MX (t). Let N be a discrete rv on the non{negative N X integers with mgf MN (t). Let N be independent of the Xi 's. Dene SN = Xi . The mgf of SN i=1 is MSN (t) = MN (ln MX (t)): Proof: Consider the case that the Xi 's are non{negative integers: We know that GX (S ) = E (S X ) = E (eln S ) = E (e(ln S )X ) = MX (ln S ) =) MX (S ) = GX (eS ) =) MS (t) = GS (et ) Th:=4:6:6 GN (GX (et )) = GN (MX (t)) = MN (ln MX (t)) X N N In the general case, i.e., if the Xi0 s are not non{negative integers, we need results from Section 4.7 (conditional expectation) to proof this Theorem. 103 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 32: Friday 11/12/1999 Scribe: Jurgen Symanzik, Hanadi B. Eltahir 4.7 Conditional Expectation In Section 4.1, we established that the conditional pmf of X given Y = yj (for PY (yj ) > 0) is a (x;y) pmf. For continuous rv's X and Y , when fY (y) > 0, fX jY (x j y) = fX;Y fY (y) , and fX;Y and fY are continuous, then fX jY (x j y) is a pdf and it is the conditional pdf of X given Y = y. Denition 4.7.1: Let X; Y be rv's on ( ; L; P ). Let h be a Borel{measurable function. Assume that E (h(X )) exists. Then the conditional expectation of h(X ) given Y , i.e., E (h(X ) j Y ), is a rv that takes the value E (h(X ) j y). It is dened as 8X > h(x)P (X = x j Y = y); if (X; Y ) is discrete and P (Y = y) > 0 > < x2X E (h(X ) j y) = > Z 1 > : h(x)fX jY (x j y)dx; if (X; Y ) is continuous and fY (y) > 0 ;1 Note: (i) The rv E (h(X ) j Y ) = g(Y ) is a function of Y as a rv. (ii) The usual properties of expectations apply to the conditional expectation: (a) E (c j Y ) = c 8c 2 IR. (b) E (aX + b j Y ) = aE (X j Y ) + b 8a; b 2 IR. (c) If g1 ; g2 are Borel{measurable functions and if E (g1 (X )); E (g2 (X )) exist, then E (a1 g1 (X ) + a2g2 (X ) j Y ) = a1 E (g1 (X ) j Y ) + a2 E (g2 (X ) j Y ) 8a1; a2 2 IR. (d) If X 0 then E (X j Y ) 0. (e) If X1 X2 then E (X1 j Y ) E (X2 j Y ). (iii) Moments are dened in the usual way. If E (j X jr ) < 1, then E (X r j Y ) exists and is the rth conditional moment of X given Y . 104 Example 4.7.2: Recall Example 4.1.12: ( 2; 0 < x < y < 1 0; otherwise The conditional pdf's fY jX (y j x) and fX jY (x j y) have been calculated as: fY jX (y j x) = 1 ;1 x for x < y < 1 (where 0 < x < 1) and fX jY (x j y) = y1 for 0 < x < y (where 0 < y < 1). fX;Y (x; y) = So, E (X j y) = and Zyx 0 y dx = y 2 Z1 1 2 1 1 1 ; x2 1 + x 1 y E (Y j x) = 1 ; x y dy = 1 ; x 2 = 2 1 ; x = 2 : x x Therefore, we get the rv's E (X j Y ) = Y2 and E (Y j X ) = 1+2X . Theorem 4.7.3: If E (h(X )) exists, then EY (EX jY (h(X ) j Y )) = E (h(X )): Proof: Continuous case only: EY (E (h(X ) j Y )) = Z1 E (h(x) j y)fY (y)dy ;1 Z1Z1 = h(x)fX jY (x j y)fY (y)dxdy ;1 ;1 Z1 Z1 = h(x) fX;Y (x; y)dydx ;1 ;1 Z1 = h(x)fX (x)dx ;1 = E (h(X )) Theorem 4.7.4: If E (X 2 ) exists, then V arY (E (X j Y )) + EY (V ar(X j Y )) = V ar(X ): 105 Proof: V arY (E (X j Y )) + EY (V ar(X j Y )) = EY ((E (X j Y ))2 ) ; (EY (E (X j Y )))2 + EY (E (X 2 j Y ) ; (E (X j Y ))2 ) = EY ((E (X j Y ))2 ) ; (E (X ))2 + E (X 2 ) ; EY ((E (X j Y ))2 ) = E (X 2 ) ; (E (X ))2 = V ar(X ) Note: If E (X 2 ) exists, then V ar(X ) V arY (E (X j Y )). V ar(X ) = V arY (E (X j Y )) i X = g(Y ). The inequality directly follows from Theorem 4.7.4. For equality, it is necessary that EY (V ar(X j Y )) = EY ((X ; E (X j Y ))2 j Y ) = 0 which holds if X = E (X j Y ) = g(Y ). If X; Y are independent, FX jY (x j y) = FX (x) 8x. Thus, if E (h(X )) exists, then E (h(X ) j Y ) = E (h(X )). 4.8 Inequalities and Identities Lemma 4.8.1: Let a; b be positive numbers and p; q > 1 such that p1 + 1q = 1 (i.e., pq = p + q and q = p;p 1 ). Then it holds that 1 ap + 1 bq ab p q with equality i ap = bq . Proof: Fix b. Let g(a) = p1 ap + 1q bq ; ab =) g0 (a) = ap;1 ; b =! 0 =) b = ap;1 =) bq = a(p;1)q = ap g00 (a) = (p ; 1)ap;2 > 0 106 Since g00 (a) > 0, this is really a minimum. The minimum value is obtained for b = ap;1 and it is 1 ap + 1 (ap;1 )q ; aap;1 = 1 ap + 1 ap ; ap = ap( 1 + 1 ; 1) = 0: p q p q p q Since g00 (a) > 0, the minimum is unique and g(a) 0. Therefore g(a) + ab = p1 ap + 1q bq ab. Theorem 4.8.2: Holders Inequality Let X; Y be 2 rv's. Let p; q > 1 such that p1 + 1q = 1 (i.e., pq = p + q and q = p;p 1 ). Then it holds that 1 1 E (j XY j) (E (j X jp)) p (E (j Y jq )) q : Proof: In Lemma 4.8.1, let a = jX j jY j . 1 and b = 1 (E (jX jp )) p (E (jY jq )) q p q Lemma =)4:8:1 1p Ej(XjXj jp ) + 1q E (jYjYj jq ) jXY j 1 1 (E (jX j )) p (E (jY jq )) q p Taking expectations on both sides of this inequality, we get E (jXY j) 1 = p1 + 1q 1 1 q q p p (E (jX j )) (E (jY j )) 1 1 The result follows immediately when multiplying both sides with (E (j X jp )) p (E (j Y jq )) q . Note: Note that Theorem 4.5.7 (ii) (Cauchy{Schwarz{Inequality) is a special case of Theorem 4.8.2 with p = q = 2. 107 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 33: Monday 11/15/1999 Scribe: Jurgen Symanzik, Bill Morphet Theorem 4.8.3: Minkowski's Inequality Let X; Y be 2 rv's. Then it holds for 1 p < 1 that 1 1 1 (E (j X + Y jp )) p (E (j X jp )) p + (E (j Y jp )) p : Proof: E (j X + Y jp) = E (j X + Y j j X + Y jp;1 ) E ((j X j + j Y j) j X + Y jp;1 ) = E (j X j j X + Y jp;1) + E (j Y j j X + Y jp;1 ) Th:4:8:2 () = 1 1 ( E (j X jp ) ) p ( E ( (j X + Y jp;1 )q ) ) q 1 1 + ( E (j Y jp ) ) p ( E ( (j X + Y jp;1 )q ) ) q (A) 1 1 1 ( E (j X jp ) ) p + ( E (j Y jp ) ) p ( E (j X + Y jp ) ) q 1 Divide the left and right side of this inequality by (E (j X + Y jp )) q . p ))1; q1 = (E (j X + Y jp )) p1 , and the result on the right The result on the left side is ( E ( j X + Y j 1 1 side is ( E (j X jp ) ) p + ( E (j Y jp) ) p . Therefore, Theorem 4.8.3 holds. In (A), we dene q = p;p 1 . Therefore, p1 + 1q = 1 and (p ; 1)q = p. () holds since p1 + 1q = 1, a condition to meet Holder's Inequality. Denition 4.8.4: A function g(x) is convex if g(x + (1 ; )y) g(x) + (1 ; )g(y) 8x; y 2 IR 80 < < 1: Note: (i) Geometrically, a convex function falls above all of its tangent lines. Also, a connecting line between any pairs of points (x; g(x)) and (y; g(y)) in the 2{dimensional plane always falls above the curve. 108 (ii) A function g(x) is concave i ;g(x) is convex. Theorem 4.8.5: Jensen's Inequality Let X be a rv. If g(x) is a convex function, then E (g(X )) g(E (X )) given that both expectations exist. Proof: Construct a tangent line l(x) to g(x) at the (constant) point x0 = E (X ): l(x) = ax + b for some a; b 2 IR The function g(x) is a convex function and falls above the tangent line l(x) =) g(x) ax + b 8x 2 IR =) E (g(X )) E (aX + b) = aE (X ) + b = l(E (X )) tangent= point g(E (X )) Therefore, Theorem 4.8.5 holds. Note: Typical convex functions g are: (i) g1 (x) =j x j ) E (j X j) j E (X ) j. (ii) g2 (x) = x2 ) E (X 2 ) (E (X ))2 ) V ar(X ) 0. (iii) g3 (x) = x1p for x > 0; p > 0 ) E ( X1p ) (E (X1 ))p ; for p = 1: E ( X1 ) E (1X ) (iv) Other convex functions are xp for x > 0; p 1; x for > 1; ; ln(x) for x > 0; etc. (v) Recall that if g is convex and dierentiable, then g00 (x) 0 8x. (vi) If the function g is concave, the direction of the inequality in Jensen's Inequality is reversed, i.e., E (g(X )) g(E (X )). Example 4.8.6: Given the real numbers a1 ; a2 ; : : : ; an > 0, we dene arithmetic mean : aA = n1 (a1 + a2 + : : : + an) = n1 109 n X i=1 ai 1 geometric mean : aG = (a1 a2 : : : an) = n n !1 Y i=1 ai n 1 = X harmonic mean : aH = 1 1 1 1 n 1 1 1 n a1 + a2 + : : : + an n Let X be a rv that takes values a1 ; a2 ; : : : ; an > 0 with probability n1 each. (i) aA aG : ln(aA ) n X ln n1 ai = ! i=1 ln(E (X )) ln concave E (ln(X )) n 1 X = n ln(ai ) = i=1 n 1X ln(a ) = n i=1 i n 1 ln( Y a) = i i=1 n Y 1 ln(( ai ) n ) i=1 ln(aG ) n = = Taking the anti{log of both sides gives aA aG . (ii) aA aH : 1 aA = 1 1 n n X i=1 ai 1 E (X ) 1=X convex E( 1 ) = X = = = Inverting both sides gives aA aH . n 11 X i=1n ai 1 + 1 + ::: 1 n a1 a2 an 1 1 aH 110 i=1 ai (iii) aG aH : ; ln(aH ) ln(a;H1 ) ln( a1 ) 1H 1 1 1 = ln n ( a + a + : : : + a ) 1 2 n 1 = ln(E ( X )) ln concave E (ln( 1 )) = = = = = n 1 X i=1 X 1) ln( n a n 1X n i=1 n 1X i ln( a1 ) i n i=1 ; ln ai n Y ; n1 ln( ai) i=1 n Y = ; ln( ai ) 1 i=1 = ; ln aG Multiplying both sides with ;1 gives ln aH ln aG . Then taking the anti{log of both sides gives aH aG . = n In summary, aH aG aA . Note that it would have been sucient to prove steps (i) and (iii) only to establish this result. However, step (ii) has been included to provide another example how to apply Theorem 4.8.5. Theorem 4.8.7: Covariance Inequality Let X be a rv with nite mean . (i) If g(x) is non{decreasing, then E (g(X )(X ; )) 0 if this expectation exists. (ii) If g(x) is non{decreasing and h(x) is non{increasing, then E (g(X )h(X )) E (g(X ))E (h(X )) if all expectations exist. 111 (iii) If g(x) and h(x) are both non{decreasing or if g(x) and h(x) are both non{increasing, then E (g(X )h(X )) E (g(X ))E (h(X )) if all expectations exist. Proof: Homework 112 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 34: Wednesday 11/17/1999 Scribe: Jurgen Symanzik, Rich Madsen 5 Particular Distributions 5.1 Multivariate Normal Distributions Denition 5.1.1: A rv X has a (univariate) Normal distribution, i.e., X N (; 2 ) with 2 IR and > 0, i it has the pdf 1 2 fX (x) = p 1 2 e; 22 (x;) : 2 X has a standard Normal distribution i = 0 and 2 = 1, i.e., X N (0; 1). Note: If X N (; 2 ), then E (X ) = and V ar(X ) = 2 . If X1 N (1 ; 12 ); X2 N (2 ; 22 ) and c1 ; c2 2 IR, then Y = c1 X1 + c2 X2 N (c1 1 + c2 2 ; c21 12 + c22 22 ). Denition 5.1.2: A 2{rv (X; Y ) has a bivariate Normal distribution i there exist constants a11 ; a12 ; a21 ; a22 ; 1 ; 2 2 IR and iid N (0; 1) rv's Z1 and Z2 such that X = 1 + a11 Z1 + a12 Z2 ; Y = 2 + a21 Z1 + a22Z2 : If we dene ! ! ! ! A = a11 a12 ; = 1 ; X = X ; Z = Z1 ; a21 a22 2 Y Z2 then we can write X = AZ + : Note: E (X ) = 1 + a11 E (Z1 ) + a12 E (Z2 ) = 1 and E (Y ) = 2 + a21 E (Z1 ) + a22 E (Z2 ) = 2 . The marginal distributions are X N (1 ; a211 + a212 ) and Y N (2 ; a221 + a222 ). Thus, X and Y have (univariate) Normal marginal densities or degenerate marginal densities (which correspond to Dirac distributions) if ai1 = ai2 = 0. 113 Theorem 5.1.3: Dene g : IR2 ! IR2 as g(x) = Cx + d. If X is a bivariate Normal rv, then g (X ) also is a bivariate Normal rv. Proof: g(X ) = CX + d = C (AZ + ) + d = (|CA {z }) Z + (C + d) another matrix | {z } another vector ~ + ~ which represents another bivariate Normal distribution = AZ Note: 1 2 = Cov(X; Y ) = Cov(a11 Z1 + a12 Z2 ; a21 Z1 + a22 Z2 ) = a11 a21 Cov(Z1 ; Z1 ) + (a11 a22 + a12 a21 )Cov(Z1 ; Z2 ) + a12 a22 Cov(Z2 ; Z2 ) = a11 a21 + a12 a22 since Z1 ; Z2 are iid N (0; 1) rv's. Denition 5.1.4: The variance{covariance matrix of (X; Y ) is a a = AA0 = 11 12 a21 a22 ! a11 a21 a12 a22 ! a211 + a212 a11 a21 + a12 a22 = a11 a21 + a12 a22 a221 + a222 ! ! 12 1 2 : = 1 2 22 Theorem 5.1.5: Assume that 1 > 0; 2 > 0 and j j< 1. Then the joint pdf of X = (X; Y ) = AZ + (as dened in Denition 5.1.2) is 1 1 0 ; 1 fX (x) = p exp ; 2 (x ; ) (x ; ) 2 j j x ; 1 2 x ; 1 y ; 2 y ; 2 2!! 1 1 p = exp ; 2(1 ; 2 ) ; 2 1 2 + 2 21 2 1 ; 2 1 Proof: The mapping Z ! X is 1{to{1: X = AZ + 114 =) Z = A;1 (X ; ) J = jA;1 j = jA1 j jAj = q (requirement is that A is invertible.) q q q q q jAj2 = jAj jAT j = jAAT j = jj = 1222 ; 212 22 = 12 1 ; 2 We can use this result to get to the second line of the theorem: 2 2 z1 z2 1 1 ; ; 2 2 p p e e fZ (z ) = 2 2 1 = 21 e; 2 (zT z) As already stated, the mapping from Z to X is 1{to{1, so we can apply Theorem 4.3.5: 1 exp(; 1 (x ; )T (A;1 )T A;1 (x ; )) p fX (x) = | {z;1 } 2 2 jj 1 1 () p exp(; 2 (x ; )T ;1(x ; )) = 2 jj This proves the 1st line of the Theorem. Step () holds since (A;1 )T A;1 = (AT );1 A;1 = (AAT );1 = ;1 : The second line of the Theorem is based on the following transformations: q q jj = 12 1 ; 2 ;1 = 1 jj 0 = @ =) fX (x) = 22 ;1 2 ;12 12 1 12 (1;2 ) ; 1 2 (1;2 ) 1p ; 1 2 (1;2 ) 1 22 (1;2 ) 21 2 1 ; 2 ! 1 A 1 exp ; 2(1 ; 2 ) x ; 1 2 1 x ; 1 y ; 2 y ; 2 2!! ; 2 + 1 2 2 Note: In the situation of Theorem 5.1.5, we say that (X; Y ) N (1 ; 2 ; 12 ; 22 ; ). Theorem 5.1.6: If (X; Y ) has a non{degenerate N (1 ; 2 ; 12 ; 22 ; ) distribution, then the conditional distribution of X given Y = y is N (1 + 1 (y ; 2 ); 12 (1 ; 2 )): 2 115 Proof: Homework Example 5.1.7: Let rv's (X1 ; Y1 ) be N (0; 0; 1; 1; 0) with pdf f1 (x; y) and (X2 ; Y2 ) be N (0; 0; 1; 1; ) with pdf f2 (x; y). Let (X; Y ) be the rv that corresponds to the pdf fX;Y (x; y) = 21 f1 (x; y) + 21 f2 (x; y): (X; Y ) is a bivariate Normal rv i = 0. However, the marginal distributions of X and Y are always N (0; 1) distributions. See also Rohatgi, page 229, Remark 2. 116 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 35: Friday 11/19/1999 Scribe: Jurgen Symanzik, Bill Morphet Theorem 5.1.8: The mgf MX (t) of a bivariate Normal rv X = (X; Y ) is MX (t) = MX;Y (t1 ; t2 ) = exp(0 t + 12 t0 t) = exp 1 t1 + 2t2 + 12 (12 t21 + 22 t22 + 21 2 t1 t2 ) : Proof: The mgf of a univariate Normal rv X N (; 2 ) will be used to develop the mgf of a bivariate Normal rv X = (X; Y ): MX (t) = E (exp(tX )) = Z1 exp(tx) p 1 ;1 Z1 1 p = ;1 22 Z1 1 p = ;1 22 Z1 1 p = ;1 22 = = = = 22 1 2 exp ; 22 (x ; ) dx exp ; 21 2 [;22 tx + (x ; )2 ] dx 1 2 2 2 exp ; 22 [;2 tx + (x ; 2x + )] dx exp ; 21 2 [x2 ; 2( + 2 t)x + ( + t2 )2 ; ( + t2 )2 + 2 ] dx 1 Z 1 1 1 2 2 2 2 2 p exp ; 22 [;( + t ) + ] exp ; 22 [x ; ( + t )] dx {z } | ;1 22 2 2 pdf of N ( + t ; ), that integrates to 1 1 2 2 2 4 2 exp ; 22 [; ; 2t ; t + ] 2 ; t2 4 ! ; 2 t exp ;22 exp t + 12 2 t2 Bivariate Normal mgf: MX;Y (t1 ; t2 ) = Z1Z1 ;1 ;1 exp(t1 x + t2 y) fX;Y (x; y) dx dy 117 = = (A) = (B) = = = = (C ) = = = Z1Z1 exp(t1 x) exp(t2 y) fX (x) fY jX (y j x) dy dx ;1 ;1 Z 1 Z 1 exp(t2 y) fY jX (y j x) dy exp(t1 x) fX (x) dx ;1 ;1 Z 1 Z 1 exp(t2 y) 2! ! ; ( y ; ) X p p exp 22 (1 ; 2) dy exp(t1 x) fX (x) dx ;1 ;1 2 1 ; 2 2 2 j X = 2 + 2 (x ; 1) 1 Z1 exp X t2 + 21 22 (1 ; 2 )t22 exp(t1 x) fX (x) dx ;1 Z1 exp [2 + 2 (x ; 1 )]t2 + 12 22 (1 ; 2 )t22 + t1 x fX (x) dx ;1 1 Z1 1 2 2 2 2 2 exp 2 t2 + t2 x ; 1 t2 + 2 2 (1 ; )t2 + t1 x fX (x) dx ;1 1 1 1 Z 1 exp 2 22 (1 ; 2 )t22 + t2 2 ; 2 1 t2 exp (t1 + 2 t2 )x fX (x) dx ;1 1 1 1 1 2 2 2 2 2 2 2 2 exp 2 2 (1 ; )t2 + t2 2 ; 1 t2 exp 1 (t1 + t2 ) + 2 1 (t1 + t2 ) 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 exp 2 2 t2 ; 2 2 t2 + 2 t2 ; 1 t2 + 1 t1 + 1 t2 + 2 1 t1 + 1 2 t1 t2 + 2 2 t2 1 1 ! 22 22 exp 1 t1 + 2 t2 + 1 t1 + 2 t2 2+ 21 2 t1 t2 (A) follows from Theorem 5.1.6 since Y j X N (X ; 22 (1 ; 2 )). (B ) follows when we apply our calculations of the mgf of a N (; 2 ) distribution to a N (X ; 22 (1 ; 2 )) distribution. (C ) holds since the integral represents MX (t1 + 21 t2 ). Corollary 5.1.9: Let (X; Y ) be a bivariate Normal rv. X and Y are independent i = 0. Denition 5.1.10: Let Z be a k{rv of k iid N (0; 1) rv's. Let A 2 IRkk be a k k matrix, and let 2 IRk be a k{dimensional vector. Then X = AZ + has a multivariate Normal distribution with mean vector and variance{covariance matrix = AA0 . 118 Note: (i) If is non{singular, X has the joint pdf fX (x) = (2)k=2 (1j j)1=2 exp ; 12 (x ; )0 ;1(x ; ) : (ii) If is singular, then X ; takes values in a linear subspace of IRk with probability 1. (iii) If is non{singular, then X has mgf MX (t) = exp(0 t + 21 t0t): Theorem 5.1.11: The components X1 ; : : : ; Xk of a normally distributed k{rv X are independent i Cov(Xi ; Xj ) = 0 8i; j = 1; : : : ; k; i 6= j . Theorem 5.1.12: Let X = (X1 ; : : : ; Xk )0 . X has a k{dimensional Normal distribution i every linear function of X , i.e., X 0 t = t1 X1 + t2 X2 + : : : + tk Xk , has a univariate Normal distribution. Proof: The Note following Denition 5.1.1 states that any linear function of two Normal rv's has a univariate Normal distribution. By induction on k, we can show that every linear function of X , i.e., X 0 t, has a univariate Normal distribution. Conversely, if X 0 t has a univariate Normal distribution, we know from Theorem 5.1.8 that 1 0 0 2 MX 0 t (s) = exp E (X t) s + 2 V ar(X t) s = exp 0 ts + 12 t0 ts2 1 0 0 =) MX 0 t (1) = exp t + 2 t t = MX (t) By uniqueness of the mgf and Note (iii) that follows Denition 5.1.10, X has a multivariate Normal distribution. 119 5.2 Exponential Familty of Distributions Denition 5.2.1: Let # be an interval on the real line. Let ff (; ) : 2 #g be a family of pdf's (or pmf's). We assume that the set fx : f (x; ) > 0g is independent of , where x = (x1 ; : : : ; xn ). We say that the family ff (; ) : 2 #g is a one{parameter exponential family if there exist real{valued functions Q() and D() on # and Borel{measurable functions T (X ) and S (X ) on IRn such that f (x; ) = exp(Q()T (x) + D() + S (x)): Note: We can also write f (x; ) as f (x; ) = h(x)c() exp(T (x)) where h(x) = exp(S (x)), = Q(), and c() = exp(D(Q;1 ())), and call this the exponential family in canonical form for a natural parameter . Denition 5.2.2: Let # IRk be a k{dimensional interval. Let ff (; ) : 2 #g be a family of pdf's (or pmf's). We assume that the set fx : f (x; ) > 0g is independent of , where x = (x1 ; : : : ; xn ). We say that the family ff (; ) : 2 #g is a k{parameter exponential family if there exist real{valued functions Q1 (); : : : Qk () and D() on # and Borel{measurable functions T1 (X ); : : : ; Tk (X ) and S (X ) on IRn such that ! k X Qi ()Ti (x) + D() + S (x) : f (x; ) = exp i=1 Note: Similar to the Note following Denition 5.2.1, we can express the k{parameter exponential family in canonical form for a natural k 1 parameter vector = (1 ; : : : ; k )0 . Example 5.2.3: Let X N (; 2 ) with both parameters and 2 unknown. We have: ! 1 2 1 2 f (x; ) = p 2 exp ; 22 (x ; ) = exp ; 21 2 x2 + 2 x ; 22 ; 12 ln(22 ) 2 = (; 2 ) # = f(; 2 ) : 2 IR; 2 > 0g 120 Therefore, Q1 () T1 (x) Q2 () T2 (x) D() S (x) = = = = ; 21 2 x2 2 x 2 = ; 22 ; 12 ln(22 ) = 0 Thus, this is a 2{parameter exponential family. 121 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 36: Monday 11/22/1999 Scribe: Jurgen Symanzik, Hanadi B. Eltahir 6 Limit Theorems Motivation: I found this slide from my Stat 250, Section 003, \Introductory Statistics" class (an undergraduate class I taught at George Mason University in Spring 1999): What does this mean at a more theoretical level??? 122 6.1 Modes of Convergence Denition 6.1.1: Let X1 ; : : : ; Xn be iid rv's with common cdf FX (x). Let T = T (X ) be any statistic, i.e., a Borel{ measurable function of X that does not involve the population parameter(s) #, dened on the support X of X . The induced probability distribution of T (X ) is called the sampling distribution of T (X ). Note: (i) Commonly used statistics are: n X Sample Mean: X n = n1 Xi i=1 X Sample Variance: Sn2 = n;1 1 (Xi ; X n )2 n i=1 Sample Median, Order Statistics, Min, Max, etc. (ii) Recall that if X1 ; : : : ; Xn are iid and if E (X ) and V ar(X ) exist, then E (X n ) = = E (X ), E (Sn2 ) = 2 = V ar(X ), and V ar(X n) = n2 . (iii) Recall that if X1 ; : : : ; Xn are iid and if X has mgf MX (t) or characteristic function X (t) then MX n (t) = (MX ( nt ))n or X n (t) = (X ( nt ))n . Note: Let fXn g1 n=1 be a sequence of rv's on some probability space ( ; L; P ). Is there any meaning behind the expression nlim !1 Xn = X ? Not immediately under the usual denitions of limits. We rst need to dene modes of convergence for rv's and probabilities. Denition 6.1.2: 1 Let fXn g1 n=1 be a sequence of rv's with cdf's fFn gn=1 and let X be a rv with cdf F . If Fn (x) ! F (x) d X ) or at all continuity points of F , we say that Xn converges in distribution to X (Xn ;! L X ), or F converges weakly to F (F ;! w F ). Xn converges in law to X (Xn ;! n n Example 6.1.3: Let Xn N (0; n1 ). Then Z x exp ; 21 nt2 q 2 dt Fn (x) = ;1 123 n Z pnx exp(; 1 s2) p 2 ds = 2 ;1 p = ( nx) ( 8 > < (1) =11; if x > 0 =) Fn (x) ! > (0) = 2 ; if x = 0 : (;1) = 0; if x < 0 1; x 0 the only point of discontinuity is at x = 0. Everywhere else, 0; x < 0 p ( nx) = Fn (x) ! FX (x). d X , where P (X = 0) = 1, or X ;! d 0 since the limiting rv here is degenerate, i.e., it So, Xn ;! n has a Dirac(0) distribution. If FX (x) = Example 6.1.4: In this example, the sequence fFn g1 n=1 converges pointwise to something that is not a cdf: Let Xn Dirac(n), i.e., P (Xn = n) = 1. Then, ( Fn(x) = 0; x < n 1; x n d X. It is Fn (x) ! 0 8x which is not a cdf. Thus, there is no rv X such that Xn ;! Example 6.1.5: 1 1 Let fXn g1 n=1 be a sequence of rv's such that P (Xn = 0) = 1 ; n and P (Xn = n) = n and let X Dirac(0), i.e., P (X = 0) = 1. It is 8 > < 0; 1 x < 0 Fn (x) = 1 ; n ; 0 x < n > : 1; xn FX (x) = w F but It holds that Fn ;! X ( 0; x < 0 1; x 0 E (Xnk ) = nk;1 6! E (X k ) = 0: Thus, convergence in distribution does not imply convergence of moments/means. 124 Note: Convergence in distribution does not say that the Xi 's are close to each other or to X . It only means that their cdf's are (eventually) close to some cdf F . The Xi 's do not even have to be dened on the same probability space. Example 6.1.6: d Let X and fXn g1 n=1 be iid N (0; 1). Obviously, Xn ;! X but nlim !1 Xn 6= X . Theorem 6.1.7: 1 Let X and fXn g1 n1 =1 be discrete rv's with support X and fXn gn=1 , respectively. Dene the count[ able set A = X [ Xn = fak : k = 1; 2; 3; : : : g. Let pk = P (X = ak ) and pnk = P (Xn = ak ). Then n=1 d X. it holds that pnk ! pk 8k i Xn ;! Theorem 6.1.8: 1 Let X and fXn g1 n=1 be continuous rv's with pdf's f and ffn gn=1 , respectively. If fn(x) ! f (x) for d almost all x as n ! 1 then Xn ;! X . Theorem 6.1.9: d Let X and fXn g1 n=1 be rv's such that Xn ;! X . Let c 2 IR be a constant. Then it holds: d X + c. (i) Xn + c ;! d cX . (ii) cXn ;! d aX + b. (iii) If an ! a and bn ! b, then an Xn + bn ;! Proof: Part (iii): Suppose that a > 0; an > 0. Let Yn = an Xn + bn and Y = aX + b. It is Likewise, FY (y) = P (Y < y) = P (aX + b < y) = P (X < y ;a b ) = FX ( y ;a b ): FY (y) = FX ( y ;a bn ): n n n If y is a continuity point of FY , y;a b is a continuity point of FX . Since an ! a; bn ! b and FXn (x) ! FX (x), it follows that FYn (y) ! FY (y) for every continuity point y of FY . Thus, d aX + b. an Xn + bn ;! 125 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 37: Wednesday 11/24/1999 Scribe: Jurgen Symanzik, Rich Madsen Denition 6.1.10: Let fXn g1 n=1 be a sequence of rv's dened on a probability space ( ; L; P ). We say that Xn p converges in probability to a rv X (Xn ;! X , P- nlim !1 Xn = X ) if nlim !1 P (j Xn ; X j> ) = 0 8 > 0: Note: The following are equivalent: nlim !1 P (j Xn ; X j> ) = 0 , nlim !1 P (j Xn ; X j ) = 1 , nlim !1 P (f! :j Xn (!) ; X (!) j> )) = 0: If X is degenerate, i.e., P (X = c) = 1, we say that Xn is consistent for c. For example, let Xn such that P (Xn = 0) = 1 ; n1 and P (Xn = 1) = n1 . Then P (j Xn j> ) = ( 1 n; 0<<1 0; 1 p Therefore, nlim !1 P (j Xn j> ) = 0 8 > 0. So Xn ;! 0, i.e., Xn is consistent for 0. Theorem 6.1.11: p p (i) Xn ;! X () Xn ; X ;! 0. p p (ii) Xn ;! X; Xn ;! Y =) P (X = Y ) = 1. p p p (iii) Xn ;! X; Xm ;! X =) Xn ; Xm ;! 0 as n; m ! 1. p p p (iv) Xn ;! X; Yn ;! Y =) Xn Yn ;! X Y. p p (v) Xn ;! X; k 2 IR a constant =) kXn ;! kX . p p r (vi) Xn ;! k; k 2 IR a constant =) Xnr ;! k 8r 2 IN . p p p (vii) Xn ;! a; Yn ;! b; a; b 2 IR =) Xn Yn ;! ab. p p (viii) Xn ;! 1 =) Xn;1 ;! 1. 126 p a p p (ix) Xn ;! a; Yn ;! b; a 2 IR; b 2 IR ; f0g =) XYnn ;! b. p p (x) Xn ;! X; Y an arbitrary rv =) XnY ;! XY . p p p (xi) Xn ;! X; Yn ;! Y =) Xn Yn ;! XY . Proof: See Rohatgi, page 244{245 for partial proofs. Theorem 6.1.12: p p Let Xn ;! X and let g be a continuous function on IR. Then g(Xn ) ;! g(X ). Proof: Preconditions: 1.) X rv =) 8 > 0 9k = k() : P (jX j > k) < 2 2.) g is continuous on IR =) g is also uniformly continuous (see Denition of u.c. in Theorem 3.3.3 (iii)) on [;k; k] =) 9 = (; k) : jX j k; jXn ; X j < ) jg(Xn ) ; g(X )j < Let A = fjX j kg = f! : jX (!)j kg B = fjXn ; X j < g = f! : jXn (!) ; X (!)j < g C = fjg(Xn ) ; g(X )j < g = f! : jg(Xn (!)) ; g(X (!))j < g If ! 2 A \ B :) =2) !2C =) A \ B C =) C C (A \ B )C = AC [ B C =) P (C C ) P (AC [ B C ) P (AC ) + P (B C ) Now: P (jg(Xn ) ; g(X )j ) P| (jX{zj > k}) + 2 by 1.) p 2 for nn0 (;;k) since Xn ;! X for n n0(; ; k) 127 P| (jXn ;{zX j }) Corollary 6.1.13: p p Let Xn ;! c; c 2 IR and let g be a continuous function on IR. Then g(Xn ) ;! g(c). Theorem 6.1.14: p d X. Xn ;! X =) Xn ;! Proof: p Xn ;! X , P (jXn ; X j > ) ! 0 as n ! 1 8 > 0 It holds: P (X x ; ) = P (X x ; ; jXn ; X j ) + P (X x ; ; jXn ; X j > ) (A) P (Xn x) + P (jXn ; X j > ) (A) holds since X x ; and Xn within of X , thus Xn x. Similarly, it holds: P (Xn x) P (X x + ) + P (jXn ; X j > ) = P (Xn x; j Xn ; X j ) + P (Xn x; j Xn ; X j> ) Combining the 2 inequalities from above gives: P (X x ; ) ; P| (jXn ;{zX j > }) P| (Xn{z x}) P (X x + ) + P| (jXn ;{zX j > }) !0 as n!1 Therefore, =Fn (x) !0 as n!1 P (X x ; ) Fn(x) P (X x + ) as n ! 1: Since the cdf's Fn () are not necessarily left continuous, we get the following result for # 0: P (X < x) Fn (x) P (X x) = FX (x) Let x be a continuity point of F . Then it holds: F (x) = P (X < x) Fn (x) F (x) =) Fn (x) ! F (x) d X =) Xn ;! 128 Theorem 6.1.15: Let c 2 IR be a constant. Then it holds: p d c () X ;! Xn ;! c: n Example 6.1.16: In this example, we will see that p d X 6=) X ;! Xn ;! X n for some rv X . Let Xn be identically distributed rv's and let (Xn ; X ) have the following joint distribution: Xn 0 1 X 0 1 0 12 12 1 0 1 2 2 1 1 1 2 2 d X since all have exactly the same cdf, but for any 2 (0; 1), it is Obviously, Xn ;! P (j Xn ; X j> ) = P (j Xn ; X j= 1) = 1 8n; so nlim 6 p X. !1 P (j Xn ; X j> ) 6= 0. Therefore, Xn ;! Theorem 6.1.17: 1 Let fXn g1 n=1 and fYn gn=1 be sequences of rv's and X be a rv dened on a probability space ( ; L; P ). Then it holds: p d X; j X ; Y j;! d X: Yn ;! 0 =) Xn ;! n n Proof: Similar to the proof of Theorem 6.1.14. See also Rohatgi, page 253, Theorem 14. 129 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 38: Monday 11/29/1999 Scribe: Jurgen Symanzik, Bill Morphet Theorem 6.1.18: Slutsky's Theorem 1 Let (Xn )1 n=1 and (Yn )n=1 be sequences of rv's and X be a rv dened on a probability space ( ; L; P ). Let c 2 IR be a constant. Then it holds: p d X; Y ;! d X + c. (i) Xn ;! c =) Xn + Yn ;! n p d X; Y ;! d cX . (ii) Xn ;! c =) Xn Yn ;! n p If c = 0, then also Xn Yn ;! 0. p d X; Y ;! d X if c 6= 0. (iii) Xn ;! c =) XYnn ;! n c Proof: p Th:6:1:11(i) p (i) Yn ;! c () Yn ; c ;! 0 p =) Yn ; c = Yn + (Xn ; Xn ) ; c = (Xn + Yn ) ; (Xn + c) ;! 0 (A) 6:1:9(i) d X Th:= d X + c (B ) Xn ;! ) Xn + c ;! Combining (A) and (B ), it follows from Theorem 6.1.17: d X +c Xn + Yn ;! (ii) Case c = 0: 8 > 0 8k > 0, it is P (j Xn Yn j> ) = P (j XnYn j> ; Yn k ) + P (j Xn Yn j> ; Yn > k ) P (j Xn k j> ) + P (Yn > k ) P (j Xn j> k) + P (j Yn j> k ) p d X and Y ;! Since Xn ;! 0, it follows n nlim !1P (j Xn Yn j> ) P (j Xn j> k) ! 0 as Therefore, p Xn Yn ;! 0: 130 k ! 1: Case c 6= 0: p p d X and Y ;! Since Xn ;! c, it follows Xn Yn ; cXn = Xn (Yn ; c) ;! 0. n p =) Xn Yn ;! cXn Th:=6) :1:14 X Y ;! d cX n n n d cX by Theorem 6.1.9 (ii), it follows from Theorem 6.1.17: Since cXn ;! d cX Xn Yn ;! p (iii) Let Zn ;! 1 and let Yn = cZn . c6=0 1 =) Yn = Z1n 1c Th:6:1:11(v;viii) 1 p 1 =) Yn ;! c With part (ii) above, it follows: p 1 d X and 1 ;! Xn ;! Yn c d X =) XYnn ;! c Denition 6.1.19: r Let (Xn )1 n=1 be a sequence of rv's such that E (j Xn j ) < 1 for some r > 0. We say that Xn r X ) if E (j X jr ) < 1 and converges in the rth mean to a rv X (Xn ;! nlim !1 E (j Xn ; X jr ) = 0: Example 6.1.20: 1 1 Let (Xn )1 n=1 be a sequence of rv's dened by P (Xn = 0) = 1 ; n and P (Xn = 1) = n . r 0 8r > 0. It is E (j Xn jr ) = n1 8r > 0. Therefore, Xn ;! Note: 1 X) The special cases r = 1 and r = 2 are called convergence in absolute mean for r = 1 (Xn ;! ms X or X ;! 2 X ). and convergence in mean square for r = 2 (Xn ;! n Theorem 6.1.21: p r X for some r > 0. Then X ;! Assume that Xn ;! X. n Proof: Using Markov's Inequality (Corollary 3.5.2), it holds for any > 0: E (j Xn ; X jr ) P (j X ; X j ) n r 131 r X =) lim E (j X ; X jr ) = 0 Xn ;! n n!1 E (j Xn ; X jr ) = 0 =) nlim P ( j X ; X j ) lim n !1 n!1 r p =) Xn ;! X Example 6.1.22: 1 1 Let (Xn )1 n=1 be a sequence of rv's dened by P (Xn = 0) = 1 ; nr and P (Xn = n) = nr for some r > 0. p For any > 0, P (j Xn j> ) ! 0 as n ! 1; so Xn ;! 0. s 0. But E (j X jr ) = 1 6! 0 as n ! 1; For 0 < s < r, E (j Xn js ) = nr1;s ! 0 as n ! 1; so Xn ;! n so Xn ;! 6 r 0. Theorem 6.1.23: r X , then it holds: If Xn ;! r r (i) nlim !1 E (j Xn j ) = E (j X j ); and s X for 0 < s < r. (ii) Xn ;! Proof: (i) For 0 < r 1, it holds: E (j Xn jr ) = E (j Xn ; X + X jr ) E (j Xn ; X jr + j X jr ) =) E (j Xn jr ) ; E (j X jr ) E (j Xn ; X jr ) r r r =) nlim !1 E (j Xn j ) ; nlim !1 E (j X j ) nlim !1 E (j Xn ; X j ) = 0 r r =) nlim !1 E (j Xn j ) E (j X j ) (A) Similarly, E (j X jr ) = E (j X ; Xn + Xn jr ) E (j Xn ; X jr + j Xn jr ) =) E (j X jr ) ; E (j Xn jr ) E (j Xn ; X jr ) r r r =) nlim !1 E (j X j ) ; nlim !1 E (j Xn j ) nlim !1 E (j Xn ; X j ) = 0 r =) E (j X jr ) nlim !1 E (j Xn j ) (B ) Combining (A) and (B ) gives r r nlim !1 E (j Xn j ) = E (j X j ) 132 For r > 1, it follows from Minkowski's Inequality (Theorem 4.8.3): [E (j X ; Xn + Xn jr )] r1 [E (j X ; Xn jr )] 1r + [E (j Xn jr )] 1r =) [E (j X jr )] r1 ; [E (j Xn jr )] 1r [E (j X ; Xn jr )] 1r r r 1r r 1r =) [E (j X jr )] 1r ; nlim !1[E (j Xn ; X j )] = 0 since Xn ;! X !1[E (j Xn j )] nlim 1 r r =) [E (j X jr )] 1r nlim !1[E (j Xn j )] (C ) Similarly, [E (j Xn ; X + X jr )] 1r [E (j Xn ; X jr )] 1r + [E (j X jr )] 1r r r 1r r 1r r 1r =) nlim !1[E (j Xn j )] ; nlim !1[E (j X j )] nlim !1[E (j Xn ; X j )] = 0 since Xn ;! X 1 1 r r r r =) nlim !1[E (j Xn j )] [E (j X j )] (D) Combining (C ) and (D) gives r 1r r 1r nlim !1[E (j Xn j )] = [E (j X j )] r r =) nlim !1 E (j Xn j ) = E (j X j ) (ii) For 1 s < r, it follows from Lyapunov's Inequality (Theorem 3.5.4): [E (j Xn ; X js )] 1s [E (j Xn ; X jr )] 1r =) E (j Xn ; X js ) [E (j Xn ; X jr )] rs r s r sr =) nlim !1 E (j Xn ; X j ) nlim !1[E (j Xn ; X j )] = 0 since Xn ;! X s X =) Xn ;! An additional proof is required for 0 < s < r 1. 133 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lectures 39 & 40: Wednesday 12/1/1999 & Friday 12/3/1999 Scribe: Jurgen Symanzik, Hanadi B. Eltahir Denition 6.1.24: Let fXn g1 n=1 be a sequence of rv's on ( ; L; P ). We say that Xn converges almost surely to 1 a:s: X ) or X converges with probability 1 to X (X w:p: a rv X (Xn ;! n n ;! X ) or Xn converges strongly to X i P (f! : Xn (!) ! X (!) as n ! 1g) = 1: Note: An interesting characterization of convergence with probability 1 and convergence in probability can be found in Parzen (1960) \Modern Probability Theory and Its Applications" on page 416 (see Handout). Example 6.1.25: Let = [0; 1] and P a uniform distribution on . Let Xn (!) = ! + !n and X (!) = !. For ! 2 [0; 1), !n ! 0 as n ! 1. So Xn (!) ! X (!) 8! 2 [0; 1). However, for ! = 1, Xn (1) = 2 6= 1 = X (1) 8n, i.e., convergence fails at ! = 1. a:s: X . Anyway, since P (f! : Xn (!) ! X (!) as n ! 1g) = P (f! 2 [0; 1)g) = 1, it is Xn ;! Theorem 6.1.26: p a:s: X =) X ;! Xn ;! X. n Proof: Choose > 0 and > 0. Find n0 = n0 (; ) such that P Since 1 \ n=n0 1 \ n=n0 ! fj Xn ; X j g 1 ; : fj Xn ; X j g fj Xn ; X j g 8n n0, it is P (fj Xn ; X j g) P 1 \ n=n0 ! fj Xn ; X j g 1 ; 8n n0: p Therefore, P (fj Xn ; X j g) ! 1 as n ! 1. Thus, Xn ;! X. 134 Example 6.1.27: p a:s: X : Xn ;! X 6=) Xn ;! Let = (0; 1] and P a uniform distribution on . Dene An by A1 = (0; 12 ]; A2 = ( 21 ; 1] A3 = (0; 14 ]; A4 = ( 41 ; 12 ]; A5 = ( 21 ; 34 ]; A6 = ( 43 ; 1] A7 = (0; 18 ]; A8 = ( 81 ; 14 ]; : : : Let Xn (!) = IAn (!). p It is P (j Xn ; 0 j ) ! 0 8 > 0 since Xn is 0 except on An and P (An ) # 0. Thus Xn ;! 0. But P (f! : Xn (!) ! 0g) = 0 (and not 1) because any ! keeps being in some An beyond any n0 , i.e., Xn (!) looks like 0 : : : 010 : : : 010 : : : 010 : : :, so Xn ;! 6 a:s: 0. Example 6.1.28: r X 6=) X ;! a:s: X : Xn ;! n Let Xn be independent rv's such that P (Xn = 0) = 1 ; n1 and P (Xn = 1) = n1 . r 0 8r > 0. It is E (j Xn ; 0 jr ) = E (j Xn jr ) = E (j Xn j) = n1 ! 0 as n ! 1, so Xn ;! But P (Xn = 0 8m n n0 ) = n0 Y + 1 ) : : : ( n0 ; 2 )( n0 ; 1 ) = m ; 1 (1 ; n1 ) = ( mm; 1 )( mm+ 1 )( m m+2 n ;1 n n 0 n=m 0 0 As n0 ! 1, it is P (Xn = 0 8m n n0 ) ! 0 8m, so Xn ;! 6 a:s: 0. Example 6.1.29: a:s: X 6=) X ;! r X: Xn ;! n Let = [0; 1] and P a uniform distribution on . Let An = [0; ln1n ]. Let Xn (!) = nIAn (!) and X (!) = 0. a:s: 0. It holds that 8! > 0 9n0 : ln1n0 < ! =) Xn (!) = 0 8n > n0 and P (! = 0) = 0. Thus, Xn ;! 6 r X. But E (j Xn ; 0 jr ) = lnnrn ! 1 8r > 0, so Xn ;! 135 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 41: Monday 12/6/1999 Scribe: Jurgen Symanzik 6.2 Weak Laws of Large Numbers Theorem 6.2.1: WLLN: Version I 2 Let fXi g1 in=1 be a sequence of iid rv's with mean E (Xi ) = and variance V ar(Xi ) = < 1. Let X X n = n1 Xi. Then it holds i=1 nlim !1 P (j X n ; j ) = 0; p i.e., X n ;! . Proof: By Markov's Inequality, it holds for all > 0: 2 P (j X n ; j ) E((Xn2;) ) = V ar(2X n ) = n22 ;! 0 as n ! 1 Note: For iid rv's with nite variance, X n is consistent for . A more general way to derive a \WLLN" follows in the next Denition. Denition 6.2.2: n X Let fXi g1 Xi. We say that fXi g obeys the WLLN with i=1 be a sequence of rv's. Let Tn = i=1 respect to a sequence of norming constants fBi g1 i=1 , Bi > 0; Bi " 1, if there exists a sequence 1 of centering constants fAi gi=1 such that p Bn;1(Tn ; An) ;! 0: Theorem 6.2.3: Let fXi g1 rv's with E (Xi ) = i and V ar(Xi ) = i2 , i 2 IN . i=1 be a sequence of pairwise uncorrelated n n n X 2 X X If i ! 1 as n ! 1, we can choose An = i and Bn = i2 and get i=1 n X i=1 i=1 (Xi ; i ) n X i=1 i2 136 i=1 p ;! 0: Proof: By Markov's Inequality, it holds for all > 0: P (j n X i=1 Xi ; n X i=1 i j> n X i=1 n X i2 ) E (( (Xi ; ))2 ) i=1 n X 2 ( i2)2 i=1 1 ;! 0 as n ! 1 = X n 2 i2 i=1 Note: To obtain Theorem 6.2.1, we choose An = n and Bn = n2 . Theorem 6.2.4: n X 1 Let fXi g1 be a sequence of rv's. Let X = Xi . A necessary and sucient condition for n i=1 n i=1 fXi g to obey the WLLN with respect to Bn = n is that ! as n ! 1. Proof: Rohatgi, page 258, Theorem 2. 2 E Xn 2 ! 0 1 + Xn Example 6.2.5: Let (X1 ; : : : ; Xn ) be jointly Normal with E (Xi ) = 0, E (Xi2 ) = 1 for all i, and Cov(Xi ; Xj ) = if j i ; j j= 1 and Cov(Xi; Xj ) = 0 if j i ; j j> 1. Then, Tn N (0; n + 2(n ; 1)) = N (0; 2 ). It is 2 X n E 1 + X 2n ! = = = = !0 p 0 =) X n ;! Tn2 E n2 + T 2 n ! Z 1 x2 ; 2 x dx 2 2 dx j y = ; dy = e 2 2 2 0 n + x Z 1 2 y2 ; 22 dy p2 e 2 0 n2 + 2 y2 Z 1 (n + 2(n ; 1))y2 ; 22 dy p2 e 2 2 2 0 n + (n + 2(n ; 1))y Z1 2 n + 2(n ; 1) 2 e; 22 dy p y n2 | 0 2{z } =1; since Var of N (0;1) distribution as n ! 1 p2 x y y y 137 Note: We would like to have a WLLN that just depends on means but does not depend on the existence of nite variances. To approach this, we consider the following: Let fXi g1 i=1 be a sequence of rv's. Let Tn = Xic = Let Tnc = n X i=1 Xic and mn = n X i=1 ( n X i=1 Xi . We truncate each Xi at c > 0 and get Xi; j Xi j c 0; otherwise E (Xic ). Lemma 6.2.6: For Tn , Tnc and mn as dened in the Note above, it holds: P (j Tn ; mn j> ) P (j Tnc ; mn j> ) + n X i=1 P (j Xi j> c) 8 > 0 Proof: It holds for all > 0: P (j Tn ; mn j> ) = P (j Tn ; mn j> and j Xi j c 8i 2 f1; : : : ; ng) + P (j Tn ; mn j> and j Xi j> c for at least one i 2 f1; : : : ; ng () P (j Tnc ; mn j> ) + P (j Xi j> c for at least one i 2 f1; : : : ; ng) n X P (j Tnc ; mn j> ) + P (j Xi j> c) i=1 () holds since Tnc = Tn when j Xi j c 8i 2 f1; : : : ; ng. Note: If the Xi 's are identically distributed, then P (j Tn ; mn j> ) P (j Tnc ; mn j> ) + nP (j X1 j> c) 8 > 0: If the Xi 's are iid, then c2 P (j Tn ; mn j> ) nE ((X2 1 ) ) + nP (j X1 j> c) 8 > 0 (): Note that P (j Xi j> c) = P (j X1 j> c) 8i 2 IN if the Xi 's are identically distributed and that E ((Xic )2 ) = E ((X1c )2 ) 8i 2 IN if the Xi 's are iid. 138 Theorem 6.2.7: Khintchine's WLLN Let fXi g1 i=1 be a sequence of iid rv's with nite mean E (Xi ) = . Then it holds: p X = 1 T ;! n n n Proof: If we take c = n and replace by n in () in the Note above, we get X1n )2 ) + nP (j X j> n): P (j Tn ; mn j> n) E ((n 1 2 Since E (j X1 j) < 1, it is nPZ(j1X1 j> n) ! 0 as n ! 1 by Theorem 3.1.9. From Corollary 3.1.12 we know that E (j X j ) = x;1 P (j X j> x)dx. Therefore, 0 E ((X1n )2 ) = 2 = 2 (+) Zn Z0A xP (j X1n j> x)dx Zn n xP (j X1 j> x)dx + 2 xP (j X1n j> x)dx 0 A Zn K + dx A K + n In (+), A is chosen suciently large such that xP (j X1n j> x) < 2 8x A for an arbitrary constant > 0 and K > 0 a constant. Therefore, E ((X1n )2 ) k + n2 n2 2 Since is arbitrary, we can make the right hand side of this last inequality arbitrarily small for suciently large n. Since E (Xi ) = 8i, it is mnn ! as n ! 1. Note: Theorem 6.2.7 meets the previously stated goal of not having a nite variance requirement. 139 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 42: Wednesday 12/8/1999 Scribe: Jurgen Symanzik 6.3 Strong Laws of Large Numbers Denition 6.3.1: n X Let fXi g1 be a sequence of rv's. Let T = Xi . We say that fXi g obeys the SLLN with n i=1 i=1 1 respect to a sequence of norming constants fBi gi=1 , Bi > 0; Bi " 1, if there exists a sequence of centering constants fAi g1 i=1 such that a:s: 0: Bn;1(Tn ; An) ;! Note: Unless otherwise specied, we will only use the case that Bn = n in this section. Theorem 6.3.2: a:s: X () lim P (sup j X ; X j> ) = 0 8 > 0. Xn ;! n!1 mn m Proof: (see also Rohatgi, page 249, Theorem 11) a:s: X implies X ; X ;! a:s: 0. Thus, we have to prove: WLOG, we can assume that X = 0 since Xn ;! n a:s: 0 () lim P (sup j X j> ) = 0 8 > 0 Xn ;! n!1 mn m \=)": Choose > 0 and dene An () = f sup j Xm j> g mn C = fnlim !1 Xn = 0g We know that P (C ) = 1 and therefore P (C c) = 0. 1 \ Let Bn () = C \ An (). Note that Bn+1 () Bn () and for the limit set Bn () = . It follows n=1 that 1 \ lim P ( B ( )) = P ( Bn ()) = 0: n n!1 n=1 140 We also have P (Bn()) = P (An \ C ) = 1 ; P (C c [ Acn ) = 1 ; P| ({zC c}) ;P (Acn ) + P| (C c{z\ ACn }) =0 =0 = P (An ) =) nlim !1 P (An ()) = 0 \(=": Assume that nlim !1 P (An ()) = 0 8 > 0 and dene D() = fnlim !1 j Xn j> g. Since D() An() 8n 2 IN , it follows that P (D()) = 0 8 > 0. Also, C c = fnlim !1 Xn 6= 0g =) 1 ; P (C ) =) Xn a:s: 0 ;! 1 X k=1 1 [ 1 fnlim !1 j Xn j> k g: k=1 P (D( k1 )) = 0 Note: a:s: 0 implies that 8 > 0 8 > 0 9n 2 IN : P ( sup j X j> ) < . (i) Xn ;! 0 n (ii) Recall that for a given sequence of events fAn g1 n=1 , A = nlim !1An = nlim !1 1 [ k=n nn0 Ak = 1 [ 1 \ n=1 k=n Ak is the event that innitely many of the An occur. We write P (A) = P (An i:o:) where i:o: stands for \innitely often". (iii) Using the terminology dened in (ii) above, we can rewrite Theorem 6.3.2 as a:s: 0 () P (j X j> i:o:) = 0 8 > 0: Xn ;! n 141 Theorem 6.3.3: Borel{Cantelli Lemma (i) 1st BC{Lemma: 1 X 1 Let fAn gn=1 be a sequence of events such that P (An ) < 1. Then P (A) = 0. n=1 (ii) 2nd BC{Lemma: 1 X Let fAn g1 P (An ) = 1. Then P (A) = 1. n=1 be a sequence of independent events such that n=1 Proof: (i): 1 [ P (A) = P (nlim !1 k=n Ak ) = nlim !1 P ( 1 [ k=n nlim !1 1 X k=n = nlim !1 (ii): We have Ac = = 0 1 \ 1 [ n=1 k=n Ak ) P (Ak ) 1 X k=1 P (Ak ) ; nX ;1 k=1 ! P (Ak ) Ack . Therefore, P (Ac) = P (nlim !1 If we choose n0 > n, it holds that 1 \ k=n Therefore, P( 1 \ k=n Ack ) 1 \ k =n Ack ) = nlim !1 P ( Ack n0 \ k=n n0lim !1 P ( = n0lim !1 = n0 \ k=n n 0 Y =) P (A) = 1 142 Ack ) (1 ; P (Ak )) k=n n0lim !1 exp 0 k=n Ack ): Ack : indep: 1 \ ; n0 X k=n P (Ak ) ! Example 6.3.4: Independence is necessary for 2nd BC{Lemma: Let = (0; 1) and P a uniform distribution on . Let An = I(0; n1 ) (!). Therefore, 1 X n=1 P (An ) = 1 1 X n=1 n = 1: But for any ! 2 , An occurs only for 1; 2; : : : ; b !1 c, where b !1 c denotes the largest integer (\oor") that is !1 . Therefore, P (A) = P (An i:o:) = 0. Lemma 6.3.5: Kolmogorov's Inequality Let fXi g1 be a sequence of independent rv's with common mean 0 and variances i2 . Let Tn = i =1 n X Xi . Then it holds: i=1 n X i2 P (1max j T j ) i=12 kn k 8 > 0 Proof: See Rohatgi, page 268, Lemma 2. Lemma 6.3.6: Kronecker's Lemma 1 X If Xi converges to s < 1 and Bn " 1, then it holds: i=1 n 1 X Bn Bk Xk ! 0 i=1 Proof: See Rohatgi, page 269, Lemma 3. Theorem 6.3.7: Cauchy Criterion a:s: X () lim P (sup j X Xn ;! n+m ; Xn j ) = 1 8 > 0. n!1 m Proof: See Rohatgi, page 270, Theorem 5. 143 Theorem 6.3.8: 1 1 X X If V ar(Xn ) < 1, then (Xn ; E (Xn )) converges almost surely. n=1 n=1 Proof: See Rohatgi, page 272, Theorem 6. 144 Stat 6710 Mathematical Statistics I Fall Semester 1999 Lecture 43: Friday 12/10/1999 Scribe: Jurgen Symanzik Corollary 6.3.9: Let fXi g1 of independent rv's. Let fBi g1 i=1 be a sequence i=1 , Bi > 0; Bi " 1, a sequence of norming n 1 V ar(X ) X X i constants. Let Tn = Xi . If B 2 < 1 then it holds: i=1 i=1 i Tn ; E (Tn ) ;! a:s: 0 Bn Proof: This Corollary follows directly from Theorem 6.3.8 and Lemma 6.3.6. Lemma 6.3.10: Equivalence Lemma n n 0 g1 be sequences of rv's. Let Tn = X Xi and Tn0 = X Xi0 . Let fXi g1 and f X i=1 i i=1 i=1 1 X i=1 P (Xi 6= Xi0 ) < 1, then the series fXi g and fXi0 g are tail{equivalent and Tn and i=1 Tn0 are convergence{equivalent, i.e., for Bn " 1 the sequences B1 Tn and B1 Tn0 converge on the If the series n same event and to the same limit, except for a null set. Proof: See Rohatgi, page 266, Lemma 1. Lemma 6.3.11: Let X be a rv with E (j X j) < 1. Then it holds: 1 X n=1 P (j X j n) E (j X j) 1 + Proof: Continuous case only: Let X have a pdf f . Then it holds: E (j X j) = =) 1 X k=0 Z1 ;1 j x j f (x)dx = kP (k j X j k + 1) E (j X j) 1 X 1 Z X 1 X n=1 P (j X j n) k=0 kjxjk+1 j x j f (x)dx (k + 1)P (k j X j k + 1) k=0 145 n It is 1 X k=0 1 X k X kP (k j X j k + 1) = k=0 n=1 1 X 1 X = n=1 k=n 1 X = Similarly, 1 X k=0 (k + 1)P (k j X j k + 1) = = n=1 1 X n=1 1 X n=1 P (k j X j k + 1) P (k j X j k + 1) P (j X j n) P (j X j n) + 1 X k=0 P (k j X j k + 1) P (j X j n) + 1 Theorem 6.3.12: Kolmogorov's SLLN n X Let fXi g1 be a sequence of iid rv's. Let T = Xi . Then it holds: n i=1 i=1 Tn = X ;! a:s: < 1 () E (j X j) < 1 (and then = E (X )) n n Proof: \=)": a:s: < 1. It is Suppose that X n ;! Tn = =) n X Xi = i=1 Xn = n nX ;1 i=1 Xi + Xn = Tn;1 + Xn . Tn ; n ; 1 Tn;1 ;! a:s: 0 n | {zn } |n {z ; 1} |{z} a:s: ;! !1 a:s: ;! By 1st Borel{Cantelli Lemma, we must have Lemma =)6:3:11 E (j X 1 X n=1 P (j Xn j n) < 1. j) < 1 Th: 6:2:7 (WLLN ) p =) X n ;! E (X ) p a:s: , it holds that X ;! Since X n ;! . Therefore, it must hold that = E (X ). n 146 \(=": Let E (j X j) < 1. Dene truncated rv's: Xk0 = Tn0 = Then it holds: ( Xk ; if j Xk j k 0; otherwise n X k=1 Xk0 0 X 0n = Tnn 1 X k=1 P (Xk 6= Xk0 ) = iid = Lemma 6:3:11 1 X k=1 1 X k=1 P (j Xk j k) P (j X j k) < E (j X j) < 1 By Lemma 6.3.10, it follows that Tn and Tn0 are convergence{equivalent. Thus, it is sucient to a:s: E (X ). prove that X 0n ;! We now establish the conditions needed in Corollary 6.3.9. It is V ar(Xn0 ) E ((Xn0 )2 ) = = =) Zn x2 fX (x)dx ;n nX ;1 Z x2 fX (x)dx k j x j <k +1 k=0 nX ;1 (k + 1)2 P (k j X j< k + 1) k=0 1 n 1 1 2 X 0 ) X X (k + 1) P (k j X j< k + 1) V ar ( X n n2 n2 n=1 n=1 k=0 1 X n (k + 1)2 X 2 P (k j X n=1 k=1 n 1 1 X 2 P (0 j X j< 1) n=1 n 1 1! 1 X () X 2 (k + 1) P (k j X j< k + 1) = n2 + 2P (0 j X j< 1) (A) = j< k + 1) + n=k k=1 147 () holds since 1 1 X 2 1:65 < 2 and the rst two sums can be rearranged as follows: = 2 6 n=1 n n 1 1 1 1; 2; 3; : : : 2 1; 2 =) 2 2; 3; : : : 3 1; 2; 3 3 3; : : : It is n k k .. . .. . .. . .. . 1 1 X 1 + 1 + 1 + ::: = 2 k2 (k + 1)2 (k + 2)2 n=k n k12 + k(k 1+ 1) + (k + 1)(1 k + 2) + : : : 1 X 1 = k12 + n ( n ; 1) n=k+1 From Bronstein, page 30, # 7, we know that 1 = 1 1 2 + 2 1 3 + 3 1 4 + : : : + n(n1+ 1) + : : : 1 X 1 = 1 1 2 + 2 1 3 + 3 1 4 + : : : + (k ;11) k + n=k+1 n(n ; 1) =) 1 X 1 ; 1 ; 1 ;::: ; 1 = 1 ; 12 23 34 (k ; 1) k n=k+1 n(n ; 1) 1 = 12 ; 2 1 3 ; 3 1 4 ; : : : ; (k ;11) k = 13 ; 3 1 4 ; : : : ; (k ;11) k = 14 ; : : : ; (k ;11) k = ::: = 1 k =) 1 1 1 X 1 + X 1 2 2 k n=k+1 n(n ; 1) n=k n = k12 + k1 148 k2 Using this result in (A), we get 1 1 1 2 X 0 ) 2 X (k + 1) P (k j X j< k + 1) + 2P (0 j X j< 1) V ar ( X n n2 k n=1 k=1 = 2 1 X k=0 kP (k j X j< k + 1) + 4 +2 1 1 X k=1 k 1 X k=1 P (k j X j< k + 1) P (k j X j< k + 1) + 2P (0 j X j< 1) (B) 2E (j X j) + 4 + 2 + 2 < 1 To establish (B ), we use an inequality from the Proof of Lemma 6.3.11. Thus, the conditions needed in Corollary 6.3.9 are met. It follows that 1 T 0 ; 1 E (T 0 ) ;! a:s: 0 (C ) n n n n Since E (Xn0 ) ! E (X ) as n ! 1, it follows by Kronecker's Lemma (6.3.6) that n1 E (Tn0 ) ! E (X ). Thus, when we replace n1 E (Tn0 ) by E (X ) in (C ), we get: 1 T 0 ;! a:s: E (X ) Lemma 6:3:10 1 T ;! a:s: E (X ) = ) n n n n Merry Xmas and a Happy New Millennium! 149