MAT4300 ————————————————————————— 23/08-2010 Curriculum for MAT4300 is chapters 1-13 including 19 and parts of 17,18. There are two main areas of study: measures and integration. Measure We wish to measure “the size” of sets -lengths -areas -volumes -mass -electrical charges -probabilities Notationwise we have Ω: which is the set of all possible outcomes, and A is a specific set and a subset of Ω. We have P (A) which is the probability that A will happen. In the sense of a measure the probability is the size of the set. Integration We introduce a new way to integrate, i.e finding the area/volume under the curve. Consider the function z = f (x, y), ∞ X αi · m(Ai ) ≈ i=−∞ Z f dm This is the basic idea of the Lebesgue-integral which provides a stronger integration concept than Riemann. The Lebesgue integral is defined on any space with a measure. 1 Set operations A set is a collection of objects, usually from a predefined universe usually denoted by Ω. A and B are subsets of Ω as shown in (i) in the image below, and are written as A, B ⊆ Ω. For objects or elements we have x ∈ A which means x is included in A, as shown in (ii) and y 6∈ A. When a set A is a strict subset of B, A ⊂ B as shown in (iii), this means ∀x ∈ A ⇒ x ∈ B: a ll the elements in A are elements in B. If A ⊂ B and B ⊂ A, then A = B. This is a very useful trick. Further set operations are A ∪ B: the collection of all elements included in at least one of A and B. (i) A ∩ B: the collection of all elements included in both A and B. (ii) A\B = a ∈ A | a 6∈ B : the collections of elements in A that are not in B. (iii) Two important concepts for measure theory: for countable sequences (which is mainly what one focuses on in measure theory): [ A1 ∪ A2 ∪ . . . = An n∈N A1 ∩ A2 ∩ . . . = \ n∈N 2 An . We have distributive laws for sets. B ∪ A1 ∩ A2 ∩ . . . ∩ An ∩ . . . = (B ∪ A1 ) ∩ (B ∪ A2 ) ∩ . . . ∩ (B ∪ An ) ∩ . . . and B ∩ A1 ∪ A2 ∪ . . . ∪ An ∩ . . . = (B ∩ A1 ) ∪ (B ∩ A2 ) ∪ . . . ∪ (B ∩ An ) ∪ . . . Proof of the second law Left side: Assume x ∈ B ∩ (A1 ∪ . . .) ⇔ x ∈ B and x ∈ Ai for at least one i. Right side: Assume x ∈ (B ∩ A1 ) ∪ . . . ⇔ x ∈ B ∩ Ai for at least one i ⇔ x ∈ B and x ∈ Ai for at least one i. DeMorgan’s Laws c A1 ∪ A2 ∪ . . . = Ac1 ∩ Ac2 ∩ . . . c A1 ∩ A2 ∩ . . . = Ac1 ∪ Ac2 ∪ . . . Proved in exercise 2.3. Cartesian Products A × B = (a, b) : a ∈ A and b ∈ B A1 × A2 × . . . × An = (a1 , a2 , . . . , an ) : a1 ∈ A1 , . . . , ai ∈ Ai , . . . an ∈ An When the same set is multiplied n times, we have a shorthand notation: n A | ×A× {z. . . × A} = A . n times Functions: f : X 7→ Y In this course we can think of functions as a rule that takes each element x ∈ X and assigns a new element f (x) ∈ Y as shown in (i). 3 If A ⊂ X we define the image of A as f (A) = f (a) : a ∈ A as shown in (ii). If B ⊂ Y we define the inverse image by f −1 (B) = a ∈ X : f (a) ∈ B as shown in the image. Injective, Surjective and Bijective functions. For f : X 7→ Y , we say f is injective if x1 6= x2 ⇒ f (x1 ) 6= f (x2 ). Two different elements x1 , x2 will always give two different elements f (x1 ), f (x2 ). We say that f : X 7→ Y is surjective if f (X) = Y , that is if the Y is completely “filled up” by f (X). It is permissible for two points in X to map to the same point in Y . 4 A function is bijective if it is both injective and surjective. We get a one-toone correspondence between X and Y . Relation between functions and sets. Result (i) f −1 A1 ∩ A2 ∩ . . . = f −1 (A1 ) ∩ f −1 (A2 ) ∩ . . . (ii) f −1 A1 ∪ A2 ∪ . . . = f −1 (A1 ) ∪ f −1 (A2 ) ∪ . . . Inverse images commutes with intersections and unions. Proof of (i) x ∈ f −1 (A1 ∩ A2 ∩ . . .) ⇔ f (x) ∈ A1 ∩ A2 ∩ . . . ⇔ x ∈ f −1 (Ai ) ∀i ⇔ x ∈ f −1 (A1 ) ∩ f −1 (A2 ) ∩ . . . This only holds for inverse images. Direct images does not commute over intersections. Result (i) f (A1 ∪ A2 ∪ . . . = f (A1 ) ∪ f (A2 ) ∪ . . . (i) f (A1 ∩ A2 ∩ . . . ⊆ f (A1 ) ∩ f (A2 ) ∩ . . . Proof of (ii) Since A1 ∩ A2 ∩ . . . ⊆ Ai ∀i ⇔ f A1 ∩ A2 ∩ . . . ⊆ f (Ai )∀i ⇔ f A1 ∩ A2 ∩ . . . ⊆ f (A1 ) ∩ f (A2 ) ∩ . . . Since the second last statement tells us that we have an inclusion in all the sets, the third statement follows. Counterexample to equality for (ii) We define X = {x1 , x2 } and Y = {y}, where we assume x1 6= x2 and we consider the function f : X 7→ Y . The only possibility we have is f (x1 ) = y and f (x2 ) = y. If we set A1 = {x1 } and A2 = {x2 } we get A1 ∩ A2 = ∅. Thus f (A1 ∩ A2 ) = f (∅) = ∅. However, f (A1 ) ∩ f (A2 ) = {y} ∩ {y} = {y}. The function is not injective since x1 6= x2 are mapped to the same point y. If we assume f is injective we get an equality in (ii). 5 Countability A set A is countable if there exists a list a1 , a2 , a3 , . . . that contain all the elements in A. Example The natural numbers. As defined they make up a satisfactory list. N = {1, 2, 3, . . .} The integers. We wish to have a beginning to the list, so we rearrange them. Z = {. . . , −2, −1, 0, 1, 2, . . .} =⇒ {0, 1, −1, 2, −2, 3, −3, . . .} Result If A and B are countable, then the product A × B is countable. Proof We define A = {a1 , a2 , a3 , . . .} B = {b1 , b2 , b3 , . . .} Then the product A × B = {(ai , bi ) : ai ∈ A, bi ∈ B}. We can write a list based on the index numbers: first all elements such that the index numbers sum to two, then three etc. A × B = {(a1 , b1 ), (a2 , b1 ), (a1 , b2 ), (a3 , b1 ), (a2 , b2 ), (a1 , b3 ), . . .} {z } | {z } | {z } | sum: 2 sum: 3 sum: 4 Corollary The set of rational numbers Q is a countable set. Proof Both Z and N are countable, so by our last result Z × N is countable. Z × N = {(a1 , b1 ), (a2 , b1 ), (a1 , b2 ), . . .} so the list of the rational numbers is as follows (all fractions are repeated infinite times, but that doesn’t matter) a1 a2 a1 Q= , , ,... b1 b1 b2 Non-countable sets can be R or even the interval (0, 1). Result The interval (0,1) is not countable. Proof 6 Assume for contradiction that (0,1) is countable so that we can make en exhaustive list. (0, 1) := {x1 , x2 , x3 , . . .}. All the xi are decimals, so we can write out the decimal expanions. x1 = x2 = x3 = .. . 0.d11 d12 d13 d14 . . . 0.d21 d22 d23 d24 . . . 0.d31 d32 d33 d34 . . . .. . xn = 0.dn1 dn2 dn3 dn4 . . . .. .. . . Now we find a number not included on this list. We choose the decimal number t. 1 dii = 6 1 t = 0.e1 e2 e3 e4 . . . where ei = 2 dii = 1 We stay away from 0’s and 9’s since they can produce the same number (like how 0.999 . . . = 1). We are left with t ∈ (0, 1) which is not included in the list. Contradiction! 7 ————————————————————————— 30/08-2010 Definition If A is a σ-algebra on a set X and µ : A 7→ [0, ∞] is a function we call µ a measure if (i) µ(∅) = 0 (ii) Whenever {Ai }i∈N is a disjoint sequence of elements in A then [ X µ Ai = µ(Ai ) i∈N i∈N Theorem - Continuity of measures Assume that µ is a measure on (X, A), then (i) For any increasing (i.e Ai ⊂ Ai+1 ) sequence {Ai }, then [ Ai = lim µ(Ai ) µ i→∞ i∈N (ii) For any decreasing sequence {Ai } of sets in A such that µ(A1 ) < ∞, we have \ Ai = lim µ(Ai ) µ i→∞ i∈N Proof (i) We set B1 = A1 , B2 = A2 \A1 , B3 = A3 \A2 , etc. We then have a disjoint sequence {Bi } of sets from {Ai } and B1 ∪ B2 ∪ . . . ∪ Bn = A1 ∪ A2 ∪ . . . ∪ An (= An ) =⇒ B1 ∪ B2 ∪ . . . ∪ Bn ∪ . . . = A1 ∪ A2 ∪ . . . ∪ An ∪ . . . . Hence µ [ i∈N =µ n [ X X µ(B ) = lim µ(Bi) = lim µ(B1 ∪. . . Bn ) = lim µ(An ) B = • i i i∈N i∈N n→∞ i=1 8 n→∞ →∞ (ii) From (i) in the image we have a graphical representation of our situation. A1 is the first, finite set and A = ∩i∈N Ai . We also observe that A1 \An ր A1 \A. By (i) of the Theorem, µ(A1 \A) = limn→∞ µ(A1 \An ). Hence µ(A1 ) − µ(A) = µ(A1 ) − limn→∞ µ(An ) since µ(A1 ) < ∞. With a little algebra limn→∞ µ(An ) = µ(A). Example Showing that the condition µ(A1 ) < ∞ cannot be removed. The function µ is the Lebesgue measure on R, or more precisely on the Borel set (the measure generalising length/area/volume). For µ([a, b]) = b − a, the length of the interval. For µ[n, ∞) = ∞ since the interval is infinitely long. T If we define An = [n∞) with µ(An ) = ∞ and consider the sequence i∈N An we eventually end up with the emptyset, so \ µ An = µ(∅) = 0. i∈N We see how important this condition is for (ii) in the theorem. Comment One can turn the theorem around to prove that if A is a σ-algebra and µ : A 7→ [0, ∞] satisfying (i) µ(∅) = 0 (ii) µ(A ∪B) · = µ(A) + µ(B) (iii) For any increasing sequence {An } of sets from A [ An ) = lim µ(An ) µ( n→∞ n∈N then µ is a measure. 9 Examples of measures (i) Counting measure. Let X be finite, and A = P(X), the power set of X. Define µ(A) = #A, which is the number of elements in set A. #A (ii) Normalised counting measure. As above except µ(A) = #X , or the (number of elements in A) divided by (the total number of elements in X), and you get the proportion of elements in each set. (iii) Dirac measure. Let (X, A) be arbitrary. Pick an element x0 ∈ X. The Dirac measure of x0 is 1 x0 ∈ A µx0 (A) = 0 x0 6∈ A The physical interpretation is in a weightless room X where only the element x0 has mass, which is 1, then all parts of the room excluding x0 will still be weightless, whereas all parts of the room including x0 will have mass 1. (iv) The n-dimensiona Lebesgue measure. If A is the Borel σ-algebra on Rn , then µ [a1 , b1 ] × [a2 , b2 ] × · · · × [an , bn ] = (b1 − a1 )(b2 − a2 ) . . . (bn − an ) which can be length, area ,or volume depending on the dimension n. For n = 2 this defines the area of a rectangle [a1 , b1 ] × [a2 , b2 ] = (b1 − a1 )(b2 − a2 ). This is illustrated in the image: 10 (v) Coin tossing. For H, T for heads and tails, we have X : the set of all infinite sequences of H’s and T ’s. We start by looking at how we should ascribe to the measure where 1 µ HT HHT . . . H ?? . . . = n. | {z } 2 n For each independent outcome there is a 1/2 probability. For that exact combination we thus get 1/2n . For probability measures we get the probability we would expect. The problem with these measures is what happens when we go from a simple set to a σ-algebra and we have to deal with measures of infinite sets. This is a problematic extension. This happens to be closely tied to the problem of uniqueness. Uniqueness of measures Definition Assume that X is a set. A collection D of subsets of X a Dynkin-system if the following conditions are satisfied (similar as for σ-algebras). (i) X ∈ D (ii) If A ∈ D =⇒ Ac ∈ D S (iii) If {An } is a disjoint sequence of elements in D then • n∈N An ∈ D Any σ-algebra is a Dynkin-system, but not the other way around; there are Dynkin-systems that are not σ-algebras. It is worth remembering that σ-algebras are the truly interesting object of study. Dynkin-systems are just a tool used to prove things about σ-algebras. Proposition Assume G is any collection of subsets of X. Then there exists a smallest Dynkin-system δ(G) containing G and G ⊂ δ(G) ⊂ σ(G). Sketch of proof Show that δ(G) = \ F : F is a Dynkin system and G ⊂ F . Show that the intersection of Dynkin systems is itself a Dynkin system. Recall that \ σ(G) = F : F is a σ-algebra and G ⊂ F . There are F ’s n the δ-intersection that are not used in the σ-intersection. It is a ’bigger’-intersection, so it is a smaller set. 11 Proposition Assume that D is a Dynkin system that is stable/closed under finite intersections. Then the D is a σ-algebra. Stable/closed: E, F ∈ D ⇒ E ∩ F ∈ D. Proof By definition, proposition (i) and (ii) are already satisfied. It is S sufficient to show that if {An } is a sequence of sets from D, then the union An ∈ D. The trick is to turn the sequence {An } into a disjoint sequence {Bn } with the same union. Starting with the situation in a in the image. We define B1 = A1 , and then B2 = A2 \A1 = A2 ∩ Ac1 . Writing this as an intersection, we already see that B2 ∈ D. Next, for the set A3 we define B3 in such a way that we only include the new information. B3 = A3 ∩ Ac2 ∩ Ac1 . Continuing in this manner, we eventually get Bn = An ∩ Acn−1 ∩ . . . ∩ Ac2 ∩ Ac1 . The sequence {Bn } is disjoint (by construction), so we get n \ i=1 Bi = n [ Ai and i=1 [ i∈N Ai = [ Bn ∈ D i∈N This is quite straight forward, but it is quite hard to show in practice. What we need is another result that states that if the generator set G is closed under intersection, then so is δ(G). This is easier to use, but also much harder to prove. To make it easier we will introduce the supporting lemma. Lemma Assume D is a Dynkin system on X. For each set D ∈ D, the collection (also called the trace of D) DD = U ⊂ X : U ∩ D ∈ D 12 is also a Dynkin system. Proof Graphical illustration: we want to see that the gray area is a Dynkin system. To do that, we must verify the three conditions for Dynkin systems. (i) X ∈ D because X ∩ D = D and by definition D ∈ D, so X ∩ D = D ∈ D. (ii) Assuming that U ∈ DD we must prove that U c ∈ DD , that is U c ∩D ∈ D. First we note that (U c ∩ D)c = (U ∩ D) ∪D · c . The first part is contained in D by assumption, and D ∈ D ⇒ D c ∈ D by property (ii) for Dynkin systems. By property (iii) of Dynkin systems, the disjoint union of elements in D is again in D, so we know that (U c ∩ D)c ∈ D. Finally, by property (ii) (U c ∩ D)c ∈ D implies U c ∩ D ∈ D. (iii) Assume Ui i∈N is a disjoint sequence of sets from DD . We need to show that ∪i∈N Ui ∈ DD Since Ui ∈ DD we have Ui ∩ D ∈ D. This means that Ui ∩ D is a disjoint sequence of elements in D, and since D is a Dynkin system, [ (Ui ∩ D) ∈ D. i∈N But ∪i∈N Ui ∩ D = ∪i∈N (Ui ∩ D) ∈ D, so we have ∪i∈N Ui ∈ DD With this we have already proved the hard part of the theorem: If G is closed under finite unions, then δ(G) = σ(G). This will be proved next time. 13 ————————————————————————— 06/09-2010 Recall: A Dynkin system is a collection of D of subsets of X such that: X ∈D (∆1 ) If D ∈ D =⇒ D c ∈ D If {Dn } is a disjoint sequence of sets in D, then (∆2 ) [ ∈D (∆3 ) n∈N Proposition A Dynkin system is a σ-algebra iff it is stable under finite intersections. Proposition If D is a Dynkin system and D ∈ D, then DD = Q ⊂ X : Q ∩ D ∈ D is a Dynkin system. Theorem If G is stable under finite intersections, then δ(G) = σ(G). Proof By previous proposition it suffices to prove that δ(G) is stable under finite intersections. Assume first that G ∈ G and consider the Dynkin system DG = Q ⊂ X : Q ∩ G ∈ D where D = δ(G). By previous proposition DG is a Dynkin system, and since G is stable under finite intersections G ⊂ DG (and we will go on to check that if Q ∈ G, then Q ∈ D). Because if Q ∈ G then Q ∩ G ∈ G ⊂ D, because G is ∩-stable, and D = δ(G). But if DG is a Dynkin system containing G, then δ(G) ⊂ DG , for δ(G) = D, since δ(G) is the smallest Dynkin system. This means that any set D = D = δ(G) belongs to DG , that is D ∩ G ⊂ D. So far; for any G and some D ∈ D then D ∩ G ∈ D. Assume now that D, E ∈ D, and we shall prove that D ∩E ∈ D. Consider DD = Q ⊂ X : Q ∩ D ∈ D . 14 By what we just proved, G ⊂ DD . Since δ(G) is the smallest Dynkin system containing G, we have D = δ(G) ⊂ DD . This means that E ∈ DD and hence, E ∩ D ∈ D. Proof II - Same proof repeated It suffices to prove that if D, E ∈ δ(G), then D ∩ E ∈ δ(G). Let G ∈ G be arbitrary, and let DG = Q : Q ∩ G = D where D = δ(G). Notice that by ∩-stability, G ⊂ DG . Any element in G ⊂ DG (which can be verified). Now choose Q ∈ G, then Q ∩ G ∈ G ⊂ D = δ(G). If the intersection is in D, then Q ∈ DG . Since DG is a Dynkin system, D = δ(G) ⊆ DG , since δ(G) is the smallest possible Dynkin system. This means that for any D ∈ D, then we have D ∈ DG , and hence D ∩ G ∈ D. (An intersection between an arbitrary element in G and D is still in D). However, this alone is not enought to prove the theorem. Second step in the proof: picking two elements in the Dynkin system: D, E ∈ D, and we must prove that D ∩ E ∈ D. DD = Q : Q ∩ D ∈ D . By what we have above: G ∩ D ∈ D for all G ∈ G, hence G ⊂ DD and moreover, D = δ(G) ⊂ DD . This means that E ∈ DD thus E ∩ D ∈ D. Theorem - Uniqueness of Measures Assume that A = σ(G) where G satisfies the following conditions: (i) G is ∩-stable. (ii) G contains an exhausting sequence: there is an increasing sequence {Gi }i∈N of sets in G such that ∪i∈N Gi = X. If µ and ν are two measures such that µ(G) = ν(G), ∀G ∈ G, (the measures agree on all sets in the generator), and µ(Gn ) = ν(Gn ) < ∞ for all n ∈ N, then µ = ν; the measures are equal on all sets in the σ-algebra. (It is always easier to check things in the generator than in the σ-algebra). Proof For each set Gi in the exhausting sequence, we define Di = A ∈ A : µ(A ∩ Gi ) = ν(A ∩ Gi ) . Assume we can prove that Di is a Dynkin system containing G as a subset: Di ⊃ δ(G) = σ(G) = A. 15 where we have equality between the generators by proposition and ∩stability, and the σ-algebra equals A by assumption. Hence for all A ∈ A, µ(A ∩ Gi ) = ν(A ∩ Gi ) for all i. Since {A ∩ Gi } ր A, and by continuity of measures; µ(A) = lim µ(A ∩ Gi ) = lim ν(A ∩ Gi ) = ν(A). i→∞ i→∞ Since A was arbitrary, this is true for all A ∈ A. We must now prove the assumption: that Di are Dynkin systems containing G. If G ∈ G, then µ(G ∩ Gi ) = ν(G ∩ Gi ) (where G ∩ Gi ∈ G by ∩-stability). Thus the inclusion is verified. We must show this is a Dynkin system. We simply verify ∆1 − ∆3 . (∆1 ) X ∈ Di by definition. µ(X ∩ Gi ) = µ(Gi) = ν(Gi ) = ν(X ∩ Gi ). (∆2 ) Assume that A ∈ Di , we must prove that Ac ∈ Di . What we know is that A ∈ Di , which means that µ(A ∩ Gi ) = ν(A ∩ Gi ) and what we need to show is that Ac ∈ Di which is µ(Ac ∩ Gi ) = ν(Ac ∩ Gi ). In the image, the dark gray area is in Di by assumption, and we want to show that the light gray area also is in Di . Using A ∩ Gi ⊂ Gi < ∞: µ(A ∩ Gi ) = µ Gi \(A ∩ Gi ) = µ(Gi ) − µ(A ∩ Gi ) c ν(Gi ) − ν(A ∩ Gi ) = ν Gi \(A ∩ Gi ) = ν(Ac ∩ Gi ). 16 (∆3 ) Assume that {An } is a disjoint sequence of sets in Ai . We need to prove that ∪n∈N An ∈ Di . What we know: µ(An ∩ Gi ) = ν(An ∩ Gi ) for all n. We need to prove that [ [ µ An ∩ Gi = ν An ∩ Gi . n∈N n∈N Beginning on the left side. [ [ Distrib. µ An ∩ Gi An ∩ Gi == µ = | {z } n∈N n∈N Disjoint X µ(An ∩ Gi ) = ν n∈N ν(An ∩ Gi ) = n∈N n∈N [ X | An ∩ Gi {z } Disjoint =ν [ n∈N An ∩ Gi Construction of Measures We know how to measure rectangles and other areas, but we now want to extend this to the Borel set or any other “big” σ-algebra. Definition A collection S of subsets of X is called a semi-ring if the following conditions are satisfied. (S1 ) ∅ ∈ S (S2 ) S, T ∈ S =⇒ S ∩ T ∈ S. (S3 ) If S, T ∈ S, then S\T = ∪· N n=1 Sn where {Sn } is a disjoint family of sets in S. A graphical illustration of the third property. The universe is a half-open rectangle, and S is all possible half open rectangles. For the difference S\T we have three disjoint, half-open rectangles, S1 , S2 and S3 . 17 The first step of the construction is to use µ to define an outer measure µ∗ defined on all subsets A ⊂ X. For A ⊂ X, let n o [ CA = Sn n∈N : Sn ∈ S and Sn ⊃ A . n∈N For the set A in X, we cover it up with balls Sn as depicted in the image: The area of any given ball is µ(Sn ). If we add up all the balls, we get (measuring from the outside) a crude approximation: X µ(Sn ) ≥ µ∗ (A). Define the outer measure µ∗ (A) by o nX ∗ µ(Sn ) : {Sn } ∈ CA . µ (A) := inf n∈N This is the best we can do with the framework set out in the theorem. We will come back and study the properties of outer measures. Caratheodory’s Extension Theorem Assume that S is a semi-ring of subsets of X and µ : S 7→ [0, ∞] (the positive, extended real line, also called R̄+ ), is a function satifying (i) µ(∅) = 0 (ii) µ [ • Sn n∈N ! = X µ(Sn ) n∈N for all disjoint sequences {Sn } in S whose union happen to be in S. Then; µ can be extended to a measure on the σ-algebra σ(S) generated by S. If there is an exhausting sequence {Sn } of sets in S such that µ(Sn ) < ∞ for all n, then the extension is unique. Will prove this next time. 18 ————————————————————————— 13/09-2010 Recall that a semiring of subsets is a collection S such that (S1 ) ∅ ∈ S (S2 ) S, T ∈ S =⇒ S ∩ T ∈ S. (S3 ) S, T ∈ S ⇒ S\T is a finite disjoint union of sets in S: S\T = S1 ∪S · 2 ∪· . . . ∪S · m. We assume that µ : S 7→ [0, ∞] satisfies (i) µ(∅) = 0. (ii) Additivity (if the union happens to be disjoint. There is no requirement that it must be) [ X µ • = µ(Si ) i∈N i∈N for all families of {Si } of sets on S whose union is in S. Our aim is to show that µ can be extended on σ(S) The outer measure µ∗ generated by µ is the function µ∗ : P (X) 7→ [0, ∞] defined by Z X [ ∗ Si ⊃ A, Si ∈ S . µ(Si ) | µ (A) = i∈N i∈N We want to prove that µ∗ restricted to σ(S) is a measure extending µ. Lemma (i) µ(∅) = 0. (ii) A ⊆ B =⇒ µ∗ (A) ≤ µ∗ (B). (iii) Sub-additivity. µ∗ [ i∈N X Ai ≤ µ∗ (Ai ) i∈N which holds for all sequences {Ai }. 19 Proof With basis in the definition of µ∗ . (i) Since ∅, ∅, ∅, . . . is a covering of ∅ we have µ∗ (∅) ≤ µ(∅) + µ(∅) + . . . = 0 =⇒ µ∗ (∅) ≤ 0. (ii) ∗ µ (A) = inf nX i∈N inf nX o µ Si ) | Si is a covering of A ≤ o µ Si ) | Si is a covering of B = µ∗ (B). i∈N We have the inequality since A ⊆ B, so µ∗ (A) ≤ µ∗ (B). (iii) (Useful argument). First we note that there is nothing to prove if we have an infinite sum. It is therefore sufficient to prove the statement when µ(Ai ) is finite. Given ε > 0 we can for each Ai find a covering {Sn }n∈N such that X ε µ(Sni ) ≤ µ∗ (Ai ) + i . 2 n∈N (We can write it this way by the definition of the infimum). Note that {Sni }i∈N,n∈N is a covering of ∪i∈N Ai , hence [ X X XX X ε µ∗ (Ai ) + ε µ∗ Ai ≤ µ(Sni ) = µ(Sni ) ≤ µ∗ (Ai ) + i = 2 n∈N i∈N i,n∈N n∈N i∈N n∈N This is true for any ε > 0, so [ X µ∗ Ai ≤ µ∗ (Ai ). i∈N n∈N Definition S∪ consist of all subsets of X which can be written as a finite, disjoint union of elements in S. Hence A ∈ S∪ ⇔ there exists disjoint set S1 , . . . , Sm ∈ S such that A = ∪· m i=1 Si . Lemma Assume that A, B ∈ S∪ , then (i) A ∩ B ∈ S∪ . (ii) A\B ∈ S∪ . (iii) A ∪ B ∈ S∪ . 20 Proof First observe that S∪ is closed under finite, disjoint unions, because if A, B ∈ S∪ are disjoint, then A = ∪· m · nj=1Tj so A ∪ B = i=1 Si and B = ∪ S1 ∪· . . . Sm ∪T · 1 . . . Tm ∈ S∪ . All S and T ’s are disjoint, so the union of A and B is also disjoint. (i) We have A = S1 ∪· . . . ∪S · m and B = T1 ∪· . . . ∪T · n. A ∩ B = S1 ∪· . . . ∪S · m T1 ∪· . . . ∪T · n To be in element in set, you must be in exactly one Si and on Tj . [ = (Si ∩ Tj ), (S2 ) =⇒ Si ∩ Tj ∈ S. i,j Si ∩ Tj is disjoint from all other Si ’s and Tj ’s, so this is a A ∩ B can be broken down into a disjoint union, and by the remark in this proof, A ∩ B ∈ S∪ . (ii) A\B = S1 ∪· . . . ∪S · m \ T1 ∪· . . . ∪T · n = S1 ∪· . . . ∪S · m \ T1 ∪· . . . ∪T · n c = \ c S1 ∪· . . . ∪S · m T1 ∩. . .∩Tnc = S1 ∩ T1c ∩. . .∩Tnc ∪· . . . ∪S · m ∩ T1c ∩. . .∩Tnc \ \ \ \ [\ = S1 ∩Tjc ∪· . . . ∪· Sm ∩Tjc = S1 \Tj ∪· . . . ∪· Sm \Tj = • (Si \Tj ) j j j j i j Now Si \Tj ∈ S∪ by (S3 ). We have ∩j Si \Tj ∈ S∪ by (i) extended to several sets and not just two. And finally we have ∪· i ∩j (Si \Tj ∈ S∪ by the remark at the beginning of the proof. (iii) For A ∪ B we divide it in three disjoint parts as illustrated in the image. We have A ∪ B = (A\B) ∪(A · ∩ B) ∪(B\A). · By parts (i) and (ii) all the three separate parts are in S∪ , and by the remark at the beginning of the proof, the disjoint union is also in S∪ , thus A ∪ B ∈ S∪ . 21 We would like to extend µ to a function µ on S∪ by A = S1 ∪· . . . ∪S · m. µ(A) = µ(S1 ) + . . . + µ(Sm ). However, we may run into a problem if we have an alternative representation for the set, like for instance A = T1 ∪· . . . ∪T · n , in which case we would have µ(A) = µ(T1 ) + . . . + µ(Tn ). Suddenly µ(A) can seemingly have two different values. Lemma Assume A ∈ S∪ can be written as the disjoint union of sets in S∪ in two different ways. A = S1 ∪· . . . ∪S · m & A = T1 ∪· . . . ∪T · n. Then µ(S1 ) + . . . + µ(Sm ) = µ(T1 ) + . . . + µ(Tn ). Proof Note that Si = Si ∩ T1 ∪· . . . ∪· Si ∩ Tn . Since µ is additive on S: µ(Si ) = n X j=1 Hence m X µ(Si ) = i=1 µ(Si ∩ Tj ), | {z } m X n X ∈S∪ µ(Si ∩ Tj ). i=1 j=1 By the exact same argument with the Si ’s and Tj ’s switched, we get n X j=1 µ(Si) = m n X X µ(Tj ∩ Si ) j=1 i=1 and since both sums are equal to the same expression, we have equality. Definition Define µ : S∪ 7→ [0, ∞] by µ(A) = µ(S1 ) + . . . + µ(Sm ) whenever A = S1 ∪· . . . ∪S · m . Observe that µ = µ on S. ∗ Note: It is µ which is the important measure. We merely introduce µ to prove things about µ∗ . 22 Lemma The measure µ is σ-additive on S∪ , that is if {Tk } is a disjoint sequence from S∪ such that T = ∪k∈N Tk ∈ S∪ , then X µ(T ) = µ(Tk ) k∈N Proof Since T ∈ S∪ , there are disjoint sets U1 , U2 , . . . , Uk such that T = ∪· ni=1 Uk . We have a slight problem: the Ti that intersect with several Uj ’s. Solution: we make a finer partition of Tk . Since Tk ∈ S∪ it can be decomposed as a disjoint union of sets in S, and being clever we may assume that each set in the decomposition is a subset of one of the U’s. For T1 ∪· . . . Sn(1) , each set Si is found within exactly one Uj . We continue T2 = Sn(1)+1 ∪· . . . ∪S · n(2) and eventually Tk = Sn(k−1)+1 ∪· . . . ∪S · n(k) . Let N(i) be the set of indices j suchSthat Sj ⊂ Ui : we collect all the Sj ’s that belong in a specific Ui . Then Ui = • j∈N (i) Sj , and by addivity of µ, this means X µ(Ui ) = µ(Sj ). j∈N (i) µ(T ) = k X µ(Ui ) = µ(Ui ) = i=1 i=1 i=1 On the other hand, k k X X µ(Tk ) = n X X µ(Sj ). j∈N (k)µ(Si ). i=n(k−1)+1 We get ∞ X k=1 hence µ(Tk ) = ∞ X n X (k)µ(Si ) = k=1 i=n(k−1)+1 µ(T ) = ∞ X k=1 23 X µ(Si), i∈N µ(Tk ) The outer measure µ∗ is an extension of µ, but this is quite difficult to prove. To do it we need µ. We start with a lemma. Lemma (i) If A ⊆ B, then µ(A) ≤ µ(B). (ii) If {An } is a sequence from S∪ and ∪n∈N An ∈ S∪ , then ∞ X µ(An ) µ ∪n An ≤ n=1 Proof Consult text book. Similar to previous proofs. Lemma The measure µ∗ is an extension of µ, that is µ∗ (S) = µ(S) for all S ∈ S. Proof Since {S, ∅, ∅, . . .} is a covering of S, µ∗ (S) ≤ µ(S) + µ(∅) + µ(∅) + . . . =⇒ µ∗ (S) ≤ µ(S). ∗ where we used the definition of µ∗ . We now need to prove that Pµ (S) ≥ µ(S). It is sufficient to prove that if {Si } is a covering of S, then µ(Si ) ≥ µ(S). µ(S) = µ(S) = µ [ i∈N a X X b X µ(Si) (S ∩ Si ≤ µ(S ∩ Si ) ≤ µ(Si ) = i∈N i∈N i∈N where a is from sub-additivity and b from S ∩ Si ⊆ Si . With this we have P shown that µ(S) ≤ µ(Si ) for any covering {Si }, and thus µ(S) ≤ µ∗ (S). We have shown the inequality both ways, which means µ∗ (S) = µ(S) for all S ∈ S. 24 ————————————————————————— 16/09-2010 We have S is a semiring, and we have the pre-measure µ : S :7→ [0, ∞], with (i) µ(∅) = 0 (ii) µ [ • Sn k∈N ! = X µ(Sn ). k∈N We defined the outer measure as: X µ∗ (A) = inf µ(Si) : {Si } covers A}. i Fact: µ and µ∗ agree on S. Definition A∗ = A ⊂ X | µ∗ (Q ∩ A) + µ∗ (Q\A) = µ∗ (A) . | {z } Q∩Ac ∗ ∗ A is a collection of µ -measurable sets. We already have, by subadditivity, µ∗ (Q ∩ A) + µ∗ (Q\A) ≥ µ∗ (Q) so all we need to prove is, µ∗ (Q ∩ A) + µ∗ (Q\A) ≤ µ∗ (Q). Lemma S ⊂ A∗ . Proof For any S ∈ S we have to prove µ∗ (Q ∩ S) + µ∗ (Q\S) = µ∗ (Q) ∀Q ⊂ X. Assume first Q = T ∈ S, then ∗ ∗ ) + µ (T \S) = µ(T ∩ S) + µ S ∪ S ∪ . . . ∪ S ≤ µ∗ (T ∩ S 1 2 m | {z } S 25 µ(T ∩ S) + X µ∗ (Si ) = µ(T ∩ S) + i X µ(Si ))µ(T ) = µ∗ (T ). i Now let Q be an arbitrary subset of X. Let {Tn } be any covering of Q, ! [ Tn ∩ S + .... µ∗ (Q ∩ S) + µ∗ (Q\S) ≤ µ∗ n∈N (proof cuts off here). Proposition A∗ is a σ-algebra, and µ∗ restricted to A∗ is a measure. Proof To prove that A∗ is a σ-algebra, we use Dynkin systems. We shall prove that A∗ is a Dynkin system closed under finite intersections. We are going to verify the following 4 properties: (∆1 ): ∅ ∈ A∗ . (∆2 ): A ∈ A∗ implies Ac ∈ A∗ . ∩-stability: A, B ∈ A∗ implies A ∩ B ∈ A∗ A∗ . (∆3 ): If {An } is a disjoint family from A∗ , then ∪· n An ∈ A∗ . Verifying the first property. (∆1 ): µ∗ (Q ∩ ∅) + µ∗ (Q\∅) = µ∗ (∅) + µ∗ (Q) = 0 + µ∗ (Q) = µ∗ (Q). (∆2 ): If A ∈ A∗ , then µ∗ (Q ∩ A) + µ∗ (Q ∩ Ac ) = µ∗ (Q), hence, µ∗ (Q ∩ Ac ) + µ∗ (Q ∩ (Ac )c ) = µ∗ (Q ∩ Ac ) + µ∗ (Q ∩ A) = µ∗ (Q). Thus Ac ∈ A∗ . ∩-stability. It is easier to prove that it is closed under unions, but this is sufficient because you can use DeMorgan’s law and (∆2 ) to show that the intersections are included in A∗ . (If complements are proved, we can always use this trick). 26 Assume A, B ∈ A∗ . We want to prove that A ∪ B ∈ A∗ . Working with the definition: µ∗ Q ∩ (A ∪ B) + µ∗ Q ∩ (A ∪ B)c = | {z } Q\(A∪B) Our goal is to prove that this sum is less than or equal to µ∗ (Q). All we have to work with is that A, B ∈ A∗ . The second term is easy, we have 3 intersections which is okay. Intersections and unions combined are harder. We write A ∪ B as the disjoint intersection A ∪(B\A). · µ∗ Q ∩ (A ∪(B\A)) · + µ∗ (Q ∩ Ac ∩ B c ) = µ∗ (Q ∩ A) ∪(Q · ∩ (B\A)) + µ∗ (Q ∩ Ac ∩ B c ) ≤ µ∗ (Q ∩ A) + µ∗ (Q ∩ (B\A)) + µ∗ (Q ∩ Ac ∩ B c ) = µ∗ (Q ∩ A) + µ∗ (Q ∩ B ∩ Ac ) + µ∗ (Q ∩ Ac ∩ B c ) = µ∗ (Q ∩ A) + µ∗ (Q ∩ Ac ∩ B) + µ∗ (Q ∩ Ac ∩ B c ) = {z } | B∈A∗ µ∗ (Q ∩ A) + µ∗ (Q ∩ Ac ) = µ∗ (Q) by definition of A∗ and using that A ∈ A∗ . All that remains now is to prove ∆3 for A∗ , and after that you have to ∗ show µ A∗ , and then the proof of Caratheodory is completed. Sublemma If A1 , A2 ∈ A∗ are disjoint, then for any Q: Proof µ∗ (Q ∩ A1 ) + µ∗ (Q ∩ A2 ) = µ∗ Q ∩ (A1 ∪ A2 ) . We have µ∗ Q ∩ (A1 ∪ A2 ) , and using that A1 ∈ A∗ and the definition of A∗ , we et µ∗ ∗ [Q ∩ (A1 ∪ A2 ) ∩ A1 + µ∗ [Q ∩ (A1 ∪ A2 )] ∩ Ac1 = (this is the shaded set iin the image intersected with A1 ): 27 µ∗ (Q ∩ A1 ) + µ∗ (Q ∩ A2 ). We can extend this by induction. X ∗ µ∗ Q ∩ (A1 ∪ . . . ∪ An ) = µ (Q ∩ Ai ) i for all disjoint Ai ∈ A∗ . -final parts of proof are missing- 28 ————————————————————————— 20/09-2010 We recall that S is a semiring, which means: S1 : ∅ ∈ S S2 : S, T ∈ S =⇒ S ∩ T ∈ S S3 : S, T ∈ S =⇒ S\T = ∪· N i=1 Sn . We also have the function µ : S 7→ [0, ∞] with (i) µ(∅) = 0 (ii) µ ( ∪· ∞ n=1 Sn ) = P∞ n=1 µ(Sn ) Caratheodory’s Ext. Thm The function µ can be extended to a measure on σ(S). How do we use this theorem to construct the Lebesgue-measure on Rn ? Intuitively the Lebesgue measure is a Borel measure on Rn (rectangles/boxes) such that µ(R) = area(R) for each rectangular box with sides parallel to the axes. I 2 consists of the empty set plus all half open rectangles of the form I = [a, b) × [c, d) (this way the rectagnles fit perfectly together). We start with I 2 , and show that this is a semiring, and then we use Caratheodory’s extension theorem and get a measure on the σ-algebra generated by the semiring, which is the collection of all half-open sets: so the σ-algebra is the Borel σ-algebra. We verify that I 2 is indeed a semiring. We check the three properties. S1 : ∅ ∈ I 2 (obvious). S2 : S, T ∈ I 2 =⇒ S ∩ T ∈ I 2 . This is clear from the illustration, where we can see that the intersection is a half open rectangle with the same form as S and T (and if they are disjoint, then the intersection is the emptyset which is in I 2 by S1 . 29 S3 : The third property is also clear from an illustration. We just need to see that the difference of S and T can be composed by a finite amount of half open rectangles. So, S\T = ∪· ni=1 Si . The second step is to show that I 2 generates the σ-algebra, that is σ(I 2 ) generates B(R). To show this, it suffices to show that all open sets are in σ(I 2 ). The general idea: inside the box B(x, ε), you can find a box (half open rectangle) where the corners are rational numbers: Ix . See image: We do this for all x ∈ U, so U= [ Ix . x∈U This an unountable union, but we can make it countable: so there are only a countable number of rectangles with rational corners. Define A function µ on I 2 is defined by µ [a, b) × [c, d) = (b − a)(d − c). 30 To show that this is a true measure, we must berify the two properties, (i) µ(∅) = 0. S P∞ (ii) µ ( • ∞ n=1 µ(Sn ). n=1 Sn ) = Showing the first property is easy, but the second one is difficult. Instead of verifying it directly we replace it by (ii′ ): finite additivity: µ (A ∪B) · = µ(A) + µ(B), and (iii′ ): For {An }\∅: some decreasing sequence in I 2 , then limn→∞ µ(An ) = 0 for An ∈ I 2 . With finite additivity, proving (iii′ ) is quite easy (and it is covered in the book). We also use Caratheodory’s theorem to construct product measures. Say we have two measurable spaces (X, A, µ) and (Y, B, ν) and we want a measure on the space X × Y such that (µ × ν)(A × B) = µ(A)ν(B). We just show that the product is a semiring and we must show that the measure satisfies the two properties for measures. Chapter 7 - Measurable Mappings A measurable space (X, A) consists of a set X and a σ-algebra A of subsets of X. If we have twp measurable spaces (X, A) and (Y, B), then a map (function) T : X 7→ Y is called A − B-measurable if T −1 (B) ∈ A for all B ∈ B. If this is true for any set B, the mapping is A − B-measurable. (A function is continuous if the inverse of a open set is still open - a related concept). Backwards images are more well-behaved than forward images. Lemma Assume that (X, A) and (Y, B) are measurable spaces, and B = σ(G), B is generated by some set G. If T −1 (G) ∈ A, 31 ∀G ∈ G, then the mapping is A − B-measurable. We can check the generator instead of the σ-algebra: so this lemma makes the potential work easier. Proof Let C = C ∈ Y T −1 (C) ∈ A , the collection of sets in Y with an inverse image in the σ-algebra A. Note that G ⊂ C, and that it suffices to prove that C is a σ-algebra, because B is the smallest σ-algebra: B ⊆ C. To prove C is a σ-algebra, we check the three conditions. (i) Y ∈ C: T −1 (Y ) = X ∈ A. (ii) Assume C ∈ C which implies that T −1 (C) ∈ A. We must show that C c ∈ C c T −1 (C c ) = T −1 (C) where we used that inverse Since c images commute with complements. −1 −1 c T (C) ∈ A, then T (C) ∈ A, since it is a σ-algebra, thus C ∈ C. (iii) Assume {Cn } is a sequence of sets from C, we must prove that ∪n Cn ∈ C. Since Cn ∈ C, T −1 (C) ∈ A. It follows that ! [ [ Cn T −1 (Cn ) = T −1 n∈N n∈N and since A is a σ-algebra [ Cn ∈ C. n∈N Proposition Let (X, A), (Y, B) and (Z, C) be three measurable spaces, and assume that S : X 7→ Y and T : Y 7→ Z are measurable mappnigs (wrt the σ-algebras), then the composition R = T ◦ S (or R(X) = T S(X) is a measurable map. 32 Proof Start with a set C ∈ C, and if R−1 (C) ∈ A, then R is a measurable map. But, R−1 (C) = S −1 T −1 (C) | {z } ∈B {z } | ∈A since both T and S are measurable functions. Assume now that (X, A) and (Y, B) are measurable spaces and T : X 7→ Y is a measurable mapping. Assume also that µ is a measure on (X, A) and introduce a function µT on B by µT (B) = µ T −1 (B) . | {z } ∈A We measure the set B via X with the function µ. Proposition The mapping µT is a measure on (Y, B) and is called the image measure of µ under T . Proof To prove something is a measure, we must show it satisfies the two defining properties. (i) µT (∅) = 0 (ii) For a disjoint sequence {Bn } in B, µT ( ∪· n∈N Bn ) = Showing the first property. since µ is a measure. Def. µT (∅) = µ T −1 (∅) = µ(∅) = 0 The second property. µT [ • Bn n∈N ! =µ T −1 [ • Bn n∈N 33 !! = P n∈N µT (Bn ). µ [ • T −1 (Bn ) n∈N ! = X n∈N X µ T −1 (Bn ) = µT (Bn ) n∈N For a probability space (Ω, A, P ), and some measurable function X : Ω 7→ R (which is a measurable map), this X is called a random variable. Now PX is a measure on R such that PX (A) = P X −1 (A) = P ω ∈ Ω X(ω) ∈ A , and this set is called the distribution of the random variable X. This is an example of a setting where image construction is used. Chapter 8 - Measurable Functions Let (X, A) be any measurable space and let (R, B(R) be the real numbers with the Borel σ-algebra. A measurable map f : X 7→ R is called a measurable function. Definition The function f is measurable if f −1 (B) ∈ A for any Borel set B. Lemma A function f : X 7→ R is measurable if any one of the following conditions are satisfied. (i) f −1 (−∞, a] ∈ A ∀a ∈ R (ii) f −1 (−∞, a) ∈ A ∀a ∈ R (iii) f −1 (−∞, q] ∈ A ∀q ∈ Q (iv) f −1 (−∞, q) ∈ A ∀q ∈ Q In each case we get the Borel set. It is sufficient to check one of these. Proof In each case the sets (intervalls) in question generate the Borel σ-algebra. Lemma All continuous functions f : Rn 7→ R are B(Rn ) − B(R)-measurable. Proof If G ⊂ R is open, then f −1 (G) is also open, so f −1 (G) ∈ O(Rn ), and O(Rn ) ⊂ B(Rn ). Since G is a generator for the Borel sets in R: B(R), this shows that f is measurable. Remark: there are many important measures that are not continuous, so this theorem has some restrictions. 34 Example A measurable function that is not a continuous function. Let A ∈ A. Then the indicator function IA , which is given by 1 x∈A I= 0 x 6∈ A is measurable. Confer with image: For some intervall, we have the inverse image to the indicator function, given by: a≥1 Xc −1 I [−a, a] = A 0 ≤ a < 1 ∅ a < 0. All these sets are in the σ-algebra, thus the inverse image is in the σ-algebra, so I is measurable. 35 ————————————————————————— 27/09-2010 We introduce the extended real numbers. R = R ∪ {−∞, ∞}. In calculations we have the obvious extensions: ∞ + ∞ = ∞, ∞ + a = ∞, ∞ a>0 a·∞ = −∞ a < 0 a∈R In addition, we have the special property that only applies in measure theory: 0 · ∞ = 0. Operations that are undefined are ±∞ . ±∞ ∞ − ∞, The Borel σ-algebra on R is defined as, B(R) = A ⊆ R : A ∩ R ⊂ B(R) . Lemma A function f : X 7→ R is measurable if, and only if, (i) {x : f (x) < α} is measurable. (ii) {x : f (x) ≤ α} is measurable. Simple Functions A simple function on (X, A) is a function of the form f (x) = ∞ X ai IAi (x) n=1 where A ∈ A for i = 1, . . . , n. A simple function is always measurable. Lemma Every simple function has a standard form representation. 36 Proof Proved in the text book. This is important since the standard form is unique, and we don’t measure the same function with different size. Observation Any measurable function f : X 7→ R can be written as the difference between two non-negative, measurable functions. We have f = f + − f − as shown in the image: where we define f + = max{f (x), 0} f − = − min{f (x), 0} Each of these are measurable functions, which is easily seen from the definition. Another useful equality is |f | = f + + f − . Proposition Any measurable function f : X 7→ R can be obtained as the limit of a sequence of simple functions {un }, such that |un (x)| ≤ |u(x)| for all un and all x. If f is non-negative we can choose the sequence {un } to be increasing. Proof It is enough to show that the theorem for non-negative f , because if f is not non-negative we can treat f + and f − separately. The general idea can be illustrated in an image. 37 We slice up the function using 2n parts on the y-axis, and construct a step (n) function. The set Ak is given the value 2kn (the minimal y-axis value). In the next step we get 2n+1 parts, which means every part on the y-axis is split in half. Since we used the minimal y-axis value in the previous step, the new partition gets a higher value this time, and we get an increasing value as we make the partition finer. We define un (x) = n −1 2X k=0 k I (n) . 2 n Ak As explained, this is an increasing function. Every measurable function can be approximated by step functions (simple functions). This concept will be very important for Lebesgue integration. Proposition Assume that {un } is a sequence of measurable functions. Then sup un , inf un , limun = lim sup un , limu∈N = lim inf un u∈N u∈N u∈N u∈N u∈N and, if it exists limu∈N un are all measurable functions. Proof We begin with the ordinary, pointwise supremum supu∈N un . By the definition of the supremum (for a fixed x): sup un = sup un (x) : u ∈ N . u∈N We consider {x : sup un (x) > α} = u∈N [ {x : un (x) > α} n∈N which is the same as writing = [ {x : un ≤ α}c . n∈N Now the set we have complemented in the last step is a collection of measurable functions, so the set is measurable. But the complement of a collection of measurable sets must also be measurable, so the sup is measurable. The result for the infimum is proved in a similar way. We now take a closer look at the lim sup. We need this concept for when we have an oscillating function. For some oscillating sequence αn we get the lim sup and lim inf as shown in the image: 38 and we note that when the sequence has a limit, i.e does not converge, the lim sup and lim inf coincide. The definition is: lim sup = inf sup αn . n→∞ k→∞ n≥k so we stop at k and look at the supremum from there and out. Now, by the first part of the proof, we know that the supremum is measurable, so the inner part of the lim sup is measurable, and we also know the infimum is measurable, so the lim sup is measurable. The result for the lim inf is similar. If the limit exists, we also consider limn→∞ un , but as mentioned, this coincides with both lim inf and lim sup, and since they are measurable so is the limit. Proposition If u and v are measurable functions, then so are u ± v, u · v max(u, v) and min(u, v) whenever they exist/are defined. Proof If u and v are simple functions, so are the resulting functions, and they are again measurable, since simple functions ⇒ measurable. If u and v are general, measurable functions, we approximate them by simple functions {un } and {vn }, and we get u ± v = lim un ± vn , n→∞ max(u, v) = lim max(un , vn ) uv = lim un vn n→∞ min(u, v) = lim min(un , vn ). n→∞ n→∞ Since limits are measurable functions, so are all these functions. This kind of argument will be used a lot later. Consider the case where we have the two measurable spaces (X, A) and (Y, B), and some measurable mapping T : X 7→ Y . In addition we have a function u : X 7→ R. Sometimes we want to have a function, w say, that mimicks u and maps: w : Y 7→ R. Confer with image: 39 If such a function w exists, u is measurable wrt to the σ-algebra σ(T ) generated by T , because u−1 (B) = (w ◦ T )−1 (B) = T −1 w −1 (B) | {z } ∈B | {z } ∈σ(T ) Factorization Lemma If (X, A) and (Y, B) are two measurable spaces, and T : X 7→ Y is measurable wrt A and B, then for any σ(T )-measurable function u : X 7→ R, there is a B-measurable function w from T 7→ R such that u = w ◦ T . Proof Assume first that u is the indicator function u = IA . Since u is σ(T )measurable, then A ∈ σ(T ) and because A = T −1 (A) for some set B ∈ B, but then we can take w = IB . P IfP u is a simple function, u = ∞ n=1 ai IAn we use the same trick to get w= an IBn . We have shown this is true for simple functions, and we approximate those to more general functions. If u is a general σ(T )-measurable function, we approximate by a sequence of simple functions, such that un (x) → u(x). We know that for each n there is a simple function w such that wn ◦ T = un . We let w = lim sup un , since the limit may not exist, but the lim sup always does. w(T (x)) = sup lim wn (T (x)) = sup lim un (x) = lim un (x) = u(x). n→∞ n→∞ n→∞ Where we used the construction of wn and the fact that un converges. This is also a much used trick we will see again. 40 ————————————————————————— 30/09-2010 Chapter 9 - Integration -Aim of this lecture R We are going to define integrals of the form f dµ for all measures µ and ”all” measurable functions f . We do this in steps. (i) f is the indicator function. (ii) f is a positive, simple function. (iii) f is a positive, measurable function. (iv) f is an integrable functions. (Not just for measurable functions, since we can’t integrate all measurable functions). •(i) When f is the indicator function, f = IA for some measurable A, we have the case as shown in the image: Thinking in terms of classical integrals, we would want the volume of the ”box” indicated in the image. The volume of this is simply the area of A times the heigth, which is 1. If we use the Lebesgue measure, area(A) = µ(A). For the indicator function we define, Z f dµ = µ(A). Pn •(ii) We now assume that f = i=0 ai IAi , that is a simple function on standard form. Looking at the image: 41 and using the same reasoning as for the indicator function (i), we would like to define this as Z n X f dµ = ai µ(Ai). i=0 However, before we make this a formal definition, there is a potential problem we must adress. What if there P is another standard representation (they are, after all, not unique) f = nj=0 bj IBj , in which case we would have Z f dµ = m X bj µ(Bj ). j=0 If this is different from the first integral, we have a problem, since we have inconsistency. We have to show that the volume of the function f is independent of the standard representation. Lemma P P Assume that ni=0 ai IAi and nj=0 bi IBj are standard representations of the same function f . Then, n X ai µ(Ai ) = m X bj µ(Bj ). j=0 i=0 Proof What we have with these standard representations, is two ways of partitioning the set X, as illustrated in the image. We make a new, finer, partition and show that the volumes of the different partitions coincide. Since B1 , . . . , Bm is a partition of X, we can write Ai = Ai ∩ B0 ∪· Ai ∩ B1 ∪· . . . ∪· Ai ∩ Bm Since X = B0 ∪B · 1 ∪· . . . ∪B · m , and Ai = Ai ∩ X. This means we can write µ(Ai ) = m X µ(Ai ∩ Bj ). j=0 42 Using this, we can rewrite the proposed definition of the integral (volume): n X ai µ(Ai ) = i=1 n X i=0 ai m X µ(Ai ∩ Bj ) = j=0 n X m X ai µ(Ai ∩ Bj ). i=0 j=0 Using the exact same approach we get a similar expression for the other representation: m m X n X X bj µ(Bj ) = bj µ(Ai ∩ Bj ). j=0 j=0 i=0 The order of the sums are different, but that doesn’t matter since they are finite sums. The main difference here is ai and bj . From the definitions of simple functions, ai is the value of f at Ai , and bj is the value of f on Bj . If we are on the intersection Ai ∩ Bj then f can only have one value so ai = bj . If they are disjoint we get µ(∅) = 0 so the values don’t matter. This verifies the uniqueness of the sums. m X j=0 bj µ(Bj ) = n m X X bj µ(Ai ∩ Bj ) = n X m X ai µ(Ai ∩ Bj ) = i=0 j=0 j=0 i=0 n X ai µ(Ai ) i=1 Definition If f is a positive, simple function with standard representation f= n X ai I( Ai ) i=1 then we define the integral of f wrt the underlying measure µ to be: Z n X f dµ = ai µ(Ai). i=1 Proposition Assume f, g are positive, simple functions. Then R R (i) af dµ = a f dµ for all real a > 0. R R R (ii) (f + g)dµ = f dµ + gdµ. R R (iii) If f ≤ g, then f dµ ≤ gdµ. Proof (i) By the definition of simple functions: Z Z n n X X af dµ = a · ai µ(Ai ) = a ai µ(Ai ) = a f dµ. i=0 i=0 43 Pn Pm (ii) Assume that f = a I and g = i A i i=0 j=0 bj IBj are standard representations. Then, the right side of the proposition is: Z f dµ + Z gdµ = n X ai µAi + i=0 m X bj µBj . j=0 We have to show that the left side is equal to this. To compute the integral of f + g, we need a standard representation of f + g. Observe that the sets Ai ∩ Bj gives us a partition of X (just like in the previous proof), and on each Ai ∩ Bj , f has aj and g has bj , so f + g has a constant value on that set: consequently we have a standard representation of f + g: X f +g = (ai + bj )IAi ∩Bj . i,j By our definition: Z (f + g)dµ = X (ai + bj )µ(Ai ∩ Bj ) i,j and we split it up X ai µ(Ai ∩Bj )+ i,j X bj µ(Ai∩Bj ) = i,j = n X i=0 ai µ(Ai ) + n X i=0 m X ai m X j=0 bj µ(Bj ) = j=0 µ(Ai ∩Bj )+ m X bj j=0 Z f dµ + Z n X µ(Ai ∩Bj ) i=0 gdµ. (iii) This is easily proved using a simple trick. g = f + (g − f ). Since g ≥ f , g − f is positive. Z Z Z Z Z (ii) gdµ = f + (g − f ) dµ = f dµ + (g − f )dµ ≥ f dµ. | {z } ≥0 44 We are now ready for the third step: integration of positive, measurable functions. Definition R If f is a positive, measurable function, define the integral f dµ of f wrt µ by Z o nZ f dµ = sup gdµ g ≤ f, g positive, simple funcion . So for some function f , we approximate with the largest possible step function g that is smaller than or equal to f . In a sense this is similar like the Riemann integral (except we can do the same over far more complex funtions). Remark: We now have two seperate definitions for the integral, but this is okay since the definitions agree. In the text book they have worked their way around this by designating the first integral definition by Iµ . (For this definition you use that f itself is one of the g’s and use proposition (iii)). ————————————————————————— 04/10-2010 Positive, simple functions f= X ai IAi , Z f dµ = X ai µ(Ai ). Positive, measurable functions Z Z gdµ = sup f dµ | f ≤ g, f’s are positive, simple functions Question: If fn → f , will R fn dµ → R f dµ? 45 Monotone Convergence Theorem (Beppo-Levi’s Theorem) Assume that {un } is an increasing sequence of positive, measurable functions. Then the limit Z Z un dµ = lim un dµ. lim n→∞ n→∞ (If we have an increasing sequence of positive functions, you can interchange limits and integral signs). Proof ————————— Since fn ≤ limn→∞ fn , we have by a previous result, (if a function is greater or equal, so is its integral) Z Z fn dµ ≤ lim fn dµ, ∀n. n→∞ hence, since this is true for all n, Z Z fn dµ ≤ lim fn dµ. lim n→∞ n→∞ This was the easy inequality. We now want to show the opposite inequality, Z Z fn dµ ≥ lim fn dµ, lim n→∞ n→∞ but to show this it suffices to show that Z Z fn dµ ≥ gdµ, lim n→∞ for any simple function g ≤ limn→∞ fn , or alternatively, limn→∞ fn = Psup g, and if it holds for all such g it also holds for the sup. Assume that g = ai IAi is a positive, simple function and g ≤ limn→∞ fn . We let 0 < α < 1 (which we think as something close to 1) and consider the set Bnα = {x ∈ X | αg(x) ≤ fn (x)} . Note that {Bnα }n∈N ր X as n gets larger. The set Bnα gets larger since the fn gets larger, and when they do the set of points x that satisfy the inequality grows. This means that Z Z Z Z fn dµ. αg · IBnα dµ ≤ fn · IBnα dµ ≤ fn dµ ≤ lim n→∞ or, in short, Z αg · IBnα dµ ≤ lim n→∞ 46 Z fn dµ. Now we consider what happens as n → ∞ and α → 1. Note that Z Z X Z X αgIBnα dµ = α ai IAi IBnα dµ = α ai IAi ∩Bn∞ dµ and by the definition of simple functions, =α X ai µ(Ai ∩ n→∞ Bnα ) −→ α X ai µ(Ai ) = α Z gdµ. Where we used that as n → ∞, Bnα → X, so Ai ∩ Bnα → Ai ∩ X = Ai . We thus have that Z Z αgIBnα dµ ≤ lim fn dµ ∀α, n n→∞ and since this is true for all n, we can take the limit, so Bnα and we are left with Z Z α gdµ ≤ lim fn dµ ∀α n→∞ Again, by similar reasoning, this is true for all α, so we can make it come arbitrarily close to 1, or in effect let α → 1, and get Z Z gdµ ≤ lim fn dµ. n→∞ A useful converse result to the theorem. Corollary Assume that f is a positive, measurable function and that fn is any increasing sequence of simple functions converging to f . Then Z Z f dµ = lim fn dµ. n→∞ Proof : Special case of the theorem. By the definition Z Z gdµ | g simple, positive functions . f dµ = sup R R We only require that gn ր f , and then we can write f dµ = limn→∞ gn dµ. Since we can freely choose such a sequence, this gives us an extra degree of freedom when working with converging integrals. 47 Proposition Assume that f, g are positive, measurable functions. Then R R (i) αf dµ = α f dµ. R R R (ii) (f + g)dµ = f dµ + gdµ. R R (iii) If f ≤ g, then f dµ ≤ gdµ. Proof (i), (iii): Follows immediately from the definition. (ii) Let un and vn be increasing sequences of step functions converging to f and g. (We know such sequences exist). By the corollary, Z Z Z Z vn dµ. f dµ = lim un dµ gdµ = lim n→∞ n→∞ Also, since un + vn ր f + g, we also have (by the same corollary) Z Z (f + g)dµ = lim (un + vn )dµ. n→∞ Hence, Z (f + g)dµ = lim n→∞ = lim n→∞ Z Z (un + vn )dµ = lim n→∞ un dµ + lim n→∞ Z vn dµ = Z Z un dµ + f dµ + Z Z vn dµ gdµ. Fatou’s Lemma If {un } is a sequence of positive functions, then Z Z lim inf un dµ ≤ lim inf un dµ. n→∞ n→∞ (We have weaker conditions in this case, so we don’t know if the limit exists: so we take the inf in addition to the lim). We only have an inequality, and show a counterexample to the equality. Counterexample We define the function 1 |x| ≥ n un (x) = 0 |x| < n which has a graph as shown in the following image. 48 In this particular case the limit does exist, but lim inf un (x) n→∞ = limn→∞ un (x) = 0 since when n becomes infinitely large, the interval which R is 0 also becomes infinitely long, so lim inf un (x)dµ = 0. But, if we look at n→∞ R any function in the sequence, un (x)dµ = ∞, since we are looking at the area under the graph, and for, say n = 3, we have infinite areas under the graph for R n > |3|. Since this is true for at least one element in the sequence, lim inf un (x)dµ = ∞, and we have the counterexample. n→∞ Proof (of Fatou’s Lemma) By the definition of the lim inf, lim inf un = lim inf un n→∞ n→∞ k≥n where inf k≥n un is an increasing sequence, since we are taking the infimum of a smaller and smaller set. We set fk (x) = inf k≥n un , and have that fk (x) is an increasing sequence, in which case we can apply the monotone convergence theorem to {fn (x)} and get Z Z lim fn dµ = lim fn dµ. n→∞ n→∞ But, limn→∞ fk (x) = lim inf un , so we can rewrite this as n→∞ Z lim inf un dµ = lim n→∞ n→∞ Z fn dµ. (Using that fn (x) ≤ un (x), and that lim inf = lim when the limit exists): Z Z = lim inf f dµ ≤ lim inf un dµ. n→∞ n→∞ Fatou’s lemma is weak due to the inequality, but at the same time strong since the conditions are so weak. It is typically used as a first step, opening up the usage for other theorems. 49 Chapter 10 - Integrable Functions Recall that we can split f as f + − f − . Definition A measurable function f : X 7→ R is called integrable if Z Z + f dµ < ∞ and f − dµ < ∞, and we define the integral Z Z Z + f dµ := f dµ − f − dµ. (We assume they are finite to avoid ∞ − ∞ in this definition. Some books allow one of them to be infinite which allows ±∞ integrals, but we will not). The set of integrable functions is denoted L1 (µ) and will be a key concept for a while. Proposition This gives us an easy way to verify integrability. The following are equivalent for a measurable function f . (Propery (iv) is a useful trick). R (i) f is integrable or ( f dµ is finite). R R (ii) f + and f − are integrable (or f + dµ and f − dµ are finite). R (iii) |f | is integrable (or |f |dµ is finite). (iv) There exists a function ω ∈ L1 (µ) such that |f | ≤ ω. Proof Suffices to show (i)⇒(ii)⇒(iii)⇒(iv)⇒(i). (i)⇒(ii): This is just a slight rephrasing of the definition. (ii)⇒(iii): We know that + − 1 f , f ∈ L (µ) =⇒ Z + f dµ, Z f − dµ < ∞. Since |f | = f + + f − |, we have Z Z Z + |f |dµ = f dµ + f − dµ < ∞, | {z } | {z } <∞ <∞ so |f | is a positive function with finite integral, thus f ∈ L1 (µ). 50 (iii)⇒(iv): We simply choose ω = |f | and since |f | ≤ |f | is true the implication follows. (iv)⇒(i): We have Z + f dµ ≤ Z |f |dµ ≤ Z ωdµ < ∞ since, by assumption, ω is integrable. Z Z Z − f dµ ≤ |f |dµ ≤ ωdµ < ∞ Both f + and f − are finite, so f ∈ L1 (µ). Lemma Assume f is a measurable function, and that h and g are positive, integrable functions such that f = h − g. Then f is integrable and Z Z Z f dµ = hdµ − gdµ. Proof We have |f | ≤ |h| + |g| = h + g (by the triangleRinequality and are R since they R positive), and we know that h+g ∈ L1 (µ) since (h+g)dµ = hdµ+ gdµ < ∞ (by assumption they are integrable). By (iv) of the previous proposition, f is integrable. It remains to show that f can be written as h − g. Combining f = h − g and f = f + − f − , we get f + − f − ⇒ f + + g = f − + h, hence, Z Z + f dµ + + Z gdµ = Z Z Z − f dµ + − Z Z hdµ f dµ + gdµ = f dµ + hdµ Z Z Z Z Z Z Z + − f dµ − f dµ = hdµ − gdµ =⇒ f dµ = hdµ − gdµ. Theorem Assume that f, g ∈ L1 (µ). Then (i) αf ∈ L1 (µ) for all α ∈ R and Z Z αf dµ = α f dµ. 51 (ii) f + g ∈ L1 (µ) and Z Z Z (f + g)dµ = f dµ + gdµ. (iii) max(f, g), min(f, g) ∈ L1 (µ). R R (iv) f ≤ g =⇒ f dµ ≤ gdµ (v) Rather intuitive result: Z Z f dµ ≤ |f |dµ. (In the first integral we can have positive and negative parts of the function that cancel each other out, while on the right side we take the absolute value of the function before we integrate). Proof - Only (ii), the rest is left to the student. (ii) Let f = f + − f − and g = g + − g − , then (f + g) = f + + g + + f − + g − | {z } | {z } pos. in L1 (µ) pos. in L1 (µ) By assumption each part is finite, so we can apply the previous lemma with h = (f + + g + ) and g ∗ = (f − + g − ). We get f + g ∈ L1 (µ) and Z Z Z + + (f + g)dµ = (f + g )dµ − (f − + g − )dµ = Z Z Z + (i) Since α ∈ R and not R, α < ∞ so R − Z f dµ + g dµ − f dµ − g − dµ = Z Z Z Z Z Z + − + − f dµ − f dµ + g dµ − g dµ = f dµ + gdµ + αf dµ < ∞ and thus integrable. By definition of the integral, Z Z Z Z Z Z + − + − αf dµ = αf dµ − αf dµ = α f dµ − f dµ = α f dµ. (iii) Both max(f, g) and min(f, g) return one of the functions f, g which are integrable by assumption. 52 (iv) We use the trick, g = f + (g − f ), and write out the integral. Z Z Z Z Z (ii) gdµ = f + (g − f )dµ = f dµ + (g − f )dµ ≥ f dµ. R since g ≥ f , we have (g − f ) ≥ 0 and then (g − f )dµ ≥ 0 and the inequality is apparent. (v) Since |f | = f + + f − , we can write Z Z Z Z (ii) + − + |f |dµ = f + f dµ = f dµ + f − dµ whereas Z Z Z Z Z f dµ = f + dµ − f − dµ ≤ f + dµ + f − dµ and now we can remove the absolute value, since f + and f − are positive. Using the first step we conclude the proof. Z Z Z Z (ii) + − + − = f dµ + f dµ = f + f dµ = |f |dµ. 53 ————————————————————————— 07/10-2010 We have previously defined + Z f <∞ 1 f dµ, f ∈ L (µ) ⇐⇒ f− < ∞ Definition If A ∈ A, we define Z Z f dµ = A IA · f dµ. We only consider the the function f over the set A and set it to 0 everywhere else. Graphical illustration. Note that if f ∈ L1 (µ), then IA f ∈ L1 (µ), simply because |IA | ≤ |f | (bigger without being restricted to the set A, and by assumption f ∈ L1 (µ) and then we can apply previous result with f = ω). Proposition If f is a positive, measurable function, then Z ν(A) = f dµ A is a measure on A. Proof Since ν(A) ≥ 0 ∀A ∈ A, all we need to show are the two axioms for measures. (i) ν(∅) = 0. (ii) ν [ • Ai i∈N ! = for all disjoint sequences of sets from A. 54 X i∈N ν(Ai ) (i) Observe that ν(∅) = Z f dµ = ∅ Z I∅ f dµ = Z 0dµ = 0. (The zero function, 0 is a simple function). (ii) Assume that we have a disjoint sequence of measurable sets{Ai }. Note that ! Z Z [ ν • Ai = f dµ = I∪· ∞ f dµ (1) i Ai ∪· ∞ i Ai i∈N X i∈N ν(Ai ) = lim n→∞ n X ν(Ai ) = lim i=1 n→∞ n Z X IAi f dµ = lim n→∞ i=1 Z X n i=1 IAi f dµ = lim n→∞ Z I∪· ni Ai f dµ (2) We must show that (1) and (2) are equal. Note that I∪ni Ai f ր I∪∞ f i Ai as n → ∞, since, if we call define the left side as fn , we see that f1 = f (A1 ), f2 = f (A1 ) + f (A2 ) etc, so we have an increasing sequence! By the monotone convergence theorem, we can interchange limits and integrals, so Z Z Z I∪· ni Ai f dµ = lim I∪· ni Ai f dµ = I∪· ∞ f dµ lim i Ai n→∞ n→∞ and using (1) and (2) with this, we get that ! [ X ν • Ai = ν(Ai ) i∈N i∈N (You need to apply some theorem when you are faced with interchanging limits and integrals, so you must look close at the elements you are working with and see if you can use a theorem). Concept of ’Almost Everywhere’ In measure theory we can’t distringuish between two sets that only differ on a null set. Example If f, g : X 7→ R, we say that f and g are equal almost everywhere if there is a set N of measure zero such that {x : f (x) 6= g(x)} ⊂ N. 55 In general a property holds almost everywhere (abbreviated a.e) if the set where it fails is contained in a set of measure zero. Example fn → f a.e. What does this mean? It means the set {x : fn (x) 6→ f (x)} is contained in a null set. Proposition For u ∈ L1 (µ) Z |u|dµ ⇐⇒ u = 0 a.e Proof Assume that u = 0 a.e, which P means that the set {x : ux) 6= 0} has measure zero (def of ’a.e’). If f = ai IAi is a simple function on standard form less than |u|, then f must be 0 a.e, and hence Z X f dµ = ai µ(Ai) = 0. Whenever ai 6= 0 we are on the null set which means µ(Ai ) = 0. So by the definition of the integral, Z Z |u|dµ = sup f dµ : f ≤ |u|, f positive, simple funcion . Assume R |u|dµ = 0. Note that in particular, An = {x : |u(x)| ≥ 1 } n 1 has measure 0. (Because R 1 n IAn is a1 simple function less than the measurable function |u(x)|, and n IAn dµ = n µ(An )). But then the set {x : |u(x)| = 6 0} = [ An , n∈N but this is a union of null sets, which is again a null set. By subadditivity, X µ{x : |u(x)| = 6 0} ≤ µ(An ) = 0. | {z } n∈N 56 0 Corollary If f ∈ L1 (µ) and µ(N) = 0, then Z f dµ = 0. N (The integral over a null set is 0). Proof Z f dµ = N Z IN f dµ ≤ Z Prop. |In f |dµ = 0. Corollary If u, v ∈ L1 (µ) and u = v a.e, then Z Z udµ = vdµ. Proof Z Z Z Z udµ − vdµ = (u − v)dµ ≤ |u − v|dµ Prop. = 0 since |u − v| = 6 0 on a set of measure 0, or alternatively, |u − v| = 0 a.e. 57 ————————————————————————— 11/10-2010 (X, A, µ) is a measure space. Proposition If f, g are measurable functions and Z Z f dµ = gdµ, A ∀A ∈ A A then f = g a.e. Proof Assume for a contradiction this is not the case. Then, 0 < µ {x : f (x) 6= g(x)} = µ {x : f (x) > g(x)} + µ {x : g(x) > f (x)} (the first set is the union of the two on the right, and since they are disjoint we can add their measures). Since these sets are strictly greater than 0 one of them must have positive measure, say the one for f (x) > g(x). (The following reasoning can be used on the other case, where we just exchange f and g). This set can be rewritten: [ 1 {x : f (x) > g(x)} = x : f (x) > g(x) + . n n∈N At least one of the sets on the right side must have a positive measure. 1 x : f (x) > g(x) + > 0 for some n. n But the, the indicator function multiplied by n1 , (to shorten the notation we define θ = {x : f (x) > g(x) + n1 }: n1 Iθ is a simple function majorized by f (x) − g(x) > n1 : since f (x) > g(x) + n1 we have f (x) − g(x) > n1 . This means, Z 1 1 > 0. [f (x) − g(x)] dµ ≥ µ x : f (x) > g(x) + n n θ Hence, since both f and g are measurable functions which means we can split the integral, Z Z Z Z f dµ − gdµ > 0 =⇒ f dµ > gdµ. θ θ θ θ which contradicts the assumption theorem. (This theorem also applies for a sub-σ-algebra C ⊂ A). 58 Chapter 11 Basic question: When can we interchange limits and integrals? Z Z lim un dµ = lim un dµ. n→∞ n→∞ Fatou’s Lemma Z lim inf un dµ = lim inf n→∞ n→∞ Z un dµ (un positive) This lemma is an example of when we can change integrals and limits. Generally the measurability of the limit is never the problem, only that a function can be too ”big”. In the Riemann theory it can be both too big and too irregular. Monotone Convergence Theorem (Generalized) Assume {un } is an increasing sequence of functions in L1 (µ), and let u = R limn→∞ un . Then u ∈ L1 (µ) if and only if limn→∞ un dµ < ∞ and in that case Z Z Z Z udµ = lim un dµ =⇒ lim udµ = lim un dµ. n→∞ n→∞ n→∞ (Compared to the first version, Beppo-Levi, we do not assume the functionals are positive, only that the sequence is increasing). Proof We reduce this to the Beppo-Levi case. Note that {un − u1 } is an increasing sequence of positive functions converging to u − u1 . Hence, by Beppo Levi’s theorem, Z Z Z Z Z B.L (u−u1 )dµ = lim (un −u1 ) = lim (un −u1 )dµ = lim un dµ− u1 dµ. n→∞ R n→∞ n→∞ (*) un dµ = ∞, then u cannot be integrable, so we have Note that if limn→∞ proved the sufficiency. R R If limn→∞ un dµ < ∞, then (u − u1 )dµ < ∞, and since it is a finite, positive function it is measurable, so u−u1 ∈ L1 (µ). But then u = (u−u1 )+u1 are both two measurable functions, and their sum is also measurable which means u ∈ L1 (µ), which proves the necessity. Returning now to (*), where we verified Z Z Z (u − u1) = lim un dµ − u1 dµ =⇒ n→∞ 59 we can now use that we know the terms in the integral are finite and measurable. Since we know they are finite we can cancel them. Z Z Z Z Z Z un dµ un dµ−u1 dµ = udµ = lim udµ−u1 dµ = lim n→∞ n→∞ Note: we cannot cancel out terms unless we know they are finite. If we do we may reach contradictions, like for instance ∞ − a = ∞ − b, which is true since both sides are infinite. But if we cancel the infinity terms we are left with −a = −b which is incorrect. We now look at a very important convergence theorem, and the one which is most used in practice. Lebesgue’s Dominated Convergence Theorem Assume that {un } is a sequence of functions in L1 (µ) converging a.e to a function u. Assume further that there is a function w ∈ L1 (µ) such that |un | ≤ w for all n, then Z Z Z Z un dµ un dµ =⇒ lim un dµ = lim udµ = lim n→∞ n→∞ Also, lim n→∞ Z n→∞ |u − un |dµ = 0 (which is a slightly stronger L1 (µ)-convergence. Before we prove this, we take a look at why we need dominated convergence. Consider the function given by 1 if x ∈ [n, n + 1) un (x) = 0 otherwise. The graph to this function is something like this: 60 R We note that limn→∞ dµ = 0, because sooner or later we will pass n, and from there on out we always have 0. On the other hand, for every n, we have ∈ un dµ = 1, since we have an interval of size 1 where the function is 1. So Z Z Z lim un (x)dµ = u(x)dµ = 0 but un dµ = 1, ∀n. n→∞ In this case the usual convergence theorems don’t work, but we could use DCT. Proof Note that 2w − |u − un | ≥ 0. Since |un | ≤ w and |u| ≤ |w| and w ∈ L1 (µ), then u ∈ L1 (µ), or u is integrable. This means that Z Z Z Z udµ − un dµ = (u − un )dµ ≤ |u − un |dµ. R It suffices to prove that limn→∞ |u − un |dµ = 0. We use Fatous’ lemma: Z Z F.L 2wdµ = lim inf (2w − |u − un |)dµ ≤ n→∞ lim inf n→∞ Z (2w − |u − un |)dµ = Z 2wdµ − lim sup n→∞ Z |u − un |dµ. where we in the last step used that both terms are measurable so we can split up the integral, 2w has no n-dependance so we can remove the lim inf, and the lim inf becomes − lim sup. We have thus established, Z Z Z 2wdµ ≤ 2wdµ − lim sup |u − un |dµ. n→∞ and since the integral of absolute values must be positive, the only possibility for this inequality is that we have equality and that the integral we are subtracting is 0. So, we have verified that Z lim sup |u − un |dµ = 0. n→∞ Since the integral must be positive, and the lim sup is 0, then so is the lim inf and when they agree we have convergence, so Z |u − un |dµ = 0 lim n→∞ Applications We see some applications of the convergence theorem. 61 Proposition Let u : (a, b) × X 7→ R (the function maps from an intervall and a set in the measurable space X to the real numbers) satisfies (i) x 7→ u(t, x) is L1 (µ) for all t ∈ (a, b). (ii) t 7→ u(t, x) is continuous for all x. (iii) There is a w ∈ L1 (µ) such that |u(t, x)| ≤ w(x) for all t and x. Then the function v(t), defined as v(t) = is continuous. Z u(t, x)dµ(x) Proof It is sufficient to show that v(tn ) → v(t) for all sequences tn → t, i.e we want to verify Z Z ? u(tn , x) dµ(x) −→ | {z } un (x) u(t, x) dµ(x) | {z } u(x) where we use the underbraced versions since t is fixed. Now we have the conditions in the dominated convergence theorem, especially due to assumption (iii), so we know that Z Z un dµ(x) = u(x)dµ(x) lim n→∞ Proposition Let u : (a, b) × X 7→ R satisfies (i) x 7→ u(t, x) is in L1 (µ) for all t. (ii) t 7→ u(t, x) is continuosuly differentiable at all points (a, b). | ≤ w(x). (iii) There is a function w ∈ L1 (µ) such that | ∂u ∂t Then the function Z v(t) = u(t, x)dµ(x) is differentiable and (we differentiate under the integral sign) Z ∂u ′ (t, x)dµ(x). v (t) = ∂t So the proposition basically tells us that Z Z d ∂u u(t, x)dx = (t, x)dµ(x) dt ∂t 62 Proof The conclusion means that we have Z v(s) − v(t) ∂u lim = (t, x)dt. s→t s−t ∂t We look at sequences of functions converging to, and not the limit-definition of the derivative, so it suffices to prove that Z ∂u v(tn ) − v(t) lim = (t, x)dµ(x) n→∞ tn − t ∂t for all sequences tn → t. Using the definition of v(t), we get R R u(tn , x)dµ(x) − u(t, x)dµ(x) u(tn , x) − u(t, x) v(tn ) − v(t) = lim = lim dµ(x). lim n→∞ n→∞ n→∞ tn − t tn − t tn − t From calculus/real analysis we see we can use the mean value theorem, so for some intermediate point cn we have that this equals ∂u (c , x), and when ∂t n ∂u cn → t we get ∂t (t, x) which, by assumption, is bounded by w. We apply the Lebesuge dominated convergence theorem, and get Z ∂u (t, x)dµ(x) = ∂t Lebesgue vs Riemann We consider the classical example of the function that is not Riemann integrable but Lebesgue integrable. For f : [0, 1] 7→ R, we define 1 if x is rational . f (x) = 0 otherwise The upper and lower step functions in the Riemann integral always converge to 1 and 0 respectively, so it is not Riemann integrable. It is Lebesgue integrable since it is 0Ra.e (the rational numbers are counable, and countable sets are null sets), so [0,1] f dµ = 0. In a more general version, 1 if x can be written on the form m for k ≤ n k fn (x) = . 0 otherwise In this case every fn (x) is Riemann integrable (since we can partition them as step functions over the rational points), but as fn → f , f is not Riemann integrable as we saw over. When we take the limits of functionals that are Riemann integrable, we get problems with the limit, which is never the case with Lebesgue integrals. In view of how important limits are as fundamental analysis wholly depend on them, this becomes on of the most important reasons to why we discard Riemann integrals and use Lebesgue integrals instead. 63 ————————————————————————— 18/10-2010 64