Chapter 4 Integration 4.18. By the definition of the Lebesgue integral it suffices to prove this when f is a simple function. In fact, this is easily reduced to the case where f = 1A for A ∈ B([0, 1]d ). So we wish to prove that if A is measurable then there exists continuous functions gn such that �gn − 1A �p → 0. For the sake of pedagogy, we will concentrate on the case d = 1. The general case follows similarly, and is left to the student. Let A denote the algebra of finite disjoint unions of sets of the form (a, b] where 0 � a < b � 1. If A ∈ A, then it is easy to see that there exist continuous functions gn such that: (i) 0 � gn (x) � 1; (ii) gn = 1 on A; (iii) gn = 0 off of A1/n = {x ∈ (0, 1] : dist(x, A) � 1/n}; and (iv) gn → 1A almost everywhere (in fact, at all but two points). In particular, by the bounded convergence theorem, �gn − 1A �p → 0 for all p � 1. Choose and fix a small ε > 0. For any� A ∈ B((0, 1]) we can find A1 , A2 , . . . ∈ ∞ ∞ A such that: (i) A ⊂ ∪∞ A ; and (ii) i=1 µ(Ai )−ε � µ(A) � µ(∪i=1 Ai ). i=1 i See the proof of Carathéodory extension theorem. Also, we can find continuous functions gi such that �gi − 1Ai �p � ε2−i . Without loss of generality, the Ai ’s are disjoint. [Else we can consider, in their place, Bi ’s, where B1 = A1 , B2�= A2 \ A1 , B3 = A3 \ (A1 ∪ A2 ), etc. ∞ ∞ All are in A.] Therefore, 1Ai . Moreover, by Minkowski’s �1∞∪i=1 Ai = �i=1 inequality, �1∪∞ − i=1 gi �p � ∞ i=1 �1Ai − gi �p � ε. Therefore, i=1 Ai � � �∞ � ∞ � � � � � � � � gi � � ε + �1A − 1∪∞i=1 Ai �p = ε + m Ai \ A � 2ε, � 1A − � � i=1 i=1 p where m denotes the Lebegsue measure. Because ε is arbitrary, this has the desired effect. 4.19. I will prove this for k = 1 in order to simplify the notation. [You should try to extend this argument to general k > 1.] It suffices to prove that for 9 10 CHAPTER 4. INTEGRATION all N > 1, lim �N ε↓0 −N (4.1) |f(x + ε) − f(x)|p dx = 0. Indeed, the fact that |a + b|p � 2p (|a|p + |b|p ) for all p ∈ (1 , ∞) and a, b ∈ R implies that � |f(x + ε) − f(x)|p dx [−N,N]c � (|f(x + ε)|p + |f(x)|p ) dx � 2p [−N,N]c � � 2p+1 |f(y)|p dy uniformly for ε ∈ (0 , 1). [−N+1,N−1]c Therefore, (1.1) would imply that � �∞ p p+1 |f(x + ε) − f(x)| dx � 2 lim sup ε↓0 |f(y)|p dy [−N+1,N−1]c −∞ for all N > 1. Let N ↑ ∞ and appeal to the monotone convergence theorem to deduce the result from (1.1). Now, (1.1) is trivial when f is continuous, because �N |f(x + ε) − f(x)|p dx � 2N sup |f(x + ε) − f(x)|p . −N −N�x�N This proves the result in the case that f is continuous. In the general case we approximate f in Lp (Rk ) by a continuous �∞ one: For every η > 0 there exist a continuous function fη such that −∞ |f(x) − fη (x)|p dx � η [this follows, for instance, from Lusin’s theorem in your real analysis course]. Therefore, �∞ �∞ |f(x + ε) − f(x)|p dx � 2p+1 η + |fη (x + ε) − fη (x)|p dx. −∞ −∞ The right-most term vanishes in the limit as ε → 0. The result follows since η is arbitrary. 4.20. For all r > 0, 1 − r � e−r . Therefore, � �n x2 2 1− � e−x /2 , 2n for all n, and the left-hand side also converges to the right-hand side as n → ∞. Therefore, by the dominated convergence theorem the integral �∞ of the problem converges to −∞ exp(−x2 /2) dx. It remains to compute the integral. Here is one way: �n �n � √n � � � √ �1 n √ 1 x2 y2 1 − dx = 2 n 1 − dy = 2n x (1−x)−1/2 dx. √ 2n 2 − n 0 1/2 11 Evidently, √ n � 1/2 n x (1 − x) −1/2 0 Thus, dx � √ −n n2 � 1/2 0 (1 − x)−1/2 dx → 0. �n � √n � √ �1 n x2 1 − dx ∼ 2n x (1 − x)−1/2 dx √ 2n − n 0 √ Γ (n + 1)Γ (1/2) = 2n Γ (n + 32 ) √ nn+(1/2) e−n β ∼ 2nπ (n + 12 )n+1 e−n−(1/2) β √ → 2π. 4.30. Let f(x) = eξx ; f is convex because f �� (x) = ξ2 eξx � 0. Therefore, for all λ ∈ [0, 1] and a, b ∈ R, eξλa+ξ(1−λ)b � λeξa + (1 − λ)eξb . Choose and fix some x ∈ [−c, c], and define a = −c, b = c, and λ = (c − x)/(2c) to obtain the hint’s inequality. In particular, with probability one, � � � � c+X c−X eξX � eξc + e−ξc . 2c 2c Take expectations—and recall that EX = 0—to deduce that E[eξX ] � cosh(ξc). We claim that for all z ∈ R, cosh(z) � ez 2 /2 (4.2) . If so, then E[eξX ] � eξ c /2 as asserted. The same holds if we replace X by −X. Because of the elementary inequality e|z| � ez + e−z —valid for 2 all z ∈ R—we have E[eξ|X| ] � 2ez /2 , whence the exercise. It remains to prove (1.2). Choose and fix some z ∈ R. By Taylor expansions, 2 2 cosh(z) = 1 + ∞ � z2j z2 z4 z6 + + + ··· = , 2 4! 6! (2j)! j=0 whereas ez 2 /2 =1+ ∞ � z2j z2 z4 z6 + + + ··· = . 2 4 · 2! 8 · 3! 2j · j! j=0 Because (2j)! = j! · (j + 1) · (j + 2) · · · (2j) � j! · 2 · 2 · · · 2 = 2j j! (j � 1), (1.2) follows. The second part follows from the first, and the simple inequality, eξ|X| � eξX + e−ξX . 12 CHAPTER 4. INTEGRATION Chapter 5 Product Spaces 5.1. 1. By definition, F is closed under complementation, denumerable unions, and denumerable intersections. 2. Because ∅ is denumerable, P(∅) = 0. If A is denumerable, then Ac is uncountable. Therefore, P(Ac ) = 1 = 1 − P(A). If A1 , A2 , . . . ∈ F are disjoint, then we can divide them into two groups: Let D be the collection of all i � 1 such that Ai is denumerable; U := N \ D denotes the indices i for which Ai is uncountable. Then, P(∪∞ i=1 Ai ) = P(∪i∈D Ai ) + P(∪i∈U Ai ). Now ∪i∈D Ai is denumerable by Cantor’s axiom of denumerable choice. Therefore, it has probability zero. On the other hand, ∪i∈U Ai is uncountable and has measure one. There∞ fore, P(∪∞ i=1 Ai ) = 1 if U �= ∅ and �P(∪i=1 Ai ) = 0 if U = ∅. It remains to prove that if U �= ∅ then i∈U P(Ai ) = 1. Suppose U �= ∅. Then it suffices to prove that U has one element only. Suppose it has at least two elements i and j. Because Ai and Aj are disjoint, then Ai ⊂ Acj . But this would be a contradiction because Ai is uncountable and Acj is denumerable. 3. Let Ω be all non-negative ordinals; c the power of the continuum; S := {x ∈ Ω : x � c}; and E := {(x, y) ∈ S × S : x < y}. As in Example 5.10, define, for all x, y ∈ S, Ex := {b ∈ S : (x, b) ∈ E} and y E := {a ∈ S : (a, y) ∈ E}. Then Ex is denumerable, whereas y E is uncountable. 5.2. For simplicity consider the case n = 1. Choose and fix a positive integer k. Let Ii,k denote the half-open interval (i/k, (i + 1)/k]. Note that Ii,k × Ii,k is a disjoint cover of the diagonal D. Therefore, (µ × µ)(D) � ∞ � i=−∞ (µ × µ)(Ii,k × Ii,k ) = 13 ∞ � i=−∞ (µ(Ii,k ))2 . 14 CHAPTER 5. PRODUCT SPACES �∞ On the other hand, µ(R) = i=−∞ µ(Ii,k ). Because µ does not charge singletons, limk→∞ µ(Ii,k ) = 0 for every fixed i. Therefore, the dominated convergence theorem implies that (µ × µ)(D) = 0. 5.7. By symmetry, if the integrals each converge then �1 �1 0 �1 �1 f dx dy = − 0 0 f dy dx. 0 So it suffices to consider one of the two double-integrals, and show that it exists and is non-zero. Fix y ∈ [0, 1], and note that �1 0 f dx = 1 y2 �1 0 (x/y)2 − 1 1 dx = ((x/y)2 + 1)2 y � 1/y 0 w2 − 1 dw (w2 + 1)2 (w = x/y). But it is easy to see that � −w 1 + w2 �� = w2 − 1 . (1 + w2 )2 �1 Therefore, 0 f(x , y) dx = −(1 + y2 )−1 . Integrate once more to find that �1 �1 �1 �1 �1 �1 0 0 f dx dy = −π/4. Thus, 0 0 f dy dx = (π/4) �= −(π/4) = 0 0 f dx dy. �1 �1 �1 5.8. For all y �= 0 −1 f(x , y) dx = 0 by symmetry. Therefore, −1 −1 f dx dy = �1 �1 0. Similarly, −1 −1 f dy dx = 0. On the other hand, by Fubini–Tonelli for positive functions, �1 �1 −1 −1 |xy| dx dy = 4 (x2 + y2 )2 �2 �1 �1 y 0 �1 0 = ∞. y 0 �1 x dx dy (x2 + y2 )2 y2 dz dy z2