Solutions Mims 2e
Solutions Mims 2e
Solutions Mims 2e
(2nd edition)
Solution Manual
René L. Schilling
Acknowledgement. I am grateful for the help of Dr. Franziska Kühn in the preparation of this solution
manual. Dr. Björn Böttcher, Dr. Julian Hollender and Dr. Franziska Kühn contributed problems and
solutions to this collection. Max Ziegenbalg helped to improve the solutions.
2
Contents
1 Prologue.
Solutions to Problems 1.1–1.5 7
3 𝜎-Algebras.
Solutions to Problems 3.1–3.16 21
4 Measures.
Solutions to Problems 4.1–4.22 31
5 Uniqueness of measures.
Solutions to Problems 5.1–5.13 49
6 Existence of measures.
Solutions to Problems 6.1–6.14 59
7 Measurable mappings.
Solutions to Problems 7.1–7.13 73
8 Measurable functions.
Solutions to Problems 8.1–8.26 81
3
R.L. Schilling: Measures, Integrals & Martingales
18 Hausdorff measure.
Solutions to Problems 18.1–18.7 223
23 Martingales.
Solutions to Problems 23.1–23.16 273
25 Martingales in action.
Solutions to Problems 25.1–25.15 289
27 Conditional expectations.
Solutions to Problems 27.1–27.19 319
4
Solution Manual. Last update 28th October 2021
5
1 Prologue.
Solutions to Problems 1.1–1.5
Problem 1.1 Solution: We have to calculate the area of an isosceles triangle of side-length 𝑟, base 𝑏,
height ℎ and opening angle 𝜙 ∶= 2𝜋∕2𝑗 . From elementary geometry we know that
cos 𝜙2 = ℎ
𝑟
and sin 𝜙2 = 𝑏
2𝑟
so that
1 𝑟2
area (triangle) = ℎ𝑏 = 𝑟2 cos 𝜙2 sin 𝜙2 = sin 𝜙.
2 2
Since we have lim𝜙→0 sin 𝜙
𝜙
= 1 we find
𝑟2
area (circle) = lim 2𝑗 sin 2𝜋
𝑗→∞ 2 2𝑗
sin 2𝜋
= 𝑟 𝜋 lim 2𝜋2
2 𝑗
𝑗→∞
2𝑗
= 𝑟2 𝜋
and each interval 𝐼𝑡1 ,…,𝑡𝑖 has length 2−𝑖 . We have used this when calculating 𝓁(𝐶𝑛+1 )
1 1 1
𝓁(𝐶𝑛+1 ) = 𝓁[0, 1] − 20 × − 21 × 2 − ⋯ − 2𝑛 × 𝑛+1
3 1 3 3
(note that we have removed 2𝑛 intervals of length 3−𝑛−1 ). If we let 𝑛 → ∞, we get for all removed
intervals
(∞ ) ∞
⋃ ⋃ ∑ 1
𝓁 𝐼𝑡1 ,…,𝑡𝑖 = 2𝑖−1 × = 1.
𝑖=1 𝑡1 ,…,𝑡𝑖 ∈{0,2} 𝑖=1
3𝑖
The last line requires 𝜎-additivity. (Just in case: you will see in the next chapter that the number
of removed intervals is indeed countable).
■■
7
R.L. Schilling: Measures, Integrals & Martingales
Problem 1.3 Solution: We record the lenghts of the removed pieces in each step
This means that the modified Cantor set does have a length! Consequently it cannot be empty.
■■
Problem 1.4 Solution: In each step the total length is increased by the factor 4∕3, since we remove
the middle interval (relative length 1∕3) and replace it by two copies constituting the sides of an
equilateral triangle (relative length 2∕3). Thus,
( )𝑛 ( )𝑛
4 4 4
𝓁(𝐾𝑛 ) = × 𝓁(𝐾𝑛−1 ) = ⋯ = 𝓁(𝐾0 ) = .
3 3 3
In particular, lim𝑛→∞ 𝓁(𝐾𝑛 ) = ∞.
Again 𝜎-additivity comes in in the form of a limit (compare with Problem 1.2).
■■
Problem 1.5 Solution: In each step the total area is decreased by the factor 3∕4, since we remove the
middle triangle (relative area 1∕4). Thus,
√
( )𝑛 ( )𝑛 3
3 3 3
area(𝑆𝑛 ) = × area(𝑆𝑛−1 ) = ⋯ = area(𝑆0 ) = .
4 4 4 4
In particular, area(𝑆) = lim𝑛→∞ area(𝑆𝑛 ) = 0.
Again 𝜎-additivity comes in in the form of a limit (compare with Problem 1.2). Notice that 𝑆 is
not empty as it contains the vertices of all black triangles (see figure) of each stage.
■■
8
2 The pleasures of counting.
Solutions to Problems 2.1–2.22
(i) We have
𝑥 ∈ 𝐴 ⧵ 𝐵 ⇐⇒ 𝑥 ∈ 𝐴 and 𝑥 ∉ 𝐵
⇐⇒ 𝑥 ∈ 𝐴 and 𝑥 ∈ 𝐵 𝑐
⇐⇒ 𝑥 ∈ 𝐴 ∩ 𝐵 𝑐 .
(iii) Using (i), de Morgan’s laws (*) and the fact that (𝐶 𝑐 )𝑐 = 𝐶 gives
(i)
𝐴 ⧵ (𝐵 ⧵ 𝐶) = 𝐴 ∩ (𝐵 ∩ 𝐶 𝑐 )𝑐
(∗)
= 𝐴 ∩ (𝐵 𝑐 ∪ 𝐶)
= (𝐴 ∩ 𝐵 𝑐 ) ∪ (𝐴 ∩ 𝐶)
(i)
= (𝐴 ⧵ 𝐵) ∪ (𝐴 ∩ 𝐶).
9
R.L. Schilling: Measures, Integrals & Martingales
= 𝐴 ∩ 𝐵𝑐 ∩ 𝐴 ∩ 𝐶 𝑐
(i)
= (𝐴 ⧵ 𝐵) ∩ (𝐴 ⧵ 𝐶)
(𝐴 ∪ 𝐵) ⧵ 𝐶 = (𝐴 ∪ 𝐵) ∩ 𝐶 𝑐
= (𝐴 ∩ 𝐶 𝑐 ) ∪ (𝐵 ∩ 𝐶 𝑐 )
= (𝐴 ⧵ 𝐶) ∪ (𝐵 ⧵ 𝐶).
■■
𝐴 ⧵ 𝐶 ⊂ (𝐴 ⧵ 𝐵) ∪ (𝐵 ⧵ 𝐶). (*)
𝐴 ⧵ 𝐶 = (𝐴 ⧵ 𝐶) ∩ 𝑋
= (𝐴 ∩ 𝐶 𝑐 ) ∩ (𝐵 ∪ 𝐵 𝑐 )
= (𝐴 ∩ 𝐶 𝑐 ∩ 𝐵) ∪ (𝐴 ∩ 𝐶 𝑐 ∩ 𝐵 𝑐 )
⊂ (𝐵 ∩ 𝐶 𝑐 ) ∪ (𝐴 ∩ 𝐵 𝑐 )
= (𝐵 ⧵ 𝐶) ∪ (𝐴 ⧵ 𝐵).
(𝐴 ∪ 𝐵 ∪ 𝐶) ⧵ (𝐴 ∩ 𝐵 ∩ 𝐶)
= (𝐴 ∪ 𝐵 ∪ 𝐶) ∩ (𝐴 ∩ 𝐵 ∩ 𝐶)𝑐
= (𝐴 ▵ 𝐵) ∪ (𝐵 ▵ 𝐶)
■■
Problem 2.3 Solution: It is clearly enough to prove (2.3) as (2.2) follows if 𝐼 contains 2 points.
De Morgan’s identities state that for any index set 𝐼 (finite, countable or not countable) and any
collection of subsets 𝐴𝑖 ⊂ 𝑋, 𝑖 ∈ 𝐼, we have
( )𝑐 ( )𝑐
⋃ ⋂ ⋂ ⋃
(a) 𝐴𝑖 = 𝐴𝑐𝑖 and (b) 𝐴𝑖 = 𝐴𝑐𝑖 .
𝑖∈𝐼 𝑖∈𝐼 𝑖∈𝐼 𝑖∈𝐼
10
Solution Manual. Last update 28th October 2021
⇐⇒ ∀ 𝑖 ∈ 𝐼 ∶ 𝑎 ∉ 𝐴𝑖
⇐⇒ ∀ 𝑖 ∈ 𝐼 ∶ 𝑎 ∈ 𝐴𝑐𝑖
⋂
⇐⇒ 𝑎 ∈ 𝐴𝑐𝑖 ,
𝑖∈𝐼
⇐⇒ ∃ 𝑖0 ∈ 𝐼 ∶ 𝑎 ∉ 𝐴𝑖0
⇐⇒ ∃ 𝑖0 ∈ 𝐼 ∶ 𝑎 ∈ 𝐴𝑐𝑖
0
⋃
⇐⇒ 𝑎 ∈ 𝑐
𝐴𝑖 .
𝑖∈𝐼
■■
(i) The inclusion 𝑓 (𝐴 ∩ 𝐵) ⊂ 𝑓 (𝐴) ∩ 𝑓 (𝐵) is always true since 𝐴 ∩ 𝐵 ⊂ 𝐴 and 𝐴 ∩ 𝐵 ⊂ 𝐵 imply
that 𝑓 (𝐴 ∩ 𝐵) ⊂ 𝑓 (𝐴) and 𝑓 (𝐴 ∩ 𝐵) ⊂ 𝑓 (𝐵), respectively. Thus, 𝑓 (𝐴 ∩ 𝐵) ⊂ 𝑓 (𝐴) ∩ 𝑓 (𝐵).
Furthermore, 𝑦 ∈ 𝑓 (𝐴)⧵𝑓 (𝐵) means that there is some 𝑥 ∈ 𝐴 but 𝑥 ∉ 𝐵 such that 𝑦 = 𝑓 (𝑥),
that is: 𝑦 ∈ 𝑓 (𝐴 ⧵ 𝐵). Thus, 𝑓 (𝐴) ⧵ 𝑓 (𝐵) ⊂ 𝑓 (𝐴 ⧵ 𝐵).
To see that the converse inclusions cannot hold we consider some non injective 𝑓 . Take
𝑋 = [0, 2], 𝐴 = (0, 1), 𝐵 = (1, 2), and 𝑓 ∶ [0, 2] → R with 𝑥 → 𝑓 (𝑥) = 𝑐 (𝑐 is some
constant). Then 𝑓 is not injective and
Observe that
11
R.L. Schilling: Measures, Integrals & Martingales
and
and, finally,
𝑥 ∈ 𝑓 −1 (𝐶 ⧵ 𝐷) ⇐⇒ 𝑓 (𝑥) ∈ 𝐶 ⧵ 𝐷
⇐⇒ 𝑓 (𝑥) ∈ 𝐶 and 𝑓 (𝑥) ∉ 𝐷
⇐⇒ 𝑥 ∈ 𝑓 −1 (𝐶) and 𝑥 ∉ 𝑓 −1 (𝐷)
⇐⇒ 𝑥 ∈ 𝑓 −1 (𝐶) ⧵ 𝑓 −1 (𝐷).
■■
1𝐴∩𝐵 (𝑥) = 1 ⇐⇒ 𝑥 ∈ 𝐴 ∩ 𝐵
⇐⇒ 𝑥 ∈ 𝐴, 𝑥 ∈ 𝐵
⇐⇒ 1𝐴 (𝑥) = 1 = 1𝐵 (𝑥)
⎧
⎪1𝐴 (𝑥) ⋅ 1𝐵 (𝑥) = 1
⇐⇒ ⎨
⎪min{1𝐴 (𝑥), 1𝐵 (𝑥)} = 1
⎩
1𝐴∪𝐵 (𝑥) = 1 ⇐⇒ 𝑥 ∈ 𝐴 ∪ 𝐵
⇐⇒ 𝑥 ∈ 𝐴 or 𝑥 ∈ 𝐵
⇐⇒ 1𝐴 (𝑥) + 1𝐵 (𝑥) ⩾ 1
⎧
⎪min{1𝐴 (𝑥) + 1𝐵 (𝑥), 1} = 1
⇐⇒ ⎨
⎪max{1𝐴 (𝑥), 1𝐵 (𝑥)} = 1
⎩
(iii) Since 𝐴 = (𝐴 ∩ 𝐵) ⊍ (𝐴 ⧵ 𝐵) we see that 1𝐴∩𝐵 (𝑥) + 1𝐴⧵𝐵 (𝑥) can never have the value 2, thus
part (ii) implies
and all we have to do is to subtract 1𝐴∩𝐵 (𝑥) on both sides of the equation.
12
Solution Manual. Last update 28th October 2021
(iv) With the same argument that we use in (iii) and with the result of (iii) we get
(vii) We have
Thus,
(viii) One possibility is to mimic the proof of (vii). We prefer to argue like this: using (iii) and de
Morgan’s identities we get
■■
and this expression is 1 if, and only if, 𝑥 is either in 𝐴 or 𝐵 but not in both sets. Thus
1𝐴 ▵ 𝐵 = |1𝐴 − 1𝐵 |.
13
R.L. Schilling: Measures, Integrals & Martingales
1𝐴 ▵(𝐵 ▵ 𝐶) = 1𝐴 + 1𝐵 ▵ 𝐶 − 21𝐴 1𝐵 ▵ 𝐶
( )
= 1𝐴 + 1𝐵 + 1𝐶 − 21𝐵 1𝐶 − 21𝐴 1𝐵 + 1𝐶 − 21𝐵 1𝐶
= 1𝐴 + 1𝐵 + 1𝐶 − 21𝐵 1𝐶 − 21𝐴 1𝐵 − 21𝐴 1𝐶 + 41𝐴 1𝐵 1𝐶
1𝐴 ▵(𝐵 ▵ 𝐶) = 1(𝐴 ▵ 𝐵) ▵ 𝐶 .
𝐴 ∩ (𝐵 ▵ 𝐶) = (𝐴 ∩ 𝐵) ▵(𝐴 ∩ 𝐶).
For this we use again indicator functions and the rules from (i) and Problem 2.5:
14
Solution Manual. Last update 28th October 2021
■■
𝑓 surjective ⇐⇒ ∀ 𝐵 ⊂ 𝑌 ∶ 𝑓 ◦𝑓 −1 (𝐵) = 𝐵
⇐⇒ ∀ 𝐵 ⊂ 𝑌 ∶ 𝑓 ◦𝑓 −1 (𝐵) ⊃ 𝐵.
and we have equality in the last step if, and only if, we can guarantee that every 𝑦 ∈ 𝐵 is of the form
𝑦 = 𝑓 (𝑥) for some 𝑥. Since this must hold for all sets 𝐵, this amounts to saying that 𝑓 (𝑋) = 𝑌 , i.e.
that 𝑓 is surjective. The second equivalence is clear since our argument shows that the inclusion
‘⊂’ always holds.
𝑓 −1 ([−1, 1]) = [0, 1] and 𝑓 ◦𝑓 −1 ([−1, 1]) = 𝑓 ([0, 1]) = [0, 1] ⊊ [−1, 1].
𝑓 injective ⇐⇒ ∀ 𝐴 ⊂ 𝑋 ∶ 𝑓 −1 ◦𝑓 (𝐴) = 𝐴
⇐⇒ ∀ 𝐴 ⊂ 𝑋 ∶ 𝑓 −1 ◦𝑓 (𝐴) ⊂ 𝐴.
since 𝑥 ∈ 𝐴 always entails 𝑓 (𝑥) ∈ 𝑓 (𝐴). The reverse is, for non-injective 𝑓 , wrong since then
there might be some 𝑥0 ∉ 𝐴 but with 𝑓 (𝑥0 ) = 𝑓 (𝑥) ∈ 𝑓 (𝐴) i.e. 𝑥0 ∈ 𝑓 −1 ◦𝑓 (𝐴) ⧵ 𝐴. This means
that we have equality in (∗) if, and only if, 𝑓 is injective. The second equivalence is clear since
our argument shows that the inclusion ‘⊃’ always holds.
■■
Problem 2.8 Solution: Assume that for 𝑥, 𝑦 we have 𝑓 ◦𝑔(𝑥) = 𝑓 ◦𝑔(𝑦). Since 𝑓 is injective, we
conclude that
15
R.L. Schilling: Measures, Integrals & Martingales
𝑔(𝑥) = 𝑔(𝑦) ⇐⇒ 𝑥 = 𝑦
■■
• Call the set of odd numbers 𝒪. Every odd number is of the form 2𝑘 − 1 where 𝑘 ∈ N. We
are done, if we can show that the map 𝑓 ∶ N → 𝒪, 𝑘 → 2𝑘 − 1 is bijective. Surjectivity is
clear as 𝑓 (N) = 𝒪. For injectivity we take 𝑖, 𝑗 ∈ N such that 𝑓 (𝑖) = 𝑓 (𝑗). The latter means
that 2𝑖 − 1 = 2𝑗 − 1, so 𝑖 = 𝑗, i.e. injectivity.
An alternative approach would be to write out Z × N (the swap of Z and N is for notational
reasons—since the map 𝛽((𝑗, 𝑘)) ∶= (𝑘, 𝑗) from Z × N to N × Z is bijective, the cardinality
does not change) in the following form
and going through the array, starting with (0, 1), then (1, 1) → (1, 2) → (0, 2) → (−1, 2) →
⨆
(−1, 1), then (2, 1) → (2, 2) → (2, 3) → (1, 3) → ... in clockwise oriented -shapes down,
left, up.
The proof of #(N × N) = #N can be easily adapted—using some pretty obvious notational
changes—to show that the Cartesian product of any two countable sets of cardinality #N has
again cardinality #N. Applying this 𝑚 − 1 times we see that #Q𝑛 = #N.
⋃
• 𝑚∈N Q𝑚 is a countable union of countable sets, hence countable, cf. Theorem 2.6.
16
Solution Manual. Last update 28th October 2021
■■
Problem 2.10 Solution: Following the hint it is clear that 𝛽 ∶ N → N × {1}, 𝑖 → (𝑖, 1) is a bijection
and that 𝚥 ∶ N × {1} → N × N, (𝑖, 1) → (𝑖, 1) is an injection. Thus, #N ⩽ #(N × N).
⋃
On the other hand, N × N = 𝑗∈N N × {𝑗} which is a countable union of countable sets, thus
#(N × N) ⩽ #N.
■■
■■
Problem 2.12 Solution: Assume that the set {0, 1}N were indeed countable and that {𝑠𝑗 }𝑗∈N was an
enumeration: each 𝑠𝑗 would be a sequence of the form (𝑑1𝑗 , 𝑑2𝑗 , 𝑑3𝑗 , ..., 𝑑𝑘𝑗 , ...) with 𝑑𝑘𝑗 ∈ {0, 1}. We
could write these sequences in an infinite list of the form:
⎧
⎪0, if 𝑑𝑚𝑚 = 1
𝑒𝑚 ∶= ⎨ .
⎪1, if 𝑑𝑚𝑚 = 0
⎩
Since 𝑆 differs from 𝑠𝓁 exactly at position 𝓁, 𝑆 cannot be in the above list, thus, the above list did
not contain all 0-1-sequences, hence a contradiction.
■■
1 1
𝑓 (𝑥) ∶= − .
1−𝑥 𝑥
This function is obviously continuous and we have lim𝑥→0 𝑓 (𝑥) = −∞ and lim𝑥→1 𝑓 (𝑥) = +∞.
By the intermediate value theorem we have therefore 𝑓 ((0, 1)) = R, i.e. surjectivity.
1 1
Since 𝑓 is also differentiable and 𝑓 ′ (𝑥) = + 2 > 0, we see that 𝑓 is strictly increasing,
(1 − 𝑥) 2 𝑥
hence injective, hence bijective.
■■
17
R.L. Schilling: Measures, Integrals & Martingales
⋃ ⋃
Problem 2.14 Solution: Since 𝐴1 ⊂ 𝑖∈N 𝐴𝑖 it is clear that 𝔠 = #𝐴1 ⩽ # 𝑖∈N 𝐴𝑖 . On the other
hand, #𝐴𝑖 = 𝔠 means that we can map 𝐴𝑖 bijectively onto R and, using Problem 2.13, we map R
⋃ ⋃
bijectively onto (0, 1) or (𝑖 − 1, 𝑖). This shows that # 𝑖∈N 𝐴𝑖 ⩽ # 𝑖∈N (𝑖 − 1, 𝑖) ⩽ #R = 𝔠. Using
Theorem 2.7 finishes the proof.
■■
Problem 2.15 Solution: Since we can write each 𝑥 ∈ (0, 1) as an infinite dyadic fraction (o.k. if it
is finite, fill it up with an infinite tail of zeroes !), the proof of Theorem 2.8 shows that #(0, 1) ⩽
#{0, 1}N .
On the other hand, thinking in base-4 expansions, each element of {1, 2}N can be interpreted
as a unique base-4 fraction (having no 0 or 3 in its expansion) of some number in (0, 1). Thus,
#{1, 2}N ⩽ #N.
But #{1, 2}N = #{0, 1}N and we conclude with Theorem 2.7 that #(0, 1) = #{0, 1}N .
■■
Problem 2.16 Solution: Just as before, expand 𝑥 ∈ (0, 1) as an 𝑛-adic fraction, then interpret each
element of {1, 2, … , 𝑛 + 1}N as a unique (𝑛 + 1)-adic expansion of a number in (0, 1) and observe
that #{1, 2, … , 𝑛 + 1}N = {0, 1, … , 𝑛}N .
■■
Problem 2.17 Solution: Take a vector (𝑥, 𝑦) ∈ (0, 1) × (0, 1) and expand its coordinate entries 𝑥, 𝑦
as dyadic numbers:
𝑥 = 0.𝑥1 𝑥2 𝑥3 … , 𝑦 = 0.𝑦1 𝑦2 𝑦3 … .
Then 𝑧 ∶= 0.𝑥1 𝑦1 𝑥2 𝑦2 𝑥3 𝑦3 … is a number in (0, 1). Conversely, we can ‘zip’ each 𝑧 = 0.𝑧1 𝑧2 𝑧3 𝑧4 … ∈
(0, 1) into two numbers 𝑥, 𝑦 ∈ (0, 1) by setting
𝑥 ∶= 0.𝑧2 𝑧4 𝑧6 𝑧8 … , 𝑦 ∶= 0.𝑧1 𝑧3 𝑧5 𝑧7 …
Since we have a bijection between (0, 1) ↔ R it is clear that we have also a bijection between
(0, 1) × (0, 1) ↔ R × R.
■■
Problem 2.18 Solution: We have seen in Problem 2.18 that #{0, 1}N = #{1, 2}N = 𝔠. Obviously,
{1, 2}N ⊂ NN ⊂ RN and since we have a bijection between (0, 1) ↔ R one extends this (using
coordinates) to a bijection between (0, 1)N ↔ RN . Using Theorem 2.9 we get
18
Solution Manual. Last update 28th October 2021
Problem 2.19 Solution: Let 𝐹 ∈ ℱ with #𝐹 = 𝑛 Then we can write 𝐹 as a tuple of length 𝑛 (having
⋃
𝑛 pairwise different entries...) and therefore we can interpret 𝐹 as an element of 𝑚∈N N𝑚 . In
⋃ ⋃
this sense, ℱ → 𝑚∈N N𝑚 and #ℱ ⩽ 𝑚∈N N𝑚 = #N since countably many countable sets are
again countable. Since N ⊂ ℱ we get #ℱ = #N by Theorem 2.7.
𝑎𝑀 −1
2 𝑎𝑀 − 1 ∑
𝜙(𝐴) ⩾ 𝜙({𝑎𝑀 }) ⩾ 2 𝑎𝑀
> = 2𝑗
2−1 𝑗=1
= 𝜙({1, 2, 3, … 𝑎𝑀 − 1})
⩾ 𝜙(𝐵),
which cannot be the case since we assumed 𝜙(𝐴) = 𝜙(𝐵). Thus, 𝑎𝑀 = 𝑏𝑁 . Now consider
recursively the next elements, 𝑎𝑀−1 and 𝑏𝑁−1 and the same conclusion yields their equality etc.
The process stops after min{𝑀, 𝑁} steps. But if 𝑀 ≠ 𝑁, say 𝑀 > 𝑁, then 𝐴 would contain at
least one more element than 𝐵, hence 𝜙(𝐴) > 𝜙(𝐵), which is also a contradiction. This, finally
shows that 𝐴 = 𝐵, hence that 𝜙 is injective.
On the other hand, each natural number can be expressed in terms of finite sums of powers of
base-2, so that 𝜙 is also surjective.
Thus, #ℱ = #N.
■■
Problem 2.20 Solution: (Let ℱ be as in the previous exercise.) Observe that the infinite sets from
𝒫 (N), ℐ ∶= 𝒫 (N) ⧵ ℱ can be surjectively mapped onto {0, 1}N : if {𝑎1 , 𝑎2 , 𝑎3 , …} = 𝐴 ⊂ N,
then define an infinite 0-1-sequence (𝑏1 , 𝑏2 , 𝑏3 , …) by setting 𝑏𝑗 = 0 or 𝑏𝑗 = 1 according to whether
𝑎𝑗 is even or odd. This is a surjection of 𝒫 (N) onto {0, 1}N and so #𝒫 (N) ⩾ #{0, 1}N . Call this
map 𝛾 and consider the family 𝛾 −1 (𝑠), 𝑠 ∈ {0, 1}N in ℐ , consisting of obviously disjoint infinite
subsets of N which lead to the same 0-1-sequence 𝑠. Now choose from each family 𝛾 −1 (𝑠) a
representative, call it 𝑟(𝑠) ∈ ℐ . Then the map 𝑠 → 𝑟(𝑠) is a bijection between {0, 1}N and a subset
of ℐ , the set of all representatives. Hence, ℐ has at least the same cardinality as {0, 1}N and as
such a bigger cardinality than N.
■■
19
R.L. Schilling: Measures, Integrals & Martingales
Problem 2.21 Solution: Denote by Θ the map 𝒫 (N) ∋ 𝐴 → 1𝐴 ∈ {0, 1}N . Let 𝛿 = (𝑑1 , 𝑑2 , 𝑑3 , …) ∈
{0, 1}N and define 𝐴(𝛿) ∶= {𝑗 ∈ N ∶ 𝑑𝑗 = 1}. Then 𝛿 = (1𝐴(𝛿) (𝑗))𝑗∈N showing that Θ is surject-
ive.
1𝐴 = 1𝐵 ⇐⇒ 1𝐴 (𝑗) = 1𝐵 (𝑗) ∀𝑗 ∈ N ⇐⇒ 𝐴 = 𝐵.
■■
(𝐴 ∩ 𝐵) ∪ (𝐴′ ∩ 𝐵 ′ ) = (𝐴 ∪ 𝐴′ ) ∩ (𝐴 ∪ 𝐵 ′ ) ∩ (𝐵 ∪ 𝐴′ ) ∩ (𝐵 ∪ 𝐵 ′ )
and since this rule carries over to the infinite case, we get the formula from the problem by ‘mul-
tiplying out’ the countable union
while
⋂ ⋃
𝑥∈ 𝐴𝑖(𝑘)
𝑘
𝑖=(𝑖(𝑘))𝑘∈N ∈{0,1}N 𝑘∈N
⋃
⇐⇒ ∀𝑖 = (𝑖(𝑘))𝑘∈N ∈ {0, 1}N ∶ 𝑥 ∈ 𝐴𝑖(𝑘)
𝑘
𝑘∈N
(**)
𝑖(𝑘0 )
⇐⇒ ∀𝑖 = (𝑖(𝑘))𝑘∈N ∈ {0, 1} ∃ 𝑘0 ∈ N ∶ 𝑥 ∈ 𝐴𝑘
N
0
Clearly, (*) implies (**). On the other hand, assume that (**) holds but that (*) is wrong, i.e.
suppose that for every 𝑛 we have that either 𝑥 ∈ 𝐴0𝑛 or 𝑥 ∈ 𝐴1𝑛 or 𝑥 is in neither of 𝐴0𝑛 , 𝐴1𝑛 . Thus
we can construct a uniquely defined sequence 𝑖(𝑛) ∈ {0, 1}, 𝑛 ∈ N, by setting
⎧
⎪0 if 𝑥 ∈ 𝐴0𝑛 ;
⎪
𝑖(𝑛) = ⎨1 if 𝑥 ∈ 𝐴1𝑛 ;
⎪
⎪0 if 𝑥 ∉ 𝐴0𝑛 and 𝑥 ∉ 𝐴1𝑛 .
⎩
Define by 𝑖′ (𝑛) ∶= 1 − 𝑖(𝑛) the ‘complementary’ 0-1-sequence. Then
⋃ ⋃
but 𝑥 ∉
′
𝑥∈ 𝐴𝑖(𝑛)
𝑛 𝐴𝑛𝑖 (𝑛)
𝑛 𝑛
■■
20
3 𝜎-Algebras.
Solutions to Problems 3.1–3.16
𝐴1 ∩ … ∩ 𝐴𝑁 ∩𝐴𝑁+1 = 𝐵 ∩ 𝐴𝑁+1 ∈ 𝒜 .
⏟⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏟
=∶𝐵∈𝒜
Alternative: Of course, the last argument also goes through for 𝑁 sets:
■■
(iv) Let us assume that 𝐵 ≠ ∅ and 𝐵 ≠ 𝑋. Then 𝐵 𝑐 ∉ {∅, 𝐵, 𝑋}. Since with 𝐵 also 𝐵 𝑐 must be
contained in a 𝜎-algebra, the family {∅, 𝐵, 𝑋} cannot be one.
(vi) Set 𝒜𝐸 ∶= {𝐸 ∩ 𝐴 ∶ 𝐴 ∈ 𝒜 }. The key observation is that all set operations in 𝒜𝐸 are now
relative to 𝐸 and not to 𝑋. This concerns mainly the complementation of sets! Let us check
(Σ1 )–(Σ3 ).
21
R.L. Schilling: Measures, Integrals & Martingales
(vii) Note that 𝑓 −1 interchanges with all set operations. Let 𝐴, 𝐴𝑗 , 𝑗 ∈ N be sets in 𝒜 . We know
that then 𝐴 = 𝑓 −1 (𝐴′ ), 𝐴𝑗 = 𝑓 −1 (𝐴′𝑗 ) for suitable 𝐴, 𝐴′𝑗 ∈ 𝒜 ′ . Since 𝒜 ′ is, by assumption
a 𝜎-algebra, we have
∅ = 𝑓 −1 (∅) ∈ 𝒜 as ∅ ∈ 𝒜 ′
( )𝑐
𝐴𝑐 = 𝑓 −1 (𝐴′ ) = 𝑓 −1 (𝐴′ ) ∈ 𝒜 as 𝐴′ ∈ 𝒜 ′
𝑐 𝑐
( )
⋃ ⋃ ⋃ ⋃
𝐴𝑗 = 𝑓 −1 (𝐴′𝑗 ) = 𝑓 −1 𝐴′𝑗 ∈ 𝒜 as 𝐴′𝑗 ∈ 𝒜 ′
𝑗∈N 𝑗∈N 𝑗∈N 𝑗∈N
Problem 3.3 Solution: Denote by Σ = 𝜎({𝑥}, 𝑥 ∈ R). Let 𝒜 be the 𝜎-algebra defined in Ex-
ample 3.3(v). It is clear that {𝑥} ∈ 𝒜 , and so Σ ⊂ 𝒜 . On the other hand, if 𝐴 ∈ 𝒜 , then either 𝐴
or 𝐴𝑐 is countable. Wlog assume that 𝐴 is countable. Then 𝐴 is a countable union of singletons,
as such 𝐴 ∈ Σ as well as 𝐴𝑐 ∈ Σ. This means 𝒜 ⊂ Σ.
■■
22
Solution Manual. Last update 28th October 2021
All in all one should have 16 sets (some of them could be empty or 𝑋 or appear several times,
depending on how much 𝐴 differs from 𝐵). That’s it, but the trouble is: is this construction correct?
Here is a somewhat more systematic procedure:
Since 𝒜 is stable under intersections, it is also clear that all atoms are disjoint sets! Now we can
make up every set from 𝒜 as union (finite or countable) of such atoms. The task at hand is to
find atoms if 𝐴, 𝐵 are given. This is easy: the atoms of our future 𝜎-algebra must be: 𝐴 ⧵ 𝐵,
𝐵 ⧵ 𝐴, 𝐴 ∩ 𝐵, (𝐴 ∪ 𝐵)𝑐 . (Test it: if you make a picture, this is a tesselation of our space 𝑋 using
disjoint sets and we can get back 𝐴, 𝐵 as union! It is also minimal, since these sets must appear in
𝜎({𝐴, 𝐵}) anyway.) The crucial point is now:
Proof. The question is how many different unions we can make out of 𝑁 sets. Simple answer:
( )
we find 𝑁𝑗 , 0 ⩽ 𝑗 ⩽ 𝑁 different unions involving exactly 𝑗 sets (𝑗 = 0 will, of course, produce
∑ (𝑁 )
the empty set) and they are all different as the atoms were disjoint. Thus, we get 𝑁 𝑗=0 𝑗 =
(1 + 1)𝑁 = 2𝑁 different sets.
It is clear that they constitute a 𝜎-algebra.
This answers the above question. The number of atoms depends obviously on the relative position
of 𝐴, 𝐵: do they intersect, are they disjoint etc. Have fun with the exercises and do not try to find
𝜎-algebras generated by three or more sets..... (By the way: can you think of a situation in [0, 1]
with two subsets given and exactly four atoms? Can there be more?)
■■
23
R.L. Schilling: Measures, Integrals & Martingales
■■
Problem 3.7 Solution: We follow the hint. Since #𝒜 = #N, the following set is a countable inter-
section of measurable sets, hence itself in 𝒜 :
⋂
∀𝑥 ∈ 𝑋 ∶ 𝐴(𝑥) ∶= 𝐴 ∈ 𝒜. (*)
𝐴∈𝒜 ,𝐴∋𝑥
Remark: A 𝜎-algebra may have no non-empty atoms at all! Here is an example (which I learned
from Julian Hollender). Let 𝐼 be an uncountable set, e.g. 𝐼 = [0, 1], and consider Ω = {0, 1}𝐼 . We
can construct a 𝜎-algebra on Ω in the following way: Let 𝐾 ⊂ 𝐼 and define 𝑃𝐾 ∶ {0, 1}𝐼 → {0, 1}𝐾
the coordinate projection. A cylinder set or finitely based set with basis 𝐾 ⊂ 𝐼 is a set of the form
𝑃𝐾−1 (𝐵) where #𝐾 < ∞ and 𝐵 ⊂ {0, 1}𝐾 . Now consider the 𝜎-algebra 𝒜 ∶= 𝜎({cylinder sets})
on {0, 1}𝐼 . Intuitively, 𝐴 ∈ 𝒜 is of the form 𝑃𝐿−1 (𝐵) where 𝐿 is countable. (The proof as such
is not obvious, a possible source is Lemma 4.5 in Schilling & Partzsch: Brownian Motion. De
Gruyter, Berlin 2012.) Assume that 𝐴0 ∈ 𝒜 were an atom. Then 𝐴0 has the basis 𝐿. Take
𝑖 ∈ 𝐼 ⧵ 𝐿, consider 𝐿′ = 𝐿 ∪ {𝑖} and construct a set 𝑃𝐿−1′ (𝐵 ′ ) where 𝐵 ′ = 𝐵 × {0}, say. Then
𝑃𝐿−1′ (𝐵 ′ ) ⊂ 𝐴0 and 𝑃𝐿−1′ (𝐵 ′ ) ∈ 𝒜 .
■■
Problem 3.8 Solution: We begin with an example: Let 𝑋 = (0, 1] and 𝒜 = ℬ(0, 1] be the Borel
sets. Define
( )
𝒜𝑛 ∶= 𝜎 ((𝑗 − 1)2−𝑛 , 𝑗2−𝑛 ] , 𝑗 = 1, 2, … , 2𝑛
24
Solution Manual. Last update 28th October 2021
On the other hand, every open set (𝑎, 𝑏) ⊂ [0, 1] is a countable union of sets from 𝒜∞ :
⋃
(𝑎, 𝑏) = 𝐼
𝐼∈𝒜∞ ,𝐼⊂(𝑎,𝑏)
which follows from the fact that the dyadic numbers are dense in (0, 1]. (If you want it more
elementary, then approximate 𝑎 and 𝑏 from the right and left, respectively, by dyadic numbers
and construct the approximating intervals by hand....). If, for example, 𝑎 and 𝑏 are irrational, then
(𝑎, 𝑏) ∉ 𝒜∞ . This shows that 𝒜∞ cannot be a 𝜎-algebra.
Argument 2: Since #𝒜𝑛 = 2𝑛 we see that #𝒜∞ = #N. But Problem 3.7 tells us that 𝒜∞ can’t be a
𝜎-algebra.
A. Broughton and B.W. Huff: A comment on unions of sigma-fields. Am. Math. Monthly 84 (1977)
553–554.
Since the 𝒜𝑛 are strictly increasing, we may assume that 𝒜1 ≠ {∅, 𝑋}. Recall also the notion of a
trace 𝜎-Algebra
𝐵 ∩ 𝒜𝑛 ∶= {𝐵 ∩ 𝐴 ∶ 𝐴 ∈ 𝒜𝑛 }.
Step 1. Claim: There exists a set 𝐸 ∈ 𝒜1 such that (𝐸 ∩ 𝒜𝑛+1 ) ⧵ (𝐸 ∩ 𝒜𝑛 ) ≠ ∅ for infinitely many
𝑛 ∈ N.
To see this, assume – to the contrary – that for some 𝑛 and some 𝐵 ∈ 𝒜1 we have
If 𝑈 ∈ 𝒜𝑛+1 ⧵ 𝒜𝑛 , then
𝑈= (𝐵 ∩ 𝑈 ) ∪ (𝐵 𝑐 ∩ 𝑈 )
⏟⏟⏟ ⏟⏞⏟⏞⏟
∈𝐵∩𝒜𝑛+1 =𝐵∩𝒜𝑛 ⊂𝒜𝑛 ∈𝐵 𝑐 ∩𝒜𝑛+1 =𝐵 𝑐 ∩𝒜𝑛 ⊂𝒜𝑛
Step 2. Let 𝐸 be the set from Step 1 and denote by 𝑛1 , 𝑛2 , … a sequence for which the assertion
in Step 1 holds. Then
ℱ𝑘 ∶= 𝐸 ∩ 𝒜𝑛𝑘 , 𝑘∈N
25
R.L. Schilling: Measures, Integrals & Martingales
is a strictly increasing sequence of 𝜎-Algebras over the set 𝐸. Again we may assume that ℱ1 ≠
{∅, 𝐸} As in Step 1, we find some 𝐸1 ∈ ℱ1 such that 𝐸1 is not trivial (i.e. 𝐸1 ≠ ∅ and 𝐸1 ≠ 𝐸)
and (𝐸1 ∩ ℱ𝑘+1 ) ⧵ (𝐸1 ∩ ℱ𝑘 ) ≠ ∅ holds for infinitely many 𝑘.
Step 3. Now we repeat Step 2 and construct recursively a sequence of 𝜎-algebras 𝒜𝑖1 ⊂ 𝒜𝑖2 ⊂
𝒜𝑖3 … and a sequence of sets 𝐸1 ⊃ 𝐸2 ⊃ 𝐸3 … such that
Step 4. The sets 𝐹𝑘 ∶= 𝐸𝑘 ⧵ 𝐸𝑘+1 have the property that they are disjoint and 𝐹𝑘 ∈ 𝒜𝑖𝑘+1 ⧵ 𝒜𝑖𝑘 .
Since the 𝜎-algebras are increasing, we have
⋃ ⋃
𝒜𝑛 = 𝒜𝑖𝑘
𝑛∈N 𝑘∈N
which means that we can restrict ourselves to a subsequence. This means that we can assume that
𝑖𝑘 = 𝑘.
Step 5. Without loss of generality we can identify 𝐹𝑘 with {𝑘} and assume that the 𝒜𝑛 are 𝜎-
algebras on N such that {𝑘} ∈ 𝒜𝑘+1 ⧵ 𝒜𝑘 . Let 𝐵𝑛 the smallest set in 𝒜𝑛 such that 𝑛 ∈ 𝐵𝑛 . Then
𝑛 ∈ 𝐵𝑛 ⊂ {𝑛, 𝑛 + 1, 𝑛 + 2, …} and 𝐵𝑛 ≠ {𝑛}. Moreover
𝑚 ∈ 𝐵𝑛 ⇐⇒ 𝐵𝑚 ⊂ 𝐵𝑛 since 𝑚 ∈ 𝐵𝑛 ∩ 𝐵𝑚 ∈ 𝒜𝑚 .
Now define 𝑛1 = 1 and pick 𝑛𝑘+1 recursively: 𝑛𝑘+1 ∈ 𝐵𝑛𝑘 such that 𝑛𝑘+1 ≠ 𝑛𝑘 . Then 𝐵𝑛1 ⊃ 𝐵𝑛2 ⊃
… . Set 𝐸 = {𝑛2 , 𝑛4 , 𝑛6 , …}. If 𝒜∞ were a 𝜎-algebra, then 𝐸 ∈ 𝒜𝑛 for some 𝑛, thus 𝐸 ∈ 𝒜𝑛2𝑘 for
some 𝑘. Then {𝑛2𝑘 , 𝑛2𝑘+2 , …} ∈ 𝒜𝑛2𝑘 and thus 𝐵𝑛2𝑘 ⊂ {𝑛2𝑘 , 𝑛2𝑘+2 , …}. This contradicts the fact
𝑛2𝑘+1 ∈ 𝐵𝑛2𝑘 .
■■
i.e. 𝑈 ∩ 𝑉 ∈ 𝒪. For finitely many, say 𝑁, sets, the same argument works. Notice that already
for countably many sets we will get a problem as the radius ℎ ∶= min{𝛿𝑗 ∶ 𝑗 ∈ N} is not
necessarily any longer > 0.
26
Solution Manual. Last update 28th October 2021
𝒪2 Let 𝐼 be any (finite, countable, not countable) index set and (𝑈𝑖 )𝑖∈𝐼 ⊂ 𝒪 be a family of open
⋃
sets. Set 𝑈 ∶= 𝑖∈𝐼 𝑈𝑖 . For 𝑥 ∈ 𝑈 we find some 𝑗 ∈ 𝐼 with 𝑥 ∈ 𝑈𝑗 , and since 𝑈𝑗 was open,
⋃
we find some 𝛿𝑗 > 0 such that 𝐵𝛿𝑗 (𝑥) ⊂ 𝑈𝑗 . But then, trivially, 𝐵𝛿𝑗 (𝑥) ⊂ 𝑈𝑗 ⊂ 𝑖∈𝐼 𝑈𝑖 = 𝑈
proving that 𝑈 is open.
The family 𝒪 𝑛 cannot be a 𝜎-algebra since the complement of an open set 𝑈 ≠ ∅, ≠ R𝑛 is closed.
■■
⋂
Problem 3.10 Solution: Let 𝑋 = R and set 𝑈𝑘 ∶= (− 𝑘1 , 𝑘1 ) which is an open set. Then 𝑘∈N 𝑈𝑘 =
{0} but a singleton like {0} is closed and not open.
■■
Problem 3.11 Solution: We know already that the Borel sets ℬ = ℬ(R) are generated by any of the
following systems:
Here is just an example (with the dense set 𝐷 = Q) how to solve the problem. Let 𝑏 > 𝑎. Since
(−∞, 𝑏) ⧵ (−∞, 𝑎) = [𝑎, 𝑏) we get that
■■
Problem 3.12 Solution: Let B ∶= {𝐵𝑟 (𝑥) ∶ 𝑥 ∈ R𝑛 , 𝑟 > 0} and let B′ ∶= {𝐵𝑟 (𝑥) ∶ 𝑥 ∈ Q𝑛 , 𝑟 ∈
Q+ }. Clearly,
B′ ⊂ B ⊂ 𝒪 𝑛
⇐⇒ 𝜎(B′ ) ⊂ 𝜎(B) ⊂ 𝜎(𝒪 𝑛 ) = ℬ(R𝑛 ).
27
R.L. Schilling: Measures, Integrals & Martingales
⋃
Indeed, 𝑈 ⊃ 𝐵∈B′ , 𝐵⊂𝑈 𝐵 follows by the very definition of the union. Conversely, if 𝑥 ∈ 𝑈 we
use the fact that 𝑈 is open, i.e. there is some 𝐵𝜖 (𝑥) ⊂ 𝑈 . Without loss of generality we can assume
that 𝜖 is rational, otherwise we replace it by some smaller rational 𝜖. Since Q𝑛 is dense in R𝑛 we
can find some 𝑞 ∈ Q𝑛 with |𝑥 − 𝑞| < 𝜖∕3 and it is clear that 𝐵𝜖∕3 (𝑞) ⊂ 𝐵𝜖 (𝑥) ⊂ 𝑈 . This shows
⋃
that 𝑈 ⊂ 𝐵∈B′ , 𝐵⊂𝑈 𝐵.
Since #B′ = #(Q𝑛 × Q) = #N, formula (∗) entails that
Σ ∶= {𝐵 ⊂ 𝑋 ∶ 𝐴 ∩ 𝐵 ∈ 𝜎(𝐴 ∩ 𝒪)}.
It is not hard to see that Σ is a 𝜎-algebra and that 𝒪 ⊂ Σ. Thus ℬ(𝑋) = 𝜎(𝒪) ⊂ Σ which
means that
Notice that this argument does not really need that 𝐴 ∈ ℬ(𝑋). If, however, 𝐴 ∈ ℬ(𝑋) we
have in addition to 𝐴 ∩ ℬ(𝑋) = ℬ(𝐴) that
ℬ(𝐴) = {𝐵 ⊂ 𝐴 ∶ 𝐵 ∈ ℬ(𝑋)}
■■
28
Solution Manual. Last update 28th October 2021
is itself a MC. Note, that the intersection is non-void as the power set 𝒫 (𝑋) is (trivially)
a MC which contains ℱ . By construction, see also the argument of Theorem 3.4, 𝔪(ℱ )
is a minimal MC containing ℱ .
(ii) Define
𝒟 ∶= {𝐹 ∈ 𝔪(ℱ ) ∶ 𝐹 𝑐 ∈ 𝔪(ℱ )}.
(iii) We follow the hint. Because of the ∩-stability of ℱ we get ℱ ⊂ Σ. Let us check that Σ
is a MC:
(Σ3 ) First we show that 𝔪(ℱ ) is ∪-stable: since 𝔪(ℱ ) is ∩-stable – by (iii) – we get
𝐶, 𝐷 ∈ 𝔪(ℱ ) ⇐⇒ 𝐶 ⧵ 𝐷 = 𝐶 ∩ 𝐷𝑐 ∈ 𝔪(ℱ )
and so
(Σ2 ) [ ]
𝐶, 𝐷 ∈ 𝔪(ℱ ) ⇐⇐⇐⇐⇐⇐⇐⇒
⇐ 𝐶 ∪ 𝐷 = 𝑋 ⧵ (𝑋 ⧵ 𝐶) ⧵ 𝐷 ∈ 𝔪(ℱ ).
(Σ1 )
29
R.L. Schilling: Measures, Integrals & Martingales
■■
Problem 3.15 Solution: By definition, ℳ = 𝔪(𝒪) is the monotone class generated by the open sets.
Note that
⋃{ }
∀𝑈 ∈ 𝒪 ∶ 𝑈 = 𝐵𝑟 (𝑥) ∶ 𝑟 ∈ Q+ , 𝑥 ∈ Q𝑑 , 𝐵𝑟 (𝑥) ⊂ 𝑈
⋃ ⋃{ }
= 𝐵𝑟 (𝑥) ∶ 𝑟 ∈ Q+ , 𝑥 ∈ Q𝑑 , 𝐵𝑟−1∕𝑛 (𝑥) ⊂ 𝑈
𝑛∈N
which means that we can write every 𝑈 ∈ 𝒪 as a union of countably many closed sets (i.e. it is
a so-called 𝐹𝜎 -set). Since unions of finitely many closed sets are still closed, we can arrange the
latter union to be an increasing union. Using the de-Morgan laws, this means that every closed set
𝐶 ∈ 𝒞 ⇐⇒ 𝐶 𝑐 ∈ 𝒪 can be written as a countable intersection of (decreasing) open sets.
Since 𝔪(𝒪) is stable under countable intersections of its members, we get 𝒪 ∪ 𝒞 ⊂ 𝔪(𝒪) ⊂
𝔪(𝒪 ∪ 𝒞 ), hence 𝔪(𝒪) = 𝔪(𝒪 ∪ 𝒞 ). Please note that 𝒪 ∪ 𝒞 = {𝐴 ∶ 𝐴 ∈ 𝒪 or 𝐴 ∈ 𝒞 }.
Since 𝒪 ∪ 𝒞 is stable under the formation of complements, the monotone class 𝔪(𝒪 ∪ 𝒞 ) is stable
under the formation of complements (see Problem 3.14 (ii)), thus it is a 𝜎-algebra containing 𝒪
and 𝒞 .
The answer to the additional question is: yes, we can omit the monotonicity in the countable inter-
section and union. The argument is as follows: Problem 3.14 still works without the monotonicity
(giving a slightly different notion of monotone class), and so the above proof goes through!
■■
⋃
Problem 3.16 Solution: Write Σ ∶= {𝜎(𝒞 ) ∶ 𝒞 ⊂ ℱ , 𝒞 is a countable sub-family}.
• Clearly, ∅ ∈ Σ.
• If (𝑆𝑛 )𝑛⩾0 ⊂ Σ are countably many sets, then 𝑆𝑛 ∈ 𝜎(𝒞𝑛 ) for some countable 𝒞𝑛 ⊂ ℱ and
⋃
each 𝑛 ⩾ 0. Set 𝒞 ∶= 𝑛 𝒞𝑛 . This is again countable and we get 𝑆𝑛 ∈ 𝜎(𝒞 ) for all 𝑛, hence
⋃ ⋃
𝑛 𝑆𝑛 ∈ 𝜎(𝒞 ) and so 𝑛 𝑆𝑛 ∈ Σ.
■■
30
4 Measures.
Solutions to Problems 4.1–4.22
We use induction in 𝑁 ∈ N. The hypothesis is clear, for the start (𝑁 = 2) see Proposition
4.3(i). Induction step: take 𝑁+1 disjoint sets 𝐴1 , … , 𝐴𝑁+1 ∈ 𝒜 , set 𝐵 ∶= 𝐴1 ⊍…⊍𝐴𝑁 ∈ 𝒜
and use the induction start and the hypothesis to conclude
(iv) To get an idea what is going on we consider first the case of three sets 𝐴, 𝐵, 𝐶. Applying the
formula for strong additivity thrice we get
We prove this formula by induction. The induction start is just the formula from Proposition
4.3(iv), the hypothesis is given above. For the induction step we observe that
∑ ∑ ∑
= +
𝜎⊂{1,…,𝑛+1} 𝜎⊂{1,…,𝑛,𝑛+1} 𝜎⊂{1,…,𝑛,𝑛+1}
#𝜎=𝑘 #𝜎=𝑘, 𝑛+1∉𝜎 #𝜎=𝑘, 𝑛+1∈𝜎
∑ ∑ (∗)
= +
𝜎⊂{1,…,𝑛} 𝜎 ′ ⊂{1,…,𝑛}
#𝜎=𝑘 #𝜎 ′ =𝑘−1, 𝜎∶=𝜎 ′ ∪{𝑛+1}
31
R.L. Schilling: Measures, Integrals & Martingales
Having this in mind we get for 𝐵 ∶= 𝐴1 ∪ … ∪ 𝐴𝑛 and 𝐴𝑛+1 using strong additivity and the
induction hypothesis (for 𝐴1 , … , 𝐴𝑛 resp. 𝐴1 ∩ 𝐴𝑛+1 , … , 𝐴𝑛 ∩ 𝐴𝑛+1 )
We use induction in 𝑁 ∈ N. The hypothesis is clear, for the start (𝑁 = 2) see Proposition
4.3(v). Induction step: take 𝑁 + 1 sets 𝐵1 , … , 𝐵𝑁+1 ∈ 𝒜 , set 𝐶 ∶= 𝐵1 ∪ … ∪ 𝐵𝑁 ∈ 𝒜 and
use the induction start and the hypothesis to conclude
■■
32
Solution Manual. Last update 28th October 2021
∑
= 𝛿𝑥 (𝐴𝑗0 ) + 𝛿𝑥 (𝐴𝑗 )
𝑗≠𝑗0
∑
= 𝛿𝑥 (𝐴𝑗 ).
𝑗∈N
⋃
If 𝑥 ∉ 𝑗∈N 𝐴𝑗 , then 𝑥 ∉ 𝐴𝑗 for every 𝑗 ∈ N, hence
( )
⋃ ∑
𝛿𝑥 𝐴𝑗 =0=0+0+0+…= 𝛿𝑥 (𝐴𝑗 ).
𝑗∈N 𝑗∈N
⎧
⎪0, if #𝐴 ⩽ #N
(ii) The measure 𝛾 is defined on (R, 𝒜 ) by 𝛾(𝐴) ∶= ⎨ where 𝒜 ∶= {𝐴 ⊂ R ∶
⎪1, if #𝐴𝑐 ⩽ #N
⎩
#𝐴 ⩽ #N or #𝐴𝑐 ⩽ #N}. (Note that #𝐴 ⩽ #N if, and only if, #𝐴𝑐 = #R ⧵ 𝐴 > #N.)
⋃
(M2 ) Let (𝐴𝑗 )𝑗∈N be pairwise disjoint 𝒜 -sets. If all of them are countable, then 𝐴 ∶= 𝑗∈N
is countable and we get
( )
⋃ ∑
𝛾 𝐴𝑗 = 𝛾(𝐴) = 0 = 𝛾(𝐴𝑗 ).
𝑗∈N 𝑗∈N
If at least one 𝐴𝑗 is not countable, say for 𝑗 = 𝑗0 , then 𝐴 ⊃ 𝐴𝑗0 is not countable and therefore
𝛾(𝐴) = 𝛾(𝐴𝑗0 ) = 1. Assume we could find some other 𝑗1 ≠ 𝑗0 such that 𝐴𝑗0 , 𝐴𝑗1 are not
countable. Since 𝐴𝑗0 , 𝐴𝑗1 ∈ 𝒜 we know that their complements 𝐴𝑐𝑗 , 𝐴𝑐𝑗 are countable, hence
0 1
𝐴𝑐𝑗 ∪𝐴𝑐𝑗 is countable and, at the same time, ∈ 𝒜 . Because of this, (𝐴𝑐𝑗 ∪𝐴𝑐𝑗 )𝑐 = 𝐴𝑗0 ∩𝐴𝑗1 = ∅
0 1 0 1
cannot be countable, which is absurd! Therefore there is at most one index 𝑗0 ∈ N such that
𝐴𝑗0 is uncountable and we get then
( )
⋃ ∑
𝛾 𝐴𝑗 = 𝛾(𝐴) = 1 = 1 + 0 + 0 + … = 𝛾(𝐴𝑗0 ) + 𝛾(𝐴𝑗 ).
𝑗∈N 𝑗≠𝑗0
⎧
⎪#𝐴, if 𝐴 is finite
(iii) We have an arbitrary measurable space (𝑋, 𝒜 ) and the measure |𝐴| = ⎨ .
⎪∞, else
⎩
(M2 ) Let (𝐴𝑗 )𝑗∈N be a sequence of pairwise disjoint sets in 𝒜 . Case 1: All 𝐴𝑗 are finite and
⋃
only finitely many, say the first 𝑘, are non-empty, then 𝐴 = 𝑗∈N 𝐴𝑗 is effectively a finite
union of 𝑘 finite sets and it is clear that
∑
|𝐴| = |𝐴1 | + … + |𝐴𝑘 | + |∅| + |∅| + … = |𝐴𝑗 |.
𝑗∈N
33
R.L. Schilling: Measures, Integrals & Martingales
⋃
Case 2: All 𝐴𝑗 are finite and infinitely many are non-void. Then their union 𝐴 = 𝑗∈N 𝐴𝑗
is an infinite set and we get
∑
|𝐴| = ∞ = |𝐴𝑗 |.
𝑗∈N
⋃
Case 3: At least one 𝐴𝑗 is infinite, and so is then the union 𝐴 = 𝑗∈N 𝐴𝑗 . Thus,
∑
|𝐴| = ∞ = |𝐴𝑗 |.
𝑗∈N
∑
(iv) On a countable set Ω = {𝜔1 , 𝜔2 , …} we define for a sequence (𝑝𝑗 )𝑗∈N ⊂ [0, 1] with 𝑗∈N 𝑝𝑗 =
1 the set function
∑ ∑
𝑃 (𝐴) = 𝑝𝑗 = 𝑝𝑗 𝛿𝜔𝑗 (𝐴), 𝐴 ⊂ Ω.
𝑗∶𝜔𝑗 ∈𝐴 𝑗∈N
The change in the order of summation needs justification; one possibility is the argument
used in the solution of Problem 4.7(ii). (Note that the reordering theorem for absolutely
convergent series is not immediately applicable since we deal with a double series!)
■■
• On (R, ℬ(R)) the function 𝛾 is not be a measure, since we can take the sets 𝐴 = (1, ∞),
𝐵 = (−∞, −1) which are disjoint, not countable and both have non-countable complements.
Hence, 𝛾(𝐴) = 𝛾(𝐵) = 1. On the other hand, 𝐴 ⊍ 𝐵 is non-countable and has non-countable
complement, [−1, 1]. So, 𝛾(𝐴 ⊍ 𝐵) = 1. This contradicts the additivity: 𝛾(𝐴 ⊍ 𝐵) = 1 ≠
2 = 𝛾(𝐴) + 𝛾(𝐵). Notice that the choice of the 𝜎-algebra 𝒜 avoids exactly this situation. ℬ
is the wrong 𝜎-algebra for 𝛾.
34
Solution Manual. Last update 28th October 2021
• On Q (and, actually, any possible 𝜎-algebra thereon) the problem is totally different: if 𝐴
is countable, then 𝐴𝑐 = Q ⧵ 𝐴 is also countable and vice versa. This means that 𝛾(𝐴) is,
according to the definition, both 1 and 0 which is, of course, impossible. This is to say: 𝛾 is
not well-defined. 𝛾 makes only sense on a non-countable set 𝑋.
■■
But as soon as 𝒜 contains one set 𝐴 which is trivial (i.e. either ∅ or 𝑋), we have actually
𝐴𝑐 ∈ 𝒜 which is also non-trivial. Thus,
(ii) If we equip R with a 𝜎-algebra which contains sets such that both 𝐴 and 𝐴𝑐 can be infinite
(the Borel 𝜎-algebra would be such an example: 𝐴 = (−∞, 0) ⇐⇒ 𝐴𝑐 = [0, ∞)), then 𝜈 is
not well-defined. The only type of sets where 𝜈 is well-defined is, thus,
■■
Problem 4.5 Solution: Denote by 𝜆 one-dimensional Lebesgue measure and consider the Borel sets
⋂
𝐵𝑘 ∶= (𝑘, ∞). Clearly 𝑘 𝐵𝑘 = ∅, 𝑘 ∈ N, so that 𝐵𝑘 ↓ ∅. On the other hand,
■■
Problem 4.6 Solution: Mind the typo in the problem: it should read “infinite mass” – otherwise the
problem is pointless.
35
R.L. Schilling: Measures, Integrals & Martingales
(Since Z×N is countable, Problem 4.7 shows that this object is indeed a measure!) Obviously, any
interval [𝑎, 𝑏) of length 𝑏−𝑎 > 2 contains some integer, say 𝑚 ∈ [𝑎, 𝑏) so that [𝑚−1∕2, 𝑚) ⊂ [𝑎, 𝑏),
thus
∑ 1
𝜇[𝑎, 𝑏) ⩾ 𝜇[𝑚 − 1∕2, 𝑚) = = ∞.
𝑘∈N
2𝑘
On the other hand, the sequence of sets
𝑛
⋃ [ 1)
𝐵𝑛 ∶= 𝑘 − 1, 𝑘 − 2𝑛
𝑘=−𝑛
⋃
satisfies 𝜇(𝐵𝑛 ) < ∞ and 𝑛 𝐵𝑛 = R.
Solution 2: Set 𝜇(𝐵) ∶= #(𝐵 ∩ Q), 𝐵 ∈ ℬ(R), i.e. the counting measure of the rationals in R.
Clearly, 𝜇[𝑎, 𝑏) = ∞ for every (non-empty) interval with 𝑎 < 𝑏. On the other hand, if (𝑞𝑘 )𝑘∈N
is an enumeration of Q, the sets 𝐵𝑛 ∶= (R ⧵ Q) ∪ {𝑞1 , … , 𝑞𝑛 } satisfy
𝐵𝑛 ↑ R and 𝜇(𝐵𝑛 ) = 𝑛,
i.e. 𝜇 is 𝜎-finite.
■■
(M2 ) Let (𝐴𝑗 )𝑗∈N ⊂ 𝒜 be mutually disjoint sets. Then we can use the 𝜎-additivity of 𝜇, 𝜈 to
get
( ) ( ) ( )
⋃ ⋃ ⋃
𝜌 𝐴𝑗 = 𝑎𝜇 𝐴𝑗 + 𝑏𝜈 𝐴𝑗
𝑗∈N 𝑗∈N 𝑗∈N
∑ ∑
=𝑎 𝜇(𝐴𝑗 ) + 𝑏 𝜈(𝐴𝑗 )
𝑗∈N 𝑗∈N
∑( )
= 𝑎𝜇(𝐴𝑗 ) + 𝑏𝜇(𝐴𝑗 )
𝑗∈N
∑
= 𝜌(𝐴𝑗 ).
𝑗∈N
Since all quantities involved are positive and since we allow the value +∞ to be attained,
there are no convergence problems.
∑
(ii) Since all 𝛼𝑗 are positive, the sum 𝑗∈N 𝛼𝑗 𝜇𝑗 (𝐴) is a sum of positive quantities and, allowing
the value +∞ to be attained, there is no convergence problem. Thus, 𝜇 ∶ 𝒜 → [0, ∞] is
well-defined. Before we check (𝑀1 ), (𝑀2 ) we prove the following
36
Solution Manual. Last update 28th October 2021
Proof. Observe that we have 𝛽𝑚𝑛 ⩽ sup𝑗∈N sup𝑖∈N 𝛽𝑖𝑗 for all 𝑚, 𝑛 ∈ N. The right-hand side
is independent of 𝑚 and 𝑛 and we may take the 𝑠𝑢𝑝 over all 𝑛
and then, with the same argument, take the sup over all 𝑚
The opposite inequality, ‘⩾’, follows from the same argument with 𝑖 and 𝑗 interchanged.
∑ ∑
(M1 ) We have 𝜇(∅) = 𝑗∈N 𝛼𝑗 𝜇𝑗 (∅) = 𝑗∈N 𝛼𝑗 ⋅ 0 = 0.
(M2 ) Take pairwise disjoint sets (𝐴𝑖 )𝑖∈N ⊂ 𝒜 . Then we can use the 𝜎-additivity of each of
the 𝜇𝑗 ’s to get
( ) ( )
⋃ ∑ ⋃
𝜇 𝐴𝑖 = 𝛼𝑗 𝜇𝑗 𝐴𝑖
𝑖∈N 𝑗∈N 𝑖∈N
𝑁
∑ ∑ ( )
= lim 𝛼𝑗 𝜇𝑗 𝐴𝑖
𝑁→∞
𝑗=1 𝑖∈N
𝑁 𝑀
∑ ∑ ( )
= lim 𝛼𝑗 lim 𝜇 𝑗 𝐴𝑖
𝑁→∞ 𝑀→∞
𝑗=1 𝑖=1
𝑁 𝑀
∑ ∑ ( )
= lim lim 𝛼 𝑗 𝜇 𝑗 𝐴𝑖
𝑁→∞ 𝑀→∞
𝑗=1 𝑖=1
𝑀 𝑁
∑∑ ( )
= sup sup 𝛼 𝑗 𝜇 𝑗 𝐴𝑖
𝑁∈N 𝑀∈N 𝑗=1 𝑖=1
where we use that the limits are increasing limits, hence suprema. By our lemma:
( ) 𝑀 𝑁
⋃ ∑ ∑ ( )
𝜇 𝐴𝑖 = sup sup 𝛼 𝑗 𝜇 𝑗 𝐴𝑖
𝑖∈N 𝑀∈N 𝑁∈N 𝑖=1 𝑗=1
𝑀 𝑁
∑ ∑ ( )
= lim lim 𝛼 𝑗 𝜇 𝑗 𝐴𝑖
𝑀→∞ 𝑁→∞
𝑖=1 𝑗=1
𝑀
∑ ∑ ( )
= lim 𝛼𝑗 𝜇𝑗 𝐴𝑖
𝑀→∞
𝑖=1 𝑗∈N
𝑀
∑ ( )
= lim 𝜇 𝐴𝑖
𝑀→∞
𝑖=1
∑ ( )
= 𝜇 𝐴𝑖 .
𝑖∈N
■■
37
R.L. Schilling: Measures, Integrals & Martingales
where we use finite additivity for (1) and 𝜎-subaddtitivity for (2).
■■
Problem 4.9 Solution: Set 𝜈(𝐴) ∶= 𝜇(𝐴 ∩ 𝐹 ). We know, by assumption, that 𝜇 is a measure on
(𝑋, 𝒜 ). We have to show that 𝜈 is a measure on (𝑋, 𝒜 ). Since 𝐹 ∈ 𝒜 , we have 𝐹 ∩ 𝐴 ∈ 𝒜 for
all 𝐴 ∈ 𝒜 , so 𝜈 is well-defined. Moreover, it is clear that 𝜈(𝐴) ∈ [0, ∞]. Thus, we only have to
check
(M2 ) Let (𝐴𝑗 )𝑗∈N ⊂ 𝒜 be a sequence of pairwise disjoint sets. Then also (𝐴𝑗 ∩ 𝐹 )𝑗∈N ⊂ 𝒜 are
pairwise disjoint and we can use the 𝜎-additivity of 𝜇 to get
( ) ( ) ( )
⋃ ⋃ ⋃
𝜈 𝐴𝑗 =𝜇 𝐹 ∩ 𝐴𝑗 = 𝜇 (𝐹 ∩ 𝐴𝑗 )
𝑗∈N 𝑗∈N 𝑗∈N
∑
= 𝜇(𝐹 ∩ 𝐴𝑗 )
𝑗∈N
∑
= 𝜈(𝐴𝑗 ).
𝑗∈N
■■
Problem 4.10 Solution: Since 𝑃 is a probability measure, 𝑃 (𝐴𝑐𝑗 ) = 1−𝑃 (𝐴𝑗 ) = 0. By 𝜎-subadditivity,
( )
⋃ ∑
𝑃 𝐴𝑐𝑗 ⩽ 𝑃 (𝐴𝑐𝑗 ), = 0
𝑗∈N 𝑗∈N
■■
38
Solution Manual. Last update 28th October 2021
■■
0 ⩽ 𝜆({𝑥}) ⩽ 𝜆((𝑥 − 𝑘1 , 𝑥 + 𝑘1 )) = (𝑥 + 𝑘1 ) − (𝑥 − 𝑘1 ) = 2
←←←←←←←←←→
𝑘 𝑘→∞
← 0.
39
R.L. Schilling: Measures, Integrals & Martingales
b) Take again an enumeration Q = {𝑞1 , 𝑞2 , 𝑞3 , …}, fix 𝜖 > 0 and define 𝐶(𝜖) as stated in
the problem. Then we have 𝐶(𝜖) ∈ ℬ and Q ⊂ 𝐶(𝜖). Using the monotonicity and
𝜎-subadditivity of 𝜆 we get
( )
0 ⩽ 𝜆(Q) ⩽ 𝜆 𝐶(𝜖)
( )
⋃
−𝑘 −𝑘
=𝜆 [𝑞𝑘 − 𝜖2 , 𝑞𝑘 + 𝜖2 )
𝑘∈N
∑ ( )
⩽ 𝜆 [𝑞𝑘 − 𝜖2−𝑘 , 𝑞𝑘 + 𝜖2−𝑘 )
𝑘∈N
∑
= 2 ⋅ 𝜖 ⋅ 2−𝑘
𝑘∈N
1
2
= 2𝜖 1
= 2𝜖.
1− 2
which is impossible.
■■
Problem 4.14 Solution: Without loss of generality we may assume that 𝑎 ≠ 𝑏; set 𝜇 ∶= 𝛿𝑎 + 𝛿𝑏 .
Then 𝜇(𝐵) = 0 if, and only if, 𝑎 ∉ 𝐵 and 𝑏 ∉ 𝐵. Since {𝑎}, {𝑏} and {𝑎, 𝑏} are Borel sets, all null
sets of 𝜇 are given by
{ }
𝒩𝜇 = 𝐵 ⧵ {𝑎, 𝑏} ∶ 𝐵 ∈ ℬ(R) .
(This shows that, in some sense, null sets can be fairly large!).
■■
Problem 4.15 Solution: Let us write 𝔑 for the family of all (proper and improper) subsets of 𝜇 null
sets. We note that sets in 𝔑 can be measurable (that is: 𝑁 ∈ 𝒜 ) but need not be measurable.
(i) Since ∅ ∈ 𝔑, we find that 𝐴 = 𝐴 ∪ ∅ ∈ 𝒜 for every 𝐴 ∈ 𝒜 ; thus, 𝒜 ⊂ 𝒜 . Let us check that
𝒜 is a 𝜎-algebra.
𝐴∗ 𝑐 = (𝐴 ∪ 𝑁)𝑐 = 𝐴𝑐 ∩ 𝑁 𝑐
= 𝐴𝑐 ∩ 𝑁 𝑐 ∩ (𝑀 𝑐 ∪ 𝑀)
40
Solution Manual. Last update 28th October 2021
= (𝐴𝑐 ∩ 𝑁 𝑐 ∩ 𝑀 𝑐 ) ∪ (𝐴𝑐 ∩ 𝑁 𝑐 ∩ 𝑀)
= (𝐴𝑐 ∩ 𝑀 𝑐 ) ∪ (𝐴𝑐 ∩ 𝑁 𝑐 ∩ 𝑀)
(Σ3 ) Let (𝐴∗𝑗 )𝑗∈N be a sequence of 𝒜 -sets. From its very definition we know that each
𝐴∗𝑗 = 𝐴𝑗 ∪ 𝑁𝑗 for some (not necessarily unique!) 𝐴𝑗 ∈ 𝒜 and 𝑁𝑗 ∈ 𝔑. So,
( ) ( )
⋃ ⋃ ⋃ ⋃
𝐴∗𝑗 = (𝐴𝑗 ∪ 𝑁𝑗 ) = 𝐴𝑗 ∪ 𝑁𝑗 =∶ 𝐴 ∪ 𝑁.
𝑗∈N 𝑗∈N 𝑗∈N 𝑗∈N
Thus, 𝐴 ∪ 𝑁 ∈ 𝒜 .
(ii) As already mentioned in part (i), 𝐴∗ ∈ 𝒜 could have more than one representation, e.g.
𝐴 ∪ 𝑁 = 𝐴∗ = 𝐵 ∪ 𝑀 with 𝐴, 𝐵 ∈ 𝒜 and 𝑁, 𝑀 ∈ 𝔑. If we can show that 𝜇(𝐴) = 𝜇(𝐵)
then the definition of 𝜇̄ is independent of the representation of 𝐴∗ . Since 𝑀, 𝑁 are not
necessarily measurable but, by definition, subsets of (measurable) null sets 𝑀 ′ , 𝑁 ′ ∈ 𝒜 we
find
𝐴 ⊂ 𝐴 ∪ 𝑁 = 𝐵 ∪ 𝑀 ⊂ 𝐵 ∪ 𝑀 ′,
𝐵 ⊂ 𝐵 ∪ 𝑀 = 𝐴 ∪ 𝑁 ⊂ 𝐴 ∪ 𝑁′
and since 𝐴, 𝐵, 𝐵∪𝑀 ′ , 𝐴∪𝑁 ′ ∈ 𝒜 , we get from monotonicity and subadditivity of measures
(M2 ) Let (𝐴∗𝑗 )𝑗∈N ⊂ 𝒜 be a sequence of pairwise disjoint sets. Then 𝐴∗𝑗 = 𝐴𝑗 ∪ 𝑁𝑗 for some
𝐴𝑗 ∈ 𝒜 and 𝑁𝑗 ∈ 𝔑. These sets are also mutually disjoint, and with the arguments in
(i) we see that 𝐴∗ = 𝐴 ∪ 𝑁 where 𝐴∗ ∈ 𝒜 , 𝐴 ∈ 𝒜 , 𝑁 ∈ 𝔑 stand for the unions of
41
R.L. Schilling: Measures, Integrals & Martingales
𝐴∗𝑗 , 𝐴𝑗 and 𝑁𝑗 , respectively. Since 𝜇̄ does not depend on the special representation of
𝒜 -sets, we get
( ) ( )
⋃ ⋃
𝜇̄ 𝐴∗𝑗 = 𝜇(𝐴
̄ ∗
) = 𝜇(𝐴) = 𝜇 𝐴𝑗
𝑗∈N 𝑗∈N
∑
= 𝜇(𝐴𝑗 )
𝑗∈N
∑
= ̄ ∗𝑗 )
𝜇(𝐴
𝑗∈N
know that 𝜇(𝑀) = 0. Moreover, we know from the definition of 𝔑 that 𝑁 ⊂ 𝑁 ′ for some
𝑁 ′ ∈ 𝒜 with 𝜇(𝑁 ′ ) = 0. This entails
𝐵 ⊂ 𝑀∗ = 𝑀 ∪ 𝑁 ⊂ 𝑀 ∪ 𝑁′ ∈ 𝒜
and 𝜇(𝑀 ∪ 𝑁 ′ ) ⩽ 𝜇(𝑀) + 𝜇(𝑁 ′ ) = 0.
𝐴 ⊂ 𝐴∗ = 𝐴 ∪ 𝑁 ⊂ 𝐴 ∪ 𝑁 ′ =∶ 𝐵 ∈ 𝒜
and that 𝜇(𝐵 ⧵𝐴) = 𝜇((𝐴∪𝑁 ′ )⧵𝐴) ⩽ 𝜇(𝑁 ′ ) = 0. (Note that (𝐴∪𝑁 ′ )⧵𝐴 = (𝐴∪𝑁 ′ )∩𝐴𝑐 =
𝑁 ′ ∩ 𝐴𝑐 ⊂ 𝑁 ′ and that equality need not hold!).
■■
and denote, without further mentioning, by 𝐹 , 𝐹𝑗 resp. 𝑁, 𝑁𝑗 sets from ℱ resp. 𝒩 . Since 𝐹 ▵ ∅ =
𝐹 , ∅ ▵ 𝑁 = 𝑁 and 𝐹 ▵ 𝑁 ∈ 𝜎(ℱ , 𝒩 ) we get
ℱ , 𝒩 ⊂ Σ ⊂ 𝜎(ℱ , 𝒩 ) (*)
42
Solution Manual. Last update 28th October 2021
and the first assertion follows if we can show that Σ is a 𝜎-algebra. In this case, we can apply the
𝜎-operation to the inclusions (*) and get
which is just
𝜎(ℱ , 𝒩 ) ⊂ Σ ⊂ 𝜎(ℱ , 𝒩 ).
∈ Σ;
(Σ3 ): We begin by a few simple observations, namely that for all 𝐹 ∈ ℱ and 𝑁, 𝑁 ′ ∈ 𝒩
𝐹 ∪ 𝑁 = 𝐹 ▵ (𝑁 ⧵ 𝐹 ) ∈ Σ; (a)
⏟⏞⏟⏞⏟
∈𝒩
𝐹 ⧵ 𝑁 = 𝐹 ▵ (𝑁 ∩ 𝐹 ) ∈ Σ; (b)
⏟⏞⏟⏞⏟
∈𝒩
𝑁 ⧵ 𝐹 = 𝑁 ▵ (𝐹 ∩ 𝑁) ∈ Σ; (c)
⏟⏞⏟⏞⏟
∈𝒩
( ) ( )
(𝐹 ▵ 𝑁) ∪ 𝑁 = 𝐹 ▵ 𝑁 ▵ 𝑁 ′ ⧵ (𝐹 ▵ 𝑁)
′
( )
= 𝐹 ▵ 𝑁 ▵(𝑁 ′ ⧵ (𝐹 ▵ 𝑁)) ∈ Σ, (d)
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
∈𝒩
43
R.L. Schilling: Measures, Integrals & Martingales
as well as
⋃ ⋃
∅⊂ (𝑁𝑗 ⧵ 𝐹𝑗 ) ⊂ 𝑁𝑗 = 𝑁
𝑗∈N 𝑗∈N
Since
𝐴 ∪ 𝑁 = 𝐴 ▵ (𝑁 ⧵ 𝐴)
⏟⏟⏟
∈𝒩
𝒜 ⊂ 𝜎(𝒜 , 𝒩 ).
𝒜 ⊂𝒜 and 𝒩 ⊂𝒜
𝐴∗ ▵ 𝐴 = 𝐴 ▵ 𝑁 ▵ 𝐴 = (𝐴 ▵ 𝐴) ▵ 𝑁 = 𝑁.
Note that this result would also follow directly from 4.15 since we know from there that 𝐴∗ = 𝐴∪𝑁
so that
𝐴∗ ▵ 𝐴 = (𝐴 ∪ 𝑁) ▵ 𝐴 = 𝐴 ▵(𝑁 ⧵ 𝐴) ▵ 𝐴 = 𝑁 ⧵ 𝐴
■■
44
Solution Manual. Last update 28th October 2021
Problem 4.18 Solution: Denote the completion by ℬ ∗ and write 𝒩𝑥 for all subsets of Borel null sets
of 𝛿𝑥 . Clearly,
𝒩𝑥 = {𝐴 ⊂ R𝑛 ∶ 𝑥 ∉ 𝐴}.
Recall from Problem 4.15(i) that ℬ ∗ contains all sets of the form 𝐵 ∪𝑁 with 𝐵 ∈ ℬ and 𝑁 ∈ 𝒩𝑥 .
Now let 𝐶 ⊂ R𝑛 be any set. If 𝑥 ∈ 𝐶, then write
𝐶 = {𝑥} ∪ (𝐶 ⧵ {𝑥}) ∈ ℬ ∗ ;
⏟⏟⏟ ⏟⏞⏞⏟⏞⏞⏟
∈ℬ ∈ 𝒩𝑥
Otherwise, 𝑥 ∉ 𝐶 and
𝐶 = 𝐶 ⧵ {𝑥} = ∅ ∪ (𝐶 ⧵ {𝑥}) ∈ ℬ ∗ .
⏟⏟⏟ ⏟⏞⏞⏟⏞⏞⏟
∈ℬ ∈ 𝒩𝑥
■■
(i) Since ℬ is a 𝜎-algebra, it is closed under countable (disjoint) unions of its elements, thus 𝜈
inherits the properties (M1 ), (M2 ) directly from 𝜇.
(ii) Yes [yes], since the full space 𝑋 ∈ ℬ so that 𝜇(𝑋) = 𝜈(𝑋) is finite [resp. = 1].
(iii) No, 𝜎-finiteness is also a property of the 𝜎-algebra. Take, for example, Lebesgue measure 𝜆
on the Borel sets (this is 𝜎-finite) and consider the 𝜎-algebra 𝒞 ∶= {∅, (−∞, 0), [0, ∞), R}.
Then 𝜆| is not 𝜎-finite since there is no increasing sequence of 𝒞 -sets having finite measure.
|
|𝒞
■■
Problem 4.20 Solution: By definition, 𝜇 is 𝜎-finite if there is an increasing sequence (𝐵𝑗 )𝑗∈N ⊂ 𝒜
such that 𝐵𝑗 ↑ 𝑋 and 𝜇(𝐵𝑗 ) < ∞. Clearly, 𝐸𝑗 ∶= 𝐵𝑗 satisfies the condition in the statement of the
problem.
Remark: A small change in the above argument allows to take pairwise disjoint sets 𝐸𝑗 .
■■
45
R.L. Schilling: Measures, Integrals & Martingales
(i) Fix 𝜖 > 0 and choose for 𝐴 ∈ Σ sets 𝑈 ∈ 𝒪, 𝐹 ∈ ℱ such that 𝐹 ⊂ 𝐴 ⊂ 𝑈 and 𝜇(𝑈 ⧵𝐹 ) < 𝜖.
Set 𝑈 ′ ∶= 𝐹 𝑐 ∈ 𝒪 and 𝐹 ′ ∶= 𝑈 𝑐 ∈ ℱ . Then we have
𝐹 ′ ⊂ 𝐴𝑐 ⊂ 𝑈 ′ and 𝑈 ′ ⧵ 𝐹 ′ = 𝐹 𝑐 ⧵ 𝑈 𝑐 = 𝐹 𝑐 ∩ 𝑈 = 𝑈 ⧵ 𝐹
Denote by 𝑑(𝑥, 𝑦) the distance of two points 𝑥, 𝑦 ∈ 𝑋 and write 𝐵1∕𝑛 (0) for the open ball
{𝑦 ∈ 𝑋 ∶ 𝑑(𝑦, 0) < 1𝑛 }. As in the solution of Problem 3.14(ii) we see that 𝑈𝑛 ∶= 𝐹 +𝐵1∕𝑛 (0)
is a sequence of open sets such that 𝑈𝑛 ↓ 𝐹 . Because of the continuity of measures we get
← 0 and since ℱ ∋ 𝐹 ⊂ 𝐹 ⊂ 𝑈𝑛 ∈ 𝒪, this means that ℱ ⊂ Σ.
𝜇(𝑈𝑛 ⧵ 𝐹 ) ←←←←←←←←←→
𝑛→∞
(ii) Fix 𝜖 > 0 and pick for 𝐴𝑗 ∈ Σ, 𝑗 = 1, 2, open sets 𝑈𝑗 and closed sets 𝐹𝑗 such that 𝐹𝑗 ⊂
𝐴 ⊂ 𝑈𝑗 and 𝜇(𝑈𝑗 ⧵ 𝐹𝑗 ) < 𝜖. Then 𝐹1 ∩ 𝐹2 and 𝑈1 ∩ 𝑈2 are again closed resp. open, satisfy
𝐹1 ∩ 𝐹2 ⊂ 𝐴1 ∩ 𝐴2 ⊂ 𝑈1 ∩ 𝑈2 as well as
( ) ( )
𝜇 (𝑈1 ∩ 𝑈2 ) ⧵ (𝐹1 ∩ 𝐹2 ) = 𝜇 (𝑈1 ∩ 𝑈2 ) ∩ (𝐹1𝑐 ∪ 𝐹2𝑐 )
( )
= 𝜇 [(𝑈1 ∩ 𝑈2 ) ⧵ 𝐹1 ] ∪ [(𝑈1 ∩ 𝑈2 ) ⧵ 𝐹2 ]
( ) ( )
⩽ 𝜇 (𝑈1 ∩ 𝑈2 ) ⧵ 𝐹1 + 𝜇 (𝑈1 ∩ 𝑈2 ) ⧵ 𝐹2
< 2𝜖.
(iii) Fix 𝜖 and pick for a given sequence (𝐴𝑗 )𝑗∈N ⊂ Σ open sets 𝑈𝑗 and closed sets 𝐹𝑗 such that
(iv) For any Borel set 𝐵 ∈ Σ and any 𝜖 > 0 we can find open and closed sets 𝑈𝜖 and 𝐹𝜖 , respect-
ively, such that 𝐹𝜖 ⊂ 𝐵 ⊂ 𝑈𝜖 and
46
Solution Manual. Last update 28th October 2021
Thus,
Thus,
= sup 𝜇(𝐾)
𝐾⊂𝐵,𝐾 cpt
and since 𝜇(𝐾) ⩽ 𝜇(𝐵) for 𝐾 ⊂ 𝐵 and sup𝐾⊂𝐵,𝐾 cpt 𝜇(𝐾) ⩽ 𝜇(𝐵) are obvious, we are
finished.
(vi) Assume now that 𝜇 is 𝜎-finite. Let (𝐵𝑛 )𝑛∈N ⊂ ℬ be an exhausting sequence for 𝑋 such
that 𝜇(𝐵𝑛 ) < ∞. Then the measures 𝜇𝑛 (𝐵) ∶= 𝜇(𝐵 ∩ 𝐵𝑛 ) defined on ℬ are finite and
regular according to part (iv). Since we may interchange any two suprema (cf. the solution
of Problem 4.7) we get
= sup sup 𝜇𝑛 (𝐹 )
𝐹 ⊂𝐵, 𝐹 ∈ℱ 𝑛
= sup 𝜇(𝐹 ).
𝐹 ⊂𝐵, 𝐹 ∈ℱ
■■
47
R.L. Schilling: Measures, Integrals & Martingales
Let (𝑑𝑘 )𝑘 be an enumeration of the dense set 𝐷 ⊂ 𝑋 and write 𝜌 for the metric in 𝑋 and 𝐾𝑟 (𝑥) ∶=
{𝑦 ∈ 𝑋 ∶ 𝜌(𝑥, 𝑦) ⩽ 𝑟} for the closed ball with centre 𝑥 and radius 𝑟.
𝜖
∀𝜖 > 0 ∃ 𝑘(𝑛) ∈ N ∶ 𝜇(𝐹𝑛 ) + ⩾ 𝜇(𝑋)
2𝑛
if 𝐹𝑛 ∶= 𝐾1∕𝑛 (𝑑1 ) ∪ ⋯ ∪ 𝐾1∕𝑛 (𝑑𝑘(𝑛) ). Setting
⋂
𝐾 ∶= 𝐾𝜖 ∶= 𝐹𝑛
𝑛
it is clear that 𝐾 is closed. Moreover, since 𝐾 is, for every 1∕𝑛, covered by finitely many balls of
radius 1∕𝑛, to wit,
we see that 𝐾 is compact. Indeed, if (𝑥𝑗 )𝑗 ⊂ 𝐾 is a sequence, there is a subsequence (𝑥𝑛𝑗 )𝑗 which
is completely contained in one of the balls 𝐾1∕𝑛 (𝑑1 ), … , 𝐾1∕𝑛 (𝑑𝑘(𝑛) ). Passing iteratively to sub-
sub-etc. sequences we find a subsequence (𝑦𝑗 )𝑗 ⊂ (𝑥𝑗 )𝑗 which is contained in a sequence of closed
balls 𝐾1∕𝑛 (𝑐𝑛 ) (𝑐𝑛 is a suitable element from 𝐷). Thus (𝑦𝑗 )𝑗 is a Cauchy sequence and converges,
because of completeness, to an element 𝑥∗ which is, as the 𝐹𝑛 are closed, in every 𝐹𝑛 , hence in 𝐾.
Thus 𝐾 is (sequentially) compact.
Since
( )
⋃ ∑ ∑ 𝜖
𝜇(𝑋 ⧵ 𝐾) = 𝜇 𝑋 ⧵ 𝐹𝑛 ⩽ 𝜇(𝑋 ⧵ 𝐹𝑛 ) ⩽ = 𝜖,
𝑛 𝑛 𝑛
2𝑛
we have found a sequence of compact sets 𝐾𝑛 such that 𝜇(𝐾𝑛 ) → 𝜇(𝑋) (note that the 𝐾𝑛 need not
‘converge’ 𝑋 as a set!). Obviously, 𝐾𝑛 ∩𝐹 is compact for every closed 𝐹 and we have 𝜇(𝐾𝑛 ∩𝐹 ) →
𝜇(𝐹 ), hence
Now we can use the argument from the proof of Problem 4.22(v).
■■
48
5 Uniqueness of measures.
Solutions to Problems 5.1–5.13
Problem 5.1 Solution: Since 𝑋 ∈ 𝒟 and since complements are again in 𝒟 , we have ∅ = 𝑋 𝑐 ∈ 𝒟 .
If 𝐴, 𝐵 ∈ 𝒟 are disjoint, we set 𝐴1 ∶= 𝐴, 𝐴2 ∶= 𝐵, 𝐴𝑗 ∶= ∅ ∀𝑗 ⩾ 3. Then (𝐴𝑗 )𝑗∈N ⊂ 𝒟 is a
sequence of pairwise disjoint sets, and by (D3 ) we find that
⨃
𝐴⊍𝐵 = 𝐴𝑗 ∈ 𝒟 .
𝑗∈N
Since (Σ1 ) = (D3 ), (Σ2 ) = (D2 ) and since (Σ3 ) ⇐⇒ (D3 ), it is clear that every 𝜎-algebra is also a
Dynkin system; that the converse is, in general, wrong is seen in Problem 5.2.
■■
Problem 5.2 Solution: Consider (D3 ) only, as the other two conditions coincide: (Σ𝑗 ) = (Δ𝑗 ), 𝑗 =
1, 2. We show that (Σ3 ) breaks down even for finite unions. If 𝐴, 𝐵 ∈ 𝒟 are disjoint, it is clear that
𝐴, 𝐵 and also 𝐴 ⊍ 𝐵 contain an even number of elements. But if 𝐴, 𝐵 have non-void intersection,
and if this intersection contains an odd number of elements, then 𝐴 ∪ 𝐵 contains an odd number
of elements. Here is a trivial example:
𝐴 = {1, 2} ∈ 𝒟 , 𝐵 = {2, 3, 4, 5} ∈ 𝒟 ,
whereas
𝐴 ∪ 𝐵 = {1, 2, 3, 4, 5} ∉ 𝒟 .
Problem 5.3 Solution: We verify the hint first. Using de Morgan’s laws we get
49
R.L. Schilling: Measures, Integrals & Martingales
■■
(i) Since the 𝜎-algebra 𝒜 is also a Dynkin system, it is enough to prove 𝛿(𝒟 ) = 𝒟 for any
Dynkin system 𝒟 . By definition, 𝛿(𝒟 ) is the smallest Dynkin system containing 𝒟 , thus
𝒟 ⊂ 𝛿(𝒟 ). On the other hand, 𝒟 is itself a Dynkin system, thus, because of minimality,
𝒟 ⊃ 𝛿(𝒟 ).
(ii) Clearly, 𝒢 ⊂ ℋ ⊂ 𝛿(ℋ ). Since 𝛿(ℋ ) is a Dynkin system containing 𝒢 , the minimality of
𝛿(𝒢 ) implies that 𝛿(𝒢 ) ⊂ 𝛿(ℋ ).
(iii) Since 𝜎(𝒢 ) is a 𝜎-algebra, it is also a Dynkin system. Since 𝒢 ⊂ 𝜎(𝒢 ) we conclude (again,
by minimality) that 𝛿(𝒢 ) ⊂ 𝜎(𝒢 ).
■■
Problem 5.5 Solution: Clearly, 𝛿({𝐴, 𝐵}) ⊂ 𝜎({𝐴, 𝐵}) is always true.
If 𝐴 ∩ 𝐵 ≠ ∅, then
{
𝛿({𝐴, 𝐵}) = ∅, 𝐴, 𝐴𝑐 , 𝐵, 𝐵 𝑐 , 𝑋}
■■
Problem 5.6 Solution: Some authors call families of sets satisfying (D1 ), (D′2 ), (D′3 ) monotone classes
(this is not the standard definition!). We will use this convention locally for this solution only.
i.e., ℱ is ⊍-stable. This and (D′3 ) yield (D3 ); (D2 ) is a special case of (D′2 ).
Conversely every Dynkin system 𝒟 is a monotone class in the sense of this problem:
(D2 )
⇐ 𝑁 𝑐 ∩ 𝑀 = 𝑀 ⧵ 𝑁 = ∅ and
𝑀, 𝑁 ∈ 𝒟 , 𝑀 ⊂ 𝑁 ⇐⇐⇐⇐⇐⇐⇐⇐⇒ 𝑁 ⧵ 𝑀 = (𝑁 𝑐 ⊍ 𝑀)𝑐 ∈ 𝒟 ,
(D3 )
■■
50
Solution Manual. Last update 28th October 2021
Problem 5.7 Solution: We prove the hint first. Let (𝐺𝑗 )𝑗∈N ⊂ 𝒢 as stated in the problem, i.e.
satisfying (1) and (2), and define the sets 𝐹𝑁 ∶= 𝐺1 ∪ … ∪ 𝐺𝑁 . As 𝒢 ⊂ 𝒜 , it is clear that
𝐹𝑁 ∈ 𝒜 (but not necessarily in 𝒢 ...). Moreover, it is clear that 𝐹𝑁 ↑ 𝑋.
We begin with a more general assertion: For any finite union of 𝒢 -sets 𝐴1 ∪ … ∪ 𝐴𝑁 we have
𝜇(𝐴1 ∪ … ∪ 𝐴𝑁 ) = 𝜈(𝐴1 ∪ … ∪ 𝐴𝑁 ).
where we use the induction hypothesis twice, namely for the union of the 𝑁 𝒢 -sets 𝐴1 , … , 𝐴𝑁 as
well as for the 𝑁 𝒢 -sets 𝐴1 ∩ 𝐴𝑁+1 , … , 𝐴𝑁 ∩ 𝐴𝑁+1 . The induction is complete.
( ) ( )
If 𝜇 (𝐴1 ∪ ⋯ ∪ 𝐴𝑁 ) ∩ 𝐴𝑁+1 = ∞, hence 𝜈 (𝐴1 ∪ ⋯ ∪ 𝐴𝑁 ) ∩ 𝐴𝑁+1 = ∞, there is nothing to
show since the monotinicity of measures entails
In particular we see that 𝜇(𝐹𝑁 ) = 𝜈(𝐹𝑁 ), 𝜈(𝐹𝑁 ) ⩽ 𝜈(𝐺1 ) + … + 𝜈(𝐺𝑁 ) < ∞ by subadditivity,
and that (think!) 𝜇(𝐺 ∩ 𝐹𝑁 ) = 𝜈(𝐺 ∩ 𝐹𝑁 ) for any 𝐺 ∈ 𝒢 (just work out the intersection, similar
to the step in the induction....). This shows that on the ∩-stable system
51
R.L. Schilling: Measures, Integrals & Martingales
Remark. The last step shows that we only need the induction for sets from 𝒢 with finite 𝜇-, hence
𝜈-measure. Therefore, the extended discussion on finiteness is actually not needed, if the induction
is only used for the sequences (𝐺𝑖 )𝑖 and (𝐹𝑛 )𝑛 .
■■
Problem 5.8 Solution: Intuition: in two dimensions we have rectangles. Take 𝐼, 𝐼 ′ ∈ 𝒥 . Call the
lower left corner of 𝐼 𝑎 = (𝑎1 , 𝑎2 ), the upper right corner 𝑏 = (𝑏1 , 𝑏2 ), and do the same for 𝐼 ′ using
𝑎′ , 𝑏′ . This defines a rectangle uniquely. We are done, if 𝐼 ∩ 𝐼 ′ = ∅. If not (draw a picture!) then
we get an overlap which can be described by taking the right-and-upper-most of the two lower left
corners 𝑎, 𝑎′ and the left-and-lower-most of the two upper right corners 𝑏, 𝑏′ . That does the trick.
𝑥 = (𝑥1 , … , 𝑥𝑛 ) ∈ 𝐼 ⇐⇒ 𝑥𝑗 ∈ [𝑎𝑗 , 𝑏𝑗 ) ∀𝑗 = 1, 2, … , 𝑛
⇐⇒ 𝑎𝑗 ⩽ 𝑥𝑗 < 𝑏𝑗 ∀𝑗 = 1, 2, … , 𝑛
and the same holds for 𝑥 ∈ 𝐼 ′ (same 𝑥, but 𝐼 ′ —no typo). Clearly 𝑎𝑗 ⩽ 𝑥𝑗 < 𝑏𝑗 , and, at the same
time 𝑎′𝑗 ⩽ 𝑥𝑗 < 𝑏′𝑗 holds exactly if
This shows that 𝐼 ∩ 𝐼 ′ is indeed a ‘rectangle’, i.e. in 𝒥 . This could be an empty set (which happens
if 𝐼 and 𝐼 ′ do not meet).
■■
Problem 5.9 Solution: First we must make sure that 𝑡 ⋅ 𝐵 is a Borel set if 𝐵 ∈ ℬ. We consider first
rectangles 𝐼 = [[𝑎, 𝑏)) ∈ 𝒥 where 𝑎, 𝑏 ∈ R𝑛 . Clearly, 𝑡 ⋅ 𝐼 = [[𝑡𝑎, 𝑡𝑏)) where 𝑡𝑎, 𝑡𝑏 are just the scaled
vectors. So, scaled rectangles are again rectangles, and therefore Borel sets. Now fix 𝑡 > 0 and set
It is not hard to see that ℬ𝑡 is itself a 𝜎-algebra and that 𝒥 ⊂ ℬ𝑡 ⊂ ℬ(R𝑛 ). But then we get
showing that ℬ𝑡 = ℬ(R𝑛 ), i.e. scaled Borel sets are again Borel sets.
52
Solution Manual. Last update 28th October 2021
Now define a new measure 𝜇(𝐵) ∶= 𝜆𝑛 (𝑡 ⋅ 𝐵) for Borel sets 𝐵 ∈ ℬ(R𝑛 ) (which is, because of the
above, well-defined). For rectangles [[𝑎, 𝑏)) we get, in particular,
( )
𝜇[[𝑎, 𝑏)) = 𝜆𝑛 (𝑡 ⋅ [[𝑎, 𝑏)) = 𝜆𝑛 [[𝑡𝑎, 𝑡𝑏))
𝑛
∏ ( )
= (𝑡𝑏𝑗 ) − (𝑡𝑎𝑗 )
𝑗=1
𝑛
∏ ( )
= 𝑡 ⋅ 𝑏𝑗 − 𝑎𝑗
𝑗=1
𝑛
∏ ( )
=𝑡 ⋅𝑛
𝑏𝑗 − 𝑎𝑗
𝑗=1
= 𝑡𝑛 𝜆𝑛 [[𝑎, 𝑏))
which shows that 𝜇 and 𝑡𝑛 𝜆𝑛 coincide on the ∩-stable generator 𝒥 of ℬ(R𝑛 ), hence they’re the
same everywhere. (Mind the small gap: we should make the mental step that for any measure
𝜈 a positive multiple, say, 𝑐 ⋅ 𝜈, is again a measure—this ensures that 𝑡𝑛 𝜆𝑛 is a measure, and we
need this in order to apply Theorem 5.7. Mind also that we need that 𝜇 is finite on all rectangles
(obvious!) and that we find rectangles increasing to R𝑛 , e.g. [−𝑘, 𝑘) × … × [−𝑘, 𝑘) as in the proof
of Theorem 5.8(ii).)
■■
Problem 5.10 Solution: Define 𝜈(𝐴) ∶= 𝜇◦𝜃 −1 (𝐴). Obviously, 𝜈 is again a finite measure. Moreover,
since 𝜃 −1 (𝑋) = 𝑋, we have
■■
Problem 5.11 Solution: The necessity of the condition is trivial since 𝒢 ⊂ 𝜎(𝒢 ) = ℬ, resp., ℋ ⊂
𝜎(ℋ ) = 𝒞 .
Obviously, 𝜇 and 𝜈 are finite measures on ℬ having mass 𝑃 (𝐻) such that 𝜇 and 𝜈 coincide on
the ∩-stable generator 𝒢 ∪ {𝑋} of ℬ. Note that this generator contains the exhausting sequence
𝑋, 𝑋, 𝑋, …. By the uniqueness theorem for measures, Theorem 5.7, we conclude
𝜇 = 𝜈 on the whole of ℬ.
53
R.L. Schilling: Measures, Integrals & Martingales
Then the same argument as before shows that 𝜌 = 𝜏 on 𝒞 and, since 𝐵 ∈ ℬ was arbitrary, the
claim follows.
■■
𝒟 ∶= {𝐴 ∈ 𝒜 ∶ ∀ 𝜖 > 0 ∃ 𝐺 ∈ 𝒢 ∶ 𝜇(𝐴 ▵ 𝐺) ⩽ 𝜖}
is a Dynkin system.
(D2 ) Assume that 𝐴 ∈ 𝒟 . For every 𝜖 > 0 there is some 𝐺 ∈ 𝒢 such that 𝜇(𝐴 ▵ 𝐺) ⩽ 𝜖.
From
𝐴𝑐 ▵ 𝐺𝑐 = (𝐺𝑐 ⧵ 𝐴𝑐 ) ∪ (𝐴𝑐 ⧵ 𝐺𝑐 )
= (𝐺𝑐 ∩ 𝐴) ∪ (𝐴𝑐 ∩ 𝐺)
= (𝐴 ⧵ 𝐺) ∪ (𝐺 ⧵ 𝐴)
= 𝐴▵𝐺
(D3 ) Let (𝐴𝑗 )𝑗∈N ⊂ 𝒟 be a sequence of mutually disjoint sets and 𝜖 > 0. Since 𝜇 is a finite
measure, we get ( )
∑ ⨃
𝜇(𝐴𝑗 ) = 𝜇 𝐴𝑗 < ∞,
𝑗∈N 𝑗∈N
54
Solution Manual. Last update 28th October 2021
𝑁 ∞
⨃ ⨃
⊂ (𝐴𝑗 ∩ 𝐺𝑗 ) ∪
𝑐
𝐴𝑗 .
𝑗=1 𝑗=𝑁+1
Thus,
(( ) ) (𝑁 ∞
)
⨃ ⋃ ⨃
𝜇 𝐴𝑗 ▵𝐺 ⩽𝜇 (𝐴𝑗 ▵ 𝐺𝑗 ) ∪ 𝐴𝑗
𝑗∈N 𝑗=1 𝑗=𝑁+1
𝑁 ∞
∑ ∑ ( )
⩽ 𝜇(𝐴𝑗 ▵ 𝐺𝑗 ) + 𝜇 𝐴𝑗
𝑗=1 𝑗=𝑁+1
𝑁
∑
⩽ 𝜖2−𝑗 + 𝜖
𝑗=1
⩽ 𝜖 + 𝜖.
⨃
Since 𝜖 > 0 is arbitrary, we conclude that 𝑗∈N 𝐴𝑗 ∈ 𝒟.
𝒜 = 𝜎(𝒢 ) = 𝛿(𝒢 ) ⊂ 𝒟 .
we find, just as in (i), that 𝒟 ′ is a Dynkin system. The rest of the proof is as before.
⋃ (⋃ )
(iii) “⇐”: Let 𝐴 ∈ 𝒜 such that 𝐴 ⊂ 𝑛∈N 𝐼𝑛 and 𝜇 𝑛∈N 𝐼𝑛 ⩽ 𝜖. Because of the monotonicity
of measures we get ( )
⋃
𝜇(𝐴) ⩽ 𝜇 𝐼𝑛 ⩽ 𝜖,
𝑛∈N
and so 𝜇(𝐴) = 0.
⋃ ⋃
“⇒”: Set 𝒦 ∶= {𝐴 ⊂ R𝑛 ∶ ∃(𝐼𝑘 )𝑘∈N ⊂ 𝒥 ∶ 𝐴 = 𝑘 𝐼𝑘 or 𝐴𝑐 = 𝑘 𝐼𝑘 } and observe that
𝐼 ∈𝒦 ⇒ 𝐼𝑐 ∈ 𝒦 . Define, furthermore,
𝒟 ∶= {𝐴 ⊂ R𝑛 ∶ ∀ 𝜖∃ 𝐽 , 𝐾 ∈ 𝒦 , 𝐽 ⊂ 𝐴 ⊂ 𝐾, 𝜇(𝐾 ⧵ 𝐽 ) ⩽ 𝜖}.
55
R.L. Schilling: Measures, Integrals & Martingales
(D2 ) Pick 𝐴 ∈ 𝒟 and 𝜖 > 0. Then there are 𝐽 , 𝐾 ∈ 𝒦 such that 𝐽 ⊂ 𝐴 ⊂ 𝐾 and
𝜇(𝐾 ⧵ 𝐽 ) ⩽ 𝜖. From 𝐽 𝑐 , 𝐾 𝑐 ∈ 𝒦 , 𝜇(𝐾 𝑐 ⧵ 𝐽 𝑐 ) = 𝜇(𝐽 ⧵ 𝐾) ⩽ 𝜖 and 𝐽 𝑐 ⊃ 𝐴𝑐 ⊃ 𝐾 𝑐 we
get immediately 𝐴𝑐 ∈ 𝒟 .
(D3 ) Let (𝐴𝑗 )𝑗∈N ⊂ 𝒟 be a sequence of mutually disjoint sets and 𝜖 > 0. Pick 𝐽𝑗 ∈ 𝒦 and
𝐾𝑗 ∈ 𝒦 such that 𝐽𝑗 ⊂ 𝐴𝑗 ⊂ 𝐾𝑗 , 𝜇(𝐾𝑗 ⧵ 𝐽𝑗 ) ⩽ 𝜖2−𝑗 and set
⋃ ⋃
𝐽 ∶= 𝐴𝑗 𝐾 ∶= 𝐾𝑗 .
𝑗∈N 𝑗∈N
⩽ 𝜖.
⨃
Thus, 𝑗 𝐴𝑗 ∈ 𝒟 .
Now let 𝐴 be a set satisfying 𝜇(𝐴) = 0. Therefore, for every 𝜖 > 0 there is a set 𝐾𝜖 = 𝐾 ∈ 𝒦
⋃ ⋃
such that 𝐴 ⊂ 𝐾 and 𝜇(𝐾) < 𝜖. If 𝐾 = 𝑖 𝐼𝑖 , we are done. If 𝐾 𝑐 = 𝑖 𝐼𝑖 we have to argue like
this: Let 𝐽 ∶= 𝐽𝑅 ∶= [−𝑅, 𝑅)𝑑 ∈ 𝒥 . Then
𝑘
⋂ ⋂ ⋂ ⋂⋂
𝐾= 𝐼𝑖𝑐 and 𝐽 ∩ 𝐾 = 𝐼𝑖𝑐 ∩𝐽 = 𝐽 ⧵ 𝐼𝑖 = 𝐽 ⧵ 𝐼𝑖
𝑖 𝑖 𝑖 𝑘 𝑖=1
⋂𝑘
and each set 𝐽 ⧵ 𝐼𝑖 is a finite union of sets from 𝒥 (since 𝒥 is a semiring), hence 𝑖=1 𝐽 ⧵ 𝐼𝑖 is a
finite union of sets from 𝒥 . Since 𝜇(𝐽 ∩ 𝐾) ⩽ 𝜇(𝐾) ⩽ 𝜖, a continuity-of-measure argument shows
⋂ ⋂
that there exists some 𝑘 such that 𝐽 ∩ 𝐾 ⊂ 𝑘𝑖=1 𝐽 ⧵ 𝐼𝑖 and 𝜇( 𝑘𝑖=1 𝐽 ⧵ 𝐼𝑖 ) ⩽ 2𝜖.
If we pick 𝜖 = 𝜖∕2𝑅 , we see that we can cover 𝐴 ∩ [−𝑅, 𝑅)𝑑 by a countable union of 𝒥 -sets, call
their union 𝑈𝑅 , such that 𝜇(𝑈𝑅 ) ⩽ 𝜖∕2𝑅 . Finally,
∑
𝜇(𝐴) ⩽ 𝜇(𝑈𝑅 ) ⩽ 𝜖
𝑅∈N
56
Solution Manual. Last update 28th October 2021
■■
(i) mind the misprint: we also need stability of ℳ under finite intersections. Clearly, any
𝜎-algebra is also a monotone class. Conversely, if ℳ is a monotone class such that 𝑀 ∈
ℳ ⇐⇒ 𝑀 𝑐 ∈ ℳ, then the condition (Σ2 ) holds, while (Σ1 ) is satisfied by the very definition
of a monotone class. If ℳ is also stable under finite intersections, we get 𝑀, 𝑁 ∈ ℳ ⇐⇒
𝑀 ∪ 𝑁 = (𝑀 𝑐 ∩ 𝑁 𝑐 )𝑐 ∈ ℳ, so (Σ3 ) follows from the stability under finite unions and the
stability of monotone classes under increasing limits of sets.
(ii) Since 𝜎(𝒢 ) is a monotone class containing 𝒢 , we have – by minimality – that 𝔪(𝒢 ) ⊂ 𝜎(𝒢 ).
On the other hand, by the monotone class theorem, we get 𝒢 ⊂ 𝔪(𝒢 ) ⇐⇒ 𝜎(𝒢 ) ⊂ 𝔪(𝒢 )
which means that 𝔪(𝒢 ) = 𝜎(𝒢 ).
■■
57
6 Existence of measures.
Solutions to Problems 6.1–6.14
If 𝑥 ⩽ 𝑦 < 0, we have [𝑦, 0) ⊂ [𝑥, 0) and so 0 ⩽ −𝐹𝜇 (𝑦) = 𝜇[𝑦, 0) ⩽ 𝜇[𝑥, 0) = −𝐹𝜇 (𝑥), i.e.
𝐹𝜇 (𝑥) ⩽ 𝐹𝜇 (𝑦) ⩽ 0.
Left-continuity: Let us deal with the case 𝑥 ⩾ 0 only, the case 𝑥 < 0 is analogous (and
even easier). Assume first that 𝑥 > 0. Take any sequence 𝑥𝑘 < 𝑥 and 𝑥𝑘 ↑ 𝑥 as 𝑘 → ∞.
Without loss of generality we can assume that 0 < 𝑥𝑘 < 𝑥. Then [0, 𝑥𝑘 ) ↑ [0, 𝑥) and using
Proposition 4.3 (continuity of measures) implies
If 𝑥 = 0 we must take a sequence 𝑥𝑘 < 0 and we have then [𝑥𝑘 , 0) ↓ [0, 0) = ∅. Again by
Proposition 4.3, now (𝑖𝑖𝑖′ ), we get
We remark that, since for a sequence 𝑦𝑘 ↓ 𝑦, 𝑦𝑘 > 𝑦 we have [0, 𝑦𝑘 ) ↓ [0, 𝑦], and not [0, 𝑦),
we cannot expect right-continuity in general.
(ii) Since 𝒥 = {[𝑎, 𝑏), 𝑎 ⩽ 𝑏} is a semi-ring (cf. the remark preceding Proposition 6.3 or Propos-
ition 6.5) it is enough to check that 𝜈𝐹 is a premeasure on 𝒥 . This again amounts to showing
(M1 ) and (M2 ) relative to 𝒥 (mind you: 𝜈𝐹 is not a measure as 𝒥 is not a 𝜎-algebra....).
(ii) Let 𝑎 ⩽ 𝑏 ⩽ 𝑐 so that [𝑎, 𝑏), [𝑏, 𝑐) ∈ 𝒥 are disjoint sets and [𝑎, 𝑐) = [𝑎, 𝑏) ⊍ [𝑏, 𝑐) ∈ 𝒥
(the latter is crucial). Then we have
59
R.L. Schilling: Measures, Integrals & Martingales
is an open cover of the compact interval [𝑎, 𝑏 − 𝜖]. Thus, there exists a finite open
subcover, hence some 𝑁 ∈ N such that
𝑁 𝑁
⋃ ⋃
(𝑎𝑛 − 𝜖𝑛 , 𝑏𝑛 ) ⊃ [𝑎, 𝑏 − 𝜖] ⇐⇒ [𝑎𝑛 − 𝜖𝑛 , 𝑏𝑛 ) ⊃ [𝑎, 𝑏 − 𝜖).
𝑛=1 𝑛=1
First note that we can de- and increase 𝑎𝑛 ⩾ 𝑎′𝑛 and 𝑏𝑛 ⩽ 𝑏′𝑛 such that
𝑁 𝑁
⨃ ⨃
[𝑎𝑛 , 𝑏𝑛 ) ⊂ [𝑎′𝑛 , 𝑏′𝑛 ) = [𝑎, 𝑏)
𝑛=1 𝑛=1
Now we choose 𝜖 and 𝜖𝑛 . For any given 𝜂 > 0 we can find 𝜖 > 0 and 𝜖𝑛 > 0 such that
𝜂
𝜈𝐹 [𝑏 − 𝜖, 𝑏) = 𝐹 (𝑏) − 𝐹 (𝑏 − 𝜖) ⩽
2
𝜂
and 𝜈𝐹 [𝑎𝑛 − 𝜖𝑛 , 𝑎𝑛 ) = 𝐹 (𝑎𝑛 ) − 𝐹 (𝑎𝑛 − 𝜖𝑛 ) ⩽ 2−𝑛
2
here we use the left-continuity of 𝐹 . Thus,
𝑁 𝑁
∑ 𝜂 ∑ −𝑛 𝜂
0 ⩽ 𝜈𝐹 [𝑎, 𝑏) − 𝜈𝐹 [𝑎𝑛 , 𝑏𝑛 ) ⩽ + 2 ⩽ 𝜂.
𝑛=1
2 𝑛=1 2
60
Solution Manual. Last update 28th October 2021
This means that we find at least one extension. Uniqueness follows since
(iii) Now let 𝜇 be a measure with 𝜇[−𝑛, 𝑛) < ∞. The latter means that the function 𝐹𝜇 (𝑥),
as defined in part (i), is finite for every 𝑥 ∈ R. Now take this 𝐹𝜇 and define, as in (ii) a
(uniquely defined) measure 𝜈𝐹𝜇 . Let us see that 𝜇 = 𝜈𝐹𝜇 . For this, it is enough to show
equality on the sets of type [𝑎, 𝑏) (since such sets generate the Borel sets and the uniqueness
theorem applies....)
If 0 ⩽ 𝑎 ⩽ 𝑏,
If 𝑎 ⩽ 𝑏 ⩽ 0,
= 𝜇[𝑎, 0) − 𝜇[𝑏, 0)
( )
= 𝜇 [𝑎, 0) ⧵ [𝑏, 0)
= 𝜇[𝑎, 𝑏) ✓
If 𝑎 ⩽ 0 ⩽ 𝑏,
61
R.L. Schilling: Measures, Integrals & Martingales
= 𝜇({𝑥}) = 0
which means that 𝐹 (𝑥) = 𝐹 (𝑥+) (𝑥+ indicates the right limit), i.e. 𝐹 is right-continuous at
𝑥, hence continuous, as 𝐹 is left-continuous anyway.
■■
62
Solution Manual. Last update 28th October 2021
■■
Problem 6.3 Solution: We know already that ℬ[0, ∞) is a 𝜎-algebra (it is a trace 𝜎-algebra) and, by
definition,
{ }
Σ = 𝐵 ∪ (−𝐵) ∶ 𝐵 ∈ ℬ[0, ∞)
Since the structure 𝐵 ∪ (−𝐵) is stable under complementation and countable unions it is clear that
Σ is indeed a 𝜎-algebra.
One possibility to extend 𝜇 defined on Σ would be to take 𝐵 ∈ ℬ(R) and define 𝐵 + ∶= 𝐵 ∩ [0, ∞)
and 𝐵 − ∶= 𝐵 ∩ (−∞, 0) and to set
which is obviously a measure. We cannot expect uniqueness of this extension since Σ does not
generate ℬ(R)—not all Borel sets are symmetric.
■■
(i) Assume first that 𝜇∗ (𝑄) < ∞. By the definition of the infimum we find for every 𝜖 > 0 a
⋃
sequence (𝐵𝑗𝜖 )𝑗∈N ⊂ 𝒜 such that 𝐵 𝜖 ∶= 𝑗 𝐵𝑗𝜖 ⊃ 𝑄 and, because of 𝜎-subadditivity,
∑
𝜇(𝐵 𝜖 ) − 𝜇∗ (𝑄) ⩽ 𝜇(𝐵𝑗𝜖 ) − 𝜇∗ (𝑄) ⩽ 𝜖.
𝑗
⋂
Set 𝐵 ∶= 𝑘𝐵
1∕𝑘 ∈ 𝒜 . Then 𝐵 ⊃ 𝑄 and 𝜇(𝐵) = 𝜇 ∗ (𝐵) = 𝜇 ∗ (𝑄).
𝐵 ⧵ 𝑁 ⊃ 𝐵 ⧵ (𝐵 ⧵ 𝑄) = 𝐵 ∩ [(𝐵 ∩ 𝑄𝑐 )𝑐 ] = 𝐵 ∩ [𝐵 𝑐 ∪ 𝑄]
=𝐵∩𝑄
= 𝑄.
So,
If 𝜇∗ (𝑄) = ∞, we take the exhausting sequence (𝐴𝑘 )𝑘∈N ⊂ 𝒜 with 𝐴𝑘 ↑ 𝑋 and 𝜇(𝐴𝑘 ) < ∞
and set 𝑄𝑘 ∶= 𝐴𝑘 ∩ 𝑄 for every 𝑘 ∈ N. By the first part we can find sets 𝐵𝑘 ∈ 𝒜 with
63
R.L. Schilling: Measures, Integrals & Martingales
𝑁𝑘 = 𝑁 ∩ 𝐵𝑘 ⊂ (𝐵 ⧵ 𝑄) ∩ 𝐵𝑘 = 𝐵𝑘 ⧵ 𝑄 = 𝐵𝑘 ⧵ 𝑄𝑘 .
∑
Thus 𝜇(𝑁𝑘 ) = 0 and, by 𝜎-subadditivity, 𝜇(𝑁) ⩽ ∞ 𝑘=1 𝜇(𝑁𝑘 ) = 0.
(ii) Define 𝜇̄ ∶= 𝜇 ∗ | ∗ . We know from Theorem 6.1 that 𝜇̄ is a measure on 𝒜 ∗ and, because of
|
|𝒜
the monotonicity of 𝜇 ∗ , we know that for all 𝑁 ∗ ∈ 𝒜 ∗ with 𝜇(𝑁
̄ ∗ ) we have
∀ 𝑀 ⊂ 𝑁 ∗ ∶ 𝜇 ∗ (𝑀) ⩽ 𝜇∗ (𝑁 ∗ ) = 𝜇(𝑁
̄ ∗ ) = 0.
∀ 𝑄 ⊂ 𝑋 ∶ 𝜇 ∗ (𝑄) = 𝜇∗ (𝑄 ∩ 𝑀) + 𝜇∗ (𝑄 ⧵ 𝑀).
𝒜 ∗ = {𝐴 ∪ 𝑁 ∶ 𝐴 ∈ 𝒜 , 𝑁 ∈ 𝔑} (*)
We are going to use both equalities and show ‘⊃’ in (∗) and ‘⊂’ in (∗∗) (which is enough
since, cf. Problem 4.15 asserts the equality of the right-hand sides of (∗), (∗∗)!).
‘⊃’: By part (ii), subsets of 𝒜 -null sets are in 𝒜 ∗ so that every set of the form 𝐴 ∪ 𝑁 with
𝐴 ∈ 𝒜 and 𝑁 being a subset of an 𝒜 null set is in 𝒜 ∗ .
64
Solution Manual. Last update 28th October 2021
‘⊂’: By part (i) we find for every 𝐴∗ ∈ 𝒜 ∗ some 𝐴 ∈ 𝒜 such that 𝐴 ⊃ 𝐴∗ and 𝐴⧵𝐴∗ is an 𝒜 ∗
null set. By the same argument we get 𝐵 ∈ 𝒜 , 𝐵 ⊃ (𝐴∗ )𝑐 and 𝐵 ⧵ (𝐴∗ )𝑐 = 𝐵 ∩ 𝐴∗ = 𝐴∗ ⧵ 𝐵 𝑐
is an 𝒜 ∗ null set. Thus,
𝐵 𝑐 ⊂ 𝐴∗ ⊂ 𝐴
and
( ) ( ) ( ) ( )
𝐴 ⧵ 𝐵 𝑐 ⊂ 𝐴 ⧵ 𝐴∗ ∪ 𝐴∗ ⧵ 𝐵 𝑐 = 𝐴 ⧵ 𝐴∗ ∪ 𝐵 ⧵ (𝐴∗ )𝑐
■■
Problem 6.5 Solution: Since, by assumption, 𝑚 is an additive set function such that 0 ⩽ 𝑚(𝑋) ⩽
𝜇(𝑋) < ∞, it is enough to show (cf. Lemma 4.9) that 𝑚 is continuous at ∅ and 𝑚(∅) = 0.
Remark. In order to be self-contained, let us check that any additive set function 𝑚 on a Boolean
algebra ℬ is a pre-measure (i.e. sigma-additive) if it is continuous at ∅:
⋃
Let (𝐵𝑛 )𝑛∈N ⊂ ℬ be a sequence of mutually disjoint sets and 𝐵 ∶= 𝑛∈N 𝐵𝑛 ∈ ℬ. From
𝐵1 ⊍ … ⊍ 𝐵𝑛 ∈ ℬ we get
𝐴𝑛 ∶= 𝐵 ⧵ (𝐵1 ⊍ … ⊍ 𝐵𝑛 ) = 𝐵 ∩ (𝐵1 ⊍ … ⊍ 𝐵𝑛 )𝑐 ∈ ℬ.
⏟⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏟
∈ℬ
■■
(i) A little geometry first: a solid, open disk of radius 𝑟, centre 0 is the set 𝐵𝑟 (0) ∶= {(𝑥, 𝑦) ∈
R2 ∶ 𝑥2 +𝑦2 < 𝑟2 }. The 𝑛-dimensional analogue is clearly {𝑥 ∈ R𝑛 ∶ 𝑥21 +𝑥22 +…+𝑥2𝑛 < 𝑟2 }
(including 𝑛 = 1 where it reduces to an interval). We want to inscribe a box into a ball.
65
R.L. Schilling: Measures, Integrals & Martingales
[ )
Claim: 𝑄𝜖 (0) ∶= ⊂ 𝐵2𝜖 (0). Indeed,
𝑛 𝜖 𝜖
× − ,
√ √
𝑗=1 𝑛 𝑛
𝜖2 𝜖2 𝜖2
𝑥 ∈ 𝑄𝜖 (0) ⇐⇒ 𝑥21 + 𝑥22 + … + 𝑥2𝑛 ⩽ + +…+ < (2𝜖)2
𝑛 𝑛 𝑛
⇐⇒ 𝑥 ∈ 𝐵2𝜖 (0),
(ii) For closed sets this is, in general, wrong. Trivial counterexample: the singleton {0} is closed,
it is Borel (take a countable sequence of nested rectangles, centered at 0 and going down to
{0}) and the Lebesgue measure is zero.
To get strictly positive Lebesgue measure, one possibility is to have interior points, i.e. closed
sets which have non-empty interior do have positive Lebesgue measure.
■■
⋃
(ii) The hint says it all: 𝐻 is contained in the union 𝑦 + 𝑘∈N 𝐴𝑘 for some 𝑦 and we have
𝜆2 (𝐴𝑘 ) = (2𝜖 2−𝑘 )⋅(2𝑘) = 4⋅𝜖⋅𝑘2−𝑘 . Using the 𝜎-subadditivity and monotonicity of measures
(the 𝐴𝑘 ’s are clearly not disjoint) as well as the translational invariance of the Lebesgue
measure we get
( ) ∞ ∞
∞ ∑ ∑
2 2
0 ⩽ 𝜆 (𝐻) ⩽ 𝜆 ∪ 𝐴𝑘 ⩽ 𝜆(𝐴𝑘 ) = 4 ⋅ 𝜖 ⋅ 𝑘2−𝑘 = 𝐶𝜖
𝑘=1
𝑘=1 𝑘=1
∑∞
where 𝐶 is the finite (!) constant 4 𝑘=1 𝑘2
−𝑘 (check convergence!). As 𝜖 was arbitrary, we
can let it → 0 and the claim follows.
(iii) 𝑛-dimensional version of (i): We have 𝐼 = × (𝑎𝑗 , 𝑏𝑗 ). Set 𝐼𝑘 ∶= × [𝑎𝑗 + 𝑘1 , 𝑏𝑗 ). Then 𝐼𝑘 ↑ 𝐼
𝑛 𝑛
𝑗=1 𝑗=1
as 𝑘 → ∞ and we have (write 𝜆 = 𝜆𝑛 , for short)
𝑛 ( ) ∏ 𝑛
∏ 1 ( )
𝜆(𝐼) = lim 𝜆(𝐼𝑘 ) = lim 𝑏𝑗 − 𝑎𝑗 − = 𝑏𝑗 − 𝑎𝑗 .
𝑘→∞ 𝑘→∞
𝑗=1
𝑘 𝑗=1
66
Solution Manual. Last update 28th October 2021
𝑛-dimensional version of (ii): The changes are obvious: 𝐴𝑘 = [−𝜖2−𝑘 , 𝜖2−𝑘 ) × [−𝑘, 𝑘)𝑛−1
∑
and 𝜆𝑛 (𝐴𝑘 ) = 2𝑛 ⋅ 𝜖 ⋅ 2−𝑘 ⋅ 𝑘𝑛−1 . The rest stays as before, since the sum ∞𝑘=1 𝑘
𝑛−1 2−𝑘 still
= 𝑃 ({𝑦1 , … , 𝑦𝑁 }) ⩽ 𝑃 (R) = 1
so 𝑁
𝑘
⩽ 1, i.e. 𝑁 ⩽ 𝑘, and the claim in the hint (about the maximal number of atoms of given
size) is shown.
Now denote, as in the hint, the atoms with measure of size [ 𝑘1 , 𝑘−1
1
) by 𝑦(𝑘)
1
, … 𝑦(𝑘)
𝑁(𝑘)
where
𝑁(𝑘) ⩽ 𝑘 is their number. Since
⋃ [1 )
, 1
𝑘 𝑘−1
= (0, ∞)
𝑘∈N
(no matter whether the sum is over a finite or countably infinite set of 𝑗’s) is indeed a measure
on R. But more is true: for any Borel set 𝐴
∑
𝜈(𝐴) = 𝑃 ({𝑥𝑗 }) ⋅ 𝛿𝑥𝑗 (𝐴)
𝑗
67
R.L. Schilling: Measures, Integrals & Martingales
∑
= 𝑃 ({𝑥𝑗 })
𝑗∶𝑥𝑗 ∈𝐴
showing that 𝜇(𝐴) ∶= 𝑃 (𝐴) − 𝜈(𝐴) is a positive number for each Borel set 𝐴 ∈ ℬ. This
means that 𝜇 ∶ ℬ → [0, ∞]. Let us check 𝑀1 and 𝑀2 . Using 𝑀1 , 𝑀2 for 𝑃 and 𝜈 (for them
they are clear, as 𝑃 , 𝜈 are measures!) we get
which is 𝑀2 for 𝜇.
■■
68
Solution Manual. Last update 28th October 2021
∑
= 2 ⋅ 2−𝑘 ⋅ 2 ⋅ 𝑘
𝑘∈N
∑
=4 𝑘 ⋅ 2−𝑘 < ∞.
𝑘∈N
It remains to check that an open rectangle is an open set. For this take any open rectangle
𝑅 = (𝑎, 𝑏) × (𝑐, 𝑑) and pick (𝑥, 𝑦) ∈ 𝑅. Then we know that 𝑎 < 𝑥 < 𝑏 and 𝑐 < 𝑦 < 𝑑 and
since we have strict inequalities, we have that the smallest distance of this point to any of the
four boundaries (draw a picture!) ℎ ∶= min{|𝑎 − 𝑥|, |𝑏 − 𝑥|, |𝑐 − 𝑦|, |𝑑 − 𝑦|} > 0. This means
that a square around (𝑥, 𝑦) with side-length 2ℎ is inside 𝑅 and what we’re going to do is to
inscribe into this virtual square an open disk with radius ℎ and centre (𝑥, 𝑦). Since the circle
is again in 𝑅, we are done. The equation for this disk is
Thus,
√
|𝑥′ − 𝑥| ⩽ |𝑥 − 𝑥′ |2 + |𝑦 − 𝑦′ |2 < ℎ
√
and |𝑦′ − 𝑦| ⩽ |𝑥 − 𝑥′ |2 + |𝑦 − 𝑦′ |2 < ℎ
(iii) No, this is impossible. Since we are in one dimension, pathwise connectedness forces us to
go between points in a straight, uninterrupted line. Since the set is unbounded, this means
that we must have a line of the sort (𝑎, ∞) or (−∞, 𝑏) in our set and in both cases Lebesgue
measure is infinite. In all dimensions 𝑛 > 1, see part (ii) for two dimensions, we can, however,
construct pathwise connected, unbounded open sets with finite Lebesgue measure.
■■
Problem 6.10 Solution: Fix 𝜖 > 0 and let {𝑞𝑗 }𝑗∈N be an enumeration of Q ∩ [0, 1]. Then
⋃( )
𝑈 ∶= 𝑈𝜖 ∶= 𝑞𝑗 − 𝜖2−𝑗−1 , 𝑞𝑗 + 𝜖2−𝑗−1 ∩ [0, 1]
𝑗∈N
■■
Problem 6.11 Solution: Assume first that for every 𝜖 > 0 there is some open set 𝑈𝜖 ⊃ 𝑁 such that
𝜆(𝑈𝜖 ) ⩽ 𝜖. Then
69
R.L. Schilling: Measures, Integrals & Martingales
Attention: A construction along the lines of Problem 3.15, hint to part (ii), using open sets 𝑈 𝛿 ∶=
𝑁 + 𝐵𝛿 (0) is, in general not successful:
• it is not clear that 𝑈 𝛿 has finite Lebesgue measure (o.k. one can overcome this by considering
𝑁 ∩ [−𝑘, 𝑘] and then letting 𝑘 → ∞...)
• 𝑈 𝛿 ↓ 𝑁̄ and not 𝑁 (unless 𝑁 is closed, of course). If, say, 𝑁 is a dense set of [0, 1], this
approach leads nowhere.
■■
⋃∞
Problem 6.12 Solution: Observe that the sets 𝐶𝑘 ∶= 𝑗=𝑘 𝐴𝑗 , 𝑘 ∈ N, decrease as 𝑘 → ∞—we
admit less and less sets in the union, i.e. the union becomes smaller. Since 𝑃 is a probability
measure, 𝑃 (𝐶𝑘 ) ⩽ 1 and therefore Lemma 4.9 applies and shows that
(∞ ∞ ) (∞ )
⋂ ⋃ ⋂
𝑃 𝐴𝑗 = 𝑃 𝐶𝑘 = lim 𝑃 (𝐶𝑘 ).
𝑘=1 𝑗=𝑘 𝑘=1 𝑘→∞
■■
Case 1: 𝑎, 𝑏 ∈ [0, 1). Then [0, 1) is the best possible cover of (𝑎, 𝑏), thus 𝜇 ∗ (𝑎, 𝑏) = 𝜇[0, 1) =
1
2
.
Case 2: 𝑎, 𝑏 ∈ [1, 2). Then [1, 2) is the best possible cover of (𝑎, 𝑏), thus 𝜇 ∗ (𝑎, 𝑏) = 𝜇[1, 2) =
1
2
.
70
Solution Manual. Last update 28th October 2021
Case 3: 𝑎 ∈ [0, 1), 𝑏 ∈ [1, 2). Then [0, 1) ⊍ [1, 2) is the best possible cover of (𝑎, 𝑏), thus
𝜇∗ (𝑎, 𝑏) = 𝜇[0, 1) + 𝜇[1, 2) = 1.
And in the case of a singleton {𝑎} the best possible cover is always either [0, 1) or [1, 2) so
that 𝜇 ∗ ({𝑎}) = 1
2
for all 𝑎.
(ii) Assume that (0, 1) ∈ 𝒜 ∗ . Since 𝒜 ⊂ 𝒜 ∗ , we have [0, 1) ∈ 𝒜 ∗ , hence {0} = [0, 1) ⧵ (0, 1) ∈
𝒜 ∗ . Since 𝜇∗ (0, 1) = 𝜇 ∗ ({0}) = 21 , and since 𝜇∗ is a measure on 𝒜 ∗ (cf. Step 4 in the proof
of Theorem 6.1), we get
1 1 1
= 𝜇[0, 1) = 𝜇 ∗ [0, 1) = 𝜇∗ (0, 1) + 𝜇 ∗ {0} = + = 1
2 2 2
leading to a contradiction. Thus neither (0, 1) nor {0} are elements of 𝒜 ∗ .
■■
Problem 6.14 Solution: Since 𝒜 ⊂ 𝒜 ∗ , the only interesting sets (to which one could extend 𝜇) are
those 𝐵 ⊂ R where both 𝐵 and 𝐵 𝑐 are uncountable. By definition,
{∑ ⋃ }
𝛾 ∗ (𝐵) = inf 𝛾(𝐴𝑗 ) ∶ 𝐴𝑗 ∈ 𝒜 , 𝐴𝑗 ⊃ 𝐵 .
𝑗 𝑗
The infimum is obviously attained for 𝐴𝑗 = R, so that 𝛾 ∗ (𝐵) = 𝛾 ∗ (𝐵 𝑐 ) = 1. On the other hand,
since 𝛾 ∗ is necessarily additive on 𝒜 ∗ , the assumption that 𝐵 ∈ 𝒜 ∗ leads to a contradiction:
Thus, 𝒜 = 𝒜 ∗ .
■■
71
7 Measurable mappings.
Solutions to Problems 7.1–7.13
Problem 7.1 Solution: We have 𝜏𝑥−1 (𝑧) = 𝑧 + 𝑥. According to Lemma 7.2 we have to check that
Problem 7.2 Solution: We had Σ′ = {𝐴′ ⊂ 𝑋 ′ ∶ 𝑇 −1 (𝐴′ ) ∈ 𝒜 } where 𝒜 was a 𝜎-algebra of subsets
of 𝑋. Let us check the properties (Σ1 )–(Σ3 ).
(Σ1 ) Take ∅ ⊂ 𝑋 ′ . Then 𝑇 −1 (∅) = ∅ ∈ 𝒜 , hence ∅ ∈ Σ′ .
( )𝑐
(Σ2 ) Take any 𝐵 ∈ Σ′ . Then 𝑇 −1 (𝐵) ∈ 𝒜 and therefore 𝑇 −1 (𝐵 𝑐 ) = 𝑇 −1 (𝐵) ∈ 𝒜 since all
set operations interchange with inverse maps and since 𝒜 is a 𝜎-algebra. This shows that
𝐵 𝑐 ∈ Σ′ .
(Σ3 ) Take any sequence (𝐵𝑗 )𝑗∈N ⊂ Σ′ . Then, using again the fact that 𝒜 is a 𝜎-algebra, 𝑇 −1 (∪𝑗 𝐵𝑗 ) =
⋃ −1 ⋃
𝑗 𝑇 (𝐵𝑗 ) ∈ 𝒜 which proves that 𝑗 𝐵𝑗 ∈ Σ .
′
■■
73
R.L. Schilling: Measures, Integrals & Martingales
On the other hand, 𝑇 −1 is not measurable: the set 𝐴 = {𝑘; 𝑘 ⩽ 0} is contained in 𝒜 , but
𝑇 (𝐴) = {𝑘 ∶ 𝑘 ⩽ 2} ∉ 𝒜 (use 2 = 2 ⋅ 1 ∈ 𝐴, but 2 ⋅ 1 + 1 = 3 ∉ 𝐴).
■■
(i) First of all we remark that 𝑇𝑖−1 (𝒜𝑖 ) is itself a 𝜎-algebra, cf. Example 3.3(vii).
(ii) From part (i) we know that 𝜎(𝑇𝑖 , 𝑖 ∈ 𝐼) necessarily contains 𝑇𝑖−1 (𝒜𝑖 ) for every 𝑖 ∈ 𝐼. Since
⋃ −1 (⋃ )
𝑇
𝑖 𝑖 (𝒜𝑖 ) is, in general, not a 𝜎-algebra, we have 𝜎 𝑇
𝑖 𝑖
−1 (𝒜 ) ⊂ 𝜎(𝑇 , 𝑖 ∈ 𝐼). On the
𝑖 𝑖
⋃ −1
other hand, each 𝑇𝑖 is, because of 𝑇𝑖 (𝒜𝑖 ) ⊂ 𝑖 𝑇𝑖 (𝒜𝑖 ) ⊂ 𝜎(𝑇𝑖 , 𝑖 ∈ 𝐼) measurable w.r.t.
−1
(⋃ )
𝜎 𝑇 −1 (𝒜 ) and this proves the claim.
𝑖 𝑖 𝑖
■■
(i), (ii)
Since an indicatior function can only assume the values 0 and 1, the claimed equality
follows for the value 0 by negating the previously shown equivalence.
(iii) “⇒”: Assume that 𝑇 is measurable. We have 𝑇 −1 (𝐴′ ) ∈ 𝒜 ∀𝐴′ ∈ 𝒜 ′ and since 𝒜 is
a 𝜎-algebra, we conclude
𝑇 −1 (𝐴′ ) ∈ 𝒜 ∀𝐴′ ∈ 𝒜 ′ ,
i.e., 𝑇 is measurable.
(iii) Theorem 7.6 shows that image measures are measures. By the definition of 𝑇 , we have
𝑇 −1 (𝐸 ′ ) = 𝐸 and 𝜈◦𝑇 −1 (𝐸 ′ ) < ∞, resp., 𝜈◦𝑇 −1 (𝐸 ′ ) = 1 follows from the definition
of image measures.
The image measure obtained from a 𝜎-finite measure need not be 𝜎-finite!
74
Solution Manual. Last update 28th October 2021
■■
For the converse consider 𝑇 ∶ (𝑋, 𝜎(𝑇 −1 (𝒢 ))) → (𝑌 , 𝜎(𝒢 )). By the very choice of the 𝜎-algebras
and since 𝑇 −1 (𝒢 ) ⊂ 𝜎(𝑇 −1 (𝒢 )) we find that 𝑇 is 𝜎(𝑇 −1 (𝒢 ))∕𝜎(𝒢 ) measurable—mind that we
only have to check measurability at a generator (here: 𝒢 ) in the image region. Thus,
Alternative: We have
For the converse, set Σ ∶= {𝐺 ∈ 𝜎(𝒢 ) ∶ 𝑇 −1 (𝐺) ∈ 𝜎(𝑇 −1 (𝒢 ))}. It is not hard to see that Σ is
itself a 𝜎-algebra and that 𝒢 ⊂ Σ ⊂ 𝜎(𝒢 ). Thus, 𝜎(𝒢 ) = Σ and so 𝑇 −1 (𝜎(𝒢 )) ⊂ 𝜎(𝑇 −1 (𝒢 )).
■■
Now
( )
∀ 𝑖 ∈ 𝐼 ∶ (𝑇𝑖 ◦𝑓 )−1 (𝒜𝑖 ) ⊂ ℱ ⇐⇒ ∀ 𝑖 ∈ 𝐼 ∶ 𝑓 −1 𝑇𝑖−1 (𝒜𝑖 ) ⊂ ℱ
(⋃ )
⇐⇒ 𝑓 −1 𝑇𝑖−1 (𝒜𝑖 ) ⊂ ℱ
𝑖∈𝐼
[ (⋃ )]
(∗)
−1 −1
⇐⇒ 𝜎 𝑓 𝑇𝑖 (𝒜𝑖 ) ⊂ ℱ
𝑖∈𝐼
(∗∗) ( [⋃ ])
⇐⇒ 𝑓 −1 𝜎 𝑇𝑖−1 (𝒜𝑖 ) ⊂ ℱ.
𝑖∈𝐼
Only (*) and (**) are not immediately clear. The direction ‘⇐⇐’ in (*) is trivial, while ‘ ⇐⇒ ’ follows
if we observe that the right-hand side, ℱ , is a 𝜎-algebra. The equivalence (**) is another case of
Problem 7.6 (see there for the solution!).
■■
Problem 7.8 Solution: Using the notation of the foregoing Problem 7.7 we put
75
R.L. Schilling: Measures, Integrals & Martingales
Since each 𝜋𝑗 is continuous, we have 𝜎(𝜋1 , … , 𝜋𝑚 ) ⊂ ℬ(R𝑚 ) so that Problem 7.7 applies and
proves
𝑓 is ℬ(R𝑚 )-measurable ⇐⇒
𝑓𝑗 = 𝜋𝑗 ◦𝑓 is ℬ(R)-measurable for all 𝑗 = 1, 2, … , 𝑚.
Remark. We will see, in fact, in Chapter 14 (in particular in Theorem 14.17) that we have the
equality 𝜎(𝜋1 , … , 𝜋𝑚 ) = ℬ(R𝑚 ).
■■
Problem 7.9 Solution: In general the direct image 𝑇 (𝒜 ) of a 𝜎-algebra is not any longer a 𝜎-algebra.
(Σ1 ) and (Σ3 ) hold, but (Σ2 ) will, in general, fail. Here is an example: Take 𝑋 = 𝑋 ′ = N, take
any 𝜎-algebra 𝒜 other than {∅, N} in N, and let 𝑇 ∶ N → N, 𝑇 (𝑗) = 1 be the constant map. Then
𝑇 (∅) = ∅ but 𝑇 (𝐴) = {1} whenever 𝐴 ≠ ∅. Thus, {1} = 𝑇 (𝐴𝑐 ) ≠ [𝑇 (𝐴)]𝑐 = N ⧵ {1} but equality
would be needed if 𝑇 (𝒜 ) were a 𝜎-algebra. This means that Σ2 fails.
Necessary and sufficient for 𝑇 (𝒜 ) to be a 𝜎-algebra is, clearly, that 𝑇 −1 is a measurable map
𝑇 −1 ∶ 𝑋 ′ → 𝑋.
Warning. Direct images of measurable sets behave badly – even if the mapping is good. For
example, the continuous (direct) image of a Borel set need not be Borel! (It is, however, analytic
or Souslin).
■■
𝑡 ⋅ 𝐵 = 𝑚−1
1∕𝑡
(𝐵)
𝑗=1
from the uniqueness theorem for measures.
■■
The crunching points in this argument are the steps (#) and (##).
76
Solution Manual. Last update 28th October 2021
(#) This is o.k. since 𝐹 was continuous, and the intermediate value theorem for continuous
functions tells us that intervals are mapped to intervals. So, no problem here, just a little
thinking needed.
(##) This is more subtle. We have defined image measures only for inverse maps, i.e. for
expressions of the type 𝜆1 ◦𝐺−1 where 𝐺 was measurable. So our job is to see that 𝐹 can
be obtained in the form 𝐹 = 𝐺−1 where 𝐺 is measurable. In other words, we have to
invert 𝐹 . The problem is that we need to understand that, if 𝐹 (𝑥) is flat on some interval
(𝑎, 𝑏) inversion becomes a problem (since then 𝐹 −1 has a jump—horizontals become
verticals in inversions, as inverting is somehow the mirror-image w.r.t. the 45-degree
line in the coordinate system.).
So, if there are no flat bits, then this means that 𝐹 is strictly increasing, and it is clear
that 𝐺 exists and is even continuous there.
If we have a flat bit, let’s say exactly if 𝑥 ∈ [𝑎, 𝑏] and call 𝐹 (𝑥) = 𝐹 (𝑎) = 𝐹 (𝑏) = 𝐶
for those 𝑥; clearly, 𝐹 −1 jumps at 𝐶 and we must see to it that we take a version of 𝐹 −1 ,
say one which makes 𝐹 −1 left-continuous at 𝐶—note that we could assign any value
from [𝑎, 𝑏] to 𝐹 −1 (𝐶)—which is accomplished by setting 𝐹 −1 (𝐶) = 𝑎. (Draw a graph
to illustrate this!)
𝑦0 ∈ {𝐺 ⩾ 𝜆} ⇐⇒ 𝐺(𝑦0 ) ⩾ 𝜆
def
⇐⇒ inf {𝐹 ⩾ 𝑦0 } ⩾ 𝜆
(‡)
⇐⇒ 𝐹 (𝜆) ⩽ 𝑦0
⇐⇒ 𝑦0 ∈ [𝐹 (𝜆), ∞).
{𝐺 ⩾ 𝜆} = [𝐹 (𝜆), ∞) ∈ ℬ(R)
which shows that 𝐺 is measurable. Even more: it shows that 𝐺−1 (𝑥) ∶= inf {𝐺 ⩾ 𝜆} =
𝐹 (𝑥). Thus, 𝜆1 ◦𝐹 = 𝜆1 ◦𝐺−1 = 𝜇 is indeed an image measure of 𝜆1 .
(ii) We have 𝐹 (𝑥) = 𝐹𝛿0 (𝑥) = 1(0,∞) (𝑥) and its left-continuous inverse 𝐺(𝑦) in the sense of part
77
R.L. Schilling: Measures, Integrals & Martingales
(i) is given by
⎧
⎪+∞, 𝑦>1
⎪
𝐺(𝑦) = ⎨0, 0<𝑦⩽1.
⎪
⎪−∞, 𝑦⩽0
⎩
This function is clearly measurable (use ℬ̄ to accommodate ±∞) and so the claim holds in
this case. Observe that in this case 𝐹 is not any longer continuous but only left-continuous.
■■
78
Solution Manual. Last update 28th October 2021
(vi) Fix 𝑛 and let 𝑘 = 0, 1, 2, … , 3𝑛−1 − 1. We saw in (c) that at step 𝑛 we remove the intervals
𝐹𝑛 , i.e. the intervals of the form
( ) ( )
3𝑘 + 1 3𝑘 + 2
, = 0. ∗∗∗ … ∗ 1 000 … , 0. ∗∗∗ … ∗ 2 000 …
3𝑛 3𝑛 ⏟⏞⏞⏞⏟⏞⏞⏞⏟ ⏟⏞⏞⏞⏟⏞⏞⏞⏟
𝑛 𝑛
where we use the ternary representation of 𝑥. These are exactly the numbers in [0, 1] whose
ternary expansion has a 1 at the 𝑛th digit. As 0. ∗∗∗ … ∗ 1 = 0. ∗∗∗ … ∗ 022222 … has
two representations, the left endpoint stays in. Since we do this for every step 𝑛 ∈ N, the
claim follows.
(vii) Take 𝑡 ∈ 𝐶 with ternary representation 𝑡 = 0.𝑡1 𝑡2 𝑡3 … 𝑡𝑗 …, 𝑡𝑗 ∈ {0, 2} and map it to the
binary number 𝑏 = 0. 21 with digits 𝑏𝑗 = ∈ {0, 1}. This gives a bijection between
𝑡 𝑡2 𝑡3 𝑡𝑗 𝑡𝑗
2 2
… 2 2
𝐶 and [0, 1], i.e. both have ‘as infinitely many’ points, i.e. #𝐶 = #[0, 1]. Despite of that
which is, by the way, another proof for the fact that 𝜎-additivity for the Lebesgue measure
does not extend to general uncountable unions.
■■
∀𝐸 ∈ ℰ ∶ 𝐸 ∪ ∅ ∈ ℰ ⋓ ℱ ⇐⇒ ℰ ⊂ ℰ ⋓ ℱ
and
∀𝐹 ∈ ℱ ∶ ∅ ∪ 𝐹 ∈ ℰ ⋓ ℱ ⇐⇒ ℱ ⊂ ℰ ⋓ ℱ
ℰ ∪ ℱ = {∅, 𝐴, 𝐵, 𝐴𝑐 , 𝐵 𝑐 , 𝑋}
while
ℰ ⋓ ℱ = {∅, 𝐴, 𝐵, 𝐴𝑐 , 𝐵 𝑐 , 𝐴 ∪ 𝐵, 𝐴𝑐 ∪ 𝐵 𝑐 , 𝐴 ∪ 𝐵 𝑐 , 𝐴𝑐 ∪ 𝐵, 𝑋}.
79
R.L. Schilling: Measures, Integrals & Martingales
so that
which proves
■■
80
8 Measurable functions.
Solutions to Problems 8.1–8.26
Problem 8.1 Solution: We remark, first of all, that {𝑢 ⩾ 𝛼} = 𝑢−1 ([𝑥, ∞)) and, similarly, for the
other sets. Now assume that {𝑢 ⩾ 𝛽} ∈ 𝒜 for all 𝛽. Then
( ))
⋃[ 1
−1 −1
{𝑢 > 𝛼} = 𝑢 ((𝛼, ∞)) = 𝑢 𝛼 + 𝑘, ∞
𝑘∈N
⋃ ([ ))
= 𝑢 −1
𝛼 + 𝑘1 , ∞
𝑘∈N
⋃
= {𝑢 ⩾ 𝛼 + 𝑘1 } ∈ 𝒜
𝑘∈N ⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
by assumption ∈ 𝒜
since 𝒜 is a 𝜎-algebra.
Conversely, assume that {𝑢 > 𝛽} ∈ 𝒜 for all 𝛽. Then
( ))
⋂( 1
−1 −1
{𝑢 ⩾ 𝛼} = 𝑢 ([𝛼, ∞)) = 𝑢 𝛼 − 𝑘, ∞
𝑘∈N
⋂ (( ))
= 𝑢−1 𝛼 − 𝑘1 , ∞
𝑘∈N
⋂
= {𝑢 > 𝛼 − 𝑘1 } ∈ 𝒜 .
𝑘∈N ⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
by assumption ∈ 𝒜
we have that {𝑢 > 𝛼} ∈ 𝒜 if, and only if, {𝑢 ⩽ 𝛼} ∈ 𝒜 and the same holds for the sets {𝑢 ⩾
𝛼}, {𝑢 < 𝛼}.
■■
Problem 8.2 Solution: Recall that 𝐵 ∗ ∈ ℬ if, and only if 𝐵 ∗ = 𝐵 ∪ 𝐶 where 𝐵 ∈ ℬ and 𝐶 is any of
the following sets: ∅, {−∞}, {∞}, {−∞, ∞}. Using the fact that ℬ is a 𝜎-algebra and using this
notation (that is: ℬ-sets carry an asterisk ∗ ) we see
(Σ1 ) Take 𝐵 = ∅ ∈ ℬ, 𝐶 = ∅ to see that ∅∗ = ∅ ∪ ∅ ∈ ℬ;
(Σ2 ) Let 𝐵 ∗ ∈ ℬ. Then (complements are to be taken in ℬ
(𝐵 ∗ )𝑐 = (𝐵 ∪ 𝐶)𝑐
81
R.L. Schilling: Measures, Integrals & Martingales
= 𝐵𝑐 ∩ 𝐶 𝑐
= (R ⧵ 𝐵) ∩ (R ⧵ 𝐶)
= (R ⧵ 𝐵 ∪ {−∞, +∞}) ∩ (R ⧵ 𝐶)
= ((R ⧵ 𝐵) ∩ (R ⧵ 𝐶)) ∪ ({−∞, +∞} ∩ (R ⧵ 𝐶))
= (R ⧵ 𝐵) ∪ ({−∞, +∞} ∩ (R ⧵ 𝐶))
which is again of the type ℬ-set union a set of the list ∅, {−∞}, {∞}, {−∞, ∞}, hence it is
in ℬ.
(Σ3 ) Let 𝐵𝑛∗ ∈ ℬ and 𝐵𝑛∗ = 𝐵𝑛 ∪ 𝐶𝑛 . Then
⋃ ⋃ ⋃ ⋃
𝐵∗ = 𝐵𝑛∗ = (𝐵𝑛 ∪ 𝐶𝑛 ) = 𝐵𝑛 ∪ 𝐶𝑛 = 𝐵 ∪ 𝐶
𝑛∈N 𝑛∈N 𝑛∈N 𝑛∈N
with 𝐵 ∈ ℬ and 𝐶 from the list ∅, {−∞}, {∞}, {−∞, ∞}, hence 𝐵 ∗ ∈ ℬ.
A problem is the notation ℬ = ℬ(R). While the left-hand side can easily be defined by (8.5),
ℬ(R) has a well-defined meaning as the (topological) Borel 𝜎-algebra over the set R, i.e. the 𝜎-
algebra in R which is defined via the open sets in R. To describe the open sets 𝒪(R) of R we
use require, that each point 𝑥 ∈ 𝑈 ∗ ∈ 𝒪(R) admits an open neighbourhood 𝐵(𝑥) inside 𝑈 ∗ . If
𝑥 ≠ ±∞, we take 𝐵(𝑥) as the usual open 𝜖-interval around 𝑥 with 𝜖 > 0 sufficiently small. If
𝑥 = ±∞ we take half-lines [−∞, 𝑎) or (𝑏, +∞] respectively with |𝑎|, |𝑏| sufficiently large. Thus,
𝒪(R) adds to 𝒪(R) a few extra sets and open sets are therefore of the form 𝑈 ∗ = 𝑈 ∪ 𝐶 with
𝑈 ∈ 𝒪(R) and 𝐶 being of the form [−∞, 𝑎) or (𝑏, +∞] or ∅ or R or unions thereof.
Thus, 𝒪(R) = R ∩ 𝒪(R) and therefore
ℬ(R) = R ∩ ℬ(R)
82
Solution Manual. Last update 28th October 2021
■■
Problem 8.4 Solution: Since 1𝐵 is ℬ-measurable if, and only if, 𝐵 ∈ ℬ the claim follows by taking
𝐵 ∈ ℬ such that 𝐵 ∉ 𝒜 (this is possible as ℬ ⊊ 𝒜 .
■■
∑𝑁
Problem 8.5 Solution: By definition, 𝑓 ∈ if it is a step-function of the form 𝑓 = 𝑗=0 𝑎𝑗 1𝐴𝑗 with
some 𝑎𝑗 ∈ R and 𝐴𝑗 ∈ 𝒜 . Since
∑ ∑
𝑓+ = 𝑎𝑗 1𝐴𝑗 and 𝑓 − = 𝑎𝑗 1𝐴𝑗 ,
0⩽𝑗⩽𝑁 0⩽𝑗⩽𝑁
𝑎𝑗 ⩾0 𝑎𝑗 ⩽0
The converse is also true since 𝑓𝑓+ −𝑓 − —see (8.8) or Problem 8.6—and since sums and differences
of simple functions are again simple.
■■
Now the claim follows from the elementary identities that for any two numbers 𝑎, 𝑏 ∈ R
■■
Problem 8.7 Solution: If we show that {𝑢 > 𝛼} is an open set, it is also a Borel set, hence 𝑢 is
measurable.
Let us first understand what openness means: {𝑢 > 𝛼} is open means that for 𝑥 ∈ {𝑢 > 𝛼} we find
some (symmetric) neighbourhood (a ‘ball’) of the type (𝑥 − ℎ, 𝑥 + ℎ) ⊂ {𝑢 > 𝛼}. What does this
mean? Obviously, that 𝑢(𝑦) > 𝛼 for any 𝑦 ∈ (𝑥 − ℎ, 𝑥 + ℎ) and, in other words, 𝑢(𝑦) > 𝛼 whenever
𝑦 is such that |𝑥 − 𝑦| < ℎ. And this is the hint of how to use continuity: we use it in order to find
the value of ℎ.
Since 𝑢(𝑥) > 𝛼 we know that for a sufficiently small 𝜖 we still have 𝑢(𝑥) ⩾ 𝛼 + 𝜖. Take this 𝜖 and
find the corresponding 𝛿. Then
83
R.L. Schilling: Measures, Integrals & Martingales
i.e. 𝑢(𝑦) > 𝛼 for 𝑦 such that |𝑥 − 𝑦| < 𝛿. This means, however, that ℎ = 𝛿 does the job.
■■
Problem 8.8 Solution: The minimum/maximum of two numbers 𝑎, 𝑏 ∈ R can be written in the form
1( )
min{𝑎, 𝑏} = 𝑎 + 𝑏 − |𝑎 − 𝑏|
2
1( )
max{𝑎, 𝑏} = 𝑎 + 𝑏 + |𝑎 − 𝑏|
2
which shows that we can write min{𝑥, 0} and max{𝑥, 0} as a combination of continuous functions.
As such they are again continuous, hence measurable. Thus,
■■
(ii) Let 𝑥 ∈ {sup𝑖 𝑓𝑖 < 𝜆}. Then we have 𝑓𝑗 (𝑥) ⩽ sup𝑖∈𝐼 𝑓𝑖 (𝑥) < 𝜆 for all 𝑗 ∈ 𝐼; this
⋂
means 𝑥 ∈ {𝑓𝑗 < 𝜆} for all 𝑗 ∈ 𝐼 and so 𝑥 ∈ 𝑗∈𝐼 {𝑓𝑗 < 𝜆}.
(Note: ‘⊃’ is, in general, wrong. To see this, use e.g. 𝑓𝑖 (𝑥) ∶= − 1𝑖 , 𝑖 ∈ N, and
⋂
𝜆 = 0. Then we have {sup𝑖 𝑓𝑖 < 0} = ∅ ≠ R = 𝑖 {𝑓𝑖 < 0}.)
⋃
(iii) Let 𝑥 ∈ 𝑖 {𝑓𝑖 ⩾ 𝜆}. Then there is some 𝑖0 ∈ 𝐼 such that 𝑥 ∈ {𝑓𝑖0 ⩾ 𝜆}, hence
⇐⇒ ∀𝑖 ∈ 𝐼 ∶ 𝑥 ∈ {𝑓𝑖 ⩽ 𝜆}
⋂
⇐⇒ 𝑥 ∈ {𝑓𝑖 ⩽ 𝜆}.
𝑖∈𝐼
84
Solution Manual. Last update 28th October 2021
■■
Problem 8.10 Solution: The 𝑓𝑗 are step-functions where the bases of the steps are the sets 𝐴𝑗𝑘 and
{ } { } { }
𝐴𝑗 . Since they are of the form, e.g. 𝑘2−𝑗 ⩽ 𝑢 < (𝑘 + 1)2−𝑗 = 𝑘2−𝑗 ⩽ 𝑢 ∩ 𝑢 < (𝑘 + 1)2−𝑗 ,
it is clear that they are not only in 𝒜 but in 𝜎(𝑢).
■■
and 𝑠+
𝑛 is simple, i.e. 𝑢 is measurable. As 𝑢 = 𝑢 − 𝑢 we conclude that 𝑢 = 𝑢 − 𝑢 is again
+ + − − +
{𝑢 < 𝑣} = {0 < 𝑢 − 𝑣}
etc. is measurable.
Let us be a bit more careful and consider the case where we could encounter expressions of
the type ‘∞ − ∞’. Since 𝑠𝑛 ↑ 𝑢 for simple functions (they are always R-valued...) we get
(∗) ⋂ ⋂
{𝑢 ⩽ 𝑣} = {sup 𝑠𝑛 ⩽ 𝑢} = {𝑠𝑛 ⩽ 𝑢} = {0 ⩽ 𝑢 − 𝑠𝑛 }
𝑛 𝑛 𝑛
and the latter is a union of measurable sets, hence measurable. Now {𝑢 < 𝑣} = {𝑢 ⩾ 𝑣}𝑐 and
we get measurability after switching the roles of 𝑢 and 𝑣. Finally {𝑢 = 𝑣} = {𝑢 ⩽ 𝑣}∩{𝑢 ⩾ 𝑣}
and {𝑢 ≠ 𝑣} = {𝑢 = 𝑣}𝑐 .
Let me stress the importance of ‘⩽’ in (∗) above: we use here
and this would be incorrect if we had had ‘<’, since the argument would break down at (∗∗)
(only one implication would be valid: ‘ ⇐⇒ ’).
■■
Problem 8.12 Solution: Since 𝑋 is 𝜎-finite, there is an exhausting sequence 𝐴𝑛 ↑ 𝑋 with 𝜇(𝐴𝑛 ) < ∞.
Let 𝑢 ∈ (𝒜 ).
85
R.L. Schilling: Measures, Integrals & Martingales
• From (an obvious variant of) Problem 5.12 we know that we can approximate 𝐴𝑛,𝑚 having
finite measure by some 𝐺𝑛,𝑚 ∈ 𝒢 in such a way that 𝜇{1𝐺𝑛,𝑚 ≠ 1𝐴𝑛,𝑚 } ⩽ 2−𝑛 ∕𝑀(𝑛) (note:
|1𝐴 − 1𝐵 | = 1𝐴 ▵ 𝐵 ).
Moreover,
𝑀(𝑛)
∑
𝑓𝑛 (𝑥) ∶= 𝛼𝑛,𝑚 1𝐺𝑛,𝑚 (𝑥)
𝑚=1
⋃
and since {𝑓𝑛 ≠ 𝑢𝑛 } ⊂ 𝑚 𝐺𝑛,𝑚 ▵ 𝐴𝑛,𝑚 , we get 𝜇{𝑓𝑛 ≠ 𝑢𝑛 } ⩽ 2−𝑛 .
As lim𝑛→∞ 𝑢𝑛 (𝑥) = 𝑢(𝑥) for all 𝑥, we find from the continuity of the measure (from above)
( )
⋂⋃
𝜇( lim 𝑓𝑛 ≠ 𝑢) ⩽ 𝜇 {𝑓𝑛 ≠ 𝑢𝑛 }
𝑛→∞
𝑘∈N 𝑛⩾𝑘
∞
∑
⩽ lim 𝜇{𝑓𝑛 ≠ 𝑢𝑛 }
𝑘→∞
𝑛=𝑘
∞
∑
⩽ lim 2−𝑛 = 0.
𝑘→∞
𝑛=𝑘
This shows that (𝒢 ) ∋ 𝑓𝑛 (𝑥) → 𝑢(𝑥) for all 𝑥 ∉ 𝑁 with 𝜇(𝑁) = 0.
An alternative proof can be based on the monotone class theorem. We sketch the steps below
(notation as above and in Theorem 8.15):
{ }
• Set 𝑛 ∶= 𝑢 ∈ (𝐴𝑛 ∩ 𝒜 ) ∶ ∃(𝑓𝑖 )𝑖 ⊂ (𝐴𝑛 ∩ 𝒢 ), ∃𝑁𝑛 ∈ 𝒜 , 𝜇(𝑁𝑛 ) = 0, ∀𝑥 ∉ 𝑁𝑛 ∶ 𝑓𝑖 (𝑥) → 𝑢(𝑥) .
Obviously 𝑛 is a vector space which is stable under bounded suprema (use a diagonal argu-
ment and the fact that the union of countably many null sets is again a null set).
• Observe that 1𝐴𝑛 , 1𝐴𝑛 ∩𝐴 ∈ 𝑛 for all 𝐴 ∈ 𝒜 by the result of Problem 5.12.
• Use the monotone class theorem.
• Glue together the sets 𝑛 by considering 𝑢 = lim𝑛 𝑢1𝐴𝑛 . This leads again to a countable union
of null sets.
■■
86
Solution Manual. Last update 28th October 2021
■■
Problem 8.14 Solution: It is sometimes necessary to distinguish between domain and range. We use
the subscript 𝑥 to signal the domain, the subscript 𝑦 for the range.
(It is an instructive exercise to check that 𝜎(𝑔) is indeed a 𝜎-algebra. This is, of course, clear
from the general theory since 𝜎(𝑔) = 𝑔 −1 ([0, ∞) ∩ ℬ), i.e. it is the pre-image of the trace
𝜎-algebra and pre-images of 𝜎-algebras are always 𝜎-algebras.
(iv) As warm-up we follow the hint. The set {(𝑥, 𝑦) ∶ 𝑥 + 𝑦 = 𝛼} is the line 𝑦 = 𝛼 − 𝑥 in the
𝑥-𝑦-plane, i.e. a line with slope −1 and shift 𝛼. So {(𝑥, 𝑦) ∶ 𝑥 + 𝑦 ⩾ 𝛼} would be the points
above this line and {(𝑥, 𝑦) ∶ 𝛽 ⩾ 𝑥 + 𝑦 ⩾ 𝛼} = {(𝑥, 𝑦) ∶ 𝑥 + 𝑦 ∈ [𝛼, 𝛽]} would be the points
in the strip which has the lines 𝑦 = 𝛼 − 𝑥 and 𝑦 = 𝛽 − 𝑥 as boundaries.
This set is, in an abuse of notation, 𝑦 = 𝐵 − 𝑥, i.e. these are all lines with slope −1 (135
degrees) and every possible shift from the set 𝐵—it gives a kind of stripe-pattern. To sum
up:
87
R.L. Schilling: Measures, Integrals & Martingales
(v) Again follow the hint to see that {(𝑥, 𝑦) ∶ 𝑥2 + 𝑦2 = 𝑟} is a circle, radius 𝑟, centre (0, 0). So
{(𝑥, 𝑦) ∶ 𝑥2 + 𝑦2 ⩽ 𝑟} is the solid disk, radius 𝑟, centre (0, 0) and {(𝑥, 𝑦) ∶ 𝑅 ⩾ 𝑥2 + 𝑦2 ⩾
𝑟} = {(𝑥, 𝑦) ∶ 𝑥2 + 𝑦2 ∈ [𝑟, 𝑅]} is the annulus with exterior radius 𝑅 and interior radius 𝑟
about (0, 0).
More general, take a Borel set 𝐵 ⊂ [0, ∞), 𝐵 ∈ ℬ(R), i.e. 𝐵 ∈ [0, ∞) ∩ ℬ(R) (negative
radii don’t make sense!) and observe that the set {(𝑥, 𝑦) ∶ 𝑥2 + 𝑦2 ∈ 𝐵} gives a ring-pattern
which is ‘supported’ by the set 𝐵 (i.e. we take all circles passing through 𝐵...). To sum up:
■■
Problem 8.15 Solution: Assume first that 𝑢 is injective. This means that every point in the range
𝑢(R) comes exactly from one uniquely defined 𝑥 ∈ R. This can be expressed by saying that
{𝑥} = 𝑢−1 ({𝑢(𝑥)}) — but the singleton {𝑢(𝑥)} is a Borel set in the range, so {𝑥} ∈ 𝜎(𝑢) as
𝜎(𝑢) = 𝑢−1 (𝑢(R) ∩ ℬ).
Conversely, assume that for each 𝑥 we have {𝑥} ∈ 𝜎(𝑢). Fix an 𝑥0 and call 𝑢(𝑥0 ) = 𝛼. Since 𝑢
is measurable, the set {𝑢 = 𝛼} = {𝑥 ∶ 𝑢(𝑥) = 𝛼} is measurable and, clearly, {𝑥0 } ⊂ {𝑢 = 𝛼}.
But if we had another 𝑥0 ≠ 𝑥1 ∈ {𝑢 = 𝛼} this would mean that we could never ‘produce’ {𝑥0 } on
its own as a pre-image of some set, but we must be able to do so as {𝑥0 } ∈ 𝜎(𝑢), by assumption.
Thus, 𝑥1 = 𝑥0 . To sum up, we have shown that {𝑢 = 𝛼} consists of one point only, i.e. we have
shown that 𝑢(𝑥0 ) = 𝑢(𝑥1 ) implies 𝑥0 = 𝑥1 which is just injectivity.
■■
Problem 8.16 Solution: Clearly 𝑢 ∶ R → [0, ∞). So let’s take 𝐼 = (𝑎, 𝑏) ⊂ [0, ∞). Then
𝑢−1 ((𝑎, 𝑏)) = (−𝑏, −𝑎) ∪ (𝑎, 𝑏). This shows that for 𝜇 ∶= 𝜆◦𝑢−1
( )
𝜇(𝑎, 𝑏) = 𝜆◦𝑢−1 ((𝑎, 𝑏)) = 𝜆 (−𝑏, −𝑎) ∪ (𝑎, 𝑏) = 𝜆(−𝑏, −𝑎) + 𝜆(𝑎, 𝑏)
= (−𝑎 − (−𝑏)) + (𝑏 − 𝑎) = 2(𝑏 − 𝑎) = 2𝜆((𝑎, 𝑏)).
This shows that 𝜇 = 2𝜆 if we allow only intervals from [0, ∞), i.e.
( )
𝜇(𝐼) = 2𝜆 𝐼 ∩ [0, ∞) for any interval 𝐼 ⊂ R.
Since a measure on the Borel sets is completely described by (either: open or closed or half-open
or half-closed) intervals (the intervals generate the Borel sets!), we can invoke the uniqueness
theorem to guarantee that the above equality holds for all Borel sets.
■■
88
Solution Manual. Last update 28th October 2021
(i) Because of Lemma 7.2 it is enough to check measurability for some generator. Let 𝐵 =
[𝑎, 𝑏) ∈ 𝒥 , 𝑎 < 𝑏. We have
if 𝑎, 𝑏 ⩽ 0
⎧
⎪∅
⎪ √ √
𝑄−1 (𝐵) = 𝐸 ∩ ⎨(− 𝑏, + 𝑏) if 𝑎 ⩽ 0, 𝑏 > 0
⎪( √ √ ] [√ √ )
⎪ − 𝑏, − 𝑎 ∪ 𝑎, 𝑏 if 𝑎, 𝑏 > 0
⎩
More precisely: we have already seen that 𝜈◦𝑄−1 is a measure (Theorem 7.6). Since 𝒥
is ∩-stable and 𝜈◦𝑄−1 a finite measure (𝜈 comes from a finite Lebesgue measure), we get
uniqueness from Theorem 5.7, and it enough to consider sets of the form 𝐵 = [𝑎, 𝑏) ∈ 𝒥 ,
𝑎 ⩽ 𝑏.
⎧
⎪0, 𝑏 ⩽ 0 or 𝑎 > 1
⎪ √
𝜈(𝑄−1 (𝐵)) = ⎨𝜆([0, 𝑏)), 𝑎 < 0, 𝑏 > 0
⎪ √ √
⎪𝜆([ 𝑎, 𝑏 ∧ 1)), 0 < 𝑎 < 1
⎩
⎧
⎪0, 𝑏 ⩽ 0 or 𝑎 > 1
= ⎨√ √
⎪ 𝑏 ∧ 1 − 0 ∨ 𝑎 ∧ 1, otherwise.
⎩
⎧
⎪0, 𝑏 ⩽ 0 or 𝑎 > 1
⎪ √ √
𝜈(𝑄−1 (𝐵)) = ⎨𝜆([(− 𝑏) ∨ (−1), 𝑏 ∧ 1)), 𝑎 < 0, 𝑏 > 0
⎪1 √ √ √ √
⎪ 𝜆([(− 𝑏) ∨ (−1), 𝑎) ∪ [ 𝑎, 𝑏 ∧ 1)), 0 < 𝑎 < 1
⎩2
⎧
⎪0, 𝑏 ⩽ 0 or 𝑎 > 1
=⎨ √ √
⎪2 21 𝜆([0 ∨ 𝑎 ∧ 1, 𝑏 ∧ 1)), otherwise
⎩
⎧
⎪0, 𝑏 ⩽ 0 or 𝑎 > 1
=⎨√ √
⎪( 𝑏 ∧ 1)) − (0 ∧ 𝑎 ∧ 1), otherwise
⎩
■■
89
R.L. Schilling: Measures, Integrals & Martingales
• clear, since 𝑢(𝑥 − 2) is a combination of the measurable shift 𝜏2 and the measurable function
𝑢.
• this is trivial since 𝑢 → 𝑒𝑢 is a continuous function, as such it is measurable and combinations
of measurable functions are again measurable.
• this is trivial since 𝑢 → sin(𝑢 + 8) is a continuous function, as such it is measurable and
combinations of measurable functions are again measurable.
• iterate Problem 8.13
• obviously, sgn 𝑥 = (−1) ⋅ 1(−∞,0) (𝑥) + 0 ⋅ 1{0} (𝑥) + 1 ⋅ 1(0,∞) (𝑥), i.e. a measurable function.
Using the first example, we see now that sgn 𝑢(𝑥 − 7) is a combination of three measurable
functions.
■■
Problem 8.19 Solution: Consider, for instance, 𝑇 ∶ [0, 1) → [0, 1) where 𝑇 (𝑥) = 𝑥
2
and 𝑤𝑛 ∶
[0, 1) → R with 𝑤𝑛 (𝑥) = (−1)𝑛 1 [1∕2,1) (𝑥).
■■
Problem 8.20 Solution: Let 𝐴 ⊂ R be such that 𝐴 ∉ ℬ. Then it is clear that 𝑢(𝑥) = 1𝐴 (𝑥)−1𝐴𝑐 (𝑥) is
NOT measurable (take, e.g. 𝐴 = {𝑓 = 1} which should be measurable for measurable functions),
but clearly, |𝑓 (𝑥)| = 1 and as constant function this IS measurable.
■■
Problem 8.21 Solution: We want to show that the sets {𝑢 ⩽ 𝛼} are Borel sets. We will even show
that they are intervals, hence Borel sets. Imagine the graph of an increasing function and the line
𝑦 = 𝛼 cutting through. Essentially we have three scenarios: the cut happens at a point where (a) 𝑢
is continuous and strictly increasing or (b) 𝑢 is flat or (c) 𝑢 jumps—i.e. has a gap; these three cases
are shown in the following pictures: From the three pictures it is clear that we get in any case an
✻ ✻ ✻
𝛼 𝛼 𝛼
𝛽 𝛽 𝛽
✲ ✲ ✲
𝑏 𝑎 𝑐 𝑏 𝑎 𝑏 𝑎
interval for the sub-level sets {𝑢 ⩽ 𝛾} where 𝛾 is some level (in the pic’s 𝛾 = 𝛼 or = 𝛽), you can
read off the intervals on the abscissa where the dotted lines cross the abscissa.
Now let’s look at the additional conditions: First the intuition: From the first picture, the continuous
and strictly increasing case, it is clear that we can produce any interval (−∞, 𝑏] to (−∞, 𝑎] by
looking at {𝑢 ⩽ 𝛽} to {𝑢 ⩽ 𝛼} my moving up the 𝛽-line to level 𝛼. The point is here that we get
all intervals, so we get a generator of the Borel sets, so we should get all Borel sets.
90
Solution Manual. Last update 28th October 2021
The second picture is bad: the level set {𝑢 ⩽ 𝛽} is (−∞, 𝑏] and all level sets below will only come
up to the point (−∞, 𝑐], so there is no chance to get any set contained in (𝑐, 𝑏), i.e. we cannot get
all Borel sets.
The third picture is good again, because the vertical jump does not hurt. The only ‘problem’ is
whether {𝑢 ⩽ 𝛽} is (−∞, 𝑏] or (−∞, 𝑏) which essentially depends on the property of the graph
whether 𝑢(𝑏) = 𝛽 or not, but this is not so relevant here, we just must make sure that we can get
more or less all intervals. The reason, really, is that jumps as we described them here can only
happen countably often, so this problem occurs only countably often, and we can overcome it
therefore.
So the point is: we must disallow flat bits, i.e. 𝜎(𝑢) is the Borel 𝜎-algebra if, and only, if 𝑢 is strictly
increasing, i.e. if, and only if, 𝑢 is injective. (Note that this would have been clear already from
Problem 8.15, but our approach here is much more intuitive.)
■■
Σ ∶= {𝐴 ∈ 𝒜 ∶ 𝐴 ∈ 𝜎(𝑔)}.
Σ ist a 𝜎-Algebra:
(Σ2 ) For 𝐴 ∈ Σ we have 𝐴 ∈ 𝜎(𝑔); since 𝜎(𝑔) is a 𝜎-algebra, we see that 𝐴𝑐 ∈ 𝜎(𝑔); hence,
𝐴𝑐 ∈ Σ.
⋃ ⋃
(Σ3 ) For (𝐴𝑛 )𝑛∈N ⊂ Σ we see 𝑛∈N 𝐴𝑛 ∈ 𝜎(𝑔), thus 𝑛 𝐴𝑛 ∈ Σ.
■■
Problem 8.23 Solution: Without loss of generality, assume that 𝑢 is right-continuous (left-continuity
works analogously). Approximate 𝑢 with simple functions:
2𝑛 2
∑
𝑢𝑛 (𝑥) ∶= 𝑢(𝑥𝑛𝑖+1 )1[𝑥𝑛 ,𝑥𝑛 ) (𝑥)
𝑖 𝑖+1
𝑖=1
91
R.L. Schilling: Measures, Integrals & Martingales
Indeed: For each 𝑥 ∈ R there is some 𝑁 ∈ N such that 𝑥 ∈ [−𝑁, 𝑁]. By definition, we find for
all 𝑛 ⩾ 𝑁, ( )
⌊𝑛𝑥⌋ + 1
𝑢𝑛 (𝑥) = 𝑢
𝑛
( is the smallest number of the form 𝑘𝑛 , 𝑘 ∈ Z, which exceeds 𝑥.) Because of the right-
⌊𝑛𝑥⌋+1
𝑛
continuity of 𝑢 we get 𝑢𝑛 (𝑥) → 𝑢(𝑥) as 𝑛 → ∞. Therefore, 𝑢 is Borel-measurable (pointwise limit
of measurable functions).
■■
Problem 8.24 Solution: Every linear map on a finite-dimensional vector space is continuous, hence
Borel measurable.
Note that 𝑓 ∶ R → R2 , 𝑓 (𝑥) ∶= (𝑥, 0)⊤ , is continuous, hence Borel measurable. This map is,
however, not measurable with respect to the completed Borel 𝜎-algebras:
To see this, let 𝐴 ⊂ R, 𝐴 ∉ ℬ(R), be a subset of a Lebesgue null set. For 𝐴 × {0} we see
that 𝐴 × {0} ∈ ℬ(R2 ); this follows from 𝐴 × {0} ⊂ 𝑁 ∶= R × {0} and 𝜆2 (𝑁) = 0 (cf.
Problem 4.15, Problem 6.7). On the other hand, 𝑓 −1 (𝐴 × {0}) = 𝐴 ∉ ℬ(R) ⊂ ℬ(R), i.e.
𝑓 ∶ (R, ℬ(R)) → (R2 , ℬ(R2 )) is not measurable.
■■
Problem 8.25 Solution: Without loss of generality we consider the right-continuous situation. The
left-continuous counterpart is very similar.
• Fix 𝜔 ∈ Ω. Note that it is enough to show that 𝑡 → 𝜉(𝑡, 𝜔)1[𝑎,𝑏] (𝑡) =∶ 𝜉 𝑎,𝑏 (𝑡, 𝜔) is measurable
for all 𝑎 < 𝑏.
Indeed: Because of
𝜉(𝑡, 𝜔) = lim 𝜉 −𝑅,𝑅 (𝑡, 𝜔)
𝑅→∞
the map 𝑡 → 𝜉(𝑡, 𝜔) is measurable (pointwise limit of measurable functions, cf. Corol-
lary 8.10.
In order to keep notation simple, we assume that 𝑎 = 0 and 𝑏 = 1; the general case is similar.
Define
2𝑛 −1
∑ ( 𝑖+1 )
𝜉𝑛 (𝑡, 𝜔) ∶= 𝜉 2𝑛 , 𝜔 1[ 𝑖 𝑖+1
) (𝑡).
, ∧1
𝑖=0 2𝑛 2𝑛
92
Solution Manual. Last update 28th October 2021
Consequently, it is enough to show (by Corollary 8.10) that each 𝑡 → 𝜉𝑛 (𝑡, 𝜔) is measurable.
For 𝛼 ∈ R we get
[ )
⋃ 𝑖 𝑖+1
{𝑡 ∶ 𝜉𝑛 (𝑡, 𝜔) ⩽ 𝛼} = , ∈ ℬ(R)
𝑖∈𝐼
2𝑛 2𝑛
⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
∈ℬ(R)
where
{ ( ) }
𝑖+1
𝐼 ∶= 𝑖 ∈ {0, … , 2𝑛 − 1}; 𝜉 , 𝜔 ⩽ 𝛼 .
2𝑛
This proves that 𝑡 → 𝜉𝑛 (𝑡, 𝜔) is measurable.
Indeed: The estimate ‘⩾’ is clear, i.e. we only have to show ‘⩽’. Using the definition of the
supremum, there is for each 𝜖 > 0 some 𝑠 ∈ R such that
Because of right-continuity we find some 𝑟 ∈ Q, 𝑟 > 𝑠, such that |𝜉(𝑟, 𝜔) − 𝜉(𝑠, 𝜔)| ⩽ 𝜖.
Therefore,
sup 𝜉(𝑡, 𝜔) ⩾ 𝜉(𝑟, 𝜔) ⩾ 𝜉(𝑠, 𝜔) − 𝜖 ⩾ sup 𝜉(𝑡, 𝜔) − 2𝜖.
𝑡∈Q 𝑡∈R
From (⋆) we get that the map 𝜔 → sup𝑡∈R 𝜉(𝑡, 𝜔) is measurable (as supremum of countably
many measurable functions, cf. Corollary 8.10).
■■
Problem 8.26 Solution: ‘⇐’: Assume that there are 𝒜 ∕ℬ(R)-measurable functions 𝑓 , 𝑔 ∶ 𝑋 → R
satisfying 𝑓 ⩽ 𝜙 ⩽ 𝑔 and 𝜇{𝑓 ≠ 𝑔} = 0. For any 𝑥 ∈ R we get
{𝜙 ⩽ 𝑥} = {𝜙 ⩽ 𝑥, 𝑓 = 𝑔} ∪ {𝜙 ⩽ 𝑥, 𝑓 ≠ 𝑔}
= {𝑔 ⩽ 𝑥, 𝑓 = 𝑔} ∪ {𝜙 ⩽ 𝑥, 𝑓 ≠ 𝑔} .
⏟⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏟
=∶𝐴 =∶𝑁
Since 𝑓 and 𝑔 are measurable, we see that 𝐴 ∈ 𝒜 . For 𝑁 we only get 𝑁 ⊂ {𝑓 ≠ 𝑔}, i.e. 𝑁 is a
subset of a 𝜇-null set. By the definition of 𝒜 (see Problem 4.15) we find {𝜙 ⩽ 𝑥} ∈ 𝒜 .
with 𝑐𝑖 ∈ R, 𝐴𝑖 ∈ 𝒜 (𝑖 = 1, … , 𝑛). From the definition of 𝒜 we get that the 𝐴𝑖 are of the form
𝐴𝑖 = 𝐵𝑖 + 𝑁𝑖
93
R.L. Schilling: Measures, Integrals & Martingales
Let 𝜙 be any 𝒜 ∕ℬ(R)-measurable function. Using Corollary 8.9, we get a sequence (𝜙𝑛 )𝑛∈N of
𝒜 ∕ℬ(R)-measurable simple functions such that 𝜙𝑛 (𝑥) → 𝜙(𝑥) for all 𝑥 ∈ 𝑋. By the first part of
this proof, there are 𝒜 ∕ℬ(R)-measurable funcitons 𝑓𝑛 , 𝑔𝑛 , 𝑛 ∈ N, such that 𝑓𝑛 ⩽ 𝜙𝑛 ⩽ 𝑔𝑛 and
𝜇(𝑓𝑛 ≠ 𝑔𝑛 ) = 0. Set
The functions 𝑓 and 𝑔 are again 𝒜 ∕ℬ(R)-measurable (Corollary 8.10) and we have 𝑓 ⩽ 𝜙 ⩽ 𝑔.
Moreover, ( )
⋃ ∑
𝜇(𝑓 ≠ 𝑔) ⩽ 𝜇 {𝑓𝑛 ≠ 𝑔𝑛 } ⩽ 𝜇(𝑓𝑛 ≠ 𝑔𝑛 ) = 0.
𝑛∈N 𝑛∈N
■■
94
9 Integration of positive functions.
Solutions to Problems 9.1–9.14
Problem 9.1 Solution: We know that for any two simple functions 𝑓 , 𝑔 ∈ + we have 𝐼𝜇 (𝑓 + 𝑔) =
𝐼𝜇 (𝑓 ) + 𝐼𝜇 (𝑔) (= additivity), and this is easily extended to finitely many, say, 𝑚 different positive
simple functions. Observe now that each 𝜉𝑛 1𝐴𝑛 is a positive simple function, hence
( )
𝑚
∑ 𝑚
∑ ( ) ∑𝑚 ( ) ∑ 𝑚
( )
𝐼𝜇 𝜉𝑛 1𝐴𝑛 = 𝐼𝜇 𝜉𝑛 1𝐴𝑛 = 𝜉𝑛 𝐼𝜇 1𝐴𝑛 = 𝜉𝑛 𝜇 𝐴𝑛 .
𝑛=1 𝑛=1 𝑛=1 𝑛=1
■■
Problem 9.2 Solution: We use indicator functions. Note that any fixed 𝑥 can be contained in 𝑘 ∈
()
{0, 1, … , 𝑁} of the sets 𝐴𝑛 . Then 𝑥 is contained in 𝐴1 ∪ ⋯ ∪ 𝐴𝑁 as well as in 𝑘2 of the pairs
( )
𝐴𝑛 ∪ 𝐴𝑘 where 𝑛 < 𝑘; as usual: 𝑚𝑛 = 0 if 𝑚 < 𝑛. This gives
( )
∑ 𝑘 ∑
1𝐴𝑛 =𝑘⩽1+ = 1𝐴1 ∪⋯∪𝐴𝑁 + 1𝐴𝑛 1𝐴𝑘
𝑛
2 𝑛<𝑘
∑
= 1𝐴1 ∪⋯∪𝐴𝑁 + 1𝐴𝑛 ∩𝐴𝑘 .
𝑛<𝑘
■■
(i) This follows from Properties 9.3 and Lemma 9.5 since ∫ 1𝐴 𝑑𝜇 = 𝐼𝜇 (1𝐴 ) = 𝜇(𝐴).
(ii) This follows again from Properties 9.3 and Corollary 9.7 since for 𝑢𝑛 ∈ + with 𝑢 = sup𝑛 𝑢𝑛
(note: the sup’s are increasing limits!) we have
= 𝛼 sup 𝐼𝜇 (𝑢𝑛 )
𝑛
=𝛼 𝑢 𝑑𝜇.
∫
95
R.L. Schilling: Measures, Integrals & Martingales
(iii) This follows again from Properties 9.3 and Corollary 9.7 since for 𝑢𝑛 , 𝑣𝑛 ∈ + with 𝑢 =
sup𝑛 𝑢𝑛 , 𝑣 = sup𝑛 𝑣𝑛 (note: the sup’s are increasing limits!) we have
= 𝑢 𝑑𝜇 + 𝑣 𝑑𝜇.
∫ ∫
(iv) This was shown in Step 1 of the proof of the Beppo Levi theorem 9.6
■■
Problem 9.4 Solution: Consider on the space ([−1, 0], 𝜆), 𝜆(𝑑𝑥) = 𝑑𝑥 is Lebesgue measure on [0, 1],
the sequence of ‘tent-type’ functions
⎧
⎪0, −1 ⩽ 𝑥 ⩽ − 𝑘1 ,
𝑓𝑘 (𝑥) = ⎨ ( (𝑘 ∈ N),
⎪𝑘3 𝑥 + 𝑘1 ), − 𝑘1 ⩽ 𝑥 ⩽ 0,
⎩
(draw a picture!). These are clearly monotonically increasing functions but, as a sequence, we do
not have 𝑓𝑘 (𝑥) ⩽ 𝑓𝑘+1 (𝑥) for every 𝑥! Note also that each function is integrable (with integral 12 𝑘)
but the pointwise limit is not integrable.
■■
Problem 9.5 Solution: The first part is trivial since it just says that the sequence becomes increasing
only from index 𝐾 onwards. This 𝐾 does not depend on 𝑥 but is uniform for the whole sequence.
Since we are anyway only interested in 𝑢 = lim𝑛→∞ 𝑢𝑛 = sup𝑛⩾𝐾 𝑢𝑛 , we can neglect the elements
𝑢1 , … , 𝑢𝐾 and consider only the then increasing sequence (𝑢𝑛+𝐾 )𝑛 . Then we can directly apply
Beppo Levi’s theorem, Theorem 9.6.
The other condition says that the sequence 𝑢𝑛+𝐾 (𝑥) is increasing for some 𝐾 = 𝐾(𝑥). But since 𝐾
may depend on 𝑥, we will never get some overall increasing behaviour of the sequence of functions.
Take, for example, on (R, ℬ(R), 𝜆 ∶= 𝜆1 ),
This is a sequence of symmetric tent-like functions of tents with base (−1∕𝑛, 1∕𝑛) and tip at 𝑛2
(which we take out and replace by the value 0). Clearly:
← 0 and
𝑢𝑛 (𝑥) ←←←←←←←←←→ 𝑢𝑛 (𝑥) 𝑑𝑥 = 1 ∀ 𝑛.
𝑛→∞ ∫
Moreover, if 𝑛 ⩾ 𝐾 = 𝐾(𝑥) with 𝐾(𝑥) defined to be the smallest integer > 1∕|𝑥|, then 𝑢𝑛 (𝑥) = 0
so that the second condition is clearly satisfied, but ∫ 𝑢𝑛 (𝑥) 𝑑𝑥 = 1 cannot converge to ∫ 0 𝑑𝑥 =
∫ 𝑢(𝑥) 𝑑𝑥 = 0.
■■
96
Solution Manual. Last update 28th October 2021
Problem 9.6 Solution: Following the hint we set 𝑠𝑚 = 𝑢1 + 𝑢2 + … + 𝑢𝑚 . As a finite sum of positive
∑
measurable functions this is again positive and measurable. Moreover, 𝑠𝑚 increases to 𝑠 = ∞
𝑛=1 𝑢𝑛
as 𝑚 → ∞. Using the additivity of the integral (9.8 (iii)) and the Beppo Levi theorem 9.6 we get
∞
∑
𝑢𝑛 𝑑𝜇 = sup 𝑠𝑚 𝑑𝜇 = sup 𝑠𝑚 𝑑𝜇
∫ ∫ 𝑚 𝑚 ∫
𝑛=1
= sup (𝑢 + … + 𝑢𝑚 ) 𝑑𝜇
𝑚 ∫ 1
𝑚
∑
= sup 𝑢 𝑑𝜇
𝑚 ∫ 𝑛
𝑛=1
∞
∑
= 𝑢𝑛 𝑑𝜇.
∫
𝑛=1
Conversely, assume that 9.9 is true. We want to deduce from it the validity of Beppo Levi’s theorem
9.6. So let (𝑤𝑛 )𝑛∈N be an increasing sequence of measurable functions with limit 𝑤 = sup𝑛 𝑤.
For ease of notation we set 𝑤0 ≡ 0. Then we can write each 𝑤𝑛 as a partial sum
■■
Problem 9.7 Solution: Set 𝜈(𝐴) ∶= ∫ 1𝐴 𝑢 𝑑𝜇. Then 𝜈 is a [0, ∞]-valued set function defined for
𝐴 ∈ 𝒜.
97
R.L. Schilling: Measures, Integrals & Martingales
■■
Problem 9.8 Solution: This is actually trivial: since our 𝜎-algebra is 𝒫 (N), all subsets of N are
measurable. Now the sub-level sets {𝑢 ⩽ 𝛼} = {𝑘 ∈ N ∶ 𝑢(𝑘) ⩽ 𝛼} are always ⊂ N and as such
they are ∈ 𝒫 (N), hence 𝑢 is always measurable.
■■
Problem 9.9 Solution: We have seen in Problem 4.7 that 𝜇 is indeed a measure. We follow the
instructions. First, for 𝐴 ∈ 𝒜 we get
∑ ∑
1𝐴 𝑑𝜇 = 𝜇(𝐴) = 𝜇𝑗 (𝐴) = 1𝐴 𝑑𝜇𝑗 .
∫ ∫
𝑗∈N 𝑗∈N
By the linearity of the integral, this easily extends to functions of the form 𝛼1𝐴 + 𝛽1𝐵 where
𝐴, 𝐵 ∈ 𝒜 and 𝛼, 𝛽 ⩾ 0:
(𝛼1𝐴 + 𝛽1𝐵 ) 𝑑𝜇 = 𝛼 1 𝑑𝜇 + 𝛽 1 𝑑𝜇
∫ ∫ 𝐴 ∫ 𝐵
∑ ∑
=𝛼 1𝐴 𝑑𝜇𝑗 + 𝛽 1 𝑑𝜇
∫ ∫ 𝐵 𝑗
𝑗∈N 𝑗∈N
∑
= (𝛼1𝐴 + 𝛽1𝐵 ) 𝑑𝜇𝑗
∫
𝑗∈N
and this extends obviously to simple functions which are finite sums of the above type.
∑
𝑓 𝑑𝜇 = 𝑓 𝑑𝜇𝑗 ∀𝑓 ∈ + .
∫ ∫
𝑗∈N
Finally, take 𝑢 ∈ + and take an approximating sequence 𝑢𝑛 ∈ + with sup𝑛 𝑢𝑛 = 𝑢. Then we get
by Beppo Levi (indicated by an asterisk ∗)
∞
∗ ∑
𝑢 𝑑𝜇 = sup 𝑢𝑛 𝑑𝜇 = sup 𝑢𝑛 𝑑𝜇𝑗
∫ 𝑛 ∫ 𝑛 𝑗=1 ∫
𝑚
∑
= sup sup 𝑢𝑛 𝑑𝜇𝑗
𝑛 𝑚 𝑗=1 ∫
𝑚
∑
= sup sup 𝑢𝑛 𝑑𝜇𝑗
𝑚 𝑛 𝑗=1 ∫
𝑚
∑
= sup lim 𝑢𝑛 𝑑𝜇𝑗
𝑚 𝑛 ∫
𝑗=1
𝑚
∑
= sup lim 𝑢𝑛 𝑑𝜇𝑗
𝑚 𝑗=1 𝑛 ∫
𝑚
∗ ∑
= sup lim 𝑢𝑛 𝑑𝜇𝑗
𝑚 𝑗=1 ∫ 𝑛
∞
∑
= 𝑢 𝑑𝜇𝑗
∫
𝑗=1
98
Solution Manual. Last update 28th October 2021
where we repeatedly use that all sup’s are increasing limits and that we may swap any two sup’s
(this was the hint to Problem 4.7.)
■■
Problem 9.10 Solution: Set 𝑤𝑛 ∶= 𝑢 − 𝑢𝑛 . Then the 𝑤𝑛 are a sequence of positive measurable
functions. By Fatou’s lemma we get
= 𝑢 𝑑𝜇 − lim sup 𝑢𝑛 𝑑𝜇
∫ 𝑛 ∫
(see, e.g. the rules for lim inf and lim sup in Appendix A). Thus,
= lim inf (𝑢 − 𝑢𝑛 ) 𝑑𝜇
∫ 𝑛
( )
= 𝑢 − lim sup 𝑢𝑛 𝑑𝜇
∫ 𝑛
and the claim follows by subtracting the finite value ∫ 𝑢 𝑑𝜇 on both sides.
Remark. The uniform domination of 𝑢𝑛 by an integrable function 𝑢 is really important. Have a look
at the following situation: (R, ℬ(R), 𝜆), 𝜆(𝑑𝑥) = 𝑑𝑥 denotes Lebesgue measure, and consider the
positive measurable functions 𝑢𝑛 (𝑥) = 1[𝑛,2𝑛] (𝑥). Then lim sup𝑛 𝑢𝑛 (𝑥) = 0 but lim sup𝑛 ∫ 𝑢𝑛 𝑑𝜆 =
lim sup𝑛 𝑛 = ∞ ≠ ∫ 0 𝑑𝜆.
■■
99
R.L. Schilling: Measures, Integrals & Martingales
which uses at the point # the continuity of measures, Proposition 4.3. This step uses the
finiteness of 𝜇.
The alternative would be (i) combined with the reversed Fatou lemma of Problem 9.10:
( )
𝜇 lim sup 𝐴𝑛 = 1lim sup𝑛 𝐴𝑛 𝑑𝜇
𝑛 ∫
(iv) Take the example in the remark to the solution for Problem 9.10. We will discuss it here in its
set-theoretic form: take (R, ℬ(R), 𝜆) with 𝜆 denoting Lebesgue measure 𝜆(𝑑𝑥) = 𝑑𝑥. Put
𝐴𝑛 = [𝑛, 2𝑛] ∈ ℬ(R). Then
⋂⋃ ⋂
lim sup 𝐴𝑛 = [𝑛, 2𝑛] = [𝑘, ∞) = ∅
𝑛 𝑘 𝑛⩾𝑘 𝑘
But 0 = 𝜆(∅) ⩾ lim sup𝑛 𝜆(𝐴𝑛 ) = lim sup𝑛 𝑛 = ∞ is a contradiction. (The problem is that
𝜆[𝑘, ∞) = ∞!)
■■
100
Solution Manual. Last update 28th October 2021
Assume now that (𝑋, 𝒜 , 𝜇) is 𝜎-finite with an exhausting sequence of sets (𝐵𝑛 )𝑛 ⊂ 𝒜 such that
𝐵𝑛 ↑ 𝑋 and 𝜇(𝐵𝑛 ) < ∞. Then we make the 𝐵𝑛 ’s pairwise disjoint by setting
■■
(i) We check (𝑀1 ), (𝑀2 ). Using the fact that 𝑁(𝑥, ⋅) is a measure, we find
Thus 𝑁(𝑓 + 𝑔)(𝑥) = 𝑁𝑓 (𝑥) + 𝑁𝑔(𝑥) for positive simple 𝑓 , 𝑔 ∈ + (𝒜 ). Moreover, since by
Beppo Levi (marked by an asterisk ∗) for an increasing sequence 𝑓𝑘 ↑ 𝑢
∗
sup 𝑁𝑓𝑘 (𝑥) = sup 𝑓𝑘 (𝑦) 𝑁(𝑥, 𝑑𝑦) = sup 𝑓𝑘 (𝑦) 𝑁(𝑥, 𝑑𝑦)
𝑘 𝑘 ∫ ∫ 𝑘
101
R.L. Schilling: Measures, Integrals & Martingales
and since the sup is actually an increasing limit, we see for positive measurable 𝑢, 𝑣 ∈ + (𝒜 )
and the corresponding increasing approximations via positive simple functions 𝑓𝑘 , 𝑔𝑘 :
= 𝑁𝑢(𝑥) + 𝑁𝑣(𝑥).
𝜈(𝐴) = 𝑢 ⋅ 1𝐴 𝑑𝜇 + (1 − 𝑢) ⋅ 1𝐴 𝑑𝜇 = 1𝐴 𝑑𝜇 = 𝜇(𝐴).
∫ ∫ ∫
This means that 𝜈 extends 𝜇. It also shows that 𝜈(∅) = 0. Since 𝜈 is defined for all sets from ℬ(R)
and since 𝜈 has values in [0, ∞], it is enough to check 𝜎-additivity.
For this, let (𝐴𝑛 )𝑛 ⊂ ℬ(R) be a sequence of pairwise disjoint sets. From the definitions it is clear
⨃ (⨃ )±
that the sets (𝐴𝑛 )±𝜎 are again pairwise disjoint and that
±
𝑛 (𝐴𝑛 )𝜎 = 𝑛 𝐴𝑛 𝜎 . Since each of the
set functions
𝐵 → 𝑢 ⋅ 1𝐵 𝑑𝜇, 𝐶 → (1 − 𝑢) ⋅ 1𝐶 𝑑𝜇
∫ ∫
is 𝜎-additive, it is clear that their sum 𝜈 will be 𝜎-additive, too.
The obvious non-uniqueness of the extension does not contradict the uniqueness theorem for ex-
tensions, since Σ does not generate ℬ(R)!
■■
102
10 Integrals of measurable functions.
Solutions to Problems 10.1–10.9
Problem 10.1 Solution: Let 𝑢, 𝑣 be integrable functions and 𝑎, 𝑏 ∈ R. Assume that either 𝑢, 𝑣 are
real-valued or that 𝑎𝑢 + 𝑏𝑣 makes sense (i.e. avoiding the case ‘∞ − ∞’). Then we have
|𝑎𝑢 + 𝑏𝑣| ⩽ |𝑎𝑢| + |𝑏𝑣| = |𝑎| ⋅ |𝑢| + |𝑏| ⋅ |𝑣| ⩽ 𝐾(|𝑢| + |𝑣|)
with 𝐾 = max{|𝑎|, |𝑏|}. Since the RHS is integrable (because of Theorem 10.3 and Properties
9.8) we have that 𝑎𝑢 + 𝑏𝑣 is integrable by Theorem 10.3. So we get from Theorem 10.4 that
(𝑎𝑢 + 𝑏𝑣) 𝑑𝜇 = 𝑎𝑢 𝑑𝜇 + 𝑏𝑣 𝑑𝜇 = 𝑎 𝑢 𝑑𝜇 + 𝑏 𝑣 𝑑𝜇
∫ ∫ ∫ ∫ ∫
■■
Problem 10.2 Solution: Without loss of generality we consider 𝑢 on (0, 1] (otherwise we have to
single out the point 𝑥 = 1, and this is just awkward in the notation...) We follow the hint and
show first that 𝑢(𝑥) ∶= 𝑥−1∕2 , 0 < 𝑥 ⩽ 1, is Lebesgue integrable. The idea here is to construct a
( )2
sequence of simple functions approximating 𝑢 from below. Set 𝑥𝑖 = 𝑛𝑖 , 𝑖 = 0, 1, … , 𝑛 and
𝑛−1 𝑛−1
∑ ∑ 𝑛
𝑢𝑛 (𝑥) ∶= 𝑢(𝑥𝑖+1 )1(𝑥𝑖 ,𝑥𝑖+1 ] (𝑥) = 1 (𝑥)
𝑖=0 𝑖=0
𝑖 + 1 (𝑥𝑖 ,𝑥𝑖+1 ]
This is clearly a simple function. Also 𝑢𝑛 ⩽ 𝑢 and lim𝑛→∞ 𝑢𝑛 (𝑥) = sup𝑛 𝑢𝑛 (𝑥) = 𝑢(𝑥) for all 𝑥.
and is thus finite, even uniformly in 𝑛. So, Beppo Levi’s theorem tells us that
103
R.L. Schilling: Measures, Integrals & Martingales
showing integrability.
Now 𝑢 is clearly not bounded but integrable.
■■
Problem 10.3 Solution: Clearly, 𝜈 is defined on 𝒜 and takes values in [0, ∞]. Since 1∅ ≡ 0 we have
𝜈(∅) = 1∅ ⋅ 𝑢 𝑑𝜇 = 0 𝑑𝜇 = 0.
∫ ∫
If (𝐴𝑛 )𝑛∈N ⊂ 𝒜 are mutually disjoint measurable sets, we get
∞
(⨃ )
𝜈 𝐴𝑛 = 1⨃∞ 𝐴𝑛 ⋅ 𝑢 𝑑𝜇
∫ 𝑛=1
𝑛=1
∞
∑
= 1𝐴𝑛 ⋅ 𝑢 𝑑𝜇
∫
𝑛=1
∞ ∞
∑ ∑
= 1𝐴𝑛 ⋅ 𝑢 𝑑𝜇 = 𝜈(𝐴𝑛 )
∫
𝑛=1 𝑛=1
Problem 10.4 Solution: ‘⇐⇒’: since the 𝐴𝑗 are disjoint we get the identities
∞ ∞
∑ ∑
1⨃ 𝑗 𝐴𝑗 = 1𝐴𝑗 and so 𝑢⋅ 1⨃ 𝑗 𝐴𝑗 = 𝑢 ⋅ 1𝐴𝑗 ,
𝑘=1 𝑘=1
hence |𝑢1𝐴𝑛 | = |𝑢|1𝐴𝑛 ⩽ |𝑢|1⨃𝑗 𝐴𝑗 = |𝑢1⨃𝑗 𝐴𝑗 | showing the integrability of each 𝑢1𝐴𝑛 by Theorem
10.3. By a Beppo Levi argument (Theorem 9.6) or, directly, by Corollary 9.9 we get
∞ ∞ ∞
∑ ∑ ∑
|𝑢| 𝑑𝜇 = |𝑢|1𝐴𝑗 𝑑𝜇 = |𝑢|1𝐴𝑗 𝑑𝜇
∫𝐴 𝑗 ∫ ∫
𝑗=1 𝑗=1 𝑗=1
= |𝑢|1⨃𝑗 𝐴𝑗 𝑑𝜇 < ∞.
∫
The converse direction ‘⇐⇐’ follows again from Corollary 9.9, now just the other way round:
∞ ∞
∑ ∑
|𝑢|1⨃𝑗 𝐴𝑗 𝑑𝜇 = |𝑢|1𝐴𝑗 𝑑𝜇 = |𝑢|1𝐴𝑗 𝑑𝜇
∫ ∫ ∫
𝑗=1 𝑗=1
∞
∑
= |𝑢| 𝑑𝜇 < ∞
∫𝐴𝑗
𝑗=1
Problem 10.5 Solution: For any measurable function 𝑢 we have 𝑢 ∈ 1 (𝜇) ⇐⇒ |𝑢| ∈ 1 (𝜇). This
means that we may assume that 𝑢 ⩾ 0. Since
𝑘
∑
1{2𝑛 ⩽𝑢<2𝑛+1 } 𝑢 ↑ 𝑢1{𝑢>0}
𝑛=−𝑘
104
Solution Manual. Last update 28th October 2021
i.e.
∑
𝐶⩽ 𝑢 𝑑𝜇 ⩽ 2𝐶.
∫{2𝑛 ⩽𝑢<2𝑛+1 }
𝑛∈Z
Therefore the following assertions are equivalent:
∑
𝑢 ∈ 1 (𝜇) ⇐⇒ 𝑢 𝑑𝜇 < ∞
∫{2𝑛 ⩽𝑢<2𝑛+1 }
𝑛∈Z
∑
⇐⇒ 𝐶 = 2𝑛 𝜇{2𝑛 ⩽ 𝑢 < 2𝑛+1 } < ∞.
𝑛∈Z
■■
First proof:
∞ ∞ ∞ ∞ 𝑘 ∞
∑ ∑ ∑ ∑ ∑ ∑
1{|𝑢|⩾𝑖} = 1{𝑘+1>|𝑢|⩾𝑘} = 1{𝑘+1>|𝑢|⩾𝑘} = 𝑘1{𝑘+1>|𝑢|⩾𝑘}
𝑖=1 𝑖=1 𝑘=𝑖 𝑘=1 𝑖=1 𝑘=1
and
∞ ∞
∑ ∑
𝑘1{𝑘+1>|𝑢|⩾𝑘} ⩽ |𝑢|1{𝑘+1>|𝑢|⩾𝑘} = |𝑢|1{|𝑢|⩾1}
𝑘=1 𝑘=1
and
∞ ∞
∑ ∑
𝑘1{𝑘+1>|𝑢|⩾𝑘} ⩾ (|𝑢| − 1)1{𝑘+1>|𝑢|⩾𝑘} = (|𝑢| − 1)1{|𝑢|⩾1} ⩾ |𝑢|1{|𝑢|⩾1} − 1{|𝑢|⩾0} .
𝑘=1 𝑘=1
So,
∞ ∞ ∞
∑ ∑ ∑
1{|𝑢|⩾𝑖} ⩽ |𝑢|1{|𝑢|⩾1} ⩽ |𝑢| ⩽ 1 + 1{|𝑢|⩾𝑖} = 1{|𝑢|⩾𝑖} .
𝑖=1 𝑖=1 𝑖=0
Second proof: For 𝑥 ∈ 𝑋, there is some 𝑘 ∈ N0 such that 𝑘 ⩽ |𝑢(𝑥)| < 𝑘 + 1. Therefore,
𝑥 ∈ {|𝑢| ⩾ 𝑖} ∀𝑖 ∈ {0, … , 𝑘}
and
𝑥 ∉ {|𝑢| ⩾ 𝑖} ∀𝑖 ⩾ 𝑘 + 1.
105
R.L. Schilling: Measures, Integrals & Martingales
Thus,
∑
1{|𝑢|⩾𝑖} (𝑥) = 𝑘 + 1.
𝑖∈N0
∑
and (ii) follows. If 𝑢 ∈ 1 (𝜇), then we get 𝑖⩾1 𝜇(|𝑢| ⩾ 1) < ∞. On the other hand, if 𝑢 is
∑
measurable, and 𝑖 𝜇(|𝑢| ⩾ 𝑖) < ∞, then we get ∫ |𝑢| 𝑑𝜇 < ∞, i.e. 𝑢 ∈ 1 (𝜇) and (i) follows.
The finiteness of the measure 𝜇 was only used for ∫ 1 𝑑𝜇 < ∞ or 𝜇{|𝑢| ⩾ 0} < ∞ – which is only
needed for the second estimate in (ii). Hence, the lower estimate in (ii) holds for all measures!
■■
Problem 10.7 Solution: One possibility to solve the problem is to follow the hint. We provide an
alternative (and shorter) solution.
(i) Observe that 𝑢𝑗 − 𝑣 ⩾ 0 is a sequence of positive and integrable functions. Applying Fatou’s
lemma (in the usual form) yields (observing the rules for lim inf, lim sup from Appendix A,
compare also Problem 9.10):
= lim inf 𝑢𝑗 𝑑𝜇 − 𝑣 𝑑𝜇
𝑗 ∫ ∫
and the claim follows upon subtraction of the finite (!) number ∫ 𝑣 𝑑𝜇.
(ii) Very similar to (i) by applying Fatou’s lemma to the positive, integrable functions 𝑤−𝑢𝑗 ⩾ 0:
⩽ lim inf (𝑤 − 𝑢𝑗 ) 𝑑𝜇
𝑗 ∫
= 𝑤 𝑑𝜇 − lim sup 𝑢𝑗 𝑑𝜇
∫ 𝑗 ∫
106
Solution Manual. Last update 28th October 2021
Consider Lebesgue measure on R. Put 𝑓𝑗 (𝑥) = −1[−2𝑗,−𝑗] (𝑥) and 𝑔𝑗 (𝑥) = 1[𝑗,2𝑗] (𝑥).
Then lim inf 𝑓𝑗 (𝑥) = 0 and lim sup 𝑔𝑗 (𝑥) = 0 for every 𝑥 and neither admits an integrable
minorant resp. majorant.
Remark. Here is an even stronger version of Fatou’s Lemma. For this we introduced the extended
integrable functions
{ }
1 + −
(𝜇) ∶= 𝑢 ∈ (𝒜 ) ∶ 𝑢 𝑑𝜇 < ∞, 𝑢 𝑑𝜇 < ∞
∫ ∫
{ }
1,𝑒 (𝜇) ∶= 𝑢 ∈ (𝒜 ) ∶ 𝑢+ 𝑑𝜇 ∈ [0, ∞], 𝑢− 𝑑𝜇 < ∞ .
∫ ∫
Let (𝑢𝑛 )𝑛∈N ⊂ (𝒜 ) such that 𝑢𝑛 ⩾ 𝑢 for some 𝑢 ∈ 1,𝑒 (𝜇).
iii) if lim inf 𝑛→∞ ∫ 𝑢𝑛 𝑑𝜇 < ∞, then lim inf 𝑛→∞ 𝑢𝑛 ∈ 1 (𝜇).
Proof. i) We have
⎧( )+
⎪ lim inf 𝑛 𝑢𝑛 ⩾ 𝑢+
𝑢𝑛 ⩾ 𝑢 ⇐⇒ lim inf 𝑢𝑛 ⩾ 𝑢 ⇐⇒ ⎨( )−
𝑛
⎪ lim inf 𝑛 𝑢𝑛 ⩽ 𝑢−
⎩
( )−
and so ∫ lim inf 𝑛 𝑢𝑛 𝑑𝜇 ⩽ ∫ 𝑢− 𝑑𝜇 < ∞, i.e. lim inf 𝑛 𝑢𝑛 ∈ 1,𝑒 (𝜇).
Adding on both sides ∫ 𝑢 𝑑𝜇 – this is possible since we do not get an expression of type
“∞ − ∞”, we get
lim inf 𝑢𝑛 𝑑𝜇 ⩾ lim inf 𝑢𝑛 𝑑𝜇.
𝑛 ∫ ∫ 𝑛
iii) We have
( )+ ( )−
lim inf 𝑢𝑛 𝑑𝜇 = lim inf 𝑢𝑛 + lim inf 𝑢𝑛 𝑑𝜇
∫ 𝑛 ∫ 𝑛 𝑛
⩽ lim inf 𝑢𝑛 + 𝑢− 𝑑𝜇
∫ 𝑛
= lim inf 𝑢𝑛 𝑑𝜇 + 𝑢− 𝑑𝜇
∫ 𝑛 ∫
107
R.L. Schilling: Measures, Integrals & Martingales
This proves the claim. (Note that in the inequality-step in the last formula we could have used
directly the ordinary Fatou lemma, and not step ii), as 𝑢𝑛 + 𝑢− ⩾ 0).
■■
𝑢𝑣 𝑑𝑃 = 𝑃 (𝐴 ∩ 𝐵) = 𝑃 (𝐴)𝑃 (𝐵) = 𝑢 𝑑𝑃 𝑣 𝑑𝑃 .
∫ ∫ ∫
∑ ∑
For positive, simple functions 𝑢 = 𝑗 𝛼𝑗 1𝐵𝑗 and 𝑣 = 𝑘 𝛽𝑘 1𝐶𝑘 we find
∑
𝑢𝑣 𝑑𝑃 = 𝛼𝑗 𝛽𝑘 1𝐴𝑗 1𝐵𝑘 𝑑𝑃
∫ ∫
𝑗,𝑘
∑
= 𝛼𝑗 𝛽𝑘 𝑃 (𝐴𝑗 ∩ 𝐵𝑘 )
𝑗,𝑘
∑
= 𝛼𝑗 𝛽𝑘 𝑃 (𝐴𝑗 )𝑃 (𝐵𝑘 )
𝑗,𝑘
( )( )
∑ ∑
= 𝛼𝑗 𝑃 (𝐴𝑗 ) 𝛽𝑘 𝑃 (𝐵𝑘 )
𝑗 𝑘
= 𝑢 𝑑𝑃 𝑣 𝑑𝑃 .
∫ ∫
For measurable 𝑢 ∈ + (ℬ) and 𝑣 ∈ + (𝒞 ) we use approximating simple functions 𝑢𝑘 ∈
+ (ℬ), 𝑢𝑘 ↑ 𝑢, and 𝑣𝑘 ∈ + (𝒞 ), 𝑣𝑘 ↑ 𝑣. Then, by Beppo Levi,
= 𝑢 𝑑𝑃 𝑣 𝑑𝑃 .
∫ ∫
𝑢𝑣 𝑑𝑃 = 𝑢 𝑑𝑃 𝑣 𝑑𝑃 .
∫ ∫ ∫
Counterexample: Just take 𝑢 = 𝑣 which are integrable but not square integrable, e.g. 𝑢(𝑥) =
𝑣(𝑥) = 𝑥−1∕2 . Then ∫(0,1) 𝑥−1∕2 𝑑𝑥 < ∞ but ∫(0,1) 𝑥−1 𝑑𝑥 = ∞, compare also Problem 10.2.
■■
108
Solution Manual. Last update 28th October 2021
(ii) Part (i) shows that a map ℎ ∶ 𝐸 → C is 𝒜 ∕𝒞 -measurable if, and only if, 𝑔◦ℎ ∶ 𝐸 → R2
is 𝒜 ∕ℬ(R2 )-measurable.
Indeed: The map ℎ ∶ (𝐸, 𝒜 ) → (C, 𝒞 ) is, by definition, measureable if ℎ−1 (𝐴) ∈ 𝒜 for
all 𝐴 ∈ 𝒞 . Since 𝒞 = 𝑔 −1 (ℬ(R2 )), this is the same as ℎ−1 (𝑔 −1 (𝐵)) = (𝑔◦ℎ)−1 (𝐵) ∈ 𝒜
for all 𝐵 ∈ ℬ(R2 ), hence it is the same as the measurability of 𝑔◦ℎ.
"⇒": Assume that ℎ ∶ 𝐸 → C is 𝒜 ∕𝒞 -measurable. Then we have that
( )
Re ℎ
(𝑔◦ℎ) =
Im ℎ
(𝑔 + ℎ) 𝑑𝜇 = Re(𝑔 + ℎ) 𝑑𝜇 + 𝑖 Im(𝑔 + ℎ) 𝑑𝜇
∫ ∫ ∫
= 𝑔 𝑑𝜇 + ℎ 𝑑𝜇.
∫ ∫
Note that we have used the R-linearity of the integral for real-valued functions. The
homogeneity of the complex integral is shown in a very similar way.
(iv) Since Re ℎ and Im ℎ are real, we get ∫ Re ℎ 𝑑𝜇 ∈ R and ∫ Im ℎ 𝑑𝜇 ∈ R. Therefore,
( ) ( )
Re ℎ 𝑑𝜇 = Re Re ℎ 𝑑𝜇 + 𝑖 Im ℎ 𝑑𝜇
∫ ∫ ∫
= Re ℎ 𝑑𝜇.
∫
Similarly, we see
( ) ( )
Im ℎ 𝑑𝜇 = Im Re ℎ 𝑑𝜇 + 𝑖 Im ℎ 𝑑𝜇
∫ ∫ ∫
= Im ℎ 𝑑𝜇.
∫
109
R.L. Schilling: Measures, Integrals & Martingales
(v) We follow the hint: as ∫ ℎ 𝑑𝜇 ∈ C we can pick some 𝜃 ∈ (−𝜋, 𝜋] such that 𝑒𝑖𝜃 ∫ ℎ 𝑑𝜇 ⩾
0. Thus, (iii) and (iv) entail
| |
| ℎ 𝑑𝜇| = 𝑒𝑖𝜃 ℎ 𝑑𝜇
|∫ | ∫
| | ( )
= Re 𝑒 𝑖𝜃
ℎ 𝑑𝜇
∫
= Re(𝑒𝑖𝜃 ℎ) 𝑑𝜇
∫
⩽ |𝑒𝑖𝜃 ℎ| 𝑑𝜇
∫
= |ℎ| 𝑑𝜇.
∫
(vi) We know from (ii) that ℎ ∶ (𝐸, 𝒜 ) → (C, 𝒞 ) is measurable if, and only if, Re ℎ and
Im ℎ are 𝒜 ∕ℬ(R2 )-measurable. If Re ℎ and Im ℎ are 𝜇-integrable, then so is
√
|ℎ| = (Re ℎ)2 + (Im ℎ)2 ⩽ | Re ℎ| + | Im ℎ|.
If |ℎ| ∈ 1R (𝜇), then we conclude from | Re ℎ| ⩽ |ℎ| and | Im ℎ| ⩽ |ℎ|, that Re ℎ and
Im ℎ are 𝜇-integrable.
■■
110
11 Null sets and the ‘almost everywhere’.
Solutions to Problems 11.1–11.12
Problem 11.1 Solution: True, we can change an integrable function on a null set, even by setting it to
the value +∞ or −∞ on the null set. This is just the assertion of Theorem 11.2 and its Corollaries
11.3, 11.4.
■■
Problem 11.2 Solution: We have seen that a single point is a Lebesgue null set: {𝑥} ∈ ℬ(R) for
all 𝑥 ∈ R and 𝜆({𝑥}) = 0, see e.g. Problems 4.13 and 6.7. If 𝑁 is countable, we know that
⨃
𝑁 = {𝑥𝑗 ∶ 𝑗 ∈ N} = 𝑗∈N {𝑥𝑗 } and by the 𝜎-additivity of measures
( )
⨃ ∑ ( ) ∑
𝜆(𝑁) = 𝜆 {𝑥𝑗 } = 𝜆 {𝑥𝑗 } = 0 = 0.
𝑗∈N 𝑗∈N 𝑗∈N
The Cantor set 𝐶 from Problem 7.12 is, as we have seen, uncountable but has measure 𝜆(𝐶) = 0.
This means that there are uncountable sets with measure zero.
In R2 and for two-dimensional Lebesgue measure 𝜆2 the situation is even easier: every line 𝐿 in
the plane has zero Lebesgue measure and 𝐿 contains certainly uncountably many points. That
𝜆2 (𝐿) = 0 is seen from the fact that 𝐿 differs from the ordinate {(𝑥, 𝑦) ∈ R2 ∶ 𝑥 = 0} only
by a rigid motion 𝑇 which leaves Lebesgue measure invariant (see Chapter 4, Theorem 4.7) and
𝜆2 ({𝑥 = 0}) = 0 as seen in Problem 6.7.
■■
(i) Since {|𝑢| > 𝑐} ⊂ {|𝑢| ⩾ 𝑐} and, therefore, 𝜇({|𝑢| > 𝑐}) ⩽ 𝜇({|𝑢| ⩾ 𝑐}), this follows
immediately from Proposition 11.5. Alternatively, one could also mimic the proof of this
Proposition or use part (iii) of the present problem with 𝜙(𝑡) = 𝑡, 𝑡 ⩾ 0.
(ii) This will follow from (iii) with 𝜙(𝑡) = 𝑡𝑝 , 𝑡 ⩾ 0, since 𝜇({|𝑢| > 𝑐}) ⩽ 𝜇({|𝑢| ⩾ 𝑐}) as
{|𝑢| > 𝑐} ⊂ {|𝑢| ⩾ 𝑐}.
111
R.L. Schilling: Measures, Integrals & Martingales
𝜙(|𝑢(𝑥)|)
= 1 (𝑥) 𝜇(𝑑𝑥)
∫ 𝜙(|𝑢(𝑥)|) {𝑥∶𝜙(|𝑢(𝑥)|)⩾𝜙(𝑐)}
𝜙(|𝑢(𝑥)|)
⩽ 1{𝑥∶𝜙(|𝑢(𝑥)|)⩾𝜙(𝑐)} (𝑥) 𝜇(𝑑𝑥)
∫ 𝜙(𝑐)
𝜙(|𝑢(𝑥)|)
⩽ 𝜇(𝑑𝑥)
∫ 𝜙(𝑐)
1
= 𝜙(|𝑢(𝑥)|) 𝜇(𝑑𝑥)
𝜙(𝑐) ∫
(iv) Let us set 𝑏 = 𝛼 ∫ 𝑢 𝑑𝜇. Then we follow the argument of (iii), where we use that 𝑢 and 𝑏 are
strictly positive.
(v) Using the fact that 𝜓 is decreasing we get {|𝑢| < 𝑐} = {𝜓(|𝑢|) > 𝜓(𝑐)}—mind the change
of the inequality sign—and going through the proof of part (iii) again we use there that 𝜙
increases only in the first step in a similar role as we used the decrease of 𝜓 here! This means
that the argument of (iii) is valid after this step and we get, altogether,
112
Solution Manual. Last update 28th October 2021
■■
Problem 11.4 Solution: We mimic the proof of Corollary 11.6. Set 𝑁 = {|𝑢| = ∞} = {|𝑢|𝑝 = ∞}.
⋂
Then 𝑁 = 𝑘∈N {|𝑢|𝑝 ⩾ 𝑘} and using Markov’s inequality (MI) and the ‘continuity’ of measures,
Proposition 4.3(vii), we find
( )
⋂ 4.3(vii)
𝜇(𝑁) = 𝜇 {|𝑢|𝑝 ⩾ 𝑘} = lim 𝜇({|𝑢|𝑝 ⩾ 𝑘})
𝑘→∞
𝑘∈N
MI 1
⩽ lim |𝑢|𝑝 𝑑𝜇 = 0.
𝑘→∞ 𝑘∫
⏟⏞⏞⏟⏞⏞⏟
<∞
For arctan this is not any longer true for several reasons:
• ... arctan is odd and changes sign, so there could be cancelations under the integral.
• ... even if we had no cancelations we have the problem that the points where 𝑢(𝑥) = ∞ are
now transformed to points where arctan(𝑢(𝑥)) = 𝜋
2
and we do not know how the measure
𝜇 acts under this transformation. A simple example: Take 𝜇 to be a measure of total finite
mass (that is: 𝜇(𝑋) < ∞), e.g. a probability measure, and take the function 𝑢(𝑥) which is
constantly 𝑢 ≡ +∞. Then arctan(𝑢(𝑥)) = 𝜋
2
throughout, and we get
𝜋 𝜋 𝜋
arctan 𝑢(𝑥) 𝜇(𝑑𝑥) = 𝑑𝜇 = 𝑑𝜇 = 𝜇(𝑋) < ∞,
∫ ∫ 2 2∫ 2
but 𝑢 is nowhere finite!
■■
(i) Assume that 𝑓 ∗ is 𝒜 -measurable. The problem at hand is to construct 𝒜 -measurable up-
per and lower functions 𝑔 and 𝑓 . For positive simple functions this is clear: if 𝑓 ∗ (𝑥) =
∑𝑁
𝑗=0 𝜙𝑗 1𝐵 ∗ (𝑥) with 𝜙𝑗 ⩾ 0 and 𝐵𝑗 ∈ 𝒜 , then we can use Problem 4.15(v) to find 𝐵𝑗 , 𝐶𝑗 ∈
∗
𝑗
𝒜 with 𝜇(𝐶𝑗 ⧵ 𝐵𝑗 ) = 0
113
R.L. Schilling: Measures, Integrals & Martingales
𝑓 𝑑𝜇 = 𝑓 ∗ 𝑑 𝜇̄ = 𝑔 𝑑𝜇.
∫ ∫ ∫
(ii), (iii) Assume that 𝑢∗ is 𝒜 ∗ -measurable; without loss of generality (otherwise consider pos-
itive and negative parts) we can assume that 𝑢∗ ⩾ 0. Because of Theorem 8.8 we know that
𝑓𝑘∗ ↑ 𝑢∗ for 𝑓𝑘∗ ∈ + (𝒜 ∗ ). Now choose the corresponding 𝒜 -measurable lower and upper
functions 𝑓𝑘 , 𝑔𝑘 constructed in part (i). By considering, if necessary, max{𝑓1 , … , 𝑓𝑘 } we
can assume that the 𝑓𝑘 are increasing.
= 𝑢∗ 𝑑 𝜇̄
∫
⩽ 𝑣 𝑑𝜇.
∫
= lim inf 𝑔𝑘 𝑑𝜇
𝑘 ∫
⩾ lim inf 𝑔𝑘 𝑑𝜇
∫ 𝑘
= 𝑣 𝑑𝜇
∫
⩾ 𝑢 𝑑𝜇
∫
This proves that ∫ 𝑢 𝑑𝜇 = ∫ 𝑣 𝑑𝜇 = ∫ 𝑢∗ 𝑑𝜇. This answers part (iii) by considering positive
and negative parts.
It remains to show that {𝑢 ≠ 𝑣} is a 𝜇-null set. (This does not follow from the above integral
equality, cf. Problem 11.10!) Clearly, {𝑢 ≠ 𝑣} = {𝑢 < 𝑣}, i.e. if 𝑥 ∈ {𝑢 < 𝑣} is fixed, we
deduce that, for sufficiently large values of 𝑘,
but the RHS is a countable union of 𝜇-null sets, hence a null set itself.
114
Solution Manual. Last update 28th October 2021
but {𝑣 > 𝛼}, {𝑢 > 𝛼} ∈ 𝒜 and {𝑢 > 𝛼} ⧵ {𝑣 > 𝛼} ⊂ {𝑢 ≠ 𝑣} is a 𝜇-null set. Because of
Problem 4.15 we conclude that {𝑢∗ > 𝛼} ∈ 𝒜 ∗ .
■■
Problem 11.6 Solution: Throughout the solution the letters 𝐴, 𝐵 are reserved for sets from 𝒜 .
(i) a) Let 𝐴 ⊂ 𝐸 ⊂ 𝐵. Then 𝜇(𝐴) ⩽ 𝜇(𝐵) and going to the sup𝐴⊂𝐸 and inf 𝐸⊂𝐵 proves
𝜇∗ (𝐸) ⩽ 𝜇 ∗ (𝐸).
|𝜇∗ (𝐸 𝑐 ) − 𝜇(𝐴𝑐 )| ⩽ 𝜖.
Thus,
|𝜇(𝑋) − 𝜇∗ (𝐸) − 𝜇∗ (𝐸 𝑐 )|
⩽ |𝜇∗ (𝐸) − 𝜇(𝐴)| + |𝜇∗ (𝐸 𝑐 ) − 𝜇(𝐴𝑐 )|
⩽ 2𝜖,
Now we pass on the right-hand side, separately, to the inf 𝐴⊃𝐸 and inf 𝐵⊃𝐹 , and obtain
𝜇∗ (𝐸 ∪ 𝐹 ) ⩽ 𝜇∗ (𝐸) + 𝜇 ∗ (𝐹 ).
Now we pass on the right-hand side, separately, to the sup𝐴⊂𝐸 and sup𝐵⊂𝐹 , where we
stipulate that 𝐴 ∩ 𝐵 = ∅, and obtain
𝜇∗ (𝐸 ⊍ 𝐹 ) ⩾ 𝜇∗ (𝐸) + 𝜇∗ (𝐹 ).
115
R.L. Schilling: Measures, Integrals & Martingales
but the latter is the completed 𝜎-algebra 𝒜 ∗ . That 𝜇∗ | ∗ = 𝜇∗ | ∗ = 𝜇̄ is now trivial since
| |
|𝒜 |𝒜
𝜇∗ and 𝜇∗ coincide on 𝒜 ∗ .
■■
Problem 11.7 Solution: Let 𝐴 ∈ 𝒜 and assume that there are non-measurable sets, i.e. 𝒫 (𝑋) ⊋ 𝒜 .
Take some 𝑁 ∉ 𝒜 which is a 𝜇-null set. Assume also that 𝑁 ∩ 𝐴 = ∅. Then 𝑢 = 1𝐴 and
𝑤 ∶= 1𝐴 + 2 ⋅ 1𝑁 are a.e. identical, but 𝑤 is not measurable.
This means that 𝑤 is only measurable if, e.g. all (subsets of) null sets are measurable, that is if
(𝑋, 𝒜 , 𝜇) is complete.
■■
Problem 11.8 Solution: The function 1Q is nowhere continuous but 𝑢 = 0 Lebesgue almost every-
where. That is
{𝑥 ∶ 1Q (𝑥) is discontinuous} = R
while
that is 1Q coincides a.e. with a continuous function but is itself at no point continuous!
The same analysis for 1[0,∞) yields that
which is a Lebesgue null set, but 1[0,∞) cannot coincide a.e. with a continuous function! This,
namely, would be of the form 𝑤 = 0 on (−∞, −𝛿) and 𝑤 = 1 on (𝜖, ∞) while it ‘interpolates’
somehow between 0 and 1 if −𝛿 < 𝑥 < 𝜖. But this entails that
116
Solution Manual. Last update 28th October 2021
Problem 11.9 Solution: Let (𝐴𝑗 )𝑗∈N ⊂ 𝒜 be an exhausting sequence 𝐴𝑗 ↑ 𝑋 such that 𝜇(𝐴𝑗 ) < ∞.
Set
∞
∑ 1
𝑓 (𝑥) ∶= 1𝐴𝑗 (𝑥).
𝑗=1
2𝑗 (𝜇(𝐴 𝑗) + 1)
Then 𝑓 is measurable, 𝑓 (𝑥) > 0 everywhere, and using Beppo Levi’s theorem
( ∞ )
∑ 1
𝑓 𝑑𝜇 = 1 𝑑𝜇
∫ ∫
𝑗=1
2𝑗 (𝜇(𝐴𝑗 ) + 1) 𝐴𝑗
∞
∑ 1
= 1 𝑑𝜇
𝑗=1
2𝑗 (𝜇(𝐴𝑗 ) + 1) ∫ 𝐴𝑗
∞
∑ 𝜇(𝐴𝑗 )
=
𝑗=1
2𝑗 (𝜇(𝐴𝑗 ) + 1)
∞
∑
⩽ 2−𝑗 = 1.
𝑗=1
Thus, set 𝑃 (𝐴) ∶= ∫𝐴 𝑓 𝑑𝜇. We know from Problem 9.7 that 𝑃 is indeed a measure.
If 𝑁 ∈ 𝒩𝜇 , then, by Theorem 11.2,
11.2
𝑃 (𝑁) = 𝑓 𝑑𝜇 = 0
∫𝑁
so that 𝒩𝜇 ⊂ 𝒩𝑃 .
Conversely, if 𝑀 ∈ 𝑃 , we see that
𝑓 𝑑𝜇 = 0
∫𝑀
but since 𝑓 > 0 everywhere, it follows from Theorem 11.2 that 1𝑀 ⋅ 𝑓 = 0 𝜇-a.e., i.e. 𝜇(𝑀) = 0.
Thus, 𝒩𝑃 ⊂ 𝒩𝜇 .
Remark. We will see later (cf. Chapter 20 or Chapter 25, Radon–Nikodým theorem) that 𝒩𝜇 = 𝒩𝑃
if and only if 𝑃 = 𝑓 ⋅ 𝜇 (i.e., if 𝑃 has a density w.r.t. 𝜇) such that 𝑓 > 0.
■■
Problem 11.10 Solution: Well, the hint given in the text should be good enough.
■■
𝑢 𝑑𝜇 = 𝑤 𝑑𝜇 ⇐⇒ (𝑢+ + 𝑤− ) 𝑑𝜇 = (𝑢− + 𝑤+ ) 𝑑𝜇
∫𝐶 ∫𝐶 ∫𝐶 ∫𝐶
holds for all 𝐶 ∈ 𝒞 . The right-hand side can be read as the equality of two measures 𝐴 →
∫𝐴 (𝑢+ + 𝑤− ) 𝑑𝜇, 𝐴 → ∫𝐴 (𝑢− + 𝑤+ ) 𝑑𝜇, 𝐴 ∈ 𝒜 which coincide on a generator 𝒞 which satisfies
the conditions of the uniqueness theorem of measures (Theorem 5.7). This shows that
𝑢 𝑑𝜇 = 𝑤 𝑑𝜇 ∀𝐴 ∈ 𝒜 .
∫𝐴 ∫𝐴
117
R.L. Schilling: Measures, Integrals & Martingales
The converse implication ‘⇐’ follows directly from Corollary 11.6 applied to 𝑢1𝐶 and 𝑤1𝐶 .
■■
(i) “⊂”: Let 𝑥 ∈ 𝐶𝑓 , i.e. 𝑓 (𝑥) = lim𝑛→∞ 𝑓𝑛 (𝑥) exists; in particular, (𝑓𝑛 (𝑥))𝑛∈N is Cauchy:
for all 𝑘 ∈ N there is some 𝓁 ∈ N such that
1
|𝑓𝑛 (𝑥) − 𝑓𝑚 (𝑥)| ⩽ ∀ 𝑚, 𝑛 ⩾ 𝓁.
𝑘
⋂ ⋃ ⋂
This shows that 𝑥 ∈ 𝑘∈N 𝓁∈N ∞ 1
𝑛,𝑚=𝓁 {|𝑓𝑛 (𝑥) − 𝑓𝑚 (𝑥)| ⩽ 𝑘 }.
⋂ ⋃ ⋂
“⊃”: Assume that 𝑘∈N 𝓁∈N ∞ 𝑛,𝑚=𝓁 {|𝑓𝑛 (𝑥) − 𝑓𝑚 (𝑥)| ⩽ 𝑘 }. This means that for
1
1
|𝑓𝑛 (𝑥) − 𝑓𝑚 (𝑥)| ⩽ ∀ 𝑚, 𝑛 ⩾ 𝓁.
𝑘
This shows that (𝑓𝑛 (𝑥))𝑛∈N is a Cauchy sequence in R. The claim follows since R is
complete.
Observe that
∞ ∞
⋃ ⋂
𝐴𝑘𝑛 ↑ {|𝑓𝑚 (𝑥) − 𝑓 (𝑥)| ⩽ 𝑘1 } ⊃ 𝐶𝑓
𝓁=1 𝑚=𝓁
118
Solution Manual. Last update 28th October 2021
i.e. for all 𝑥 ∈ 𝐴𝜖 and 𝑘 ∈ N there is some 𝓁(𝑥) ⩽ 𝑛(𝑘, 𝜖) such that
1
|𝑓 (𝑥) − 𝑓𝑚 (𝑥)| ⩽ ∀ 𝑚 ⩾ 𝓁(𝑥).
𝑘
Since 𝓁(𝑥) ⩽ 𝑛(𝑘, 𝜖) we get, in particular,
1
|𝑓 (𝑥) − 𝑓𝑚 (𝑥)| ⩽ ∀𝑥 ∈ 𝐴𝜖 , 𝑚 ⩾ 𝑛(𝑘, 𝜖).
𝑘
Since 𝑘 ∈ N is arbitrary, the uniform convergence 𝐴𝜖 follows.
(iv) Consider one-dimensional Lebesgue measure, set 𝑓 (𝑥) ∶= |𝑥| and 𝑓𝑛 (𝑥) ∶= |𝑥|1[−𝑛,𝑛] .
Then we have 𝑓𝑛 (𝑥) ↑ 𝑓 (𝑥) for every 𝑥, but the set {|𝑓𝑛 − 𝑓 | > 𝜖} = [−𝑛, 𝑛]𝑐 has
infinite measure for any 𝜖 > 0.
■■
119
12 Convergence theorems and their
applications.
Solutions to Problems 12.1–12.37
Because of this we find that |𝑢𝑗 − 𝑢|𝑝 ⩽ 2𝑝 𝑔 𝑝 and the right-hand side is an integrable dominating
function.
Proof alternative 1: Apply Theorem 12.2 on dominated convergence to the sequence 𝜙𝑗 ∶= |𝑢𝑗 −
𝑢|𝑝 of integrable functions. Note that 𝜙𝑗 (𝑥) → 0 and that 0 ⩽ 𝜙𝑗 ⩽ Φ where Φ = 2𝑝 𝑔 𝑝 is integrable
and independent of 𝑗. Thus,
= 0 𝑑𝜇 = 0.
∫
Since the limit lim𝑗 𝜓𝑗 exists, it coincides with lim inf 𝑗 𝜓𝑗 , and so we can use Fatou’s Lemma to
get
2𝑝 𝑔 𝑝 𝑑𝜇 = lim inf 𝜓𝑗 𝑑𝜇
∫ ∫ 𝑗→∞
⩽ lim inf 𝜓𝑗 𝑑𝜇
𝑗→∞ ∫
( )
= lim inf 2𝑝 𝑔 𝑝 − |𝑢𝑗 − 𝑢|𝑝 𝑑𝜇
𝑗→∞ ∫
121
R.L. Schilling: Measures, Integrals & Martingales
( )
= 2 𝑔 𝑑𝜇 + lim inf
𝑝 𝑝
− |𝑢𝑗 − 𝑢| 𝑑𝜇
𝑝
∫ 𝑗→∞ ∫
where we use that lim inf 𝑗 (−𝛼𝑗 ) = − lim sup𝑗 𝛼𝑗 . This shows that lim sup𝑗 ∫ |𝑢𝑗 − 𝑢|𝑝 𝑑𝜇 = 0,
hence
showing that lower and upper limit coincide and equal to 0, hence lim𝑗 ∫ |𝑢𝑗 − 𝑢|𝑝 𝑑𝜇 = 0.
■■
Problem 12.2 Solution: Assume that, as in the statement of Theorem 12.2, 𝑢𝑗 → 𝑢 and that |𝑢𝑗 | ⩽
𝑓 ∈ 1 (𝜇). In particular,
−𝑓 ⩽ 𝑢𝑗 and 𝑢𝑗 ⩽ 𝑓
𝑢 𝑑𝜇 = lim inf 𝑢𝑗 𝑑𝜇
∫ ∫ 𝑗→∞
∗
⩽ lim inf 𝑢𝑗 𝑑𝜇
𝑗→∞ ∫
⩽ lim sup 𝑢𝑗 𝑑𝜇
𝑗→∞ ∫
∗
⩽ lim sup 𝑢𝑗 𝑑𝜇 = 𝑢 𝑑𝜇.
∫ 𝑗→∞ ∫
■■
0 ⩽ 𝑓𝑘 − 𝑔𝑘 ←←←←←←←←←→
← 𝑓 − 𝑔,
𝑘→∞
0 ⩽ 𝐺𝑘 − 𝑓𝑘 ←←←←←←←←←→
← 𝐺 − 𝑓.
𝑘→∞
(𝑓 − 𝑔) 𝑑𝜇 = lim(𝑓𝑘 − 𝑔𝑘 ) 𝑑𝜇
∫ ∫ 𝑘
122
Solution Manual. Last update 28th October 2021
(𝐺 − 𝑓 ) 𝑑𝜇 = lim(𝐺𝑘 − 𝑓𝑘 ) 𝑑𝜇
∫ ∫ 𝑘
Problem 12.4 Solution: Using Beppo Levi’s theorem in the form of Corollary 9.9 we find
∞ ∞
∑ ∑
|𝑢𝑗 | 𝑑𝜇 = |𝑢𝑗 | 𝑑𝜇 < ∞, (*)
∫ ∫
𝑗=1 𝑗=1
∑∞ ∑∞
which means that the positive function 𝑗=1 |𝑢𝑗 | is finite almost everywhere, i.e. the series 𝑗=1 𝑢𝑗
converges (absolutely) almost everywhere.
∑𝑘
In order to show the second part, we want to apply dominated convergence. Set 𝑣𝑘 ∶= 𝑛=1 𝑢𝑛
and notte that
|∑ | ∞
| 𝑘 | ∑ 𝑘
∑
| |
|𝑣𝑘 | = | 𝑢𝑛 | ⩽ |𝑢𝑛 | ⩽ |𝑢𝑛 | ⩽ 𝑤 ∈ ℒ 1 (𝜇).
| 𝑛=1 | 𝑛=1
| | 𝑛=1
∑∞
Clearly, 𝑣𝑘 → 𝑢 = 𝑛=1 𝑢𝑛 as 𝑘 → ∞. Thus, we get with dominated convergence
∞ 𝑘
∑ ∑
𝑢𝑛 𝑑𝜇 = 𝑢 𝑑𝜇 = lim 𝑣𝑘 𝑑𝜇 = lim 𝑣𝑘 𝑑𝜇 = lim 𝑢𝑛 𝑑𝜇
∫ ∫ ∫ 𝑘→∞ 𝑘→∞ ∫ 𝑘→∞ ∫
𝑛=1 𝑛=1
∞
∑
= 𝑢𝑛 𝑑𝜇.
∫
𝑛=1
■■
Problem 12.5 Solution: Since 1 (𝜇) ∋ 𝑢𝑗 ↓ 0 we find by monotone convergence, Theorem 12.1,
that ∫ 𝑢𝑗 𝑑𝜇 ↓ 0. Therefore,
∞ ∞
∑ ∑
𝜎= (−1) 𝑢𝑗 and 𝑆 =
𝑗
(−1)𝑗 𝑢𝑗 𝑑𝜇 converge
∫
𝑗=1 𝑗=1
123
R.L. Schilling: Measures, Integrals & Martingales
𝑁 𝑁
∑ ∑
(−1)𝑗 𝑢𝑗 𝑑𝜇 = (−1)𝑗 𝑢𝑗 𝑑𝜇 ←←←←←←←←←←←→
← 𝑆.
∫ ∫ 𝑁→∞
𝑗=1 𝑗=1
All that remains is to show that the right-hand side converges to ∫ 𝜎 𝑑𝜇. Observe that for 𝑆𝑁 ∶=
∑𝑁
𝑗=1 (−1) 𝑢𝑗 we have
𝑗
𝑆2𝑁 ⩽ 𝑆2𝑁+2 ⩽ … ⩽ 𝑆
■■
Problem 12.6 Solution: Consider 𝑢𝑗 (𝑥) ∶= 𝑗 ⋅ 1(0,1∕𝑗) (𝑥), 𝑗 ∈ N. It is clear that 𝑢𝑗 is measurable and
Lebesgue integrable with integral
1
𝑢𝑗 𝑑𝜆 = 𝑗 =1 ∀ 𝑗 ∈ N.
∫ 𝑗
so that 0 = ∫ 𝑢 𝑑𝜆 = ∫ lim𝑗 𝑢𝑗 𝑑𝜆 ≠ 1.
The example does not contradict dominated convergence as there is no uniform dominating integ-
rable function.
Alternative: a similar situation can be found for 𝑣𝑘 (𝑥) ∶= 𝑘1 1[0,𝑘] (𝑥) and the pointwise limit 𝑣 ≡ 0.
Note that in this case the limit is even uniform and still lim𝑘 ∫ 𝑣𝑘 𝑑𝜆 = 1 ≠ 0 = ∫ 𝑣 𝑑𝜆. Again
there is no contradiction to dominated convergence as there does not exist a uniform dominating
integrable function.
■■
Problem 12.7 Solution: Using the majorant (𝑒−𝑟𝑥 ⩽ 1 ∈ 1 (𝜇), 𝑟, 𝑥 ⩾ 0) we find with dominated
convergence
■■
124
Solution Manual. Last update 28th October 2021
lim |𝑢| 𝑑𝜆 = 0.
𝑅→∞ ∫𝐵 (0)𝑐
𝑅
|𝑢| 𝑑𝜆 ⩽ 𝜖.
∫𝐵𝑅 (0)𝑐
Since 𝐾 is compact (in fact: bounded), there is some 𝑟 = 𝑟(𝑅) > 0, such that 𝑥 + 𝐾 ⊂
𝐵𝑅 (0)𝑐 for all 𝑥 satisfying |𝑥| ⩾ 𝑟. Thus, we have
(ii) Fix 𝜖 > 0. By assumption, 𝑢 is uniformly continuous. Therefore, there is some 𝛿 > 0 such
that
|𝑢(𝑦) − 𝑢(𝑥)| ⩽ 𝜖 ∀ 𝑥 ∈ R𝑛 , 𝑦 ∈ 𝑥 + 𝐾 ∶= 𝑥 + 𝐵𝛿 (0) = 𝐵𝛿 (𝑥).
Hence,
1
|𝑢(𝑥)|𝑝 = |𝑢(𝑥)|𝑝 𝑑𝜆(𝑦)
𝜆(𝐾 + 𝑥) ∫𝐾+𝑥
1 ( )𝑝
⩽ |𝑢(𝑦) − 𝑢(𝑥)| +|𝑢(𝑦)| 𝑑𝜆(𝑦).
𝜆(𝐾) ∫𝐾+𝑥 ⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
⩽𝜖
we get for 𝐶 = 2𝑝
( )
𝐶
|𝑢(𝑥)| ⩽𝑝
𝜖 𝑑𝜆(𝑦) +
𝑝
|𝑢(𝑦)| 𝑑𝜆(𝑦)
𝜆(𝐾) ∫𝐾+𝑥 ∫𝐾+𝑥
𝜆(𝐾 + 𝑥) 𝐶
⩽ 𝐶𝜖 𝑝 + |𝑢(𝑦)| 𝑑𝜆(𝑦).
𝜆(𝐾) 𝜆(𝐾) ∫𝐾+𝑥
⏟⏞⏞⏟⏞⏞⏟
1
𝜖→0
lim sup |𝑢(𝑥)|𝑝 ⩽ 𝐶𝜖 𝑝 ←←←←←←←→
← 0
|𝑥|→∞
■■
125
R.L. Schilling: Measures, Integrals & Martingales
(i) Fix 𝜖 > 0, 𝑅 > 0 and consider 𝐵 ∶= {|𝑢| ⩽ 𝑅}. By definition, sup𝑥∈𝐵 |𝑢(𝑥)| < ∞. On
the other hand, dominated convergence and Corollary 11.6 show that
In particular, we can choose 𝑅 so large, that ∫𝐵 |𝑢(𝑥)| 𝑑𝑥 < 𝜖. Using Markov’s inequality
(Proposition 11.5) yields
1
𝜆(𝐵) = 𝜆{|𝑢| ⩾ 𝑅} ⩽ |𝑢(𝑥)| 𝑑𝑥 < ∞.
𝑅∫
(ii) Fix 𝜖 > 0 and let 𝐵 ∈ ℬ(R𝑛 ) be as in (i). Further, let 𝐴 ∈ ℬ(R𝑛 ) with 𝜆(𝐴) < 𝜖. Then
we have
⩽ sup |𝑢(𝑥)| ⋅ 𝜖 + 𝜖.
𝑥∈𝐵
lim |𝑢| 𝑑𝜆 = 0.
𝜆(𝐴)→0 ∫𝐴
■■
(i) From 𝑢𝑛 ∈ 1 (𝜇) and ‖𝑢𝑛 − 𝑢‖∞ ⩽ 1 (for all sufficiently large 𝑛) we infer
lim 𝑢𝑛 𝑑𝜇 = 1 ≠ 0 = 𝑢 𝑑𝜇.
𝑛→∞ ∫ ∫
■■
126
Solution Manual. Last update 28th October 2021
Problem 12.11 Solution: Without loss of generality we assume that 𝑢 is increasing. Because of the
monotonicity of 𝑢, we find for every sequence (𝑎𝑛 )𝑛∈N ⊂ (0, 1) such that 𝑎𝑛 ↓ 0, that
1 1 1 1
lim 𝑢(𝑡𝑛 ) 𝑑𝑡 = inf 𝑢(𝑡𝑛 ) 𝑑𝑡 = inf 𝑢(𝑡𝑛 ) 𝑑𝑡 = 𝑢(0+) 𝑑𝑡 = 𝑢(0+).
𝑛→∞ ∫0 𝑛∈N ∫0 ∫0 𝑛∈N ∫0
■■
Problem 12.12 Solution: Set 𝑢𝑛 (𝑡) ∶= 𝑡𝑛 𝑢(𝑡), 𝑡 ∈ (0, 1). Since |𝑡𝑛 | ⩽ 1 for 𝑡 ∈ (0, 1), we have
Since 𝑡𝑛 ←←←←←←←←←→
← 0 for all 𝑡 ∈ (0, 1) and |𝑓 (𝑡)| < ∞ a.e. (Corollary 11.6), we have |𝑢𝑛 (𝑡)| → 0 a.e.
𝑛→∞
An application of dominated convergence (Theorem 12.2 and Remark 12.3) yields
1 1 1
lim 𝑡𝑛 𝑢(𝑡) 𝑑𝑡 = lim 𝑢𝑛 (𝑡) 𝑑𝑡 = lim 𝑢 (𝑡) 𝑑𝑡 = 0.
𝑛→∞ ∫0 𝑛→∞ ∫0 ∫0 𝑛→∞ 𝑛
⏟⏞⏟⏞⏟
0
■■
∑
Problem 12.13 Solution: From the geometric series we know that 1
1−𝑥
= 𝑛⩾0 𝑥
𝑛 for 𝑥 ∈ [0, 1).
This implies that for all 𝑡 > 0
1 1 1 ∑ ∑
−𝑡 −𝑡 𝑛
= = 𝑒 (𝑒 ) = 𝑒−𝑛𝑡
𝑒𝑡 − 1 𝑒𝑡 1 − 𝑒−𝑡 𝑛⩾0 𝑛⩾1
∑𝑘
(observe that 𝑒−𝑡 < 1 for 𝑡 > 0!). Set 𝑢𝑘 (𝑡) ∶= sin(𝑡) ⋅ 𝑛=1 𝑒
−𝑛𝑡 , then we get the estimate
|∑ |
| 𝑘 −𝑛𝑡 | ∑𝑘
∑ | sin 𝑡|
| |
|𝑢𝑘 (𝑡)| ⩽ | sin 𝑡| ⋅ | 𝑒 | = | sin 𝑡| 𝑒−𝑛𝑡 ⩽ | sin 𝑡| 𝑒−𝑛𝑡 = 𝑡 (∗)
| 𝑛=1 | 𝑒 −1
| | 𝑛=1 𝑛⩾1
for all 𝑘 ∈ N und 𝑡 > 0. Using the elementary inequalities 𝑒𝑡 − 1 ⩾ 𝑡 (𝑡 ⩾ 0) and 𝑒𝑡 − 1 ⩾ 𝑒𝑡∕2
(𝑡 ⩾ 1) we see
|𝑢𝑘 (𝑡)| ⩽ 1[0,1] (𝑡) + 𝑒−𝑡∕2 1(1,∞) (𝑡) =∶ 𝑤(𝑡).
Let us now show that 𝑤 ∈ 1 (0, ∞). This can be done with Beppo Levi’s theorem:
∞ 1 ∞
𝑤(𝑡) 𝑑𝑡 = 𝑤(𝑡) 𝑑𝑡 + 𝑤(𝑡) 𝑑𝑡
∫0 ∫0 ⏟⏟⏟ ∫1 ⏟⏟⏟
1 𝑒−𝑡∕2
𝑛 [ ]𝑛
= 1 + sup 𝑒−𝑡∕2 𝑑𝑡 = 1 + sup − 2𝑒−𝑡∕2 𝑡=1 < ∞.
𝑛∈N ∫1 𝑛∈N
127
R.L. Schilling: Measures, Integrals & Martingales
We use here that every Riemann-integrable function 𝑓 ∶ [𝑎, 𝑏] → C, −∞ < 𝑎 < 𝑏 < ∞,
is Lebesgue integrable and that Riemann and Lebesgue intgrals coincide (in this case, see The-
orem 12.8). By dominated convergence,
∞ ∞ 𝑘 ∞
sin(𝑡) ∑
𝑑𝑡 = lim 𝑢 𝑘 (𝑡) 𝑑𝑡 = lim sin(𝑡)𝑒−𝑛𝑡 𝑑𝑡.
∫0 𝑒𝑡 − 1 𝑘→∞ ∫0 𝑘→∞
𝑛=1
∫0
■■
∑
Problem 12.14 Solution: We know that the exponential function is given by 𝑒𝑧𝑥 = 𝑛⩾0
(𝑧𝑥)𝑛
𝑛!
. Thus,
𝑘
∑ (𝑧𝑥)𝑛
𝑢𝑘 (𝑥) ∶= 𝑢(𝑥) ← 𝑢(𝑥)𝑒𝑧𝑥 .
←←←←←←←←←→
𝑛=0
𝑛! 𝑘→∞
= lim 𝑢𝑘 (𝑥) 𝑑𝑥
𝑘→∞ ∫
𝑘
∑ 1
= lim (𝑧𝑥)𝑛 𝑢(𝑥) 𝑑𝑥
𝑘→∞
𝑛=0
𝑛! ∫
∞ 𝑛
∑ 𝑧
= 𝑥𝑛 𝑢(𝑥) 𝑑𝑥.
𝑛=0
𝑛! ∫
■■
128
Solution Manual. Last update 28th October 2021
Problem 12.15 Solution: We get ||∫𝐴 𝑢 𝑑𝜇|| ⩽ ∫𝐴 |𝑢| 𝑑𝜇 straight from the triangle inequality. There-
fore, it is enough to prove the second estimate. Fix 𝜖 > 0.
Solution 1: The Sombrero lemma ensures that there is a sequence (𝑢𝑛 )𝑛∈N ⊂ (𝒜 ) with |𝑢𝑛 | ⩽ |𝑢|
and lim𝑛→∞ 𝑢𝑛 = 𝑢 (Corollary 8.9). From dominated convergence we get ∫ |𝑢𝑛 − 𝑢| 𝑑𝜇 ←←←←←←←←←→
← 0;
𝑛→∞
in particular, we can choose 𝑛 ∈ N such that ∫ |𝑢𝑛 − 𝑢| 𝑑𝜇 ⩽ 𝜖. Since each 𝑢𝑛 is bounded (b/o the
definition of a simple function) we get
for any 𝐴 ∈ 𝒜 with 𝜇(𝐴) < 𝛿 ∶= 𝜖∕‖𝑢𝑛 ‖∞ . Using the triangle inequality we get
Solution 2: Obviously,
We estimate each term by itself. For the first expression on the RHS we use Beppo Levi:
|𝑢| 𝑑𝜇 ←←←←←←←←←←→
← |𝑢| 𝑑𝜇.
∫𝐴∩{|𝑢|⩾𝑅} 𝑅→∞ ∫𝐴∩{|𝑢|=∞}
By assumption, 𝑢 ∈ 1 (𝜇), we get 𝜇(|𝑢| = ∞) = 0 (see the proof of Corollaryr 11.6) and we get
with Theorem 11.2,
|𝑢| 𝑑𝜇 = 0.
∫𝐴∩{|𝑢|=∞}
Therefore, we can pick some 𝑅 > 0 with
|𝑢| 𝑑𝜇 ⩽ 𝜖.
∫𝐴∩{|𝑢|⩾𝑅}
|𝑢| 𝑑𝜇 ⩽ 𝑅 1 𝑑𝜇 ⩽ 𝑅𝜇(𝐴).
∫𝐴∩{|𝑢|<𝑅} ∫𝐴∩{|𝑢|<𝑅}
■■
Problem 12.16 Solution: Let 𝜇 be an arbitrary Borel measure on the line R and define the integral
function for some 𝑢 ∈ 1 (𝜇) through
129
R.L. Schilling: Measures, Integrals & Martingales
For any sequence 0 < 𝑙𝑗 → 𝑥, 𝑙𝑗 < 𝑥 from the left and 𝑟𝑘 → 𝑥, 𝑟𝑘 > 𝑥 from the right we find
Remark: the proof shows, by the way, that 𝐼𝜇𝑢 (𝑥) is always left-continuous at every 𝑥, no matter
what 𝜇 or 𝑢 look like.
■■
(i) We have
1
1 (𝑥) 𝑑𝑥
∫ 𝑥 [1,∞)
1
= lim 1 (𝑥) 𝑑𝑥 by Beppo Levi’s thm.
𝑛→∞ ∫ 𝑥 [1,𝑛)
1
= lim 𝑑𝑥 usual shorthand
𝑛→∞ ∫[1,𝑛) 𝑥
𝑛 𝑛
1
= lim (𝑅) 𝑑𝑥 Riemann- exists
𝑛→∞ ∫1 𝑥 ∫1
[ ]𝑛
= lim log 𝑥 1
𝑛→∞
(ii) We have
1
1 (𝑥) 𝑑𝑥
∫ 𝑥2 [1,∞)
1
= lim 1 (𝑥) 𝑑𝑥 by Beppo Levi’s thm.
𝑛→∞ ∫ 𝑥2 [1,𝑛)
1
= lim 𝑑𝑥 usual shorthand
𝑛→∞ ∫[1,𝑛) 𝑥2
𝑛 𝑛
1
= lim (𝑅) 𝑑𝑥 Riemann- exists
𝑛→∞ ∫1 𝑥2 ∫1
130
Solution Manual. Last update 28th October 2021
[ ]𝑛
= lim − 𝑥1
𝑛→∞ 1
1
= lim [1 − 𝑛
] =1<∞
𝑛→∞
=2<∞
=∞
131
R.L. Schilling: Measures, Integrals & Martingales
This shows that exp(−𝑥𝛼 ) ⩽ 1(0,1) + 𝑀𝑥−2 1[1,∞) with the right-hand side being integrable.
■■
Problem 12.19 Solution: Take 𝛼 ∈ (𝑎, 𝑏) where 0 < 𝑎 < 𝑏 < ∞ are fixed (but arbitrary). We show
that the function is continuous for these 𝛼. This shows the general case since continuity is a local
property and we can ‘catch’ any given 𝛼0 by some choice of 𝑎 and 𝑏’s.
We use the Continuity lemma (Theorem 12.4) and have to find uniform (for 𝛼 ∈ (𝑎, 𝑏)) dominating
( )3
bounds on the integrand function 𝑓 (𝛼, 𝑥) ∶= sin𝑥 𝑥 𝑒−𝛼𝑥 . First of all, we remark that | sin𝑥 𝑥 | ⩽ 𝑀
| |
| |
which follows from the fact that sin𝑥 𝑥 is a continuous function such that lim𝑥→∞ sin𝑥 𝑥 = 0 and
lim𝑥↓0 sin 𝑥
𝑥
= 1. (Actually, we could choose 𝑀 = 1...). Moreover, exp(−𝛼𝑥) ⩽ 1 for 𝑥 ∈ (0, 1)
and exp(−𝛼𝑥) ⩽ 𝐶𝑎,𝑏 𝑥−2 for 𝑥 ⩾ 1—use for this the continuity of 𝑥2 exp(−𝛼𝑥) and the fact that
lim𝑥→∞ 𝑥2 exp(−𝛼𝑥) = 0. This shows that
( )
|𝑓 (𝛼, 𝑥)| ⩽ 𝑀 1(0,1) (𝑥) + 𝐶𝑎,𝑏 𝑥−2 1[1,∞) (𝑥)
and the right-hand side is an integrable dominating function which does not depend on 𝛼—as long
as 𝛼 ∈ (𝑎, 𝑏). But since 𝛼 → 𝑓 (𝛼, 𝑥) is obviously continuous, the Continuity lemma applies and
proves that ∫(0,∞) 𝑓 (𝛼, 𝑥) 𝑑𝑥 is continuous.
■■
Problem 12.20 Solution: Fix some number 𝑁 > 0 and take 𝑥 ∈ (−𝑁, 𝑁). We show that 𝐺(𝑥) is
continuous on this set. Since 𝑁 was arbitrary, we find that 𝐺 is continuous for every 𝑥 ∈ R.
132
Solution Manual. Last update 28th October 2021
(use in the last equality that {0} is a Lebesgue null set). Thus, by a Beppo Levi-argument (and
using that Riemann=Lebesgue whenever the Riemann integral over a compact interval exists...)
𝑛
1 1
𝐺′ (0) = 𝑑𝑡 = lim (𝑅) 𝑑𝑡
∫R 1 + 𝑡 2 𝑛→∞ ∫−𝑛 1 + 𝑡2
= lim [tan−1 (𝑡)]𝑛−𝑛
𝑛→∞
= 𝜋.
■■
133
R.L. Schilling: Measures, Integrals & Martingales
(ii) Note that 𝑥 → ln 𝑥 is continuous and bounded in [𝜖, 1], thus Riemann integrable. It is easy
to see that 𝑥 ln 𝑥 − 𝑥 is a primitive for ln 𝑥. The improper Riemann integral
1
ln 𝑥 𝑑𝑥 = lim[𝑥 ln 𝑥 − 𝑥]1𝜖 = −1
∫0 𝜖→0
exists and, since ln 𝑥 is negative throughout (0, 1), improper Riemann and Lebesgue integrals
coincide. Thus, ln 𝑥 ∈ 𝐿1 (𝑑𝑥, (0, 1)).
Therefore,
|( )𝑘 |
| 1 − 𝑥 ln 𝑥| ⩽ | ln 𝑥|, ∀ 𝑥 ∈ (0, 1)
| 𝑘 |
| |
is uniformly dominated by an integrable function and we can use dominated convergence to
get
( ) ( )
𝑥 𝑘 𝑥 𝑘
lim 1− ln 𝑥 𝑑𝑥 = lim 1 − ln 𝑥 𝑑𝑥
𝑘 ∫(0,1) 𝑘 ∫(0,1) 𝑘 𝑘
= 𝑒−𝑥 ln 𝑥 𝑑𝑥
∫(0,1)
■■
Problem 12.22 Solution: Since the integrand of 𝐹 (𝑡) is continuous and bounded by the integrable
function 𝑒−𝑥 , 𝑥 > 0, it is clear that 𝐹 (𝑡) exists. With the usual approximation argument,
𝑛
𝑡 𝑡
𝑒−𝑥 𝜆(𝑑𝑥) = lim 𝑒−𝑥 2 𝑑𝑥
∫(0,∞) 𝑡2 + 𝑥2 𝑛→∞ ∫1∕𝑛 𝑡 + 𝑥2
(the right-hand side is a Riemann integral) we can use the classical (Riemann) rules to evaluate
the integral. Thus, a change of variables 𝑥 = 𝑡 ⋅ 𝑦 ⇐⇒ 𝑑𝑥 = 𝑡 𝑑𝑦 yields
𝑡
𝐹 (𝑡) = 𝑒−𝑥 𝜆(𝑑𝑥)
∫(0,∞) 𝑡2 + 𝑥2
𝑡
= 𝑒−𝑡𝑦 𝑡 𝜆(𝑑𝑦)
∫(0,∞) 𝑡 + (𝑡𝑦)2
2
134
Solution Manual. Last update 28th October 2021
1
= 𝑒−𝑡𝑦 𝜆(𝑑𝑦).
∫(0,∞) 1 + 𝑦2
Observe that
| −𝑡𝑦 1 |
|𝑒 |⩽ 1 uniformly for all 𝑡 > 0,
| 1 + 𝑦 2 | 1 + 𝑦2
| |
and that the right-hand side is Lebesgue integrable (the primitive is the arctan). Therfore, we can
use dominated convergence to conclude
1
𝐹 (0+) = lim 𝑒−𝑡𝑦 𝜆(𝑑𝑦)
𝑡↓0 ∫(0,∞) 1 + 𝑦2
1
= lim 𝑒−𝑡𝑦 𝜆(𝑑𝑦)
∫(0,∞) 𝑡↓0 1 + 𝑦2
1
= 𝜆(𝑑𝑦)
∫(0,∞) 1 + 𝑦2
𝑛
1
= lim 𝑑𝑦
1 + 𝑦2
𝑛→∞ ∫1∕𝑛
[ ]𝑛 𝜋
= lim arctan 𝑦 1∕𝑛 = .
𝑛→∞ 2
■■
Problem 12.23 Solution: For the existence of the integrals we need |𝑒−𝑖⋅𝜉 | ∈ 1 (𝜇) and |𝑒−𝑖⋅𝜉 | ⋅
|𝑢(⋅)| ∈ 1 (𝑑𝑥). Since |𝑒−𝑖⋅𝜉 | = 1, it is reasonable to require that 𝜇 is a finite measure (such that
the constant 1 is integrable) or 𝑢 ∈ 1 (𝑑𝑥). Under these assumptions, the continuity of the Fourier
transform follows directly from the continuity lemma: set
1 −𝑖 𝑥𝜉
𝑓 (𝜉, 𝑥) ∶= 𝑒 , 𝜉 ∈ R, 𝑥 ∈ R.
2𝜋
By assumption, |𝑓 (𝑥, 𝜉)| ⩽ (2𝜋)−1 ∈ 1 (𝜇) and 𝜉 → 𝑓 (𝜉, 𝑥) is continuous. Using Theorem 12.4,
we get the continuity of the map
135
R.L. Schilling: Measures, Integrals & Martingales
(i) Let 𝑡 ∈ (−𝑅, 𝑅) for some 𝑅 > 0. Since |𝜙(𝑥) − 𝑡| ⩽ |𝜙(𝑥)| + |𝑡| ⩽ |𝜙(𝑥)| + 𝑅 ∈
1 ([0, 1], 𝑑𝑥) and since 𝑡 → |𝜙(𝑥)−𝑡| is continuous, the continuity lemma, Theorem 12.4,
shows that the mapping
| |
|𝑓 (𝑡) − 𝑓 (𝑠)| ⩽ ||𝜙(𝑥) − 𝑡| − |𝜙(𝑥) − 𝑠|| 𝑑𝑥 ⩽ |𝑠 − 𝑡| 𝑑𝑥 = |𝑠 − 𝑡|,
∫[0,1] | | ∫[0,1]
= 𝑑𝑥 = 𝜆(𝜙 ⩽ 𝑡 − ℎ) ←←←←←←←←→
← 𝜆{𝜙 < 𝑡}.
∫𝜙⩽𝑡−ℎ ℎ→0
Similarly,
(𝜙(𝑥) − (𝑡 + ℎ) − (𝜙(𝑥) − 𝑡)
𝐼3 (ℎ) = 𝑑𝑥
∫𝜙⩾𝑡−ℎ ℎ
= 𝜆(𝜙 ⩾ 𝑡 + ℎ) ←←←←←←←←→
← 𝜆{𝜙 > 𝑡}.
ℎ→0
𝑓 (𝑥 + ℎ) − 𝑓 (𝑥)
lim = 𝜆{𝜙 > 𝑡} + 𝜆{𝜙 < 𝑡}.
ℎ→0 ℎ
136
Solution Manual. Last update 28th October 2021
Obviously, we have
|ℎ| |ℎ|
𝐼22 (ℎ) = 1 𝑑𝑥 = 𝜆{𝜙 = 𝑡}
ℎ ∫{𝜙=𝑡} ℎ
lim 𝐼 1 (ℎ) = 0.
ℎ→0 2
|ℎ|
lim 𝐼22 (ℎ) = 𝜆(𝜙 = 𝑡) lim
ℎ→0 ℎ→0 ℎ
■■
(i) The map 𝑡 → 𝑢(𝑡, 𝑥) ∶= 𝑥−2 sin2 (𝑥)𝑒−𝑡𝑥 is continuous on [0, ∞) and differentiable on
(0, ∞). Because of the continuity and differentiability lemmas (Theorem 12.4 and 12.5)
it is enough to find suitable majorants for the function and its derivatives. Fix 𝑡 ⩾ 0.
Using the elementary inequalities sin 𝑥
𝑥
⩽ 1 and 𝑒−𝑡𝑥 ⩽ 1 we get
1
|𝑢(𝑡, 𝑥)| ⩽ 1[0,1] (𝑥) + 1(1,∞) (𝑥) =∶ 𝑤(𝑥).
𝑥2
Since 𝑤 ∈ 1 ([0, ∞)) (cf. Beispiel 12.14), continuity follows from the continuity
lemma. Assume now that 𝑡 ∈ (𝑟, ∞) for some 𝑟 > 0. Then we get
| sin2 (𝑥) |
| −𝑡𝑥 |
|𝜕𝑡 𝑢(𝑡, 𝑥)| = | (−𝑥)𝑒 |
| 𝑥2 |
| |
⩽ 1[0,1] (𝑥) + 𝑥𝑒 1[1,∞) (𝑥) ∈ 1 ([0, ∞))
−𝑡𝑥
| sin2 (𝑥) |
| 2 −𝑡𝑥 |
|𝜕𝑡2 𝑢(𝑡, 𝑥)| = | (−𝑥) 𝑒 |
| 𝑥2 |
| |
⩽ 1[0,1] (𝑥) + 𝑥 𝑒 1[1,∞) (𝑥) ∈ 1 ([0, ∞)).
2 −𝑡𝑥
137
R.L. Schilling: Measures, Integrals & Martingales
Now the differentiability lemma shows that 𝑓 has two derivatives which are given by
∞
sin2 (𝑥) −𝑡𝑥
𝑓 ′ (𝑡) = − 𝑒 𝑑𝑥,
∫0 𝑥
∞
𝑓 ′′ (𝑡) = sin2 (𝑥)𝑒−𝑡𝑥 𝑑𝑥.
∫0
(ii) In order to calculate 𝑓 ′′ we use that Riemann and Lebesgue integrals auszurechnen
coincide if a function is Riemann integrable (Theorem 12.8).
(iii) We begin with a closed expression for 𝑓 ′ : from the fundamental theorem of (Riemann)
integration we know
𝑅
𝑓 ′ (𝑅) − 𝑓 ′ (𝑡) = 𝑓 ′′ (𝑠) 𝑑𝑠.
∫𝑡
Letting 𝑅 → ∞ we get using (ii)
𝑅
𝑓 ′ (𝑡) = − lim 𝑓 ′′ (𝑠) 𝑑𝑠
𝑅→∞ ∫𝑡
[ ]𝑅
1 1
=− lim log 𝑠 − log(𝑠2 + 4)
2 𝑅→∞ 2 𝑠=𝑡
( )
1 1 2
= log 𝑡 − log(𝑡 + 4)
2 2
1 𝑡
= log √ .
2 𝑡 +4
2
138
Solution Manual. Last update 28th October 2021
Finally,
𝑅 ∞
1 𝑠
𝑓 (𝑡) = − lim 𝑓 ′ (𝑠) 𝑑𝑠 = − log √ 𝑑𝑠.
𝑅→∞ ∫𝑡 2 ∫𝑡 𝑠2 + 4
(In this part we have again used the fact that the Lebesgue integral extends the Riemann
integral.)
■■
Problem 12.26 Solution: We follow the hint: since 𝑒−𝑡𝑥 ⩾ 0 we can use Beppo Levi to get
∞ 𝑛 𝑛
𝑒−𝑥𝑡 𝑑𝑥 = sup 𝑒−𝑥𝑡 𝑑𝑥 = lim 𝑒−𝑥𝑡 𝑑𝑥.
∫0 𝑛∈N ∫0 𝑛→∞ ∫0
Thus, 𝑒−𝑡𝑥 ∈ 1 (0, ∞) and ∫0 𝑒−𝑥𝑡 𝑑𝑥 = . Now we use the differentiability lemma, The-
∞ 1
𝑡
orem 12.5. For 𝑢(𝑡, 𝑥) ∶= 𝑒−𝑡𝑥 we have
Since 𝑎 > 0 is arbitrary, we get differentiability on (0, ∞). Iterating this argument, we inver that
we can swap derivatives of any order with the integral. Morover,
( ∞ ) ( )
𝑑𝑛 −𝑥𝑡 𝑑𝑛 1
𝑒 𝑑𝑥 =
𝑑𝑡𝑛 ∫ 𝑑𝑡𝑛 𝑡
( ∞ 0 )
(−1)𝑛 𝑛!
⇒ (−𝑥)𝑛 𝑒−𝑥𝑡 𝑑𝑥 = 𝑛+1 .
∫0 𝑡
If 𝑡 = 1, the claim follows.
■■
Problem 12.27 Solution: Throughout we fix (𝑎, 𝑏) ⊂ (0, ∞) and take 𝑡 ∈ (𝑎, 𝑏). As in Problem 12.17
we get
(i) Note that differentiability implies continuity, so it suffices to show that Γ is 𝑚 times
differentiable for every 𝑚.
Induction Hypothesis: Γ(𝑚) exists and is of the form as claimed in the statement of the
problem.
139
R.L. Schilling: Measures, Integrals & Martingales
We have now to find a uniform (for 𝑡 ∈ (𝑎, 𝑏)) integrable dominating function for
|𝜕𝑡 𝛾(𝑡, 𝑥)|. Since log 𝑥 ⩽ 𝑥 for all 𝑥 > 0 (the logarithm is a concave function!),
| −𝑥 𝑡−1 |
|𝑒 𝑥 log 𝑥| = 𝑒−𝑥 𝑥𝑡−1 log 𝑥
| |
⩽ 𝑒−𝑥 𝑥𝑡 ⩽ 𝑒−𝑥 𝑥𝑏 ⩽ 𝐶𝑏 𝑥−2 ∀ 𝑥 ⩾ 1, 𝑡 ∈ (𝑎, 𝑏)
(for the last step multiply with 𝑥2 and use that 𝑥𝜌 𝑒−𝑥 is continuous for every 𝜌 > 0 and
lim𝑥→∞ 𝑥𝜌 𝑒−𝑥 = 0 to find 𝐶𝑏 ). Moreover,
| −𝑥 𝑡−1 |
|𝑒 𝑥 log 𝑥| ⩽ 𝑥𝑎−1 | log 𝑥|
| |
1
= 𝑥𝑎−1 log ⩽ 𝐶𝑎 𝑥−1∕2 ∀ 𝑥 ∈ (0, 1), 𝑡 ∈ (𝑎, 𝑏)
𝑥
where we use the fact that lim𝑥→0 𝑥𝜌 log 𝑥1 = 0 which is easily seen by the substitution
𝑥 = 𝑒−𝑢 and 𝑢 → ∞ and the continuity of the function 𝑥𝜌 log 𝑥1 .
Induction Step 𝑚 ⇝ 𝑚 + 1: Set 𝛾 (𝑚) (𝑡, 𝑥) = 𝑒−𝑥 𝑥𝑡−1 (log 𝑥)𝑚 . We want to apply the dif-
ferentiability lemma to Γ(𝑚) (𝑥). With very much the same arguments as in the induction
start we find that 𝛾 (𝑚+1) (𝑡, 𝑥) = 𝜕𝑡 𝛾 (𝑚) (𝑡, 𝑥) exists (obvious) and satisfies the following
bounds
| −𝑥 𝑡−1 |
|𝑒 𝑥 (log 𝑥)𝑚+1 | = 𝑒−𝑥 𝑥𝑡−1 (log 𝑥)𝑚+1
| |
⩽ 𝑒−𝑥 𝑥𝑡+𝑚
⩽ 𝑒−𝑥 𝑥𝑏+𝑚
⩽ 𝐶𝑏,𝑚 𝑥−2 ∀ 𝑥 ⩾ 1, 𝑡 ∈ (𝑎, 𝑏)
| −𝑥 𝑡−1 |
|𝑒 𝑥 (log 𝑥)𝑚+1 | ⩽ 𝑥𝑎−1 | log 𝑥|𝑚+1
| |
( )
1 𝑚+1
= 𝑥𝑎−1 log
𝑥
⩽ 𝐶𝑎,𝑚 𝑥−1∕2 ∀ 𝑥 ∈ (0, 1), 𝑡 ∈ (𝑎, 𝑏)
140
Solution Manual. Last update 28th October 2021
(ii) Using a combination of Beppo Levi (indicated by ‘BL’), Riemann=Lebesgue (if the
Riemann integral over an interval exists) and integration by parts (for the Riemann in-
tegral, indicated by ‘parts’) techniques we get
log Γ(𝜆𝑡 + (1 − 𝜆)𝑠) ⩽ 𝜆 log Γ(𝑡) + (1 − 𝜆) log Γ(𝑠) ∀𝑠, 𝑡 > 0, 𝜆 ∈ (0, 1).
(iii) Alternative – direct calculuation Since log and Γ are in 𝐶 2 we can apply the convexity
criterion: log Γ is convex if, and only if, 𝑑2
𝑑𝑡2
log Γ(𝑡) ⩾ 0 holds. We have
𝑑 Γ′ (𝑡)
log Γ(𝑡) =
𝑑𝑡 Γ(𝑡)
𝑑 2 Γ(𝑡)Γ′′ (𝑡) − (Γ′ (𝑡))2
log Γ(𝑡) =
𝑑𝑡2 (Γ(𝑡))2
which is non-negative iff
!
0 ⩽ Γ(𝑡)Γ′′ (𝑡) − (Γ′ (𝑡))2
141
R.L. Schilling: Measures, Integrals & Martingales
So with the notation from part (ii), along with the dominated convergence theorem
(indicated by ‘DC’ – this is needed for Γ′ , since its integrand will take negative values,
so Beppo Levi does not apply), we get
Γ(𝑡)Γ′′ (𝑡) − (Γ′ (𝑡))2 = lim 𝑒−𝑥−𝑦 (𝑥𝑦)𝑡−1 (log 𝑦)2 𝑑𝑥 𝑑𝑦 (BL)
𝑛→∞ ∫(1∕𝑛,𝑛) ∫(1∕𝑛,𝑛)
142
Solution Manual. Last update 28th October 2021
∞
= (−1)𝑘 𝑡𝑘 𝑒−𝑡(𝑘+1) 𝑑𝑡
∫0
∞( )𝑘
𝑠 𝑑𝑠
= (−1)𝑘 𝑒−𝑠
∫0 𝑘+1 𝑘+1
( )𝑘+1 ∞
1
= (−1)𝑘 𝑠(𝑘+1)−1 𝑒−𝑠 𝑑𝑠
𝑘+1 ∫0
( )𝑘+1
1
= (−1)𝑘 Γ(𝑘 + 1).
𝑘+1
Since for 𝑥 ∈ (0, 1) the terms under the sum are all positive, we can use Beppo Levi’s theorem
and the formula Γ(𝑘 + 1) = 𝑘! to get
∞
∑ 1
−𝑥
𝑥 𝑑𝑥 = (−1)𝑘 (𝑥 ln 𝑥)𝑘 𝑑𝑥
∫(0,1)
𝑘=0
𝑘! ∫(0,1)
∞ ( )𝑘+1
∑ 1 1
= (−1)𝑘 (−1)𝑘 Γ(𝑘 + 1)
𝑘=0
𝑘! 𝑘+1
∞ ( )𝑘+1
∑ 1
=
𝑘=0
𝑘+1
∞ ( )
∑ 1 𝑛
= .
𝑛=1
𝑛
■■
Problem 12.29 Solution: Fix (𝑎, 𝑏) ⊂ (0, 1) and let always 𝑢 ∈ (𝑎, 𝑏). We have for 𝑥 ⩾ 0 and 𝐿 ∈ N0
| 𝑒𝑢𝑥 |
|𝑥𝐿 𝑓 (𝑢, 𝑥)| = |𝑥|𝐿 || 𝑥 |
|
|𝑒 + 1|
𝑒𝑢𝑥
= 𝑥𝐿 𝑥
𝑒 +1
𝑒𝑢𝑥
⩽ 𝑥𝐿 𝑥
𝑒
= 𝑥𝐿 𝑒(𝑢−1)𝑥
⩽ 1[0,1] (𝑥) + 𝑀𝑎,𝑏 1(1,∞) (𝑥) 𝑥−2
where we use that 𝑢 − 1 < 0, the continuity and boundedness of 𝑥𝜌 𝑒−𝑎𝑥 for 𝑥 ∈ [1, ∞) and 𝜌 ⩾ 0.
If 𝑥 ⩽ 0 we get
| 𝑒𝑢𝑥 |
|𝑥𝐿 𝑓 (𝑢, 𝑥)| = |𝑥|𝐿 || 𝑥 |
|
|𝑒 + 1|
= |𝑥|𝐿 𝑒−𝑢|𝑥|
⩽ 1[−1,0] (𝑥) + 𝑁𝑎,𝑏 1(−∞,1) (𝑥) |𝑥|−2 .
Both inequalities give dominating functions which are integrable; therefore, the integral ∫R 𝑥𝐿 𝑓 (𝑢, 𝑥) 𝑑𝑥
exists.
143
R.L. Schilling: Measures, Integrals & Martingales
To see 𝑚-fold differentiability, we use the Differentiability lemma (Theorem 12.5) 𝑚-times. Form-
ally, we have to use induction. Let us only make the induction step (the start is very similar!). For
this, observe that
𝑥𝑛 𝑒𝑢𝑥 𝑥𝑛+𝑚 𝑒𝑢𝑥
𝜕𝑢𝑚 (𝑥𝑛 𝑓 (𝑢, 𝑥)) = 𝜕𝑢𝑚 = 𝑥
𝑒 +1
𝑥 𝑒 +1
but, as we have seen in the first step with 𝐿 = 𝑛 + 𝑚, this is uniformly bounded by an integrable
function. Therefore, the Differentiability lemma applies and shows that
Problem 12.30 Solution: Because of the binomial formual we have (1 + 𝑥2 )𝑛 ⩾ 1 + 𝑛𝑥2 ; this yields,
in particular,
| 1 + 𝑛𝑥2 |
| |
| (1 + 𝑥2 )𝑛 | ⩽ 1.
| |
Since
1 + 𝑛𝑥2
lim = 0 ∀ 𝑥 ∈ (0, 1)
𝑛→∞ (1 + 𝑥2 )𝑛
(exponential growth is always stronger than polynomial growth!) we can use dominated conver-
gence and find
1
1 + 𝑛𝑥2
lim 𝑑𝑥 = 0.
𝑛→∞ ∫0 (1 + 𝑥2 )𝑛
■■
𝜋 | 𝑡 |
|𝑢(𝑡, 𝑥)| ⩽1(0,1) (𝑥) + || |1
| [1,∞) (𝑥)
2 | sinh 𝑥 |
𝜋 1 1
⩽ 1[0,1] (𝑥) + 1 (𝑥) ∈ 1 ((0, ∞)).
2 4 𝑥2 [1,∞)
This proves that the integral 𝑓 (𝑡) = ∫(0,∞) 𝑢(𝑡, 𝑥) 𝑑𝑥 exists. In order to check differentiability
of 𝑓 , we have to find (Theorem 12.5) a majorizing function for the derivative of the integrand.
Fix 𝑅 > 0 and let 𝑡 ∈ (𝑅−1 , 𝑅). By the chain rule
𝜕 1 1
𝑢(𝑡, 𝑥) = ( )2 sinh 𝑥
𝜕𝑡 𝑡
1+ sinh 𝑥
144
Solution Manual. Last update 28th October 2021
1
= .
𝑡2
sinh 𝑥
+ sinh 𝑥
Since 𝑥 → 1
𝑅−2 +sinh 𝑥
is continuous, there is a constant 𝐶1 > 0 such that
1
sup −2 + sinh 𝑥
⩽ 𝐶1 .
𝑥∈[0,1] 𝑅
Using 0 ⩽ sinh 𝑥 ⩽ 1 for 𝑥 ∈ [0, 1], we get
1 1
|𝜕𝑡 𝑢(𝑡, 𝑥)| ⩽ ⩽ ⩽ 𝐶1 ∀ 𝑥 ∈ [0, 1].
𝑅−2
+ sinh 𝑥 𝑅−2 + sinh 𝑥
sinh 𝑥
Similarly we get for 𝑥 > 1
1 1 2 1
|𝜕𝑡 𝑢(𝑡, 𝑥)| ⩾ =2 𝑥 = 𝑥 ∈ 1 ((1, ∞)).
sinh 𝑥 𝑒 −𝑒 −𝑥 𝑒 1 − 𝑒−2𝑥
⏟⏞⏟⏞⏟
⩽𝐶2 <∞
Therefore,
1
|𝜕𝑡 𝑢(𝑡, 𝑥)| ⩽ 𝐶1 1(0,1] (𝑥) + 2𝐶2
1 (𝑥) ∈ 1 ((0, ∞)).
𝑒𝑥 (1,∞)
Using the differentiability lemma, Theorem 12.5, we find that 𝑓 is differentiable on (𝑅−1 , 𝑅)
and that
1
𝑓 ′ (𝑡) = 𝑑𝑥 ∀ 𝑡 ∈ (𝑅−1 , 𝑅).
∫(0,∞) 𝑡2
sinh 𝑥
+ sinh 𝑥
Since 𝑅 > 0 is arbitrary, 𝑓 is differentiable on (0, ∞). That lim𝑡↓0 𝑓 ′ (𝑡) does not exist, follows
directly from the closed expresson for 𝑓 ′ in part (ii).
(ii) Note that 𝑓 (0) = 0. In order to find an expression for 𝑓 ′ , we perform the following substitu-
tion: 𝑢 = cosh 𝑥 and we get, observing that cosh2 𝑥 − sinh2 𝑥 = 1:
1 1
𝑓 ′ (𝑡) = √ √ 𝑑𝑢
∫(1,∞) 𝑡2
√ + 𝑢2 − 1 𝑢2 − 1
𝑢2 −1
1
= 𝑑𝑢.
∫(1,∞) 𝑡2 − 1 + 𝑢2
(Observe: 𝑥 → 𝑡2
1
is continuous, hence Riemann-integrable. Since we have estab-
+sinh 𝑥
lished in part (i) the existence of the Lebesgue integral, we can use Riemann integrals (b/o
sinh 𝑥
145
R.L. Schilling: Measures, Integrals & Martingales
√
• 𝑡 < 1: Then 𝐶 ∶= 1 − 𝑡2 makes sense and we get
𝑢2 + 𝑡2 − 1 = 𝑢2 − 𝐶 2 = (𝑢 + 𝐶)(𝑢 − 𝑐).
■■
𝜙(𝑚)
𝑋
(0+) = (−1)𝑚 𝑋 𝑚 𝑑P.
∫
Since the left-hand side has a finite P-integral, so has the right, i.e.
∞
(∑ (−1)𝑚+1+𝑗 𝑡𝑗 )
𝑋 𝑚+1+𝑗 𝑑P converges
∫
𝑗=0
(𝑚 + 1 + 𝑗)!
146
Solution Manual. Last update 28th October 2021
as 𝑡 → 0.
| −𝑢 𝑚−1
∑ (−𝑢)𝑘 | 𝑢𝑚
|𝑒 −
|
|⩽
| 𝑚! ∀ 𝑢 ⩾ 0. (*)
| 𝑘=0
𝑘! |
|𝑒−𝑢 − 1| ⩽ 𝑢 ∀𝑢 ⩾ 0
the start of the induction 𝑚 = 1 is clear. For the induction step 𝑚 → 𝑚 + 1 we note that
(−𝑢)𝑘 || || 𝑢 ( −𝑦 ∑ (−𝑦)𝑘 ) ||
𝑚 𝑚−1
| −𝑢 ∑
|𝑒 − = 𝑒 − 𝑑𝑦|
| 𝑘! || || ∫0 𝑘!
| 𝑘=0 𝑘=0 |
𝑚−1
𝑢| ∑ (−𝑦)𝑘 |
⩽ |𝑒−𝑦 − | 𝑑𝑦
∫0 | 𝑘! |
| 𝑘=0 |
(*) 𝑢
𝑦𝑚
⩽ 𝑑𝑦
∫0 𝑚!
𝑢𝑚+1
= ,
(𝑚 + 1)!
and the claim follows.
|𝑡|𝑚 ∫ 𝑋 𝑚 𝑑P
lim =0
𝑚→∞ 𝑚!
which, when combined with (iii), proves that
𝑚−1
∑ ∫ 𝑋 𝑘 𝑑P
𝜙𝑋 (𝑡) = lim (−1)𝑘 𝑡𝑘 .
𝑚→∞
𝑘=0
𝑘!
■■
(i) Wrong, 𝑢 is NOT continuous on the irrational numbers. To see this, just take a sequence of
rationals 𝑞𝑗 ∈ Q ∩ [0, 1] approximating 𝑝 ∈ [0, 1] ⧵ Q. Then
147
R.L. Schilling: Measures, Integrals & Martingales
(ii) True. Mind that 𝑣 is not continuous at 0, but {𝑛−1 , 𝑛 ∈ N} ∪ {0} is still countable.
(iii) True. The points where 𝑢 and 𝑣 are not 0 (that is: where they are 1) are countable sets, hence
measurable and also Lebesgue null sets. This shows that 𝑢, 𝑣 are measurable and almost
everywhere 0, hence ∫ 𝑢 𝑑𝜆 = 0 = ∫ 𝑣 𝑑𝜆.
(iv) True. Since Q ∩ [0, 1] as well as [0, 1] ⧵ Q are dense subsets of [0, 1], ALL lower resp. upper
Darboux sums are always
(for any finite partition 𝜋 of [0, 1]). Thus upper and lower integrals of 𝑢 have the value 0 resp.
1 and it follows that 𝑢 cannot be Riemann integrable.
■■
Problem 12.34 Solution: Note that every function which has finitely many discontinuities is Riemann
integrable. Thus, if {𝑞𝑗 }𝑗∈N is an enumeration of Q, the functions 𝑢𝑗 (𝑥) ∶= 1{𝑞1 ,𝑞2 ,…,𝑞𝑗 } (𝑥) are
Riemann integrable (with Riemann integral 0) while their increasing limit 𝑢∞ = 1Q is not Riemann
integrable.
■■
Problem 12.35 Solution: Of course we have to assume that 𝑢 is Borel measurable! By assumption
we know that 𝑢𝑗 ∶= 𝑢1[0,𝑗] is (properly) Riemann integrable, hence Lebesgue integrable and
𝑗 ∞
𝑢 𝑑𝜆 = 𝑢𝑗 𝑑𝜆 = (R) 𝑢(𝑥) 𝑑𝑥 ←←←←←←←←←→
← 𝑢(𝑥) 𝑑𝑥.
∫[0,𝑗] ∫[0,𝑗] ∫0 𝑗→∞ ∫0
The last limit exists because of improper Riemann integrability. Moreover, this limit is an increas-
ing limit, i.e. a ‘sup’. Since 0 ⩽ 𝑢𝑗 ↑ 𝑢 we can invoke Beppo Levi’s theorem and get
∞
𝑢 𝑑𝜆 = sup 𝑢𝑗 𝑑𝜆 = 𝑢(𝑥) 𝑑𝑥 < ∞
∫ 𝑗 ∫ ∫0
■■
√
Problem 12.36 Solution: Observe that 𝑥2 = 𝑘𝜋 ⇐⇒ 𝑥 = 𝑘𝜋, 𝑥 ⩾ 0, 𝑘 ∈ N0 . Thus, Since
sin 𝑥2 is continuous, it is on every bounded interval Riemann integrable. By a change of variables,
𝑦 = 𝑥2 , we get
√
𝑏 𝑏 𝑏
𝑑𝑦 | sin 𝑦|
| sin(𝑥2 )| 𝑑𝑥 = | sin 𝑦| √ = √ 𝑑𝑦
∫√𝑎 ∫𝑎 2 𝑦 ∫ 𝑎 2 𝑦
√
which means that for 𝑎 = 𝑎𝑘 = 𝑘𝜋 and 𝑏 = 𝑏𝑘 = (𝑘 + 1)𝜋 = 𝑎𝑘+1 the values ∫√
𝑎𝑘+1
| sin(𝑥2 )| 𝑑𝑥
𝑎𝑘
[√ √ ]
are a decreasing sequence with limit 0. Since on 𝑎𝑘 , 𝑎𝑘+1 the function sin 𝑥2 has only one
148
Solution Manual. Last update 28th October 2021
sign (and alternates its sign from interval to interval), we can use Leibniz’ convergence criterion
to see that the series
√
∑ 𝑎𝑘+1
sin(𝑥2 ) 𝑑𝑥 (*)
∫√𝑎𝑘
𝑘
The function cos 𝑥2 can be treated similarly. Alternatively, we remark that sin 𝑥2 = cos(𝑥2 − 𝜋∕2).
The functions are not Lebesgue integrable. Either we show that the series (*) does not converge
absolutely, or we argue as follows:
sin 𝑥2 = cos(𝑥2 −𝜋∕2) shows that ∫ | sin 𝑥2 | 𝑑𝑥 and ∫ | cos 𝑥2 | 𝑑𝑥 either both converge or diverge.
If they would converge (this is equivalent to Lebesgue integrability...) we would find because of
sin2 + cos2 ≡ 1 and | sin |, | cos | ⩽ 1,
∞ ∞[ ]
∞= 1 𝑑𝑥 = (sin 𝑥2 )2 + (cos 𝑥2 )2 𝑑𝑥
∫0 ∫0
∞ ∞
= (sin 𝑥2 )2 𝑑𝑥 + (cos 𝑥2 )2 𝑑𝑥
∫0 ∫0
∞ ∞
⩽ | sin 𝑥2 | 𝑑𝑥 + | cos 𝑥2 | 𝑑𝑥 < ∞,
∫0 ∫0
which is a contradiction.
■■
Problem 12.37 Solution: Let 𝑟 < 𝑠 and, without loss of generality, 𝑎 ⩽ 𝑏. A change of variables
yields
𝑠 𝑠 𝑠
𝑓 (𝑏𝑥) − 𝑓 (𝑎𝑥) 𝑓 (𝑏𝑥) 𝑓 (𝑎𝑥)
𝑑𝑥 = 𝑑𝑥 − 𝑑𝑥
∫𝑟 𝑥 ∫𝑟 𝑥 ∫𝑟 𝑥
𝑏𝑠 𝑎𝑠
𝑓 (𝑦) 𝑓 (𝑦)
= 𝑑𝑦 − 𝑑𝑦
∫𝑏𝑟 𝑦 ∫ 𝑎𝑟 𝑦
𝑏𝑠 𝑏𝑟
𝑓 (𝑦) 𝑓 (𝑦)
= 𝑑𝑦 − 𝑑𝑦
∫𝑎𝑠 𝑦 ∫ 𝑎𝑟 𝑦
Using the mean value theorem for integrals, I.12, we get
𝑠 𝑏𝑠 𝑏𝑟
𝑓 (𝑏𝑥) − 𝑓 (𝑎𝑥) 1 1
𝑑𝑥 = 𝑓 (𝜉𝑠 ) 𝑑𝑦 − 𝑓 (𝜉𝑟 ) 𝑑𝑦
∫𝑟 𝑥 ∫ 𝑎𝑠 𝑦 ∫ 𝑎𝑟 𝑦
= 𝑓 (𝜉𝑠 ) ln 𝑎𝑏 − 𝑓 (𝜉𝑟 ) ln 𝑎𝑏 .
𝑠 [ ]
𝑓 (𝑏𝑥) − 𝑓 (𝑎𝑥) 𝑠→∞
𝑑𝑥 = 𝑓 (𝜉𝑠 ) − 𝑓 (𝜉𝑟 ) ln 𝑎𝑏 ←←←←←←←←←←←→
← (𝑀 − 𝑚) ln 𝑎𝑏 .
∫𝑟 𝑥 𝑟→0
■■
149
13 The function spaces 𝑝.
Solutions to Problems 13.1–13.26
(ii) If 𝑢 ∈ 𝑝 we know that 𝑢 is measurable and ‖𝑢‖𝑝 < ∞. The inequality in (i) then shows that
151
R.L. Schilling: Measures, Integrals & Martingales
■■
Problem 13.2 Solution: This is going to be a bit messy and rather than showing the ‘streamlined’
solution we indicate how one could find out the numbers oneself. Now let 𝜆 be some number in
(0, 1) and let 𝛼, 𝛽 be conjugate indices: 1
𝛼
+ 1
𝛽
= 1 where 𝛼, 𝛽 ∈ (1, ∞). Then by the Hölder
inequality
■■
152
Solution Manual. Last update 28th October 2021
and, therefore,
| |
|‖𝑢‖𝑝 − ‖𝑣‖𝑝 | = max{‖𝑢‖𝑝 − ‖𝑣‖𝑝 , ‖𝑣‖𝑝 − ‖𝑢‖𝑝 } ⩽ ‖𝑢 − 𝑣‖𝑝 .
| |
■■
(a) Every map 𝑢 ∶ (Ω, {∅, Ω}) → (R, {∅, R}) is measurable. Indeed: 𝑢 is measurable if,
and only if, 𝑢−1 (𝐴) ∈ {∅, Ω} for all 𝐴 ∈ 𝒜 = {∅, R}. Since
(b) Every measurable map 𝑢 ∶ (Ω, {∅, Ω}) → (R, ℬ(R)) is constant. Indeed: Suppose,
𝑢 is not constant, i.e. there are 𝜔1 , 𝜔2 ∈ Ω and 𝑥, 𝑦 ∈ R, 𝑥 ≠ 𝑦, such that 𝑢(𝜔1 ) = 𝑥,
𝑢(𝜔2 ) = 𝑦. Then 𝑢−1 ({𝑥}) ∉ {∅, Ω} as 𝜔1 ∈ 𝑢−1 ({𝑥}) (and so 𝑢−1 ({𝑥}) ≠ ∅) and
𝜔2 ∉ 𝑢−1 ({𝑥}) (and so 𝑢−1 ({𝑥}) ≠ Ω).
(c) Every measurable map 𝑢 ∶ (Ω, {∅, Ω}) → (R, 𝒫 (R)) is clearly {∅, Ω}∕ℬ(R)-measurable.
From (b) we know that such functions are constant. On the other hand, constant maps
are measurable for any 𝜎-algebra. Therefore, every {∅, Ω}∕𝒫 (R)-measurable map is
constant.
⎧
⎪Ω, 𝑐1 , 𝑐2 ∈ 𝐴,
⎪
⎪𝐵, 𝑐1 ∈ 𝐴, 𝑐2 ∉ 𝐴,
𝑢−1 (𝐴) = ⎨
⎪𝐵 𝑐 , 𝑐1 ∉ 𝐴, 𝑐2 ∈ 𝐴,
⎪
⎪∅, 𝑐1 , 𝑐2 ∉ 𝐴
⎩
153
R.L. Schilling: Measures, Integrals & Martingales
By definition,
{ }
(Ω, 𝜎(𝐵), 𝜇) = 𝑢 ∶ (Ω, 𝜎(𝐵)) → (R, ℬ(R)) measurable ∶
𝑝
|𝑢| 𝑑𝜇 < ∞
𝑝
.
∫
We have already shown that the 𝜎(𝐵)-measurable maps are given by (⋆). Because of the
linearity of the integral we see that
• 𝑐1 = 0 or 𝜇(𝐵) < ∞
• 𝑐2 = 0 or 𝜇(𝐵 𝑐 ) < ∞.
In particular, every map of the form (⋆) is in 𝑝 (Ω, 𝜎(𝐵), 𝜇) if 𝜇 is a finite measure.
■■
Now use the induction hypothesis which allows us to apply the generalized Hölder inequality for
∑
𝑁 (!) factors 𝜆𝑗 ∶= 𝑝𝑗 ∕𝑝, and thus 𝑁
𝑗=1 𝜆𝑗 = 𝑝∕𝑝 = 1, to the first factor to get
−1
( )1∕𝑝
|𝑢1 ⋅ 𝑢2 ⋅ … ⋅ 𝑢𝑁 ⋅ 𝑤| 𝑑𝜇 = |𝑢1 | ⋅ |𝑢2 | ⋅ … ⋅ |𝑢𝑁 | 𝑑𝜇
𝑝 𝑝 𝑝
‖𝑢‖𝑞
∫ ∫
⩽ ‖𝑢‖𝑝1 ⋅ ‖𝑢‖𝑝2 ⋅ … ⋅ ‖𝑢‖𝑝𝑁 ‖𝑢‖𝑞 .
■■
154
Solution Manual. Last update 28th October 2021
Problem 13.6 Solution: Draw a picture similar to the one used in the proof of Lemma 13.1 (note
that the increasing function need not be convex or concave....). Without loss of generality we can
assume that 𝐴, 𝐵 > 0 are such that 𝜙(𝐴) ⩾ 𝐵 which is equivalent to 𝐴 ⩾ 𝜓(𝐵) since 𝜙 and 𝜓 are
inverses. Thus,
𝐵 𝜓(𝐵) 𝐴
𝐴𝐵 = 𝜓(𝜂) 𝑑𝜂 + 𝜙(𝜉) 𝑑𝜉 + 𝐵 𝑑𝜉.
∫0 ∫0 ∫𝜓(𝐵)
■■
Problem 13.7 Solution: Let us show first of all that 𝑝 -lim𝑘→∞ 𝑢𝑘 = 𝑢. This follows immediately
∑
from lim𝑘→∞ ‖𝑢 − 𝑢𝑘 ‖𝑝 = 0 since the series ∞𝑘=1 ‖𝑢 − 𝑢𝑘 ‖𝑝 converges.
Now we want to show that 𝑢 is the a.e. limit of the original sequence. For this we mimic the trick
from the Riesz–Fischer theorem 13.7 and show that the series
∞ 𝐾
∑ ∑
(𝑢𝑗+1 − 𝑢𝑗 ) = lim (𝑢𝑗+1 − 𝑢𝑗 ) = lim 𝑢𝐾
𝐾→∞ 𝐾→∞
𝑗=0 𝑗=0
(again we agree on 𝑢0 ∶= 0 for notational convenience) makes sense. So let us employ Lemma
13.6 used in the proof of the Riesz–Fischer theorem to get
‖∑ ‖ ‖∑ ‖
‖∞ ‖ ‖∞ ‖
‖ (𝑢𝑗+1 − 𝑢𝑗 )‖ ⩽ ‖ |𝑢𝑗+1 − 𝑢𝑗 |‖
‖ ‖ ‖ ‖
‖ 𝑗=0 ‖ ‖ ‖
‖ ‖𝑝 ‖ 𝑗=0 ‖𝑝
∞
∑
⩽ ‖𝑢𝑗+1 − 𝑢𝑗 ‖𝑝
𝑗=0
∞
∑( )
⩽ ‖𝑢𝑗+1 − 𝑢‖𝑝 + ‖𝑢 − 𝑢𝑗 ‖𝑝
𝑗=0
<∞
155
R.L. Schilling: Measures, Integrals & Martingales
∑
where we use Minkowski’s inequality, the function 𝑢 from above and the fact that ∞
𝑗=1 ‖𝑢𝑗 −𝑢‖𝑝 <
∑∞
∞ along with ‖𝑢1 ‖𝑝 < ∞. This shows that lim𝐾→∞ 𝑢𝐾 (𝑥) = 𝑗=0 (𝑢𝑗+1 (𝑥) − 𝑢𝑗 (𝑥)) exists almost
everywhere.
We still have to show that lim𝐾→∞ 𝑢𝐾 (𝑥) = 𝑢(𝑥). For this we remark that a subsequence has
necessarily the same limit as the original sequence—whenever both have limits, of course. But
then,
∞
∑
𝑢(𝑥) = lim 𝑢𝑘(𝑗) (𝑥) = lim 𝑢𝑘 (𝑥) = (𝑢𝑗+1 (𝑥) − 𝑢𝑗 (𝑥))
𝑗→∞ 𝑘→∞
𝑗=0
■■
1
|𝑢𝑛(𝑗) |𝑝 𝑑𝜆 = 𝑛(𝑗)𝑝 = 𝑛(𝑗)𝑝−1 ←←←←←←←←←→
← 𝑐
∫ 𝑛(𝑗) 𝑗→∞
with 𝑐 = 1 in case 𝑝 = 1 and 𝑐 = ∞ if 𝑝 > 1. This shows that the 𝑝 -limit of this subsequence—let
us call it 𝑤 if it exists at all—cannot be (not even a.e.) 𝑢 = 0.
lim ̃
𝑢𝑘(𝑗) (𝑥) = 𝑤(𝑥).
𝑗
Since the full sequence lim𝑛 𝑢𝑛 (𝑥) = 𝑢(𝑥) = 0 has a limit, this shows that the sub-sub-sequence
limit 𝑤(𝑥) = 0 almost everywhere—a contradiction. Thus, 𝑤 does not exist in the first place.
■■
Problem 13.9 Solution: Using Minkowski’s and Hölder’s inequalities we find for all 𝜖 > 0
for all 𝑛 ⩾ 𝑁𝜖 . We use here that the sequence (‖𝑢𝑘 ‖𝑝 )𝑘∈N is bounded. Indeed, by Minkowski’s
inequality
■■
156
Solution Manual. Last update 28th October 2021
Case 1: 𝑢𝑛 → 𝑢 in 2 . This means that (𝑢𝑛 )𝑛∈N is an 2 Cauchy sequence, i.e. that lim𝑚,𝑛→∞ ‖𝑢𝑛 −
𝑢𝑚 ‖22 = 0. On the other hand, we get from the lower triangle inequality for norms
so that also lim𝑛→∞ ‖𝑢𝑛 ‖22 = lim𝑚→∞ ‖𝑢𝑚 ‖22 = ‖𝑢‖22 . Using (*) we find
= 2‖𝑢‖22 .
Case 2: Assume that lim𝑛,𝑚→∞ ∫ 𝑢𝑛 𝑢𝑚 𝑑𝜇 = 𝑐 for some number 𝑐 ∈ R. By the very definition of
this double limit, i.e.
| |
∀𝜖 > 0 ∃ 𝑁𝜖 ∈ N ∶ | 𝑢 𝑢 𝑑𝜇 − 𝑐 | < 𝜖 ∀𝑛, 𝑚 ⩾ 𝑁𝜖 ,
|∫ 𝑛 𝑚 |
| |
we see that lim𝑛→∞ ∫ 𝑢𝑛 𝑢𝑛 𝑑𝜇 = 𝑐 = lim𝑚→∞ ∫ 𝑢𝑚 𝑢𝑚 𝑑𝜇 hold (with the same 𝑐!). Therefore, again
by (*), we get
i.e. (𝑢𝑛 )𝑛∈N is a Cauchy sequence in 2 and has, by the completeness of this space, a limit.
■■
Problem 13.11 Solution: Use the exponential series to conclude from the positivity of ℎ and 𝑢(𝑥)
that
∞
∑ ℎ 𝑗 𝑢𝑗 ℎ𝑁 𝑁
exp(ℎ𝑢) = ⩾ 𝑢 .
𝑗=0
𝑗! 𝑁!
157
R.L. Schilling: Measures, Integrals & Martingales
(i) We have to show that |𝑢𝑛 (𝑥)|𝑝 ∶= 𝑛𝑝𝛼 (𝑥 + 𝑛)−𝑝𝛽 has finite integral—measurability is clear
since 𝑢𝑛 is continuous. Since 𝑛𝑝𝛼 is a constant, we have only to show that (𝑥 + 𝑛)−𝑝𝛽 is in 1 .
Set 𝛾 ∶= 𝑝𝛽 > 1. Then we get from a Beppo Levi and a domination argument
Now using that Riemann=Lebesgue on intervals where the Riemann integral exists, we get
𝑘
lim 𝑥−𝛾 𝜆(𝑑𝑥) = lim 𝑥−𝛾 𝑑𝑥
𝑘→∞ ∫(1,𝑘) 𝑘→∞ ∫1
[ ]𝑘
= lim (1 − 𝛾)−1 𝑥1−𝛾 1
𝑘→∞
( )
= (1 − 𝛾)−1 lim 𝑘1−𝛾 − 1
𝑘→∞
−1
= (𝛾 − 1) < ∞
(ii) We have to show that |𝑣𝑛 (𝑥)|𝑞 ∶= 𝑛𝑞𝛾 𝑒−𝑞𝑛𝑥 is in 1 —again measurability is inferred from
continuity. Since 𝑛𝑞𝛾 is a constant, it is enough to show that 𝑒−𝑞𝑛𝑥 is integrable. Set 𝛿 = 𝑞𝑛.
Since
and since 𝑒−𝛿𝑥 is continuous on [0, ∞), we conclude that there are constants 𝐶, 𝐶(𝛿) such
that
{ }
−𝛿𝑥 𝐶
𝑒 ⩽ min 1,
(𝛿𝑥)2
{ }
1
⩽ 𝐶(𝛿) min 1, 2
𝑥
( )
1
= 𝐶(𝛿) 1(0,1) (𝑥) + 1[1,∞) 2
𝑥
but the latter is an integrable function on (0, ∞).
■■
Problem 13.13 Solution: Without loss of generality we may assume that 𝛼 ⩽ 𝛽. We distinguish
between the case 𝑥 ∈ (0, 1) and 𝑥 ∈ [1, ∞). If 𝑥 ⩽ 1, then
1 1 1 1∕2
⩾ 𝛼 ⩾ 𝛼 = 𝛼 ∀ 𝑥 ⩽ 1;
𝑥𝛼 𝑥 + 𝑥𝛽 𝑥 + 𝑥𝛼 𝑥 + 𝑥𝛼
this shows that (𝑥𝛼 + 𝑥𝛽 )−1 is in 1 ((0, 1), 𝑑𝑥) if, and only if, 𝛼 < 1.
158
Solution Manual. Last update 28th October 2021
Similarly, if 𝑥 ⩾ 1, then
1 1 1 1∕2
⩾ 𝛼 ⩾ 𝛽 = 𝛽 ∀𝑥 ⩾ 1
𝑥𝛽 𝑥 +𝑥 𝛽 𝑥 +𝑥 𝛽 𝑥 + 𝑥𝛽
this shows that (𝑥𝛼 + 𝑥𝛽 )−1 is in 1 ((1, ∞), 𝑑𝑥) if, and only if, 𝛽 > 1.
Thus, (𝑥𝛼 + 𝑥𝛽 )−1 is in 1 (R, 𝑑𝑥) if, and only if, both 𝛼 < 1 and 𝛽 > 1.
■■
and it is clear that this is a norm for 𝑝 ⩾ 1 and, in view of Problem 13.19 it is not a norm for
𝑝 < 1 since the triangle (Minkowski) inequality fails. (This could also be shown by a direct
counterexample.
■■
Problem 13.15 Solution: Without loss of generality we can restrict ourselves to positive functions—
else we would consider positive and negative parts. Separability can obviously considered separ-
ately!
Assume that 1+ is separable and choose 𝑢 ∈ 𝑝+ . Then 𝑢𝑝 ∈ 1 and, because of separability, there
is a sequence (𝑓𝑛 )𝑛 ⊂ 𝒟1 ⊂ 1 such that
in 1 in 1
← 𝑢𝑝 ⇐⇒ 𝑢𝑝𝑛 ←←←←←←←←←→
𝑓𝑛 ←←←←←←←←←→ ← 𝑢𝑝
𝑛→∞ 𝑛→∞
if we set 𝑢𝑛 ∶= 𝑓𝑛 ∈ 𝑝 . In particular, 𝑢𝑛(𝑘) (𝑥) → 𝑢(𝑥) almost everywhere for some subsequence
1∕𝑝
in 𝑝
𝑝 ∋ 𝑢𝑛(𝑘) ←←←←←←←←←→
← 𝑢.
𝑘→∞
Obviously the separating set 𝒟𝑝 is essentially the same as 𝒟1 , and we are done.
The converse is similar (note that we did not make any assumptions on 𝑝 ⩾ 1 or 𝑝 < 1—this is
immaterial in the above argument).
■■
Problem 13.16 Solution: We have seen in the lecture that, whenever lim𝑛→∞ ‖𝑢 − 𝑢𝑛 ‖𝑝 = 0, there
is a subsequence 𝑢𝑛(𝑘) such that lim𝑘→∞ 𝑢𝑛(𝑘) (𝑥) = 𝑢(𝑥) almost everywhere. Since, by assumption,
lim𝑗→∞ 𝑢𝑗 (𝑥) = 𝑤(𝑥) a.e., we have also that lim𝑗→∞ 𝑢𝑛(𝑗) (𝑥) = 𝑤(𝑥) a.e., hence 𝑢(𝑥) = 𝑤(𝑥)
almost everywhere.
■■
159
R.L. Schilling: Measures, Integrals & Martingales
Problem 13.17 Solution: We remark that 𝑦 → log 𝑦 is concave. Therefore, we can use Jensen’s
inequality for concave functions to get for the probability measure 𝜇∕𝜇(𝑋) = 𝜇(𝑋)−1 1𝑋 𝜇
( )
𝑑𝜇 𝑑𝜇
(log 𝑢) ⩽ log 𝑢
∫ 𝜇(𝑋) ∫ 𝜇(𝑋)
( )
∫ 𝑢 𝑑𝜇
= log
𝜇(𝑋)
( )
1
= log ,
𝜇(𝑋)
and the claim follows.
■■
Assume now that ∫(0,1) 𝑢(𝑥) 𝑑𝑥 = 1. Substituting in the above inequality 𝑦 = 𝑢(𝑥) and integrating
over (0, 1) yields
Now assume that 𝛼 = ∫(0,1) 𝑢(𝑥) 𝑑𝑥. Then ∫(0,1) 𝑢(𝑥)∕𝛼 𝑑𝑥 = 1 and the above inequality gives
which is equivalent to
160
Solution Manual. Last update 28th October 2021
1
= 𝑢(𝑥) log 𝑢(𝑥) 𝑑𝑥 − log 𝛼.
𝛼 ∫(0,1)
The claim now follows by adding log 𝛼 on both sides and then multiplying by 𝛼 = ∫(0,1) 𝑢(𝑥) 𝑑𝑥.
■■
(i) Let 𝑝 ∈ (0, 1) and pick the conjugate index 𝑞 ∶= 𝑝∕(𝑝−1) < 0. Moreover, 𝑠 ∶= 1∕𝑝 ∈ (1, ∞)
and the conjugate index 𝑡, 1
𝑠
+ 1
𝑡
= 1, is given by
1
𝑠 𝑝 1
𝑡= = = ∈ (1, ∞).
𝑠−1 1
−1 1−𝑝
𝑝
𝑤𝑝
𝑢𝑝 𝑑𝜇 = 𝑢𝑝 𝑑𝜇
∫ ∫ 𝑤𝑝
( )1∕𝑠 ( )1∕𝑡
( 𝑝 𝑝 )𝑠
⩽ 𝑢 𝑤 𝑑𝜇 𝑤−𝑝𝑡 𝑑𝜇
∫ ∫
( )𝑝 ( )1−𝑝
= 𝑢𝑤 𝑑𝜇 𝑤 𝑝∕(𝑝−1)
𝑑𝜇 .
∫ ∫
(ii) This ‘reversed’ Minkowski inequality follows from the ‘reversed’ Hölder inequality in exactly
the same way as Minkowski’s inequality follows from Hölder’s inequality, cf. Corollary 13.4.
To wit:
(𝑢 + 𝑣)𝑝 𝑑𝜇 = (𝑢 + 𝑣) ⋅ (𝑢 + 𝑣)𝑝−1 𝑑𝜇
∫ ∫
= 𝑢 ⋅ (𝑢 + 𝑣)𝑝−1 𝑑𝜇 + 𝑣 ⋅ (𝑢 + 𝑣)𝑝−1 𝑑𝜇
∫ ∫
(i)
‖ ‖ ‖ ‖
⩾ ‖𝑢‖𝑝 ⋅ ‖(𝑢 + 𝑣)𝑝−1 ‖ + ‖𝑣‖𝑝 ⋅ ‖(𝑢 + 𝑣)𝑝−1 ‖ .
‖ ‖𝑞 ‖ ‖𝑞
■■
161
R.L. Schilling: Measures, Integrals & Martingales
(i) We have
𝑀𝑛 = |𝑢|𝑛 𝑑𝜇
∫
𝑛+1 𝑛−1
= |𝑢| 2 |𝑢| 2 𝑑𝜇
∫
( )1∕2 ( )1∕2
𝑛+1 𝑛−1
⩽ |𝑢| 𝑑𝜇 |𝑢| 𝑑𝜇
∫ ∫
√
= 𝑀𝑛+1 𝑀𝑛−1 .
i.e.
162
Solution Manual. Last update 28th October 2021
■■
Problem 13.21 Solution: The hint says it all.... Maybe, you have a look at the specimen solution of
Problem 13.20, too.
Case 1: ‖𝑢‖𝐿∞ < ∞. For 𝐴𝛿 ∶= {𝑢 ⩾ ‖𝑢‖∞ − 𝛿}, 𝛿 > 0, we gave 𝜇(𝐴𝛿 ) > 0 and
( )1 1
𝑝
‖𝑢‖𝑝 ⩾ (‖𝑢‖∞ − 𝛿) 𝑑𝜇 𝑝
= (‖𝑢‖∞ − 𝛿)𝜇(𝐴𝛿 ) 𝑝 .
∫𝐴𝛿
Therefore,
( 1
)
lim inf ‖𝑢‖𝑝 ⩾ lim inf (‖𝑢‖∞ − 𝛿)𝜇(𝐴𝛿 ) 𝑝 = ‖𝑢‖∞ − 𝛿.
𝑝→∞ 𝑛→∞
Since 𝛿 > 0 is arbitrary, this shows that lim inf 𝑝→∞ ‖𝑢‖𝑝 ⩾ ‖𝑢‖∞ .
is trivially true. The converse inequality follows like this: Define 𝐴𝑅 ∶= {𝑢 ⩾ 𝑅}, 𝑅 > 0. We
have 𝜇(𝐴𝑟 ) > 0 (otherwise ‖𝑢‖𝐿∞ < ∞!) and, as in the first part of the proof, we find
( )1 1
𝑝
‖𝑢‖𝑝 ⩾ 𝑝
𝑅 𝑑𝜇 = 𝑅𝜇(𝐴𝑅 ) 𝑝 .
∫𝐴 𝑅
Thus, lim inf 𝑝→∞ ‖𝑢‖𝑝 ⩾ 𝑅 and since 𝑅 > 0 is arbitrary, the claim follows:
■■
• If 𝑟 ⩽ 𝑠 ⩽ 𝑞, then ‖𝑢‖𝑟 ⩽ ‖𝑢‖𝑠 . This follows from Jensen’s inequality (Theorem 13.13) and
the fact that 𝑉 (𝑥) ∶= 𝑥𝑠∕𝑟 , 𝑥 ∈ R, is convex (cf. also Problem 13.1). In particular, ‖𝑢‖𝑟 < ∞
for all 𝑟 ∈ (0, 𝑞).
163
R.L. Schilling: Measures, Integrals & Martingales
• We have
log |𝑢| 𝑑𝜇 ⩽ log ‖𝑢‖𝑝 ∀ 𝑝 ∈ (0, 𝑞). (⋆)
∫
This follows again from Jensen’s inequality applied to the convex function 𝑉 (𝑥) ∶= − log 𝑥:
( )
− log 𝑝
|𝑢| 𝑑𝜇 ⩽ − log(|𝑢|𝑝 ) 𝑑𝜇 − 𝑝 log |𝑢| 𝑑𝜇;
∫ ∫ ∫
therefore,
( )
1
log ‖𝑢‖𝑝 = log 𝑝
|𝑢| 𝑑𝜇 ⩾ log |𝑢| 𝑑𝜇.
𝑝 ∫ ∫
Because of (⋆) it is enough to show that lim𝑝→0 ‖𝑢‖𝑝 ⩽ exp(∫ ln |𝑢| 𝑑𝜇). (Note: by the monoton-
icity of ‖𝑢‖𝑝 as 𝑝 ↓ 0 we know that the limit lim𝑝→0 ‖𝑢‖𝑝 exists.) Note that
𝑎𝑝 − 1
log 𝑎 = inf , 𝑎 > 0. (⋆⋆)
𝑝>0 𝑝
Use l’Hospital’s rule to show that lim𝑝→0 𝑎 𝑝−1 = log 𝑎.) From monotone convergence (mc) we get
𝑝
mc |𝑢|𝑝 − 1
log |𝑢| 𝑑𝜇 = inf 𝑑𝜇
∫ 𝑝>0 ∫ 𝑝
∫ |𝑢|𝑝 𝑑𝜇 − 1
= inf
𝑝>0 𝑝
𝑝
‖𝑢‖𝑝 − 1 (⋆⋆)
= inf = log ‖𝑢‖𝑝
𝑝>0 𝑝
for all 𝑝 > 0. Letting 𝑝 → 0 finishes the proof.
■■
Problem 13.23 Solution: Without loss of generality we may assume that 𝑓 ⩾ 0. We use the following
standard representation of 𝑓 , see (8.7):
𝑁
∑
𝑓= 𝜙𝑗 1𝐴𝑗
𝑗=0
with 0 = 𝜙0 < 𝜙1 < … < 𝜙𝑁 < ∞ and mutually disjoint sets 𝐴𝑗 . Clearly, {𝑓 ≠ 0} =
𝐴1 ⊍ ⋯ ⊍ 𝐴 𝑁 .
164
Solution Manual. Last update 28th October 2021
Since this integrability criterion does not depend on 𝑝 ⩾ 1, it is clear that + ∩𝑝 (𝜇) = + ∩1 (𝜇),
and the rest follows since = + − + .
■■
Problem 13.24 Solution: (i) ⇐⇒ (ii) and (iii) ⇐⇒ (iv), since Λ is concave if, and only if, 𝑉 = −Λ
is convex. Moreover, (iii) generalizes (i) and (iv) gives (ii). It is, therefore, enough to verify (iii).
Now go to the sup over all affine-linear 𝓁 below 𝑉 and the claim follows.
■■
(i) Note that Λ(𝑥) = 𝑥1∕𝑞 is concave—e.g. differentiate twice and show that it is negative—and
using Jensen’s inequality for positive 𝑓 , 𝑔 ⩾ 0 yields
165
R.L. Schilling: Measures, Integrals & Martingales
Adding on both sides ∫{𝑓 =0} (𝑓 + 𝑔)𝑝 𝑑𝜇 = ∫{𝑓 =0} 𝑔 𝑝 𝑑𝜇 yields, because of the elementary
inequality 𝐴𝑝 + 𝐵 𝑝 ⩽ (𝐴 + 𝐵)𝑝 , 𝐴, 𝐵 ⩾ 0, 𝑝 ⩾ 1,
(𝑓 + 𝑔)𝑝 𝑑𝜇
∫
[( )1∕𝑝 ( )1∕𝑝 ]𝑝 [ ]𝑝∕𝑝
⩽ 𝑝
𝑔 𝑑𝜇 + 𝑓 𝑑𝜇𝑝
+ 𝑝
𝑔 𝑑𝜇
∫{𝑓 ≠0} ∫{𝑓 ≠0} ∫{𝑓 =0}
[( )1∕𝑝 ( )1∕𝑝 ]𝑝
⩽ 𝑔 𝑝 𝑑𝜇 + 𝑓 𝑝 𝑑𝜇 .
∫ ∫
■■
Since 𝜇(𝑋) < ∞, this shows that both sides of the asserted integral inequality are finite.
Without loss of generality we may assume that 𝑎 > 0, otherwise we would consider −𝑓 instead of
𝑓.
Observe that
|𝑓 |𝑝 𝑑𝜇 ⩽ (2𝑎)𝑝−1 |𝑓 | 𝑑𝜇
∫{0<𝑓 <2𝑎} ∫{0<𝑓 <2𝑎}
⩽ (2𝑎)𝑝−1 |𝑓 | 𝑑𝜇
∫{𝑓 >0}
= (2𝑎)𝑝−1 |𝑓 | 𝑑𝜇.
∫{𝑓 <0}
In the last line we use the fact that
∫ 𝑓 𝑑𝜇=0
|𝑓 | 𝑑𝜇 = 𝑓 + 𝑑𝜇 = 𝑓 − 𝑑𝜇 = |𝑓 | 𝑑𝜇.
∫{𝑓 >0} ∫ ∫ ∫{𝑓 <0}
Thus,
|𝑓 |𝑝 𝑑𝜇 ⩽ (2𝑎)𝑝−1 |𝑓 | 𝑑𝜇
∫{0<𝑓 <2𝑎} ∫{𝑓 <0}
Moreover,
|𝑓 |𝑝 𝑑𝜇 ⩽ 2𝑝 |𝑓 − 𝑎|𝑝 𝑑𝜇, (**)
∫{𝑓 >2𝑎} ∫{𝑓 >2𝑎}
which follows from
𝑓 > 2𝑎 ⇐⇒ |𝑓 − 𝑎| = 𝑓 − 𝑎 > 𝑎.
166
Solution Manual. Last update 28th October 2021
Finally,
|𝑓 |𝑝 𝑑𝜇 ⩽ 2𝑝 |𝑓 − 𝑎|𝑝 𝑑𝜇. (***)
∫{𝑓 ⩽0} ∫{𝑓 ⩽0}
⩽ 2𝑝 |𝑓 − 𝑎|𝑝 𝑑𝜇.
∫
Solution 2 to 13.26: We need the following inequality for 𝑎, 𝑏 ∈ R which follows from Hölder’s
inequality:
Set 𝑏 = 𝑓 (𝑥). Since 𝜇(𝑋) < ∞, this shows that both sides of the claimed integral inequality are
finite.
Assume first that 𝜇(𝑋) = 1. Then we find
𝜙(𝑓 − 𝑚) 𝑑𝜇 ⩽ 𝑐𝜙 𝜙(𝑓 − 𝑎) 𝑑𝜇 ∀𝑎 ∈ R.
∫ ∫
■■
167
14 Product measures and Fubini’s theorem.
Solutions to Problems 14.1–14.20
• We have
( )
⋃ ⋃
(𝑥, 𝑦) ∈ 𝐴𝑖 × 𝐵 ⇐⇒ 𝑥 ∈ 𝐴𝑖 and 𝑦 ∈ 𝐵
𝑖 𝑖
⇐⇒ ∃ 𝑖0 ∶ 𝑥 ∈ 𝐴𝑖0 and 𝑦 ∈ 𝐵
⇐⇒ ∃ 𝑖0 ∶ (𝑥, 𝑦) ∈ 𝐴𝑖0 × 𝐵
⋃
⇐⇒ (𝑥, 𝑦) ∈ (𝐴𝑖 × 𝐵).
𝑖
• We have
( )
⋂ ⋂
(𝑥, 𝑦) ∈ 𝐴𝑖 × 𝐵 ⇐⇒ 𝑥 ∈ 𝐴𝑖 and 𝑦 ∈ 𝐵
𝑖 𝑖
⇐⇒ ∀ 𝑖 ∶ 𝑥 ∈ 𝐴𝑖 and 𝑦 ∈ 𝐵
⇐⇒ ∀ 𝑖 ∶ (𝑥, 𝑦) ∈ 𝐴𝑖 × 𝐵
⋂
⇐⇒ (𝑥, 𝑦) ∈ (𝐴𝑖 × 𝐵).
𝑖
• Using the formula 𝐴 × 𝐵 = 𝜋1−1 (𝐴) ∩ 𝜋2−1 (𝐵) (see page 135 and the fact that inverse maps
interchange with all set operations, we get
[ ] [ ]
(𝐴 × 𝐵) ∩ (𝐴′ × 𝐵 ′ ) = 𝜋1−1 (𝐴) ∩ 𝜋2−1 (𝐵) ∩ 𝜋1−1 (𝐴′ ) ∩ 𝜋2−1 (𝐵 ′ )
[ ] [ ]
= 𝜋1−1 (𝐴) ∩ 𝜋1−1 (𝐴′ ) ∩ 𝜋2−1 (𝐵) ∩ 𝜋2−1 (𝐵 ′ )
= 𝜋1−1 (𝐴 ∩ 𝐴′ ) ∩ 𝜋2−1 (𝐵 ∩ 𝐵 ′ )
= (𝐴 ∩ 𝐴′ ) × (𝐵 ∩ 𝐵 ′ ).
• Using the formula 𝐴 × 𝐵 = 𝜋1−1 (𝐴) ∩ 𝜋2−1 (𝐵) (see page 135 and the fact that inverse maps
interchange with all set operations, we get
169
R.L. Schilling: Measures, Integrals & Martingales
{[ ]𝑐 [ ]𝑐 }
= 𝜋1−1 (𝑋) ∩ 𝜋2−1 (𝐵) ∩ 𝜋1−1 (𝐴) ∪ 𝜋2−1 (𝐵)
[ ]𝑐
= (𝑋 × 𝐵) ∩ 𝜋1−1 (𝐴) ∩ 𝜋2−1 (𝐵)
[ ]𝑐
= (𝑋 × 𝐵) ∩ 𝐴 × 𝐵
= (𝑋 × 𝐵) ⧵ (𝐴 × 𝐵).
• We have
[ ]
𝐴 × 𝐵 ⊂ 𝐴′ × 𝐵 ′ ⇐⇒ (𝑥, 𝑦) ∈ 𝐴 × 𝐵 ⇐⇒ (𝑥, 𝑦) ∈ 𝐴′ × 𝐵 ′
[ ]
⇐⇒ 𝑥 ∈ 𝐴, 𝑦 ∈ 𝐵 ⇐⇒ 𝑥 ∈ 𝐴′ , 𝑦 ∈ 𝐵 ′
⇐⇒ 𝐴 ⊂ 𝐴′ , 𝐵 ⊂ 𝐵 ′ .
■■
Problem 14.2 Solution: Pick two exhausting sequences (𝐴𝑘 )𝑘 ⊂ 𝒜 and (𝐵𝑘 )𝑘 ⊂ ℬ such that
𝜇(𝐴𝑘 ), 𝜈(𝐵𝑘 ) < ∞ and 𝐴𝑘 ↑ 𝑋, 𝐵𝑘 ↑ 𝑌 . Then, because of the continuity of measures,
( )
𝜇 × 𝜈(𝐴 × 𝑁) = lim 𝜇 × 𝜈 (𝐴 × 𝑁) ∩ (𝐴𝑘 × 𝐵𝑘 )
𝑘
( )
= lim 𝜇 × 𝜈 (𝐴 ∩ 𝐴𝑘 ) × (𝑁 ∩ 𝐵𝑘 )
𝑘
[ ]
= lim 𝜇(𝐴 ∩ 𝐴𝑘 ) ⋅ 𝜈(𝑁 ∩ 𝐵𝑘 )
𝑘
⏟⏞⏞⏟⏞⏞⏟ ⏟⏞⏞⏞⏟⏞⏞⏞⏟
<∞ ⩽ 𝜈(𝑁)=0
= 0.
■■
170
Solution Manual. Last update 28th October 2021
( )
+ |𝑓 (𝑥1 , 𝑥2 )| 𝑑𝜇2 (𝑥2 ) 𝑑𝜇1 (𝑥1 ).
∫𝐸1 ⧵𝑁 ∫𝐸 2
The first integral on the right-hand side is, by Theorem 11.2 equal to 0. The second integral
is also 0, due to the definition of the set 𝑁. Using Tonelli’s theorem we see
• (a) ⇔ (c): Use the symmetry in the variables or argue as in “(a) ⇔ (b)”.
■■
Problem 14.4 Solution: Since the two expressions are symmetric in 𝑥 and 𝑦, they must coincide if
they converge. Let us, therefore only look at the left hand side.
clearly satisfies
| −𝑥𝑦 |
|𝑒 sin 𝑥| 𝜆(𝑑𝑥) ⩽ 𝑒−𝑥𝑦 𝜆(𝑑𝑥)
∫(0,∞) | | ∫(0,∞)
∞
= 𝑒−𝑥𝑦 𝑑𝑥
∫0
[ ]∞
𝑒−𝑥𝑦
= −
𝑦 𝑥=0
1
= .
𝑥
Since the integrand is continuous and has only one sign, we can use Riemann’s integral. Thus, the
integral exists. To calculate its value we observe that two integrations by parts yield
∞ ∞
|∞
𝑒−𝑥𝑦 𝑠𝑖𝑛𝑥 𝑑𝑥 = −𝑒−𝑥𝑦 cos 𝑥| − 𝑦𝑒−𝑥𝑦 cos 𝑥 𝑑𝑥
∫0 |𝑥=0 ∫0
∞
=1−𝑦 𝑒−𝑥𝑦 cos 𝑥 𝑑𝑥
∫0
( ∞ )
−𝑥𝑦 |∞ −𝑥𝑦
= 1 − 𝑦 𝑒 sin 𝑥| + 𝑦𝑒 sin 𝑥 𝑑𝑥
|𝑥=0 ∫0
∞
= 1 − 𝑦2 𝑒−𝑥𝑦 sin 𝑥 𝑑𝑥.
∫0
And if we solve this equality for the integral expression, we get
∞ ∞
1
(1 + 𝑦2 ) 𝑒−𝑥𝑦 sin 𝑥 𝑑𝑥 = 1 ⇐⇒ 𝑒−𝑥𝑦 sin 𝑥 𝑑𝑥 = .
∫0 ∫0 1 + 𝑦2
171
R.L. Schilling: Measures, Integrals & Martingales
(Here we use again that improper Riemann integrals with positive integrands coincide with Le-
besgue integrals.)
In principle, the existence and equality of iterated integrals is not good enough to guarantee the
existence of the double integral. For this one needs the existence of the absolute iterated integrals—
cf. Tonelli’s theorem 14.8. In the present case one can see that the absolute iterated integrals exist,
though:
and sin 𝑦
𝑦
is, as a bounded continuous function, Lebesgue integrable over (0, 1).
i.e. ∫𝑘𝜋
(𝑘+1)𝜋 −𝑥𝑦
𝑒 sin 𝑥 𝑑𝑥 = (−1)𝑘 𝑦21+1 (𝑒−(𝑘+1)𝜋𝑦 + 𝑒−𝑘𝜋𝑦 ).
which means that the left hand side is integrable over (1, ∞).
Thus we have
172
Solution Manual. Last update 28th October 2021
∞
sin 𝑦 2 ∑
⩽ 𝜆(𝑑𝑦) + 𝜆(𝑑𝑦) (𝑒−𝜋 )𝑘
∫(0,1] 𝑦 ∫(1,∞) 𝑦2 + 1
𝑘=0
< ∞.
By Fubini’s theorem we know that the iterated integrals as well as the double integral exist and
their values are identical.
Since the integrands are bounded and continuous, we can use Riemann integrals. Fix 𝐴 > 1 and
𝐵 > 1. Then
𝐴 𝐵 1 1 1 𝐵 1 𝐴 𝐴 𝐵
= + + +
∫0 ∫0 ∫0 ∫0 ∫0 ∫1 ∫0 ∫1 ∫1 ∫1
Now we can estimate these expressions separately: since | sin 𝑡| ⩽ |𝑡| we have
1 1 1 1
𝑓 (𝑥, 𝑦) 𝑑𝑦 𝑑𝑥 ⩽ 1 𝑑𝑥 𝑑𝑦 = 1.
∫0 ∫0 ∫0 ∫0
[ ]
1 𝐵 𝐵 1
𝑓 (𝑥, 𝑦) 𝑑𝑦 𝑑𝑥 ⩽ 𝑥𝑒−𝑥𝑦 𝑑𝑥 𝑑𝑦
∫0 ∫1 ∫1 ∫0
1 𝑒−𝐵 − 1 1
=1− + <1− .
𝑒[ 𝐵 ] 𝑒
1 𝐴 𝐴 1
𝑓 (𝑥, 𝑦) 𝑑𝑥 𝑑𝑦 ⩽ 𝑦𝑒−𝑥𝑦 𝑑𝑦 𝑑𝑥
∫0 ∫1 ∫1 ∫0
1 𝑒−𝐴 − 1 1
+ <1− .
=1−
𝑒 𝐴 𝑒
𝐴 𝐵 𝐵[ 𝐴 ]
𝑓 (𝑥, 𝑦) 𝑑𝑥 𝑑𝑦 ⩽ 𝑥𝑒−𝑥𝑦 𝑑𝑥 𝑑𝑦
∫1 ∫1 ∫1 ∫1
1 𝑒−𝐴𝐵 − 𝑒−𝐵 1
= − 𝑒−𝐴 + < .
𝑒 𝐵 𝑒
These estimates now show
∞ ∞
1
𝑒−𝑥𝑦 |sin 𝑥 sin 𝑦| 𝑑𝑥 𝑑𝑦 ⩽ 3 − .
∫0 ∫0 𝑒
■■
1
This much more elegant proof was communicated to me in July 2012 by Alvaro H. Salas from the Universidad Nacional de
Colombia, Department of Mathematics
173
R.L. Schilling: Measures, Integrals & Martingales
𝑥2 − 𝑦2 1 |1 𝜋
𝑑𝑦 𝑑𝑥 = 𝑑𝑥 = arctan 𝑥| = .
∫(0,1) ∫(0,1) (𝑥 + 𝑦 )
2 2 2 ∫(0,1) 𝑥 + 1
2 |0 4
𝑦2 − 𝑥2 𝜋
𝑑𝑦 𝑑𝑥 = −
∫(0,1) ∫(0,1) (𝑥 + 𝑦 )
2 2 2 4
and therefore the double integral can not exist. Since the existence would imply the equality of the
two above integrals. We can see this directly by
| 𝑥2 − 𝑦2 | 1 𝑥
𝑥2 − 𝑦2
| |
| 2 | 𝑑𝑦 𝑑𝑥 ⩾ 𝑑𝑦 𝑑𝑥
∫(0,1) ∫(0,1) | (𝑥 + 𝑦2 )2 | ∫0 ∫0 (𝑥2 + 𝑦2 )2
| |
1
𝑥
= 𝑑𝑥
∫0 𝑥 + 𝑥2
2
1
1 1
= 𝑑𝑥 = ∞.
2 ∫0 𝑥
■■
and {0} is a null set. Thus the iterated integrals have common value 0. But the double integral
does not exist, since for the iterated absolute integrals we get
| | 1∕|𝑦| 1
| 𝑥𝑦 | 𝑑𝑥 = 1 𝜉
𝑑𝜉 ⩾
2 𝜉
𝑑𝜉 .
|
∫(−1,1) | (𝑥 + 𝑦 ) |
2 2 2 | |𝑦| ∫0 (𝜉 + 1)2
2 |𝑦| ∫0 (𝜉 + 1)2
2
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟
<∞
Here we use the substitution 𝑥 = 𝜉|𝑦| and the fact that |𝑦| ⩽ 1, thus 1∕|𝑦| ⩾ 1. But the outer
integral is bounded below by
2
𝑑𝑦 which is divergent.
∫(−1,1) |𝑦|
■■
Problem 14.7 Solution: We use the generic notation 𝑓 (𝑥, 𝑦) for any of the integrands.
a) We have
| 1|
1 |𝑥 − 2 |
𝑓 (𝑥, 𝑦) 𝑑𝑦 = (| |
∫0 )3
𝑥 − 12
174
Solution Manual. Last update 28th October 2021
and this function is not integrable (in 𝑥) in the interval (0, 1). For 0 < 𝑦 ⩽ 1
2
we have
1
1 −𝑦 ( ) 1 ( )
2 1 −3 1 −3
𝑓 (𝑥, 𝑦) 𝑑𝑥 = 𝑥− 𝑑𝑥 + 𝑥− 𝑑𝑥 = 0.
∫0 ∫0 2 ∫ 1 +𝑦 2
2
For 1
2
⩽ 𝑦 ⩽ 1 this integral is again 0. Therefore,
( )
1 1
𝑓 (𝑥, 𝑦) 𝑑𝑥 𝑑𝑦 = 0.
∫0 ∫0
Finally,
1 1 1
| |−2
|𝑓 (𝑥, 𝑦)| 𝑑𝑦 = |𝑥 − 21 | ⇐⇒ |𝑓 (𝑥, 𝑦)| 𝑑𝑥 𝑑𝑦 = ∞.
∫0 | | ∫0 ∫0
b) We have
1 1 1[ ]𝑦=1
𝑥−𝑦 1 𝑥+𝑦
𝑑𝑦 𝑑𝑥 = 𝑑𝑥
∫0 ∫0 (𝑥2 + 𝑦2 )3∕2 ∫0 𝑥 (𝑥2 + 𝑦2 )1∕2 𝑦=0
[ ]
1
𝑥+1
= √ 𝑑𝑥
∫0 𝑥2 + 1 − 1
[ √ ]𝑥=1
𝑥 + 𝑥2 + 1
= ln √
1 + 𝑥2 + 1 − 1 𝑥=0
= ln 2.
c) Since 𝑓 is positive, Tonelli’s theorem ensures that all three integrals coincide. Let 𝑝 ≠ 1. We
get
1 1 1( ) 𝑑𝑥
1
(1 − 𝑥𝑦)−𝑝 𝑑𝑦 𝑑𝑦 = (1 − 𝑥)1−𝑝 − 1 .
∫0 ∫0 𝑝 − 1 ∫0 𝑥
This integral is finite if, and only if, 𝑝 < 2. For 𝑝 = 1 we have
1 1 1
𝑑𝑥
(1 − 𝑥𝑦)−𝑝 𝑑𝑦 𝑑𝑦 = − ln(1 − 𝑥) < ∞.
∫0 ∫0 ∫0 𝑥
175
R.L. Schilling: Measures, Integrals & Martingales
■■
(i) We have [−𝑛, 𝑛] ↑ R as 𝑛 → ∞ and 𝜆([−𝑛, 𝑛]) = 2𝑛 < ∞. This shows 𝜎-finiteness of
𝜆. Let (𝑞𝑗 )𝑗∈N be an enumeration of Q; set 𝐴𝑛 ∶= {𝑞1 , … , 𝑞𝑛 } ∪ (R ⧵ Q), then we have
𝐴𝑛 ↑ R and 𝜁Q (𝐴𝑛 ) = 𝑛 < ∞. This shows 𝜎-finiteness of 𝜁Q .
We will show that 𝜁R is not 𝜎-finite. Assume 𝜁R were 𝜎-finite. Thus, there would be
a sequence 𝐴𝑛 ↑ R, 𝑛 ∈ N, such that 𝜁R (𝐴𝑛 ) < ∞. Since 𝜁R is a counting measure,
every 𝐴𝑛 is countable. Thus, R is a countable union of countable sets, hence countable
– a contradiciton.
(ii) The rationals Q are a 𝜆 null set, hence 1𝑦 Q is for each 𝑦 a 𝜆 null set. We have
1Q (𝑥 ⋅ 𝑦) 𝜆(𝑑𝑥) = 0 ∀ 𝑦 ∈ R.
∫(0,1)
This implies
1Q (𝑥 ⋅ 𝑦) 𝑑𝜆(𝑥) 𝑑𝜁R (𝑦) = 0.
∫(0,1) ∫(0,1)
(iii) Let 𝑥 ∈ (0, 1). The set ( 𝑥1 Q) ∩ (0, 1) contains infinitely many values, so
1Q (𝑥 ⋅ 𝑦) 𝜁R (𝑑𝑦) = ∞ ∀ 𝑥.
∫(0,1)
1Q (𝑥 ⋅ 𝑦) 𝜁Q (𝑑𝑦) = 0 ∀ 𝑥 ∈ (0, 1) ⧵ Q.
∫(0,1)
1Q (𝑥 ⋅ 𝑦) 𝜁Q (𝑑𝑦) = ∞ ∀ 𝑥 ∈ (0, 1) ∩ Q.
∫(0,1)
(v) The results of (iii),(iv) do not contradict Fubini’s or Tonelli’s theorem, since these the-
orems require 𝜎-finiteness of all measures.
■■
(i) Since the integrand is positive, we can use Tonelli’s theorem and work out the integral
as an iterated integral
𝑑𝑥 𝑑𝑦
𝐼 ∶=
∫[0,∞)2 (1 + 𝑦)(1 + 𝑥2 𝑦)
176
Solution Manual. Last update 28th October 2021
( )
1 1
= 𝑑𝑥 𝑑𝑦
∫[0,∞) 1+𝑦 ∫[0,∞) 1 + 𝑥2 𝑦
√
1 arctan(𝑥 𝑦) ||∞
= √ | 𝑑𝑦
∫[0,∞) 1+𝑦 𝑦 |𝑥=0
𝜋 1 1
= [0, ∞) √ 𝑑𝑦.
2∫ 1+𝑦 𝑦
(Observe that the integrand is continuous, which enables us to use Riemann integrals
on bounded intervals. Note that ∫[0,∞) ⋯ = sup𝑛∈N ∫[0,𝑛) … because of monotone con-
√
vergence.) Using the substitution 𝑢 = 𝑦, we get
𝜋 1 |∞ 𝜋2
𝐼= 𝑑𝑢 = 𝜋 arctan(𝑢)| = .
2 ∫[0,∞) 1 + 𝑢2 | 2
|𝑢=0
as well as
1 1 1 1 ∑ −2 𝑛 ∑ −2(𝑛+1)
= = (𝑥 ) = 𝑥 , |𝑥| > 1.
𝑥2 − 1 𝑥2 1 − 𝑥−2 𝑥2 𝑛⩾0 𝑛⩾0
Thus,
ln 𝑥 ∑ ∑
𝑑𝑥 = − 𝑥2𝑛
ln 𝑥 𝑑𝑥 + 𝑥−2(𝑛+1) ln 𝑥 𝑑𝑥. (⋆)
∫(0,∞) 𝑥2 − 1 𝑛⩾0
∫(0,1)
𝑛⩾0
∫(1,∞)
(In order to swap summation and integration, we use dominated convergence!) Using
integration by parts, we find
𝑥2𝑛+1 |1 1
𝑥2𝑛 ln 𝑥 𝑑𝑥 = ln 𝑥|| − 𝑥2𝑛 𝑑𝑥
∫(0,1) 2𝑛 + 1 |𝑥=0 2𝑛 + 1 ∫(0,1)
177
R.L. Schilling: Measures, Integrals & Martingales
1
=−
(2𝑛 + 1)2
𝑥−2(𝑛+1)+1 |∞ 1
𝑥−2(𝑛+1) ln 𝑥 𝑑𝑥 = ln 𝑥|| − 𝑥−2(𝑛+1) 𝑑𝑥
∫(1,∞) −2(𝑛 + 1) + 1 |𝑥=1 −2(𝑛 + 1) + 1 ∫ (1,∞)
1 1
= = .
(−2(𝑛 + 1) + 1)2 (2𝑛 + 1)2
Inserting these results into (⋆), the claim follows from part (ii).
■■
(i) Since 𝜇 is 𝜎-finite, there is an exhausting sequence (𝐺𝑛 )𝑛∈N ⊂ ℬ(R) such that 𝜇(𝐺𝑛 ) < ∞
and 𝐺𝑛 ↑ R. For each 𝑛 ∈ N the set
{ }
1
𝐵𝑘𝑛 ∶= 𝑥 ∈ 𝐺𝑛 ; 𝜇({𝑥}) >
𝑘
is finite. Indeed: Assume there were countably infinitely many (𝑥𝑗 )𝑗∈N ⊂ 𝐵𝑘𝑛 , 𝑥𝑗 ≠ 𝑥𝑖 for
𝑖 ≠ 𝑗. Since the sets {𝑥𝑗 }, 𝑗 ∈ N, are disjoint, we conclude that
( )
∑ ∑
𝜇(𝐺𝑛 ) ⩾ 𝜇 {𝑥𝑗 } = 𝜇({𝑥𝑗 }) = ∞.
𝑗∈N 𝑗∈N
is countable and so is
⋃
𝐷= 𝐵𝑛
𝑛∈N
(ii) For the diagonal 1Δ (𝑥, 𝑦) = 1{𝑦} (𝑥)1R (𝑦) we find from Theorem 14.5:
( )
𝜇 × 𝜈(Δ) = 1{𝑦} (𝑥) 𝜇(𝑑𝑥) 𝜈(𝑑𝑦)
∫R ∫
■■
178
Solution Manual. Last update 28th October 2021
Problem 14.11 Solution: Note that the diagonal Δ ⊂ R2 is measurable, i.e. the (double) integrals
are well-defined. The inner integral on the l.h.S. satisfies
On the other hand, the inner integral on the right-hand side equals
This shows that the double integrals are not equal. This does not contradict Tonelli’s theorem since
𝜇 is not 𝜎-finite.
■■
(i) Note that, due to the countability of N and N × N there are no problems with measurability
and 𝜎-finiteness (of the counting measure).
Tonelli’s Theorem. Let (𝑎𝑗𝑘 )𝑗,𝑘∈N be a double sequence of positive numbers 𝑎𝑗𝑘 ⩾ 0. Then
∑∑ ∑∑
𝑎𝑗𝑘 = 𝑎𝑗𝑘
𝑗∈N 𝑘∈N 𝑘∈N 𝑗∈N
with the understanding that both sides are either finite or infinite.
Fubini’s Theorem. Let (𝑎𝑗𝑘 )𝑗,𝑘∈N ⊂ R be a double sequence of real numbers 𝑎𝑗𝑘 . If
∑∑ ∑∑
|𝑎𝑗𝑘 | or |𝑎𝑗𝑘 |
𝑗∈N 𝑘∈N 𝑘∈N 𝑗∈N
is finite, then all of the following expressions converge absolutely and sum to the same value:
( ) ( )
∑ ∑ ∑ ∑ ∑
|𝑎𝑗𝑘 | , |𝑎𝑗𝑘 | , |𝑎𝑗𝑘 |.
𝑗∈N 𝑘∈N 𝑘∈N 𝑗∈N (𝑗,𝑘)∈N×N
∑ ∑
(ii) Consider the (obviously 𝜎-finite) measures 𝜇𝑗 ∶= 𝑘∈𝐴𝑗 𝛿𝑘 and 𝜈 = 𝑗∈N 𝜇𝑗 . Tonelli’s
theorem tells us that
∑ ∑
|𝑥𝑘 | = |𝑥𝑘 | 𝜇𝑗 (𝑑𝑘) 𝜇(𝑑𝑗)
∫N ∫N
𝑗∈N 𝑘∈𝐴𝑗
179
R.L. Schilling: Measures, Integrals & Martingales
= |𝑥𝑘 | 𝜇(𝑑𝑘)
∫N
∑
= |𝑥𝑘 |.
𝑘∈N
■■
= 1 𝜆1 (𝑑𝑦) 𝜆1 (𝑑𝑥)
∫R ∫[0,𝑢(𝑥)]
= 𝑢(𝑥) 𝜆1 (𝑑𝑥)
∫R
(iii) Measurability follows from (i) and with the hint. Moreover,
= 1 𝜆1 (𝑑𝑦) 𝜆1 (𝑑𝑥)
∫R ∫[𝑢(𝑥),𝑢(𝑥)]
= 𝜆1 ({𝑢(𝑥)}) 𝜆1 (𝑑𝑥)
∫R
= 0 𝜆1 (𝑑𝑥)
∫R
= 0.
■■
Problem 14.14 Solution: The hint given in the text should be good enough to solve this problem....
■■
180
Solution Manual. Last update 28th October 2021
Problem 14.15 Solution: Since (i) implies (ii), we will only prove (i) under the assumption that both
(𝑋, 𝒜 , 𝜇) and (𝑌 , ℬ, 𝜈) are complete measure spaces. Note that we have to assume 𝜎-finiteness of
𝜇 and 𝜈, otherwise the product construction would not work. Pick some set 𝑍 ∈ 𝒫 (𝑋) ⧵ 𝒜 (which
is, because of completeness, not a null-set!), and some 𝜈-null set 𝑁 ∈ ℬ and consider 𝑍 × 𝑁.
= 0;
thus 𝑍×𝑁 ⊂ 𝑋×𝑁 is a subset of a measurable 𝜇×𝜈 null set, hence it should be 𝒜 ⊗ℬ-measurable,
if the product space were complete. On the other hand, because of Theorem 14.17(iii), if 𝑍 × 𝑁
is 𝒜 ⊗ ℬ-measurable, then the section
𝑦∈𝑁
𝑥 → 1𝑍×𝑁 (𝑥, 𝑦) = 1𝑍 (𝑥)1𝑁 (𝑦) = 1𝑍 (𝑥)
■■
(𝑥, 𝑘) ∈ 𝐴 ⇐⇒ 1𝐴 (𝑥, 𝑘) = 1
⇐⇒ ∃ 𝑘 ∈ N ∶ 1𝐴 (𝑥, 𝑘) = 1
⇐⇒ ∃ 𝑘 ∈ N ∶ 𝑥 ∈ 𝐵𝑘
⋃
it is clear that 𝐴 = × {𝑘}.
𝑘∈N 𝐵𝑘
∑
(ii) Let 𝑀 ∈ 𝒫 (N) and set 𝜁 ∶= 𝑗∈N 𝛿𝑗 ; we know that 𝜁 is a (𝜎-finite) measure on 𝒫 (N).
Using Tonelli’s theorem 14.8 we get
∑
𝜋(𝐵 × 𝑀) ∶= 𝜋(𝐵 × {𝑚})
𝑚∈𝑀
∑ 𝑡𝑚
∶= 𝑒−𝑡 𝜇(𝑑𝑡)
𝑚∈𝑀
∫𝐵 𝑚!
𝑡𝑚
= 𝑒−𝑡 𝜇(𝑑𝑡) 𝜁(𝑑𝑚)
∫𝑀 ∫𝐵 𝑚!
𝑡𝑚
= 𝑒−𝑡 𝜇 × 𝜁(𝑑𝑡, 𝑑𝑚)
∬𝐵×𝑀 𝑚!
181
R.L. Schilling: Measures, Integrals & Martingales
The uniqueness follows, however, from the uniqueness theorem for measures (Theorem 5.7):
the family of ‘rectangles’ of the form 𝐵 × 𝑀 ∈ ℬ[0, ∞) × 𝒫 (N) is a ∩-stable generator of
the product 𝜎-algebra ℬ[0, ∞) ⊗ 𝒫 (N) and contains an exhausting sequence, say, [0, ∞) ×
{1, 2, … 𝑘} ↑ [0, ∞) × N. But on this generator 𝜋 is (uniquely) determined by prescribing
the values 𝜋(𝐵 × {𝑚}).
■■
Problem 14.17 Solution: Assume first that 𝜆 ⩾ 0. The point here is that Corollary 14.15 does not
apply to the function 𝑠 → 𝑒−𝜆𝑠 since this function is decreasing and has the value 1 for 𝑠 = 0.
Consider therefore 𝜙(𝑠) ∶= 1 − 𝑒−𝜆𝑠 . This 𝜙 is admissible in 14.15 and we get
( ) ∞
𝜙(𝑇 ) 𝑑P = 1 − 𝑒−𝜆𝑇 𝑑P = 𝜆𝑒−𝜆𝑠 P(𝑇 ⩾ 𝑠) 𝑑𝑠.
∫ ∫ ∫0
Rearranging this equality then yields
∞
−𝜆𝑇
𝑒 𝑑P = 1 − 𝜆 𝑒−𝜆𝑠 P(𝑇 ⩾ 𝑠) 𝑑𝑠.
∫ ∫0
If 𝜆 < 0 the formula remains valid if we understand it in the sense that either both sides are finite
or both sides are infinite. The above argument needs some small changes, though. First, 𝑒−𝜆𝑠
is now increasing (which is fine) but still takes the value 1 if 𝑠 = 0. So we should change to
𝜙(𝑠) ∶= 𝑒−𝜆𝑠 − 1. Now the same calculation as above goes through. If one side is finite, so is
the other; and if one side is infinite, then the other is infinite, too. The last statement follows from
Theorem 14.13 or Corollary 14.15.
■■
the last expression is, however, a product of (combinations of) measurable functions, thus 1𝐵
is measurable and so is then 𝐵.
Without loss of generality we can assume that 𝑎 > 0, all other cases are similar.
182
Solution Manual. Last update 28th October 2021
= 𝜇(𝑑𝑥) 𝜈(𝑑𝑦)
∫(𝑎,𝑏] ∫(𝑎,𝑦]
= 𝜇(𝑎, 𝑦] 𝜈(𝑑𝑦)
∫(𝑎,𝑏]
( )
= 𝜇(0, 𝑦] − 𝜇(0, 𝑎] 𝜈(𝑑𝑦)
∫(𝑎,𝑏]
We remark at this point already that a very similar calculation (with 𝜇, 𝜈 and 𝐹 , 𝐺 inter-
changed and with an open interval rather than a semi-open interval) yields
Combining this formula with the previous one marked (+) reveals that
183
R.L. Schilling: Measures, Integrals & Martingales
(Mind that the sum is at most countable because of Lemma 14.14) from which the claim
follows.
(iv) It is clear that uniform approximation allows to interchange limiting and integration proced-
ures so that we *really* do not have to care about this. We show the formula for monomials
𝑡, 𝑡2 , 𝑡3 , ... by induction. Write 𝜙𝑛 (𝑡) = 𝑡𝑛 , 𝑛 ∈ N.
Induction start 𝑛 = 1: in this case 𝜙1 (𝑡) = 𝑡, 𝜙′1 (𝑡) = 1 and 𝜙(𝐹 (𝑠)) − 𝜙(𝐹 (𝑠−)) − Δ𝐹 (𝑠) = 0,
i.e. the formula just becomes
Induction step 𝑛 ⇝ 𝑛 + 1: Write, for brevity 𝐹 = 𝐹 (𝑠) and 𝐹− = 𝐹 (𝑠−). We have because
of (iii) with 𝐺 = 𝜙𝑛 ◦𝐹 and because of the induction assumption
= 𝐹 𝑛 𝑑𝐹 + 𝐹 𝜙′ (𝐹 ) 𝑑𝐹 +
∫(𝑎,𝑏] − ∫(𝑎,𝑏] − 𝑛 −
∑[ ] ∑
+ 𝐹− 𝜙𝑛 (𝐹 ) − 𝐹− 𝜙𝑛 (𝐹− ) − 𝐹− 𝜙′𝑛 (𝐹− )Δ𝐹 + Δ𝐹 Δ𝐹 𝑛
= 𝐹 𝑛 𝑑𝐹 + 𝐹 𝑛𝐹 𝑛−1 𝑑𝐹 +
∫(𝑎,𝑏] − ∫(𝑎,𝑏] − −
∑[ ]
+ 𝐹− 𝐹 𝑛 − 𝐹−𝑛+1 − 𝐹− 𝑛𝐹−𝑛−1 Δ𝐹 + Δ𝐹 Δ𝐹 𝑛
∑[ ]
= (𝑛 + 1)𝐹−𝑛 𝑑𝐹 + 𝐹− 𝐹 𝑛 − 𝐹−𝑛+1 − 𝑛𝐹−𝑛 Δ𝐹 + Δ𝐹 Δ𝐹 𝑛
∫(𝑎,𝑏]
∑[ ]
= 𝜙′𝑛+1 ◦𝐹− 𝑑𝐹 + 𝐹− 𝐹 𝑛 − 𝐹−𝑛+1 − 𝑛𝐹−𝑛 Δ𝐹 + Δ𝐹 Δ𝐹 𝑛
∫(𝑎,𝑏]
184
Solution Manual. Last update 28th October 2021
𝐹− 𝐹 𝑛 − 𝐹−𝑛+1 − 𝑛𝐹−𝑛 Δ𝐹 + Δ𝐹 Δ𝐹 𝑛
= (𝐹− − 𝐹 )𝐹 𝑛 + 𝐹 𝑛+1 − 𝐹−𝑛+1 − 𝑛𝐹−𝑛 Δ𝐹 + Δ𝐹 Δ𝐹 𝑛
( )
= 𝐹 𝑛+1 − 𝐹−𝑛+1 + Δ𝐹 − 𝐹 𝑛 − 𝑛𝐹−𝑛 + Δ𝐹 𝑛
( )
= 𝐹 𝑛+1 − 𝐹−𝑛+1 + Δ𝐹 − 𝐹 𝑛 − 𝑛𝐹−𝑛 + 𝐹 𝑛 − 𝐹−𝑛
𝑓 (𝑥) ✻
𝜇𝑓 (𝑡)
✻ This is the graph of the associated
distribution function 𝜇𝑓 (𝑡). It is de-
creasing and left-continuous at the
𝑚3 jump points.
𝑡-values are to be measured using Le-
besgue measure in [0, ∞).
𝑚2
( )
𝑚1 = 𝜇 [4, 5]
( )
𝑚1 𝑚2 − 𝑚1 = 𝜇 [6, 9]
( )
✲ 𝑚3 − 𝑚2 = 𝜇 [4, 5]
2 3 4 𝑡
185
R.L. Schilling: Measures, Integrals & Martingales
follows immediately from Theorem 14.13 with 𝑢 = |𝑓 | and 𝜇𝑓 (𝑡) = 𝜇({|𝑓 | ⩾ 𝑡}).
a) ...show the second equality first for (positive) simple functions and use then a (by now
standard...) Beppo Levi/monotone convergence argument to extend the result to all positive
∑
measurable functions. Assume that 𝑓 (𝑥) = 𝑁 𝑗=0 𝑎𝑗 1𝐵𝑗 (𝑥) is a positive simple function in
standard representation, i.e. 𝑎0 = 0 < 𝑎1 < ⋯ < 𝑎𝑛 < ∞ and the sets 𝐵𝑗 = {𝑓 = 𝑎𝑗 } are
pairwise disjoint. Then we have
This proves
𝑛 𝑛
∑ ∑
𝑓 𝑑𝜇 =
𝑝
𝑎𝑝𝑗 𝜇(𝐵𝑗 ) = 𝑎𝑝𝑗 𝜆1 (𝑓 ∗ = 𝑎𝑗 ) = (𝑓 ∗ )𝑝 𝑑𝜆1
∫ ∫
𝑗=0 𝑗=0
and the general case follows from the above-mentioned Beppo Levi argument.
or we can
1
b) use Theorem 14.13 once again with 𝑢 = 𝑓 ∗ and 𝜇 = 𝜆1 provided we know that
( ) ( )
𝜇 {|𝑓 | ⩾ 𝑡} = 𝜆1 {𝑓 ∗ ⩾ 𝑡} .
186
Solution Manual. Last update 28th October 2021
and therefore
( ) ( )
𝜆1 {𝜉 ⩾ 0 ∶ 𝑓 ∗ (𝜉) ⩾ 𝑡} = 𝜆1 {𝜉 ⩾ 0 ∶ 𝜇(|𝑓 | ⩾ 𝑡) ⩾ 𝜉} = 𝜇(|𝑓 | ⩾ 𝑡).
■■
Problem 14.20 Solution: (By Franzsika Kühn) Fix 𝑡 ∈ R. Applying the fundamental theorem of
calculus and Fubini’s theorem, we find
𝑡+ℎ
𝐹 (𝑡 + ℎ) − 𝐹 (𝑡) = (𝜙(𝑡 + ℎ, 𝑥) − 𝜙(𝑡, 𝑥)) 𝜇(𝑑𝑥) = 𝜕𝑡 𝜙(𝑟, 𝑥) 𝑑𝑟 𝜇(𝑑𝑥)
∫𝑋 ∫𝑋 ∫𝑡
𝑡+ℎ
= 𝜕 𝜙(𝑟, 𝑥) 𝜇(𝑑𝑥) 𝑑𝑟.
∫𝑡 ∫𝑋 𝑡
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
=∶𝑓 (𝑟)
def
𝑡+ℎ
1 1
lim (𝐹 (𝑡 + ℎ) − 𝐹 (𝑡)) = lim 𝑓 (𝑟) 𝑑𝑟 = 𝑓 (𝑡) = 𝜕𝑡 𝜙(𝑡, 𝑥) 𝜇(𝑑𝑥).
ℎ→0 ℎ ℎ→0 ℎ ∫𝑡 ∫𝑋
■■
187
15 Integrals with respect to image
measures.
Solutions to Problems 15.1–15.16
𝑢◦𝑇 𝑓 𝑑𝜇 = 𝑢 𝑓 ◦𝑇 −1 𝑑𝑇 (𝜇)
∫ ∫
is again Theorem 15.1.
The third equality finally follows again from Lemma 10.8.
■■
Problem 15.2 Solution: Observe that 𝑇𝜖 is represented by the 𝑛 × 𝑛 diagonal matrix 𝐴 with entries
𝜖. Since det 𝐴 = 𝜖 𝑛 , the claim follows from Example 15.3(iii).
■■
189
R.L. Schilling: Measures, Integrals & Martingales
Let us work out the two integrals separately. For the first expression we find using (∗)
⎧
⎪0, 𝑥 > 0 or 𝑥 > 2,
⎪ 2
𝐼1 (𝑥) = ⎨∫0𝑥 𝑦 = 𝑥 , 𝑥 ∈ [0, 1],
2
⎪ 1 1
⎪∫1−𝑥 𝑦 𝑑𝑦 =
2
(1 − (1 − 𝑥)2 ), 𝑥 ∈ [1, 2].
⎩
𝑥2 1
= 1[0,1] (𝑥) + (1 − (1 − 𝑥)2 )1[1,2] (𝑥).
2 2
A similar calculation for the second integral yields
⎧
⎪0, 𝑥 < 1 or 𝑥 > 3,
⎪
1[0,1] (𝑥 − 𝑦)1[1,2] (𝑦) = 1[𝑥−1,𝑥] (𝑦)1[1,2] (𝑦) = ⎨1[1,𝑥] (𝑦), 𝑥 ∈ [1, 2],
⎪
⎪1[𝑥−1,2] (𝑦), 𝑥 ∈ [2, 3].
⎩
This gives
⎧
⎪0, 𝑥 < 1 or 𝑥 > 3,
⎪
𝐼2 (𝑥) = ⎨∫1𝑥 (2 − 𝑦) 𝑑𝑦 = 2(𝑥 − 1) − 1 (𝑥2 − 1), 𝑥 ∈ [1, 2],
2
⎪ 2
⎪∫𝑥−1 (2 − 𝑦) 𝑑𝑦 = 2(3 − 𝑥) − 1 (4 − (1 − 𝑥)2 ), 𝑥 ∈ [2, 3]
⎩ 2
( ) ( )
1 1
= 2(𝑥 − 1) − (𝑥2 − 1) 1[1,2] (𝑥) + 2(1 + 𝑥) − (4 − (1 − 𝑥)2 ) 1[2,3] (𝑥).
2 2
Finally
( )
𝑥2 3
(1[0,1] ∗ 1[0,1] ∗ 1[0,1] )(𝑥) = 1[0,1] (𝑥) + −𝑥2 + 3𝑥 − 1 (𝑥)+
2 2 [1,2]
( )
1
2(3 − 𝑥) − (4 − (1 − 𝑥)2 ) 1[2,3] (𝑥).
2
■■
190
Solution Manual. Last update 28th October 2021
Assume that 𝑥0 ∈ (supp 𝑢 + supp 𝑤)𝑐 . Since this is an open set, there is some 𝑟 > 0 such that
𝐵𝑟 (𝑥0 ) ⊂ (supp 𝑢 + supp 𝑤)𝑐 . Pick any 𝑥 ∈ 𝐵𝑟 (𝑥0 ). For all 𝑦 ∈ supp 𝑤 we find 𝑥 − 𝑦 ∉ supp 𝑢. In
particular,
𝑢(𝑥 − 𝑦) ⋅ 𝑤(𝑦) = 0 ∀ 𝑦 ∈ supp 𝑤.
This implies that 𝑢(𝑥 − 𝑦)𝑤(𝑦) = 0 for all 𝑦 ∈ R𝑛 . From the definition of the convolution we see
that (𝑢 ∗ 𝑤)(𝑥) = 0. Since 𝑥 ∈ 𝐵𝑟 (𝑥0 ) is arbitrary, we get 𝑥0 ∉ supp(𝑢 ∗ 𝑤).
■■
Fix 𝑦 ∈ (0, ∞) and define 𝜃𝑦 ∶= 𝑦−1 𝑥. From Theorem 7.10 we know that the image
measure 𝜃𝑦 (𝜆)(𝑑𝑧) of 𝜆 is given by 𝑦𝜆(𝑑𝑧) gegeben ist, and because of Theorem 15.1 we
get
𝑑𝑥 𝑑𝑥
𝑢(𝑥𝑦−1 ) = 𝑦−1 𝑢(𝑥𝑦−1 ) −1
∫(0,∞) 𝑥 ∫(0,∞) 𝑥𝑦
𝜃𝑦 (𝜆)(𝑑𝑧)
= 𝑦−1 𝑢(𝑧)
∫(0,∞) 𝑧
(⋆⋆)
𝑑𝑧
= 𝑢(𝑧) .
∫(0,∞) 𝑧
If we insert this into (⋆), we obtain
( )
𝑑𝑧 𝑑𝑦
𝑢 ⊛ 𝑤(𝑥) 𝜇(𝑑𝑥) = 𝑢(𝑧) 𝑤(𝑦)
∫(0,∞) ∫(0,∞) ∫(0,∞) 𝑧 𝑦
= 𝑢 𝑑𝜇 𝑤 𝑑𝜇.
∫(0,∞) ∫(0,∞)
191
R.L. Schilling: Measures, Integrals & Martingales
(ii) Consider first the case 𝑝 = ∞: As |𝑢(𝑥𝑦−1 )| ⩽ ‖𝑢‖𝐿∞ (𝜇) for 𝜇-a.a. 𝑦 ∈ (0, ∞), we get
= ‖𝑤‖𝑝−1
1
|𝑢(𝑥𝑦−1 )|𝑝 |𝑤(𝑦)| 𝜇(𝑑𝑦),
∫(0,∞)
= ‖𝑤‖𝑝−1
1
|𝑢|𝑝 𝑑𝜇 |𝑤| 𝑑𝜇
∫ ∫
= ‖𝑤‖1 ‖𝑢‖𝑝𝑝 .
■■
192
Solution Manual. Last update 28th October 2021
( )
= 𝜇 𝑇 −1 (𝐵 ∩ 𝐶)
( )
= 𝜇 𝑇 −1 (𝐵) ∩ 𝑇 −1 (𝐶)
( )
= 𝜇 𝐴 ∩ 𝑇 −1 (𝐶)
( )
= 𝜇|𝐴 𝑇 −1 (𝐶)
= 𝑇 (𝜇|𝐴 )(𝐶).
■■
Problem 15.7 Solution: By definition, we find for any Borel set 𝐵 ∈ ℬ(R𝑛 )
= 1𝐵 (𝑥 + 𝑡) 𝛿𝑦 (𝑑𝑡)
∫
= 1𝐵 (𝑥 + 𝑦)
= 1𝐵 (𝑧 + 𝑡) 𝜇(𝑑𝑡)
∫
■■
= 𝜇(𝐵 − 𝑦) 𝜈(𝑑𝑦).
∫
Similarly we get
193
R.L. Schilling: Measures, Integrals & Martingales
■■
Problem 15.9 Solution: Because of Tonelli’s theorem we can iterate the very definition of ‘convolu-
tion’ of two measures, Definition 15.4(iii), and get
=𝑛 |𝜔1 | P(𝑑𝜔1 )
∫
where we use the symmetry of the iterated integrals in the integrating measures as well as the fact
that P(R𝑛 ) = ∫ P(𝑑𝜔𝑘 ) = 1. Note that we could have +∞ on either side, i.e. the integrability
condition is only important for the second assertion.
The equality ∫ 𝜔 P⋆𝑛 (𝑑𝜔) = 𝑛 ∫ 𝜔 P(𝑑𝜔) follows with same calculation (note that we do not get
an inequality as there is no need for the triangle inequality at point (*) above). The integrability
condition is now needed since the integrands are no longer positive. Note that, since 𝜔 ∈ R𝑛 ,
the above equality is an equality between vectors in R𝑛 ; this is no problem, just read the equality
coordinate-by-coordinate.
■■
Problem 15.10 Solution: Since the convolution 𝑝 → 𝑢 ⋆ 𝑝 is linear, it is enough to consider monomi-
als of the form 𝑝(𝑥) = 𝑥𝑘 . Thus, by the binomial formula,
𝑢 ⋆ 𝑝(𝑥) = 𝑢(𝑥 − 𝑦) 𝑦𝑘 𝑑𝑦
∫
= 𝑢(𝑦) (𝑥 − 𝑦)𝑘 𝑑𝑦
∫
194
Solution Manual. Last update 28th October 2021
𝑘 ( )
∑ 𝑘 𝑗
= 𝑥 𝑢(𝑦) 𝑦𝑘−𝑗 𝑑𝑦.
𝑗=0
𝑗 ∫
Since supp 𝑢 is compact, there is some 𝑟 > 0 such that supp 𝑢 ⊂ 𝐵𝑟 (0) and we get for any 𝑚 ∈ N0 ,
and in particular for 𝑚 = 𝑘 − 𝑗 or 𝑚 = 𝑘, that
| |
| 𝑢(𝑦)𝑦𝑚 𝑑𝑦| ⩽ ‖𝑢‖∞ |𝑦|𝑚 𝑑𝑦
|∫ | ∫
| | supp 𝑢
⩽ ‖𝑢‖∞ 𝑟𝑚 𝑑𝑦
∫𝐵𝑟 (0)
= 2𝑟 ⋅ 𝑟𝑚 ⋅ ‖𝑢‖∞
which is clearly finite. This shows that 𝑢 ⋆ 𝑝 exists and that it is a polynomial.
■■
Problem 15.11 Solution: That the convolution 𝑢 ⋆ 𝑤 is bounded and continuous follows from The-
orem 15.8.
■■
Problem 15.12 Solution: (This solution is written for 𝑢 ∈ 𝐶𝑐 (R𝑛 ) and 𝑤 ∈ 𝐶 ∞ (R𝑛 )).
Let 𝜕𝑖 = 𝜕∕𝜕𝑥𝑖 denote the partial derivative in direction 𝑥𝑖 where 𝑥 = (𝑥1 , … , 𝑥𝑛 ) ∈ R𝑛 . Since
𝑤 ∈ 𝐶 ∞ ⇐⇒ 𝜕𝑖 𝑤 ∈ 𝐶 ∞ ,
𝜕 𝜕
𝑢(𝑦)𝑤(𝑥 − 𝑦) 𝑑𝑦 = 𝑢(𝑦) 𝑤(𝑥 − 𝑦) 𝑑𝑦
𝜕𝑥𝑖 ∫ ∫ 𝜕𝑥𝑖
( )
= 𝑢(𝑦) 𝜕𝑥𝜕 𝑤 (𝑥 − 𝑦) 𝑑𝑦
∫ 𝑖
195
R.L. Schilling: Measures, Integrals & Martingales
= 𝑢 ⋆ 𝜕𝑖 𝑤(𝑥).
■■
Problem 15.13 Solution: Let 𝜒𝑡 be a Friedrichs mollifier. From Lemma 15.10 we know
| |
|𝑢(𝑥) − 𝑢 ∗ 𝜒𝑡 (𝑥)| = || (𝑢(𝑥) − 𝑢(𝑦))𝜒𝑡 (𝑥 − 𝑦) 𝑑𝑦||
|∫ |
−𝑛
( 𝑥−𝑦 )
⩽ |𝑢(𝑥) − 𝑢(𝑦)|𝑡 𝜒 𝑡 𝑑𝑦
∫
In the last step we use the integrable dominating function 2‖𝑢‖∞ 𝜒(𝑢).
■■
Problem 15.14 Solution: The measurability considerations are just the same as in Theorem 15.6, so
we skip this part.
By assumption,
1 1 1
+ =1+ ;
𝑝 𝑞 𝑟
We can rewrite this as
[ ] [ ]
1 1 1 1 1
+ − + − = 1. (*)
𝑟 𝑝 𝑟 𝑞 𝑟
⏟⏞⏟⏞⏟ ⏟⏞⏟⏞⏟
= 1− 1𝑞 ∈[0,1) = 1− 1𝑝 ∈[0,1)
Now write the integrand appearing in the definition of 𝑢 ⋆ 𝑤(𝑥) in the form
[ ] [ ] [ ]
|𝑢(𝑥 − 𝑦)𝑤(𝑦)| = |𝑢(𝑥 − 𝑦)|𝑝∕𝑟 |𝑤(𝑦)|𝑞∕𝑟 ⋅ |𝑢(𝑥 − 𝑦)|1−𝑝∕𝑟 ⋅ |𝑤(𝑦)|1−𝑞∕𝑟
and apply the generalized Hölder inequality (cf. Problem 13.5) with the exponents from (*):
196
Solution Manual. Last update 28th October 2021
[ ]1 [ ]1−1 [ ]1−1
𝑟 𝑝 𝑟 𝑞 𝑟
⩽ |𝑢(𝑥 − 𝑦)| |𝑤(𝑦)| 𝑑𝑦
𝑝 𝑞
|𝑢(𝑥 − 𝑦)| 𝑑𝑦𝑝 𝑞
|𝑤(𝑦)| 𝑑𝑦 .
∫ ∫ ∫
Raising this inequality to the 𝑟th power we get, because of the translation invariance of Lebesgue
measure,
[ ]
|𝑢 ⋆ 𝑤(𝑥)|𝑟 ⩽ |𝑢(𝑥 − 𝑦)|𝑝 |𝑤(𝑦)|𝑞 𝑑𝑦 ‖𝑢‖𝑟−𝑝 𝑟−𝑞
𝑝 ⋅ ‖𝑤‖𝑞
∫
= |𝑢|𝑝 ⋆ |𝑤|𝑞 (𝑥) ⋅ ‖𝑢‖𝑟−𝑝 𝑟−𝑞
𝑝 ⋅ ‖𝑤‖𝑞 .
Now we integrate this inequality over 𝑥 and use Theorem 15.6 for 𝑝 = 1 and the integral
Thus,
■■
Problem 15.15 Solution: For 𝑁 = 1 the inequality is trivial, for 𝑁 = 2 it is in line with Problem
15.14 with 𝑝 = 𝑞.
Let us, first of all, give a heuristic derivation of this result which explains how one arrives at the
particular form for the value of 𝑝 = 𝑝(𝑟, 𝑁). We may assume that 𝑁 ⩾ 2. Set 𝐹𝑗 ∶= 𝑓𝑗 ⋆ … ⋆ 𝑓𝑁
for 𝑗 = 1, 2, … 𝑁 − 1. Then
‖𝑓1 ⋆ ⋯ ⋆ 𝑓𝑁 ‖𝑟
⩽ ‖𝑓1 ‖𝑝 ‖𝐹2 ‖𝑞2 = ‖𝑓1 ‖𝑝 ‖𝑓2 ⋆ 𝐹3 ‖𝑞2
(1 )
by Pr. 15.14 where 1
𝑟
+1= 1
𝑝
+ 1
𝑞2
= 𝑝
−1 + 1
𝑞2
+1
197
R.L. Schilling: Measures, Integrals & Martingales
where 𝑝 = 𝑁𝑟
(𝑁−1)𝑟+1
and 𝑞 is given by
1 1 1 (𝑁 − 1)𝑟 + 1 1 1 1 1
+1= + = + =1+ − +
𝑟 𝑞 𝑞 𝑁𝑟 𝑞 𝑞 𝑁 𝑁𝑟
so that
𝑁𝑟
𝑞= .
𝑁 +𝑟−1
Using the induction hypothesis we now get
( )
‖𝑓1 ⋆ ⋯ ⋆ 𝑓𝑁 ‖𝑟 ⩽ ‖𝑓1 ‖𝑝 ⋅ ‖𝑓2 ⋆ ⋯ ⋆ 𝑓𝑁 ‖𝑞 ⩽ ‖𝑓1 ‖𝑝 ⋅ ‖𝑓2 ‖𝑠 ⋯ ‖𝑓𝑁 ‖𝑠
198
Solution Manual. Last update 28th October 2021
(i)
2𝜋 2𝜋
𝑢 ⋆ 𝑣(𝑥) = 1R (𝑥 − 𝑦) sin 𝑦 𝑑𝑦 = sin 𝑦 𝑑𝑦 = 0 ∀ 𝑥.
∫0 ∫0
(ii) Since all functions 𝑢, 𝑣, 𝑤, 𝜙 are continuous, we can use the usual rules for the (Riemann)
integral and get, using integration by parts and the fundamental theorem of integral calculus,
𝑥
𝑑
𝑣 ⋆ 𝑤(𝑥) = 𝑑𝑥
𝜙(𝑥 − 𝑦) 𝜙(𝑡) 𝑑𝑡 𝑑𝑥
∫ ∫−∞
( 𝑑 ) 𝑦
= − 𝑑𝑦 𝜙(𝑥 − 𝑦) 𝜙(𝑡) 𝑑𝑡 𝑑𝑥
∫ ∫−∞
𝑦
𝑑
= 𝜙(𝑥 − 𝑦) 𝑑𝑦 𝜙(𝑡) 𝑑𝑡 𝑑𝑥
∫ ∫−∞
= 𝜙(𝑥 − 𝑦) 𝜙(𝑦) 𝑑𝑦
∫
= 𝜙 ⋆ 𝜙(𝑥).
If 𝑥 ∈ (0, 4𝜋), then 𝑥 − 𝑦 ∈ (0, 2𝜋) for some suitable 𝑦 = 𝑦= and even for all 𝑦 from an
interval (𝑦0 − 𝜖, 𝑦0 + 𝜖) ⊂ (0, 2𝜋). Since 𝜙 is positive with support [0, 2𝜋], the positivity
follows.
(iii) Obviously,
(𝑖)
(𝑢 ⋆ 𝑣) ⋆ 𝑤 = 0 ⋆ 𝑤 = 0
while
= 𝑣 ⋆ 𝑤(𝑦) 𝑑𝑦
∫
= 𝜙 ⋆ 𝜙(𝑦) 𝑑𝑦
∫
> 0.
Note that 𝑤 is not an (𝑝th power, 𝑝 < ∞) integrable function so that we cannot use Fubini’s
theorem to prove associativity of the convolution.
■■
199
16 Jacobi’s transformation theorem.
Solutions to Problems 16.1–16.12
for closed sets 𝐶𝑘 resp. 𝐶𝑘𝑖 . Since complements of closed sets are open, we find, using the rules
for (countable) unions and intersections that
𝑛 𝑛 𝑛
⋂ ⋂ ⋃ ⋃ ⋂
(i) 𝐹𝑖 = 𝐶𝑘𝑖 = 𝐶𝑘𝑖 .
𝑖=1 𝑖=1 𝑘∈N 𝑘∈N 𝑖=1
⏟⏟⏟
closed set
⋃ ⋃⋃ ⋃
(ii) 𝐹𝑖 = 𝐶𝑘𝑖 = 𝐶𝑘𝑖 .
𝑖∈N 𝑖∈N 𝑘∈N (𝑖,𝑘)∈N×N
⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
countable union!
⋂ ⋂ ⋂ [ ]𝑐 ⋂ [ ]𝑐
Moreover, 𝐹𝑖𝑐 = 𝐶𝑘𝑖 = 𝐶𝑘𝑖 .
𝑖∈N 𝑖∈N 𝑘∈N (𝑖,𝑘)∈N×N
⏟⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏟
countable intersection!
[ ]𝑐
⋃ ⋃ ⋂
(iii) 𝐹 = 𝐶𝑘 ⇐⇒ 𝐹 = 𝑐
𝐶𝑘 = 𝐶𝑘𝑐 .
𝑘∈N 𝑘∈N 𝑘∈N ⏟⏟⏟
open
⋃
(iv) Set 𝑐1 ∶= 𝐶 and 𝐶𝑖 = ∅, 𝑖 ⩾ 2. Then 𝐶 = 𝐶𝑖 is an 𝐹𝜎 -set.
𝑖∈N
■■
Problem 16.2 Solution: Write 𝜆 = 𝜆𝑛 and ℬ = ℬ(R𝑛 ). Fix 𝐵 ∈ ℬ. According to Lemma 16.12
there are sets 𝐹 ∈ 𝐹𝜎 and 𝐺 ∈ 𝐺𝛿 such that
𝐶1 ∪ ⋯ ∪ 𝐶𝑁 ⊂ 𝐵 ⊂ 𝑈1 ∩ ⋯ ∩ 𝑈𝑀
and
(*)
| |
|𝜆(𝑈1 ∩ ⋯ ∩ 𝑈𝑀 ) − 𝜆(𝐵)| ⩽ 𝜖,
| |
201
R.L. Schilling: Measures, Integrals & Martingales
(**)
| |
|𝜆(𝐵) − 𝜆(𝐶1 ∪ ⋯ ∪ 𝐶𝑁 )| ⩽ 𝜖.
| |
Since finite unions of closed sets are closed and finite intersections of open sets are open, (*) proves
outer regularity while (**) proves inner regularity (w.r.t. close sets).
To see inner regularity with compact sets, we note that the closed set 𝐶 ′ ∶= 𝐶1 ∪ ⋯ ∪ 𝐶𝑁 is
approximated by the following compact sets
𝐾𝓁 ∶= 𝐵𝓁 (0) ∩ 𝐶 ′ ↑ 𝐶 ′ as 𝓁 → ∞
and, because of the continuity of measures, we get for suitably large 𝐿 = 𝐿𝜖 ∈ N that
| |
|𝜆(𝐾𝐿 ) − 𝜆(𝐶1 ∪ ⋯ ∪ 𝐶𝑁 )| ⩽ 𝜖
| |
| |
|𝜆(𝐾𝐿 ) − 𝜆(𝐵)| ⩽ 2𝜖.
| |
■■
Solution 1: Following the hint we get (with the notation of Problem 11.6)
̄ ∗ ) = 𝜆∗ (𝐵 ∗ )
𝜆(𝐵) = 𝜆(𝐵
= inf 𝜆(𝐴) (by 11.6)
ℬ∋𝐴⊃𝐵 ∗
= inf
′ ∗
𝜆(𝑈 ′ ) (by 16.2)
𝑈 ⊃𝐵
̄ ∗ ) = 𝜆∗ (𝐵 ∗ )
𝜆(𝐵) = 𝜆(𝐵
= sup 𝜆(𝐴) (by 11.6)
ℬ∋𝐴⊂𝐵 ∗
202
Solution Manual. Last update 28th October 2021
Solution 2: (without Problem 11.6). Using the definition of the completion we get
⩽ sup 𝜆(𝐶)
𝐶⊂𝐵 ∗
⩽ sup 𝜆(𝐶 ′′ )
𝐶 ′′ ⊂𝐵∪𝑁
= 𝜆(𝐵 ∪ 𝑁)
= 𝜆(𝐵)
as well as
⩽ inf ∗ 𝜆(𝑈 )
𝑈 ⊃𝐵
⩽ inf 𝜆(𝑈 ′′ )
𝑈 ′′ ⊃𝐵∪𝑁
= 𝜆(𝐵 ∪ 𝑁)
= 𝜆(𝐵).
■■
where 𝑥𝑖 , 𝑦𝑖 ∈ {0, 2}. In order to enforce uniqueness, we only want to have truly infinite
sums, i.e. we use 0.002222 … instead of 0.01000 … etc.
203
R.L. Schilling: Measures, Integrals & Martingales
(ii) Let 𝛼(𝑥, 𝑦) = 𝑥−𝑦 as in the hint. This is a Lipschitz (Hölder-1) continuous map from R2 → R
and it has the following property: 𝐶 × 𝐶 → 𝛼(𝐶, 𝐶) = [−1, 1]. But 𝐶 × 𝐶 is a Lebesgue null
set in R2 while 𝜆1 [−1, 1] = 2. This situation cannot occur in Corollary 16.14.
■■
(i) Obviously, 𝒢 ⊂ ℬ[0, ∞). On the other hand, 𝜎(𝒢 ) contains all open intervals of the form
⋃[ ) [ )
(𝛼, 𝛽) = 𝛼 − 1𝑛 , ∞ ⧵ 𝛽, ∞ , 0⩽𝛼<𝛽<∞ (*)
𝑛∈N
Thus,
so that 𝑈 ∩ [0, ∞) ∈ 𝒪 ∩ [0, ∞) is indeed a countable union of sets of the form (*) and (**).
Thus,
(ii) That 𝜇 is a measure follows from Lemma 10.8 (for a proof, see the online section ‘additional
material’). Since
−1
𝜌(𝐵) = 𝜇(𝑇1∕5 (𝐵)) = 𝑇1∕5 (𝜌)(𝐵)
Since
204
Solution Manual. Last update 28th October 2021
■■
so that
( )
̃ 𝑟)
𝜕 Φ(𝑥,
̃ 𝑟) =
𝐷Φ(𝑥,
𝜕(𝑥, 𝑟)
′
⎛1 − 𝑟 𝜕 √ 𝑓 (𝑥) 𝑓 ′ (𝑥) + 𝑟 𝜕𝑥
𝜕
√
1 ⎞
𝜕𝑥 1+[𝑓 ′ (𝑥)]2 1+[𝑓 ′ (𝑥)]2 ⎟
=⎜ ′
⎜ − √ 𝑓 (𝑥) √
1 ⎟
⎝ ′ 2
1+[𝑓 (𝑥)] 1+[𝑓 ′ (𝑥)]2 ⎠
205
R.L. Schilling: Measures, Integrals & Martingales
and
′ 𝑓 ′′
− √𝑓
𝜕 1 1+[𝑓 ′ ]2
√ = .
𝜕𝑥 1 + [𝑓 ′ (𝑥)]2 1 + [𝑓 ′ ]2
1 + [𝑓 ′ ]2 𝑟 𝑓 ′′
=√ −
1 + [𝑓 ′ ]2 1 + [𝑓 ]
′ 2
√ 𝑟 𝑓 ′′
= 1 + [𝑓 ′ ]2 −
1 + [𝑓 ′ ]2
If 𝑥 is from a compact set, say [𝑐, 𝑑], we can, because of the continuity of 𝑓 , 𝑓 ′ and
𝑓 ′′ , achieve that for sufficiently small values of |𝑟| < 𝜖 we get that det 𝐷Φ
̃ > 0, i.e. Φ
̃
is a local 𝐶 1 -diffeomorphism.
(v) The set is a ‘tubular’ neighbourhood of radius 𝑟 around the graph Γ𝑓 for 𝑥 ∈ [𝑐, 𝑑].
Measurability follows, since Φ
̃ is a diffeomorphism, from the fact that the set 𝐶(𝑟) is
the image of the cartesian product of measurable sets.
(vi) Because of part (iv) we have, for fixed 𝑥 and sufficiently small values of 𝑟, that the
determinant is positive so that
1 | ̃ 𝑠)|| 𝜆1 (𝑑𝑠)
lim |det 𝐷Φ(𝑥,
𝑟↓0 2𝑟 (−𝑟,𝑟)
∫ | |
|√ ′′ |
= lim
1 | 1 + (𝑓 ′ (𝑥))2 − 𝑠 𝑓 (𝑥) | 𝜆1 (𝑑𝑠)
|
𝑟↓0 2𝑟 ∫(−𝑟,𝑟) | 1 + (𝑓 ′ (𝑥))2 ||
(√ )
1 𝑠 𝑓 ′′ (𝑥)
= lim 1 + (𝑓 ′ (𝑥))2 − 𝜆1 (𝑑𝑠)
𝑟↓0 2𝑟 ∫(−𝑟,𝑟) 1 + (𝑓 ′ (𝑥))2
1 √
= lim 1 + (𝑓 ′ (𝑥))2 𝜆1 (𝑑𝑠)
𝑟↓0 2𝑟 ∫(−𝑟,𝑟)
1 𝑠 𝑓 ′′ (𝑥)
− lim 𝜆1 (𝑑𝑠)
𝑟↓0 2𝑟 ∫(−𝑟,𝑟) 1 + (𝑓 ′ (𝑥))2
√ 𝑓 ′′ (𝑥) 1
= 1 + (𝑓 ′ (𝑥))2 − lim 𝑠 𝜆1 (𝑑𝑠)
1 + (𝑓 (𝑥))
′ 2 𝑟↓0 2𝑟 ∫ (−𝑟,𝑟)
206
Solution Manual. Last update 28th October 2021
√
= 1 + (𝑓 ′ (𝑥))2
| ̃ 0)||.
= |det 𝐷Φ(𝑥,
| |
(vii) We have
1
1 (𝑥, 𝑦) 𝜆2 (𝑑𝑥, 𝑑𝑦)
2𝑟 ∫R2 𝐶(𝑟)
1
= 1 ̃ −1 (𝑥, 𝑦) 𝜆2 (𝑑𝑥, 𝑑𝑦)
2𝑟 ∫R2 Φ(Φ (𝐶)×(−𝑟,𝑟))
1
(Thm 16.4)
| ̃ 𝑠)||𝜆2 (𝑑𝑧, 𝑑𝑠)
= 1Φ−1 (𝐶)×(−𝑟,𝑟) (𝑧, 𝑠) |det 𝐷Φ(𝑧,
2𝑟 R2
∫ | |
[ ]
1
(Tonelli)
| |
̃ 𝑠)|𝜆 (𝑑𝑠) 𝜆1 (𝑑𝑧).
1
= 1 −1 (𝑧) |det 𝐷Φ(𝑧,
∫R Φ (𝐶) 2𝑟 ∫(−𝑟,𝑟) | |
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
←←←←𝑟↓0
←←→ ̃
← |det 𝐷Φ(𝑧,0)|
Since Φ−1 (𝐶) is a bounded subset of R, we can use the result of part (vii) and domin-
ated convergence and the proof is finished.
(viii) This follows from (i)–(iii) and the fact that
| √
̃ 0)|| = 1 + (𝑓 ′ (𝑥))2
|det 𝐷Φ(𝑥,
| |
and the geometrical meaning of the weighted area 1
2𝑟
𝜆2 (𝐶(𝑟))—recall that 𝐶(𝑟) was a
tubular neighbourhood of the graph.
■■
(ii) This is the formula from part (i) with Φ = 𝜃𝑟 ; observe that 𝜃𝑟 (R𝑛 ) = R𝑛 .
(iii) The equality
207
R.L. Schilling: Measures, Integrals & Martingales
■■
■■
det𝐷Φ(𝑟, 𝜃, 𝜔)
( )
−𝑟 sin 𝜃 cos 𝜔 −𝑟 cos 𝜃 sin 𝜔
= sin 𝜔 det
𝑟 cos 𝜃 cos 𝜔 −𝑟 sin 𝜃 sin 𝜔
( )
cos 𝜃 cos 𝜔 −𝑟 sin 𝜃 cos 𝜔
+ 𝑟 cos 𝜔 det
sin 𝜃 cos 𝜔 𝑟 cos 𝜃 cos 𝜔
( )
= sin 𝜔 𝑟2 sin2 𝜃 cos 𝜔 sin 𝜔 + 𝑟2 cos2 𝜃 cos 𝜔 sin 𝜔
( )
+ 𝑟 cos 𝜔 𝑟 cos2 𝜃 cos2 𝜔 + 𝑟 sin2 𝜃 cos2 𝜔
■■
208
Solution Manual. Last update 28th October 2021
(ii) We have to show that 𝐵(𝑥, 𝑦)Γ(𝑥 + 𝑦) = Γ(𝑥)Γ(𝑦). Using polar coordinates in (i) we see
∞ 2𝜋
2
Γ(𝑥)Γ(𝑦) = 4 𝑒−𝑟 𝑟2𝑥+2𝑦−1 (cos 𝜙)2𝑥−1 (sin 𝜙)2𝑦−1 𝑑𝜙 𝑑𝑟
∫𝑟=0 ∫𝜙=0
( ∞ ) ( 𝜋∕2 )
=4 −𝑟2 2𝑥+2𝑦−1
𝑒 𝑟 𝑑𝑟 (cos 𝜙) 2𝑥−1
(sin 𝜙) 2𝑦−1
𝑑𝜙 . (⋆)
∫𝑟=0 ∫𝜙=0
Setting 𝑠 ∶= 𝑟2 we see
∞ ∞
2 1 1
𝑒−𝑟 𝑟2𝑥+2𝑦−1 𝑑𝑟 = 𝑒−𝑠 𝑠(𝑥+𝑦)−1 𝑑𝑠 = Γ(𝑥 + 𝑦).
∫𝑟=0 2 ∫𝑠=0 2
Change variables in the second integral of (⋆) according to 𝑡 = cos2 𝜙 and use sin2 𝜙 +
cos2 𝜙 = 1. This yields
𝜋∕2 1
1 1
(cos 𝜙)2𝑥−1 (sin 𝜙)2𝑦−1 𝑑𝜙 = 𝑡2𝑥−1 (1 − 𝑡)2𝑦−1 𝑑𝑡 = 𝐵(𝑥, 𝑦).
∫𝜙=0 2 ∫0 2
■■
Thus,
𝑥𝑚 𝑦𝑛 𝑑𝜆2 (𝑥, 𝑦)
∬
‖𝑥‖2 +‖𝑦‖2 <1
1 2𝜋
= 𝑟𝑛+𝑚+1 cos𝑚 𝜃 sin𝑛 𝜃 𝑑𝑟 𝑑𝜃
∫0 ∫0
( 1 )( 2𝜋 )
= 𝑟 𝑛+𝑚+1
𝑑𝑟 cos 𝜃 sin 𝜃 𝑑𝜃
𝑚 𝑛 (*)
∫0 ∫0
( 2𝜋 )
𝑟𝑚+𝑛+2 ||𝑟=1
= cos 𝜃 sin 𝜃 𝑑𝜃
𝑚 𝑛
𝑚 + 𝑛 + 2 ||𝑟=0 ∫0
2𝜋
1
= cos𝑚 𝜃 sin𝑛 𝜃 𝑑𝜃.
𝑚 + 𝑛 + 2 ∫0
209
R.L. Schilling: Measures, Integrals & Martingales
Since the l.h.s. of the expression (*) is symmetric in 𝑚 and 𝑛, so is the r.h.s. and we get
𝑥𝑚 𝑦𝑛 𝑑𝜆2 (𝑥, 𝑦) = 0
∬
‖𝑥‖2 +‖𝑦‖2 <1
Introducing planar polar coordinates yields, as seen above, for even 𝑚 and 𝑛,
𝜋∕2
4
4 𝑥𝑚 𝑦𝑛 𝑑𝜆2 (𝑥, 𝑦) = cos𝑚 𝜃 sin𝑛 𝜃 𝑑𝜃
∬ 𝑚 + 𝑛 + 2 ∫0
‖𝑥‖2 +‖𝑦‖2 <1
𝑥>0, 𝑦>0
1
4 𝑚−1 𝑛−1
= (1 − 𝑡2 ) 2 (𝑡2 ) 2 𝑡 𝑑𝑡
𝑚 + 𝑛 + 2 ∫0
√ √
where we use the substitution 𝑡 = sin 𝜃 and cos 𝜃 = 1 − sin2 𝜃 = 1 − 𝑡2 . A further substitution
𝑠 = 𝑡2 yields
1
2 𝑚−1 𝑛−1
= (1 − 𝑠) 2 𝑠 2 𝑑𝑠
𝑚 + 𝑛 + 2 ∫0
1
2 𝑚+1 𝑛+1
= (1 − 𝑠) 2 −1 𝑠 2 −1 𝑑𝑠
𝑚 + 𝑛 + 2 ∫0
2 ( 𝑛+1 )
= 𝐵 𝑚+1 ,
𝑚+𝑛+2 2 2
which is Euler’s Beta function. There is a well-known relation between the Euler Beta- and Gamma
functions:
Γ(𝑥)Γ(𝑦)
𝐵(𝑥, 𝑦) = (*)
Γ(𝑥 + 𝑦)
210
Solution Manual. Last update 28th October 2021
so that, finally,
⎧0
⎪ 𝑚 or 𝑛 odd;
( 𝑚+1 ) ( 𝑛+1 )
⎪ Γ Γ
2 ⎪ 2
(2 𝑛+𝑚+2 )2 else
𝑥 𝑦 𝑑𝜆 (𝑥, 𝑦) = ⎨
𝑚 𝑛 𝑚+𝑛+2 Γ
∬ ⎪ ( 𝑚+1 ) (2 𝑛+1 )
‖𝑥‖2 +‖𝑦‖2 <1 Γ Γ
⎪ 2 2
⎪ = ( 𝑛+𝑚+4 )
⎩ Γ
2
Let us briefly sketch the proof of (*): our calculation shows that
𝜋∕2
𝐵(𝑥, 𝑦) = 2 sin2𝑥−1 𝜃 cos2𝑦−1 𝜃 𝑑𝜃;
∫0
multiplying this formula with 𝑟2𝑥+2𝑦−1 𝑒−𝑟 , integrating w.r.t. 𝑟 over (0, ∞) and changing variables
2
■■
211
17 Dense and determining sets.
Solutions to Problems 17.1–17.9
Problem 17.1 Solution: Let 𝑓 ∈ 𝑝 (𝜇) and fix 𝜖 > 0. It is enough to show that there is some ℎ ∈
such that ‖𝑓 − ℎ‖𝑝 ⩽ 𝜖. Since is dense in 𝑝 (𝜇), there exists some 𝑔 ∈ satisfying ‖𝑓 − 𝑔‖𝑝 ⩽
𝜖∕2. On the other hand, as is dense in , there is some ℎ ∈ such that ‖𝑔 − ℎ‖𝑝 ⩽ 𝜖∕2. Now
the triangle inequality gives
𝜖 𝜖
‖𝑓 − ℎ‖𝑝 ⩽ ‖𝑓 − 𝑔‖𝑝 + ‖𝑔 − ℎ‖𝑝 ⩽ + .
2 2
■■
(i) Continuity follows from the continuity of the function 𝑥 → 𝑑(𝑥, 𝐴), cf. (17.1). Clearly,
0 ⩽ 𝑢𝑘 ⩽ 1 and 𝑢𝑘 |𝐾 = 1 and 𝑢|𝑈𝑘𝑐 = 0. Since 𝑈𝐾 ↓ 𝐾, we get 𝑢𝑘 ↓ 1𝐾 . Since 𝑈 𝑘 is
closed and bounded, it is clear that 𝑈 𝑘 is compact, i.e. supp 𝑢𝑘 is compact.
(iii) We have 𝜇(𝐾) = 𝜈(𝐾) for all compact sets 𝐾 ⊂ R𝑛 and the compact sets generate
the Borel 𝜎-algebra. In particular, this holds for [−𝑘, 𝑘]𝑛 ↑ R𝑛 , so that the conditions
for the uniqueness theorem for measures (Theorem 5.7) are satisfied. We conclude that
𝜇 = 𝜈.
(iv) Since each 𝑥 has a compact neighbourhood, we can choose 𝑘 so large that 𝐵1∕𝑘 (𝑥)
⋃
becomes compact. In particular, 𝐾 ⊂ 𝑥∈𝐾 𝐵1∕𝑘(𝑥) (𝑥) is an open cover. We can choose
each 𝑘(𝑥) so large, that 𝐵1∕𝑘(𝑥) (𝑥) has a compact closure. Since 𝐾 is compact, we find
⋃
finitely many 𝑥𝑖 such that 𝐾 ⊂ 𝑖 𝐵1∕𝑘(𝑥𝑖 ) (𝑥𝑖 ) = 𝑈𝑘 where 𝑘 ∶= max𝑖 𝑘𝑖 . In particular,
⋃
𝑈 𝑘 ⊂ 𝑖 𝐵1∕(𝑥𝑖 ) (𝑥𝑖 ) is compact. This produces a sequence of 𝑈𝑘 ↓ 𝐾. The rest follows
almost literally as in the previous steps.
■■
(i) We have to show that ‖𝜏ℎ 𝑓 ‖𝑝𝑝 = ‖𝑓 ‖𝑝 for all 𝑝 ∈ 𝑝 (𝑑𝑥). This is an immediate con-
sequence of the invariance of Lebesgue measure under translations:
213
R.L. Schilling: Measures, Integrals & Martingales
(ii) We show the assertion first for 𝑓 ∈ 𝐶𝑐 (R). If 𝑓 ∈ 𝐶𝑐 (R), then 𝐾 ∶= supp 𝑓 is compact.
Pick 𝑅 > 0 in such a way that 𝐾 + 𝐵1 (0) ⊂ 𝐵𝑅 (0). Since limℎ→0 𝑓 (𝑥 − ℎ) = 𝑓 (𝑥) and
Now take 𝑓 ∈ 𝑝 (𝑑𝑥). Since 𝐶𝑐 (R) is dense in 𝑝 (𝑑𝑥), cf. Theorem 17.8, there is a
sequence (𝑓𝑛 )𝑛∈N ⊂ 𝐶𝑐 (R) such that ‖𝑓𝑛 − 𝑓 ‖𝑝 → 0. From part (i) we get
← 2‖𝑓𝑛 − 𝑓 ‖𝑝 ←←←←←←←←←→
←←←←←←←←→ ← 0.
ℎ→0 𝑛→∞
This finishes the proof of the first assertion. The second claim follows in a similar way.
Consider first 𝑓 ∈ 𝐶𝑐 (R) and 𝐾 ∶= supp 𝑓 . Since 𝐾 is compact, there is some 𝑅 > 0
with (ℎ + 𝐾) ∩ 𝐾 = ∅1 for all ℎ > 𝑅. If ℎ > 𝑅, then
and so
= |𝑓 (𝑦)|𝑝 𝑑𝑦 + |𝑓 (𝑥)|𝑝 𝑑𝑥
∫𝐾 ∫𝐾
= 2‖𝑓 ‖𝑝𝑝 .
This proves the assertion for 𝑓 ∈ 𝐶𝑐 (R), and the general case follows via density as in the
first part of (ii).
■■
1 | 𝑥+ℎ |
| |
|𝑀ℎ 𝑓 (𝑥)| 𝑑𝑥 = | 𝑓 (𝑡) 𝑑𝑡| 𝑑𝑥
∫ 2ℎ ∫ |∫𝑥−ℎ |
| |
ℎ
1
⩽ |𝑓 (𝑥 + 𝑡)| 𝑑𝑥 𝑑𝑡 ⩽ ‖𝑓 ‖1
2ℎ ∫−ℎ ∫
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟
∫ |𝑓 (𝑦)| 𝑑𝑦=‖𝑓 ‖1
214
Solution Manual. Last update 28th October 2021
(ii) Assume first that 𝑓 ∈ 𝐶𝑐 (R). Because of the continuity of the function 𝑓 we find
ℎ
1
|𝑀ℎ 𝑓 (𝑥) − 𝑓 (𝑥)| ⩽ |𝑓 (𝑥 + 𝑡) − 𝑓 (𝑥)| 𝑑𝑥 ⩽ sup |𝑓 (𝑥 + 𝑡) − 𝑓 (𝑥)| ←←←←←←←←→
← 0
2ℎ ∫−ℎ 𝑡∈[−ℎ,ℎ] ℎ→0
for all 𝑥 ∈ R. Since the support of 𝑓 , 𝐾 ∶= supp 𝑓 , is compact, there is some 𝑅 > 0
such that 𝐾 + 𝐵1 (0) ⊆ 𝐵𝑅 (0). For ℎ < 1 we get 𝑀ℎ 𝑓 (𝑥) = 0 = 𝑓 (𝑥) if 𝑥 ∉ 𝐵𝑅 (0). Since
|𝑀ℎ 𝑓 (𝑥)| ⩽ |𝑓 (𝑥)| for 𝑥 ∈ R, we get
|𝑀ℎ 𝑓 (𝑥) − 𝑓 (𝑥)| = |𝑀ℎ 𝑓 (𝑥) − 𝑓 (𝑥)|1𝐵 (𝑥) ⩽ 2‖𝑓 ‖∞ 1𝐵 (𝑥) ∈ 1 (𝑑𝑥).
𝑅 (0) 𝑅 (0)
i.e. the claim is true for any 𝑓 ∈ 𝐶𝑐 (R). Now we take a general 𝑓 ∈ 1 (𝑑𝑥). Because of
Theorem 17.8 there is a sequence (𝑓𝑛 )𝑛∈N ⊂ 𝐶𝑐 (R) such that ‖𝑓𝑛 − 𝑓 ‖1 → 0. Therefore,
← 2‖𝑓𝑛 − 𝑓 ‖1 ←←←←←←←←←→
←←←←←←←←→ ← 0.
ℎ→0 𝑛→∞
■■
(i) Let 𝐴 ∈ ℬ(𝑋) such that 𝑓 ∶= 1𝐴 ∈ 𝑝 (𝜇). Clearly, 𝜇(𝐴) < ∞ and because of the
outer regularity of 𝜇 there is an open set 𝑈 ⊂ 𝑋 such that 𝐴 ⊂ 𝑈 and 𝜇(𝑈 ) < ∞.
Literally as in the proof of Lemma 17.3 we can construct some 𝜙𝜖 ∈ 𝐶Lip (𝑋) ∩ 𝑝 (𝜇)
with ‖𝑓 − 𝜙𝜖 ‖𝑝 ⩽ 𝜖 (just replace in the proof 𝐶𝑏 (𝑋) with 𝐶Lip (𝑋)).
(ii) If 𝑓 ∈ 𝑝 (𝜇), then the Sombrero lemma shows that there is a sequence of simple func-
tions (𝑓𝑛 )𝑛∈N satisfying 0 ⩽ 𝑓𝑛 ⩽ 𝑓 , 𝑓𝑛 ↑ 𝑓 . Using the monotone convergence
theorem, we see ∫ (𝑓 − 𝑓𝑛 )𝑝 𝑑𝜇 ↓ 0; in particular, there is some 𝑛 ∈ N such that
‖𝑓𝑛 − 𝑓 ‖𝑝 ⩽ 𝜖. Using linearity and the result of part (i), we get some 𝜙𝜖 ∈ 𝐶Lip (𝑋)
such that ‖𝑓𝑛 − 𝜙𝜖 ‖𝑝 ⩽ 𝜖. Therefore,
‖𝑓 − 𝜙𝜖 ‖𝑝 ⩽ ‖𝑓 − 𝑓𝑛 ‖𝑝 + ‖𝑓𝑛 − 𝜙𝜖 ‖𝑝 ⩽ 2𝜖.
■■
215
R.L. Schilling: Measures, Integrals & Martingales
(i) Let (𝑥𝑛 )𝑛∈N be a countable dense subset of 𝑋. By assumption, each 𝑥𝑛 has a relatively
compact open neighbourhood: 𝑥𝑛 ∈ 𝑉𝑛 and 𝑉 𝑛 is compact. Since 𝐵1∕𝑘 (𝑥𝑛 ) ⊂ 𝑉𝑛 for
sufficiently large values of 𝑘 ⩾ 𝑘0 (𝑥𝑛 ), we see that the balls 𝐵1∕𝑘 (𝑥𝑛 ), 𝑘 ⩾ 𝑘0 (𝑥𝑛 ), are
also relatively compact. Thus,
is a sequence of relatively compact, open sets. For any open set 𝑈 ⊂ 𝑋 we find
⋃
𝑈= 𝑈𝑛 .
𝑛∈N
𝑈𝑛 ⊂𝑈
(The inclusion ‘⊃’ is obvious. In order to see ‘⊂’ we observe that for any 𝑥 ∈ 𝑈 there
is some 𝑟 > 0 with 𝐵𝑟 (𝑥) ⊂ 𝑈 . Since (𝑥𝑛 )𝑛∈N is dense, we may choose 𝑛 ∈ N and
𝑘 ⩾ 𝑘0 (𝑥𝑛 ) such that 𝐵1∕𝑘 (𝑥𝑛 ) ⊂ 𝐵𝑟 (𝑥) ⊂ 𝑈 .)
(iii) Assume that 𝑈 ⊂ 𝑋 is an open set such that 𝜇(𝑈 ) < ∞ and let (𝑈𝑛 )𝑛∈N be the sequence
from part (i). Because of (i), there is a subsequence (𝑈𝑛(𝑘) )𝑘∈N ⊂ (𝑈𝑛 )𝑛∈N such that
⋃ ⋃
𝑈 = 𝑘 𝑈𝑛(𝑘) . Set 𝑊𝑛 ∶= 𝑛𝑘=1 𝑈𝑛(𝑘) and observe that 𝑊𝑛 ∈ . Since 𝑊𝑛 ↑ 𝑈 ,
Beppo Levi’s theorem shows that
‖1𝑊𝑛 − 1𝑈 ‖𝑝 ←←←←←←←←←→
← 0.
𝑛→∞
Obviously, the 𝐺𝑛 are open sets, 𝐺𝑛 ↑ 𝑋 and 𝜇(𝐺𝑛 ) < ∞ – here we use that the 𝑈𝑘 are
relatively compact and that 𝜇 is finite on compact sets. This means that the assumptions
of Theorem H.3 are satisfied, and we see that 𝜇 is outer regular.
Let 𝐵 ∈ ℬ(𝑋), 𝜇(𝐵) < ∞ and fix 𝜖 > 0. Since 𝜇 is outer regular, there is a sequence
of open sets (𝑈𝑛 )𝑛∈N such that 𝑈𝑛 ⊃ 𝐵 and 𝜇(𝑈𝑛 ) < ∞. By monotone convergence,
‖1𝑈𝑛 − 1𝐵 ‖𝑝 → 0 as 𝑛 → ∞. Pick 𝑛 ∈ N such that ‖1𝑈𝑛 − 1𝐵 ‖𝑝 ⩽ 𝜖. Because of (iii),
there is some 𝐷 ∈ with ‖1𝑈𝑛 − 1𝐷 ‖𝑝 ⩽ 𝜖. Consequently,
(v) By definition, ⊂ 𝑝 (𝜇), i.e. it is enough to show that for every 𝑓 ∈ 𝑝 (𝜇) and
𝜖 > 0 there is some 𝐷 ∈ such that ‖𝑓 − 1𝐷 ‖𝑝 ⩽ 𝜖. Using the Sombrero lemma
(Corollary 8.9) and the dominated convergence theorem we can construct a sequence
216
Solution Manual. Last update 28th October 2021
𝑁
∑
𝑓𝑛 (𝑥) = 𝑐𝑗 1𝐵𝑗 (𝑥)
𝑗=1
■■
(i) Assume first that 𝐴 is an open set. Without loss of generality 𝐴 ≠ ∅. Fix 𝜖 > 0. Since
{ }
𝑥 ∈ 𝐴 ∶ 𝑑(𝑥, 𝐴𝑐 ) < 1𝑛 ↓ ∅ as 𝑛 → ∞
In order to show the claim for any Borel set 𝐴 ∈ ℬ(𝑋), we proceed as in the proof of
Lemma 17.3: let 𝑈 ⊂ 𝑋, 𝜇(𝑈 ) < ∞, and define
As in the proof of Lemma 17.3 we see that 𝒟 is a Dynkin system. By construction, the
open sets are contained in 𝒟 , and so ℬ(𝑈 ) ⊂ 𝒟 .
If 𝐴 ∈ ℬ(𝑋) is an arbitrary Borel set with 1𝐴 ∈ 𝑝 (𝜇), we have 𝜇(𝐴) < ∞. Since 𝜇
is outer regular, there exists an open set 𝑈 ⊂ 𝑋 such that 𝐴 ⊂ 𝑈 and 𝜇(𝑈 ) < ∞. Since
𝐴 ∈ ℬ(𝑈 ) ⊂ 𝒟 , the claim follows.
(ii) Let 𝑓 ∈ 𝑝 (𝜇), 0 ⩽ 𝑓 ⩽ 1, and fix 𝜖 > 0. Without loss of generality we may assume
that ‖𝑓 ‖∞ = 1, otherwise we would use 𝑓 ∕‖𝑓 ‖∞ . The (proof of the) Sombrero lemma
(Theorem 8.8) shows that
𝑛2𝑛 −1 2𝑛 −1
∑ 𝑘 0⩽𝑓 ⩽1 ∑ 𝑘
𝑓𝑛 ∶= 1{ 𝑘 ⩽𝑓 < 𝑘+1 } + 𝑛1{𝑓 >𝑛} = 1{ 𝑘 ⩽𝑓 < 𝑘+1 } , 𝑛 ∈ N,
𝑘=0
2 𝑛
2𝑛 2𝑛
𝑘=0
2 𝑛
2𝑛 2𝑛
217
R.L. Schilling: Measures, Integrals & Martingales
2𝑛−1 −1 { }
⋃ 2𝑘 + 1 2𝑘 + 2
𝐴𝑗 = ⩽ 𝑓 < .
𝑘=0
2𝑗 2𝑗
Observe that 1𝐴𝑗 ⩽ 2𝑗 𝑓 ∈ 𝑝 (𝜇). Because of part (i), there is for every 𝑗 ⩾ 1 a function
𝜙𝑗,𝜖 ∈ 𝐶𝑏 (𝑋) ∩ 𝑝 (𝜇) such that
and
𝜖 𝜖
‖𝜙𝑗,𝜖 − 𝜙𝑗 ‖𝑝 ⩽ , 𝜇{𝜙𝑗,𝜖 ≠ 𝜙𝑗 } ⩽ 𝑗 ‖𝜙𝑗,𝜖 ‖∞ ⩽ ‖𝜙𝑗 ‖∞ ⩽ 1.
2 𝑗 2
∑
The function 𝜙𝜖 ∶= enjoys all required properties:
𝜙𝑗,𝜖
𝑗⩾1 2𝑗
∑ ∑
• ‖𝜙𝜖 ‖∞ ⩽ = 1 = ‖𝑓 ‖∞ .
‖𝜙𝑗,𝜖 ‖∞ 1
𝑗⩾1 2𝑗
⩽ 𝑗⩾1 2𝑗
∑ ∑
• ‖𝜙𝜖 − 𝑓 ‖𝑝 ⩽ 𝑗⩾1 21𝑗 ‖𝜙𝑗,𝜖 − 𝜙𝑗 ‖𝑝 ⩽ 𝜖 𝑗⩾1 21𝑗 ⩽ 𝜖. In particular, 𝜙𝜖 ∈ 𝑝 (𝜇).
∑ ∑
• 𝜇{𝜙𝜖 ≠ 𝑓 } ⩽ 𝑗⩾1 𝜇{𝜙𝑗,𝜖 ≠ 𝜙𝑗 } ⩽ 𝑗⩾1 𝜖2−𝑗 = 𝜖.
(iii) Observe, first of all, that the theorem holds for all 𝑔 ∈ 𝑝 (𝜇) with 0 ⩽ 𝑔 ⩽ ‖𝑔‖∞ < ∞;
for this, apply part (ii) to 𝑔∕‖𝑔‖∞ . Without loss of generality we may assume for such
𝑔 that 𝜙𝜖 ⩾ 0; otherwise we would consider 𝜙̃𝜖 ∶= 𝜙𝜖 ∨ 0.
218
Solution Manual. Last update 28th October 2021
and
{ }
‖𝜓𝜖 ‖∞ ⩽ ‖𝑓 − ‖∞ , 𝜇 𝑓 − ≠ 𝜓𝜖 ⩽ 𝜖 and ‖𝑓 − − 𝜓𝜖 ‖𝑝 ⩽ 𝜖.
as well as
‖Φ𝜖 ‖∞ ⩽ max{‖𝑓 + ‖∞ , ‖𝑓 − ‖∞ } = ‖𝑓 ‖∞
(this step requires that 𝜙𝜖 ⩾ 0 and 𝜓𝜖 ⩾ 0). The triangle inequality yields
‖𝑓 − Φ𝜖 ‖𝑝 ⩽ ‖𝑓 + − 𝜙𝜖 ‖𝑝 + ‖𝑓 − − 𝜓𝜖 ‖𝑝 ⩽ 2𝜖.
(iv) Fix 𝑓 ∈ 𝑝 (𝜇) and 𝜖 > 0. Using the Markov inequality we get
1
𝜇{|𝑓 | ⩾ 𝑅} ⩽ |𝑓 |𝑝 𝑑𝜇.
𝑅𝑝 ∫
In particular, we can pick a sufficiently large 𝑅 > 0 such that 𝜇{|𝑓 | ⩾ 𝑅} ⩽ 𝜖. Using
monotone convergence, we see
|𝑓 |𝑝 𝑑𝜇 < 𝜖
∫{|𝑓 |>𝑅}
‖𝜙𝜖 − 𝑓 ‖𝑝𝑝
⩽ ‖𝜙𝜖 − 𝑓𝑅 ‖𝑝𝑝 + 𝐼1 + 𝐼2 .
𝐼1 = (𝑓 − 𝑅)𝑝 𝑑𝜇 + (−𝑅 − 𝑓 )𝑝 𝑑𝜇
∫{𝑓 >𝑅}∩{𝜙𝜖 =𝑓𝑅 } ∫{𝑓 <−𝑅}∩{𝜙𝜖 =𝑓𝑅 }
⩽ 𝑓 𝑝 𝑑𝜇 + (−𝑓 )𝑝 𝑑𝜇
∫{𝑓 >𝑅}∩{𝜙𝜖 =𝑓𝑅 } ⏟⏟⏟ ∫{𝑓 <−𝑅}∩{𝜙𝜖 =𝑓𝑅 } ⏟⏟⏟
|𝑓 |𝑝 |𝑓 |𝑝
⩽ |𝑓 |𝑝 𝑑𝜇 < 𝜖.
∫{|𝑓 |>𝑅}
219
R.L. Schilling: Measures, Integrals & Martingales
|∑ | (∑ )1 ( 𝑛 )1
| 𝑛 | 𝑛 𝑝 ∑ 𝑞
| 𝑥𝑗 ⋅ 𝑦𝑗 | ⩽ |𝑥 | 𝑝
⋅ |𝑦 | 𝑞
| | 𝑗 𝑗
| 𝑗=1 |
| | 𝑗=1 𝑗=1
■■
Problem 17.8 Solution: We see immediately that ∫𝑎 𝑝(𝑥)𝑓 (𝑥) 𝑑𝑥 = 0 for all polynomials 𝑝. Fix
𝑏
𝑔 ∈ 𝐶[𝑎, 𝑏] and 𝜖 > 0. By Weierstraß’ theorem, there is some polynomial 𝑝 such that ‖𝑔−𝑝‖∞ ⩽ 𝜖.
Therefore,
| 𝑏 | | 𝑏 𝑏 |
| |
| 𝑔(𝑥)𝑓 (𝑥) 𝑑𝑥| = || (𝑔(𝑥) − 𝑝(𝑥))𝑓 (𝑥) 𝑑𝑥 + 𝑝(𝑥)𝑓 (𝑥) 𝑑𝑥 ||
|∫𝑎 | | ∫𝑎 ∫𝑎 |
| |
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟
=0
𝑏
⩽ |𝑝(𝑥) − 𝑔(𝑥)| |𝑓 (𝑥)| 𝑑𝑥
∫𝑎 ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟
⩽𝜖
𝑏
⩽𝜖 |𝑓 (𝑥)| 𝑑𝑥.
∫𝑎
From this we conclude that
𝑏
𝑔(𝑥)𝑓 (𝑥) 𝑑𝑥 = 0 ∀𝑔 ∈ 𝐶[𝑎, 𝑏].
∫𝑎
220
Solution Manual. Last update 28th October 2021
Define measures 𝜇± by 𝜇 ± (𝑑𝑥) ∶= 1[𝑎,𝑏] (𝑥)1{±𝑓 >0} (𝑥) 𝑑𝑥. Then ∫ 𝑔 𝑑𝜇+ = ∫ 𝑔 𝑑𝜇− for all
𝑔 ∈ 𝐶[𝑎, 𝑏]. According to Theorem 17.12, 𝐶[𝑎, 𝑏] is a determining set, and so 𝜇 + = 𝜇 − . This is
only possible if 𝜇 = 0, hence 𝑓 = 0 Lebesgue a.e.
■■
(i) First of all, we note that it is enough to know that the polynomials are uniformly dense
in the set 𝐶[−1, 1]. This follows immediately from the observation that any function in
𝐶[0, 1] can be mapped onto 𝐶[𝑎, 𝑏] using the affine transform 𝑎 + 𝑡(𝑏 − 𝑎), 𝑡 ∈ [0, 1] –
and vice versa. Fix 𝑢 ∈ 𝐶[−1, 1] and define a sequence of polynomials (𝑝𝑛 )𝑛∈N by
( )𝑛
1 𝑥2
𝑝𝑛 (𝑥) ∶= −1 , 𝑥 ∈ R,
𝑐𝑛 16
where 𝑐𝑛 ∶= ∫−4 (𝑥2 ∕16−1)𝑛 𝑑𝑥. Since 𝑢 ∈ 𝐶[−1, 1], there is some ̃
𝑢 ∈ 𝐶(R) such that
4
| |
|𝑢 (𝑥) − ̃
𝑢(𝑥)|| = || (̃
𝑢(𝑥 − 𝑦) − ̃ 𝑝𝑛 (𝑦) 𝑑𝑦||
𝑢(𝑥))̃
| 𝑛
|∫ |
⩽ [ 𝑢(𝑥
] |̃ − 𝑦) − ̃
𝑢(𝑥)|̃
𝑝𝑛 (𝑦) 𝑑𝑦
∫ − 𝑅1 , 𝑅1
+ [ 𝑢(𝑥
] |̃ − 𝑦) − ̃
𝑢(𝑥)|̃
𝑝𝑛 (𝑦) 𝑑𝑦
∫R⧵ − 𝑅1 , 𝑅1
=∶ 𝐼1 (𝑥) + 𝐼2 (𝑥)
𝐼1 (𝑥) ⩽ [
sup ]
𝑢(𝑥 − 𝑦) − ̃
|̃ 𝑢(𝑥)| [ ̃𝑛 (𝑦) 𝑑𝑦
]𝑝
∫ − 𝑅1 , 𝑅1
𝑦∈ − 𝑅1 , 𝑅1
221
R.L. Schilling: Measures, Integrals & Martingales
⩽ [
sup ]
𝑢(𝑥 − 𝑦) − ̃
|̃ 𝑢(𝑥)|
𝑦∈ − 𝑅1 , 𝑅1
← 0
←←←←←←←←←←→
𝑅→∞
𝐼2 (𝑥) ⩽ 2‖̃
𝑢‖∞ [ ̃𝑛 (𝑦) 𝑑𝑦.
]𝑝
∫R⧵ − 𝑅1 , 𝑅1
Since 𝑝̃𝑛 (𝑦) ↓ 0 for all 𝑦 ≠ 0, we can use the monotone convergence theorem to conclude
that 𝐼2 ←←←←←←←←←→
← 0 uniformly in 𝑥. This proves the claim.
𝑛→∞
(ii) Fix 𝑢 ∈ 𝐶𝑐 [0, ∞). Since 𝑢 has compact support, 𝑢(𝑥) = 0 for large 𝑥; in particular,
𝑢◦(− log)(𝑥) = 0 if 𝑥 is small. Therefore,
⎧
⎪𝑢◦(− log)(𝑥), 𝑥 ∈ (0, 1]
⎨
⎪0, 𝑥 = 0,
⎩
defines a continuous function on [0, 1]. According to (i), there is a sequence of polyno-
mials (𝑝𝑛 )𝑛∈N with 𝑝𝑛 → 𝑢◦(− log) uniformly.
(iii) For 𝑝(𝑥) ∶= 𝑥𝑛 we obviously have 𝑝(𝑒−𝑡 ) = 𝑒−𝑛𝑡 = 𝜖𝑛 (𝑡) and, by assumption,
Using the linearity of the integral, this equality extends to arbitrary polynomials 𝑝.
Assume that 𝑢 ∈ 𝐶𝑐 [0, ∞) and (𝑝𝑛 )𝑛∈N as in (ii). Since 𝑝𝑛 converges uniformly to
𝑢◦(− log), we can interchange integration and limit to get
= 𝑢 𝑑𝜈.
∫
■■
222
18 Hausdorff measure.
Solutions to Problems 18.1–18.7
Problem 18.1 Solution: This is clear from the monotonicity of the infimum and the fact that there
are more 𝒫 -𝛿-covers than 𝒞 -𝛿-covers, i.e. we have
𝜙 𝜙
𝛿,𝒫 (𝐴) ⩽ 𝛿,𝒞 (𝐴).
■■
Problem 18.2 Solution: From the proof of Corollary 18.10 we know, using the monotonicity of
measures
𝜙
(𝐴) = 𝜙 (𝐺) = lim 𝜙 (𝑈 𝑘 )
𝑘→∞
𝑈𝑘 ⊃ 𝐴 { } 𝑈 ⊃𝐴 𝜙
⩾ inf 𝜙 (𝑈 ) ∶ 𝑈 ⊃ 𝐴, 𝑈 open ⩾ (𝐴).
When using the monotonicity we must make sure that 𝜙 (𝑈 𝑘 ) < ∞ – this we can enforce by
𝑈 𝑘 ⇝ 𝑈 𝑘 ∩ 𝑈 (where 𝑈 is the open set with finite Hausdorff measure).
For counting measure this is clearly violated: Any open set 𝑈 ⊃ 𝐴 ∶= {𝑎} has infinitely many
points! Nevertheless 𝐴 is itself a 𝐺𝛿 -set.
■■
⋂
Problem 18.3 Solution: By Corollary 18.10 there are open sets 𝑈𝑖 such that 𝐻 ∶= 𝑖 𝑈𝑖 ⊃ 𝐵 and
𝜙 (𝐻 ⧵ 𝐵) = 0 or 𝜙 (𝐻) = 𝜙 (𝐵). Now we can write each 𝑈𝑖 as an 𝐹𝜎 -set:
⋃
𝑈𝑖 = 𝐵𝑟∕2 (𝑥)
𝐵𝑟 (𝑥)⊂𝑈𝑖 ,𝑥∈𝑈𝑖
is indeed a countable union of closed sets, since 𝑈𝑖 ⊂ 𝑋 contains a countable dense subset. So we
have
⋃
𝑈𝑖 = 𝐹𝑖𝑘 for closed sets 𝐹𝑖𝑘 .
𝑘
Without loss of generality we may assume that the sets 𝐹𝑖𝑘 increase in 𝑘, otherwise we would
consider 𝐹𝑖1 ∪ ⋯ ∪ 𝐹𝑖𝑘 . By the continuity of measure (here we require the measurability of 𝐵!)
we have
223
R.L. Schilling: Measures, Integrals & Martingales
𝜙 (𝐵 ⧵ 𝐹𝑖𝑘(𝑖) ) ⩽ 𝜖∕2𝑖 , 𝑖 ∈ N.
⋂
Consider the closed set 𝐹 = 𝑖 𝐹𝑖𝑘(𝑖) and observe that
∑ ∑ 𝜖
𝜙 (𝐹 ) ⩾ 𝜙 (𝐹 ∩ 𝐵) ⩾ 𝜙 (𝐵) − 𝜙 (𝐵 ⧵ 𝐹𝑖𝑘(𝑖) ) ⩾ 𝜙 (𝐵) − = 𝜙 (𝐵) − 𝜖.
𝑖 𝑖
2 𝑖
⋂
Since 𝐹 ⊂ 𝑖 𝑈𝑖 , we get
( )
⋂
(𝐹 ⧵ 𝐵) ⩽
𝜙 𝜙
𝑈𝑖 ⧵ 𝐵 = 𝜙 (𝐻 ⧵ 𝐵) = 0.
𝑖
⋂
By Corollary 18.10, the set 𝐹 ⧵ 𝐵 is contained in a 𝐺𝛿 -set 𝐺 = 𝑖 𝑉𝑖 (where the 𝑉𝑖 are open sets)
such that 𝜙 (𝐺) =0= 𝜙 (𝐹 ⧵ 𝐵). Thus,
⋃ ⋃
𝐹 ⧵𝐺 =𝐹 ∩ 𝑉𝑖𝑐 = 𝐹 ∩ 𝑉𝑖𝑐
𝑖 𝑖 ⏟⏟⏟
closed
𝜙 (𝐹 ⧵ 𝐺) ⩾ 𝜙 (𝐹 ) − 𝜙 (𝐺) ⩾ 𝜙 (𝐵) − 𝜖.
Now consider 𝜖 = 1
𝑛
and take unions of the thus obtained 𝐹𝜎 -sets. But, clearly, countable unions
of 𝐹𝜎 -sets are still 𝐹𝜎 .
■■
Problem 18.4 Solution: Fix 𝐴 ⊂ R𝑛 . We have to show that for any 𝑄 ⊂ R𝑛 the equality
#𝑄 = #(𝑄 ∩ 𝐴) + #(𝑄 ⧵ 𝐴)
Case 1: #𝑄 = ∞. Then at least one of the terms #(𝑄 ∩ 𝐴), #(𝑄 ⧵ 𝐴) on the right-hand side must
be infinite, so the equality is clear.
Case 2: #𝑄 < ∞. Then both sets (𝑄 ∩ 𝐴), (𝑄 ⧵ 𝐴) are finite and, as such, they are metrically separ-
0
ated. Therefore we can use the fact that (𝐴) = #(𝐴) is a metric outer measure (Theorem 18.5)
to get equality.
■■
Problem 18.5 Solution: Use Lemma 18.17 to see 0 ⩽ dim 𝐵 ⩽ dim R𝑛 as 𝐵 ⊂ R𝑛 . From
Example 18.18 we know that dim R𝑛 = 𝑛.
If 𝐵 contains an open set 𝑈 (or a set of non-zero Lebesgue measure), we see 𝑛 (𝐵) ⩾ 𝑛 (𝑈 ) > 0;
intersect with a large open ball 𝐾 to make sure that 𝑛 (𝐵 ∩ 𝐾) < ∞ and 𝑈 ∩ 𝐾 ⊂ 𝐵 ∩ 𝐾. This
shows 𝑛 = dim (𝐵 ∩ 𝐾) ⩽ dim (𝐵) ⩽ 𝑛.
■■
224
Solution Manual. Last update 28th October 2021
Problem 18.6 Solution: By self-similarity, we see for the Sierpinski triangle of generation 𝑖, 𝑆 𝑖−1
and its follow-up stage 𝑆 𝑖 = 𝑆1𝑖 ∪ 𝑆2𝑖 ∪ 𝑆3𝑖 that the 𝑆𝑘𝑖 ’s are scaled versions of 𝑆 with a factor 12 .
So,
and dividing by 𝑠 (𝑆 𝑖−1 ) and solving the equality 1 = 3 ⋅ 2−𝑠 ⇐⇒ 2𝑠 = 3 ⇐⇒ 𝑠 = log 3∕ log 2
Koch’s snowflake 𝑆 has in each subsequent generation stage 4 new parts, each scaled by 1∕3, so
and dividing by 𝑠 (𝑆) and solving the equality 1 = 4 ⋅ 3−𝑠 ⇐⇒ 3𝑠 = 4 ⇐⇒ 𝑠 = log 4∕ log 3.
■■
Letting 𝜖 → 0 yields
𝜙 𝜙 𝜙(𝑥) 𝜓 𝜙(𝑥) 𝜓
(𝐴) = lim 𝜖 (𝐴) ⩽ lim sup (𝐴) = lim sup (𝐴) = 0.
𝜖→0 𝜖→0 𝑥⩽𝜖 𝜓(𝑥) 𝑥→0 𝜓(𝑥)
■■
225
19 The Fourier transform.
Solutions to Problems 19.1–19.9
(a) By definition,
1
[−1,1] (𝜉) =
1̂ 1 (𝑥)𝑒−𝑖 𝑥𝜉 𝑑𝑥
2𝜋 ∫ [−1,1]
[ −𝑖 𝑥𝜉 ]1
1 𝑒
= −
2𝜋 𝑖 𝜉 𝑥=−1
1 1 ( 𝑖𝜉 )
= 𝑒 − 𝑒−𝑖 𝜉
2𝜋 𝑖 𝜉
1 sin 𝜉
=
𝜋 𝜉
1 1
[−1,1] (0) =
1̂ 1 (𝑥) 𝑑𝑥 = .
2𝜋 ∫ [−1,1] 𝜋
(b) The convolution theorem, Theorem 19.11, shows that 𝑓̂ ̂ Because of part (a)
∗ 𝑔 = (2𝜋)𝑓̂ ⋅ 𝑔.
we get
( )2 2
1 sin 𝜉 2 sin 𝜉
ℱ (1[−1,1] ∗ 1[−1,1] )(𝜉) = (2𝜋) = .
𝜋 𝜉 𝜋 𝜉2
1 1
= .
2𝜋 1 + 𝑖 𝜉
227
R.L. Schilling: Measures, Integrals & Martingales
Thus,
(e) From (d) and ℱ ◦ℱ 𝑢(𝑥) = (2𝜋)−1 𝑢(−𝑥) (cf. Corollary 19.24) we find
( )
1 (𝑑) 1 1
ℱ (𝜉) = 𝜋 ⋅ ℱ ◦ℱ (𝑒−|⋅| )(𝜉) = 𝑒−|−𝜉| = 𝑒−|𝜉| .
1 + 𝑥2 2 2
The first expression is as in part (a). For the second integral we use integration by parts:
1 [ ]1 1
sin(𝑥𝜉) 1
𝑥 cos(𝑥𝜉) 𝑑𝑥 = 𝑥 − sin(𝑥𝜉) 𝑑𝑥
∫0 𝜉 𝑥=0 𝜉 ∫0
[ ]1
sin(𝜉) 1 cos(𝑥𝜉)
= −
𝜉 𝜉 𝜉 𝑥=0
sin(𝜉) cos(𝜉) 1
= − + 2.
𝜉 𝜉2 𝜉
Thus,
( )
1 sin 𝜉 1 sin 𝜉 cos 𝜉 1 1 1 − cos 𝜉
ℱ (1[−1,1] (1 − | ⋅ |))(𝜉) = − − 2 + 2 = .
𝜋 𝜉 𝜋 𝜉 𝜉 𝜉 𝜋 𝜉2
(g) By definition,
( ∞ 𝑘
) ∞ 𝑘 ∞
∑ 𝑡 −𝑡 1 ∑ 𝑡 −𝑡 1 ∑ 𝑡𝑘 −𝑡 −𝑖 𝑘𝜉
−𝑖 𝑥𝜉
ℱ 𝑒 𝛿𝑘 (𝜉) = 𝑒 𝑒 𝛿𝑘 (𝑑𝑥) = 𝑒 𝑒 .
𝑘=0
𝑘! 2𝜋 ∫ 𝑘=0
𝑘! 2𝜋 𝑘=0 𝑘!
228
Solution Manual. Last update 28th October 2021
|𝑢 + 𝑣|2 = (𝑢 + 𝑣)(𝑢 + 𝑣)
= (𝑢 + 𝑣)(𝑢̄ + 𝑣)
̄
= 𝑢𝑢̄ + 𝑢𝑣̄ + 𝑣𝑢̄ + 𝑣𝑣̄
= |𝑢|2 + 2 Re 𝑢𝑣̄ + |𝑣|2
̂ 𝑣(𝜉) 𝑑𝜉 = (2𝜋)−𝑛
𝑢(𝜉)̂ 𝑢(𝜉)̌
̂ 𝑣(𝜉) 𝑑𝜉
∫ ∫
229
R.L. Schilling: Measures, Integrals & Martingales
19.12 [ ]
= (2𝜋)−𝑛 𝑢(𝑥)ℱ ̌
𝑣 (𝑥) 𝑑𝑥
∫
19.9
= (2𝜋)−𝑛 𝑢(𝑥)𝑣(𝑥) 𝑑𝑥.
∫
■■
= 𝑒−𝑖𝑥𝜉 𝜇(𝑑𝑥)
∫
= 𝑒−𝑖𝑥𝜉 𝜇(𝑑𝑥)
∫
= 𝜒(𝜉).
Therefore, 𝜒 is real-valued. On the other hand, the above calculation shows that
This means that 𝜒 = 𝜒 entails ℱ 𝜇 = ℱ 𝜇̃, and so 𝜇 = 𝜇̃ because of the injectivity of the Fourier
transform.
■■
Problem 19.4 Solution: From linear algebra we know that a symmetric positive definite matrix has
a unique symmetric positive square root, i.e. there is some 𝐵 ∈ R𝑛×𝑛 which is symmetric and
√
positive definite such that 𝐵 2 = 𝐴. Since det(𝐵 2 ) = (det 𝐵)2 , we see that det 𝐵 = det 𝐴 > 0.
Now we change coordinates according to 𝑦 ∶= 𝐵𝑥
If we set
1 ( )
𝑔1∕2 (𝑥) ∶= exp −|𝑥|2 ,
𝜋 𝑛∕2
cf. Example 19.2(iii), then the calculation from above gives
𝜋 𝑛∕2 ( )
ℱ (𝑒−⟨⋅,𝐴⋅⟩ )(𝜉) = √ ℱ (𝑔1∕2 ) 𝐵 −1 𝜉 .
det 𝐴
230
Solution Manual. Last update 28th October 2021
Finally, since 𝐵 −1 = (𝐵 −1 )⊤ ,
we infer that ( )
−⟨⋅,𝐴⋅⟩ 1 1 1 ⟨𝜉, 𝐴−1 𝜉⟩
ℱ (𝑒 )(𝜉) = √ exp − .
det 𝐴 2 (2𝜋)
𝑛∕2 𝑛∕2 4
■■
2 ∕2
𝑢(𝜉)𝑒−𝑡|𝜉|
̂ 𝑑𝜉 = (2𝜋) 𝑢(𝜉)̂
̂ 𝑔𝑡 (𝜉) 𝑑𝜉
∫ ∫
= 𝑢(𝑥)𝑔𝑡 (𝑥) 𝑑𝑥
∫
2 ∕2𝑡
= 𝑢(𝑥)(2𝜋𝑡)−1∕2 𝑒−𝑥 𝑑𝑥
∫
2 ∕2
= (2𝜋)−1∕2 𝑢(𝑡𝑦)𝑒−𝑦 𝑑𝑦
∫
⩽ 𝑐‖𝑢‖∞ .
(In fact, 𝑐 = 1, see Example 14.11). Now let 𝑡 ↑ 0 using monotone convergence and use that, by
assumption, ̂
𝑢 ⩾ 0.
■■
Problem 19.6 Solution: We follow the hint and find using Fubini’s theorem
( )𝑛 1∕𝑅 1∕𝑅
𝑅
2 ⋯ (1 − 𝑒𝑖⟨𝑥,𝜉⟩ ) 𝜇(𝑑𝑥) 𝑑𝜉1 … 𝑑𝜉𝑑
2 ∫−1∕𝑅 ∫−1∕𝑅 ∫R𝑛
( )𝑛 1∕𝑅 1∕𝑅
𝑅
=2 ⋯ (1 − 𝑒𝑖⟨𝑥,𝜉⟩ ) 𝑑𝜉1 … 𝑑𝜉𝑑 𝜇(𝑑𝑥)
∫R𝑛 2 ∫−1∕𝑅 ∫−1∕𝑅
1∕𝑅 1∕𝑅
𝑅 𝑅
=2 … (1 − 𝑒𝑖⟨𝑥,𝜉⟩ ) 𝑑𝜉1 … 𝑑𝜉𝑑 𝜇(𝑑𝑥)
∫R𝑛 2 ∫−1∕𝑅 2 ∫−1∕𝑅
( )
1∕𝑅 1∕𝑅
𝑅 𝑅
=2 1− … 𝑒𝑖⟨𝑥,𝜉⟩
𝑑𝜉1 … 𝑑𝜉𝑑 𝜇(𝑑𝑥)
∫R𝑛 2 ∫−1∕𝑅 2 ∫−1∕𝑅
( 𝑛
)
∏ 1∕𝑅
𝑅
=2 1− 𝑒𝑖𝑥𝑛 𝜉𝑛 𝑑𝜉𝑛 𝜇(𝑑𝑥)
∫R𝑛 2 ∫−1∕𝑅 𝑛=1
231
R.L. Schilling: Measures, Integrals & Martingales
( 𝑛 [ ]𝜉 =1∕𝑅 )
∏ 𝑅 𝑒𝑖𝑥𝑛 𝜉𝑛 𝑛
=2 1− 𝜇(𝑑𝑥)
∫R𝑛
𝑛=1
2 𝑖𝑥𝑛 𝜉𝑛 =−1∕𝑅
( 𝑛
)
∏ 𝑒𝑖𝑥𝑛 ∕𝑅 − 𝑒−𝑖𝑥𝑛 ∕𝑅
=2 1− 𝜇(𝑑𝑥)
∫R𝑛
𝑛=1
2𝑖𝑥𝑛 ∕𝑅
( 𝑛
)
∏ sin(𝑥𝑛 ∕𝑅)
=2 1− 𝜇(𝑑𝑥)
∫R𝑛
𝑛=1
𝑥𝑛 ∕𝑅
( 𝑛
)
∏ sin(𝑥𝑛 ∕𝑅)
⩾2 1− 𝜇(𝑑𝑥).
∫R𝑛 ⧵[−2𝑅,2𝑅]𝑛
𝑛=1
𝑥𝑛 ∕𝑅
In the last step we use that the integrand is positive since | sin 𝑦∕𝑦| ⩽ 1. Observe that
and so
𝑛
∏ sin(𝑥𝑛 ∕𝑅) 1
⩽
𝑛=1
𝑥𝑛 ∕𝑅 2
hence
( )𝑛 1∕𝑅 1∕𝑅
𝑅
2 ⋯ (1 − 𝑒𝑖⟨𝑥,𝜉⟩ ) 𝜇(𝑑𝑥) 𝑑𝜉1 … 𝑑𝜉𝑑
2 ∫−1∕𝑅 ∫−1∕𝑅 ∫R𝑛
( 𝑛
)
∏ sin(𝑥𝑛 ∕𝑅)
⩾2 1− 𝜇(𝑑𝑥)
∫R𝑛 ⧵[−2𝑅,2𝑅]𝑛
𝑛=1
𝑥𝑛 ∕𝑅
( )
1
⩾2 1− 𝜇(𝑑𝑥)
∫R𝑛 ⧵[−2𝑅,2𝑅]𝑛 2
⩾ 𝜇(𝑑𝑥).
∫R𝑛 ⧵[−2𝑅,2𝑅]𝑛
Remark. A similar inequality exists for the Fourier transform (instead of the inverse Fourier
transform). This has the form
( )
𝜇 (R𝑛 ⧵ [−2𝑅, 2𝑅]𝑛 ) ⩽ 2(𝜋𝑅)𝑛 𝜇̂(0) − Re 𝜇̂(𝜉) 𝑑𝜉.
∫[−1∕𝑅,1∕𝑅]𝑛
■■
232
Solution Manual. Last update 28th October 2021
|∑ |2
1 | 𝑛 |
= | 𝜆𝑗 𝑒 −𝑖 𝑥𝜉𝑗 | 𝑑𝜇(𝑥) ⩾ 0.
|
(2𝜋) ∫ | 𝑗=1
𝑛 |
|
| |
Note that this already implies that 𝜙(−𝜉) = 𝜙(𝜉). The argument is as follows: If we
∑
have for a matrix (𝑎𝑗𝑘 ) that 𝑗𝑘 𝑎𝑗𝑘 𝜆𝑗 𝜆̄ 𝑗 ⩾ 0, then
∑ ∑ ∑ ∑
0⩽ 𝑎𝑗𝑘 𝜆𝑗 𝜆̄ 𝑘 = 𝑎𝑗𝑘 𝜆𝑗 𝜆̄ 𝑘 = 𝑎𝑗𝑘 𝜆̄ 𝑗 𝜆𝑘 = 𝑎𝑘𝑗 𝜆̄ 𝑘 𝜆𝑗
𝑗𝑘 𝑗𝑘 𝑗𝑘 𝑘𝑗
which means that 𝑎𝑗𝑘 = 𝑎𝑘𝑗 . Apply this to the matrix 𝑎𝑗𝑘 = 𝜙(𝜉𝑗 − 𝜉𝑘 ) with 𝑚 = 2 and
𝜉1 = 𝜉 and 𝜉2 = 0 to infer that 𝜙(𝜉) = 𝜙(−𝜉).
(ii) We want to use the differentiability lemma for parameter-dependent integrals. For this
we define
1 −𝑖 𝑥𝜉
𝑢(𝜉, 𝑥) ∶= 𝑒 .
(2𝜋)𝑛
Since 𝜇 is a finite measure and |𝑢(𝑥, 𝜉)| ⩽ (2𝜋)−𝑛 , we find 𝑢(𝜉, ⋅) ∈ 𝐿1 (𝜇). Moreover,
1
𝜕𝜉𝑗 𝜙(𝜉) = 𝜕𝜉𝑗 𝑢(𝜉, 𝑥) 𝜇(𝑑𝑥) = (−𝑖𝑥𝑗 )𝑒−𝑖 𝑥𝜉 𝜇(𝑑𝑥).
∫ (2𝜋)𝑛 ∫
Iterating this argument, we see that 𝜕 𝛼 𝜙 exists for any 𝛼 ∈ N𝑛0 such that |𝛼| ⩽ 𝑚.
(iii) We follow the hint and consider first the case 𝑑 = 1 and 𝑛 = 1. We can rewrite the
expression 𝜙(2ℎ) − 2𝜙(0) + 𝜙(−2ℎ) using Fourier transforms:
1
𝜙(2ℎ) − 2𝜙(0) + 𝜙(−2ℎ) = (𝑒−𝑖 2ℎ𝑥 − 2 + 𝑒𝑖 2ℎ𝑥 ) 𝜇(𝑑𝑥)
2𝜋 ∫
1
= (cos(2ℎ𝑥) − 1)𝜇(𝑑𝑥).
𝜋∫
L’Hospital’s theorem applies and gives
1 − cos(2𝑦) 𝑦→0 1
← .
←←←←←←←←→
4𝑦2 2
1 1 − cos(2ℎ𝑥)
𝑥2 𝜇(𝑑𝑥) = 𝑥2 lim 𝜇(𝑑𝑥)
∫ 2 ∫ ℎ→0 4(ℎ𝑥)2
1
⩽ lim inf 2 (1 − cos(2ℎ𝑥)) 𝜇(𝑑𝑥)
ℎ→0 4ℎ ∫
1 (
= −𝜋 lim inf 2 𝜙(2ℎ) − 2𝜙(0) + 𝜙(−2ℎ))
ℎ→0 4ℎ
233
R.L. Schilling: Measures, Integrals & Martingales
If 𝑛 ⩾ 1, we use induction. Assume that 𝜙 ∈ 𝐶 2𝑛 (R) and that the assertion has been
proved for 𝑛−1. Since 𝜙 ∈ 𝐶 2𝑛 (R) ⇒ 𝜙 ∈ 𝐶 2(𝑛−1) , we see by the induction assumption
that ∫ |𝑥|2(𝑛−1) 𝑑𝜇(𝑥) < ∞. Thus, 𝜈(𝑑𝑥) ∶= 𝑥2(𝑛−1) 𝜇(𝑑𝑥) is a measure and
1
𝜈̂(𝜉) = 𝑥2(𝑛−1) 𝑒−𝑖 𝑥𝜉 𝑑𝜇(𝑥)
2𝜋 ∫
1 1 𝑑 2(𝑛−1)
= 𝑒−𝑖 𝑥𝜉 𝑑𝜇(𝑥).
2𝜋 (−𝑖)2(𝑛−1) 𝑑𝜉 2(𝑛−1) ∫
Consequently, we see that 𝜈̂ ∈ 𝐶 2 (R). The first part of the proof (𝑛 = 1) gives
(iv) Assume that 𝑧 ∈ C𝑛 . If 𝐾 ∶= supp 𝜇 is compact, then we get, because of the continuity
of 𝑒−𝑖 𝑧𝑥 , that 𝑀 ∶= sup𝑥∈𝐾 |𝑒−𝑖 𝑧𝑥 | < ∞. From
𝑢 𝑑𝜇 = 𝑢 𝑑𝜇 for any 𝑢 ⩾ 0
∫ ∫supp 𝜇
we conclude that
|𝑒−𝑖 𝑧𝑥 | 𝑑𝜇(𝑥) ⩽ 𝑀𝜇(R𝑛 ) < ∞,
∫
i.e.
1
𝜙(𝑧) = 𝑒−𝑖 𝑧𝑥 𝑑𝜇(𝑥)
(2𝜋)𝑛 ∫
is well-defined. Setting
𝑛
1 ∑ (−𝑖 𝑧𝑥)𝑘
𝑢𝑛 (𝑥) ∶= , 𝑥 ∈ R𝑛 ,
(2𝜋)𝑛 𝑘=0 𝑘!
we get
𝑛
1 ∑ |𝑧𝑥|𝑘 1 |𝑧𝑥| 1
|𝑢𝑛 (𝑥)| ⩽ ⩽ 𝑒 ⩽ sup 𝑒|𝑧𝑥| < ∞.
(2𝜋)𝑛 𝑘=0 𝑘! (2𝜋)𝑛 (2𝜋)𝑛 𝑥∈𝐾
Since 𝜇 is a finite measure, we can use the dominated convergence theorem to get
■■
234
Solution Manual. Last update 28th October 2021
∫𝐵 𝑒𝑖𝑥∕𝑛 𝑑𝑥 = 0 that 1𝐵 𝑒𝑖⋅∕𝑛 ∈ 1 (𝑑𝑥). As |𝑒𝑖 𝑥∕𝑛 | = 1, we get 𝜆1 (𝐵) < ∞. By dominated
convergence
0 = lim 𝑒𝑖 𝑥∕𝑛 𝑑𝑥 = lim 𝑒𝑖 𝑥∕𝑛 𝑑𝑥 = 𝜆1 (𝐵).
𝑛→∞ ∫𝐵 ∫𝐵 𝑛→∞
⏟⏞⏞⏟⏞⏞⏟
1
Alternative solution: Set 𝑓 (𝑥) ∶= 1𝐵 (𝑥); by assumption, 𝑓̂(1∕𝑛) = 0. Since the Fourier trans-
form is continous, cf. 19.3, we get
( )
1
𝑓̂(0) = lim 𝑓̂ = 0.
𝑛→∞ 𝑛
On the other hand, 𝑓̂(0) = (2𝜋)−1 𝜆1 (𝐵).
■■
with 𝑝𝑗 ∶= 𝜇( 2𝜋
𝜉
𝑗). From the definition of the Fourier transform we get
1
𝜇(𝜂)
̂ = 𝑒−𝑖 𝑥𝜂 𝜇(𝑑𝑥)
2𝜋 ∫
[ ( ) ]
1 ∑ 2𝜋
= 𝑝 exp −𝑖 𝑗 𝜂
2𝜋 𝑗∈Z 𝑗 𝜉
⇒: From 𝜇(𝜉)
̂ ̂ we conclude
= 𝜇(0)
2𝜋(𝜇(0)
̂ − 𝜇(𝜉))
̂ = (1 − 𝑒−𝑖 𝑥𝜉 ) 𝜇(𝑑𝑥) = 0.
∫
In particular, ∫ (1 − 𝑒−𝑖 𝑥𝜉 ) 𝜇(𝑑𝑥) ∈ R, i.e.
𝜇 {𝑥 ∈ R; 1 − cos(𝑥𝜉) > 0} = 0.
Consequently, ) (
2𝜋
0 = 𝜇 {𝑥 ∈ R; cos(𝑥𝜉) ≠ 1} = 𝜇 R ⧵ Z .
𝜉
235
R.L. Schilling: Measures, Integrals & Martingales
(ii) Because of |̂
𝜇(𝜉1 )| = 𝜇̂(0) there is some 𝑧1 ∈ R such that
𝜇̂(𝜉1 ) = 𝜇̂(0)𝑒𝑖 𝑧1 𝜉1 .
Therefore,
1
𝑒−𝑖 𝜉1 (𝑥+𝑧1 ) 𝜇(𝑑𝑥) = 𝜇(0).
̂
2𝜋 ∫
Observe that the left-hand side is just the Fourier transform of the measure 𝜈(𝐵) ∶= 𝜇(𝐵 −
𝑧1 ), 𝐵 ∈ ℬ(R), and so
̂ 1 ) = 𝜇(0)
𝜈(𝜉 ̂ = 𝜈(0).
̂
Setting ( ) ( )
2𝜋 2𝜋
𝐴 ∶= 𝑧1 + Z ∩ 𝑧2 + Z
𝜉1 𝜉2
we see that 𝜇(R ⧵ 𝐴) = 0. Let us show that 𝐴 contains at most one element: Assume, on
the contrary, that there are two distinct points in 𝐴, then there are 𝑛, 𝑛′ ∈ Z and 𝑚, 𝑚′ ∈ Z
such that
2𝜋 2𝜋 ′
𝑧1 + 𝑛 = 𝑧2 + 𝑛,
𝜉1 𝜉2
2𝜋 2𝜋 ′
𝑧1 + 𝑚 = 𝑧2 + 𝑚.
𝜉1 𝜉2
2𝜋 2𝜋 ′
(𝑛 − 𝑚) = (𝑛 − 𝑚′ )
𝜉1 𝜉2
𝜉 𝑛′ − 𝑚′
⇒ 2 = ∈ Q.
𝜉1 𝑛−𝑚
■■
236
20 The Radon–Nikodým theorem.
Solutions to Problems 20.1–20.9
Using the Radon–Nikodým theorem we conclude that there exists a measurable function 𝑓 ∈
+ (𝒜 ) such that 𝜈 = 𝑓 ⋅ 𝜇. Assume that 𝑓 > 1 on a set of positive 𝜇-measure. Without loss
of generality we may assume that the set has finite measure, otherwise we would consider the
intersection 𝐴𝑘 ∩ {𝑓 > 1} with some exhausting sequence 𝐴𝑘 ↑ 𝑋 and 𝜇(𝐴𝑘 ) < ∞.
Then, for sufficiently small 𝜖 > 0 we know that 𝜇({𝑓 ⩾ 1 + 𝜖}) > 0 and so
𝜈({𝑓 ⩾ 1 + 𝜖}) = 𝑓 𝑑𝜇
∫{𝑓 ⩾1+𝜖}
⩾ (1 + 𝜖) 𝑑𝜇
∫{𝑓 ⩾1+𝜖}
⩾ (1 + 𝜖)𝜇({𝑓 ⩾ 1 + 𝜖})
⩾ 𝜇({𝑓 ⩾ 1 + 𝜖})
which is impossible.
■■
Problem 20.2 Solution: Because of our assumption both 𝜇 ≪ 𝜈 and 𝜈 ≪ 𝜇 which means that we
know
𝜈 = 𝑓𝜇 and 𝜇 = 𝑔𝜈
𝜈 = 𝑓 𝜇 = 𝑓 ⋅ 𝑔𝜈
237
R.L. Schilling: Measures, Integrals & Martingales
Problem 20.3 Solution: Take Lebesgue measure 𝜆 ∶= 𝜆1 on (R, ℬ(R)) and the function 𝑓 (𝑥) ∶=
𝑥 + ∞ ⋅ 1[0,1]𝑐 (𝑥). Then 𝑓 ⋅ 𝜆 is certainly not 𝜎-finite.
■■
■■
■■
Problem 20.6 Solution: (i) If 𝐹 is AC, continuity is trivial, just take 𝑁 = 2 in the very definition of
AC functions.
To see that 𝐹 is also BV, we take 𝜖 = 1 and choose 𝛿 > 0 such that for any subcollection 𝑎 ⩽
∑ ∑
𝑥1 < 𝑦1 < ⋯ < 𝑥𝑁 < 𝑦𝑁 ⩽ 𝑏 with 𝑛 (𝑦𝑛 − 𝑥𝑛 ) < 𝛿 we have 𝑛 |𝐹 (𝑦𝑛 ) − 𝐹 (𝑥𝑛 )| < 1. Let
𝑀 = [(𝑏−𝑎)∕𝛿]+1 and 𝑎𝑖 = 𝑎+𝑖(𝑏−𝑎)∕𝑀 for 𝑖 = 0, 1, … , 𝑀. Clearly, 𝑎𝑖 −𝑎𝑖−1 = (𝑏−𝑎)∕𝑀 < 𝛿
and, in particular, 𝑉 (𝑓 , [𝑎𝑖−1 , 𝑎𝑖 ]) < 1 for all 𝑖 = 0, 1, … 𝑀. Thus,
𝑀
∑
𝑉 (𝑓 ; [𝑎, 𝑏]) ⩽ 𝑉 (𝑓 , [𝑎𝑖−1 , 𝑎𝑖 ]) < 𝑀.
𝑖=1
(ii) Following the hint, we see that 𝑓 is increasing. Define 𝑔 ∶= 𝐹 − 𝑓 . We have to show that 𝑔 is
increasing. Let 𝑥 < 𝑦. Obviously,
𝑉 (𝑓 ; [𝑎, 𝑥]) + 𝐹 (𝑦) − 𝐹 (𝑥) ⩽ 𝑉 (𝑓 ; [𝑎, 𝑥]) + |𝐹 (𝑦) − 𝐹 (𝑥)| ⩽ 𝑉 (𝑓 ; [𝑎, 𝑦])
(since the points 𝑥 < 𝑦 can be added to extend any partition of [𝑎, 𝑥] to give a partition of [𝑎, 𝑦]).
This gives 𝑔(𝑥) ⩽ 𝑔(𝑦).
𝜖
|𝑓 | 𝑑𝜆 < .
∫{|𝑓 |>𝑅} 2
This is possible since 𝑓 is integrable: use, e.g. monotone convergence. Now pick 𝑥1 < 𝑦1 < 𝑥2 <
∑
𝑦2 < ⋯ < 𝑥𝑁 < 𝑦𝑁 with 𝑁 𝑛=1 |𝑦𝑛 − 𝑥𝑛 | < 𝛿 where 𝛿 = 𝛿(𝜖) ∶= 𝜖∕(2𝑅) with the 𝑅 we’ve just
chosen. Then
238
Solution Manual. Last update 28th October 2021
(iv) Write 𝐹 = 𝑓1 − 𝑓2 with 𝑓𝑖 increasing (see part (ii)). From (ii) we know that we can pick
𝑓1 (𝑥) = 𝑉 (𝐹 , [𝑎, 𝑥]). Since 𝐹 is absolutely continuous, so is 𝑓1 , hence 𝑓2 . This follows from the
observation that
Since the 𝑓𝑖 are continuous, the set-functions 𝜇𝑖 [𝑎, 𝑥) ∶= 𝑓𝑖 (𝑥)−𝑓𝑖 (𝑎) are pre-measures and extend
to measures on the Borel 𝜎-algebra – see also Problem 6.1.
Now let 𝑁 be a Lebesgue null-set. For every 𝛿 > 0 we can cover 𝑁 by finitely many intervals
∑
[𝑥𝑖 , 𝑦𝑖 ] such that 𝑖 (𝑦𝑖 − 𝑥𝑖 ) < 𝛿. Without loss of generality we can make the intervals non-
overlapping and their length is still < 𝛿. Since the 𝑓𝑖 are AC, we find for every 𝜖 some 𝛿 such
that
∑
|𝑓𝑖 (𝑦𝑛 ) − 𝑓𝑖 (𝑥𝑛 )| < 𝜖.
𝑛
In particular,
∑
𝜇𝑖 (𝑁) ⩽ 𝜇𝑖 ([𝑥𝑛 , 𝑦𝑛 ]) < 𝜖,
𝑛
which shows that the Lebesgue null-set is also a 𝜇𝑖 -null set, i.e. 𝜇𝑖 ≪ 𝜆 and therefore the claim
follows from the Radon–Nikodým theorem.
■■
Problem 20.7 Solution: This problem is somewhat ill-posed. We should first embed it into a suitable
context, say, on the measurable space (R, ℬ(R)). Denote by 𝜆 = 𝜆1 one-dimensional Lebesgue
measure. Then
1[1,2] 𝜈 ≪ 𝜇
while
1(2,3] 𝜈⊥𝜇.
It is interesting to note how ‘big’ the null-set of ambiguity for the Lebesgue decomposition is—it
is actually R ⧵ [0, 3] a, from a Lebesgue (i.e. 𝜆) point of view, huge and infinite set, but from a
𝜇-𝜈-perspective a negligible, namely null, set.
■■
239
R.L. Schilling: Measures, Integrals & Martingales
Problem 20.8 Solution: Since we deal with a bounded measure we can use 𝐹 (𝑥) ∶= 𝜇(−∞, 𝑥) rather
than the more cumbersome definition for 𝐹 employed in Problem 6.1 (which is good for locally
finite measures!).
𝜇 = 𝜇◦ + 𝜇⊥ where 𝜇◦ ≪ 𝜆, 𝜇⊥ ⊥𝜆.
Now define 𝜇2 ∶= 𝜇◦ and 𝐹2 ∶= 𝜇 ◦ (−∞, 𝑥). We have to prove property (2). For this we observe
that 𝜇 ◦ is a finite measure (since 𝜇 ◦ ⩽ 𝜇 and that, therefore, 𝜇 ◦ = 𝑓 ⋅ 𝜆 with a function 𝑓 ∈ 𝐿1 (𝜆).
Thus, for every 𝑅 > 0
= 𝑓 (𝑡) 𝜆(𝑑𝑡)
∫(𝑥𝑗 ,𝑦𝑗 )
to confirm that
𝑁
∑
|𝐹2 (𝑦𝑗 ) − 𝐹2 (𝑥𝑗 )| ⩽ 𝜖
𝑗=1
Now consider the measure 𝜇 ⊥ . Its distribution function 𝐹 ⊥ (𝑥) ∶= 𝜇 ⊥ (−∞, 𝑥) is increasing, left-
continuous but not necessarily continuous. Such a function has, by Lemma 14.14 at most countably
many discontinuities (jumps), which we denote by 𝐽 . Thus, we can write
𝜇⊥ = 𝜇1 + 𝜇3
240
Solution Manual. Last update 28th October 2021
𝜇1 is clearly a measure (the sum being countable) with 𝜇1 ⩽ 𝜇 ⊥ and so is, therefore, 𝜇2 (since the
defining difference is always positive). The corresponding distribution functions are
∑
𝐹1 (𝑥) ∶= Δ𝐹 (𝑦)
𝑦∈𝐽 ,𝑦<𝑥
It is clear that 𝐹2 is increasing and, more importantly, continuous so that the problem is solved.
It is interesting to note that our problem shows that we can decompose every left- or right-continuous
monotone function into an absolutely continuous and singular part and the singular part again into
a continuous and discontinuous part:
where
𝑔sd —is a monotone discontinuous (even: pure jump), but nevertheless left- or right-continuous,
and singular function.
■■
(i) In the following picture 𝐹1 is represented by a black line, 𝐹2 by a grey line and 𝐹3 is a dotted
black line.
1
|𝐹𝑛 (𝑥) − 𝐹𝑛+1 (𝑥)| ⩽
2𝑛+1
since we modify 𝐹𝑛 only on a set 𝐼𝑛+1
𝓁
by replacing a diagonal line by a combination of
diagonal-flat-diagonal and all this happens only within a range of 2−𝑛 units. Since the flat bit
is in the middle, we get that the maximal deviation between 𝐹𝑛 and 𝐹𝑛+1 is at most 1
2
⋅ 2−𝑛 .
Just look at the pictures!
241
R.L. Schilling: Measures, Integrals & Martingales
(iv) Let 𝐶 denote the Cantor set. Then for 𝑥 ∈ [0, 1] ⧵ 𝐶 we find 𝑛 and 𝓁 such that 𝑥 ∈ 𝐼𝑛𝓁 (which
is an open set!) and, since on those pieces 𝐹𝑛 and 𝐹 do not differ any more
where we use that 𝐹𝑛 |𝐼𝑛𝓁 is constant. Since 𝜆(𝐶) = 0 (see Problem 7.12) we have 𝜆([0, 1] ⧵
𝐶) = 1 so that 𝐹 ′ exists a.e. and satisfies 𝐹 ′ = 0 a.e.
(v) We have 𝐼𝑛𝓁 = (𝑎𝓁 , 𝑏𝓁 ) (we suppress the dependence of 𝑎𝓁 , 𝑏𝓁 on 𝑛 with, because of our
ordering of the middle-thirds sets (see the problem):
and
2𝑛 −1
∑[ ]
𝐹 (𝑏𝓁 ) − 𝐹 (𝑎𝓁 ) = 𝐹 (𝑏2𝑛 −1 ) − 𝐹 (𝑎1 ) ←←←←←←←←←→
← 𝐹 (1) − 𝐹 (0) = 1
𝑛→∞
𝓁=1
This leads to a contradiction since, because of the first equality, the sum
2𝑛 −1
∑[ ]
𝐹 (𝑎𝓁 ) − 𝐹 (𝑏𝓁−1 )
𝓁=1
■■
242
Solution Manual. Last update 28th October 2021
1✻
3
4
1
2
1
4
✲
1 2
3 3
1
243
21 Riesz representation theorems.
Solutions to Problems 21.1–21.7
‖𝑓 ⋅ 𝑔‖1 ⩽ ‖𝑓 ‖𝑝 ‖𝑔‖𝑞 ⩽ ‖𝑓 ‖𝑝 .
Therefore { }
‖𝑓 ‖𝑝 ⩾ sup 𝑓 𝑔 𝑑𝜇 ∶ 𝑔 ∈ 𝐿 (𝜇), ‖𝑔‖𝑞 ⩽ 1
𝑞
.
∫
For the converse inequality ‘⩽’ we use 𝑔 ∶= sgn(𝑓 ) ⋅ |𝑓 |𝑝−1 . Since 𝑞 = 𝑝
𝑝−1
, we have
𝑓 𝑔 𝑑𝜇 ⩾ ‖𝑓 ‖𝑝 − 𝜖.
∫
Since is dense, there is some ℎ ∈ with ‖𝑔 − ℎ‖𝑞 ⩽ 𝜖. The Hölder inequality now
shows
𝑓 ℎ 𝑑𝜇 = 𝑓 (ℎ − 𝑔) 𝑑𝜇 + 𝑓 𝑔 𝑑𝜇
∫ ∫ ∫
⩾ −‖𝑓 ‖𝑝 ‖ℎ − 𝑔‖𝑞 + 𝑓 𝑔 𝑑𝜇
∫
⩾ −‖𝑓 ‖𝑝 𝜖 + 𝑓 𝑔 𝑑𝜇
∫
⩾ −‖𝑓 ‖𝑝 𝜖 + ‖𝑓 ‖𝑝 − 𝜖
= ‖𝑓 ‖𝑝 (1 − 𝜖) − 𝜖.
245
R.L. Schilling: Measures, Integrals & Martingales
(iii) If 𝑓 𝑔 ∈ 𝐿1 (𝜇) for all 𝑔 ∈ 𝐿𝑞 (𝜇), then 𝐼𝑓 (𝑔) ∶= ∫ |𝑓 |𝑔 𝑑𝜇 is a positive linear functional
on 𝐿𝑞 (𝜇). From Theorem 21.5 we know that there exists a unique 𝑓̃ ∈ 𝐿𝑞 (𝜇) such that
Therefore, 𝑓 = 𝑓̃ ∈ 𝐿𝑞 (𝜇).
■■
(i) We use a classical diagonal argument (as in the proof of Theorem 21.18). Let (𝑔𝑛 )𝑛∈N
denote an enumeration of 𝑞 . Hölder’s inequality (13.5) tells us
( )
| |
| 𝑢 𝑔 𝑑𝜇 | ⩽ ‖𝑢 ‖ ‖𝑔 ‖ ⩽ sup ‖𝑢𝑛 ‖𝑝 ‖𝑔𝑖 ‖𝑞
|∫ 𝑛 𝑖 | 𝑛 𝑝 𝑖 𝑞
| | 𝑛∈N
lim 𝑢𝑖+1
𝑛 𝑔𝑖+1 𝑑𝜇
𝑛→∞ ∫
lim 𝑢𝑖𝑛 𝑔𝑘 𝑑𝜇
𝑛→∞ ∫
exists for all 𝑘 = 1, 2, … , 𝑖. Thus, for the diagonal sequence 𝑣𝑛 ∶= 𝑢𝑛𝑛 the limits
lim𝑛→∞ ∫ 𝑣𝑛 𝑔𝑖 𝑑𝜇 exist for each 𝑖 ∈ N.
(ii) Let 𝑔 ∈ 𝐿𝑞 (𝜇) and (𝑢𝑛(𝑖) )𝑖∈N be the diagonal sequence constructed in (i). Since R is
( )
complete, it is enough to show that ∫ 𝑢𝑛(𝑖) 𝑔 𝑑𝜇 𝑖∈N is a Cauchy sequence. Fix 𝜖 > 0.
By assumption, 𝑞 is dense in 𝐿𝑞 (𝜇), i.e. there exists some ℎ ∈ 𝑞 such that ‖𝑔 −ℎ‖𝑞 ⩽
𝜖. Part (i) shows that we can take 𝑁 ∈ N with
| |
| 𝑢 ℎ 𝑑𝜇 −
|∫ 𝑛(𝑖) 𝑢𝑛(𝑘) ℎ 𝑑𝜇|| ⩽ 𝜖 ∀𝑖, 𝑘 ⩾ 𝑁. (⋆)
| ∫ |
Hölder’s inequality and the triangle inequality show
| |
| 𝑢𝑛(𝑖) 𝑔 𝑑𝜇 − 𝑢𝑛(𝑘) 𝑔 𝑑𝜇||
|∫ ∫
| |
| |
= || (𝑢𝑛(𝑖) − 𝑢𝑛(𝑘) )(𝑔 − ℎ) 𝑑𝜇 + (𝑢𝑛(𝑖) − 𝑢𝑛(𝑘) )ℎ 𝑑𝜇||
|∫ ∫ |
| | | |
⩽ || (𝑢𝑛(𝑖) − 𝑢𝑛(𝑘) )(𝑔 − ℎ) 𝑑𝜇 || + || (𝑢𝑛(𝑖) − 𝑢𝑛(𝑘) )ℎ 𝑑𝜇||
|∫ | |∫ |
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏟
⩽𝜖 b/o (⋆)
246
Solution Manual. Last update 28th October 2021
𝐼(𝑔) ∶= lim 𝑢+
𝑛(𝑖)
𝑔 𝑑𝜇, and 𝐽 (𝑔) ∶= lim 𝑢−
𝑛(𝑖)
𝑔 𝑑𝜇
𝑖→∞ ∫ 𝑖→∞ ∫
exist for all 𝑔 ∈ 𝐿𝑞 (𝜇). Indeed: From (i),(ii) we see that there is a subsequence such
that 𝐼(𝑔) exists for all 𝑔 ∈ 𝐿𝑞 (𝜇). Thinning out this subsequence once again, we see
that 𝐽 (𝑔) exists for all 𝑔 ∈ 𝐿𝑞 (𝜇). Since 𝐼 and 𝐽 are positive linear functionals on
𝐿𝑞 (𝜇), Theorem 21.5 proves that there are unique functions 𝑣, 𝑤 ∈ 𝐿𝑞 (𝜇), 𝑣, 𝑤 ⩾ 0
representing these functionals:
= (𝑣 − 𝑤)𝑔 𝑑𝜇.
∫
The claim follows if we use 𝑢 ∶= 𝑣 − 𝑤 ∈ 𝐿𝑞 (𝜇).
■■
Since 𝜇̂𝑖 (−𝜉) = 𝜇̂𝑖 (𝜉), this also holds for the limit
This shows that 𝜙 is positive semidefinite. If 𝑚 = 1 resp. 𝑚 = 2, we see that the matrices
( )
( ) 𝜙(0) 𝜙(−𝜉)
𝜙(0) and
𝜙(𝜉) 𝜙(0)
247
R.L. Schilling: Measures, Integrals & Martingales
are positve hermitian for all 𝜉 ∈ R𝑛 . Since determinants of positive hermitian matrices
are positive, we find 𝜙(0) ⩾ 0 and
(ii) First of all we show that the limit exists. Pick 𝑢 ∈ 𝐶𝑐∞ (R𝑛 ). Because of Theorem 19.23,
−1 𝑢 ∈ (R𝑛 ) and we can use Plancherel’s theorem (Theorem 19.12), to get
Since |𝜇̂𝑖 (𝜉)| ⩽ 𝜇̂𝑖 (0) → 𝜙(0) is uniformly bounded, we can use dominated convergence
and find that
Λ(𝑢) ∶= lim 𝑢 𝑑𝜇𝑖 = −1 𝑢(𝜉)𝜙(𝜉) 𝑑𝜉
𝑖→∞ ∫ ∫
is well-defined. The linearity of Λ follows from the linearity of the integral Moreover, if
𝑢 ⩾ 0, then
Λ𝑢 = lim 𝑢 𝑑𝜇𝑖 ⩾ 0.
𝑖→∞ ∫
|Λ𝑢| ⩽ lim sup |𝑢| 𝑑𝜇𝑖 ⩽ ‖𝑢‖∞ lim sup 𝜇𝑖 (R𝑛 ) = (2𝜋)𝑛 𝜙(0)‖𝑢‖∞ .
𝑖→∞ ∫ 𝑖→∞ ⏟⏟⏟
(2𝜋)𝑛 𝜇̂𝑖 (0)
Since 𝐶𝑐∞ (R𝑛 ) is uniformly dense in 𝐶𝑐 (R𝑛 ), (see Problem 15.13, the proof resembles the
argument of Theorem 15.11), we can extend Λ to a positive linear functional on 𝐶𝑐 (R𝑛 ):
For 𝑢 ∈ 𝐶𝑐 (R𝑛 ) we take (𝑢𝑖 )𝑖∈N ⊂ 𝐶𝑐∞ (R𝑛 ), such that ‖𝑢𝑖 − 𝑢‖∞ → 0. Since
(note that 𝜇̌𝑖 (𝜉) = (2𝜋)𝑛 𝜇̂𝑖 (−𝜉)). With the dominated convergence theorem we get
248
Solution Manual. Last update 28th October 2021
⩽ 2(2𝜋)𝑛 𝜖
for 𝑅 ⩾ 𝛿1 . In particular we find for 𝑖 ⩾ 𝑛0 (𝜖), 𝜇𝑖 (R𝑛 ⧵ [−𝑅, 𝑅]𝑛 ) ⩽ 3(2𝜋)𝑛 𝜖. In order to
get 𝜇𝑖 (R𝑛 ⧵ [−𝑅, 𝑅]𝑛 ) ⩽ 3(2𝜋)𝑛 𝜖 for 𝑖 = 1, … , 𝑛0 (𝜖), we can increase 𝑅, if needed.
(v) Let (𝜒𝑘 )𝑘∈N ⊂ 𝐶𝑐 (R𝑛 ) be a sequence of functions such that 0 ⩽ 𝜒𝑘 ⩽ 1 and 𝜒𝑘 ↑ 1R𝑛
(use, e.g. Urysohn functions, cf. page 239, or construct the 𝜒𝑘 directly). Because of (iii)
we have
𝜒 𝑑𝜇 = Λ(𝜒𝑘 ) ⩽ (2𝜋)𝑛 𝜙(0).
∫ 𝑘
The monotone convergence theorem shows that 𝜇 is a finite measure:
Moreover, 𝑀 ∶= sup𝑖∈N 𝜇𝑖 (R𝑛 ) < ∞ since 𝜇𝑖 (R𝑛 ) = (2𝜋)𝑛 𝜇̂𝑖 (0) → 𝜙(0). It remains to
show that 𝜇𝑖 converges weakly to 𝜇. First of all,
𝑢 𝑑𝜇𝑖 ←←←←←←←←←→
← 𝑢 𝑑𝜇 ∀𝑢 ∈ 𝐶𝑐 (R𝑛 ). (⋆)
∫ 𝑖→∞ ∫
Let 𝑢 ∈ 𝐶𝑐 (R𝑛 ). Since 𝐶𝑐∞ (R𝑛 ) is dense in 𝐶𝑐 (R𝑛 ), there is a sequence (𝑓𝑘 )𝑘∈N ⊂
𝐶𝑐∞ (R𝑛 ) such that ‖𝑓𝑘 − 𝑢‖∞ → 0. Thus,
| |
| 𝑢 𝑑𝜇𝑖 − 𝑢 𝑑𝜇||
|∫ ∫
| |
| | | | | |
⩽ || (𝑢 − 𝑓𝑘 ) 𝑑𝜇𝑖 || + || 𝑓𝑘 𝑑𝜇𝑖 − 𝑓𝑘 𝑑𝜇|| + || (𝑓𝑘 − 𝑢) 𝑑𝜇 ||
|∫ | |∫ ∫ | |∫ |
| |
⩽ ‖𝑢 − 𝑓𝑘 ‖∞ 𝜇𝑖 (R ) + || 𝑓𝑘 𝑑𝜇𝑖 −
𝑛
𝑓𝑘 𝑑𝜇|| + ‖𝑓𝑘 − 𝑢‖∞ 𝜇(R𝑛 )
| ∫ ∫ |
| |
⩽ ‖𝑢 − 𝑓𝑘 ‖∞ (𝑀 + 𝜇(R𝑛 )) + || 𝑓𝑘 𝑑𝜇𝑖 − 𝑓𝑘 𝑑𝜇 ||
| ∫ ∫ |
(ii)
← ‖𝑢 − 𝑓𝑘 ‖∞ (𝑀 + 𝜇(R𝑛 )) ←←←←←←←←←→
←←←←←←←←←→ ← 0.
𝑖→∞ 𝑘→∞
Assume that 𝑓 ∈ 𝐶𝑏 (R𝑛 ). For 𝜖 > 0, Party (iv) shows that there is some 𝑅 > 0 such that
with 𝐾 ∶= [−𝑅, 𝑅]𝑛
𝜇𝑖 (𝐾𝑛𝑐 ) = 𝜇𝑖 (R𝑛 ⧵ 𝐾) ⩽ 𝜖.
249
R.L. Schilling: Measures, Integrals & Martingales
(vi) Let (𝜇𝑘 )𝑘∈N be a weakly convergent sequence of finite measures. Define 𝑓 (𝑥) ∶= 𝑒−𝑖𝑥⋅𝜉 ,
𝜉 ∈ R𝑛 , we get
1 1
𝜇̂𝑘 (𝜉) = 𝑒−𝑖𝑥⋅𝜉 𝑑𝜇𝑘 (𝑥) ←←←←←←←←←→
← 𝑒−𝑖𝑥⋅𝜉 𝜇(𝑑𝑥) = 𝜇̂(𝜉),
(2𝜋)𝑛 ∫ 𝑘→∞ (2𝜋)𝑛 ∫
i.e. the Fourier transforms converge pointwise. From part (iv) we know that the sequence
(𝜇𝑘 )𝑘∈N is tight. For 𝜖 > 0 there is some 𝑅 > 0 such that 𝜇𝑘 (R𝑛 ⧵ 𝐾) ⩽ 𝜖 for 𝐾 ∶=
[−𝑅, 𝑅]𝑛 . Without loss of generality we can enlarge 𝑅 to make sure that 𝜇(R𝑛 ⧵ 𝐾) ⩽ 𝜖,
too. Because of the (uniform) continuity of the function R ∋ 𝑟 → 𝑒𝑖𝑟 on compact sets,
there is some 𝛿 > 0 such that
1 1
𝜇𝑘 (𝜉) − 𝜇̂𝑘 (𝜂)| ⩽
|̂ |𝑒𝑖𝜉⋅𝑥 − 𝑒𝑖𝜂⋅𝑥 | 𝜇𝑘 (𝑑𝑥) = |𝑒𝑖(𝜉−𝜂)⋅𝑥 − 1| 𝜇𝑘 (𝑑𝑥)
(2𝜋)𝑛 ∫ (2𝜋)𝑛 ∫
1 1
= |𝑒𝑖(𝜉−𝜂)⋅𝑥 − 1| 𝜇𝑘 (𝑑𝑥) + |𝑒𝑖(𝜉−𝜂)⋅𝑥 − 1| 𝜇 (𝑑𝑥)
(2𝜋) 𝐾 ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟
𝑛 ∫ (2𝜋)𝑛 ∫𝐾 𝑐 ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟ 𝑘
⩽𝜖 ⩽2
𝜇𝑘 (R𝑛 ) 2
⩽ 𝜖+ 𝜇 (𝐾 𝑐 )
(2𝜋)𝑛 (2𝜋)𝑛 𝑖
1
⩽ (𝑀 + 2)𝜖
(2𝜋)𝑛
where 𝑀 ∶= sup𝑘∈N 𝜇𝑘 (R𝑛 ) < ∞. This proves the equicontinuity of the sequence
𝜇𝑘 )𝑘∈N .
(̂
𝜇(𝜉) − 𝜇̂(𝜂)| ⩽ 𝜖
|̂ ∀|𝜉 − 𝜂| ⩽ 𝛿.
𝜇𝑘 (𝜂) − 𝜇̂(𝜂)| ⩽ |̂
|̂ 𝜇𝑘 (𝜂) − 𝜇̂𝑘 (𝜉)| +|̂
𝜇𝑘 (𝜉) − 𝜇̂(𝜉)| + |̂
𝜇(𝜉) − 𝜇̂(𝜂)|
⏟⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏟ ⏟⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏟
⩽𝜖 ⩽𝜖
⇐⇒ sup |̂
𝜇𝑘 (𝜂) − 𝜇̂(𝜂)| ⩽ 2𝜖 + |̂
𝜇𝑘 (𝜉) − 𝜇̂(𝜉)| ←←←←←←←←←→
← 2𝜖 ←←←←←←←→
← 0.
𝜂∈𝐵𝛿 (𝜉) 𝑘→∞ 𝜖→0
Here we use that 𝜇̂𝑘 converges pointwise to 𝜇̂, cf. (vi). The calculation shows that 𝜇̂𝑘 con-
verges locally uniformly to 𝜇̂. Since locally uniform convergence is the same as uniform
convergence on compact sets, we are done.
■■
250
Solution Manual. Last update 28th October 2021
(i) Since 𝜇 is a finite measure, the continuity of 𝜇̂ follows directly from the continuity
lemma, Theorem 12.4 (cf. also 19.3). In order to show positive definiteness, pick 𝑚 ∈
N, 𝜉1 , … , 𝜉𝑚 ∈ R𝑛 and 𝜆1 , … , 𝜆𝑚 ∈ C. We get
𝑚 𝑚
∑ 1 ∑
𝜙(𝜉𝑗 − 𝜉𝑘 )𝜆𝑗 𝜆̄ 𝑘 = 𝜆 𝜆̄ 𝑒−𝑖𝑥⋅(𝜉𝑗 −𝜉𝑘 ) 𝜇(𝑑𝑥)
𝑗,𝑘=1
(2𝜋)𝑛 𝑗,𝑘=1 𝑗 𝑘 ∫
𝑚
1 ∑
= 𝜆 𝜆̄ 𝑒−𝑖𝑥⋅𝜉𝑗 𝑒−𝑖𝑥⋅𝜉𝑘 𝜇(𝑑𝑥)
(2𝜋)𝑛 𝑗,𝑘=1 𝑗 𝑘 ∫
(𝑚 )( 𝑚 )
1 ∑ ∑
= 𝜆𝑗 𝑒−𝑖𝑥⋅𝜉𝑗 𝜆𝑘 𝑒−𝑖𝑥⋅𝜉𝑘 𝜇(𝑑𝑥)
(2𝜋)𝑛 ∫ 𝑗=1 𝑘=1
|∑ |2
1 |𝑚 |
= | 𝜆 𝑒−𝑖𝑥⋅𝜉𝑗 | 𝜇(𝑑𝑥) ⩾ 0.
|
(2𝜋)𝑛 ∫ | 𝑗=1 𝑗 |
|
| |
(ii) For 𝑚 = 1 and 𝜉 = 0 the definition of positive definiteness implies that the matrix (𝜙(0))
is positive definite; in particular, 𝜙(0) ⩾ 0.
∑
If we have for a matrix (𝑎𝑖𝑘 ) that 𝑖𝑘 𝑎𝑖𝑘 𝜆𝑖 𝜆̄ 𝑗 ⩾ 0, then
∑ ∑ ∑ ∑
0⩽ 𝑎𝑖𝑘 𝜆𝑖 𝜆̄ 𝑘 = 𝑎𝑖𝑘 𝜆𝑖 𝜆̄ 𝑘 = 𝑎𝑖𝑘 𝜆̄ 𝑖 𝜆𝑘 = 𝑎𝑘𝑖 𝜆̄ 𝑘 𝜆𝑖
𝑖𝑘 𝑖𝑘 𝑖𝑘 𝑘𝑖
which means that 𝑎𝑖𝑘 = 𝑎𝑘𝑖 . Apply this to the matrix 𝑎𝑖𝑘 = 𝜙(𝜉𝑖 − 𝜉𝑘 ) with 𝑚 = 2 and
𝜉1 = 𝜉 and 𝜉2 = 0 to infer that 𝜙(𝜉) = 𝜙(−𝜉). Moreover, the matrix
( )
𝜙(0) 𝜙(−𝜉)
𝜙(𝜉) 𝜙(0)
0 ⩽ 𝜙(0)2 − 𝜙(−𝜉)𝜙(𝜉).
𝑘→∞ ∫𝐼 𝑘 ∫𝐼 𝑘
𝑚,𝑗∈N 𝑚 𝑗
∑ ( 𝑘 2
)( 𝑘 2
)
𝜙(𝜉𝑗𝑘 − 𝜉𝑚𝑘 ) 𝑘−𝑛 𝑒𝑖𝑥⋅𝜉𝑗 𝑒−2𝜖|𝜉𝑗 | 𝑘−𝑛 𝑒𝑖𝑥⋅𝜉𝑗 𝑒−2𝜖|𝜉𝑗 |
𝑘 𝑘
= lim
𝑘→∞
𝑚,𝑗∈N
251
R.L. Schilling: Measures, Integrals & Martingales
⩾ 0.
we obtain
( ( ))
𝜈𝜖 (𝑥) = 𝑒𝑖𝑥⋅𝜉 𝑒𝑖𝑥⋅𝜂 𝑒−2𝜖|𝜂|2 −2𝜖|𝜉|2 𝑑𝜉 𝑑𝜂
∬
( 2 2
)
= 𝑒𝑖𝑥⋅(𝜉−𝜂) 𝑒−|𝜉−𝜂| −|𝜉+𝜂| 𝑑𝜉 𝑑𝜂.
∬
leads to
1 2 2
𝜈𝜖 (𝑥) = 𝜙(𝑡)𝑒𝑖𝑥⋅𝑡 𝑒−𝜖(|𝑡| +|𝑠| ) 𝑑𝑡 𝑑𝑠
|det 𝐴| ∬
1 2
= 𝜙(𝑡)𝑒−𝜖|𝑡| 𝑒𝑖𝑥⋅𝑡 𝑑𝑡
𝑐∫
1
= 𝜙 (𝑡)𝑒𝑖𝑥⋅𝑡 𝑑𝑡. (⋆)
𝑐∫ 𝜖
(iv) Define ( )
1 |𝑥|2
𝑔𝑡 (𝑥) ∶= exp − .
(2𝜋𝑡)𝑛∕2 2𝑡
Applying Theorem 19.12 for the finite measure 𝜇(𝑑𝑥) ∶= 𝑒−𝑡|𝑥| 𝑑𝑥 yields
2
𝑡 2 (⋆) 1 𝑡 2 1 𝑡 2
𝜈𝜖 (𝑥)𝑒− 2 |𝑥| 𝑑𝑥 = −1 (𝜙𝜖 )(𝑥)𝑒− 2 |𝑥| 𝑑𝑥 = 𝜙𝜖 (𝜉) −1 (𝑒− 2 |⋅| )(𝜉) 𝑑𝜉
∫ 𝑐 ∫ 𝑐 ∫
for all 𝑡 > 0 (observe: 𝜙𝜖 ∈ 𝐿1 (R𝑛 )). Example 19.2(iii) shows (𝑔𝑡 )(𝑥) = (2𝜋)−𝑛 exp(−𝑡|𝑥|2 ∕2).
Therefore, −1 (𝑒− 2 |⋅| )(𝜉) = (2𝜋)𝑛 𝑔𝑡 (𝜉). Since |𝜙(𝜉)| ⩽ 𝜙(0) and ∫ 𝑔𝑡 (𝑥) 𝑑𝑥 = 1 we
𝑡 2
thus get
𝑡 2 (2𝜋)𝑛 (2𝜋)𝑛
𝜈𝜖 (𝑥)𝑒− 2 |𝑥| 𝑑𝑥 = 𝜙𝜖 (𝜉)𝑔𝑡 (𝜉) 𝑑𝜉 ⩽ 𝜙(0).
∫ 𝑐 ∫ 𝑐
Fatou’s lemma (Theorem 9.11) finally shows
1 2
𝜈𝜖 (𝑥) 𝑑𝑥 = lim 𝜈𝜖 (𝑥)𝑒− 2𝑘 |𝑥| 𝑑𝑥
∫ ∫ 𝑘→∞
1 2
⩽ lim inf 𝜈𝜖 (𝑥)𝑒− 2𝑘 |𝑥| 𝑑𝑥
𝑘→∞ ∫
(2𝜋)𝑛
⩽ 𝜙(0).
𝑐
Since 𝜈𝜖 ⩾ 0, see (iii), this means that 𝜈𝜖 ∈ 𝐿1 (R𝑛 ).
252
Solution Manual. Last update 28th October 2021
(v) Parts (iii) and (iv) show that 𝜇̂𝜖 = 𝜙𝜖 for the finite measure 𝜇𝜖 (𝑑𝑥) ∶= 𝑐𝜈𝜖 (𝑥) 𝑑𝑥. Since
𝜙𝜖 → 𝜙, Lévy’s continuity theorem (Problem 21.3) shows that there exists a measure
𝜇 which is the weak limit of the family 𝜇𝜖 as 𝜖 → 0 and 𝜇̂ = 𝜙.
■■
(i) Since uniform convergence preserves continuity, we see that every 𝑢 ∈ 𝐶𝑐 (𝑋) is con-
tinuous. By construction, the set {|𝑢| ⩾ 𝜖} is compact since there is some 𝑢𝜖 ∈ 𝐶𝑐 (𝑋)
such that ‖𝑢 − 𝑢𝜖 ‖∞ < 𝜖. This means that 𝑢 vanishes at infinity. In particular 𝐶𝑐 (𝑋) ⊂
𝐶∞ (𝑋).
Conversely, if 𝑢 ∈ 𝐶∞ (𝑋) and 𝜖 > 0, there is some compact set 𝐾𝜖 such that |𝑢| ⩽ 𝜖
outside of 𝐾𝜖 . Now we use Urysohn’s lemma and construct a function 𝜒𝜖 ∈ 𝐶𝑐 (𝑋) such
that 1𝐾𝜖 ⩽ 𝜒𝜖 ⩽ 1. Then we get 𝑢𝜖 ∶= 𝜒𝜖 𝑢 ∈ 𝐶𝑐 (𝑋) as well as
|𝑢 − 𝑢𝜖 | = (1 − 𝜒𝜖 )|𝑢| ⩽ 𝜖
(ii) It is obvious that 𝐶∞ (𝑋) is a vector space and that ‖∙‖∞ is a norm in this space. The
completeness follows from part (i) since 𝐶∞ (𝑋) = 𝐶𝑐 (𝑋) = 𝐶𝑐 (𝑋).
(iii) Let 𝑢 ∈ 𝐶∞ (𝑋) and 𝜖 > 0. Urysohn’s lemma shows that there is a 𝜒 ∈ 𝐶𝑐 (𝑋),
0 ⩽ 𝜒 ⩽ 1, such that |𝑢| ⩽ 𝜖 on the set {𝜒 < 1} = {𝜒 = 1}𝑐 . Therefore,
| |
| 𝑢 𝑑𝜇𝑛 − 𝑢 𝑑𝜇||
|∫ ∫
| |
| | | |
⩽ || 𝑢𝜒 𝑑𝜇𝑛 − 𝑢𝜒 𝑑𝜇|| + || 𝑢(1 − 𝜒) 𝑑𝜇𝑛 − 𝑢(1 − 𝜒) 𝑑𝜇 ||
|∫ ∫ | |∫ ∫ |
| | [ ]
⩽ || 𝑢𝜒 𝑑𝜇𝑛 − 𝑢𝜒 𝑑𝜇|| + 𝜖 𝜇𝑛 (𝑋) + 𝜇(𝑋)
|∫ ∫ |
21.16 | |
⩽ || 𝑢𝜒 𝑑𝜇𝑛 − 𝑢𝜒 𝑑𝜇|| + 2𝜖 sup 𝜇𝑚 (𝑋).
|∫ ∫ | 𝑚∈N
| |
lim sup || 𝑢 𝑑𝜇𝑛 − 𝑢 𝑑𝜇|| ⩽ 2𝜖 sup 𝜇𝑚 (𝑋) ←←←←←←←→
← 0.
𝑛→∞ |∫ ∫ | 𝑚∈N 𝜖→0
■■
(i) First we consider 𝑢 ∈ 𝐶𝑐∞ (R𝑛 ). According to Theorem 19.23, −1 𝑢 ∈ (R𝑛 ), and
Plancherel’s theorem (Theorem 19.12) gives
253
R.L. Schilling: Measures, Integrals & Martingales
Since |𝜇̂𝑖 (𝜉)| ⩽ 𝜇̂𝑖 (0) → 𝜙(0) is uniformly bounded, we can use the dominated conver-
gence theorem to see
i.e. 𝑀 ∶= sup𝑖 𝜇𝑖 (R𝑛 ) < ∞. Assume now that 𝑢 ∈ 𝐶𝑐 (𝑋). Since 𝐶𝑐∞ (R𝑛 ) is dense in
𝐶𝑐 (R𝑛 ) (with respect to uniform convergence, cf. Problem 15.13), there is a sequence
(𝑢𝑘 )𝑘∈N ⊂ 𝐶𝑐∞ (R𝑛 ) such that ‖𝑢𝑘 − 𝑢‖∞ → 0. Thus,
| |
| 𝑢 𝑑𝜇 − 𝑢 𝑑𝜇𝑗 ||
|∫ 𝑖 ∫
| |
| | | | | |
⩽ || (𝑢 − 𝑢𝑘 ) 𝑑𝜇𝑖 || + || (𝑢 − 𝑢𝑘 ) 𝑑𝜇𝑗 || + || 𝑢𝑘 𝑑𝜇𝑖 − 𝑢𝑘 𝑑𝜇𝑗 ||
|∫ | |∫ | |∫ ∫ |
( ) | |
⩽ ‖𝑢 − 𝑢𝑘 ‖∞ 𝜇𝑖 (R𝑛 ) + 𝜇𝑗 (R𝑛 ) + || 𝑢𝑘 𝑑𝜇𝑖 − 𝑢 𝑑𝜇 |
|∫ ∫ 𝑘 𝑗 ||
| |
⩽ 2‖𝑢 − 𝑢𝑘 ‖∞ 𝑀 + || 𝑢𝑘 𝑑𝜇𝑖 − 𝑢𝑘 𝑑𝜇𝑗 ||
|∫ ∫ |
← 2‖𝑢 − 𝑢𝑘 ‖∞ 𝑀 ←←←←←←←←←→
←←←←←←←←←←←→ ← 0.
𝑖,𝑗→∞ 𝑘→∞
( )
This shows that ∫ 𝑢 𝑑𝜇𝑖 𝑖∈N
is a Cauchy sequence in R. Thus, the limit Λ(𝑢) ∶=
lim𝑖→∞ ∫ 𝑢 𝑑𝜇𝑖 exists. Since convergent sequences are bounded, we see
| |
sup || 𝑢 𝑑𝜇𝑖 || < ∞.
𝑖∈N |∫ |
Since 𝑢 ∈ 𝐶𝑐 (R𝑛 ) ⇐⇒ |𝑢| ∈ 𝐶𝑐 (R𝑛 ), we get
i.e. the sequence (𝜇𝑖 )𝑖∈N is vaguely bounded. According to Theorem 21.18, (𝜇𝑖 )𝑖∈N has
a vaguely convergent subsequence 𝜇𝑛(𝑖) → 𝜇.
(ii) We can use part (i) for any subsequence of (𝜇𝑖 )𝑖∈N . We will show the the subsequential
limits do not depend on the subsequence. Pick any two subsequences (𝜇𝑛(𝑖) )𝑖∈N and
v v
(𝜇𝑚(𝑖) )𝑖∈N of (𝜇𝑖 )𝑛∈N and assume that 𝜇𝑛(𝑖) ←←←←→ ← 𝜈. By definition, we find
← 𝜇, 𝜇𝑚(𝑖) ←←←←→
for all 𝑢 ∈ 𝐶𝑐 (R𝑛 )
On the other hand, we have seen in (i) that Λ(𝑢) = lim𝑖→∞ ∫ 𝑢 𝑑𝜇𝑖 . Thus,
𝑢 𝑑𝜇 = Λ(𝑢) = 𝑢 𝑑𝜈.
∫ ∫
254
Solution Manual. Last update 28th October 2021
Since this holds for all 𝑢 ∈ 𝐶𝑐 (R𝑛 ), we can use the regularity of the measures 𝜇 and
𝜈 to conclude that 𝜇 = 𝜈. Since the limit does not depend on the subsequence, we
already have vague convergence of the full sequence (𝜇𝑖 )𝑖∈N . (Compare this with the
following subsequence principle: A sequence (𝑎𝑖 )𝑖∈N ⊂ R converges if, and only if,
every subsequence of (𝑎𝑖 )𝑖∈N has a convergent subsequence, and all subsequential limits
coincide.)
(iii) In view of Theorem 21.17 it is enough to show that the sequence (𝜇𝑖 )𝑖∈N is tight
(observe, that 𝜇̌𝑖 (𝜉) = (2𝜋)𝑛 𝜇̂𝑖 (−𝜉)). Now we can use dominated convergence to get
for all 𝑅 ⩾ 𝛿1 . In particular, we find 𝜇𝑖 (R𝑛 ⧵ [−𝑅, 𝑅]𝑛 ) ⩽ 3(2𝜋)𝑛 𝜖 for 𝑖 ⩾ 𝑛0 (𝜖). In
order to ensure 𝜇𝑖 (R𝑛 ⧵ [−𝑅, 𝑅]𝑛 ) ⩽ 3(2𝜋)𝑛 𝜖 for 𝑖 = 1, … , 𝑛0 (𝜖), we can enlarge 𝑅, if
need be.
■■
𝑢 𝑑𝜇𝑛 = 𝑢 𝑑𝜇𝑛
∫𝐵 ∫𝐵∩supp 𝑢
we can assume, without loss of generality, that 𝐵 is contained in a compact set. Denote by 𝐾 ∶= 𝐵
the closure of 𝐵 and by 𝑈 ∶= 𝐵 ◦ the open interior of 𝐵. Moreover, we can assume that 𝑢 ⩾ 0 –
otherwise we consider 𝑢± separately.
According to Urysohn’s lemma (Lemma B.2 or (21.6) & (21.7)), there are sequences (𝑤𝑘 )𝑘∈N ⊂
𝐶𝑐 (𝑋), (𝑣𝑘 )𝑘∈N ⊂ 𝐶𝑐 (𝑋), 0 ⩽ 𝑣𝑘 ⩽ 1, 0 ⩽ 𝑤𝑘 ⩽ 1, with 𝑤𝑘 ↑ 1𝑈 and 𝑣𝑘 ↓ 1𝐾 . By assumption
v
← 𝜇 and so
𝜇𝑛 ←←←←→
255
R.L. Schilling: Measures, Integrals & Martingales
■■
256
22 Uniform integrability and Vitali’s
convergence theorem.
Solutions to Problems 22.1–22.17
Thus,
Equivalently,
⋃ ⋂
{lim 𝑢𝑗 = 0}𝑐 = {sup |𝑢𝑗 | > 1∕𝑘}.
𝑗
𝑘∈N 𝑁∈N 𝑗⩾𝑁
and, since countable unions of null sets are again null sets, we conclude that
■■
257
R.L. Schilling: Measures, Integrals & Martingales
and since
⋃ ⋂⋃ def
{|𝑢𝑗 | > 𝜖} ↓ {|𝑢𝑗 | > 𝜖} = lim sup{|𝑢𝑗 | > 𝜖}
𝑗⩾𝑘 𝑘∈N 𝑗⩾𝑘 𝑗→∞
This, and the result of Problem 22.1 show that either of the following two equivalent conditions
( )
lim 𝜇 sup |𝑢𝑗 | ⩾ 𝜖 = 0 ∀ 𝜖 > 0;
𝑘→∞ 𝑗⩾𝑘
( )
𝜇 lim sup{|𝑢𝑗 | ⩾ 𝜖} = 0 ∀ 𝜖 > 0;
𝑗→∞
■■
Since
we see that
(since, otherwise |𝑢𝑗 − 𝑢𝑘 | ⩽ 𝜖 + 𝜖 = 2𝜖). Thus, we get for every measurable set 𝐴 with finite
𝜇-measure that
( )
𝜇 {|𝑢𝑗 − 𝑢𝑘 | > 2𝜖} ∩ 𝐴
[ ]
⩽ 𝜇 ({|𝑢𝑗 − 𝑢| > 𝜖} ∩ 𝐴) ∪ ({|𝑢𝑘 − 𝑢| > 𝜖} ∩ 𝐴)
[ ] [ ]
⩽ 𝜇 {|𝑢𝑗 − 𝑢| > 𝜖} ∩ 𝐴 + 𝜇 {|𝑢𝑘 − 𝑢| > 𝜖} ∩ 𝐴
258
Solution Manual. Last update 28th October 2021
(Note that 𝑁𝑗 may depend on 𝓁, but we suppress this dependency as 𝓁 is fixed.) By enlarging
𝑁𝑗 , if needed, we can always assume that
Consequently, there is an exceptional set 𝐸𝑗 ⊂ 𝐴𝓁 with 𝜇(𝐸𝑗 ∩ 𝐴𝓁 ) < 2−𝑗 such that
⋃
and, if 𝐸𝑖∗ ∶= 𝑗⩾𝑖 𝐸𝑗 we have 𝜇(𝐸𝑖 ∩ 𝐴𝓁 ) ⩽ 2 ⋅ 2−𝑖 as well as
so that lim𝑗 𝑢𝑁𝑗 exists uniformly on 𝐴𝓁 ⧵ 𝐸𝑖∗ for all 𝑖. Since 𝜇(𝐸𝑖∗ ∩ 𝐴𝓁 ) < 2 ⋅ 2−𝑖 we conclude
that
for some 𝑢(𝓁) . Since, however, a.e. limits are unique (up to a null set, that is) we know that
𝑢(𝓁) = 𝑢(𝑚) a.e. on 𝐴𝓁 ∩𝐴𝑚 so that there is a (up to null sets) unique limit function 𝑢 satisfying
Thus, we have found a candidate for the limit of our Cauchy sequence. In fact, since
we have
𝜇({|𝑢𝑘 − 𝑢| > 𝜖} ∩ 𝐴𝓁 )
⩽ 𝜇({|𝑢𝑘 − 𝑢𝑁𝑗 | > 𝜖} ∩ 𝐴𝓁 ) + 𝜇({|𝑢𝑁𝑗 − 𝑢| > 𝜖} ∩ 𝐴𝓁 )
and the first expression on the right-hand side tends to zero (as 𝑘, 𝑁(𝑗) → ∞) because of the
assumption, while the second term tends to zero (as 𝑁(𝑗) → ∞) because of (*))
■■
259
R.L. Schilling: Measures, Integrals & Martingales
This means, however, that potential a.e. and 𝑝 -limits must be 𝑓 ≡ 0, too. Since for
every 𝑥
so that the 𝑝 -limit does not exist. The pointwise limit, however, exists since
(iii) The shape of 𝑔𝑛 is that of a triangle with base [0, 1∕𝑛]. Thus, for every 𝜖 > 0,
1
𝜆(|ℎ𝑛 | > 𝜖) ⩽ 𝜆[0, 1∕𝑛] = 𝑛
𝑎𝑝𝑛
|ℎ𝑛 |𝑝 𝑑𝜆 = 𝑎𝑝𝑛 21 𝜆[0, 1∕𝑛] =
∫ 2𝑛
we have 𝑝 -convergence if, and only if, the sequence 𝑎𝑝𝑛 ∕𝑛 tends to zero as 𝑛 → ∞.
We have, however, always a.e. convergence since the support of the function ℎ𝑛 is
[0, 1∕𝑛] and this shrinks to {0} which is a null set. Thus,
lim 𝑎𝑛 (1 − 𝑛𝑥)+ = 0
𝑛
except, possibly, at 𝑥 = 0.
■■
260
Solution Manual. Last update 28th October 2021
Note that
so that
{|𝑎𝑢𝑗 + 𝑏𝑤𝑗 − 𝑎𝑢 − 𝑏𝑤| > 2𝜖} ⊂ {|𝑢𝑗 − 𝑢| > 𝜖∕|𝑎|} ∪ {|𝑤𝑗 − 𝑤| > 𝜖∕|𝑏|}.
we get
Finally, since
( )
1
max 𝑢𝑗 , 𝑤𝑗 = 𝑢𝑗 + 𝑤𝑗 + |𝑢𝑗 − 𝑤𝑗 |
2
we get max 𝑢𝑗 , 𝑤𝑗 → max 𝑢, 𝑤 by using rules (i) and (iv) several times. The minimum is treated
similarly.
■■
Problem 22.6 Solution: The hint is somewhat misleading since this construction is not always pos-
sible (or sensible). Just imagine R with the counting measure. Then 𝑋𝜎𝑓 would be all of R...
What I had in mind when giving this hint was a construction along the following lines:
Consider Lebesgue measure 𝜆 in R and define 𝑓 ∶= 1𝐹 + ∞1𝐹 𝑐 where 𝐹 = [−1, 1] (or any other
set of finite Lebesgue measure). Then 𝜇 ∶= 𝑓 ⋅ 𝜆 is a not 𝜎-finite measure. Moreover, Take any
sequence 𝑢𝑛 ←←←←→
← 𝑢 converging in 𝜆-measure. Then
𝜆
since all sets 𝐴 with 𝜇(𝐴) < ∞ are contained in 𝐹 and 𝜆(𝐹 ) = 𝜇(𝐹 ) < ∞. Thus, 𝑢𝑛 ←←←←←→
𝜇
← 𝑢.
However, changing 𝑢 arbitrarily on 1𝐹 𝑐 also yields a limit point in 𝜇-measure since, as mentioned
above, all sets of finite 𝜇-measure are within 𝐹 .
This pathology cannot happen in a 𝜎-finite measure space, cf. Lemma 22.6.
■■
261
R.L. Schilling: Measures, Integrals & Martingales
|𝑢 − 𝑢𝑗 | 𝑑𝜇 = |𝑢 − 𝑢𝑗 | 𝑑𝜇
∫ ∫𝐴
= |𝑢 − 𝑢𝑗 | 𝑑𝜇 + |𝑢 − 𝑢𝑗 | 𝑑𝜇
∫𝐴∩{|𝑢−𝑢𝑗 |⩽𝜖} ∫𝐴∩{|𝑢−𝑢𝑗 |>𝜖}
⩽ 𝜖 𝑑𝜇 + (|𝑢| + |𝑢𝑗 |) 𝑑𝜇
∫𝐴∩{|𝑢−𝑢𝑗 |⩽𝜖} ∫𝐴∩{|𝑢−𝑢𝑗 |>𝜖}
( )
⩽ 𝜖𝜇(𝐴) + 2𝐶 𝜇 𝐴 ∩ {|𝑢 − 𝑢𝑗 | > 𝜖}
← 𝜖𝜇(𝐴)
←←←←←←←←←→
𝑗→∞
← 0.
←←←←←←←→
𝜖→0
|𝑢𝑗 | 𝑑𝜆 = 𝜆[𝑗, 𝑗 + 1] = 1 ≠ 0
∫
so that the limit—if it exists—cannot be 𝑢 ≡ 0. Since this is, however, the canonical
candidate, we conclude that there is no 1 convergence.
(iii) The limit depends on the set 𝐴 which is fixed. This means that we are, essentially,
dealing with a finite measure space.
■■
Problem 22.8 Solution: A pseudo-metric is symmetric (𝑑2 ) and satisfies the triangle inequality (𝑑3 ).
(i) First we note that 𝜌𝜇 (𝜉, 𝜂) ∈ [0, 1] is well-defined. That it is symmetric (𝑑2 ) is obvi-
ous. For the triangle inequality we observe that for three random variables 𝜉, 𝜂, 𝜁 and
numbers 𝜖, 𝛿 > 0 we have
|𝜉 − 𝜁| ⩽ |𝜉 − 𝜂| + |𝜂 − 𝜁 |
implying that
so that
𝜌P (𝜉, 𝜁) ⩽ 𝜖 + 𝛿.
262
Solution Manual. Last update 28th October 2021
𝜌P (𝜉𝑗 , 𝜉) ←←←←←←←←←→
← 0 ⇐⇒ ∃ (𝜖𝑗 )𝑗 ⊂ R+ ∶ P(|𝜉 − 𝜉𝑗 | > 𝜖𝑗 ) ⩽ 𝜖𝑗
𝑗→∞
Thus, for given 𝜖 > 0 we pick 𝑁 = 𝑁(𝜖) such that 𝜖 > 𝜖𝑗 for all 𝑗 ⩾ 𝑁 (possible as
𝜖𝑗 → 0). Then we find
← 0. Then
Conversely, assume that 𝜉𝑗 ←←←←←→
P
(iii) We have
(ii) P
𝜌(𝜉𝑗 , 𝜉𝑘 ) ←←←←←←←←←←←←←←→
← 0 ⇐⇒ 𝜉𝑗 − 𝜉𝑘 ←←←←←←←←←←←←←←→
← 0
𝑗,𝑘→∞ 𝑗,𝑘→∞
P22.3 P
⇐⇒ ∃ 𝜉 ∶ 𝜉𝑘 ←←←←←←←←←→
← 𝜉
𝑘→∞
(ii)
⇐⇒ ∃ 𝜉 ∶ 𝜌(𝜉, 𝜉𝑘 ) ←←←←←←←←←←←←→
← 0
𝑘→∞
𝑥+𝑦 𝑥 𝑦 𝑥 𝑦
= + ⩽ +
1+𝑥+𝑦 1+𝑥+𝑦 1+𝑥+𝑦 1+𝑥 1+𝑦
and
⎧
⎪𝑥 + 𝑦 = (𝑥 ∧ 1) + (𝑦 ∧ 1) if 𝑥 + 𝑦 ⩽ 1;
(𝑥 + 𝑦) ∧ 1 = ⎨
⎪1 ⩽ (𝑥 ∧ 1) + (𝑦 ∧ 1) if 𝑥 + 𝑦 ⩾ 1.
⎩
This means that both 𝑔P and 𝑑P satisfy the triangle inequality, that is (𝑑3 ). Symmetry,
i.e. (𝑑2 ), is obvious.
𝑥 𝑥
⩽𝑥∧1⩽2
1+𝑥 1+𝑥
263
R.L. Schilling: Measures, Integrals & Martingales
which shows that 𝑔P and 𝑑P have the same Cauchy sequences. Moreover, for all 𝜖 ⩽ 1,
𝑑P (𝜉, 𝜂) = |𝜉 − 𝜂| ∧ 1 𝑑P + |𝜉 − 𝜂| ∧ 1 𝑑P
∫|𝜉−𝜂|>𝜖 ∫|𝜉−𝜂|⩽𝜖
⩽ 1 𝑑P + 𝜖 𝑑P
∫|𝜉−𝜂|>𝜖 ∫|𝜉−𝜂|⩽𝜖
⩽ P(|𝜉 − 𝜂| > 𝜖) + 𝜖,
■■
(ii) WLOG we show that (𝑢𝑛 )𝑛 contains an a.e. convergent subsequence. Let (𝐴𝑘 )𝑘 be as in
the hint and fix 𝑖. By (i) we know that |𝑢 − 𝑢𝑛 | ∧ 1𝐴𝑖 → 0 in 𝐿1 . By Corollary 13.8 we
see that there is a subsequence 𝑢(𝑖)
𝑛 such that |𝑢 − 𝑢𝑛 | ∧ 1𝐴𝑖 → 0 almost everywhere.
(𝑖)
(iii) We are now in the setting of Corollary 13.8: |𝑢𝑛 |, |𝑢| ⩽ 𝑤 for some 𝑤 ∈ 𝑝 (𝜇) and 𝑢𝑛 ←←→
𝜇
←
𝑢. Thus, every subsequence (𝑢′𝑛 )𝑛 ⊂ (𝑢𝑛 )𝑛 converges in measure to the same limit 𝑢 and
a.e.
by (ii) there is some (𝑢′′𝑛 )𝑛 ⊂ (𝑢′𝑛 )𝑛 such that 𝑢′′𝑛 ←←←←←→
← 𝑢. Now we can use the dominated
convergence theorem (Theorem 12.2 or Theorem 13.9) to show that lim𝑛 ‖𝑢′′𝑛 −𝑢‖𝑝 = 0.
264
Solution Manual. Last update 28th October 2021
Assume now that 𝑢𝑛 does not converge to 𝑢 in 𝐿𝑝 . This means that lim sup𝑛 ‖𝑢𝑛 − 𝑢‖𝑝 >
0, i.e. there is some subsequence such that lim inf 𝑛 ‖𝑢′𝑛 − 𝑢‖𝑝 > 0. On the other hand,
there is some (𝑢′′𝑛 )𝑛 ⊂ (𝑢′𝑛 )𝑛 such that
■■
Problem 22.10 Solution: Note that the sets 𝐴𝑗 are of finite 𝜇-measure. Observe that the functions
𝑓𝑗 ∶= 𝑢1𝐴𝑗
• converge in 𝜇-measure to 𝑓 ≡ 0:
sup |𝑓𝑗 | 𝑑𝜇 = 0
𝑗 ∫{|𝑓𝑗 |>|𝑢|}
■■
A sequence (𝑥𝑛 )𝑛 ⊂ R converges to 0 if, and only if, every subsequence (𝑥𝑛𝑘 )𝑘 contains
some sub-subsequence (̃
𝑥𝑛𝑘 )𝑘 which converges to 0.
Necessity is again trivial. Sufficiency: assume that (𝑥𝑛 )𝑛 does not converge to 0. Then
the sequence (min{|𝑥𝑛 |, 1})𝑛 is bounded and still does not converge to 0. Since this
sequence is bounded, it contains a convergent subsequence (𝑥𝑛𝑘 )𝑘 with some limit 𝛼 ≠ 0.
But then (𝑥𝑛𝑘 )𝑘 cannot contain a sub-subsequence (̃
𝑥𝑛𝑘 )𝑘 which is a null sequence.
Note that (unless we are in a 𝜎-finite measure space) the exceptional set may depend on
the testing set 𝐴.
265
R.L. Schilling: Measures, Integrals & Martingales
𝑥𝑛 ∶= 𝜇({|𝑢𝑛 − 𝑢| > 𝜖} ∩ 𝐴) →
̸ 0.
Since the whole sequence (𝑥𝑛 )𝑛 is bounded (by 𝜇(𝐴)) there exists some subsequence
(𝑥𝑛𝑘 )𝑘 given by (𝑢𝑛𝑘 )𝑘 such that
This contradicts, however, the fact that 𝑥𝑛𝑘 has itself a subsequence converging to zero.
(iii) Fix some set 𝐴 of finite 𝜇-measure. All conclusions below take place relative to resp.
on this set only.
This means, however, that every subsequence (Φ◦𝑢𝑛𝑘 )𝑘 of (Φ◦𝑢𝑛 )𝑛 has a sub-subsequence
𝑢𝑛𝑘 )𝑘 which converges a.e. to Φ◦𝑢. Thus, part (ii) says that Φ◦𝑢𝑛 ←←←←←→
𝜇
(Φ◦̃ ← Φ◦𝑢.
■■
Problem 22.12 Solution: Since and are uniformly integrable, we find for any given 𝜖 > 0
functions 𝑓𝜖 , 𝑔𝜖 ∈ 1+ such that
|𝑓𝑗 | 𝑑𝜇 = |𝑓𝑗 | 𝑑𝜇 = 0
∫{|𝑓𝑗 |>𝑓 } ∫∅
(ii) Instead of {𝑓1 , … , 𝑓𝑁 } (which is uniformly integrable because of (i)) we show that
∪ is uniformly integrable.
⎧
⎪∫ |𝑤| 𝑑𝜇 ⩽ 𝜖 if 𝑤 ∈
|𝑤| 𝑑𝜇 ⩽ ⎨ {|𝑤|>𝑓𝜖 }
∫{|𝑤|>ℎ𝜖 } ⎪∫{|𝑤|>𝑔 } |𝑤| 𝑑𝜇 ⩽ 𝜖 if 𝑤 ∈ .
⎩ 𝜖
266
Solution Manual. Last update 28th October 2021
|𝑓 + 𝑔| 𝑑𝜇
∫
{|𝑓 +𝑔|>ℎ𝜖 }
( )
⩽ |𝑓 | + |𝑔| 𝑑𝜇 + |𝑓 | ∨ |𝑔| 𝑑𝜇 + |𝑓 | ∨ |𝑔| 𝑑𝜇
∫ ∫ ∫
{|𝑓 |>ℎ𝜖 } {|𝑓 |>ℎ𝜖 } {|𝑓 |⩽ℎ𝜖 }
∩{|𝑔|>ℎ𝜖 } ∩{|𝑔|⩽ℎ𝜖 } ∩{|𝑔|>ℎ𝜖 }
= |𝑓 | 𝑑𝜇 + |𝑔| 𝑑𝜇 + |𝑓 | 𝑑𝜇 + |𝑔| 𝑑𝜇
∫ ∫ ∫ ∫
{|𝑓 |>ℎ𝜖 } {|𝑓 |>ℎ𝜖 } {|𝑓 |>ℎ𝜖 } {|𝑓 |⩽ℎ𝜖 }
∩{|𝑔|>ℎ𝜖 } ∩{|𝑔|>ℎ𝜖 } ∩{|𝑔|⩽ℎ𝜖 } ∩{|𝑔|>ℎ𝜖 }
⩽ |𝑓 | 𝑑𝜇 + |𝑔| 𝑑𝜇 + |𝑓 | 𝑑𝜇 + |𝑔| 𝑑𝜇
∫ ∫ ∫ ∫
{|𝑓 |>ℎ𝜖 } {|𝑔|>ℎ𝜖 } {|𝑓 |>ℎ𝜖 } {|𝑔|>ℎ𝜖 }
⩽ |𝑓 | 𝑑𝜇 + |𝑔| 𝑑𝜇 + |𝑓 | 𝑑𝜇 + |𝑔| 𝑑𝜇
∫ ∫ ∫ ∫
{|𝑓 |>𝑓𝜖 } {|𝑔|>𝑔𝜖 } {|𝑓 |>𝑓𝜖 } {|𝑔|>𝑔𝜖 }
⩽ 4𝜖
• 𝑡 ⇝ ,
• (1 − 𝑡) ⇝ ,
• 𝑡𝑓𝜖 ⇝ 𝑓𝜖 ,
• (1 − 𝑡)𝑓𝜖 ⇝ 𝑔𝜖 ,
and observe that the calculation is uniform for all 𝑡 ∈ [0, 1].
(v) Without loss of generality we can assume that is convex, i.e. coincides with its convex
hull.
Let 𝑢 be an element of the 1 -closure of (the convex hull of) . Then there is a sequence
(𝑓𝑗 )𝑗 ⊂ ∶ lim ‖𝑢 − 𝑓𝑗 ‖1 = 0.
𝑗
267
R.L. Schilling: Measures, Integrals & Martingales
[ ]
∪ {|𝑢 − 𝑓𝑗 | ⩽ 𝑓𝜖 } ∩ {|𝑓𝑗 | > 𝑓𝜖 }
that
|𝑢| 𝑑𝜇
∫
{|𝑢|>𝑓𝜖 }
⩽ |𝑢 − 𝑓𝑗 | 𝑑𝜇 + |𝑓𝑗 | 𝑑𝜇
∫ ∫
{|𝑢−𝑓𝑗 |>𝑓𝜖 } {|𝑢−𝑓𝑗 |>𝑓𝜖 }
∩{|𝑓𝑗 |>𝑓𝜖 } ∩{|𝑓𝑗 |>𝑓𝜖 }
+ |𝑢 − 𝑓𝑗 | ∨ |𝑓𝑗 | 𝑑𝜇 + |𝑢 − 𝑓𝑗 | ∨ |𝑓𝑗 | 𝑑𝜇
∫ ∫
{|𝑢−𝑓𝑗 |>𝑓𝜖 } {|𝑢−𝑓𝑗 |⩽𝑓𝜖 }
∩{|𝑓𝑗 |⩽𝑓𝜖 } ∩{|𝑓𝑗 |>𝑓𝜖 }
⩽ ‖𝑢 − 𝑓𝑗 ‖1 + |𝑓𝑗 | 𝑑𝜇 + ‖𝑢 − 𝑓𝑗 ‖1 + |𝑓𝑗 | 𝑑𝜇
∫ ∫
{|𝑓𝑗 |>𝑓𝜖 } {|𝑓𝑗 |>𝑓𝜖 }
⩽ 2 ‖𝑢 − 𝑓𝑗 ‖1 + 2𝜖
← 2𝜖.
←←←←←←←←←→
𝑗→∞
■■
sup |𝑓 | 𝑑𝜇
∫{sup1⩽𝑗⩽𝑘 |𝑓𝑗 |>𝑤𝜖 } 1⩽𝑗⩽𝑘 𝑗
𝑘
∑
⩽ |𝑓𝓁 | 𝑑𝜇
∫{sup1⩽𝑗⩽𝑘 |𝑓𝑗 |>𝑤𝜖 }∩{|𝑓𝓁 |=sup1⩽𝑗⩽𝑘 |𝑓𝑗 |}
𝓁=1
𝑘
∑
⩽ |𝑓𝓁 | 𝑑𝜇
∫{|𝑓𝓁 |>𝑤𝜖 }
𝓁=1
𝑘
∑
⩽ 𝜖
𝓁=1
= 𝑘 𝜖.
Therefore,
sup |𝑓𝑗 | 𝑑𝜇
∫ 1⩽𝑗⩽𝑘
⩽ sup |𝑓 | 𝑑𝜇 + sup |𝑓 | 𝑑𝜇
∫{sup1⩽𝑗⩽𝑘 |𝑓𝑗 |⩽𝑤𝜖 } 1⩽𝑗⩽𝑘 𝑗 ∫{sup1⩽𝑗⩽𝑘 |𝑓𝑗 |>𝑤𝜖 } 1⩽𝑗⩽𝑘 𝑗
268
Solution Manual. Last update 28th October 2021
⩽ 𝑤𝜖 𝑑𝜇 + 𝑘𝜖
∫
and we get
1 1
lim sup |𝑓𝑗 | 𝑑𝜇 ⩽ lim 𝑤 𝑑𝜇 + 𝜖 = 𝜖
𝑘→∞ 𝑘∫ 1⩽𝑗⩽𝑘 𝑘→∞ 𝑘∫ 𝜖
■■
Problem 22.14 Solution: Since the function 𝑢 ≡ 𝑅, 𝑅 > 0, is integrable w.r.t. the probability measure
P, we get
|𝑢𝑗 |𝑝−1
|𝑢𝑗 | 𝑑P ⩽ |𝑢𝑗 | 𝑑P
∫{|𝑢𝑗 |>𝑅} ∫{|𝑢𝑗 |>𝑅} 𝑅𝑝−1
1
= |𝑢𝑗 |𝑝 𝑑P
𝑅𝑝−1 ∫{|𝑢𝑗 |>𝑅}
1
⩽ |𝑢𝑗 |𝑝 𝑑P
𝑅𝑝−1 ∫
1
⩽ sup |𝑢𝑘 |𝑝 𝑑P
𝑅𝑝−1 𝑘 ∫
1
= sup ‖𝑢𝑘 ‖𝑝𝑝
𝑅𝑝−1 𝑘
which converges to zero as 𝑅 → ∞. This proves uniform integrability.
Counterexample:
Consider, for example, the probability space ((0, 1), ℬ(0, 1), 𝑑𝑥) and the sequence
𝑢𝑗 ∶= 𝑗 ⋅ 1(0,1∕𝑗) .
Then 𝑢𝑗 → 0 pointwise (everywhere!), hence in measure. This is also the expected 1 limit, if it
exists. Moreover,
‖𝑢𝑗 ‖1 = 𝑢𝑗 𝑑𝑥 = 1
∫
which means that 𝑢𝑗 cannot converge in 1 to the expected limit 𝑢 ≡ 0, i.e. it does not converge in
1 .
We can verify this fact also directly: for 𝑅 > 0 and all 𝑗 > 𝑅 we get
|𝑢𝑗 | 𝑑𝑥 = 𝑢𝑗 𝑑𝑥 = 1
∫{|𝑢𝑗 |>𝑅} ∫
269
R.L. Schilling: Measures, Integrals & Martingales
which proves
and (𝑢𝑗 )𝑗 cannot be uniformly integrable (in view of the equivalent characterizations of uniform
integrability on finite measure spaces, cf. Theorem 22.9)
■■
= |𝑓 | 𝑑𝜇,
∫{|𝑓 |>𝑘}
and, since 2𝑗 ⩾ 𝑗 + 1 for all 𝑗 ∈∈ N, also
∞ ∞
∑ ∑
2 𝑗 𝜇(𝑗 < |𝑓 | ⩽ 𝑗 + 1) = 2𝑗 𝜇(𝑗 < |𝑓 | ⩽ 𝑗 + 1)
𝑗=𝑘 𝑗=𝑘
∞
∑
= 2𝑗 𝑑𝜇
∫{𝑗<|𝑓 |⩽𝑗+1}
𝑗=𝑘
∞
∑
⩾ |𝑓 | 𝑑𝜇
∫{𝑗<|𝑓 |⩽𝑗+1}
𝑗=𝑘
= |𝑓 | 𝑑𝜇.
∫{|𝑓 |>𝑘}
This shows that
∞
∑
|𝑓 | 𝑑𝜇 ⩽ 2 𝑗 𝜇(𝑗 < |𝑓 | ⩽ 𝑗 + 1) ⩽ 2 |𝑓 | 𝑑𝜇
∫{|𝑓 |>𝑘} ∫{|𝑓 |>𝑘}
𝑗=𝑘
This proves the claim (since we are in a finite measure space where 𝑢 ≡ 𝑘 is an integrable function!)
■■
Problem 22.16 Solution: Fix 𝜖 > 0. By assumption there is some 𝑤 = 𝑤𝜖 ∈ 1+ such that
sup |𝑓𝑖 | 𝑑𝜇 ⩽ 𝜖.
𝑖 ∫{|𝑓𝑖 |>𝑤}
Since |𝑢𝑖 | ⩽ |𝑓𝑖 | we infer that {|𝑢𝑖 | > 𝑤} ⊂ {|𝑓𝑖 | > 𝑤}, and so
■■
270
Solution Manual. Last update 28th October 2021
This implies that uniform integrability of the family implies that the condition of Problem 22.17
holds. On the other hand,
⩽ (2|𝑢| − 𝑔) 𝑑𝜇
∫{|𝑢|⩾𝑔}
⩽ (2|𝑢| − 𝑔) 𝑑𝜇
∫{2|𝑢|⩾𝑔}
( )
=2 |𝑢| − 12 𝑔 𝑑𝜇
∫{|𝑢|⩾ 1 𝑔}
2
( [1 ] )
=2 |𝑢| − 2
𝑔 ∧ |𝑢| 𝑑𝜇
∫{|𝑢|⩾ 1 𝑔}
2
and since 𝑔 ∈ 1 if, and only if, 12 𝑔 ∈ 1 , we see that the condition given in Problem 22.17 entails
uniform integrability.
In finite measure spaces this conditions is simpler: constants are integrable functions in finite
measure spaces; thus we can replace the condition given in Problem 22.17 by
■■
271
23 Martingales.
Solutions to Problems 23.1–23.16
𝑢0 𝑑𝜇 = 𝑢1 𝑑𝜇
∫∅ ∫∅
always holds, it is clear that the calculation and choice in (*) is necessary and sufficient for the
claim.
■■
Problem 23.2 Solution: We consider only the martingale case, the other two cases are similar.
(a) Since ℬ𝑗 ⊂ 𝒜𝑗 we get
𝑢𝑗 𝑑𝜇 = 𝑢𝑗+1 𝑑𝜇 ∀ 𝐴 ∈ 𝒜𝑗
∫𝐴 ∫𝐴
⇐⇒ 𝑢𝑗 𝑑𝜇 = 𝑢𝑗+1 𝑑𝜇 ∀ 𝐵 ∈ ℬ𝑗
∫𝐵 ∫𝐵
showing that (𝑢𝑗 , ℬ𝑗 )𝑗 is a martingale.
(b) It is clear that the above implication cannot hold if we enlarge 𝒜𝑗 to become 𝒞𝑗 . Just consider
the following ‘extreme’ case (to get a counterexample): 𝒞𝑗 = 𝒜 for all 𝑗. Any martingale
(𝑢𝑗 , 𝒞 )𝑗 must satisfy,
𝑢𝑗 𝑑𝜇 = 𝑢𝑗+1 𝑑𝜇 ∀ 𝐴 ∈ 𝒜.
∫𝐴 ∫𝐴
Considering the sets 𝐴 ∶= {𝑢𝑗 < 𝑢𝑗+1 } ∈ 𝒜 and 𝐴′ ∶= {𝑢𝑗 > 𝑢𝑗+1 } ∈ 𝒜 we conclude that
and, similarly 𝜇({𝑢𝑗 < 𝑢𝑗+1 }) = 0 so that 𝑢𝑗 = 𝑢𝑗+1 almost everywhere and for all 𝑗. This
means that, if we start with a non-constant martingale (𝑢𝑗 , 𝒜𝑗 )𝑗 , then this can never be a
martingale w.r.t. the filtration (𝒞𝑗 )𝑗 .
273
R.L. Schilling: Measures, Integrals & Martingales
■■
Problem 23.3 Solution: For the notation etc. we refer to Problem 4.15. Since the completion 𝒜 𝑗 is
given by
{ }
𝒜 𝑗 = 𝜎(𝒜𝑗 , 𝒩 ), 𝒩 ∶= 𝑀 ⊂ 𝑋 ∶ ∃ 𝑁 ∈ 𝒜 , 𝑁 ⊃ 𝑀, 𝜇(𝑁) = 0
we find that for all 𝐴∗𝑗 ∈ 𝒜𝑗∗ there exists some 𝐴𝑗 ∈ 𝒜𝑗 such that
𝐴∗𝑗 ⧵ 𝐴𝑗 ∪ 𝐴𝑗 ⧵ 𝐴∗𝑗 ∈ 𝒩 .
Writing 𝜇̄ for the unique extension of 𝜇 onto 𝒜 (and thus onto 𝒜 𝑗 for all 𝑗) we get for 𝐴∗𝑗 , 𝐴𝑗 as
above
| | | |
| | | |
| 𝑢 𝑑 𝜇
̄ − 𝑢 𝑑𝜇 | = | 𝑢 𝑑 𝜇
̄ − 𝑢 𝑑 𝜇
̄ |
|∫ ∗ 𝑗 ∫𝐴 𝑗 𝑗 | | 𝑗 𝑗 |
| 𝐴𝑗 | |∫𝐴∗𝑗 ∫𝐴 𝑗 |
| | | |
| |
= || (1𝐴∗ − 1𝐴𝑗 )𝑢𝑗 𝑑 𝜇̄ ||
|∫ 𝑗
|
| |
⩽ |1 ∗ − 1𝐴𝑗 | 𝑢𝑗 𝑑 𝜇̄
∫ | 𝐴𝑗 |
⩽ 1𝑁 𝑢𝑗 𝑑𝜇 = 0
∫
for a suitable 𝜇-null-set 𝑁 ⊃ 𝐴∗𝑗 ⧵ 𝐴𝑗 ∪ 𝐴𝑗 ⧵ 𝐴∗𝑗 . This proves that
𝑢𝑗 𝑑 𝜇̄ = 𝑢𝑗 𝑑𝜇
∫𝐴 ∗ ∫𝐴𝑗
𝑗
and we see easily from this that (𝑢𝑗 , 𝒜𝑗∗ )𝑗 is again a (sub-, super-)martingale if (𝑢𝑗 , 𝒜𝑗 )𝑗 is a (sub-,
super-)martingale.
■■
Problem 23.4 Solution: To see that the condition is sufficient, set 𝑘 = 𝑗 + 1. For the necessity,
assume that 𝑘 = 𝑗 + 𝑚. Since 𝒜𝑗 ⊂ 𝒜𝑗+1 ⊂ ⋯ ⊂ 𝒜𝑗+𝑚 = 𝒜𝑘 we get from the submartingale
property
𝑢𝑗 𝑑𝜇 ⩾ 𝑢𝑘 𝑑𝜇 ∀ 𝑗 < 𝑘, ∀ 𝐴 ∈ 𝒜𝑗
∫𝐴 ∫𝐴
resp.
𝑢𝑗 𝑑𝜇 = 𝑢𝑘 𝑑𝜇 ∀ 𝑗 < 𝑘, ∀ 𝐴 ∈ 𝒜𝑗 .
∫𝐴 ∫𝐴
■■
274
Solution Manual. Last update 28th October 2021
Problem 23.5 Solution: We have 𝒮𝑗 = {𝐴 ∈ 𝒜𝑗 ∶ 𝜇(𝐴) < ∞} and we have to check conditions
(S1 )–(S3 ) for a semiring, cf. page 39. Indeed
∅ ∈ 𝒜𝑗 , 𝜇(∅) = 0 ⇐⇒ ∅ ∈ 𝒮𝑗 ⇐⇒ (𝑆1 );
and
and
Since 𝒮𝑗 ⊂ 𝒜𝑗 also 𝜎(𝒮𝑗 ) ⊂ 𝒜𝑗 . On the other hand, if 𝐴 ∈ 𝒜𝑗 with 𝜇(𝐴) = ∞ we can, because
of 𝜎-finiteness find a sequence (𝐴𝑘 )𝑘 ⊂ 𝒜0 ⊂ 𝒜𝑗 such that 𝜇(𝐴𝑘 ) < ∞ and 𝐴𝑘 ↑ 𝑋. Thus,
⋃
𝐴𝑘 ∩ 𝐴 ∈ 𝒮𝑗 for all 𝑘 and 𝐴 = 𝑘 (𝐴𝑘 ∩ 𝐴). This shows that 𝒜𝑗 ⊂ 𝜎(𝒮𝑗 ).
The rest of the problem is identical to remark 23.2(i) when combined with Lemma 16.6.
■■
Problem 23.6 Solution: Using Lemma 17.2 we can approximate 𝑢𝑗 ∈ 2 (𝒜𝑗 ) by simple functions in
∑
(𝒜𝑗 ), i.e. with functions of the form 𝑓𝑗𝓁 = 𝑚 𝑐𝑗𝓁,𝑚 1𝐴𝓁,𝑚 (the sum is a finite sum!) where 𝑐𝑗𝓁 ∈ R
𝑗
and 𝐴𝓁𝑗 ∈ 𝒜𝑗 . Using the Cauchy–Schwarz inequality we also see that
𝓁→∞
(𝑓𝑗𝓁 − 𝑢𝑗 )𝑢𝑗 𝑑𝜇 ⩽ ‖𝑓𝑗𝓁 − 𝑢𝑗 ‖𝐿2 ⋅ ‖𝑢𝑗 ‖𝐿2 ←←←←←←←←←←←→
← 0.
∫ 𝑗 fixed
1𝐴𝓁,𝑚 𝑢𝑘 𝑑𝜇 = 1𝐴𝓁,𝑚 𝑢𝑗 𝑑𝜇 ∀ 𝓁, 𝑚
∫ 𝑗 ∫ 𝑗
and therefore
𝑓𝑗𝓁 𝑢𝑘 𝑑𝜇 = 𝑓𝑗𝓁 𝑢𝑗 𝑑𝜇 ∀𝓁
∫ ∫
and since the limit 𝓁 → ∞ exists
■■
Problem 23.7 Solution: Since the 𝑓𝑗 ’s are bounded, it is clear that (𝑓 ∙ 𝑢)𝑘 is integrable. Now take
𝐴 ∈ 𝒜𝑘 . Then
𝑘+1
∑
(𝑓 ∙ 𝑢)𝑘+1 𝑑𝜇 = 𝑓𝑗−1 (𝑢𝑗 − 𝑢𝑗−1 ) 𝑑𝜇
∫𝐴 ∫𝐴
𝑗=1
275
R.L. Schilling: Measures, Integrals & Martingales
= (𝑓 ∙ 𝑢)𝑘 + 𝑓𝑘 (𝑢𝑘+1 − 𝑢𝑘 ) 𝑑𝜇
∫𝐴
= 1𝐴 ⋅ 𝑓𝑘 𝑢𝑘 𝑑𝜇 − 1𝐴 ⋅ 𝑓𝑘 𝑢𝑘 𝑑𝜇
∫ ∫
=0
(𝑓 ∙ 𝑢)𝑘+1 𝑑𝜇 = (𝑓 ∙ 𝑢)𝑘 𝑑𝜇 ∀ 𝐴 ∈ 𝒜𝑘 .
∫𝐴 ∫𝐴
■■
2
𝑆𝑛+1 − 𝑆𝑛2 = (𝑆𝑛 + 𝜉𝑛+1 )2 − 𝑆𝑛2 = 𝜉𝑛+1
2
+ 𝜉𝑛+1 𝑆𝑛 .
2
(𝑆𝑛+1 − 𝑆𝑛2 ) 𝑑P = 2
𝜉𝑛+1 𝑑P + 𝜉𝑛+1 𝑆𝑛 𝑑P
∫𝐴 ∫𝐴 ∫𝐴
⩾ 𝜉𝑛+1 𝑆𝑛 𝑑P
∫𝐴
= 𝜉𝑛+1 (1𝐴 𝑆𝑛 ) 𝑑P
∫
= 𝜉 𝑑P 1𝐴 𝑆𝑛 𝑑P
∫ 𝑛+1 ∫
⏟⏞⏞⏞⏟⏞⏞⏞⏟
=0
= 0.
=𝑛 𝜉12 𝑑P
∫
276
Solution Manual. Last update 28th October 2021
so that 𝜅 ∶= ∫ 𝜉12 𝑑P is a reasonable candidate for the assertion. Using the calculation of
part (i) of this problem we see
2
[𝑆𝑛+1 − 𝜅(𝑛 + 1)] − [𝑆𝑛2 − 𝜅𝑛] = 𝜉𝑛+1
2
+ 𝜉𝑛+1 𝑆𝑛 − 𝜅
and integrating over ∫𝐴 … 𝑑P for any 𝐴 ∈ 𝒜𝑛 gives, just as in (i), because of independence
of 1𝐴 and 𝜉𝑛+1 resp. 1𝐴 𝑆𝑛 and 𝜉𝑛+1
( )
2
[𝑆𝑛+1 − 𝜅(𝑛 + 1)] − [𝑆𝑛2 − 𝜅𝑛] 𝑑P
∫𝐴
2
= 1𝐴 ⋅ 𝜉𝑛+1 𝑑P + 𝜉𝑛+1 𝑑P 1𝐴 ⋅ 𝑆𝑛 𝑑P − 𝜅 𝑑P
∫ ∫ ∫ ∫𝐴
2
= P(𝐴) 𝜉𝑛+1 𝑑P − 𝜅 𝑑P
∫ ∫𝐴
=0
■■
2 2
𝑀𝑛+1 − 𝑀𝑛 = 𝜉𝑛+1 + 𝑆𝑛 𝜉𝑛+1 − 𝜎𝑛+1 .
(𝑀𝑛+1 − 𝑀𝑛 ) 𝑑P
∫𝐴
2 2
= 𝜉𝑛+1 𝑑P + 𝑆𝑛 𝜉𝑛+1 𝑑P − 𝜎𝑛+1 𝑑P
∫𝐴 ∫𝐴 ∫𝐴
= P(𝐴) 𝜉 2 𝑑P + 𝑆𝑛 𝑑P 2
𝜉 𝑑P −𝜎𝑛+1 P(𝐴)
∫Ω 𝑛+1 ∫𝐴 ∫Ω 𝑛+1
⏟⏞⏞⏞⏟⏞⏞⏞⏟ ⏟⏞⏞⏞⏟⏞⏞⏞⏟
2
= 𝜎𝑛+1 =0
= 0,
where we use the independence of 1𝐴 and 𝜉𝑛+1 resp. of 1𝐴 𝑆𝑛 and 𝜉𝑛+1 and the hint given in the
statement of the problem.
■■
which shows that (𝑢𝑛 , 𝒜𝑛 )𝑛 is a martingale, hence (𝑢2𝑛 , 𝒜𝑛 )𝑛 is a submartingale—cf. Example 23.3(vi).
Now
∑ ∑
𝑢2𝑛 𝑑𝜇 = 𝑑𝑗2 𝑑𝜇 + 2 𝑑𝑗 𝑑𝑘 𝑑𝜇
∫ 𝑗
∫ ∫
𝑗<𝑘
277
R.L. Schilling: Measures, Integrals & Martingales
but, just as in Problem 23.6, we can approximate 𝑑𝑗 by 𝒜𝑗 -measurable simple functions (𝑓𝑗𝓁 )𝓁∈N
which shows, since ∫𝐴 𝑑𝑘 𝑑𝜇 = 0 for any 𝐴 ∈ 𝒜𝑗 and 𝑘 > 𝑗:
𝑑𝑗 𝑑𝑘 𝑑𝜇 = lim 𝑓𝑗𝓁 𝑑𝑘 𝑑𝜇 = 0.
∫ 𝓁 ∫
■■
■■
Problem 23.12 Solution: A solution in a more general context can be found in Example 25.4 on page
297 of the textbook.
■■
𝑢𝑗 𝑑𝜇 ⩾ 𝑢𝑗+1 𝑑𝜇 ∀𝑗 ∈ N, 𝐴 ∈ 𝒜𝑗 .
∫𝐴 ∫𝐴
0= 𝑢𝑘 𝑑𝜇 ⩾ 𝑢𝑘+1 𝑑𝜇 ⩾ 0
∫𝑋 ∫𝑋
■■
278
Solution Manual. Last update 28th October 2021
𝐴 ∈ 𝒜𝜏 ⇐⇒ 𝐴 ∈ 𝒜 and ∀ 𝑗 ∶ 𝐴 ∩ {𝜏 ⩽ 𝑗} ∈ 𝒜𝑗 .
Thus,
• ∅ ∈ 𝒜𝜏 is obvious;
• if 𝐴 ∈ 𝒜𝜏 , then
𝐴𝑐 ∩ {𝜏 ⩽ 𝑗} = {𝜏 ⩽ 𝑗} ⧵ 𝐴 = {𝜏 ⩽ 𝑗} ⧵ (𝐴 ∩ {𝜏 ⩽ 𝑗}) ∈ 𝒜𝑗
⏟⏟⏟ ⏟⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏟
∈ 𝒜𝑗 ∈ 𝒜𝑗
thus 𝐴𝑐 ∈ 𝒜𝜏 .
• if 𝐴𝓁 ∈ 𝒜𝜏 , 𝓁 ∈ N, then
[ ]
⋃ ⋃[ ]
𝐴𝓁 ∩ {𝜏 ⩽ 𝑗} = 𝐴𝓁 ∩ {𝜏 ⩽ 𝑗} ∈ 𝒜𝑗
𝓁 𝓁 ⏟⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏟
∈ 𝒜𝑗
⋃
thus 𝐴𝓁 ∈ 𝒜𝜏 .
■■
∀ 𝑛 ∈ N0 ∶ {𝜏 ⩽ 𝑛} ∈ 𝒜𝑛 .
{𝜏 < 𝑛} = {𝜏 ⩽ 𝑛 − 1} ∈ 𝒜𝑛−1 ⊂ 𝒜𝑛
{𝜏 = 𝑛} = {𝜏 ⩽ 𝑛} ⧵ {𝜏 < 𝑛} ∈ 𝒜𝑛 .
{𝜏 ⩽ 𝑘} = {𝜏 = 0} ∪ {𝜏 = 1} ∪ ⋯ ∪ {𝜏 = 𝑘} ∈ 𝒜0 ∪ ⋯ ∪ 𝒜𝑘 ⊂ 𝒜𝑘 .
■■
Problem 23.16 Solution: Since 𝜎 ∧ 𝜏 ⩽ 𝜎 and 𝜎 ∧ 𝜏 ⩽ 𝜏, we find from Lemma 23.6 that
ℱ𝜎∧𝜏 ⊂ ℱ𝜎 ∩ ℱ𝜏 .
𝐴 ∩ {𝜎 ⩽ 𝑗} ∈ ℱ𝑗 and 𝐴 ∩ {𝜏 ⩽ 𝑗} ∈ ℱ𝑗 ∀ 𝑗 ∈ N0 .
Thus,
( )
𝐴 ∩ {𝜎 ∧ 𝜏 ⩽ 𝑗} = 𝐴 ∩ {𝜎 ⩽ 𝑗} ∪ {𝜏 ⩽ 𝑗} ∈ ℱ𝑗
279
24 Martingale convergence theorems.
Solutions to Problems 24.1–24.9
Problem 24.1 Solution: We have 𝜏0 = 0 which is clearly a stopping time and since
it is clear that
The claim follows by induction once we have shown that 𝜎𝑘 and 𝜏𝑘 are stopping times for a generic
value of 𝑘. Since the structure of their definitions are similar, we show this for 𝜎𝑘 only.
{𝜎𝑘 > 𝑁} = ∅ ∈ 𝒜𝑁 .
■■
Problem 24.2 Solution: Theorem 24.7 becomes for supermartingales: Let (𝑢𝓁 )𝓁∈−N be a backwards
supermartingale and assume that 𝜇|𝒜−∞ is 𝜎-finite. Then lim𝑗→∞ 𝑢−𝑗 = 𝑢−∞ ∈ (−∞, ∞] exists a.e.
Moreover, 𝐿1 -lim𝑗→∞ 𝑢−𝑗 = 𝑢−∞ if, and only if, sup𝑗 ∫ 𝑢−𝑗 𝑑𝜇 < ∞; in this case (𝑢𝓁 , 𝒜𝓁 )𝓁∈−N is
a supermartingale and 𝑢−∞ is finitely-valued.
Using this theorem the claim follows immediately from the supermartingale property:
■■
281
R.L. Schilling: Measures, Integrals & Martingales
Problem 24.3 Solution: Corollary 24.3 shows pointwise a.e. convergence. Using Fatou’s lemma we
get
⩾ lim inf 𝑢𝑗 𝑑𝜇
∫ 𝑗→∞
= 𝑢∞ 𝑑𝜇 ⩾ 0
∫
so that 𝑢∞ = 0 a.e.
Moreover, since ∫ 𝑢𝑗 𝑑𝜇 ←←←←←←←←←→
← 0 = ∫ 𝑢∞ 𝑑𝜇, Theorem 24.6 shows that 𝑢𝑗 → 𝑢∞ in 𝐿1 -sense.
𝑗→∞
■■
Problem 24.4 Solution: From 𝐿1 -lim𝑗→∞ 𝑢𝑗 = 𝑓 we conclude that sup𝑗 ∫ |𝑢𝑗 | 𝑑𝜇 < ∞ and we get
that lim𝑗→∞ 𝑢𝑗 exists a.e. Since 𝐿1 -convergence also implies a.e. convergence of a subsequence,
the limiting functions must be the same.
■■
Problem 24.5 Solution: The quickest solution uses the famous Chung-Fuchs result that a simple
random walk (this is just 𝑆𝑗 ∶= 𝜉1 + ⋯ + 𝜉𝑗 with 𝜉𝑘 iid Bernoulli 𝑝 = 𝑞 = 12 ) does not converge
and that −∞ = lim inf 𝑗 𝑆𝑗 < lim sup𝑗 𝑆𝑗 = ∞ a.e. Knowing this we are led to
1
𝑃 (𝑢𝑗 converges) = 𝑃 (𝜉0 + 1 = 0) = .
2
It remains to show that 𝑢𝑗 is a martingale. For 𝐴 ∈ 𝜎(𝜉1 , … , 𝜉𝑗 ) we get
= 𝑢𝑗 𝑑𝑃 + (𝜉0 + 1) 𝑑𝑃 𝜉𝑗+1 𝑑𝑃
∫𝐴 ∫𝐴 ∫Ω
= 𝑢𝑗 𝑑𝑃
∫𝐴
where the last step follows because of independence.
If you do not know the Chung-Fuchs result, you could argue as follows: assume that for some finite
random variable 𝑆 the limit 𝑆𝑗 (𝜔) → 𝑆(𝜔) takes place on a set 𝐴 ⊂ Ω. Since the 𝜉𝑗 ’s are iid, we
have
𝜉2 + 𝜉3 + ⋯ → 𝑆
and
𝜉1 + 𝜉2 + ⋯ → 𝑆
which means that 𝑆 and 𝑆 + 𝜉1 have the same probability distribution. But this entails that 𝑆 is
necessarily ±∞, i.e., 𝑆𝑗 cannot have a finite limit.
■■
282
Solution Manual. Last update 28th October 2021
Finally, lim sup𝑛 1𝐴𝑛 = 1lim sup𝑛 𝐴𝑛 = 1 a.e. while, by Fatou’s lemma
i.e., lim inf 𝑛 1𝐴𝑛 = 0 a.e. This means that 1𝐴𝑛 does not have a limit as 𝑛 → ∞.
(iii) For 𝐴 ∈ 𝒜𝑛 we have because of independence
𝑀𝑛+1 𝑑𝜆
∫𝐴
= 𝑀𝑛 𝑑𝜆.
∫𝐴
(iv) We have
{𝑀𝑛+1 ≠ 0}
= {𝑀𝑛+1 ≠ 0, 𝜉𝑛2 +1 = −1} ∪ {𝑀𝑛+1 ≠ 0, 𝜉𝑛2 +1 = +1}
⊂ 𝐴𝑛 ∪ {𝑀𝑛 ≠ 0, 𝜉𝑛2 +1 = +1}.
283
R.L. Schilling: Measures, Integrals & Martingales
(v) By definition,
so that
This shows that for 𝑥 ∈ {lim𝑛 𝑀𝑛 exists} the limit lim𝑛 1𝐴𝑛 (𝑥) exists. But, because of
(ii), the latter is a null set, so that the pointwise limit of 𝑀𝑛 cannot exist.
1
𝜆(𝑀𝑛+𝑘 ≠ 0) ⩽ 2𝑘
𝜆(𝑀𝑛 ≠ 0) + 𝜆(𝐴𝑛 ) + ⋯ 𝜆(𝐴𝑛+𝑘−1 )
1
⩽ 2𝑘
+ 𝜆(𝐴𝑛 ) + ⋯ 𝜆(𝐴𝑛+𝑘−1 ).
■■
{ }
Problem 24.7 Solution: Note that for 𝐴 ∈ {1}, {2}, … , {𝑛}, {𝑛 + 1, 𝑛 + 2, …} we have
𝜉𝑛+1 𝑑𝑃 = (𝑛 + 2)1[𝑛+2,∞)∩N 𝑑𝑃
∫𝐴 ∫𝐴
⎧
⎪0 if 𝐴 is a singleton
=⎨
⎪∫ (𝑛 + 2)1[𝑛+2,∞)∩N 𝑑𝑃 else
⎩ [𝑛+1,∞)∩N
and in the second case we have
(𝑛 + 2)1[𝑛+2,∞)∩N 𝑑𝑃 = (𝑛 + 2) 𝑑𝑃
∫[𝑛+1,∞)∩N ∫[𝑛+2,∞)∩N
∞
∑
= (𝑛 + 2) 𝑃 ({𝑗})
𝑗=𝑛+2
∞ ( )
∑ 1 1
= (𝑛 + 2) −
𝑗=𝑛+2
𝑗 𝑗+1
= 1,
𝜉𝑛 𝑑𝑃 = (𝑛 + 1)1[𝑛+1,∞)∩N 𝑑𝑃
∫𝐴 ∫𝐴
284
Solution Manual. Last update 28th October 2021
⎧
⎪0 if 𝐴 is a singleton
=⎨
⎪∫ (𝑛 + 1)1[𝑛+1,∞)∩N 𝑑𝑃 = 1 else
⎩ [𝑛+1,∞)∩N
so that
𝜉𝑛+1 𝑑𝑃 = 𝜉𝑛 𝑑𝑃
∫𝐴 ∫𝐴
for all 𝐴 from a generator of the 𝜎-algebra which contains an exhausting sequence. This shows,
by Remark 23.2(i) that (𝜉𝑛 )𝑛 is indeed a martingale.
are obvious.
■■
= 𝑢2𝑗 𝑑𝜇 − 𝑢2𝑗−1 𝑑𝜇
∫ ∫
which means that
𝑁
∑
𝑢2𝑁 𝑑𝜇 = (𝑢𝑗 − 𝑢𝑗−1 )2 𝑑𝜇
∫ ∫
𝑗=1
285
R.L. Schilling: Measures, Integrals & Martingales
← 1𝐴𝑘 𝑢
1𝐴𝑘 𝑢𝑛 ←←←←←←←←←→
𝑛→∞
Now we use Fatou’s lemma and the result of part (ii) to get
⩽ lim sup (𝑢 − 𝑢𝑗 )2 𝑑𝜇
𝑗 ∫
∞
∑
⩽ lim sup (𝑢𝓁 − 𝑢𝓁−1 )2 𝑑𝜇
𝑗 ∫
𝓁=𝑗+1
=0
∑∞
since, by 𝐿2 -boundedness, 𝑘=1 ∫ (𝑢𝑘 − 𝑢𝑘−1 )2 𝑑𝜇 < ∞.
(iv) Since 𝜇(𝜉) < ∞, constants are integrable and we find using the Cauchy–Schwarz and
Markov inequalities
√
√
|𝑢𝑘 | 𝑑𝜇 ⩽ 𝜇(|𝑢𝑘 | > 𝑅) ⋅ 𝑢2𝑘 𝑑𝜇
∫|𝑢𝑘 |>𝑅 ∫
√ √
1
⩽ 𝑢2𝑘 𝑑𝜇 ⋅ 𝑢2𝑘 𝑑𝜇
𝑅 ∫ ∫
1
⩽ sup 𝑢2 𝑑𝜇
𝑅 𝑘 ∫ 𝑘
from which we get uniform integrability; the claim follows now from parts (i)–(iii) and
Theorem 24.6.
■■
286
Solution Manual. Last update 28th October 2021
∑∞
Problem 24.8 now shows that 2
𝑗=1 𝑦𝑗 < ∞ means that the martingale (𝜉𝑛 )𝑛 is 𝐿2 -
bounded, i.e. 𝜉𝑛 converges a.e. The converse follows from part (iii).
(ii) This follows with the same arguments as in part (i) with 𝒜𝑛 = 𝜎(𝜉1 , … , 𝜉𝑛 ).
2
𝑀𝑛+1 𝑑𝑃 = (𝑆𝑛+1 − 𝐴𝑛+1 ) 𝑑𝑃
∫𝐴 ∫𝐴
= (𝑆𝑛2 − 𝐴𝑛 ) 𝑑𝑃 + 2
(2𝜉𝑛+1 𝑆𝑛 + 𝜉𝑛+1 2
− 𝜎𝑛+1 ) 𝑑𝑃
∫𝐴 ∫𝐴
2 2
= 𝑀𝑛 𝑑𝑃 + (2𝜉𝑛+1 𝑆𝑛 + 𝜉𝑛+1 − 𝜎𝑛+1 ) 𝑑𝑃
∫𝐴 ∫𝐴
But, because of independence,
2 2
(2𝜉𝑛+1 𝑆𝑛 + 𝜉𝑛+1 − 𝜎𝑛+1 ) 𝑑𝑃
∫𝐴
2 2
= 2𝜉𝑛+1 𝑑𝑃 𝑆𝑛 𝑑𝑃 + 𝑃 (𝐴) 𝜉𝑛+1 𝑑𝑃 − 𝑃 (𝐴)𝜎𝑛+1
∫𝐴 ∫Ω ∫
2 2
= 0 + 𝑃 (𝐴)𝜎𝑛+1 − 𝑃 (𝐴)𝜎𝑛+1
= 0.
Now define
where we use, for the estimate of 𝑀𝜏−1 , the definition of 𝜏 for the last estimate. Since
(𝑀𝑛∧𝜏 )𝑛 is a martingale, this gives
2
(𝑆𝑛∧𝜏 − 𝐴𝑛∧𝜏 ) 𝑑𝑃 = (𝑆02 − 𝐴0 ) 𝑑𝑃 = 0
∫ ∫
so that
2
𝐴𝑛∧𝜏 𝑑𝑃 = 𝑆𝑛∧𝜏 𝑑𝑃 ⩽ (𝜅 + 𝐶)2
∫ ∫
287
R.L. Schilling: Measures, Integrals & Martingales
uniformly in 𝑛.
𝐴𝜏 𝑑𝑃 ⩽ (𝜅 + 𝐶)2 < ∞
∫
∑
which means that 𝐴𝜏 < ∞ almost surely. But since 𝑗 𝜉𝑗 converges almost surely,
𝑃 (𝜏 = ∞) = 1 for sufficiently large 𝜅, and we are done.
■■
288
25 Martingales in action.
Solutions to Problems 25.1–25.15
Problem 25.1 Solution: This problem is intimately linked with problem 25.7.
Without loss of generality we assume that 𝜇 and 𝜈 are finite measures, the case for 𝜎-finite 𝜇 and
arbitrary 𝜈 is exactly as in the proof of Theorem 25.2.
Let (𝐴𝑖 )𝑖 be as described in the problem and define the finite 𝜎-algebras 𝒜𝑛 ∶= 𝜎(𝐴1 , … , 𝐴𝑛 ).
Using the hint we can achieve that
( )
𝒜𝑛 = 𝜎 𝐶1𝑛 , … , 𝐶𝓁(𝑛)
𝑛
⨃
with mutually disjoint 𝐶𝑖𝑘 ’s and 𝓁(𝑛) ⩽ 2𝑛 + 1 and 𝑛
𝑖 𝐶𝑖 = 𝑋. Then the construction of Example
25.4 yields a countably-indexed martingale since the 𝜎-algebras 𝒜𝑖 are increasing.
This means, that the countable version of the martingale convergence theorem is indeed enough
for the proof.
■■
Problem 25.2 Solution: “⇒”: Assume first that (25.1) holds, i.e. that 𝜈 ≪ 𝜇. If 𝜇(𝐴 ▵ 𝐵) = 0 for
some 𝐴, 𝐵 ∈ 𝒜 we get 𝜈(𝐴 ▵ 𝐵) = 0. By definition,
so that
Assume that 𝜈(𝐴) < ∞. Then 𝜈(𝐴 ∩ 𝐵) ⩽ 𝜈(𝐴) < ∞ and we see that
“⇐”: Assume now that the condition stated in the problem is satisfied. If 𝑁 ∈ 𝒜 is any 𝜇-null
set, we choose 𝐴 ∶= 𝑁 and 𝐵 ∶= ∅ and observe that 𝐴 ▵ 𝐵 = 𝑁. Thus,
289
R.L. Schilling: Measures, Integrals & Martingales
■■
Problem 25.3 Solution: Using simply the Radon–Nikodým theorem, Theorem 25.2, gives
with a measurable function 𝑥 → 𝑝𝑡 (𝑥); it is, however, far from being clear that (𝑡, 𝑥) → 𝑝𝑡 (𝑥) is
jointly measurable.
A slight variation of the proof of Theorem 25.2 allows us to incorporate parameters provided the
families of measures are measurable w.r.t. these parameters. Following the hint we set (notation
as in the proof of 25.2)
∑ 𝜈𝑡 (𝐴)
𝑝𝛼 (𝑡, 𝑥) ∶= 𝐼 (𝑥)
𝜇 (𝐴) 𝐴
𝐴∈𝛼 𝑡
(𝑡, 𝑥) → 𝑝(𝑡, 𝑥)
exists (say, in 𝐿1 , 𝑡 being fixed) then the limiting function is again jointly measurable.
Using exactly the arguments of the proof of Theorem 25.2 with 𝑡 fixed we can confirm that this
limit exists and defines a jointly measurable function with the property that
Because of the a.e. uniqueness of the Radon–Nikodým density the functions 𝑝(𝑡, 𝑥) and 𝑝𝑡 (𝑥) co-
incide, for every 𝑡 a.e. as functions of 𝑥; without additional assumptions on the nature of the de-
pendence on the parameter, the exceptional set may, though, depend on 𝑡!
■■
= 1𝑁 (𝑥 + 𝑦) 𝜆(𝑑𝑥)𝜈(𝑑𝑦)
∬
Tonelli
= 1𝑁 (𝑥 + 𝑦) 𝜈(𝑑𝑦)𝜆(𝑑𝑥)
∬
290
Solution Manual. Last update 28th October 2021
= 𝜈(𝑁 − 𝑦) 𝜆(𝑑𝑦).
∫
Therefore, 𝜈(𝑁 −𝑦) = 0 for Lebesgue almost every 𝑦, i.e. there is some 𝑥0 such that 𝜈(𝑁 +𝑥0 ) = 0.
Now we use the quasi-invariance to get 𝜈(𝑁) = 𝜈((𝑁 + 𝑥0 ) − 𝑥0 ) = 0.
𝜆 ≪ 𝜈 . We show that 𝜈(𝑁) = 0 ⇐⇒ 𝜆(𝑁) = 0. Let 𝑁 ∈ ℬ(R𝑛 ) be a null set for the measure 𝜈.
Similar to the first part of the proof we get
■■
Problem 25.5 Solution: Have a look at the respective solutions for Chapter 20.
■■
Problem 25.6 Solution: We write 𝑢± for the positive resp. negative parts of 𝑢 ∈ 1 (𝒜 ), i.e. 𝑢 =
𝑢+ − 𝑢− and 𝑢± ⩾ 0. Fix such a function 𝑢 and define
𝜈 ± (𝐹 ) ∶= 𝑢± (𝑥) 𝜇(𝑑𝑥), ∀𝐹 ∈ ℱ.
∫𝐹
∀ 𝑁 ∈ ℱ , 𝜇(𝑁) = 0 ⇐⇒ 𝜈 ± (𝑁) = 𝑢± 𝑑𝜇 = 0
∫𝑁
which means that 𝜈 ± ≪ 𝜇. By the Radon–Nikodým theorem we find (up to null-sets unique)
positive functions 𝑓 ± ∈ 1 (ℱ ) such that
𝜈 ± (𝐹 ) = 𝑓 ± 𝑑𝜇 ∀𝐹 ∈ ℱ.
∫𝐹
𝑢ℱ 𝑑𝜇 = 𝑢 𝑑𝜇 ∀𝐹 ∈ ℱ.
∫𝐹 ∫𝐹
𝑤 𝑑𝜇 = 𝑢 𝑑𝜇 ∀𝐹 ∈ ℱ.
∫𝐹 ∫𝐹
Since then
𝑢ℱ 𝑑𝜇 = 𝑤 𝑑𝜇 ∀𝐹 ∈ ℱ.
∫𝐹 ∫𝐹
0= (𝑤 − 𝑢ℱ ) 𝑑𝜇
∫{𝑤>𝑢ℱ }
291
R.L. Schilling: Measures, Integrals & Martingales
which is only possible if 𝜇({𝑤 > 𝑢ℱ }) = 0. Similarly we conclude that 𝜇({𝑤 < 𝑢ℱ }) = 0 from
which we get 𝑤 = 𝑢ℱ almost everywhere.
𝑢𝑗 𝑑𝜇 ⩽ 𝑢𝑗+1 𝑑𝜇 ∀𝐴 ∈ 𝒜𝑗 , ∀ 𝑗.
∫𝐴 ∫𝐴
almost everywhere, ∀ 𝑗.
𝒜
𝑗
𝑢𝑗 ⩽ 𝑢𝑗+1
The direction ‘⇒’ is clear. To see ‘⇐’ we fix 𝑗 and observe that, since
𝒜
𝑢𝑗 𝑑𝜇 ⩽ 𝑢𝑗+1 𝑑𝜇 = 𝑗
𝑢𝑗+1 𝑑𝜇 ∀𝐴 ∈ 𝒜𝑗 ,
∫𝐴 ∫𝐴 ∫𝐴
𝒜
0⩽ (𝑢𝑗+1
𝑗
− 𝑢𝑗 ) 𝑑𝜇
∫ 𝒜𝑗
{𝑢𝑗+1 <𝑢𝑗 }
■■
Problem 25.7 Solution: Since both 𝜇 and 𝜈 are 𝜎-finite, we can restrict ourselves, using the technique
of the Proof of Theorem 25.2 to the case where 𝜇 and 𝜈 are finite. All we have to do is to pick an
exhaustion (𝐾𝓁 )𝓁 , 𝐾𝓁 ↑ 𝑋 such that 𝜇(𝐾𝓁 ), 𝜇(𝐾𝓁 ) < ∞ and to consider the measures 1𝐾𝓁 𝜇 and
1𝐾𝓁 𝜈 which clearly inherit the absolute continuity from 𝜇 and 𝜈.
𝜇𝑗 ≪ 𝜈𝑗 ⇐⇒ 𝜇𝑗 = 𝑢𝑗 ⋅ 𝜈𝑗
so that all the (𝑢𝑗 )𝑗 are 𝜈-integrable. Using exactly the same argument as at the beginning of the
proof of Theorem 25.2 (ii)⇒(i), we get that (𝑢𝑗 )𝑗 is even uniformly 𝜈-integrable. Finally, (𝑢𝑗 )𝑗 is a
martingale (given the measure 𝜈), since for 𝑗, 𝑗 + 1 and 𝐴 ∈ 𝒜𝑗 we have
= 𝑑𝜇𝑗 (𝐴 ∈ 𝒜𝑗 )
∫𝐴
292
Solution Manual. Last update 28th October 2021
= 𝑢𝑗 𝑑𝜈𝑗 (𝜇𝑗 = 𝑢𝑗 ⋅ 𝜈𝑗 )
∫𝐴
= 𝑢𝑗 𝑑𝜈
∫𝐴
and we conclude that 𝑢𝑗 → 𝑢∞ a.e. and in 𝐿1 (𝜈) for some limiting function 𝑢∞ which is still 𝐿1 (𝜈)
⋃
and also 𝒜∞ ∶= 𝜎( 𝑗∈N 𝒜𝑗 )-measurable. Since, by assumption, 𝒜∞ = 𝒜 , this argument shows
also that
𝜇 = 𝑢∞ ⋅ 𝜈
■■
Problem 25.8 Solution: We can assume that V𝜉𝑗 < ∞, otherwise the inequality would be trivial.
Note that the random variables 𝜉𝑗 − E𝜉𝑗 , 𝑗 = 1, 2, … , 𝑛 are still independent and, of course,
centered (= mean-zero). Thus, by Example 23.3(x) we get that
𝑘
∑
𝑀𝑘 ∶= (𝜉𝑗 − E𝜉𝑗 ) is a martingale
𝑗=1
and, because of Example 23.3(v), (|𝑀𝑘 |)𝑘 is a submartingale. Applying (25.10) in this situation
proves the claimed inequality since
where we use, for the last equality, what probabilists call Theorem of Bienaymé for the independent
random variables 𝜉𝑗 :
𝑛
∑ [ ]
E(𝑀𝑛2 ) = E (𝜉𝑗 − E𝜉𝑗 )(𝜉𝑘 − E𝜉𝑘 )
𝑗,𝑘=1
𝑛
∑ [ ] ∑ [ ] [ ]
= E (𝜉𝑗 − E𝜉𝑗 )2 + E (𝜉𝑗 − E𝜉𝑗 ) E (𝜉𝑘 − E𝜉𝑘 ) (by independence)
𝑗=𝑘=1 𝑗≠𝑘
𝑛
∑ [ ]
= E (𝜉𝑗 − E𝜉𝑗 )2
𝑗=𝑘=1
𝑛
∑ [ ]
= E 𝑀𝑗2
𝑗=1
𝑛
∑
= V𝑀𝑗 .
𝑗=1
■■
293
R.L. Schilling: Measures, Integrals & Martingales
Problem 25.10 Solution: Only the first inequality needs proof. Note that
Problem 25.11 Solution: Let (𝐴𝑘 )𝑘 ⊂ 𝒜0 be an exhausting sequence, i.e. 𝐴𝑘 ↑ 𝑋 and 𝜇(𝐴𝑘 ) < ∞.
Since (𝑢𝑗 )𝑗 is 𝐿1 -bounded, we know that
( )1∕𝑞 ( )1∕𝑞
|1𝐴𝑘 𝑢𝑗 | 𝑑𝜇 ⩽ 𝜇(𝐴𝑘 ) ⋅ ‖𝑢𝑗 ‖𝑝 ⩽ 𝑐 𝜇(𝐴𝑘 )
∫
294
Solution Manual. Last update 28th October 2021
uniformly for all 𝑗 ∈ N. This means that the martingale (1𝐴𝑘 𝑢𝑗 )𝑗 (see the solution to Problem
24.8) is 𝐿1 -bounded and we get, as in Problem 24.8 that for some unique function 𝑢
For each 𝑘 ∈ N the martingale (1𝐴𝑘 𝑢𝑗 )𝑗 is also uniformly integrable: using Hölder’s and Markov’s
inequalities we arrive at
1𝐴𝑘 |𝑢𝑗 | 𝑑𝜇 ⩽ 1 |𝑢 | 𝑑𝜇
∫{1𝐴 |𝑢 |>1𝐴𝑘 𝑅} ∫{|𝑢𝑗 |>𝑅} 𝐴𝑘 𝑗
𝑘 𝑗
( )1∕𝑞
⩽ 𝜇{|𝑢𝑗 | > 𝑅} ‖𝑢𝑗 ‖𝑝
( )1∕𝑞
1
⩽ ‖𝑢 ‖𝑝 ‖𝑢𝑗 ‖𝑝
𝑅𝑝 𝑗 𝑝
𝑐 𝑝∕𝑞+1
⩽
𝑅𝑝∕𝑞
and the latter tends, uniformly for all 𝑗, to zero as 𝑅 → ∞. Since 1𝐴𝑘 ⋅ 𝑅 is integrable, the claim
follows.
Thus, Theorem 24.6 applies and shows that for 𝑢∞ ∶= 𝑢 and every 𝑘 the family (𝑢𝑗 1𝐴𝑘 )𝑗∈N∪{∞} is
a martingale. Because of Example 23.3(vi) (|𝑢𝑗 |𝑝 1𝐴𝑘 )𝑗∈N∪{∞} is a submartingale and, therefore,
for all 𝑘 ∈ N
295
R.L. Schilling: Measures, Integrals & Martingales
= sup |𝑢𝑗 |𝑝 𝑑𝜇
𝑗 ∫
■■
= |𝑓 | 𝑑𝜆𝑛 < ∞
∫
we get from the martingale convergence theorem 24.2 that
𝑓∞ ∶= lim 𝑓𝑘
𝑘
exists almost everywhere and that 𝑓∞ ∈ 1 (ℬ). The above calculation shows, on top of that, that
for any set 𝑄 ∈ 𝒜𝑘[0]
𝑓𝑘 𝑑𝜆𝑛 = 𝑓 𝑑𝜆𝑛
∫𝑄 ∫𝑄
and
This allows us to show that (𝑓𝑘 )𝑘∈N is uniformly integrable. Indeed, fix 𝑅, take some 𝑤 ∈ 𝐿1 (𝒜1[0] )
with 𝑤 > 0 (you can construct this easily using a convergent series with steps of type 𝑧 + [0, 2−1 )𝑛 ,
𝑧 ∈ Z𝑛 and suitable weights), and take 𝑄 = {|𝑓𝑘 | > 𝑅𝑤} to get
= |𝑓 | 𝑑𝜆𝑛 + |𝑓 | 𝑑𝜆𝑛
∫{|𝑓𝑘 |>𝑅𝑤}∩{|𝑓 |>𝑅𝑤∕2} ∫{|𝑓𝑘 |>𝑅𝑤}∩{|𝑓 |⩽𝑅𝑤∕2}
⩽ |𝑓 | 𝑑𝜆𝑛 + |𝑓 | 𝑑𝜆𝑛
∫{|𝑓 |>𝑅𝑤∕2} ∫{|𝑓𝑘 |>𝑅𝑤⩾2|𝑓 |}
1
⩽ |𝑓 | 𝑑𝜆𝑛 + |𝑓 | 𝑑𝜆𝑛
∫{|𝑓 |>𝑅𝑤∕2} 2 ∫{|𝑓𝑘 |>𝑅𝑤⩾2|𝑓 |} 𝑘
1
⩽ |𝑓 | 𝑑𝜆𝑛 + |𝑓 | 𝑑𝜆𝑛
∫{|𝑓 |>𝑅𝑤∕2} 2 ∫{|𝑓𝑘 |>𝑅𝑤} 𝑘
296
Solution Manual. Last update 28th October 2021
The rhS is uniform for all 𝑘, so we can use dominated convergence and let 𝑅 → ∞ to get
This shows (Theorem 24.6) that 𝑓𝑘 → 𝑓∞ in 𝐿1 and a.e. In particular, we get for any 𝑘 ∈ N and
𝑄 ∈ 𝒜𝑘[0]
𝐿1 -limit ∀𝑚 ⩾ 𝑘, def. of 𝑓𝑘
𝑓∞ 𝑑𝜆𝑛 = lim 𝑓𝑚 𝑑𝜆𝑛 = 𝑓 𝑑𝜆𝑛
∫𝑄 𝑚⩾𝑘,𝑚→∞ ∫𝑄 ∫𝑄
⋃
Thus, ∫𝑄 𝑓∞ 𝑑𝜆𝑛 = ∫𝑄 𝑓 𝑑𝜆𝑛 for all 𝑄 ∈ 𝑘 𝒜𝑘 .
[0]
This is a ∩-stable system, so the equality also
(⋃ )
holds for 𝑄 ∈ 𝜎 𝒜
𝑘 𝑘
[0]
= ℬ.
Taking 𝑄 = {𝑓 > 𝑓∞ } and 𝑄 = {𝑓 < 𝑓∞ } – both are measurable sets since 𝑓 , 𝑓∞ are measurable
– shows 𝑓 = 𝑓∞ a.e. Thus, (𝑓𝑘 )𝑘∈N∪{∞} is UI.
■■
Problem 25.13 Solution: As one would expect, the derivative at 𝑥 turns out to be 𝑢(𝑥). This is seen
as follows (without loss of generality we can assume that 𝑦 > 𝑥):
( )
| 1 |
| 𝑢(𝑡) 𝑑𝑡 − 𝑢(𝑡) 𝑑𝑡 − 𝑢(𝑥)||
|𝑥 − 𝑦 ∫ ∫
| [𝑎,𝑥] [𝑎,𝑥] |
| 1 ( ) |
= || 𝑢(𝑡) − 𝑢(𝑥) 𝑑𝑡||
| 𝑥 − 𝑦 ∫[𝑥,𝑦] |
1 | |
⩽ |𝑢(𝑡) − 𝑢(𝑥)| 𝑑𝑡
|𝑥 − 𝑦| ∫[𝑥,𝑦] | |
1 | |
⩽ |𝑥 − 𝑦| sup |𝑢(𝑡) − 𝑢(𝑥)|
|𝑥 − 𝑦| 𝑡∈[𝑥,𝑦] | |
| |
= sup |𝑢(𝑡) − 𝑢(𝑥)|
𝑡∈[𝑥,𝑦] | |
and the last expression tends to 0 as |𝑥 − 𝑦| → 0 since 𝑢 is uniformly continuous on compact sets.
If 𝑢 is not continuous but merely of class 𝐿1 , we have to refer to Lebesgue’s differentiation theorem,
Theorem 25.20, in particular formula (25.19) which reads in our case
1
𝑢(𝑥) = lim 𝑢(𝑡) 𝑑𝑡
𝑟→0 2𝑟 ∫(𝑥−𝑟,𝑥+𝑟)
■■
Problem 25.14 Solution: We follow the hint: first we remark that by Lemma 14.14 we know that
𝑓 has at most countably many discontinuities. Since it is monotone, we also know that 𝐹 (𝑡) ∶=
𝑓 (𝑡+) = lim𝑠>𝑡,𝑠→𝑡 𝑓 (𝑠) exists and is finite for every 𝑡 and that {𝑓 ≠ 𝐹 } is at most countable (since
it is contained in the set of discontinuities of 𝑓 ), hence a Lebesgue null set.
297
R.L. Schilling: Measures, Integrals & Martingales
write 𝜇 = 𝜇 ◦ + 𝜇⊥ with 𝜇◦ ≪ 𝜆 and 𝜇 ⊥ ⊥𝜆. By Corollary 25.22 𝐷𝜇⊥ = 0 a.e. and 𝐷𝜇 ◦ exists a.e.
and we get a.e.
𝜇(𝑥 − 𝑟, 𝑥 + 𝑟) 𝜇◦ (𝑥 − 𝑟, 𝑥 + 𝑟)
𝐷𝜇(𝑥) = lim = lim +0
𝑟→0 2𝑟 𝑟→0 2𝑟
and we can set 𝑓 ′ (𝑥) = 𝐷𝜇(𝑥) which is a.e. defined. Where it is not defined, we put it equal to 0.
Now we get
= 𝑑𝜇
∫(𝑎,𝑏)
⩾ 𝑑𝜇 ◦
∫(𝑎,𝑏)
= 𝐷𝜇(𝑥) 𝜆(𝑑𝑥)
∫(𝑎,𝑏)
= 𝑓 ′ (𝑥) 𝜆(𝑑𝑥).
∫(𝑎,𝑏)
The above estimates show that we get equality if 𝑓 is continuous and also absolutely continuous
w.r.t. Lebesgue measure.
■■
Problem 25.15 Solution: Without loss of generality we may assume that 𝑓𝑗 (𝑎) = 0, otherwise we
would consider the (still increasing) functions 𝑥 → 𝑓𝑗 (𝑥) − 𝑓𝑗 (𝑎) resp. their sum 𝑥 → 𝑠(𝑥) − 𝑠(𝑎).
The derivatives are not influenced by this operation. As indicated in the hint call 𝑠𝑛 (𝑥) ∶= 𝑓1 (𝑥) +
⋯ + 𝑓𝑛 (𝑥) the 𝑛th partial sum. Clearly, 𝑠, 𝑠𝑛 are increasing
Note that the exceptional null-sets depend originally on the function 𝑠𝑛 etc. but we can consider
their (countable!!) union and get thus a universal exceptional null set 𝐸. This shows that the
formally differentiated series
∞
∑
𝑓𝑗′ (𝑥) converges for all 𝑥 ∉ 𝐸.
𝑗=1
Since the sequence of partial sums is increasing, it will be enough to check that
298
Solution Manual. Last update 28th October 2021
Since, by assumption the sequence 𝑠𝑘 (𝑥) → 𝑠(𝑥) we can choose a subsequence 𝑛𝑘 in such a way
that
Since
the series
∞ ∞
∑ ∑
(𝑠(𝑥) − 𝑠𝑛𝑘 (𝑥)) ⩽ 2−𝑘 < ∞ ∀ 𝑥 ∈ [𝑎, 𝑏].
𝑘=1 𝑘=1
By the first part of the present proof, we can differentiate this series term-by-term and get that
∞
∑
(𝑠′ (𝑥) − 𝑠′𝑛 (𝑥)) converges ∀ 𝑥 ∈ (𝑎, 𝑏) ⧵ 𝐸
𝑘
𝑘=1
299
26 Abstract Hilbert space.
Solutions to Problems 26.1–26.19
∑
Problem 26.1 Solution: If we set 𝜇 = 𝛿1 + ⋯ + 𝛿𝑛 , 𝑋 = {1, 2, … , 𝑛}, 𝒜 = 𝒫 (𝑋) or 𝜇 = 𝑗∈N 𝛿𝑗 ,
𝑋 = N, 𝒜 = 𝒫 (𝑋), respectively, we can deduce 26.5(i) and (ii) from 26.5(iii).
Let us, therefore, only verify (iii). Without loss of generality (see the complexification of a real
inner product space in Problem 26.3) we can consider the real case where 𝐿2 = 𝐿2R .
■■
(i) We prove it for the complex case—the real case is simpler. Observe that
Thus,
1
0 < (𝑢, 𝑢) = 4
‖2𝑣‖2 = ‖𝑣‖2 ⇐⇒ 𝑣 ≠ 0.
(𝑆𝑃1 ): is clear.
(iii) Using at the point (*) below the parallelogram identity, we have
301
R.L. Schilling: Measures, Integrals & Martingales
1 ( )
= 2
‖(𝑢 + 𝑤) + (𝑣 + 𝑤)‖2 − ‖(𝑢 − 𝑤) + (𝑣 − 𝑤)‖2
∗
[ ( )]
= 21 2 ‖𝑢 + 𝑤‖2 + ‖𝑣 + 𝑤‖2 − ‖𝑢 − 𝑤‖2 − ‖𝑣 − 𝑤‖2
= 4(𝑢, 𝑤) + 4(𝑣, 𝑤)
(v) By the lower triangle inequality for norms we get for any 𝑠, 𝑡 ∈ R
| |
|‖𝑡𝑣 ± 𝑤‖ − ‖𝑠𝑣 ± 𝑤‖| ⩽ ‖(𝑡𝑣 ± 𝑤) − (𝑠𝑣 ± 𝑤)‖
| |
= ‖(𝑡 − 𝑠)𝑣‖
= |𝑡 − 𝑠| ⋅ ‖𝑣‖.
This means that the maps 𝑡 → 𝑡𝑣 ± 𝑤 are continuous and so is 𝑡 → (𝑡𝑣, 𝑤) as the sum
of two continuous maps. If 𝑡 ∈ R is arbitrary, we pick a sequence (𝑞𝑗 )𝑗∈N ⊂ Q such
that lim𝑗 𝑞𝑗 = 𝑡. Then
so that
■■
Problem 26.3 Solution: This is actually a problem on complexification of inner product spaces... .
Since 𝑣 and 𝑖𝑤 are vectors in 𝑉 ⊕ 𝑖𝑉 and since ‖𝑣‖ = ‖ ± 𝑖𝑣‖, we get
( )
(𝑣, 𝑖𝑤)R = 14 ‖𝑣 + 𝑖𝑤‖2 − ‖𝑣 − 𝑖𝑤‖2
( )
= 14 ‖𝑖(𝑤 − 𝑖𝑣)‖2 − ‖(−𝑖)(𝑤 + 𝑖𝑣)‖2
( )
= 14 ‖𝑤 − 𝑖𝑣‖2 − ‖𝑤 + 𝑖𝑣‖2 (*)
= (𝑤, −𝑖𝑣)R
= −(𝑤, 𝑖𝑣)R .
302
Solution Manual. Last update 28th October 2021
In particular,
and we get
Finally, for real 𝛼, 𝛽 ∈ R the linearity property of the real scalar product shows that
(𝛼𝑢 + 𝛽𝑣, 𝑤)C = 𝛼(𝑢, 𝑤)R + 𝛽(𝑣, 𝑤)R + 𝑖𝛼(𝑢, 𝑖𝑤)R + 𝑖𝛽(𝑣, 𝑖𝑤)R
= 𝛼(𝑢, 𝑤)C + 𝛽(𝑣, 𝑤)C .
Therefore to get the general case where 𝛼, 𝛽 ∈ C we only have to consider the purely imaginary
case:
∗
(𝑖𝑣, 𝑤)C = (𝑖𝑣, 𝑤)R + 𝑖(𝑖𝑣, 𝑖𝑤)R = −(𝑣, 𝑖𝑤)R − 𝑖(𝑣, −𝑤)R
= −(𝑣, 𝑖𝑤)R + 𝑖(𝑣, 𝑤)R
( )
= 𝑖 𝑖(𝑣, 𝑖𝑤)R + (𝑣, 𝑤)R
= 𝑖(𝑣, 𝑤)C ,
where we use twice the identity (*). This shows complex linearity in the first coordinate, while
skew-linearity follows from the conjugation rule (𝑣, 𝑤)C = (𝑤, 𝑣)C .
■■
Problem 26.4 Solution: The parallelogram law (stated for 𝐿1 ) would say:
( 1 )2 ( 1 )2 ( 1 )2 ( 1 )2
|𝑢 + 𝑤| 𝑑𝑥 + |𝑢 − 𝑤| 𝑑𝑥 =2 |𝑢| 𝑑𝑥 + 2 |𝑤| 𝑑𝑥 .
∫0 ∫0 ∫0 ∫0
If 𝑢 ± 𝑤, 𝑢, 𝑤 have always only ONE sign (i.e. +ve or −ve), we could leave the modulus signs |∙|
away, and the equality would be correct! To show that there is no equality, we should therefore
choose functions where we have some sign change. We try:
303
R.L. Schilling: Measures, Integrals & Martingales
12 + ( 41 )2 = 17
16
≠ 1 = 2( 12 )2 + 2( 21 )2 .
We conclude, in particular, that 𝐿1 cannot be a Hilbert space (since in any Hilbert space the Par-
allelogram law is true....).
■■
‖𝑣 + 𝜃 𝑗 𝑤‖2 = ⟨𝑣 + 𝜃 𝑗 𝑤, 𝑣 + 𝜃 𝑗 𝑤⟩
= ⟨𝑣, 𝑣⟩ + ⟨𝑣, 𝜃 𝑗 𝑤⟩ + ⟨𝜃 𝑗 𝑤, 𝑣⟩ + ⟨𝜃 𝑗 𝑤, 𝜃 𝑗 𝑤⟩
= ⟨𝑣, 𝑣⟩ + 𝜃 −𝑗 ⟨𝑣, 𝑤⟩ + 𝜃 𝑗 ⟨𝑤, 𝑣⟩ + 𝜃 𝑗 𝜃 −𝑗 ⟨𝑤, 𝑤⟩
= ⟨𝑣, 𝑣⟩ + 𝜃 −𝑗 ⟨𝑣, 𝑤⟩ + 𝜃 𝑗 ⟨𝑤, 𝑣⟩ + ⟨𝑤, 𝑤⟩.
Therefore,
𝑛
1∑ 𝑗
𝜃 ‖𝑣 + 𝜃 𝑗 𝑤‖2
𝑛 𝑗=1
𝑛 𝑛 𝑛 𝑛
1∑ 𝑗 1∑ 1 ∑ 2𝑗 1∑ 𝑗
= 𝜃 ⟨𝑣, 𝑣⟩ + ⟨𝑣, 𝑤⟩ + 𝜃 ⟨𝑤, 𝑣⟩ + 𝜃 ⟨𝑤, 𝑤⟩
𝑛 𝑗=1 𝑛 𝑗=1 𝑛 𝑗=1 𝑛 𝑗=1
= 0 + ⟨𝑣, 𝑤⟩ + 0 + 0
304
Solution Manual. Last update 28th October 2021
(iii) Since the function 𝜙 → 𝑒𝑖𝜙 ‖𝑣 + 𝑒𝑖𝜙 𝑤‖2 is bounded and continuous, the integral exists as
a (proper) Riemann integral, and we can use any Riemann sum to approximate the integral,
see 12.6–12.12 in Chapter 12 or Corollary I.6 and Theorem I.8 of Appendix I. Before we do
that, we change variables according to 𝜓 = (𝜙 + 𝜋)∕2𝜋 so that 𝑑𝜓 = 𝑑𝜙∕2𝜋 and
1 ‖ ‖2 ‖ ‖2
𝑒𝑖𝜙 ‖𝑣 + 𝑒𝑖𝜙 𝑤‖ 𝑑𝜙 = − 𝑒2𝜋𝑖𝜓 ‖𝑣 − 𝑒2𝜋𝑖𝜓 𝑤‖ 𝑑𝜓.
2𝜋 ∫(−𝜋,𝜋] ‖ ‖ ∫(0,1] ‖ ‖
Now using equidistant Riemann sums with step 1∕𝑛 and nodes 𝜃𝑛𝑗 = 𝑒2𝜋𝑖⋅ 𝑛 ⋅𝑗 , 𝑗 = 1, 2, … , 𝑛
1
= ⟨𝑣, 𝑤⟩.
■■
‖𝑣 + 𝑤‖2 = ⟨𝑣 + 𝑤, 𝑣 + 𝑤⟩
= ⟨𝑣, 𝑣⟩ + ⟨𝑣, 𝑤⟩ + ⟨𝑤, 𝑣⟩ + ⟨𝑤, 𝑤⟩
= ‖𝑣‖2 + ⟨𝑣, 𝑤⟩ + ⟨𝑣, 𝑤⟩ + ‖𝑤‖2
= ‖𝑣‖2 + 2 Re⟨𝑣, 𝑤⟩ + ‖𝑤‖2 .
Thus
■■
Problem 26.7 Solution: Let (ℎ𝑘 )𝑘 ⊂ such that lim𝑘 ‖ℎ𝑘 − ℎ‖ = 0. By the triangle inequality
■■
|⟨𝑔, ℎ⟩ − ⟨̃
𝑔 , ℎ⟩|| ⩽ ||⟨𝑔 − 𝑔̃, ℎ⟩|| ⩽ ‖ℎ‖ ⋅ ‖̃
𝑔 − 𝑔‖
|
which proves continuity. Incidentally, this calculation shows also that, since 𝑔 → ⟨𝑔, ℎ⟩ is linear,
it would have been enough to check continuity at the point 𝑔 = 0 (think about it!).
■■
305
R.L. Schilling: Measures, Integrals & Martingales
Problem 26.9 Solution: Definiteness (𝑁1 ) and positive homogeneity (𝑁2 ) are obvious. The triangle
inequality reads in this context (𝑔, 𝑔 ′ , ℎ, ℎ′ ∈ ):
Since
( )1∕𝑝 ([ ]𝑝 [ ]𝑝 )1∕𝑝
‖𝑔 + 𝑔 ′ ‖𝑝 + ‖ℎ + ℎ′ ‖𝑝 ⩽ ‖𝑔‖‖𝑔 ′ ‖ + ‖ℎ‖ + ‖ℎ′ ‖
we can use the Minkowski inequality for sequences resp. in R2 —which reads for numbers 𝑎, 𝐴, 𝑏, 𝐵 ⩾
0
( )1∕𝑝 ( )1∕𝑝 ( 𝑝 )1∕𝑝
(𝑎 + 𝑏)𝑝 + (𝐴 + 𝐵)𝑝 ⩽ 𝑎𝑝 + 𝐴𝑝 + 𝑏 + 𝐵𝑝
Problem 26.10 Solution: For the scalar product we have for all 𝑔, 𝑔 ′ , ℎ, ℎ′ ∈ such that ‖𝑔 − 𝑔 ′ ‖2 +
‖ℎ − ℎ′ ‖2 < 1
| | [ ]1∕2
|⟨𝑔 − 𝑔 ′ , ℎ − ℎ′ ⟩| ⩽ ‖𝑔 − 𝑔 ′ ‖ ⋅ ‖ℎ − ℎ′ ‖ ⩽ ‖𝑔 − 𝑔 ′ ‖2 + ‖ℎ − ℎ′ ‖2
| |
where we use the elementary inequality
1 √
𝑎𝑏 ⩽ (𝑎2 + 𝑏2 ) ⩽ 𝑎2 + 𝑏2 ⩽ 𝑎2 + 𝑏2 .
2 ⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
if 𝑎2 +𝑏2 ⩽1
[ ]1∕2
Since (𝑔, ℎ) → ‖𝑔‖2 + ‖ℎ‖2 is a norm on × we are done.
Essentially the same calculation applies to (𝑡, ℎ) → 𝑡 ⋅ ℎ.
■■
Problem 26.11 Solution: Assume that has a countable maximal ONS, say (𝑒𝑗 )𝑗 . Then, by defini-
tion, every vector ℎ ∈ can be approximated by a sequence made up of finite linear combinations
of the (𝑒𝑗 )𝑗 :
𝑛(𝑘)
∑
ℎ𝑘 ∶= 𝛼𝑗 ⋅ 𝑒𝑗
𝑗=1
(note that 𝛼𝑗 = 0 is perfectly possible!). In view of problem 26.10 we can even assume that the 𝛼𝑗
are rational numbers. This shows that the set
{∑ 𝑛 }
∶= 𝛼𝑗 ⋅ 𝑒𝑗 ∶ 𝛼𝑗 ∈ Q, 𝑛 ∈ N
𝑗=1
306
Solution Manual. Last update 28th October 2021
Problem 26.12 Solution: Let us, first of all, show that for a closed subspace 𝐶 ⊂ we have 𝐶 =
(𝐶 ⊥ )⊥ .
Because of Lemma 26.12 we know that 𝐶 ⊂ (𝐶 ⊥ )⊥ and that 𝐶 ⊥ is itself a closed linear subspace
of . Thus,
𝐶 ⊕ 𝐶 ⊥ = = 𝐶 ⊥ ⊕ (𝐶 ⊥ )⊥ .
also. Further
{𝑤}⊥⊥ = span(𝑤)
This shows that no subsequence (𝑒𝑗 )𝑗∈⟋ can ever be a Cauchy sequence, i.e. it cannot
converge.
If ℎ ∈ we get from Bessel’s inequality 26.19 that the series
∑
|⟨𝑒𝑗 , ℎ⟩|2 ⩽ ‖ℎ‖2
𝑗
is finite, i.e. converges. Thus the sequence with elements ⟨𝑒𝑗 , ℎ⟩ must converge to 0 as
𝑗 → ∞.
307
R.L. Schilling: Measures, Integrals & Martingales
and
1 ( ) ( )
|𝑐2 (𝓁𝑗1 )| ⩽ ⇐⇒ ∃ 𝑐2 (𝓁𝑗2 ) 𝑗 ⊂ 𝑐2 (𝓁𝑗1 ) 𝑗 ∶ lim 𝑐2 (𝓁𝑗2 ) = 𝛾2
2 𝑗
and, recursively,
1 ( ) ( )
|𝑐𝑘 (𝓁𝑗𝑘−1 )| ⩽ ⇐⇒ ∃ 𝑐𝑘 (𝓁𝑗𝑘 ) 𝑗 ⊂ 𝑐𝑘 (𝓁𝑗𝑘−1 ) 𝑗 ∶ lim 𝑐𝑘 (𝓁𝑗𝑘 ) = 𝛾𝑘
𝑘 𝑗
𝑐𝑘 (𝓁𝑚𝑚 ) ←←←←←←←←←←→
← 𝛾𝑘 ∀ 𝑘 ∈ N.
𝑚→∞
308
Solution Manual. Last update 28th October 2021
(iii) 𝑅 cannot be compact since (𝑒𝑗 )𝑗 ⊂ 𝑅 does not have any convergent subsequence, see
part (i).
𝑅 is bounded since 𝑟 ∈ 𝑅 if, and only if, there is some 𝑗 ∈ N such that
1
‖𝑟 − 𝑒𝑗 ‖ ⩽ ⩽ 1.
𝑗
Thus, every 𝑟 ∈ 𝑅 is bounded by
‖𝑟‖ ⩽ ‖𝑟 − 𝑒𝑗 ‖ + ‖𝑒𝑗 ‖ ⩽ 2.
∑
𝓁
ℎ𝓁 = 𝛿𝑗 𝑒𝑗 ∈ 𝑆
𝑗=1
■■
309
R.L. Schilling: Measures, Integrals & Martingales
Finally,
‖ℎ‖ = sup |⟨𝑔, ℎ⟩| ⩽ sup |⟨𝑔, ℎ⟩| ⩽ sup ‖𝑔‖ ⋅ ‖ℎ‖ ⩽ ‖ℎ‖.
𝑔, ‖𝑔‖=1 𝑔, ‖𝑔‖⩽1 𝑔, ‖𝑔‖⩽1
⟨𝑑𝑗 , ℎ⟩ ⟨ℎ, ℎ⟩
lim = = ‖ℎ‖
𝑗 ‖𝑑𝑗 ‖ ‖ℎ‖
■■
Problem 26.15 Solution: Let 𝑥, 𝑦 ∈ span{𝑒𝑗 , 𝑗 ∈ N}. By definition, there exist numbers 𝑚, 𝑛 ∈ N
and ‘coordinates’ 𝜉1 , … , 𝜉𝑚 , 𝜂1 , … , 𝜂𝑛 ∈ K such that
𝑚 𝑛
∑ ∑
𝑥= 𝜉𝑗 𝑒𝑗 and 𝑦= 𝜂𝑘 𝑒𝑘 .
𝑗=1 𝑘=1
𝜉𝑚+1 ∶= 0, … , 𝜉𝑛 ∶= 0
■■
310
Solution Manual. Last update 28th October 2021
∑∞
(i) Since 2
𝑗=1 𝑎𝑗 = ∞ there is some number 𝑗1 ∈ N such that
𝑗1
∑
𝑎2𝑗 > 1.
𝑗=1
∑
Since the remaining tail of the series 𝑗>𝑗1 𝑎2𝑗 = ∞ we can construct recursively a
strictly increasing sequence (𝑗𝑘 )𝑘∈N0 ⊂ N, 𝑗0 ∶= 1, such that
∑
𝑎2𝑗 > 1 where 𝐽𝑘 ∶= (𝑗𝑘 , 𝑗𝑘+1 ] ∩ N.
𝑗∈𝐽𝑘
1
𝛾𝑘 ∶= √ .
∑ 2
𝑘 𝑎
𝑗∈𝐽𝑘 𝑗
Then
∑ ∑∑
𝑏2𝑗 = 𝛾𝑘2 𝑎2𝑗
𝑗 𝑘 𝑗∈𝐽𝑘
∑ ∑
= 𝛾𝑘2 𝑎2𝑗
𝑘 𝑗∈𝐽𝑘
∑
∑ 𝑗∈𝐽𝑘 𝑎2𝑗
= ∑
𝑘 𝑘2 𝑗∈𝐽𝑘 𝑎2𝑗
∑ 1
= 2
< ∞.
𝑘 𝑘
Moreover, since
∑
𝑗∈𝐽𝑘 𝑎2𝑗
√∑ ⩾ 1,
2
𝑗∈𝐽𝑘 𝑎𝑗
we get
∑ ∑∑
𝑎𝑗 𝑏𝑗 = 𝛾𝑘 𝑎2𝑗
𝑗 𝑘 𝑗∈𝐽𝑘
∑ ∑
= 𝛾𝑘 𝑎2𝑗
𝑘 𝑗∈𝐽𝑘
∑
∑1 𝑗∈𝐽𝑘 𝑎2𝑗
= √∑
𝑘
𝑘
𝑗∈𝐽𝑘 𝑎2𝑗
∑1
⩾ = ∞.
𝑘
𝑘
(iii) We want to show (note that we renamed 𝛽 ∶= 𝑎 and 𝛼 ∶= 𝑏 for notational reasons) that
for any sequence 𝛼 = (𝛼𝑗 )𝑗 we have:
∀ 𝛽 ∈ 𝓁 2 ∶ ⟨𝛼, 𝛽⟩ < ∞ ⇐⇒ 𝛼 ∈ 𝓁 2 .
311
R.L. Schilling: Measures, Integrals & Martingales
∑
Assume, to the contrary, that 𝛼 ∉ 𝓁 2 . Then 𝑗 𝛼𝑗2 = ∞ and, by part (i), we can find
a sequence of 𝑗𝑘 with the properties described in (i). Because of part (ii) there is a
sequence 𝛽 = (𝛽𝑗 )𝑗 ∈ 𝓁 2 such that the scalar product ⟨𝛼, 𝛽⟩ = ∞. This contradicts our
assumption, i.e. 𝛼 should have been in 𝓁 2 in the first place.
(iv) Since, by Theorem 26.24 every separable Hilbert space has a basis (𝑒𝑗 )𝑗∈N ⊂ , we
can identify ℎ ∈ with the sequence of ‘coordinates’ (⟨ℎ, 𝑒𝑗 ⟩)𝑗∈N and it is clear that
(iii) implies (iv).
■■
(i) Since 𝑃 2 = 𝑃 is obvious by the uniqueness of the minimizing element, this part follows
already from Remark 26.15.
∀ ℎ ∈ ∶ ⟨𝑢, ℎ⟩ = ⟨𝑣, ℎ⟩ ⇐⇒ 𝑢 = 𝑣.
so that 𝑢 = 𝑣.
‖𝑃 ℎ‖2 = ⟨𝑃 ℎ, 𝑃 ℎ⟩ = ⟨𝑃 2 ℎ, ℎ⟩ = ⟨𝑃 ℎ, ℎ⟩ ⩽ ‖𝑃 ℎ‖ ⋅ ‖ℎ‖
Closedness of 𝑃 (): Note that 𝑓 ∈ 𝑃 () if, and only if, 𝑓 = 𝑃 ℎ for some ℎ ∈ .
Since 𝑃 2 = 𝑃 we get
𝑓 = 𝑃 ℎ ⇐⇒ 𝑓 − 𝑃 ℎ = 0
⇐⇒ 𝑓 − 𝑃 2 ℎ = 0
⇐⇒ 𝑓 − 𝑃 𝑓 = 0
⇐⇒ 𝑓 ∈ (id −𝑃 )−1 ({0})
312
Solution Manual. Last update 28th October 2021
and since 𝑃 is continuous and {0} is a closed set, (id −𝑃 )−1 ({0}) is closed and the above
line shows 𝑃 () = (id −𝑃 )−1 ({0}) is closed.
⟨𝑃 ℎ − ℎ, 𝑓 ⟩ = ⟨𝑃 ℎ, 𝑓 ⟩ − ⟨ℎ, 𝑓 ⟩
= ⟨ℎ, 𝑃 𝑓 ⟩ − ⟨ℎ, 𝑓 ⟩
= ⟨ℎ, 𝑓 ⟩ − ⟨ℎ, 𝑓 ⟩ = 0.
ℎ ∈ 𝒩 ⊥ ⇐⇒ 𝑦 = 𝑃 ℎ − ℎ ∈ 𝒩
⇐⇒ 𝑃 ℎ = ℎ + 𝑦 with ℎ⊥𝑦. (*)
Thus,
We conclude that
ℎ ∈ 𝒩 ⊥ ⇐⇒ 𝑃 ℎ − ℎ = 0 ⇐⇒ 𝑃 ℎ = ℎ ⇐⇒ ℎ ∈ 𝑃 ()
thus
ℎ = 𝑃 ℎ = 𝑥⊥ ⇐⇒ 𝑃 () ⊂ 𝒩 ⊥ .
⟨𝑃 ℎ − ℎ, 𝑃 ℎ⟩ = 0
■■
313
R.L. Schilling: Measures, Integrals & Martingales
(ii) Let 𝑢 ∈ 𝐿2 (𝜇) and set 𝑤𝑗 ∶= 𝑤1𝐴1 ∪⋯∪𝐴𝑗 . Since (𝐴1 ∪ ⋯ ∪ 𝐴𝑗 )𝑐 = 𝐴𝑐1 ∩ ⋯ ∩ 𝐴𝑐𝑗 ↓ ∅
we get by dominated convergence
‖𝑢 − 𝑤𝑗 ‖22 = 𝑢2 𝑑𝜇 = 𝑢2 𝑑𝜇 ←←←←←←←←←→
← 0.
∫(𝐴1 ∪⋯∪𝐴𝑗 )𝑐 ∫𝐴𝑐 ∩⋯∩𝐴𝑐 𝑗→∞
1 𝑗
■■
where 𝐴0 ∶= (𝐴1 ∪⋯∪𝐴𝑛 )𝑐 and 1∕∞ ∶= 0. This follows simply from the consideration
∑
that 𝑢𝑛 , as an element of 𝐿2 (𝒜𝑛 ), must be of the form 𝑛𝑗=0 𝛼𝑗 ⋅ 1𝐴𝑗 while the 𝛼𝑗 ’s are
calculated as
Clearly this is a linear map and 𝑢𝑛 ∈ 𝐿2 (𝒜𝑛 ). Orthogonality follows because all the
𝐴0 , … , 𝐴𝑛 are disjoint so that
⟨ 𝑛
∑ 𝑛
∑ ⟩
⟨𝑢 − 𝑢𝑛 , 𝑢𝑛 ⟩ = 𝑢 − 𝛼𝑗 1𝐴𝑗 , 𝛼𝑘 1𝐴𝑘
𝑗=0 𝑘=0
314
Solution Manual. Last update 28th October 2021
𝑛
∑
= (𝑢 − 𝛼𝑗 )𝛼𝑗 𝑑𝜇
∫
𝑗=0 𝐴𝑗
𝑛 ( )
∑
= 𝛼𝑗 𝑢 𝑑𝜇 − 𝜇(𝐴𝑗 )𝛼𝑗2
∫𝐴𝑗
𝑗=0
𝑛
∑
= 0 = 0.
𝑗=0
(iii) We have
{ 𝑛 𝑛 }
∑ ∑
2 2
𝐿 (𝒜𝑛 ) = ⊥
𝑢− 𝛼𝑗 1𝐴𝑗 = (𝑢 − 𝛼𝑗 )1𝐴𝑗 ∶ 𝑢 ∈ 𝐿 (𝜇)
𝑗=0 𝑗=0
𝐸 𝒜𝑛 𝑢 𝑑𝜇 = 𝐸 𝒜𝑛+1 𝑢 𝑑𝜇, ∀ 𝐴0 , 𝐴1 , … , 𝐴𝑛 .
∫𝐴 𝑗 ∫𝐴𝑗
Thus
𝐸 𝒜𝑛+1 𝑢 𝑑𝜇 = 𝑢 𝑑𝜇 ∀ 𝑗 = 1, 2, … , 𝑛.
∫𝐴 𝑗 ∫𝐴𝑗
𝐸 𝒜𝑛+1 𝑢 𝑑𝜇
∫𝐴0
( ∫𝐴 𝑢 𝑑𝜇 ∫𝐴 𝑢 𝑑𝜇 )
0 ⧵𝐴𝑛+1
= + 1𝐴0 ⧵𝐴𝑛+1
𝑛+1
1𝐴𝑛+1 𝑑𝜇
∫𝐴0 𝜇(𝐴𝑛+1 ) 𝜇(𝐴0 ⧵ 𝐴𝑛+1 )
∫𝐴 𝑢 𝑑𝜇 ∫𝐴 𝑢 𝑑𝜇
0 ⧵𝐴𝑛+1
= 𝜇(𝐴0 ∩ 𝐴𝑛+1 ) + 𝜇(𝐴0 ⧵ 𝐴𝑛+1 )
𝑛+1
= 𝑢 𝑑𝜇 + 𝑢 𝑑𝜇
∫𝐴𝑛+1 ∫𝐴0 ⧵𝐴𝑛+1
= 𝑢 𝑑𝜇.
∫𝐴0
Remark. It is, actually, better to show that for 𝑢𝑛 ∶= 𝐸 𝒜𝑛 𝑢 the sequence (𝑢2𝑛 )𝑛 is a
sub-Martingale. (The advantage of this is that we do not have to assume that 𝑢 ∈ 𝐿1
and that 𝑢 ∈ 𝐿2 is indeed enough....). O.k.:
315
R.L. Schilling: Measures, Integrals & Martingales
We have
and
𝑛
∑ 𝑑𝜇 𝑑𝜇
𝐸 𝒜𝑛 𝑢 = 1𝐴𝑗 𝑢 + 1𝐴𝑛 𝑢
𝑗=1
∫𝐴 𝑗 𝜇(𝐴𝑗 ) 0 ∫ 𝑛
𝐴 𝜇(𝐴𝑛0 )
0
𝑛+1
∑ 𝑑𝜇 𝑑𝜇
𝐸 𝒜𝑛+1 𝑢 = 1𝐴𝑗 𝑢 + 1𝐴𝑛+1 𝑢
𝑗=1
∫𝐴 𝑗 𝜇(𝐴𝑗 ) 0 ∫𝐴𝑛+1 𝜇(𝐴𝑛+1 )
0 0
with the convention that 1∕∞ = 0. Since the 𝐴𝑗 ’s are mutually disjoint,
𝑛 [ ]2 [ ]2
( 𝒜 )2 ∑ 𝑑𝜇 𝑑𝜇
𝐸 𝑢 =
𝑛 1𝐴𝑗 𝑢 + 1𝐴𝑛 𝑢
∫𝐴𝑗 𝜇(𝐴𝑗 ) 0 ∫𝐴𝑛 𝜇(𝐴𝑛0 )
𝑗=1 0
𝑛+1 [ ]2 [ ]2
( )2 ∑ 𝑑𝜇 𝑑𝜇
𝐸 𝒜𝑛+1
𝑢 = 1𝐴𝑗 𝑢 + 1𝐴𝑛+1 𝑢 .
∫𝐴𝑗 𝜇(𝐴𝑗 ) 0 ∫𝐴𝑛+1 𝜇(𝐴𝑛+1 )
𝑗=1 0 0
( )2 ( )2
We have to show that 𝐸 𝒜𝑛 𝑢 = 𝑢2𝑛 ⩽ 𝑢2𝑛+1 = 𝐸 𝒜𝑛+1 𝑢 . If 𝜇(𝐴𝑛+1
0
) = ∞ this follows
trivially since in this case
𝑛 [ ]2
( 𝒜 )2 ∑ 𝑑𝜇
𝐸 𝑢 =
𝑛 1𝐴𝑗 𝑢
∫𝐴𝑗 𝜇(𝐴𝑗 )
𝑗=1
𝑛+1 [ ]2
( 𝒜 )2 ∑ 𝑑𝜇
𝐸 𝑛+1 𝑢 = 1𝐴𝑗 𝑢 .
∫𝐴𝑗 𝜇(𝐴𝑗 )
𝑗=1
If 𝜇(𝐴𝑛+1
0
) < ∞ we get
( 𝒜 )2 ( 𝒜 )2
𝐸 𝑛 𝑢 − 𝐸 𝑛+1 𝑢
[ ]2 [ ]2
𝑑𝜇 𝑑𝜇
= 1𝐴𝑛 𝑢 − 1 𝐴𝑛+1 ∫ 𝑢
𝐴0 𝜇(𝐴0 ) 𝐴𝑛+1 𝜇(𝐴𝑛+1 )
0 ∫ 𝑛 𝑛
[ ]2
𝑑𝜇
+ 1𝐴𝑛+1 𝑢
0 ∫𝐴𝑛+1 𝜇(𝐴𝑛+1 )
0 0
([ ]2 [ ]2 )
𝑑𝜇 𝑑𝜇
= 1𝐴𝑛+1 𝑢 − 𝑢
∫𝐴𝑛+1 𝜇(𝐴𝑛0 ) ∫𝐴𝑛+1 𝜇(𝐴𝑛+1 )
([ ]2 [ ]2 )
𝑑𝜇 𝑑𝜇
+ 1𝐴𝑛+1 𝑢 − 𝑢
0 ∫𝐴𝑛+1 𝜇(𝐴𝑛0 ) ∫𝐴𝑛+1 𝜇(𝐴𝑛+1 )
0 0 0
316
Solution Manual. Last update 28th October 2021
(v) Set 𝑢𝑛 ∶= 𝐸 𝒜𝑛 𝑢. Since (𝑢𝑛 )𝑛 is a martingale, 𝑢2𝑛 is a submartingale. In fact, (𝑢2𝑛 )𝑛 is even
uniformly integrable. For this we remark that
𝑛
∑ 𝜇(𝑑𝑥)
𝑢𝑛 = 1𝐴𝑗 𝑢(𝑥) + 1𝐴𝑛 𝑢(𝑥) 𝑓 𝑟𝑎𝑐𝜇(𝑑𝑥)𝜇(𝐴𝑛0 )
𝑗=1
∫𝐴𝑗 𝜇(𝐴𝑗 ) 0 ∫ 𝑛
𝐴 0
is in 𝐿2 (𝒜∞ ). Only integrability is a problem: since the 𝐴𝑗 ’s are mutually disjoint, the
square of the series defining 𝑣 factorizes, i.e.
( ∞ )2
∑ 𝜇(𝑑𝑥)
2
𝑣 (𝑦) 𝜇(𝑑𝑦) = 1𝐴𝑗 (𝑦) 𝑢(𝑥) 𝜇(𝑑𝑦)
∫ ∫
𝑗=1
∫𝐴𝑗 𝜇(𝐴𝑗 )
∞ ( )2
∑ 𝜇(𝑑𝑥)
= 1𝐴𝑗 (𝑦) 𝜇(𝑑𝑦) 𝑢(𝑥)
𝑗=1
∫ ∫𝐴𝑗 𝜇(𝐴𝑗 )
∞
∑ 𝜇(𝑑𝑥)
⩽ 1𝐴𝑗 (𝑦) 𝜇(𝑑𝑦) 𝑢2 (𝑥)
𝑗=1
∫ ∫𝐴𝑗 𝜇(𝐴𝑗 )
∞
∑
= 𝑢2 (𝑥) 𝜇(𝑑𝑥)
∫𝐴𝑗
𝑗=1
= 𝑢2 (𝑥) 𝜇(𝑑𝑥)
∫
where we use Beppo Levi’s theorem (twice) and Jensen’s inequality. In fact,
𝑣 = 𝐸 𝒜∞ 𝑢.
𝑢2𝑛 𝑑𝜇 ⩽ 𝑢2𝑛 𝑑𝜇
∫{𝑢2 >(2𝑣)2 } ∫𝐴 𝑛
𝑛 0
⩽ 𝑢2 𝑑𝜇
∫𝐴 𝑛
0
← 0
←←←←←←←←←→
𝑛→∞
317
R.L. Schilling: Measures, Integrals & Martingales
i.e. 𝑣 = 𝑢∞ on all sets of the ∩-stable generator of 𝒜∞ which can easily be extended to
contain an exhausting sequence 𝐴1 ⊍ ⋯ ⊍ 𝐴𝑛 of sets of finite 𝜇-measure.
(if 𝜇(𝐴𝑛0 ) = ∞, then 𝑞0 = 0) is countable and dense in 𝐷 so that the claim follows.
■■
318
27 Conditional expectations.
Solutions to Problems 27.1–27.19
Eℋ E𝒢 𝑢 = Eℋ 𝑢.
E𝒢 Eℋ 𝑢 = Eℋ 𝑢.
■■
Problem 27.2 Solution: Note that by the Markov inequality 𝜇{𝑢 > 1} ⩽ ∫ 𝑢2 𝑑𝜇 < ∞, i.e. 𝑢1{𝑢>1}
is an integrable function (use Cauchy-Schwarz).
We have
(*) assumption
1𝜇{𝑢 > 1} = 1 𝑑𝜇 < 𝑢 𝑑𝜇 ⩽ 𝜇{𝑢 > 1}.
∫{𝑢>1} ∫{𝑢>1}
In the step marked (*) we really (!) need that 𝜇{𝑢 > 1} > 0 — otherwise we could not get a strict
inequality. Thus, 𝜇{𝑢 > 1} < 𝜇{𝑢 > 1} which is a contradiciton. Therefore, 𝜇{𝑢 > 1} = 0 and
we have 𝑢 ⩽ 1 a.e.
If you are unhappy with strict inequalities, you can extend the argument as follows: By assumption
⋃
𝜇{𝑢 > 1} > 0. Since {𝑢 > 1} = 𝑛⩾1 {𝑢 ⩾ 1+1∕𝑛}, there is some 𝑁 such that 𝜇{𝑢 ⩾ 1+1∕𝑛} > 0
for all 𝑛 ⩾ 𝑁 — use a continuity of measure argument. Now we get for all 𝑛 ⩾ 𝑁
{ }
1 𝑑𝜇 = 𝜇 𝑢 ⩾ 1 + 1𝑛
∫{𝑢⩾1+ 1 }
𝑛
( ) { }
< 1 + 1𝑛 𝜇 𝑢 ⩾ 1 + 1𝑛
( )
= { 1 + 1𝑛 𝑑𝜇
∫ 𝑢⩾1+ 1 }
𝑛
⩽ { 𝑢 𝑑𝜇.
∫ 𝑢⩾1+ 1 }
𝑛
Observe that
∞
∑
1 𝑑𝜇 = 1 𝑑𝜇 + { 1 𝑑𝜇
∫{𝑢>1} ∫{1+1∕𝑛⩽𝑢<1+1∕(𝑛−1)} ∫ 𝑢⩾1+ 1 }
𝑛=𝑁+1 𝑁
319
R.L. Schilling: Measures, Integrals & Martingales
∞
∑
𝑢 𝑑𝜇 + { 1 𝑑𝜇
∫{1+1∕𝑛⩽𝑢<1+1∕(𝑛−1)} ∫ 𝑢⩾1+ 1 }
⩽
𝑛=𝑁+1 𝑁
∞
∑
< 𝑢 𝑑𝜇 + { 𝑢 𝑑𝜇
∫{1+1∕𝑛⩽𝑢<1+1∕(𝑛−1)} ∫ 𝑢⩾1+ 1 }
𝑛=𝑁+1 𝑁
= 𝑢 𝑑𝜇.
∫{𝑢>1}
With our assumption we thus get the contradiction 𝜇{𝑢 > 1} < 𝜇{𝑢 > 1}.
(𝑢 − 1) 𝑑𝜇 ⩽ 0.
∫{𝑢>1}
(𝑢 − 1) 𝑑𝜇 ⩾ 0.
∫{𝑢>1}
Therefore, ∫{𝑢>1} (𝑢 − 1) 𝑑𝜇 = 0 and we see that (𝑢 − 1)1{𝑢>1} = 0 a.e., hence 1{𝑢>1} = 0 a.e.
■■
Problem 27.3 Solution: Note that, since E𝒢 is (currently...) only defined for 𝐿2 -functions the prob-
lem implicitly requires that 𝑓 ∈ 𝐿2 (𝒜 , 𝜇). (A look at the next section reveals that this is not
really necessary...). Below we will write ⟨∙, ∙⟩𝐿2 (𝜇) resp. ⟨∙, ∙⟩𝐿2 (𝜈) to indicate which scalar product
is meant.
We begin with a general consideration: Let 𝑢, 𝑤 be functions such that 𝑢2 , 𝑣2 ∈ 𝐿2 (𝜇). Then we
have |𝑢 ⋅ 𝑤| ⩽ 12 (𝑢2 + 𝑤2 ) ∈ 𝐿2 (𝜇) and, using again the elementary inequality
𝑥2 𝑦2
|𝑥𝑦| ⩽ +
2 2
√ √
for 𝑥 = |𝑢|∕ E𝒢𝜇 (𝑢2 ) and 𝑦 = |𝑤|∕ E𝒢𝜇 (𝑤2 ) we conclude that on 𝐺𝑛 ∶= {E𝒢𝜇 (𝑢2 ) > 1
𝑛
} ∩
1
{E𝒢𝜇 (𝑤2 ) > 𝑛
}
[ ]
|𝑢| ⋅ |𝑤| 𝑢2 𝑤2
√ √ 1𝐺𝑛 ⩽ + 1𝐺𝑛 .
E𝒢𝜇 (𝑤2 ) E𝒢𝜇 (𝑤2 ) 2 E𝒢𝜇 (𝑢2 ) 2 E𝒢𝜇 (𝑤2 )
Multiplying through with the denominator of the lhS and letting 𝑛 → ∞ gives
( ) √ √
| 𝒢 |
|E𝜇 (𝑢𝑤)|1𝐺∗ ⩽ E𝒢𝜇 |𝑢𝑤| 1𝐺∗ ⩽ E𝒢𝜇 (𝑢2 ) E𝒢𝜇 (𝑤2 )
| |
on the set 𝐺∗ ∶= 𝐺𝑢 ∩ 𝐺𝑤 ∶= {E𝒢𝜇 𝑢2 > 0} ∩ {E𝒢𝜇 𝑤2 > 0}.
320
Solution Manual. Last update 28th October 2021
(i) Set 𝐺∗ ∶= {E𝒢𝜇 𝑓 > 0} and 𝐺𝑛 ∶= {E𝒢𝜇 𝑓 > 1𝑛 }. Clearly, using the Markov inequality,
E𝒢𝜇 (𝑓 𝑢)
𝑃 𝑢 ∶= 1𝐺 ∗ .
E𝒢𝜇 𝑓
( )
Let us show that 𝑃 ∈ 𝐿2 (𝜈). Set 𝐺√𝑓 𝑢 ∶= {E𝒢𝜇 𝑓 ⋅ 𝑢2 > 0}. Then, for bounded
𝑢 ∈ 𝐿2 (𝜈)
‖ E𝜇 (𝑓 𝑢)
𝒢
‖2
‖ 1 ‖
‖ E𝒢 𝑓 𝐺 ∗ ∩𝐺√ ∩𝐺√
𝑓‖ 2
‖ 𝜇 𝑓𝑢
‖𝐿 (𝜈)
[ 𝒢 ]2
E𝜇 (𝑓 𝑢)
= 𝑑𝜈
∫𝐺∗ ∩𝐺√ ∩𝐺√ [E𝒢 𝑓 ]2
𝑓𝑢 𝑓 𝜇
[ 𝒢 ]2
E𝜇 (𝑓 𝑢)
= 𝑓 𝑑𝜇
∫𝐺∗ ∩𝐺√ ∩𝐺√ [E𝒢 𝑓 ]2
𝑓𝑢 𝑓 𝜇
[ 𝒢 ]2
E𝜇 (𝑓 𝑢)
= E𝒢𝜇 𝑓 𝑑𝜇
∫𝐺∗ ∩𝐺√ ∩𝐺√ [E𝒢 𝑓 ]2
𝑓𝑢 𝑓 𝜇
[ 𝒢 ]2
E𝜇 (𝑓 𝑢)
= 𝑑𝜇
∫𝐺∗ ∩𝐺√ ∩𝐺√ E𝒢𝜇 𝑓
𝑓𝑢 𝑓
[ 𝒢 [√ √ ]]2
E𝜇 𝑓 ( 𝑓 𝑢)
= 𝑑𝜇
∫𝐺∗ ∩𝐺√ ∩𝐺√ E𝒢𝜇 𝑓
𝑓𝑢 𝑓
[ ]
E𝒢𝜇 𝑓 ⋅ E𝒢𝜇 𝑓 𝑢2
⩽ 𝑑𝜇
∫𝐺∗ ∩𝐺√ ∩𝐺√ E𝒢𝜇 𝑓
𝑓𝑢 𝑓
[ ]
= E𝒢𝜇 𝑓 𝑢2 𝑑𝜇
∫𝐺∗ ∩𝐺√ ∩𝐺√𝑓
𝑓𝑢
[ ]
= sup 1𝐺𝑛 ∩𝐺√ 𝒢
∩𝐺√𝑓 E𝜇 𝑓 𝑢2 𝑑𝜇
𝑛 ∫ 𝑓𝑢
321
R.L. Schilling: Measures, Integrals & Martingales
‖√ ‖2
⩽ 𝑓 𝑢2 𝑑𝜇 = ‖ 𝑓 𝑢‖ 2 = ‖𝑢‖2𝐿2 (𝜈) < ∞.
∫ ‖ ‖𝐿 (𝜇)
E𝒢𝜇 (𝑓 𝑢) 𝑑𝜇
∫𝐺𝑛 ∩{𝑓 <𝑛}∩{E𝒢 (𝑓 𝑢2 )=0}
= 𝑓 𝑢 𝑑𝜇
∫𝐺𝑛 ∩{E𝒢 (√𝑓 𝑢)=0}
√ √
⩽ 𝑓 𝑑𝜇 𝑓 𝑢2 𝑑𝜇
∫𝐺𝑛 ∩{𝑓 <𝑛} ∫𝐺𝑛 ∩{E𝒢 (𝑓 𝑢2 )=0}
√ √
= 𝑓 𝑑𝜇 E𝒢𝜇 𝑓 𝑢2 𝑑𝜇
∫𝐺𝑛 ∩{𝑓 <𝑛} ∫𝐺𝑛 ∩{E𝒢 (𝑓 𝑢2 )=0}
=0
for all bounded 𝑢 ∈ 𝐿2 (𝜈), hence – through extension by continuity – for all 𝑢 ∈ 𝐿2 (𝜈).
(iii) Since
⟨ ⟩
𝑢 − 𝑃 𝑢, 𝑃 𝑢 𝐿2 (𝜈)
⟨ ⟩
= 𝑓 𝑢 − 𝑓 𝑃 𝑢, 𝑃 𝑢 𝐿2 (𝜇)
⟨ E𝒢𝜇 (𝑓 𝑢) E𝒢𝜇 (𝑓 𝑢) ⟩
= 𝑓𝑢 − 𝑓 1 𝐺 ∗ , 1 𝐺 ∗
E𝒢𝜇 𝑓 E𝒢𝜇 𝑓 𝐿2 (𝜇)
⟨ [ E𝒢𝜇 (𝑓 𝑢) ] E𝒢𝜇 (𝑓 𝑢) ⟩
= E𝒢𝜇 𝑓 𝑢 − 𝑓 1 𝐺 ∗ , 1𝐺 ∗
E𝒢𝜇 𝑓 E𝒢𝜇 𝑓 𝐿2 (𝜇)
⟨ [ E𝒢𝜇 (𝑓 𝑢) ] E𝒢𝜇 (𝑓 𝑢) ⟩
= E𝒢𝜇 (𝑓 𝑢) − E𝒢𝜇 𝑓 1 𝐺 ∗ , 1 𝐺 ∗
E𝒢𝜇 𝑓 E𝒢𝜇 𝑓 𝐿2 (𝜇)
⟨ E𝒢𝜇 (𝑓 𝑢) E𝒢𝜇 (𝑓 𝑢) ⟩
= E𝒢𝜇 (𝑓 𝑢) − E𝒢𝜇 (𝑓 ) 1𝐺 ∗ , 1𝐺 ∗
E𝒢𝜇 𝑓 E𝒢𝜇 𝑓 𝐿2 (𝜇)
⟨ E𝒢𝜇 (𝑓 𝑢) ⟩
= E𝒢𝜇 (𝑓 𝑢) − E𝒢𝜇 (𝑓 𝑢)1𝐺∗ , 1 𝐺∗
E𝒢𝜇 𝑓 𝐿2 (𝜇)
=0
which shows that 𝑃 is the (uniquely determined) orthogonal projection onto 𝐿2 (𝜈, 𝒢 ),
i.e. 𝑃 = E𝒢𝜈 .
322
Solution Manual. Last update 28th October 2021
(iv) The condition that 𝑓 1𝐺∗ is 𝒢 -measurable will do. Indeed, since 𝐺∗ ∈ 𝒢 :
E𝒢𝜇 𝑓 = E𝒢𝜈 𝑓
E𝒢𝜇 (𝑓 2 ) ( 𝒢 )2
E𝒢𝜇 𝑓 = 1{E𝒢 𝑓 >0} ⇐⇒ E𝜇 𝑓 = E𝒢𝜇 (𝑓 2 )1{E𝒢 𝑓 >0}
E𝒢𝜇 𝑓 𝜇 𝜇
( )2
⇐⇒ E𝒢𝜇 𝑓 = E𝒢𝜇 (𝑓 2 ).
Thus,
[( )2 ]
E𝒢𝜇 𝑓 − E𝒢𝜇 𝑓 = 0,
⋃
which means that on the set 𝐺∗ = 𝑛 𝐺𝑛 with 𝜇(𝐺𝑛 ) < ∞, see above,
■■
Problem 27.4 Solution: Since 𝒢 = {𝐺1 , … , 𝐺𝑛 } such that the 𝐺𝑗 ’s form a mutually disjoint partition
of the whole space 𝑋, we have
{ 𝑛 }
∑
2
𝐿 (𝒢 ) = 𝛼𝑗 1𝐺𝑗 ∶ 𝛼𝑗 ∈ R .
𝑗=1
It is, therefore, enough to determine the values of the 𝛼𝑗 . Using the symmetry and idempotency of
the conditional expectation we get for 𝑘 ∈ {1, 2, … , 𝑛}
■■
323
R.L. Schilling: Measures, Integrals & Martingales
Problem 27.5 Solution: We follow the hint. Let 𝑢 ∈ 𝐿𝑝 (𝜇) and define 𝑢𝑛 = [(−𝑛) ∨ 𝑢 ∧ 𝑛]1{|𝑢|⩾1∕𝑛} .
Clearly, 𝑢𝑛 is bounded, and by the Markov inequality (11.4)
Since 𝑢𝑛 → 𝑢 a.e., dominated convergence (use the majorant |𝑢|𝑝 ) shows that 𝑢𝑛 → 𝑢 in 𝐿𝑝 . Thus,
we see as in the remark before Theorem 27.5 that (𝑇 𝑢𝑛 )𝑛∈N is a Cauchy sequence in 𝐿𝑝 (𝜇), i.e. the
limit 𝐿𝑝 -lim𝑛 𝑇 𝑢𝑛 exists. If (𝑤𝑛 )𝑛 is a further approximating sequence such that 𝑤𝑛 → 𝑢 in 𝐿𝑝 (𝜇),
we get
which shows that lim𝑛 𝑇 𝑢𝑛 = lim𝑛 𝑇 𝑤𝑛 , i.e. 𝑇̃ 𝑢 ∶= lim𝑛 𝑇 𝑢𝑛 (as an 𝐿𝑝 -limit) is well-defined since
it is independent of the approximating sequence. Linearity is clear from the linearity of the limit.
Assume now that 0 ⩽ 𝑢𝑛 ↑ 𝑢 where 𝑢𝑛 ∈ 𝐿𝑝 (𝜇) ∩ 𝐿2 (𝜇). By the first part, 𝑇̃ 𝑢 = lim𝑛 𝑇 𝑢𝑛 in 𝐿𝑝 ,
so there is a subsequence such that 𝑇̃ 𝑢 = lim𝑘 𝑇 𝑢𝑛 a.e. Because of monotonicity we have
𝑘
So,
■■
Problem 27.6 Solution: Let 𝐺𝑢 ∶= {E𝒢 |𝑢|𝑝 > 0}, 𝐺𝑤 ∶= {E𝒢 |𝑤|𝑞 > 0} and 𝐺 ∶= 𝐺𝑢 ∩ 𝐺𝑤 .
Following the hint we get
|𝑢| |𝑤| |𝑢|𝑝 |𝑢|𝑞
[ ]1∕𝑝 [ ⩽ 1 +
]1∕𝑞 𝐺 𝑝 E𝒢 (|𝑢|𝑝 ) 𝐺 𝑞 E𝒢 (|𝑤|𝑞 ) 1𝐺
1
E𝒢 (|𝑢|𝑝 ) E𝒢 (|𝑤|𝑞 )
Since 1𝐺 is bounded and 𝒢 -measurable, we can apply E𝒢 on both sides of the above inequality
and get
324
Solution Manual. Last update 28th October 2021
=0
■■
Problem 27.7 Solution: In this problem it is helpful to keep the distinction between E𝒢 defined on
𝐿2 (𝒜 ) and the extension 𝐸 𝒢 defined on 𝐿𝒢 (𝒜 ).
Since 𝜇|𝒜 is 𝜎-finite we can find an exhausting sequence of sets 𝐴𝑛 ↑ 𝑋 with 𝜇(𝐴𝑛 ) < ∞. Setting
( ) ( )
for 𝑢, 𝑤 ∈ 𝐿𝒢 (𝒜 ) with 𝑢𝐸 𝒢 𝑤 ∈ 𝐿1 (𝒜 ) 𝑢𝑛 ∶= (−𝑛) ∨ 𝑢 ∧ 𝑛 ⋅ 1𝐴𝑛 and 𝑤𝑛 ∶= (−𝑛) ∨ 𝑤 ∧ 𝑛 ⋅ 1𝐴𝑛
we have found approximating sequences such that 𝑢𝑛 , 𝑤𝑛 ∈ 𝐿1 (𝒜 ) ∩ 𝐿∞ (𝒜 ) and, in particular,
∈ 𝐿2 (𝒜 ).
(iii): For 𝑢, 𝑤 ⩾ 0 we find by monotone convergence, using the properties listed in Theorem 27.4:
⟨𝐸 𝒢 𝑢, 𝑤⟩ = lim⟨E𝒢 𝑢𝑛 , 𝑤⟩
𝑛
= lim lim⟨E𝒢 𝑢𝑛 , 𝑤𝑚 ⟩
𝑛 𝑚
= lim lim⟨𝑢𝑛 , E𝒢 𝑤𝑚 ⟩
𝑛 𝑚
= lim⟨𝑢𝑛 , 𝐸 𝒢 𝑤⟩
𝑛
= ⟨𝑢, 𝐸 𝒢 𝑤⟩.
⟨𝐸 𝒢 𝑢, 𝑤⟩ = ⟨𝐸 𝒢 𝑢+ , 𝑤+ ⟩ − ⟨𝐸 𝒢 𝑢− , 𝑤+ ⟩ − ⟨𝐸 𝒢 𝑢+ , 𝑤− ⟩ + ⟨𝐸 𝒢 𝑢− , 𝑤− ⟩
(iv): we have
𝑢 = 𝑤 ⇐⇒ 𝑢𝑗 = 𝑤𝑗 ∀ 𝑗 ⇐⇒ E𝒢 𝑢𝑗 = E𝒢 𝑤𝑗 ∀𝑗
325
R.L. Schilling: Measures, Integrals & Martingales
and we get
𝐸 𝒢 𝑢 = lim E𝒢 𝑢𝑗 = lim E𝒢 𝑤𝑗 = 𝑤.
𝑗 𝑗
(ix): we have
0 ⩽ 𝑢 ⩽ 1 ⇐⇒ 0 ⩽ 𝑢𝑛 ⩽ 1 ∀𝑛
⇐⇒ 0 ⩽ E𝒢 𝑢𝑛 ⩽ 1 ∀𝑛
⇐⇒ 0 ⩽ 𝐸 𝒢 𝑢 = lim E𝒢 𝑢𝑛 ⩽ 1.
𝑛
(x):
𝑢 ⩽ 𝑤 ⇐⇒ 0 ⩽ 𝑤 − 𝑢 ⇐⇒ 0 ⩽ 𝐸 𝒢 (𝑢 − 𝑤) = 𝐸 𝒢 𝑢 − 𝐸 𝒢 𝑤.
(xi):
| |
±𝑢 ⩽ |𝑢| ⇐⇒ ±𝐸 𝒢 𝑢 ⩽ 𝐸 𝒢 |𝑢| ⇐⇒ |𝐸 𝒢 𝑢| ⩽ 𝐸 𝒢 |𝑢|.
| |
■■
Problem 27.8 Solution: (Mind the typo in the hint: E𝒢 = E𝒢 should read E𝒢 = 𝐸 𝒢 .) Assume first
that 𝜇|𝒢 is 𝜎-finite and denote by 𝐺𝑘 ∈ 𝒢 , 𝐺𝑘 ↑ 𝑋 and 𝜇(𝐺𝑘 ) < ∞ an exhausting sequence. Then
1𝐺𝑘 ∈ 𝐿2 (𝒢 ), 1𝐺𝑘 ↑ 1 and
𝐸 𝒢 1 = sup E𝒢 𝑢𝑘 = 1,
𝑘
i.e. there is a sequence 𝑔𝑘 ∶= E𝒢 𝑢𝑘 ∈ 𝐿2 (𝒢 ) such that 𝑔𝑘 ↑ 1. Set 𝐺𝑘 ∶= {𝑔𝑘 > 1 − 1∕𝑘} and
observe that 𝐺𝑘 ↑ 𝑋 as well as
1
𝜇(𝐺𝑘 ) ⩽ 1 2
𝑔𝑘2 𝑑𝜇
(1 − ) ∫
𝑘
1
= ‖E𝒢 𝑢𝑘 ‖2𝐿2
(1 − 𝑘1 )2
1
⩽ 1 2
‖𝑢𝑘 ‖2𝐿2
(1 − 𝑘
)
< ∞.
326
Solution Manual. Last update 28th October 2021
If 𝒢 is not 𝜎-finite, e.g. if 𝒢 = {∅, 𝐺, 𝐺𝑐 , 𝑋} where 𝜇(𝐺) < ∞ and 𝜇(𝐺𝑐 ) = ∞ we find that
𝐿2 (𝒢 ) = {𝑐1𝐺 ∶ 𝑐 ∈ R}
which means that 𝐸 𝒢 1 = 1𝐺 since for every 𝐴 ⊂ 𝐺𝑐 , 𝐴 ∈ 𝒜 and 𝜇(𝐴) < ∞ we find
𝐸 𝒢 1𝐴⊍𝐺 = 𝐸 𝒢 (1𝐴 + 1𝐺 ) = 𝐸 𝒢 1𝐴 + 𝐸 𝒢 1𝐺 = 𝐸 𝒢 1𝐴 + 1𝐺
hence 𝑐 = 0 or 𝐸 𝒢 1𝐴 = 0.
𝐸 𝒢 1 = 1𝐺 ⩽ 1
is best possible.
■■
Problem 27.9 Solution: For this problem it is helpful to distinguish between E𝒢 (defined on 𝐿2 ) and
the extension 𝐸 𝒢 .
Without loss of generality we may assume that 𝑔 ⩾ 0—otherwise we would consider positive and
negative parts separately. Since 𝑔 ∈ 𝐿𝑝 (𝒢 ) we have that
E𝒢 𝑔𝑗 = 𝑔𝑗 ⇐⇒ 𝐸 𝒢 = sup E𝒢 𝑔𝑗 = sup 𝑔𝑗 = 𝑔.
𝑗 𝑗
■■
Problem 27.10 Solution: For this problem it is helpful to distinguish between E𝒢 (defined on 𝐿2 )
and the extension 𝐸 𝒢 .
Eℋ E𝒢 𝑢𝑗 = Eℋ 𝑢𝑗 ↑ 𝐸 ℋ 𝑢
327
R.L. Schilling: Measures, Integrals & Martingales
while
( )
Eℋ E𝒢 𝑢𝑗 ↑ 𝐸 ℋ sup E𝒢 𝑢𝑗 = 𝐸 ℋ 𝐸 𝒢 𝑢.
𝑗
■■
𝐸 𝒜𝑛 𝑢(𝑥) 𝑑𝑥 = 𝑐𝑘 .
∫[𝑘−1,𝑘)
Since
𝐸 𝒜𝑛 𝑢(𝑥) 𝑑𝑥 = ⟨𝐸 𝒜𝑛 𝑢, 1[𝑘−1,𝑘) ⟩
∫[𝑘−1,𝑘)
= ⟨𝑢, 𝐸 𝒜𝑛 1[𝑘−1,𝑘) ⟩
= ⟨𝑢, 1[𝑘−1,𝑘) ⟩
= 𝑢(𝑥) 𝑑𝑥
∫[𝑘−1,𝑘)
we get
𝑛
∑
𝐸 𝒜𝑛 𝑢(𝑥) = 𝑢(𝑡) 𝑑𝑡1[𝑗−1,𝑗) (𝑥).
∫[𝑗−1,𝑗)
𝑗=1
■■
Problem 27.12 Solution: For this problem it is helpful to distinguish between E𝒢 (defined on 𝐿2 )
and the extension 𝐸 𝒢 .
If 𝜇(𝑋) = ∞ and if 𝒢 = {∅, 𝑋}, then 𝐿1 (𝒢 ) = {0} which means that 𝐸 𝒢 𝑢 = 0 for any 𝑢 ∈ 𝐿1 (𝒜 ).
Thus for integrable functions 𝑢 > 0 and 𝜇|𝒢 not 𝜎-finite we can only have ‘⩽’.
If 𝜇|𝒢 is 𝜎-finite and if 𝐺𝑗 ↑ 𝑋, 𝐺𝑗 ∈ 𝒢 , 𝜇(𝐺𝑗 ) < ∞ is an exhausting sequence, we find for any
𝑢 ∈ 𝐿1+ (𝒜 )
𝐸 𝒢 𝑢 𝑑𝜇 = sup 𝐸 𝒢 𝑢 𝑑𝜇
∫ 𝑗 ∫𝐺𝑗
= sup⟨𝐸 𝒢 𝑢, 1𝐺𝑗 ⟩
𝑗
328
Solution Manual. Last update 28th October 2021
= sup⟨𝑢, 𝐸 𝒢 1𝐺𝑗 ⟩
𝑗
= sup⟨𝑢, 1𝐺𝑗 ⟩
𝑗
= ⟨𝑢, 1⟩
= 𝑢 𝑑𝜇.
∫
If 𝜇|𝒢 is not 𝜎-finite and if 𝑢 ⩾ 0, we perform a similar calculation with an exhausting sequence
𝐴𝑗 ∈ 𝒜 , 𝐴𝑗 ↑ 𝑋, 𝜇(𝐴𝑗 ) < ∞ (it is implicit that 𝜇|𝒜 is 𝜎-finite as otherwise the conditional
expectation would not be defined!):
𝐸 𝒢 𝑢 𝑑𝜇 = sup 𝐸 𝒢 𝑢 𝑑𝜇
∫ 𝑗 ∫𝐴𝑗
= sup⟨𝐸 𝒢 𝑢, 1𝐴𝑗 ⟩
𝑗
= sup⟨𝑢, 𝐸 𝒢 1𝐴𝑗 ⟩
𝑗
⩽ ⟨𝑢, 1⟩
= 𝑢 𝑑𝜇.
∫
■■
we get
( )
E𝒢 inf 𝑢𝑗 ⩽ E𝒢 𝑢𝑚 ∀𝑚 ⩾ 𝑘
𝑗⩾𝑘
thus
( )
E𝒢 inf 𝑢𝑗 ⩽ inf E𝒢 𝑢𝑚 ⩽ sup inf E𝒢 𝑢𝑚 = lim inf E𝒢 𝑢𝑚 .
𝑗⩾𝑘 𝑚⩾𝑘 𝑘 𝑚⩾𝑘 𝑚→∞
Since on the other hand the sequence inf 𝑗⩾𝑘 𝑢𝑗 increases, as 𝑘 → ∞, towards sup𝑘 inf 𝑗⩾𝑘 𝑢𝑗 we can
use the conditional Beppo Levi theorem 27.13 on the left-hand side and find
( ) ( ) ( )
E𝒢 lim inf 𝑢𝑗 = E𝒢 sup inf 𝑢𝑗 = sup E𝒢 inf 𝑢𝑗 ⩽ lim inf E𝒢 𝑢𝑚 .
𝑗→∞ 𝑘 𝑗⩾𝑘 𝑘 𝑗⩾𝑘 𝑚→∞
Proof of Corollary 27.15: Since |𝑢𝑗 | ⩽ 𝑤 we conclude that |𝑢| = lim𝑗 |𝑢𝑗 | ⩽ 𝑤 and that 2𝑤 −
|𝑢 − 𝑢𝑗 | ⩾ 0. Applying the conditional Fatou lemma 27.14 we find
( )
E𝒢 (2𝑤) = E𝒢 lim inf 2𝑤 − |𝑢 − 𝑢𝑗 |
𝑗
329
R.L. Schilling: Measures, Integrals & Martingales
( )
⩽ lim inf E𝒢 2𝑤 − |𝑢 − 𝑢𝑗 |
𝑗
Since, however,
| 𝒢 | | |
|E 𝑢𝑗 − E𝒢 𝑢| = |E𝒢 (𝑢𝑗 − 𝑢)| ⩽ E𝒢 |𝑢𝑗 − 𝑢| ←←←←←←←←←→
← 0
| | | | 𝑗→∞
■■
Problem 27.14 Solution: (i) ⇐⇒ (ii): Let 𝐴 ∈ 𝒜∞ be such that 𝜇(𝐴) < ∞. Then, by Hölder’s
inequality with 1∕𝑝 + 1∕𝑞 = 1,
| |
| 𝑢𝑗 𝑑𝜇 − 𝑢 𝑑𝜇|| ⩽ |𝑢𝑗 − 𝑢| 𝑑𝜇 ⩽ ‖𝑢𝑗 − 𝑢‖𝑝 𝜇(𝐴)1∕𝑞 ←←←←←←←←←→
← 0.
|∫ ∫ ∫
| 𝐴 𝐴 | 𝐴 𝑗→∞
Thus, if 𝑢∞ ∶= E𝒜∞ 𝑢, we find by the martingale property for all 𝑘 > 𝑗 and 𝐴 ∈ 𝒜𝑗 such that
𝜇(𝐴) < ∞
𝑢𝑗 𝑑𝜇 = 𝑢𝑘 𝑑𝜇 = lim 𝑢𝑘 𝑑𝜇 = 𝑢 𝑑𝜇 = 𝑢∞ 𝑑𝜇,
∫𝐴 ∫𝐴 𝑘→∞ ∫𝐴 ∫𝐴 ∫𝐴
and since we are in a 𝜎-finite setting, we can apply Theorem 27.12(i) and find that 𝑢𝑗 = E𝒜𝑗 𝑢∞ .
(ii) ⇐⇒ (iii): Assume first that 𝑢∞ ∈ 𝐿1 ∩ 𝐿𝑝 . Then 𝑢𝑗 = E𝒜𝑗 𝑢∞ ∈ 𝐿1 ∩ 𝐿𝑝 and Theorem 27.19(i)
shows that 𝑢𝑗 ←←←←←←←←←→
← 𝑢∞ both in 𝐿1 and a.e. In particular, we get
𝑗→∞
In the general case where 𝑢∞ ∈ 𝐿𝑝 (𝒜∞ ) we find for every 𝜖 > 0 an element 𝑢𝜖∞ ∈ 𝐿1 (𝒜∞ ) ∩
𝐿𝑝 (𝒜∞ ) such that
‖𝑢∞ − 𝑢𝜖∞ ‖𝑝 ⩽ 𝜖
(indeed, since we are working in a 𝜎-finite filtered measure space, there is an exhaustion 𝐴𝑘 ↑ 𝑋
such that 𝐴𝑘 ∈ 𝒜∞ and for large enough 𝑘 = 𝑘𝜖 the function 𝑢𝜖∞ ∶= 𝑢∞ 1𝐴𝑘 will to the job).
Similarly, we can approximate any fixed 𝜙 ∈ 𝐿𝑞 by 𝜙𝜖 ∈ 𝐿𝑞 ∩ 𝐿1 such that ‖𝜙 − 𝜙𝜖 ‖𝑞 ⩽ 𝜖.
⟨𝑢𝑗 − 𝑢∞ , 𝜙⟩
330
Solution Manual. Last update 28th October 2021
lim⟨𝑢𝑛(𝑘) − 𝑢, 𝜙⟩ = 0 ∀ 𝜙 ∈ 𝐿𝑞 .
𝑘
Then, in particular,
lim⟨𝑢𝑛(𝑘) − 𝑢, E𝒜𝑛 𝜙⟩ = 0 ∀ 𝜙 ∈ 𝐿𝑞 , 𝑛 ∈ N
𝑘
or
⟨𝑢𝑛 − E𝒜𝑛 𝑢, 𝜙⟩ = 0 ∀ 𝜙 ∈ 𝐿𝑞 , 𝑛 ∈ N.
and we conclude that 𝑢𝑛 = E𝒜𝑛 𝑢. Because of the tower property we can always replace 𝑢 by
𝑢∞ ∶= E𝒜∞ 𝑢:
(ii) ⇐⇒ (i): We show that we can take 𝑢 = 𝑢∞ . First, if 𝑢∞ ∈ 𝐿1 ∩ 𝐿∞ we find by the closability of
martingales, Theorem 27.19(i), that
331
R.L. Schilling: Measures, Integrals & Martingales
← 0
←←←←←←←←←→
𝑗→∞
Now for the general case where 𝑢∞ ∈ 𝐿𝑝 . Since we are in a 𝜎-finite setting, we can set 𝑢𝜖 ∶=
(𝑢 ⋅ 1𝐴𝑗 ) ∧ 𝑗, 𝑗 = 𝑗(𝜖) sufficiently large and 𝐴𝑗 → 𝑋 an exhausting sequence of sets from 𝒜∞ , and
can guarantee that
‖𝑢 − 𝑢𝜖 ‖𝑝 ⩽ 𝜖.
■■
Since 𝑚1 = 𝑢1 ∈ 𝐿1 𝒜1 , this shows, by induction, that 𝑚𝑘 ∈ 𝐿1 (𝒜𝑘 ). Applying 𝐸 𝒜𝑘−1 to both sides
of the displayed equality yields
Problem 27.16 Solution: Problem 27.15 shows that 𝑠𝑘 is a martingale, so that 𝑠2𝑘 is a sub-martingale
(use Jensen’s inequality for conditional expectations). Now
∑ ∑
𝑠2𝑘 𝑑𝜇 = 𝑢2𝑘 𝑑𝜇 + 2 𝑢 𝑢 𝑑𝜇
∫ 𝑗
∫ ∫ 𝑗 𝑘
𝑗<𝑘
and if 𝑗 < 𝑘
𝑢𝑗 𝑢𝑘 𝑑𝜇 = 𝐸 𝒜𝑗 (𝑢𝑗 𝑢𝑘 ) 𝑑𝜇 = 𝑢𝑗 𝐸 𝒜𝑗 (𝑢𝑘 ) 𝑑𝜇 = 0.
∫ ∫ ∫ ⏟⏟⏟
=0
■■
332
Solution Manual. Last update 28th October 2021
𝑎𝑗+1 = 𝑎𝑗 + 𝐸 𝒜𝑗 𝑢𝑗+1 − 𝑢𝑗
𝐸 𝒜𝑗 𝑢𝑗+1 ⩾ 𝑢𝑗 ⇐⇒ 𝑎𝑗+1 − 𝑎𝑗 ⩾ 0
Finally, if 𝑚𝑗 + 𝑎𝑗 = 𝑢𝑗 = 𝑚
̃ 𝑗 + 𝑎̃𝑗 are two such decompositions we find that 𝑚𝑗 − 𝑚
̃ 𝑗 = 𝑎𝑗 − 𝑎̃𝑗 is
𝒜𝑗−1 measurable. Using the martingale property we find
Martingale
̃ 𝑗 = 𝐸 𝒜𝑗−1 (𝑚𝑗 − 𝑚
𝑚𝑗 − 𝑚 ̃𝑗 ) = 𝑚𝑗−1 − 𝑚
̃ 𝑗−1
𝑚1 − 𝑚
̃ 1 = 0, 𝑚2 − 𝑚
̃ 2 = 0, 𝑚3 − 𝑚
̃ 3 = 0, …
so that 𝑚𝑗 = 𝑚
̃ 𝑗 and, consequently, 𝑎𝑗 = 𝑎̃𝑗 .
■■
Problem 27.18 Solution: Assume that 𝑀𝑘 = 𝐸 𝒜𝑘 𝑀. Then we know from Theorem 27.19 that
̃ = lim𝑘 𝑀𝑘 exists a.e. and in 𝐿1 . Moreover, ∫ 𝑀𝑘 𝑑𝑃 = 1 so that 𝑀
𝑀 ̃ cannot be trivial. On the
other hand,
■■
Problem 27.19 Solution: (Compare this problem with Problem 22.16.) Recall that in finite measure
spaces uniform integrability follows from (and is actually equivalent to)
this is true since in a finite measure space the constant function 𝑤 ≡ 𝑅 is integrable.
|𝑢𝑛 | 𝑑𝜇 ⩽ 𝐸 𝒜𝑛 𝑓 𝑑𝜇
∫{|𝑢𝑛 |>𝑅} ∫{|𝑢𝑛 |>𝑅}
= 𝑓 𝑑𝜇
∫{|𝑢𝑛 |>𝑅}
333
R.L. Schilling: Measures, Integrals & Martingales
= 𝑓 𝑑𝜇 + 𝑓 𝑑𝜇
∫{|𝑢𝑛 |>𝑅}∩{𝑓 ⩽𝑅∕2} ∫{|𝑢𝑛 |>𝑅}∩{𝑓 >𝑅∕2}
1
⩽ |𝑢 | 𝑑𝜇 + 𝑓 𝑑𝜇
∫{|𝑢𝑛 |>𝑅}∩{𝑓 ⩽𝑅∕2} 2 𝑛 ∫{|𝑢𝑛 |>𝑅}∩{𝑓 >𝑅∕2}
1
⩽ |𝑢 | 𝑑𝜇 + 𝑓 𝑑𝜇
∫{|𝑢𝑛 |>𝑅} 2 𝑛 ∫{𝑓 >𝑅∕2}
1 𝑅→∞
|𝑢𝑛 | 𝑑𝜇 ⩽ ← 0.
𝑓 𝑑𝜇 ←←←←←←←←←←←←←←←←←←←←←←←←←←←←←→
2 ∫{|𝑢𝑛 |>𝑅} ∫{𝑓 >𝑅∕2} uniformly for all 𝑛
■■
334
28 Orthonormal systems and their
convergence behaviour.
Solutions to Problems 28.1–28.11
Problem 28.1 Solution: Since 𝐽𝑘(𝛼,𝛽) is a polynomial of degree 𝑘, it is enough to show that 𝐽𝑘(𝛼,𝛽) is
orthogonal in 𝐿2 (𝐼, 𝜌(𝑥) 𝑑𝑥) to any polynomial 𝑝(𝑥) of degree 𝑗 < 𝑘. We write 𝜕 𝑘 for 𝑑𝑘
𝑑𝑥𝑘
and
𝑢(𝑥) = (𝑥 − 1)𝑘+𝛼 (𝑥 + 1)𝑘+𝛽 . Then we get by repeatedly integrating by parts
1
𝐽𝑘(𝛼,𝛽) (𝑥)𝑝(𝑥)(𝑥 − 1)𝛼 (𝑥 + 1)𝛽 𝑑𝑥
∫−1
(−1)𝑘 1
= 𝑝(𝑥)𝜕 𝑘 𝑢(𝑥) 𝑑𝑥
𝑘! 2𝑘 ∫−1
[ ]1
1
= 𝑝(𝑥) ⋅ 𝜕 𝑢(𝑥) − 𝜕 𝑝(𝑥) ⋅ 𝜕 𝑢(𝑥) + ⋯ + (−1) 𝜕 𝑝(𝑥) ⋅ 𝑢(𝑥)
𝑘−1 𝑘−2 𝑘−1 𝑘−1
−1
1
+ (−1)𝑘 𝑢(𝑥)𝜕 𝑘 𝑝(𝑥) 𝑑𝑥.
∫−1
■■
Problem 28.2 Solution: It is pretty obvious how to go about this problem. The calculations them-
selves are quite tedious and therefore omitted.
■■
Problem 28.3 Solution: Theorem 28.6: The polynomials are dense in 𝐶[𝑎, 𝑏] with respect to uniform
convergence.
Proof 2: Let 𝑓 ∈ 𝐶[𝑎, 𝑏]. Then 𝑓̃(𝑦) ∶= 𝑓 (𝑎 + (𝑏 − 𝑎)𝑦), 𝑦 ∈ [0, 1] satisfies 𝑓̃ ∈ 𝐶[0, 1] and,
because of Theorem 28.6, there is a sequence of polynomials 𝑝̃𝑛 such that
( 𝑥−𝑎 )
Define 𝑝𝑛 (𝑥) ∶= 𝑝̃𝑛 𝑏−𝑎
, 𝑥 ∈ [𝑎, 𝑏]. Clearly 𝑝𝑛 is a polynomial and we have
𝑝𝑛 (𝑦) − 𝑓̃(𝑦)|.
sup |𝑝𝑛 (𝑥) − 𝑓 (𝑥)| = sup |̃
𝑥∈[𝑎,𝑏] 𝑦∈[0,1]
335
R.L. Schilling: Measures, Integrals & Martingales
=0
and since the map [0, 1] ∋ 𝑡 → 𝑥 = (𝑏 − 𝑎)𝑡 + 𝑎 ∈ [𝑎, 𝑏] is continuous, bijective and with a
continuous inverse, we also get
■■
= 2 sin 𝑥 sin 𝑦,
and that
( ) [ ( )]
Re 𝑒𝑖(𝑥+𝑦) + 𝑒𝑖(𝑥−𝑦) = Re 𝑒𝑖𝑥 𝑒𝑖𝑦 + 𝑒−𝑖𝑦
[ ]
= Re 2𝑒𝑖𝑥 cos 𝑦
= 2 cos 𝑥 cos 𝑦.
⎧ 𝑒𝑖𝑁𝑥 |𝜋
𝜋
⎪ | = 0, if 𝑁 ≠ 0;
𝑒𝑖𝑁𝑥 𝑑𝑥 = ⎨ 𝑖𝑁 |−𝜋
if 𝑁 = 0.
∫−𝜋 ⎪2𝜋,
⎩
Thus, if 𝑘 ≠ 𝓁
𝜋 ( 𝜋 𝜋 )
2 cos 𝑘𝑥 cos 𝓁𝑥 𝑑𝑥 = Re 𝑒 𝑖(𝑘+𝓁)𝑥
𝑑𝑥 + 𝑒 𝑖(𝑘+𝓁)𝑥
𝑑𝑥 = 0
∫−𝜋 ∫−𝜋 ∫−𝜋
336
Solution Manual. Last update 28th October 2021
and if 𝑘 = 𝓁 ⩾ 1
𝜋 ( 𝜋 𝜋 )
2𝑖𝑘𝑥
2 cos 𝑘𝑥 cos 𝑘𝑥 𝑑𝑥 = Re 𝑒 𝑑𝑥 + 1 𝑑𝑥 = 2𝜋
∫−𝜋 ∫−𝜋 ∫−𝜋
and if 𝑘 = 𝓁 = 0,
𝜋
2 cos 𝑘𝑥 cos 𝑘𝑥 𝑑𝑥 = 2 𝑑𝑥 = 4𝜋.
∫−𝜋 ∫
The proof for the pure sines integral is similar while for the mixed sine-cosine integrals the integ-
rand
𝑥 → cos 𝑘𝑥 sin 𝓁𝑥
is always an odd function, the integral over the symmetric (w.r.t. the origin) interval (−𝜋, 𝜋) is
always zero.
■■
(i) We have
( 𝑖𝑥 )
𝑒 + 𝑒−𝑖𝑥 𝑘
2𝑘 cos𝑘 (𝑥) = 2𝑘
2
( 𝑖𝑥 )
𝑒 + 𝑒−𝑖𝑥 𝑘
=
2
𝑘 ( )
∑ 𝑘
= 𝑒𝑖𝑗𝑥 𝑒−𝑖(𝑘−𝑗)𝑥
𝑗=0
𝑗
𝑘 ( )
∑ 𝑘 𝑖(2𝑗−𝑘)𝑥
= 𝑒
𝑗=0
𝑗
Adding the first and last terms, second and penultimate terms, term no. 𝑗 and 𝑘 − 𝑗, etc. under
() ( 𝑘 )
the sum gives, since the binomial coefficients satisfy 𝑘𝑗 = 𝑘−𝑗 ,
– if 𝑘 = 2𝑛 is even
𝑛−1 ( ) ( )
∑ 2𝑛 2𝑛
2𝑛 2𝑛
2 cos (𝑥) = (𝑒𝑖(2𝑗−2𝑛)𝑥
+𝑒 𝑖(2𝑛−2𝑗)𝑥
)+
𝑗=0
𝑗 𝑛
𝑛 ( ) ( )
∑ 2𝑛 2𝑛
= 2 cos(2𝑗 − 2𝑛) +
𝑗=0
𝑗 𝑛
– if 𝑘 = 2𝑛 − 1 is odd
𝑛−1 ( )
∑ 2𝑛 − 1
2𝑛−1 2𝑛−1
2 cos (𝑥) = (𝑒𝑖(2𝑗−2𝑛+1)𝑥 + 𝑒𝑖(2𝑛−2𝑗−1)𝑥 )
𝑗=0
𝑗
𝑛−1 ( )
∑ 2𝑛 − 1
= 2 cos(2𝑛 − 2𝑗 − 1)𝑥.
𝑗=0
𝑗
337
R.L. Schilling: Measures, Integrals & Martingales
Adding the first and last terms, second and penultimate terms, term no. 𝑗 and 𝑘 − 𝑗, etc. under
() ( 𝑘 )
the sum gives, since the binomial coefficients satisfy 𝑘𝑗 = 𝑘−𝑗 ,
– if 𝑘 = 2𝑛 is even
– if 𝑘 = 2𝑛 − 1 is odd
(ii) We have
( )𝑘 ( )𝑘
cos 𝑘𝑥 + 𝑖 sin 𝑘𝑥 = 𝑒𝑖𝑘𝑥 = 𝑒𝑖𝑥 = cos 𝑥 + 𝑖 sin 𝑥
338
Solution Manual. Last update 28th October 2021
it is a matter of double summation and part (ii) to see that 𝑇𝑛 (𝑥) can be written like 𝑈𝑛 (𝑥).
Conversely, part (i) enables us to rewrite any expression of the form 𝑈𝑛 (𝑥) as 𝑇𝑛 (𝑥).
■■
■■
339
R.L. Schilling: Measures, Integrals & Martingales
𝜋
2
= sin 𝑥 cos 𝑗𝑥 𝑑𝑥
𝜋 ∫0
𝜋 ( )
2 1
= sin((𝑗 + 1)𝑥) − sin((𝑗 − 1)𝑥) 𝑑𝑥
𝜋 ∫0 2
[ ]𝜋
1 cos((𝑗 − 1)𝑥) cos((𝑗 + 1)𝑥)
= −
𝜋 𝑗−1 𝑗+1 0
[ ]
1 cos((𝑗 − 1)𝜋) cos((𝑗 + 1)𝜋) 1 1
= − − + .
𝜋 𝑗−1 𝑗+1 𝑗−1 𝑗+1
If 𝑗 is odd, we get 𝑎𝑗 = 0 and if 𝑗 is even, we have
[ ]
1 −1 −1 1 1 4 1
𝑎𝑗 = − − + =− .
𝜋 𝑗−1 𝑗+1 𝑗−1 𝑗+1 𝜋 (𝑗 − 1)(𝑗 + 1)
This shows that we have only evenly indexed cosines in the Fourier series.
■■
Problem 28.8 Solution: This is not as trivial as it looks in the first place! Since 𝑢 is itself a Haar
function, we have
𝑠𝑁 (𝑢, 𝑥) = 𝑢(𝑥) ∀𝑁 ∈ N
(it is actually the first Haar function) so that 𝑠𝑁 converges in any 𝐿𝑝 -norm, 1 ⩽ 𝑝 < ∞ to 𝑢.
The same applies to the right tail of the Haar wavelet expansion. The left tail, however, converges
only for 1 < 𝑝 < ∞ in 𝐿𝑝 . The reason is the calculation of Step 5 in the proof of Theorem 28.20
which goes in the case 𝑝 = 1:
Δ
E𝒜−𝑀 𝑢 = 2−𝑀 𝑢(𝑥) 𝑑𝑥1[−2𝑀 ,0) + 2−𝑀 𝑢(𝑥) 𝑑𝑥1[0,2𝑀 )
∫[−2𝑀 ,0) ∫[0,2𝑀 )
= 2−𝑀 1[0,2𝑀 ) ,
but this is not 𝐿1 -convergent to 0 as it would be required. For 𝑝 > 1 all is fine, though....
■■
Problem 28.9 Solution: Assume that 𝑢 is uniformly continuous (𝐶𝑐 and 𝐶∞ -functions are!). Since
𝐻
𝑠𝑛 (𝑢; 𝑥) = E𝒜𝑛 𝑢(𝑥)
is the projection onto the sets in 𝒜𝑛𝐻 , see e.g. Step 2 in the proof of Theorem 28.17, we have
1
𝑠𝑛 (𝑢; 𝑥) = 𝑢(𝑦) 𝑑𝑥1𝐼 (𝑥)
𝜆(𝐼) ∫𝐼
where 𝐼 is an dyadic interval from the generator of 𝒜𝑛𝐻 as in Step 2 of the proof of Theorem 28.17.
Thus, if 𝑥 is from 𝐼 we get
| 1 |
|𝑠𝑛 (𝑢; 𝑥) − 𝑢(𝑥)| = || (𝑢(𝑦) − 𝑢(𝑥)) 𝑑𝑥||
| 𝜆(𝐼) ∫𝐼 |
1
⩽ |𝑢(𝑦) − 𝑢(𝑥)| 𝑑𝑥
𝜆(𝐼) ∫𝐼
340
Solution Manual. Last update 28th October 2021
1
⩽ 𝜖 𝑑𝑥
𝜆(𝐼) ∫𝐼
=𝜖
if 𝜆(𝐼) < 𝛿 for small enough 𝛿 > 0. This follows from uniform continuity: for given 𝜖 > 0 there
is some 𝛿 > 0 such that for 𝑥, 𝑦 ∈ 𝐼 (this entails |𝑥 − 𝑦| ⩽ 𝛿!) we have |𝑢(𝑥) − 𝑢(𝑦)| ⩽ 𝜖.
The above calculation holds uniformly for all 𝑥 and we are done.
■■
Problem 28.10 Solution: The calculation for the right tail is more or less the same as in Problem
28.9. Only the left tail differs. Here we argue as in Step 5 of the proof of Theorem 28.20: if
𝑢 ∈ 𝐶𝑐 (R) we can assume that supp 𝑢 ⊂ [−𝑅, 𝑅] and we see
Δ
E𝒜−𝑀 𝑢(𝑥) = 2−𝑀 𝑢(𝑥) 𝑑𝑥1[−2𝑀 ,0) + 2−𝑀 𝑢(𝑥) 𝑑𝑥1[0,2𝑀 )
∫[−𝑅,0] ∫[0,𝑅]
⩽ 2−𝑀 𝑅 ‖𝑢‖∞ 1[−2𝑀 ,0) + 2−𝑀 𝑅 ‖𝑢‖∞ 1[0,2𝑀 )
= 2−𝑀 𝑅 ‖𝑢‖∞ 1[−2𝑀 ,2𝑀 )
𝑀→∞
⩽ 2−𝑀 𝑅 ‖𝑢‖∞ ←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←→
← 0
uniformly for all 𝑥
If 𝑢 ∈ 𝐶∞ we can use the fact that 𝐶𝑐 is dense in 𝐶∞ , i.e. we can find for every 𝜖 > 0 functions
𝑣 = 𝑣𝜖 ∈ 𝐶𝑐 and 𝑤 = 𝑤𝜖 ∈ 𝐶∞ such that
Then
| 𝒜−𝑀
Δ | | Δ | | Δ |
|E 𝑢(𝑥)| ⩽ |E𝒜−𝑀 𝑣(𝑥)| + |E𝒜−𝑀 𝑤(𝑥)|
| | | | | |
| 𝒜−𝑀
Δ | Δ
⩽ |E 𝑣(𝑥)| + E 𝒜−𝑀
‖𝑤‖∞
| |
| Δ |
⩽ |E𝒜−𝑀 𝑣(𝑥)| + 𝜖
| |
and, by the first calculation for 𝐶𝑐 -functions, the right-hand side converges, since 𝑣 ∈ 𝐶𝑐 , to 0 + 𝜖
uniformly for all 𝑥, and letting 𝜖 → 0 we conclude the proof.
■■
Problem 28.11 Solution: See the picture at the end of this solution. Since the function 𝑢(𝑥) ∶=
1[0,1∕3) (𝑥) is piecewise constant, and since for each Haar function ∫ 𝜒𝑘,𝑗 𝑑𝑥 = 0 unless 𝑗 = 𝑘 = 1,
we see that only a single Haar function contributes to the value of 𝑠𝑁 (𝑢; 13 ), namely where 1
3
∈
supp 𝜒𝑛,𝑗 .
The idea of the proof is now pretty clear: take values 𝑁 where 𝑥 = 1
3
is in the left ‘half’ of 𝜒𝑁,𝑘 ,
i.e. where 𝜒𝑁,𝑘 ( 13 ) = 1 and values 𝑀 such that 𝑥 = 1
3
is in the opposite, negative ‘half’ of 𝜒𝑀,𝓁 ,
i.e. where 𝜒𝑀,𝓁 ( 31 ) = −1. Of course, 𝑘, 𝓁 depend on 𝑥, 𝑁 and 𝑀 respectively. One should expect
that the partial sums for these different positions lead to different limits, hence different upper and
lower limits.
341
R.L. Schilling: Measures, Integrals & Martingales
The problem is to pick 𝑁’s and 𝑀’s. We begin with the simple observation that the dyadic (i.e.
base-2) representation of 1∕3 is the periodic, infinite dyadic fraction
∞
1 ∑ 1
= 0.01010101 ⋯ =
3 𝑘=1
22𝑘
Now consider those Haar functions whose support consists of intervals of the length 2−2𝑛 , i.e.
the 𝜒2𝑛,𝑗 ’s and agree that 𝑗 = 𝑗(1∕3, 𝑛) is the one value where 1
3
∈ supp 𝜒2𝑛,𝑗 . By construction
supp 𝜒2𝑛,𝑗 = [𝑑𝑛 , 𝑑𝑛 + 1∕22𝑛 ] and we get for the Haar-Fourier partial sum
1∕3
1
𝑠2𝑛 (𝑢, 31 ) − = 2𝑛 𝑑𝑥 ⋅ 𝜒2𝑛,𝑗 ( 13 )
3 ∫𝑑𝑛
(1 )
= 22𝑛 3
− 𝑑𝑛
∞
∑ 1
= 4𝑛
𝑘=𝑛+1
2 2𝑘
∞
∑ 1
= 4𝑛
𝑘=𝑛+1
4 𝑘
1
= 4𝑛 4−𝑛−1 1
1− 4
1
= .
3
The shift by −1∕3 comes from the starting ‘atypical’ Haar function 𝜒0,0 since ⟨𝑢, 𝜒0,0 ⟩ = ∫0
1∕3
𝑑𝑥 =
1
3
.
Using the next smaller Haar functions with support of length 2−2𝑛−1 , i.e. the 𝜒2𝑛+1,𝑘 ’s, we see that
with 𝑗 as above 𝜒2𝑛+1,2𝑗−1 ( 31 ) = −1 (since twice as many Haar functions appear in the run-up to
𝑑𝑛 ) and that
1
𝑠2𝑛+1 (𝑢, 31 ) −
3
[ 𝑑𝑛 +1∕22𝑛+2 1∕3 ]
= 2 𝑑𝑥 −
𝑛+1
2 𝑑𝑥 ⋅ 𝜒2𝑛+1,2𝑗−1 ( 13 )
𝑛+1
∫𝑑𝑛 ∫𝑑𝑛 +1∕22𝑛+2
[ ]
1 1 1
= 𝑑𝑛 + 2𝑛+2 − 𝑑𝑛 − + 𝑑𝑛 + 2𝑛+2 2𝑛+1 ⋅ (−2𝑛+1 )
2 3 2
[ ]
1 2
= 𝑑𝑛 − + 2𝑛+2 ⋅ (−22𝑛+2 )
3 2
342
Solution Manual. Last update 28th October 2021
( )
1
= 4 ⋅ 22𝑛 − 𝑑𝑛 − 2
3
1
=4⋅ −2 (using the result above)
3
2
=−
3
2 1
𝑠2𝑛 (𝑢; 31 ) = > − = 𝑠2𝑛+1 (𝑢, 13 )
3 3
and the claim follows since because of the above inequality,
1 2
lim inf 𝑠𝑁 (𝑢; 31 ) ⩽ − ⩽ ⩽ lim sup 𝑠𝑁 (𝑢; 13 ).
𝑁 3 3 𝑁
■■
343
R.L. Schilling: Measures, Integrals & Martingales
2𝑛+1
2 2
and 𝜒2𝑛+1,2𝑗−1
𝜒2𝑛,𝑗
2𝑛
22
1
𝑑𝑛+1 3
✲
1 1
𝑑𝑛 𝑑𝑛 + 22𝑛+1
𝑑𝑛 + 22𝑛
344