Brownian Motion

Brownian Motion (2nd edition)
An Introduction to Stochastic Processes

de Gruyter Graduate, Berlin 2014
ISBN: 978–3–11–030729–0
Solution Manual
René L. Schilling & Lothar Partzsch
Dresden, June 2014. Last update December 19, 2017

R.L. Schilling, L. Partzsch: Brownian Motion (2nd edn.)
Acknowledgement. We are grateful to Björn Böttcher, Katharina Fischer, Julian Hollender,

Franziska Kühn, Felix Lindner, Vera Schade and Michael Schwarzenberger who supported us
in the preparation of this solution manual. Daniel Tillich and Robert Baumgarth pointed out
quite a few misprints and minor errors, as well as alternative solutions.
Dresden, June 2014 René Schilling

Lothar Partzsch
2
Contents
1 Robert Brown’s new thing 5
2 Brownian motion as a Gaussian process 13
3 Constructions of Brownian motion 25
4 The canonical model 35
5 Brownian motion as a martingale 41
6 Brownian motion as a Markov process 53
7 Brownian motion and transition semigroups 65
8 The PDE connection 85
9 The variation of Brownian paths 93
10 Regularity of Brownian paths 101
11 Brownian motion as a random fractal 107
12 The growth of Brownian paths 113
13 Strassen’s functional law of the iterated logarithm 117
14 Skorokhod representation 125
15 Stochastic integrals: L2 –theory 127
16 Stochastic integrals: Beyond L2T 139
17 Itô’s formula 143
18 Applications of Itô’s formula 155
19 Stochastic differential equations 165
20 Stratonovich’s stochastic calculus 183
3
21 On diffusions 185
4
1 Robert Brown’s new thing
Problem 1.1. Solution:
(a) We show the result for Rd -valued random variables. Let ξ, η ∈ Rd . By assumption,
ξ Xn ξ X
lim E exp [i ⟨( ), ( )⟩] = E exp [i ⟨( ), ( )⟩]
n→∞ η Yn η Y
⇐⇒ lim E exp [i⟨ξ, Xn ⟩ + i⟨η, Yn ⟩] = E exp [i⟨ξ, X⟩ + i⟨η, Y ⟩]
n→∞
If we take ξ = 0 and η = 0, respectively, we see that

d
lim E exp [i⟨η, Yn ⟩] = E exp [i⟨η, Y ⟩] or Yn Ð
→Y
n→∞
d
lim E exp [i⟨ξ, Xn ⟩] = E exp [i⟨ξ, X⟩] or Xn Ð
→ X.
n→∞
Since Xn á Yn we find
E exp [i⟨ξ, X⟩ + i⟨η, Y ⟩] = lim E exp [i⟨ξ, Xn ⟩ + i⟨η, Yn ⟩]

n→∞
= lim E exp [i⟨ξ, Xn ⟩] E exp [i⟨η, Yn ⟩]

n→∞
= lim E exp [i⟨ξ, Xn ⟩] lim E exp [i⟨η, Yn ⟩]

n→∞ n→∞
= E exp [i⟨ξ, X⟩] E exp [i⟨η, Y ⟩]
and this shows that X á Y .
(b) We have
1 almost surely d
Xn = X + ÐÐÐÐÐÐÐ→ X Ô⇒ Xn Ð
→X
n n→∞
1 almost surely d
Yn = 1 − Xn = 1 − − X ÐÐÐÐÐÐÐ→ 1 − X Ô⇒ Yn Ð
→1−X
n n→∞
almost surely d
Xn + Yn = 1 ÐÐÐÐÐÐÐ→ 1 Ô⇒ Xn + Yn Ð
→ 1.
n→∞
A simple direct calculation shows that 1 − X ∼ 21 (δ0 + δ1 ) ∼ Y . Thus,

d d d
Xn Ð
→ X, Yn Ð
→ Y ∼ 1 − X, Xn + Yn Ð
→ 1.
d
Assume that (Xn , Yn ) Ð
→ (X, Y ). Since X á Y , we find for the distribution of X + Y :
X + Y ∼ 12 (δ0 + δ1 ) ∗ 21 (δ0 + δ1 ) = 41 (δ0 ∗ δ0 + 2δ1 ∗ δ0 + δ1 ∗ δ1 ) = 41 (δ0 + 2δ1 + δ2 ).
Thus, X + Y ∼/ δ0 ∼ 1 = limn (Xn + Yn ) and this shows that we cannot have that
d
(Xn , Yn ) Ð
→ (X, Y ).
5
d
(c) If Xn á Yn and X á Y , then we have Xn + Yn Ð
→ X + Y : this follows since we have
for all ξ ∈ R:
lim E eiξ(Xn +Yn ) = lim E eiξXn E eiξYn

n→∞ n→∞
= lim E eiξXn lim E eiξYn

n→∞ n→∞
= Ee iξX
Ee iξY
= E [eiξX eiξY ]
a)
= E eiξ(X+Y ) .
d
A similar (even easier) argument works if (Xn , Yn ) Ð
→ (X, Y ). Then we have
f (x, y) ∶= eiξ(x+y)
is bounded and continuous, i. e. we get directly
lim E eiξ(Xn +Yn ) lim E f (Xn , Yn ) = E f (X, Y ) = E eiξ(X+Y ) .

n→∞ n→∞
For a counterexample (if Xn and Yn are not independent), see part b).
Notice that the independence and d-convergence of the sequences Xn , Yn already

implies X á Y and the d-convergence of the bivariate sequence (Xn , Yn ). This is a
consequence of the following
Lemma. Let (Xn )n⩾1 and (Yn )n⩾1 be sequences of random variables (or random
vectors) on the same probability space (Ω, A, P). If
d d
Xn á Yn for all n ⩾ 1 and Xn ÐÐÐ→ X and Yn ÐÐÐ→ Y,
n→∞ n→∞
d
then (Xn , Yn ) ÐÐÐ→ (X, Y ) and X á Y .
n→∞
Proof. Write φX , φY , φX,Y for the characteristic functions of X, Y and the pair
(X, Y ). By assumption
lim φXn (ξ) = lim E eiξXn = E eiξX = φX (ξ).

n→∞ n→∞
A similar statement is true for Yn and Y . For the pair we get, because of independence
lim φXn ,Yn (ξ, η) = lim E eiξXn +iηYn

n→∞ n→∞
= lim E eiξXn E eiηYn

n→∞
= lim E eiξXn lim E eiηYn

n→∞ n→∞
= Ee iξX
Ee iηY
= φX (ξ)φY (η).
6
Solution Manual. Last update December 19, 2017
Thus, φXn ,Yn (ξ, η) → h(ξ, η) = φX (ξ)φY (η). Since h is continuous at the origin
(ξ, η) = 0 and h(0, 0) = 1, we conclude from Lévy’s continuity theorem that h is a
d
(bivariate) characteristic function and that (Xn , Yn ) Ð
→ (X, Y ). Moreover,
h(ξ, η) = φX,Y (ξ, η) = φX (ξ)φY (η)
which shows that X á Y .
∎∎
Problem 1.2. Solution: Using the elementary estimate

iz
∣eiz − 1∣ = ∣∫ eζ dζ∣ ⩽ sup ∣eiy ∣ ∣z∣ = ∣z∣ (*)
0 ∣y∣⩽∣z∣
we see that the function t ↦ ei⟨ξ,t⟩ , ξ, t ∈ Rd is locally Lipschitz continuous:
∣ei⟨ξ,t⟩ − ei⟨ξ,s⟩ ∣ = ∣ei⟨ξ,t−s⟩ − 1∣ ⩽ ∣⟨ξ, t − s⟩∣ ⩽ ∣ξ∣ ⋅ ∣t − s∣ for all ξ, t, s ∈ Rd ,
Thus,
E ei⟨ξ,Yn ⟩ = E [ei⟨ξ,Yn −Xn ⟩ ei⟨ξ,Xn ⟩ ]
= E [(ei⟨ξ,Yn −Xn ⟩ − 1)ei⟨ξ,Xn ⟩ ] + E ei⟨ξ,Xn ⟩ .
Since limn→∞ E ei⟨ξ,Xn ⟩ = E ei⟨ξ,X⟩ , we are done if we can show that the first term in the
last line of the displayed formula tends to zero. To see this, we use the Lipschitz continuity
of the exponential function. Fix ξ ∈ Rd .
∣ E [(ei⟨ξ,Yn −Xn ⟩ − 1)ei⟨ξ,Xn ⟩ ]∣
⩽ E ∣(ei⟨ξ,Yn −Xn ⟩ − 1)ei⟨ξ,Xn ⟩ ∣
= E ∣ei⟨ξ,Yn −Xn ⟩ − 1∣
=∫ ∣ei⟨ξ,Yn −Xn ⟩ − 1∣ d P + ∫ ∣ei⟨ξ,Yn −Xn ⟩ − 1∣ d P

∣Yn −Xn ∣⩽δ ∣Yn −Xn ∣>δ
(*)
⩽ δ ∣ξ∣ + ∫ 2dP
∣Yn −Xn ∣>δ
= δ ∣ξ∣ + 2 P (∣Yn − Xn ∣ > δ)

ÐÐÐ→ δ ∣ξ∣ ÐÐ→ 0,
n→∞ δ→0
P
where we used in the last step the fact that Xn − Yn Ð
→ 0.
∎∎
d
Problem 1.3. Solution: Recall that Yn Ð
→ Y with Y = c a.s., i. e. where Y ∼ δc for some
P
constant c ∈ R. Since the d-limit is trivial, this implies Yn Ð
→ Y . This means that both
“is this still true”-questions can be answered in the affirmative.
7
d
We will show that (Xn , Yn ) Ð
→ (Xn , c) holds – without assuming anything on the joint
distribution of the random vector (Xn , Yn ), i.e. we do not make assumption on the corre-
lation structure of Xn and Yn . Since the maps x ↦ x + y and x ↦ x ⋅ y are continuous, we
see that
lim E f (Xn , Yn ) = E f (X, c) ∀f ∈ Cb (R × R)
n→∞
implies both
lim E g(Xn Yn ) = E g(Xc) ∀g ∈ Cb (R)
n→∞
and
lim E h(Xn + Yn ) = E h(X + c) ∀h ∈ Cb (R).
n→∞
This proves (a) and (b).
In order to show that (Xn , Yn ) converges in distribution, we use Lévy’s characterization of

distributional convergence, i.e. the pointwise convergence of the characteristic functions.
This means that we take f (x, y) = ei(ξx+ηy) for any ξ, η ∈ R:
∣E ei(ξXn +ηYn ) − E ei(ξX+ηc) ∣ ⩽ ∣E ei(ξXn +ηYn ) − E ei(ξXn +ηc) ∣ + ∣E ei(ξXn +ηc) − E ei(ξX+ηc) ∣
⩽ E ∣ei(ξXn +ηYn ) − E ei(ξXn +ηc) ∣ + ∣E ei(ξXn +ηc) − E ei(ξX+ηc) ∣
⩽ E ∣eiηYn − eiηc ∣ + ∣E eiξXn − E eiξX ∣ .
d
The second expression on the right-hand side converges to zero as Xn Ð
→ X. For fixed
η we have that y ↦ eiηy is uniformly continuous. Therefore, the first expression on the
right-hand side becomes, with any > 0 and a suitable choice of δ = δ() > 0
E ∣eiηYn − eiηc ∣ = E [∣eiηYn − eiηc ∣ 1{∣Yn −c∣>δ} ] + E [∣eiηYn − eiηc ∣ 1{∣Yn −c∣⩽δ} ]
⩽ 2 E [1{∣Yn −c∣>δ} ] + E [1{∣Yn −c∣⩽δ} ]
⩽ 2 P(∣Yn − c∣ > δ) +
P -convergence as δ, are fixed
ÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐÐ→ Ð→ 0.
n→∞ ↓0
Remark. The direct approach to (a) is possible but relatively ugly. Part (b) has a
relatively simple direct proof:
Fix ξ ∈ R.
E eiξ(Xn +Yn ) − E eiξX = ( E eiξ(Xn +Yn ) − E eiξXn ) + ( E eiξXn − E eiξX ) .

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
Ðn→∞
ÐÐ→0 by d-convergence
For the first term on the right we find with the uniform-continuity argument from Prob-
lem 1.2 and any > 0 and suitable δ = δ(, ξ) that
∣ E eiξ(Xn +Yn ) − E eiξXn ∣ ⩽ E ∣eiξXn (eiξYn − 1)∣
= E ∣eiξYn − 1∣
8
⩽ + P (∣Yn ∣ > δ)
fixed
ÐÐÐ→ ÐÐ→ 0
n→∞ →0
where we use P-convergence in the penultimate step.
∎∎
Problem 1.4. Solution: Let ξ, η ∈ R and note that f (x) = eiξx and g(y) = eiηy are bounded
and continuous functions. Thus we get
i⟨(ηξ ), (X )⟩
Ee Y = E eiξX eiηY
= E f (X)g(Y )
= lim E f (Xn )g(Y )
n→∞
= lim E eiξXn eiηY

n→∞
i⟨(ηξ ), (XYn )⟩
= lim E e
n→∞
d
and we see that (Xn , Y ) Ð
→ (X, Y ).
Assume now that X = φ(Y ) for some Borel function φ. Let f ∈ Cb and pick g ∶= f ○ φ.
Clearly, f ○ φ ∈ Bb and we get
E f (Xn )f (X) = E f (Xn )f (φ(Y ))

= E f (Xn )g(Y )
ÐÐÐ→ E f (X)g(Y )
n→∞
= E f (X)f (X)
= E f 2 (X).
Now observe that f ∈ Cb Ô⇒ f 2 ∈ Cb and g ≡ 1 ∈ Bb . By assumption
E f 2 (Xn ) ÐÐÐ→ E f 2 (X).

n→∞
Thus,
E (∣f (X) − f (Xn )∣2 ) = E f 2 (Xn ) − 2 E f (Xn )f (X) + E f 2 (X)

ÐÐÐ→ E f 2 (X) − 2 E f (X)f (X) + E f 2 (X) = 0,
n→∞
L2
i. e. f (Xn ) Ð→ f (X).
Now fix > 0 and R > 0 and set f (x) = −R ∨ x ∧ R. Clearly, f ∈ Cb . Then
P(∣Xn − X∣ > )
⩽ P(∣Xn − X∣ > , ∣X∣ ⩽ R, ∣Xn ∣ ⩽ R) + P(∣X∣ ⩾ R) + P(∣Xn ∣ ⩾ R)
= P(∣f (Xn ) − f (X)∣ > , ∣X∣ ⩽ R, ∣Xn ∣ ⩽ R) + P(∣X∣ ⩾ R) + P(∣f (Xn )∣ ⩾ R)
9
⩽ P(∣f (Xn ) − f (X)∣ > ) + P(∣X∣ ⩾ R) + P(∣f (Xn )∣ ⩾ R)

⩽ P(∣f (Xn ) − f (X)∣ > ) + P(∣X∣ ⩾ R) + P(∣f (X)∣ ⩾ R/2) + P(∣f (Xn ) − f (X)∣ ⩾ R/2)
where we used that {∣f (Xn )∣ ⩾ R} ⊂ {∣f (X)∣ ⩾ R/2} ∪ {∣f (Xn ) − f (X)∣ ⩾ R/2} because of
the triangle inequality: ∣f (Xn )∣ ⩽ ∣f (X)∣ + ∣f (X) − f (Xn )∣
= P(∣f (Xn ) − f (X)∣ > ) + P(∣X∣ ⩾ R/2) + P(∣X∣ ⩾ R/2) + P(∣f (Xn ) − f (X)∣ ⩾ R/2)
= P(∣f (Xn ) − f (X)∣ > ) + 2 P(∣X∣ ⩾ R/2) + P(∣f (Xn ) − f (X)∣ ⩾ R/2)
1 4
⩽ ( 2 + 2 ) E (∣f (X) − f (Xn )∣2 ) + 2 P(∣X∣ ⩾ R/2)
R
,R fixed and f =fR ∈Cb X is a.s. R-valued
ÐÐÐÐÐÐÐÐÐÐÐÐ→ 2 P(∣X∣ ⩾ R/2) ÐÐÐÐÐÐÐÐÐÐ→ 0.
n→∞ R→∞
∎∎
Problem 1.5. Solution: Note that E δj = 0 and V δj = E δj2 = 1. Thus, E S⌊nt⌋ = 0 and
V S⌊nt⌋ = ⌊nt⌋.
(a) We have, by the central limit theorem (CLT)

√
S⌊nt⌋ ⌊nt⌋ S⌊nt⌋ CLT √
√ = √ √ ÐÐÐ→ t G1
n n ⌊nt⌋ n→∞
√
where G1 ∼ N(0, 1), hence Gt ∶= t G1 ∼ N (0, t).
(b) Let s < t. Since the δj are iid, we have, S⌊nt⌋ − S⌊ns⌋ ∼ S⌊nt⌋−⌊ns⌋ , and by the central
limit theorem (CLT)
√
S⌊nt⌋−⌊ns⌋ ⌊nt⌋ − ⌊ns⌋ S⌊nt⌋−⌊ns⌋ CLT √
√ = √ √ ÐÐÐ→ t − s G1 ∼ Gt−s .
n n ⌊nt⌋ − ⌊ns⌋ n→∞
If we know that the bivariate random variable (S⌊ns⌋ , S⌊nt⌋ −S⌊ns⌋ ) converges in distri-
bution, we do get Gt ∼ Gs + Gt−s because of Problem 1.1. But this follows again from
the lemma which we prove in part d). This lemma shows that the limit has indepen-
dent coordinates, see also part c). This is as close as we can come to Gt − Gs ∼ Gt−s ,
unless we have a realization of ALL the Gt on a good space. It is Brownian motion
which will achieve just this.
(c) We know that the entries of the vector (Xtnm − Xtnm−1 , . . . , Xtn2 − Xtn1 , Xtn1 ) are inde-
pendent (they depend on different blocks of the δj and the δj are iid) and, by the
one-dimensional argument of b) we see that
d √
Xtnk − Xtnk−1 ÐÐÐ→ tk − tk−1 Gk1 ∼ Gktk −tk−1 for all k = 1, . . . , m
n→∞
where the Gk1 , k = 1, . . . , m are standard normal random vectors.
By the lemma in part d) of the solution we even see that

d √ √
(Xtnm − Xtnm−1 , . . . , Xtn2 − Xtn1 , Xtn1 ) ÐÐÐ→ ( t1 G11 , . . . , tm − tm−1 Gm
1 )
n→∞
10
and the Gk1 , k = 1, . . . , m are independent. Thus, by the second assertion of part b)
√ √
( t1 G11 , . . . , tm − tm−1 Gm
1 ) ∼ (Gt1 , . . . , Gtm −tm−1 ) ∼ (Gt1 , . . . , Gtm − Gtm−1 ).
1 m
(d) We have the following
Lemma. Let (Xn )n⩾1 and (Yn )n⩾1 be sequences of random variables (or random
vectors) on the same probability space (Ω, A, P). If
d d
Xn á Yn for all n ⩾ 1 and Xn ÐÐÐ→ X and Yn ÐÐÐ→ Y,
n→∞ n→∞
d
then (Xn , Yn ) ÐÐÐ→ (X, Y ) and X á Y (for suitable versions of the rv’s).
n→∞
Proof. Write φX , φY , φX,Y for the characteristic functions of X, Y and the pair
(X, Y ). By assumption
lim φXn (ξ) = lim E eiξXn = E eiξX = φX (ξ).

n→∞ n→∞
A similar statement is true for Yn and Y . For the pair we get, because of independence
lim φXn ,Yn (ξ, η) = lim E eiξXn +iηYn

n→∞ n→∞
= lim E eiξXn E eiηYn

n→∞
= lim E eiξXn lim E eiηYn

n→∞ n→∞
= Ee iξX
Ee iηY
= φX (ξ)φY (η).
Thus, φXn ,Yn (ξ, η) → h(ξ, η) = φX (ξ)φY (η). Since h is continuous at the origin
(ξ, η) = 0 and h(0, 0) = 1, we conclude from Lévy’s continuity theorem that h is a
d
(bivariate) characteristic function and that (Xn , Yn ) Ð
→ (X, Y ). Moreover,
h(ξ, η) = φX,Y (ξ, η) = φX (ξ)φY (η)
which shows that X á Y .
∎∎
Problem 1.6. Solution: Necessity is clear. For sufficiency write
1 ⎛ B(t) − B( 2 ) B( 2 ) − B(s) ⎞
s+t s+t
B(t) − B(s) 1
√ =√ ⎜ √ + √ ⎟ =∶ √ (X + Y ) .
t−s 2⎝ t−s t−s ⎠ 2
2 2
By assumption X ∼ Y , X á Y and X ∼ √1 (X + Y ). This is already enough to guarantee

2
that X ∼ N(0, 1), since V X = 1, cf. Rényi [12, Chapter VI.5, Theorem 2, pp. 324–325].
11
Alternative Solution: Fix s < t and define tj ∶= s + nj (t − s) for j = 0, . . . , n. Then

√
√ n B −B
tj tj−1 t − s n Btj − Btj−1
Bt − Bs = tj − tj−1 ∑ √ = ∑ √
j=1 tj − tj−1 n j=1 tj − tj−1
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸n¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
=∶Gj
By assumption, the random variables (Gnj )j,n are identically distributed (for all j, n) and
independent (in j). Moreover, E(Gnj ) = 0 and V(Gnj ) = 1. Applying the central limit
theorem (for triangular arrays) we obtain
1 n d
√ ∑ Gnj Ð
→ G1
n j=1
where G1 ∼ N(0, 1). Thus, Bt − Bs ∼ N(0, t − s).
∎∎
Problem 1.7. Solution: Let ξ, η ∈ R. Then

− + i √ξ (G−G′ ) i √η (G+G′ )
E (eiξ Γ eiη Γ ) = E (e 2 e 2 )
√ G i √ G′
i ξ+η η−ξ
= E (e 2 e 2 )
GáG′ √ G
i ξ+η √ G′
i η−ξ
= E (e 2 ) E (e 2 )
G,G′ ∼N(0,1)
2 2
√ ]
− 12 [ ξ+η √ ]
− 12 [ η−ξ
= e 2 e 2
e− 2 ξ e− 2 η .
1 2 1 2
=
Taking η = 0 or ξ = 0 we find that Γ− ∼ N(0, 1) and Γ+ ∼ N(0, 1), respectively. Moreover,

since ξ, η are arbitrary, we conclude
− + − +
E (eiξ Γ eiη Γ ) = E (eiξ Γ ) E (eiη Γ ) Ô⇒ Γ− á Γ+ .
In the last implication we used Kac’s characterization of independence by characteristic

functions.
∎∎
12
2 Brownian motion as a Gaussian process
Problem 2.1. Solution: Let us check first that f (u, v) ∶= g(u)g(v)(1 − sin u sin v) is indeed a
probability density. Clearly, f (u, v) ⩾ 0. Since g(u) = (2π)−1/2 e−u
2 /2
is even and sin u is
odd, we get
∬ f (u, v) du dv = ∫ g(u) du ∫ g(v) dv − ∫ g(u) sin u du ∫ g(v) sin v dv = 1 − 0.
Moreover, the density fU (u) of U is
fU (u) = ∫ f (u, v) dv = g(u) ∫ g(v) dv − g(u) sin u ∫ g(v) sin v dv = g(u).
This, and a analogous argument show that U, V ∼ N(0, 1).
Let us show that (U, V ) is not a normal random variable. Assume that (U, V ) is normal,
then U + V ∼ N(0, σ 2 ), i. e.
E eiξ(U +V ) = e− 2 ξ
1 2 σ2
. (*)
On the other hand we calculate with f (u, v) that
E eiξ(U +V ) = ∬ eiξu+iξv f (u, v) du dv

2 2
= (∫ eiξu g(u) du) − (∫ eiξu g(u) sin u du)
2
1
= e−ξ − ( −iu
2
∫ e (e − e )g(u) du)
iξu iu
2i
2
1
= e−ξ − ( ∫ (ei(ξ+1)u − ei(ξ−1)u )g(u) du)
2
2i
2
1
= e−ξ − ( (e− 2 (ξ+1) − e− 2 (ξ−1) ))
2 1 2 1 2
2i
1 2 2
= e−ξ + (e− 2 (ξ+1) − e− 2 (ξ−1) )
2 1 2 1
4
1
= e−ξ + e−1 e−ξ (e−ξ − eξ ) ,
2 2 2
4
and this contradicts (*).
∎∎
Problem 2.2. Show that the covariance matrix C = (tj ∧ tk )j,k=1,...,n appearing in Theorem 2.6
is positive definite. Solution: Let (ξ1 , . . . , ξn ) ≠ (0, . . . , 0) and set t0 = 0. Then we find
from (2.12)
n n n
∑ ∑ (tj ∧ tk ) ξj ξk = ∑ (tj − tj−1 )(ξj + ⋯ + ξn ) ⩾ 0.
2
(2.1)
j=1 k=1 j=1 ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
>0
13
Equality (= 0) occurs if, and only if, (ξj + ⋯ + ξn )2 = 0 for all j = 1, . . . , n. This implies
that ξ1 = . . . = ξn = 0.
Abstract alternative: Let (Xt )t∈I be a real-valued stochastic process which has a second
moment (such that the covariance is defined!), set µt = E Xt . For any finite set S ⊂ I we
pick λs ∈ C, s ∈ S. Then
∑ Cov(Xs , Xt )λs λ̄t = ∑ E ((Xs − µs )(Xt − µt ))λs λ̄t

s,t∈S s,t∈S
⎛ ⎞
=E ∑ (Xs − µs )λs (Xt − µt )λt
⎝s,t∈S ⎠
= E ( ∑ (Xs − µs )λs ∑ (Xt − µt )λt )

s∈S t∈S
2
⎛ ⎞
= E ∣ ∑ (Xs − µs )λs ∣ ⩾ 0.
⎝ s∈S ⎠
Remark: Note that this alternative does not prove that the covariance is strictly positive
definite. A standard counterexample is to take Xs ≡ X.
∎∎
Problem 2.3. Solution: These are direct & straightforward calculations.
∎∎
Problem 2.4. Solution: Let ei = (0, . . . , 0, 1, 0 . . .) ∈ Rn be the ith standard unit vector. Then
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
i
aii = ⟨Aei , ei ⟩ = ⟨Bei , ei ⟩ = bii .
Moreover, for i ≠ j, we get by the symmetry of A and B
⟨A(ei + ej ), ei + ej ⟩ = aii + ajj + 2bij
and
⟨B(ei + ej ), ei + ej ⟩ = bii + bjj + 2bij
which shows that aij = bij . Thus, A = B.
We have
Let A, B ∈ Rn×n be symmetric matrices. If ⟨Ax, x⟩ = ⟨Bx, x⟩ for all x ∈ Rn , then A = B.
∎∎
(a) Xt = 2Bt/4 is a BM1 : scaling property with c = 1/4, cf. 2.16.
14
(b) Yt = B2t − Bt is not a BM1 , the independent increments is clearly violated:
E(Y2t − Yt )Yt = E(Y2t Yt ) − E Yt2

= E(B4t − B2t )(B2t − Bt ) − E(B2t − Bt )2
= E(B4t − B2t ) E(B2t − Bt ) − E(B2t − Bt )2
(B1)
= − E(Bt2 ) = −t ≠ 0.
(B1)
√
(c) Zt = tB1 is not a BM1 , the independent increments property is violated:
√ √ √ √ √ √
E(Zt − Zs )Zs = ( t − s) s E B12 = ( t − s) s ≠ 0.
∎∎
Problem 2.6. Solution: We use formula (2.10b).

1 1 x2 (y − x)2
(a) fB(s),B(t) (x, y) = √ exp [− ( + )] .
2π s(t − s) 2 s t−s
(b)
fB(s),B(t)∣B(1) (x, y∣B(1) = z)

fB(s),B(t),B(1) (x, y, z)
=
fB(1) (z)
1 1 x2 (y − x)2 (z − y)2 z2
= √ exp [− ( + + )] (2π)1/2 exp [ ] .
(2π)3/2 s(t − s)(1 − t) 2 s t−s 1−t 2
Thus,
1 1 x2 (y − x)2 y2
fB(s),B(t)∣B(1) (x, y ∣ B(1) = 0) = √ exp [− ( + + )] .
2π s(t − s)(1 − t) 2 s t−s 1−t
Note that
x2 (y − x)2 y2 t s 2 y2 y2 t s 2 y2
+ + = (x − y) + + = (x − y) + .
s t−s 1 − t s(t − s) t t 1 − t s(t − s) t t(1 − t)
Therefore,
E(B(s)B(t) ∣ B(1) = 0)
= ∬ xyfB(s),B(t)∣B(1) (x, y ∣ B(1) = 0) dx dy

1 ∞ 1 y2
= √ ∫ y exp [− ]×
2π s(t − s)(1 − t) y=−∞ 2 t(1 − t)
∞ 1 t s 2
×∫ x exp [− (x − y) ] dx dy
x=−∞ 2 s(t − s) t
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹√¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
√
√
s(t−s)
= 2π s
t
y
t
1 ∞ s 1 y2
=√ √ ∫ y 2 exp [− ] dy
2π t(1 − t) y=−∞ t 2 t(1 − t)
s
= t(1 − t) = s(1 − t).
t
15
(c) In analogy to part b) we get
fB(t2 ),B(t3 )∣B(t1 ),B(t4 ) (x, y ∣ B(t1 ) = u, B(t4 ) = z)

fB(t1 ),B(t2 ),B(t3 ),B(t4 ) (u, x, y, z)
=
fB(t1 ),B(t4 ) (u, z)
1
1 t1 (t4 − t1 ) 2
1 u2 (x − u)2 (y − x)2 (z − y)2
= [ ] exp [− ( + + + )] ×
2π t1 (t2 − t1 )(t3 − t2 )(t4 − t3 ) 2 t1 t2 − t1 t3 − t2 t4 − t3
1 u2 (z − u)2
× exp [ ( + )] .
2 t1 t4 − t1
Thus,
fB(t2 ),B(t3 )∣B(t1 ),B(t4 ) (x, y ∣ B(t1 ) = B(t4 ) = 0)

1
1 t1 (t4 − t1 ) 2
1 x2 (y − x)2 y2
= [ ] exp [− ( + + )] .
2π t1 (t2 − t1 )(t3 − t2 )(t4 − t3 ) 2 t2 − t1 t3 − t2 t4 − t3
Observe that
x2 (y − x)2 y2 t3 − t1 t2 − t1 2 t4 − t1
+ + = (x − y) + y2.
t2 − t1 t3 − t2 t4 − t3 (t2 − t1 )(t3 − t2 ) t3 − t1 (t3 − t1 )(t4 − t3 )
Therefore, we get (using physicists’ notation: ∫ dy h(y) ∶= ∫ h(y) dy for easier read-
ability)
∬ xy fB(t2 ),B(t3 )∣B(t1 ),B(t4 ) (x, y ∣ B(t1 ) = B(t4 ) = 0) dx dy

1 ∞ 1 t4 − t1
= ∫ dy exp [− y2] ×
2π(t4 − t3 ) y=−∞ 2 (t3 − t1 )(t4 − t3 )
y ∞ 1 t2 − t1 2 t3 − t1
×√ ∫ x exp [− (x − y) ] dx
2π(t2 − t1 )(t3 − t2 ) x=−∞ 2 t 3 − t 1 (t 2 t1 )(t3 − t2 )
−
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
t2 −t1
=√ y
2
t3 −t1 t3 −t1
t2 − t1 (t4 − t3 )(t3 − t1 ) (t2 − t1 )(t4 − t3 )

= = .
t3 − t1 t4 − t1 t4 − t1
∎∎
Problem 2.7. Solution: Let s ⩽ t. Then
C(s, t) = E(Xs Xt )
= E(Bs2 − s)(Bt2 − t)
= E(Bs2 − s)([Bt − Bs + Bs ]2 − t)
= E(Bs2 − s)(Bt − Bs )2 + 2 E(Bs2 − s)Bs (Bt − Bs ) + E(Bs2 − s)Bs2 − E(Bs2 − s)t
= E(Bs2 − s) E(Bt − Bs )2 + 2 E(Bs2 − s)Bs E(Bt − Bs ) + E(Bs2 − s)Bs2 − E(Bs2 − s)t
(B1)
= 0 ⋅ (t − s) + 2 E(Bs2 − s)Bs ⋅ 0 + E Bs4 − s E Bs2 − 0

= 2s2 = 2(s2 ∧ t2 ) = 2(s ∧ t)2 .
∎∎
16
(a) We have for s, t ⩾ 0
m(t) = E Xt = e−αt/2 E Beαt = 0.

C(s, t) = E(Xs Xt ) = e− 2 (s+t) E Beαs Beαt = e− 2 (s+t) (eαs ∧ eαt ) = e− 2 ∣t−s∣ .
α α α
(b) We have
P(X(t1 ) ⩽ x1 , . . . , X(tn ) ⩽ xn ) = P (B(eαt1 ) ⩽ eαt1 /2 x1 , . . . , B(eαtn ) ⩽ eαtn /2 xn )
Thus, the density is
fX(t1 ),...,X(tn ) (x1 , . . . , xn )

n
= ∏ eαtk /2 fB(eαt1 ),...,B(eαtn ) (eαt1 /2 x1 , . . . , eαtn /2 xn )
k=1
−1/2
αtk /2 x /2
n n
αtk /2 −n/2
e− 2 ∑k=1 (e k −e k−1 xk−1 ) /(e k −e k−1 )
1 n αt 2 αt αt
= ∏e (2π) ( ∏ (e αtk
−e αtk−1
))
k=1 k=1
−1/2
−α(tk −tk−1 )/2
xk−1 )2 /(1−eα(tk −tk−1 ) )
n
−n/2 −α(tk −tk−1 )
e− 2 ∑k=1 (xk −e
1 n
= (2π) ( ∏ (1 − e ))
k=1
(we use the convention t0 = −∞ and x0 = 0).
Remark: the form of the density shows that the Ornstein–Uhlenbeck is strictly stationary,
i. e.
(X(t1 + h), . . . , X(tn + h) ∼ (X(t1 ), . . . , X(tn )) ∀h > 0.
∎∎
Problem 2.9. Solution: Set
Σ ∶= ⋃ σ(B(t) ∶ t ∈ J)
J⊂[0,∞), J countable
Clearly,
⋃ σ(Bt ) ⊂ Σ ⊂ σ(Bt ∶ t ⩾ 0) = F∞
B def
(*)
t⩾0
The first inclusion follows from the fact that each Bt is measurable with respect to Σ.
Let us show that Σ is a σ-algebra. Obviously,
∅∈Σ and F ∈ Σ Ô⇒ F c ∈ Σ.
Let (An )n ⊂ Σ. Then, for every n there is a countable set Jn such that An ∈ σ(B(t) ∶ t ∈
Jn ). Since J = ⋃n Jn is still countable we see that An ∈ σ(B(t) ∶ t ∈ J) for all n. Since
the latter family is a σ-algebra, we find
⋃ An ∈ σ(B(t) ∶ t ∈ J) ⊂ Σ.
n
17
Since ⋃t σ(Bt ) ⊂ Σ, we get—note: F∞

B
is, by definition, the smallest σ-algebra for which
all Bt are measurable—that
F∞
B
⊂ Σ.
This shows that Σ = F∞

B
.
∎∎
Problem 2.10. Solution: Assume that the indices t1 , . . . , tm and s1 , . . . , sn are given. Let
{u1 , . . . , up } ∶= {s1 , . . . , sn } ∪ {t1 , . . . , tm }. By assumption,
(X(u1 ), . . . , X(up )) á (Y (u1 ), . . . , Y (up )).
Thus, we may thin out the indices on each side without endangering independence:
{s1 , . . . , sn } ⊂ {u1 , . . . , up } and {t1 , . . . , tm } ⊂ {u1 , . . . , up }, and so
(X(s1 ), . . . , X(sn )) á (Y (t1 ), . . . , Y (tm )).
∎∎
Problem 2.11. Solution: Since Ft ⊂ F∞ and Gt ⊂ G∞ it is clear that
F∞ á G∞ Ô⇒ Ft á Gt .
Conversely, since (Ft )t⩾0 and (Gt )t⩾0 are filtrations we find
∀F ∈ ⋃ Ft , ∀G ∈ ⋃ Gt , ∃t0 ∶ F ∈ Ft0 , G ∈ Gt0 .

t⩾0 t⩾0
By assumption: P(F ∩ G) = P(F ) P(G). Thus,
⋃ Ft á ⋃ Gt .
t⩾0 t⩾0
Since the families ⋃t⩾0 Ft and ⋃t⩾0 Gt are ∩-stable (use again the argument that we have
filtrations to find for F, F ′ ∈ ⋃t⩾0 Ft some t0 with F, F ′ ∈ Ft0 etc.), the σ-algebras generated
by these families are independent:
F∞ = σ ( ⋃ Ft ) á σ ( ⋃ Gt ) = G∞ .
t⩾0 t⩾0
∎∎
Problem 2.12. Solution: Let U ∈ Rd×d be an orthogonal matrix: U U ⊺ = id and set Xt ∶= U Bt

for a BMd (Bt )t⩾0 . Then
⎡ n ⎤ ⎡ n ⎤
⎛ ⎢ ⎥⎞ ⎛ ⎢ ⎥⎞
E exp ⎢⎢i ∑ ⟨ξj , X(tj ) − X(tj−1 )⟩⎥⎥ = E exp ⎢⎢i ∑ ⟨ξj , U B(tj ) − U B(tj−1 )⟩⎥⎥
⎝ ⎢ j=1 ⎥⎠ ⎝ ⎢ j=1 ⎥⎠
⎣ ⎦ ⎣ ⎦
⎡ n ⎤
⎛ ⎢ ⎥⎞
= E exp ⎢⎢i ∑ ⟨U ⊺ ξj , B(tj ) − B(tj−1 )⟩⎥⎥
⎝ ⎢ j=1 ⎥⎠
⎣ ⎦
18
⎡ ⎤
⎢ 1 n ⎥
= exp ⎢⎢− ∑ (tj − tj−1 )⟨U ⊺ ξj , U ⊺ ξj ⟩⎥⎥
⎢ 2 j=1 ⎥
⎣ ⎦
⎡ ⎤
⎢ 1 n ⎥
= exp ⎢⎢− ∑ (tj − tj−1 )∣ξj ∣2 ⎥⎥ .
⎢ 2 j=1 ⎥
⎣ ⎦
(Observe ⟨U ⊺ ξj , U ⊺ ξj ⟩ = ⟨U U ⊺ ξj , ξj ⟩ = ⟨ξj , ξj ⟩ = ∣ξj ∣2 ). The claim follows.
∎∎
Problem 2.13. Solution: Note that the coordinate processes b and β are independent BM1 .
√
(a) Since b á β, the process Wt = (bt + βt )/ 2 is a Gaussian process with continuous
sample paths. We determine its mean and covariance functions:
1
E Wt = √ (E bt + E βt ) = 0;
2
Cov(Ws , Wt ) = E(Ws Wt )
1
= E(bs + βs )(bt + βt )
2
1
= ( E bs bt + E βs bt + E bs βt + E βs βt )
2
1
= (s ∧ t + 0 + 0 + s ∧ t) = s ∧ t
2
where we used that, by independence, E bu βv = E bu E βv = 0. Now the claim follows
from Corollary 2.7.
(b) The process Xt = (Wt , βt ) has the following properties
• W and β are BM1

√
• E(Wt bt ) = 2−1/2 E(bt + βt )βt = 2−1/2 (E bt E βt + E βt2 ) = t/ 2 ≠ 0, i. e. W and β
are NOT independent.
This means that X is not a BM2 , as its coordinates are not independent.
The process Yt can be written as
1 ⎛bt + βt ⎞ ⎛ bt ⎞ 1 ⎛1 1 ⎞ ⎛ bt ⎞
√ =U =√ .
2 ⎝bt − βt ⎠ ⎝βt ⎠ 2 ⎝1 −1⎠ ⎝βt ⎠
Clearly, U U ⊺ = id, i. e. Problem 2.12 shows that (Yt )t⩾0 is a BM2 .
∎∎
Problem 2.14. Solution: Observe that b á β since B is a BM2 . Since
E Xt = 0
Cov(Xt , Xs ) = E Xt Xs
= E(λbs + µβs )(λbt + µβt )
= λ2 E bs bt + λµ E bs βt + λµ E bt βs + µ2 βs βt
19
= λ2 E bs bt + λµ E bs E βt + λµ E bt E βs + µ2 E βs βt
= λ2 (s ∧ t) + 0 + 0 + µ2 s ∧ t = (λ2 + µ2 )(s ∧ t).
Thus, by Corollary 2.7, X is a BM1 if, and only if, λ2 + µ2 = 1.

∎∎
Problem 2.15. Solution: Xt = (bt , βs−t − βt ), 0 ⩽ t ⩽ s, is NOT a Brownian motion: X0 =

(0, βs ) ≠ (0, 0).
On the other hand, Yt = (bt , βs−t − βs ), 0 ⩽ t ⩽ s, IS a Brownian motion, since bt and
βs−t − βs are independent BM1 , cf. Time inversion 2.15 and Theorem 2.10.
∎∎
Problem 2.16. Solution: We have

⎛ cos α sin α ⎞ ⎛ bt ⎞
Wt = U Bt⊺ = .
⎝− sin α cos α⎠ ⎝βt ⎠
The matrix U is a rotation, hence orthogonal and we see from Problem 2.12 that W is a
Brownian motion.
Generalization: take U orthogonal.
∎∎
Problem 2.17. Solution: If G ∼ N(0, Q) then Q is the covariance matrix, i. e. Cov(Gj , Gk ) =

qjk . Thus, we get for s < t
Cov(Xsj , Xtk ) = E(Xsj Xtk )

= E Xsj (Xtk − Xsk ) + E(Xsj Xsk )
= E Xsj E(Xtk − Xsk ) + sqjk
= (s ∧ t)qjk .
The characteristic function is

⊺ ξ,B ⊺ ξ∣2 ⊺ ξ⟩
E ei⟨ξ,Xt ⟩ = E ei⟨Σ t⟩
= e− 2 ∣Σ = e− 2 ⟨ξ,ΣΣ
t t
,
and the transition probability is, if Q is non-degenerate,

1 1
fQ (x) = √ exp (− ⟨x, Qx⟩) .
(2πt)n det Q 2t
If Q is degenerate, there is an orthogonal matrix U ∈ Rn×n such that
U Xt = (Yt1 , . . . , Ytk , 0, . . . , 0)⊺

´¹¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹¶
n−k
where k < n is the rank of Q. The k-dimensional vector has a nondegenerate normal
distribution in Rk .
∎∎
20
“⇒” Assume that we have (B1). Observe that the family of sets
⋃ σ(Bu1 , . . . , Bun )
0⩽u1 ⩽⋯⩽un ⩽s, n⩾1
is a ∩-stable family. This means that it is enough to show that
Bt − Bs á (Bu1 , . . . , Bun ) for all t ⩾ s ⩾ 0.
By (B1) we know that
Bt − Bs á (Bu1 , Bu2 − Bu1 , . . . , Bun − Bun−1 )
and so
⎛1 0 0 . . . 0⎞ ⎛ Bu1 ⎞ ⎛ Bu1 ⎞
⎜ ⎟⎜ ⎟ ⎜ ⎟
⎜1 1 0 . . . 0⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ Bu2 − Bu1 ⎟ ⎜ Bu2 ⎟
⎜ ⎟⎜ ⎟ ⎜ ⎟
Bt − Bs á ⎜1 1 1 . . . 0⎟ ⎜ ⎟=⎜ ⎟
⎜ ⎟ ⎜ Bu3 − Bu2 ⎟ ⎜ Bu3 ⎟
⎜ ⎟⎜ ⎟ ⎜ ⎟
⎜⋮ ⋮ ⋮ ⋱ 0⎟ ⎜ ⋮ ⎟ ⎜ ⋮ ⎟
⎜ ⎟⎜ ⎟ ⎜ ⎟
⎝1 1 1 . . . 1⎠ ⎝Bun − Bun−1 ⎠ ⎝Bun ⎠
“⇐” Let 0 = t0 ⩽ t1 < t2 < . . . < tn < ∞, n ⩾ 1. Then we find for all ξ1 , . . . , ξn ∈ Rd
E (ei ∑k=1 ⟨ξk , B(tk )−B(tk−1 )⟩ ) = E (ei⟨ξn , B(tn )−B(tn−1 )⟩ ⋅ ei ∑k=1 ⟨ξk , B(tk )−B(tk−1 )⟩ )
n n−1
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
Ftn−1 mble., hence áB(tn )−B(tn−1 )
= E (ei⟨ξn , B(tn )−B(tn−1 )⟩ ) ⋅ E (ei ∑k=1 ⟨ξk , B(tk )−B(tk−1 )⟩ )

n−1
⋮
n
= ∏ E (ei⟨ξk , B(tk )−B(tk−1 )⟩ ).
k=1
This shows (B1).
∎∎
Problem 2.19. Solution: Reflection invariance of BM, cf. 2.12, shows
τa = inf{s ⩾ 0 ∶ Bs = a} ∼ inf{s ⩾ 0 ∶ −Bs = a} = inf{s ⩾ 0 ∶ Bs = −a} = τ−a .
The scaling property 2.16 of BM shows for c = 1/a2
τa = inf{s ⩾ 0 ∶ Bs = a} ∼ inf{s ⩾ 0 ∶ aBs/a2 = a}

= inf{a2 r ⩾ 0 ∶ aBr = a}
= a2 inf{r ⩾ 0 ∶ Br = 1} = a2 τ1 .
∎∎
21
(a) Not stationary:
E Wt2 = C(t, t) = E(Bt2 − t)2 = E(Bt4 − 2tBt2 + t2 ) = 3t2 − 2t2 + t2 = 2t2 ≠ const.
(b) Stationary. We have E Xt = 0 and
E Xs Xt = e−α(t+s)/2 E Beαs Beαt = e−α(t+s)/2 (eαs ∧ eαt ) = e−α∣t−s∣/2 ,
i. e. it is stationary with g(r) = e−α∣r∣/2 .

(c) Stationary. We have E Yt = 0. Let s ⩽ t. Then we use E Bs Bt = s ∧ t to get
E Ys Yt = E(Bs+h − Bs )(Bt+h − Bt )
= E Bs+h Bt+h − E Bs+h Bt − E Bs Bt+h + E Bs Bt
= (s + h) ∧ (t + h) − (s + h) ∧ t − s ∧ (t + h) + s ∧ t
⎧
⎪
⎪
⎪0, if t > s + h ⇐⇒ h < t − s
= (s + h) − (s + h) ∧ t = ⎨
⎪
⎪
⎩h − (t − s),
⎪ if t ⩽ s + h ⇐⇒ h ⩾ t − s.
Swapping the roles of s and t finally gives: the process is stationary with g(t) =
(h − ∣t∣)+ = (h − ∣t∣) ∨ 0.
(d) Not stationary. Note that
E Zt2 = E Be2t = et ≠ const.
∎∎
Problem 2.21. Solution: Clearly, t ↦ Wt is continuous for t ≠ 1. If t = 1 we get
lim Wt (ω) = W1 (ω) = B1 (ω)

t↑1
and
lim Wt (ω) = B1 (ω) − lim tβ1/t (ω) − β1 (ω) = B1 (ω);
t↓1 t↓1
this proves continuity for t = 1.

Let us check that W is a Gaussian process with E Wt = 0 and E Ws Wt = s ∧ t. By
Corollary 2.7, W is a BM1 .
Pick n ⩾ 1 and t0 = 0 < t1 < . . . < tn .
Case 1: If tn ⩽ 1, there is nothing to show since (Bt )t∈[0,1] is a BM1 .
Case 2: Assume that tn > 1. Then we have
⎛ B1 ⎞
⎛ Wt1 ⎞ ⎛1 t1 0 0 ⋯ 0 −1⎞ ⎜ ⎟
⎜ ⎟ ⎜ ⎟⎜⎜ β1/t1 ⎟
⎟
⎜ Wt2 ⎟ ⎜1 0 t2 0 ⋯ −1⎟
⎜ ⎟=⎜ 0 ⎟⎜⎜ ⋮
⎟
⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟
⎜ ⋮ ⎟ ⎜ ⋮ ⋮ 0 t3 ⋯ ⋮ ⋮ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟ ⎜β
⎜ 1/tn ⎟
⎟
⎝Wtn ⎠ ⎝1 0 0 0 ⋯ tn −1 ⎠
⎝ β1 ⎠
22
and since
B1 á (β1/t1 , . . . , β1/tn , β1 )⊺
are both Gaussian, we see that (Wt1 , . . . , Wtn ) is Gaussian.

Further, let t ⩾ 1 and 1 ⩽ ti < tj :
E Wt = E B1 + t E β1/t − E β1 = 0
E Wti Wtj = E(B1 + ti β1/ti − β1 )(B1 + tj β1/tj − β1 )
= 1 + ti tj t−1 −1 −1
j − ti ti − tj tj + 1 = ti = ti ∧ tj .
Case 3: Assume that 0 < t1 < . . . < tk ⩽ 1 < tk+1 < . . . < tn . Then we have
⎛ Bt1 ⎞
⎛ 1 0 ⋯ 0 ⎞⎜ ⎟
⎜ ⎟⎜⎜ ⋮ ⎟ ⎟
⎛ Wt1 ⎞ ⎜
⎜ 0 ⋱ 0 ⎟⎜
⎟⎜ ⎟
⎜ ⎟ ⎜ ⎟⎜ ⋮ ⎟ ⎟
⎜ Wt2 ⎟ ⎜ ⋮ ⋱ ⋮ ⎟⎜ ⎟
⎜ ⎟ ⎜ ⎟⎜
⎟ ⎜ Btk ⎟
⎜ ⎟ ⎜ ⎟
⎜ ⋮ ⎟ ⎜ ⋯ 1 ⎟⎜ ⎟
⎜ ⎟=⎜ 0 0 ⎟⎜
⎟ ⎜ B1 ⎟
⎜ ⎟ ⎜ ⎟.
⎜Wtk ⎟ ⎜ ⋯ −1 ⎟⎜ ⎟
⎜ ⎟ ⎜ 1 tk+1 0 0 ⎟⎜
⎟ ⎜β1/tk+1 ⎟
⎜ ⎟ ⎜ ⎟
⎜ ⋮ ⎟ ⎜ −1 ⎟⎜ ⎟
⎜ ⎟ ⎜
⎜
1 0 tk+2 0 ⎟⎜
⎟⎜ ⋮ ⎟
⎝Wtn ⎠ ⎜ ⎟⎜ ⎟
⎜ ⋮ ⋮ ⋱ ⋮ ⎟⎜ ⎟
⎜ ⎟ ⎜ β1/tn ⎟⎟
⎝ 1 0 ⋯ tn −1 ⎠⎜ ⎟
⎝ β1 ⎠
Since
(Bt1 , . . . , Btk , B1 ) á (β1/tk+1 , . . . , β1/tn , β1 )
are Gaussian vectors, (Wt1 , . . . , Wtn ) is also Gaussian and we find
E Wt = 0
E Wti Wtj = E Bti (B1 + tj β1/tj − β1 ) = ti = ti ∧ tj
for i ⩽ k < j.
∎∎
Problem 2.22. Solution: The process X(t) = B(et ) has no memory since (cf. Problem 2.18)
σ(B(s) ∶ s ⩽ ea ) á σ(B(s) − B(ea ) ∶ s ⩾ ea )
and, therefore,
σ(X(t) ∶ t ⩽ a) = σ(B(s) ∶ 1 ⩽ s ⩽ ea ) á σ(B(ea+s ) − B(ea ) ∶ s ⩾ 0)

= σ(X(t + a) − X(a) ∶ t ⩾ 0).
The process X(t) ∶= e−t/2 B(et ) is not memoryless. For example, X(a + a) − X(a) is not
independent of X(a):
E(X(2a) − X(a))X(a) = E (e−a B(e2a ) − e−a/2 B(ea ))e−a/2 B(ea ) = e−3a/2 ea − e−a ea ≠ 0.
∎∎
23
Problem 2.23. Solution: The process Wt = Ba−t − Ba , 0 ⩽ t ⩽ a clearly satisfies (B0) and (B4).
For 0 ⩽ s ⩽ t ⩽ a we find
Wt − Ws = Ba−t − Ba−s ∼ Ba−s − Ba−t ∼ Bt−s ∼ N(0, (t − s) id)
and this shows (B2) and (B3).
For 0 = t0 < t1 < . . . < tn ⩽ a we have
Wtj − Wtj−1 = Ba−tj − Ba−tj−1 ∼ Ba−tj−1 − Ba−tj ∀j
and this proves that W inherits (B1) from B.
∎∎
Problem 2.24. Solution: We know from Paragraph 2.17 that
B(s)
lim tB(1/t) = 0 Ô⇒ lim =0 a.s.
t↓0 s↑∞ s
Moreover,
2
B(s) s 1 s→∞
E( ) = 2 = ÐÐÐ→ 0
s s s
i. e. we get also convergence in mean square.
Remark: a direct proof of the SLLN is a bit more tricky. Of course we have by the classical
SLLN that
Bn ∑j=1 (Bj − Bj−1 ) SLLN
n
= ÐÐÐ→ 0 a.s.
n n n→∞
But then we have to make sure that Bs /s converges. This can be done in the following
way: fix s > 0. Then there is a unique interval (n, n + 1] such that s ∈ (n, n + 1]. Thus,
Bs Bs − Bn+1 Bn+1 n + 1 supn⩽s⩽n+1 ∣Bs − Bn+1 ∣ n + 1 Bn

∣ ∣⩽∣ ∣+∣ ∣⋅ ⩽ + ∣ ∣
s s n+1 s n n n
and we have to show that the expression with the sup tends to zero. This can be done
by showing, e.g., that the L2 -limit of this expression goes to zero (using the reflection
principle) and with a subsequence argument.
∎∎
24
3 Constructions of Brownian motion
Problem 3.1. Solution: The partial sums

N −1
WN (t, ω) = ∑ Gn (ω)Sn (t), t ∈ [0, 1],
n=0
converge as N → ∞ P-a.s. uniformly for t towards B(t, ω), t ∈ [0, 1]—cf. Problem 3.3.
Therefore, the random variables
1 N −1 1 1
P -a.s.
∫ WN (t) dt = ∑ Gn ∫ Sn (t) dt ÐÐÐ→ X = ∫ B(t) dt.
0 n=0 0 N →∞ 0
1
This shows that ∫0 WN (t) dt is the sum of independent N(0, 1)-random variables, hence
itself normal and so is its limit X.
From the definition of the Schauder functions (cf. Figure 3.2) we find
1 1
∫ S0 (t) dt =
0 2
1 1 −3 j
∫ S2j +k (t) dt = 2 2 , k = 0, 1, . . . , 2j − 1, j ⩾ 0.
0 4
and this shows
1 n 2 −1 3
j
1 1
∫ W2n+1 (t) dt = G0 + ∑ ∑ 2− 2 j G2j +l .
0 2 4 j=0 l=0
Consequently, since the Gj are iid N(0, 1) random variables,
1
E∫ W2n+1 (t) dt = 0,
0
1 1 n 2 −1 −3j
j
1
V∫ W2n+1 (t) dt = + ∑ ∑ 2
0 4 16 j=0 l=0
1 1 n −2j
= + ∑2
4 16 j=0
1 1 1 − 2−2(n+1)
= +
4 16 1 − 41
1 1 4 1
ÐÐÐ→ + = .
n→∞ 4 16 3 3
This means that
∞
1 3 2 −1
j
1
X = G0 + ∑ 2− 2 j ∑ G2j +l
2 j=0 4 l=0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
∼N(0,2j )
where the series converges P-a.s. and in mean square, and X ∼ N(0, 13 ).
∎∎
25
Problem 3.2. Solution: Denote by λ Lebesgue measure on [0, 1].

(a) By the independence of the random variables Gn ∼ N(0, 1) and Parseval’s identity,
we have for M < N
⎡ N −1 ⎤
⎢ ⎥
E(∣WN (B) − WM (B)∣ ) = E ⎢ ∑ Gm Gn ⟨1B , φm ⟩L2 ⟨1B , φn ⟩L2 ⎥⎥
2 ⎢
⎢m,n=M ⎥
⎣ ⎦
N −1
= ∑ E(Gm Gn )⟨1B , φm ⟩L2 ⟨1B , φn ⟩L2
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
m,n=M
= 0 (n≠m), or = 1 (n=m)
N −1
= ∑ ⟨1B , φn ⟩2L2 ÐÐÐÐÐ→ 0.
n=M M,N →∞
This shows that W (B) = L2 (P)-limN →∞ WN (B) exists.

(b) We have E W (A)W (B) = λ(A ∩ B). This can be seen as follows: Using the Cauchy-
Schwarz inequality, we find
E(∣WN (A)WN (B) − W (A)W (B)∣)

⩽ E(∣WN (A)(WN (B) − W (B))∣) + E(∣W (B)(WN (A) − W (A))∣)
√ √ √ √
⩽ E(∣WN (A)∣2 ) E(∣WN (B) − W (B)∣2 ) + E(∣W (B)∣2 ) E(∣WN (A) − W (A)∣2 )
By part a), W (A) = L2 (P)-limN →∞ WN (A) and W (B) = L2 (P)-limN →∞ WN (B),

and therefore this calculation shows W (A)W (B) = L1 (P)-limN →∞ WN (A)WN (B).
A similar calculation as in the first part yields
E(W (A)W (B)) = lim E(WN (A)WN (B))

N →∞
N −1
= lim ∑ E(Gm Gn )⟨1A , φm ⟩L2 ⟨1B , φn ⟩L2
N →∞ m,n=0 ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
=0 (n≠m), or =1 (n=m)
N −1
= lim ∑ ⟨1A , φn ⟩L2 ⟨1B , φn ⟩L2
N →∞ n=0
= λ(A ∩ B)
where we used Parseval’s identity for the last step.

(c) We have seen in part b) that
µ(B) ∶= E(∣W (B)∣2 ) = λ(B).
Consequently, µ is a measure. In contrast, the mapping B ↦ W (B) is not non-

negative (and random), hence it does not define a measure on [0, 1]. In fact, it is
not even a signed measure: Since W (B) is a random variable (and an element of the
space L2 consisting of equivalence classes), it is only defined up to a null set, and
the null set can (and will) depend on B. This means that we run into difficulties
when we consider σ-additivity, since there are more than countably many ways to
represent a set B as a countable union of disjoint sets Bj . Thus, the exceptional null
sets may become uncontrollable . . .
26
(d) Let
l m
f (t) = ∑ fi 1Ai (t) = ∑ gj 1Bj (t)
i=1 j=1
representations of a step function f ∈ S. Without loss of generality we may assume

[0, 1] = ⋃li=1 Ai = ⋃m
j=1 Bj . Choose a common disjoint refinement of A1 , . . . , Al and
B1 , . . . , Bm , i. e. disjoint sets C1 , . . . , Cn ∈ B[0, 1] such that
Ai = ⊍ Ck and Bj = ⊍ Ck .
k∶Ck ⊂Ai k∶Ck ⊂Bj
For
hk ∶= ∑ fi = ∑ gj
i∶Ck ⊂Ai j∶Ck ⊂Bj
we have
l m n
f (t) = ∑ fi 1Ai (t) = ∑ gj 1Bj (t) = ∑ hk 1Ck (t).
i=1 j=1 k=1
Since (all limits in the next calculation are L2 (P)-limits)
N −1
W (A ∪ B) = lim ∑ Gn ⟨1A + 1B , φn ⟩L2
N →∞ n=0
N −1 N −1
= lim ∑ Gn ⟨1A , φn ⟩L2 + lim ∑ Gn ⟨1B , φn ⟩L2
N →∞ n=0 N →∞ n=0
= W (A) + W (B)
for any two disjoint sets A, B ∈ B[0, 1], we conclude
l l ⎛ ⎞
∑ fi W (Ai ) = ∑ fi W ⊍ Ck
i=1 i=1 ⎝k∶Ck ⊂Ai ⎠
l
=∑ ∑ fi W (Ck )
i=1 k∶Ck ⊂Ai
n
=∑ ∑ fi W (Ck )
k=1 i∶Ck ⊂Ai
n
= ∑ dk W (Ck )
k=1
m
= . . . = ∑ gj W (Bj ).
j=1
(e) Let f ∈ S be given by

m
f (t) = ∑ cj 1Bj (t)
j=1
where cj ∈ R, Bj ∈ B[0, 1] for j = 1, . . . , m. Using the definition of I, we get

m m
E(∣I(f )∣2 ) = ∑ ∑ cj ck E(W (Bj )W (Bk ))
j=1 k=1
27
m m
= ∑ ∑ cj ck λ(Bj ∩ Bk )
b)
j=1 k=1
2
1⎛m ⎞ 1
=∫ ∑ cj 1Bj (t) λ(dt) = ∫ ∣f ∣ dλ.
2
0 ⎝j=1 ⎠ 0
Since the family of step functions S is dense in L2 ([0, 1], λ), the isometry allows us
to extend the operator I to L2 ([0, 1], λ): Let f ∈ L2 ([0, 1], λ), then there exists a
sequence (fn )n∈N ⊂ S such that fn → f in L2 ([0, 1], λ). From
1
E(∣I(fn ) − I(fm )∣2 ) = E(∣I(fn − fm )∣2 ) = ∫ ∣fn − fm ∣2 dλ
0
we see that the sequence (I(fn ))n∈N is a Cauchy sequence in L2 (P). Therefore, the
limit
I(f ) ∶= L2 (P)- lim I(fn )
n→∞
exists. Note that the isometry implies that I(f ) does not depend on the approxi-
mating sequence (fn )n∈N . Consequently, I is well-defined.
Remark: Using the results from Section 3.1 it is clear that Wt ∶= W ([0, t]), t ∈ [0, 1], has
all properties of a Brownian motion. As usual, the continuity of the paths t ↦ Wt is not
obvious and needs arguments along the lines of, say, the Lévy–Ciesielski construction in
Section 3.2.
∎∎

(a) From the definition of the Schauder functions Sn (t), n ⩾ 0, t ∈ [0, 1], we find
0 ⩽ Sn (t) ∀n, t
1 −j/2
S2j +k (t) ⩽ S2j +k ((2k + 1)/2j+1 ) = 2−j/2 /2j+1 = 2 ∀j, k, t
2
2j −1
1 −j/2
∑ S2j +k (t) ⩽ 2 (disjoint supports!)
k=0 2
By assumption,
∃C > 0, ∃ ∈ (0, 21 ), ∀n ∶ ∣an ∣ ⩽ C ⋅ n .
Thus, we find
∞ ∞ 2j −1
∑ ∣an ∣Sn (t) ⩽ ∣a0 ∣ + ∑ ∑ ∣a2j +k ∣S2j +k (t)
n=0 j=0 k=0
∞ 2j −1
⩽ ∣a0 ∣ + ∑ ∑ C ⋅ (2j+1 ) S2j +k (t)
j=0 k=0
∞
1 −j
⩽ ∣a0 ∣ + ∑ C ⋅ 2(j+1) 2 < ∞.
j=0 2
(The series is convergent since < 1/2).

This shows that ∑∞
n=0 an Sn (t) converges absolutely and uniformly for t ∈ [0, 1].
28
√
(b) For C > 2 we find from
√ √
√ 2 1 2 1 −C 2 /2
e− 2 C log n ⩽
1 2
P (∣Gn ∣ > log n) ⩽ √ n ∀n ⩾ 3
π C log n π C
that the following series converges:
∞ √
∑ P (∣Gn ∣ > log n) < ∞.
n=1
√
By the Borel–Cantelli Lemma we find that Gn (ω) = O( log n) for almost all ω, thus
Gn (ω) = O(n ) for any ∈ (0, 1/2).
From part a) we know that the series ∑∞

n=0 Gn (ω)Sn (t) converges a.s. uniformly for
t ∈ [0, 1].
∎∎
1/p
Problem 3.4. Solution: Set ∥f ∥p ∶= ( E ∣f ∣p )
Solution 1: We observe that the space Lp (Ω, A, P; S) = {X ∶ X ∈ S, ∥d(X, 0)∥p < ∞} is

complete and that the condition stated in the problem just says that (Xn )n is a Cauchy
sequence in the space Lp (Ω, A, P; S). A good reference for this is, for example, the mono-
graph by F. Trèves [17, Chapter 46]. You will find the ‘pedestrian’ approach as Solution
2 below.
Solution 2: By assumption
∀k ⩾ 0 ∃Nk ⩾ 1 ∶ sup ∥d(XNk , Xm )∥p ⩽ 2−k .

m⩾Nk
Without loss of generality we can assume that Nk ⩽ Nk+1 . In particular

l−1
∀l>k 2
∥d(XNk , XNk+1 )∥p ⩽ 2−k Ô⇒ ∥d(XNk , XNl )∥p ⩽ ∑ 2−j ⩽ .
j=k 2k
Fix m ⩾ 1. Then we see that
∥d(XNk , Xm ) − d(XNl , Xm )∥p ⩽ ∥d(XNk , XNl )∥p ÐÐÐ→ 0.

k,l→∞
This means that that (d(XNk , Xm ))k⩾0 is a Cauchy sequence in Lp (P; R). By the com-
pleteness of the space Lp (P; R) there is some fm ∈ Lp (P; R) such that
in Lp
d(XNk , Xm ) ÐÐÐ→ fm
k→∞
and, for a subsequence (nk ) ⊂ (Nk )k we find

almost surely
d(Xnk , Xm ) ÐÐÐÐÐÐÐ→ fm .
k→∞
The subsequence nk may also depend on m. Since (nk (m))k is still a subsequence of
(Nk ), we still have d(Xnk (m) , Xm+1 ) → fm+1 in Lp , hence we can find a subsequence
29
(nk (m + 1))k ⊂ (nk (m))k such that d(Xnk (m+1) , Xm+1 ) → fm+1 a.s. Iterating this we see
that we can assume that (nk )k does not depend on m.
In particular, we have almost surely
∀ > 0 ∃L = L() ⩾ 1 ∀k ⩾ L ∶ ∣d(Xnk , Xm ) − fm ∣ ⩽ .
Moreover,
lim ∥fm ∥p = lim ∥ lim d(Xnk , Xm )∥p ⩽ lim lim ∥d(Xnk , Xm )∥p
m→∞ m→∞ k→∞ m→∞ k→∞
⩽ lim sup ∥d(Xnk , Xm )∥p = 0.

k→∞ m⩾nk
Thus, fm → 0 in Lp and, for a subsequence mk we get
∀ > 0 ∃K = K() ⩾ 1 ∀r ⩾ K ∶ ∣fmr ∣ ⩽ .
Therefore,
d(Xnk , Xnl ) ⩽ d(Xnk , Xmr ) + d(Xnk , Xmr )

⩽ ∣d(Xnk , Xmr ) − fmr ∣ + ∣d(Xnk , Xmr ) − fmr ∣ + 2∣fmr ∣.
Fix > 0 and pick r > K. Then let k, l → ∞. This gives
d(Xnk , Xnl ) ⩽ ∣d(Xnk , Xmr ) − fmr ∣ + ∣d(Xnk , Xmr ) − fmr ∣ + 2 ⩽ 4 ∀k, l ⩾ L().
Since S is complete, this proves that (Xnk )k⩾0 converges to some X ∈ S almost surely.
Remark: If we replace the condition of the Problem by
lim E (sup dp (Xn , Xm )) = 0

n→∞ m⩾n
things become MUCH simpler:
This condition says that the sequence dn ∶= supm⩾n dp (Xn , Xm ) converges in Lp (P; R) to
zero. Hence there is a subsequence (nk )k such that
lim sup d(Xnk , Xm ) = 0

k→∞ m⩾nk
almost surely. This shows that d(Xnk , Xnl ) → 0 as k, l → ∞, i. e. we find by the complete-
ness of the space S that Xnk → X.
∎∎
Problem 3.5. Solution: Fix n ⩾ 1, 0 ⩽ t1 ⩽ . . . ⩽ tn and Borel sets A1 , . . . , An . By assumption,

we know that
⎛n ⎞
P(Xt = Yt ) = 1 ∀t ⩾ 0 Ô⇒ P(Xtj = Ytj j = 1, . . . , n) = P ⋂ {Xtj = Ytj } = 1.
⎝j=1 ⎠
30
Thus,
⎛n ⎞ ⎛n n ⎞
P ⋂ {Xtj ∈ Aj } = P ⋂ {Xtj ∈ Aj } ∩ ⋂ {Xtj = Ytj }
⎝j=1 ⎠ ⎝j=1 j=1 ⎠
⎛n ⎞
= P ⋂ {Xtj ∈ Aj } ∩ {Xtj = Ytj }
⎝j=1 ⎠
⎛n ⎞
= P ⋂ {Ytj ∈ Aj } ∩ {Xtj = Ytj }
⎝j=1 ⎠
⎛n ⎞
= P ⋂ {Ytj ∈ Aj } .
⎝j=1 ⎠
∎∎
Problem 3.6. Solution: indistinguishable Ô⇒ modification:
P(Xt = Yt ∀t ⩾ 0) = 1 Ô⇒ ∀t ⩾ 0 ∶ P(Xt = Yt ) = 1.
modification Ô⇒ equivalent: see the previous Problem 3.5
Now assume that I is countable or t ↦ Xt , t ↦ Yt are (left- or right-)continuous.
modification Ô⇒ indistinquishable: By assumption, P(Xt ≠ Yt ) = 0 for any t ∈ I. Let

D ⊂ I be any countable dense subset. Then
⎛ ⎞
P ⋃ {Xq ≠ Yq } ⩽ ∑ P(Xq ≠ Yq ) = 0
⎝q∈D ⎠ q∈D
which means that P(Xq = Yq ∀q ∈ D) = 1. If I is countable, we are done. In the other case
we have, by the density of D,
P(Xt = Yt ∀t ∈ I) = P (lim Xq = lim Yq ∀t ∈ I) ⩾ P (Xq = Yq ∀q ∈ D) = 1.

D∋q D∋q
equivalent Ô⇒
/ modification: To see this let (Bt )t⩾0 and (Wt )t⩾0 be two independent
one-dimensional Brownian motions defined on the same probability space. Clearly,
these processes have the same finite-dimensional distributions, i. e. they are equivalent.
On the other hand, for any t > 0
∞ ∞
P(Bt = Wt ) = ∫ P(Bt = y) P(Wt ∈ dy) = ∫ 0 P(Wt ∈ dy) = 0.
−∞ −∞
∎∎
Problem 3.7. Solution: We use the characterization from Lemma 2.8. Its proof shows that
we can derive (2.15)
⎡ ⎤
⎢ ⎛ n ⎞⎥⎥ ⎛ 1 n ⎞
⎢
E ⎢exp i ∑ ⟨ξj , Xqj − Xqj−1 ⟩ + i⟨ξ0 , Xq0 ⟩ ⎥ = exp − ∑ ∣ξj ∣2 (qj − qj−1 )
⎢ ⎝ j=1 ⎠⎥ ⎝ 2 j=1 ⎠
⎣ ⎦
31
on the basis of (B0)–(B3) for (Bq )q∈Q∩[0,∞) and q0 = 0, q1 . . . , qn ∈ Q ∩ [0, ∞).

Now set t0 = q0 = 0 and pick t1 , . . . , tn ∈ R and approximate each tj by a rational sequence
(k) (k)
qj , k ⩾ 1. Since (2.15) holds for qj , j = 0, . . . , n and every k ⩾ 0, we can easily perform
the limit k → ∞ on both sides (on the left we use dominated convergence!) since Bt is
continuous.
This proves (2.15) for (Bt )t⩾0 , and since (Bt )t⩾0 has continuous paths, Lemma 2.8 proves
that (Bt )t⩾0 is a BM1 .
∎∎
Problem 3.8. Solution: The joint density of (W (t0 ), W (t), W (t1 )) is
1 1 1 (x1 − x)2 (x − x0 )2 x20

ft0 ,t,t1 (x0 , x, x1 ) = √ exp (− [ + + ])
(2π)3/2 (t1 − t)(t − t0 )t0 2 t1 − t t − t0 t0
while the joint density of (W (t0 ), W (t1 )) is
1 1 1 (x1 − x0 )2 x20
ft0 ,t1 (x0 , x1 ) = √ exp (− [ + ]) .
(2π) (t1 − t0 )t0 2 t1 − t0 t0
The conditional density of W (t) given (W (t0 ), W (t1 )) is
ft∣t0 ,t1 (x ∣ x1 , x2 )
ft0 ,t,t1 (x0 , x, x1 )
=
ft0 ,t1 (x0 , x1 )
(x1 −x)2 (x−x0 )2 x20
1
(2π)3/2
√ 1
exp (− 12 [ t1 −t + + t0 ])
(t1 −t)(t−t0 )t0 t−t0
=
(x1 −x0 )2 x20
1 √ 1
(2π) (t1 −t0 )t0 exp (− 12 [ t1 −t0 + t0 ])
¿
1 ÁÀ (t1 − t0 ) 1 (x1 − x)2 (x − x0 )2 (x1 − x0 )2
=√ Á exp (− [ + − ])
2π (t1 − t)(t − t0 ) 2 t1 − t t − t0 t1 − t0
¿
1 ÁÀ (t1 − t0 ) 1 (t − t0 )(x1 − x)2 + (t1 − t)(x − x0 )2 (x1 − x0 )2
=√ Á exp (− [ − ])
2π (t1 − t)(t − t0 ) 2 (t1 − t)(t − t0 ) t1 − t0
Now consider the argument in the square brackets [⋯] of the exp-function
(t − t0 )(x1 − x)2 + (t1 − t)(x − x0 )2 (x1 − x0 )2

[ − ]
(t1 − t)(t − t0 ) t1 − t0
(t1 − t0 ) t − t0 t1 − t (t1 − t)(t − t0 )
= [ (x1 − x)2 + (x − x0 )2 − (x1 − x0 )2 ]
(t1 − t)(t − t0 ) t1 − t0 t1 − t0 (t1 − t0 )2
(t1 − t0 ) t − t0 t1 − t t − t0 (t1 − t)(t − t0 ) 2
= [( + ) x2 + ( − ) x1
(t1 − t)(t − t0 ) t1 − t0 t1 − t0 t1 − t0 (t1 − t0 )2
t1 − t (t1 − t)(t − t0 ) 2
+( − ) x0
t1 − t0 (t1 − t0 )2
t − t0 t1 − t (t1 − t)(t − t0 )
−2 x1 x − 2 xx0 + 2 x1 x0 ]
t1 − t0 t1 − t0 (t1 − t0 )2
(t1 − t0 ) (t − t0 )2 2 (t1 − t)2 2
= [x2 + x + x
(t1 − t)(t − t0 ) (t1 − t0 )2 1 (t1 − t0 )2 0
32
t − t0 t1 − t (t1 − t)(t − t0 )
−2 x1 x − 2 xx0 + 2 x1 x0 ]
t1 − t0 t1 − t0 (t1 − t0 )2
(t1 − t0 ) t − t0 t1 − t 2
= [x − x1 − x0 ]
(t1 − t)(t − t0 ) t1 − t0 t1 − t0
(t1 − t0 ) t − t0 t1 − t 2
= [x − ( x1 + x0 )] .
(t1 − t)(t − t0 ) t1 − t0 t1 − t0
Set
(t1 − t)(t − t0 ) t − t0 t1 − t
σ2 = and m = x1 + x0
(t1 − t0 ) t1 − t0 t1 − t0
then our calculation shows that
1 (x − m)2
ft∣t0 ,t1 (x ∣ x1 , x2 ) = √ exp ( ).
2π σ 2σ 2
∎∎
33
4 The canonical model
Problem 4.1. Solution: Let F ∶ R → [0, 1] be a distribution function. We begin with a general
lemma: F has a unique generalized monotone increasing right-continuous inverse:
F −1 (u) = G(u) = inf{x ∶ F (x) > u}

(4.1)
[ = sup{x ∶ F (x) ⩽ u}].
We have F (G(u)) = u if F (t) is continuous in t = G(u), otherwise, F (G(u)) ⩾ u.
Indeed: For those t where F is strictly increasing and continuous, there is nothing to show.
Let us look at the two problem cases: F jumps and F is flat.
F (t) G(u)
w+
w
w−
G(u) G(v−) G(v) G(w) t u
Figure 4.1: An illustration of the problem cases
If F (t) jumps, we have G(w) = G(w+ ) = G(w− ) and if F (t) is flat, we take the right
endpoint of the ‘flatness interval’ [G(v−), G(v)] to define G (this leads to right-continuity
of G)
(a) Let (Ω, A, P) = ([0, 1], B[0, 1], du) (du stands for Lebesgue measure) and define X =
G (G = F −1 as before). Then
P({ω ∈ Ω ∶ X(ω) ⩽ x})

= λ({u ∈ [0, 1] ∶ G(u) ⩽ x})
(the discontinuities of F are countable, i. e. a Lebesgue null set)
35
= λ({t ∈ [0, 1] ∶ t ⩽ F (x)})

= λ([0, F (x)]) = F (x).
Measurability is clear because of monotonicity.

(b) Use the product construction and part a). To be precise, we do the construction for
two random variables. Let X ∶ Ω → R and Y ∶ Ω′ → R be two iid copies. We define
on the product space
(Ω × Ω′ , A ⊗ A′ , P × P′ )
the new random variables ξ(ω, ω ′ ) ∶= X(ω) and η(ω, ω ′ ) ∶= Y (ω ′ ). Then we have
• ξ, η live on the same probability space
• ξ ∼ X, η ∼ Y
P × P′ (ξ ∈ A) = P × P′ ({(ω, ω ′ ) ∈ Ω × Ω′ ∶ ξ(ω, ω ′ ) ∈ A})

= P × P′ ({(ω, ω ′ ) ∈ Ω × Ω′ ∶ X(ω) ∈ A})
= P × P′ ({ω ∈ Ω ∶ X(ω) ∈ A} × Ω′ )
= P({ω ∈ Ω ∶ X(ω) ∈ A})
= P(X ∈ A).
and a similar argument works for η.

• ξáη
P × P′ (ξ ∈ A, η ∈ B) = P × P′ ({(ω, ω ′ ) ∈ Ω × Ω′ ∶ ξ(ω, ω ′ ) ∈ A, η(ω, ω ′ ) ∈ B})

= P × P′ ({(ω, ω ′ ) ∈ Ω × Ω′ ∶ X(ω) ∈ A, Y (ω ′ ) ∈ B})
= P × P′ ({ω ∈ Ω ∶ X(ω) ∈ A} × {ω ∈ Ω′ ∶ Y (ω ′ ) ∈ B})
= P({ω ∈ Ω ∶ X(ω) ∈ A}) P′ ({ω ∈ Ω′ ∶ Y (ω ′ ) ∈ B})
= P(X ∈ A) P(Y ∈ B)
= P × P′ (ξ ∈ A) P × P′ (η ∈ B)
The same type of argument works for arbitrary products, since independence is
always defined for any finite-dimensional subfamily. In the infinite case, we have
to invoke the theorem on the existence of infinite product measures (which are
constructed via their finite marginals) and which can be seen as a particular case
of Kolmogorov’s theorem, cf. Theorem 4.8 and Theorem A.2 in the appendix.
(c) The statements are the same if one uses the same construction as above. A difficulty
is to identify a multidimensional distribution function F (x). Roughly speaking, these
are functions of the form
F (x) = P (X ∈ (−∞, x1 ] × ⋯ × (−∞, xn ])
where X = (X1 , . . . , Xn ) and x = (x1 , . . . , xn ), i. e. x is the ‘upper right’ endpoint of

an infinite rectancle. An abstract characterisation is the following
36
• F ∶ Rn → [0, 1]
• xj ↦ F (x) is monotone increasing
• xj ↦ F (x) is right continuous
• F (x) = 0 if at least one entry xj = −∞
• F (x) = 1 if all entries xj = +∞
• ∑(−1)∑k=1 k F (1 a1 + (1 − 1 )b1 , . . . , n an + (1 − n )bn ) ⩾ 0 where −∞ < aj < bj < ∞

n
and where the outer sum runs over all tuples (1 , . . . , n ) ∈ {0, 1}n
The last property is equivalent to

(1) (n) (k)
• ∆h1 ⋯∆hn F (x) ⩾ 0 ∀h1 , . . . , hn ⩾ 0 where ∆h F (x) = F (x + hek ) − F (x) and
ek is the k th standard unit vector of Rn .
In principle we can construct such a multidimensional F from its marginals using

the theory of copulas, in particular, Sklar’s theorem etc. etc. etc.
Another way would be to take (Ω, A, P) = (Rn , B(Rn ), µ) where µ is the probability
measure induced by F (x). Then the random variables Xn are just the identity maps!
The independent copies are then obtained by the usual product construction.
∎∎
Problem 4.2. Solution: Step 1: Let us first show that P(lims→t Xs exists) < 1.
Since Xr á Xs and Xs ∼ −Xs we get

√
Xr − Xs ∼ Xr + Xs ∼ N(0, s + r) ∼ s + r N(0, 1).
Thus,

P(∣Xr − Xs ∣ > ) = P (∣X1 ∣ > √ ) ÐÐÐ→ P (∣X1 ∣ > √ ) ≠ 0.
s+r r,s→t 2t
This proves that Xs is not a Cauchy sequence in probability, i. e. it does not even converge
in probability towards a limit, so a.e. convergence is impossible.
In fact we have
∞
{ω ∶ lim Xs (ω) does not exist} ⊃ ⋂ { sup ∣Xs − Xr ∣ > 0}
s→t k=1 s,r∈[t−1/k,t+1/k]
and so we find with the above calculation

P ( lim Xs does not exist) ⩾ lim P ( sup ∣Xs − Xr ∣ > 0) ⩾ P (∣X1 ∣ > √ )
s→t k s,r∈[t−1/k,t+1/k] 2t
This shows, in particular that for any sequence tn → t we have
P ( lim Xtn exists) < q < 1.

n→∞
where q = q(t) (but independent of the sequence).
37
Step 2: Fix t > 0, fix a sequence (tn )n with tn → t, and set
A = {ω ∈ Ω ∶ lim Xs (ω) exists} and A(tn ) = {ω ∈ Ω ∶ lim Xtn (ω) exists}.

s→t n→∞
Clearly, A ⊂ A(tn ) for any such sequence. Moreover, take two sequences (sn )n , (tn )n such
that sn → t and tn → t and which have no points in common; then we get by independence
and step 1
(Xs1 , Xs2 , Xs3 . . .) á (Xt1 , Xt2 , Xt3 . . .) Ô⇒ A(tn ) á A(sn )
and so, P(A) ⩽ P(A(sn ) ∩ A(tn )) = P(A(sn )) P(A(tn )) = q 2 .

By Step 1, q < 1. Since there are infinitely many sequences having all no points in common,
we get 0 ⩽ P(A) ⩽ limk→∞ q k = 0.
∎∎
Problem 4.3. Solution: Write Σ ∶= ⋃{σ(C) ∶ C ⊂ E, C is countable}.

If C ⊂ E we get σ(C) ⊂ σ(E), and so Σ ⊂ σ(E).
Conversely, it is clear that E ⊂ Σ, just take C ∶= CE ∶= {E} for each E ∈ E. If we can show
that Σ is a σ-algebra we get σ(E) ⊂ σ(Σ) = Σ and equality follows.
• Clearly, ∅ ∈ Σ.
• If S ∈ Σ, then S ∈ σ(CS ) for some countable CS ⊂ E. Moreover, S c ∈ σ(CS ), i. e. S c ∈
Σ.
• If (Sn )n⩾0 ⊂ Σ are countably many sets, then Sn ∈ σ(Cn ) for some countable Cn ⊂ E
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.
∎∎

−1
(a) Following the hint, we use E = ⋃{πK (A) ∶ A ∈ BK (E), K ⊂ I, #K < ∞}, i. e. the
cylinder sets. By definition, σ(E) = BI (E). If C ⊂ E is a countable set, then there is
a countable index set J ⊂ I such that C ⊂ πJ−1 (BJ (E)), and since the right-hand side
is a σ-algebra, σ(C) ⊂ πJ−1 (BJ (E)). This proves the claim.
(b) The previous part shows that every B ∈ BI (E) is of the form B = πJ−1 (A) with
A ∈ BJ (E), i. e. it is a cylinder with a countable base. All other indices (and since I
is not countable, there are uncountably many left) are E. Since E contains at least
two points, we see that B ≠ ∅ is uncountable.
−1
(c) Use the fact that the open cylinders πK (U ), where U ⊂ E K is open and #K < ∞,
are open sets in E I and (a).
(d) Let e ∈ E; then F ∶= {e} is a non-void compact set. By Tychonov’s theorem F I ⊂ E I
is compact and as such it is contained in the Borel σ-algebra B(E I ). But F I contains
exactly one point, i. e. it is countable. By part (b) it cannot be in BI (E).
38
(e) Assume that I is countable. Then B(E I ) is generated by countably many sets,
−1
namely πK (B(q, r)∩E) where K ⊂ I is finite and B(q, r) are open balls with rational
radii r > 0 and rational centres q ∈ E. (Note: There are only countably many
finite sets contained in a countable set I! Here the argument would break down
for uncountable index sets.) These are but the cylinder sets, i. e. they also generate
BI (E), and this proves B(E I ) ⊂ BI (E).
∎∎
Problem 4.5. Solution: Xt (ω) is a ‘random’ path starting at the randomly chosen point ω
and moving uniformly with constant speed ∣v∣ in the direction v/∣v∣. Note that only the
starting point is random and it is ‘drawn’ using the law µ or δx , i. e. in the latter case we
start a.s. at x.
The finite-dimensional distributions are for 0 ⩽ t1 < t2 < . . . < tn given by
Pµ (Xt1 ∈ dy1 , Xt2 ∈ dy2 , . . . , Xtn ∈ dyn )
=∫ µ(dx) δx+t1 v (dy1 ) ⊗ δx+t2 v (dy2 ) ⊗ ⋯ ⊗ δx+tn v (dyn ).

Rd
∎∎
39
5 Brownian motion as a martingale
(a) We have
FtB ⊂ σ(σ(X), FtB ) = σ(X, Bs ∶ s ⩽ t) = F̃t .
Let s ⩽ t. Then σ(Bt − Bs ), FsB and σ(X) are independent, thus σ(Bt − Bs ) is
independent of σ(σ(X), FB ) = F̃s . This shows that (F̃t ) is an admissible filtration s t⩾0
for (Bt )t⩾0 .
(b) Set N ∶= {N ∶ ∃M ∈ A such that N ⊂ M, P(M ) = 0}. Then we have

B
FtB ⊂ σ(FtB , N) = Ft .
From measure theory we know that (Ω, A, P) can be completed to (Ω, A∗ , P∗ ) where
A∗ ∶= {A ∪ N ∶ A ∈ A, N ∈ N},
P∗ (A∗ ) ∶= P(A) for A∗ = A ∪ N ∈ A∗ .
We find for all A ∈ B(Rd ), F ∈ Fs , N ∈ N
P∗ ({Bt − Bs ∈ A} ∩ (F ∪ N )) = P∗ (({Bt − Bs ∈ A} ∩ F ) ∪ ({Bt − Bs ∈ A} ∩ N ))

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
∈A ∈N
= P({Bt − Bs ∈ A} ∩ F )
= P(Bt − Bs ∈ A) P(F )
= P∗ (Bt − Bs ∈ A) P∗ (F ∪ N ).
B
Therefore Ft is admissible.
∎∎
Problem 5.2. Solution: Let t = t0 < . . . < tn , and consider the random variables
B(t1 ) − B(t0 ), . . . , B(tn ) − B(tn−1 ).
Using the argument of Problem 18 we see for any F ∈ Ft
E (ei ∑k=1 ⟨ξk , B(tk )−B(tk−1 )⟩ 1F ) = E (ei⟨ξn , B(tn )−B(tn−1 )⟩ ⋅ ei ∑k=1 ⟨ξk , B(tk )−B(tk−1 )⟩ 1F )
n n−1
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
Ftn−1 mble., hence áB(tn )−B(tn−1 )
= E (ei⟨ξn , B(tn )−B(tn−1 )⟩ ) ⋅ E (ei ∑k=1 ⟨ξk , B(tk )−B(tk−1 )⟩ 1F )

n−1
41
⋮
n
= ∏ E (ei⟨ξk , B(tk )−B(tk−1 )⟩ ) E 1F .
k=1
This shows that the increments are independent among themselves (use F = Ω) and
that they are all together independent of Ft (use the above calculation and the fact that
the increments are among themselves independent to combine again the ∏n1 under the
expected value)
Thus,
Ft á σ(B(tk ) − B(tk−1 ) ∶ k = 1, . . . , n)
Therefore the statement is implied by
Ft á ⋃ σ(B(tk ) − B(t) ∶ k = 1, . . . , n).

t<t1 <...<tn
n⩾1
∎∎
(a) i) E ∣Xt ∣ < ∞, since the expectation does not depend on the filtration.
ii) Xt is Ft measurable and Ft ⊂ Ft∗ . Thus Xt is Ft∗ measurable.
iii) Let N denote the set of all sets which are subsets of P-null sets. Denote by
P∗ the measure of the completion of (Ω, A, P) (compare with the solution to
Exercise 1.b)).
Let t ⩾ s. For all F ∗ ∈ Fs∗ there exist F ∈ Fs , N ∈ N such that F ∗ = F ∪ N and
∫ Xs d P∗ = ∫ Xs d P = ∫ Xt d P = ∫ Xt d P∗ .
F∗ F F F∗
Since F ∗ is arbitrary this implies that E(Xt ∣ Fs∗ ) = Xs .
(b) i) E ∣Yt ∣ = E ∣Xt ∣ < ∞.
ii) Note that {Xt ≠ Yt }, its complement and any of its subsets is in Ft∗ . Let B ∈
B(Rd ). Then we get
{Yt ∈ B} = ({Xt ∈ B} ∩ {Xt ≠ Yt }c ) ∪ {Yt ∈ B, Xt ≠ Yt } .

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
∈Ft ∈Ft∗ ∈Ft∗
iii) Similar to part a-iii). For each F ∗ ∈ Fs∗ we get
Ys d P∗ = ∫ Xs d P∗ = ∫ Xt d P∗ = ∫ Yt d P∗ ,
a)
∫
F∗ F∗ F∗ F∗
i. e. E(Yt ∣ Fs∗ ) = Ys .
∎∎
42
Problem 5.4. Solution: Let s < t and pick sn ↓ s such that s < sn < t. Then
sub-MG a.e.
E(Xt ∣ Fs+ ) ←ÐÐÐÐ E(X(t) ∣ Fsn ) ⩾ X(sn ) ÐÐÐ→ X(s+) =
continuous
X(s).
sn ↓s n→∞ paths
The convergence on the left side follows from the (sub-)martingale convergence theorem
(Lévy’s downward theorem).
∎∎
Problem 5.5. Solution: Here is a direct proof without using the hint.
We start with calculating the conditional expectations
E(Bt4 ∣ Fs )
= E ((Bt − Bs + Bs )4 ∣ Fs )
= Bs4 + 4Bs3 E(Bt − Bs ) + 6Bs2 E((Bt − Bs )2 ) + 4Bs E((Bt − Bs )3 ) + E((Bt − Bs )4 )
= Bs4 + 6Bs2 (t − s) + 3(t − s)2
= Bs4 − 6Bs2 s + 6Bs2 t + 3(t − s)2 ,
and
E(Bt2 ∣ Fs ) = E ((Bt − Bs + Bs )2 ∣ Fs )
= t − s + 2Bs E(Bt − Bs ) + Bs2
= Bs2 + t − s.
Combining these calculations, such that the term 6Bs2 t vanishes from the first formula,
we get
E (Bt4 − 6tBt2 ∣ Fs ) = Bs4 − 6sBs2 − 6t2 + 6st + 3t2 − 6st + 3s2

= Bs4 − 6sBs + 3s2 − 3t2 .
Therefore π(t, Bt ) ∶= Bt4 − 6tBt2 + 3t2 is a martingale.
∎∎
(a) Since Brownian motion has exponential moments of any order, we can use the dif-
ferentiation lemma for parameter-dependent integrals. Following the instructions we
get
d ξBt − t ξ2
e 2 = (Bt − tξ)Mtξ
dξ
d2 ξBt − t ξ2
e 2 = ((Bt − tξ)2 − t) Mtξ
dξ 2
d3 ξBt − t ξ2
e 2 = ((Bt − tξ)2 − 3t) (Bt − tξ)Mtξ
dξ 3
43
d4 ξBt − t ξ2
4
e 2 = {((Bt − tξ)2 − 3t) ((Bt − tξ)2 − t) − 2t(Bt − tξ)2 } Mtξ
dξ
and so on. The recursion n → n + 1 is pretty obvious
d d
Pn (B, ξ)Mtξ = [ Pn (B, ξ) + Pn (B, ξ)(B − tξ)] Mtξ .
dξ dξ
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
Pn+1 (B,ξ)
If we set ξ = 0 we find that Pn (b, 0)∣b=Bt is a martingale. In particular,
Bt
Bt2 − t
Bt3 − 3tBt
Bt4 − 6tBt2 + 3t2
are martingales.
(b) Part (a) shows the general recursion scheme
d
P1 (b, ξ) = b − tξ, Pn+1 (b, ξ) = Pn (b, ξ) + (b − tξ)Pn (b, ξ).
dξ
(c) Using the fact that Mt ∶= Bt4 − 6tBt2 + 3t2 is a martingale with M0 = 0 we get for the
bounded stopping times τ ∧ n by optional stopping
0 = E [Mτ ∧n ] = E [Bτ4∧n ] − 6 E [(τ ∧ n)Bτ2∧n ] + 3 E [(τ ∧ n)2 ]
and, by rearranging this equality, and with the Cauchy-Schwarz inequality
3 E [(τ ∧ n)2 ] = 6 E [(τ ∧ n)Bτ2∧n ] − E [Bτ4∧n ]

⩽ 6 E [(τ ∧ n)Bτ2∧n ]
√ √
⩽ 6 E [(τ ∧ n)2 ] E [Bτ4∧n ].
Thus, √
√
E [(τ ∧ n)2 ] ⩽ 2 E [Bτ4∧n ].
Since ∣Bτ ∧n ∣ ⩽ max{a, b}, we can use monotone convergence (on the left side) and
dominated convergence (on the right), and the first inequality follows.
The second inequality follows in a similar way: By optional stopping we get
E [Bτ4∧n ] = 6 E [(τ ∧ n)Bτ2∧n ] − 3 E [(τ ∧ n)2 ]

⩽ 6 E [(τ ∧ n)Bτ2∧n ]
√ √
⩽ 6 E [(τ ∧ n) ] E [Bτ4∧n ].
2
Thus, √ √ √
E [Bτ4∧n ] ⩽ 6 E [(τ ∧ n)2 ] ⩽ 6 E [τ 2 ]
and the estimate follows from dominated convergence (on the left).
44
∎∎
Problem 5.7. Solution: For t = 0 and all c we have
E ec∣B0 ∣ = E ec∣B0 ∣ = 1.
2
and for c ⩽ 0
E ec∣B0 ∣ ⩽ 1 E ec∣B0 ∣ ⩽ 1.
2
and
Now let t > 0 and c > 0. There exists some R > 0 such that c∣x∣ < 1
4t ∣x∣2 for all ∣x∣ > R.
Thus
E ec∣Bt ∣ = c̃ ∫ ec∣x∣ e− 2t ∣x∣ dx

1 2
ec∣x∣ e− 2t ∣x∣ dx + c̃ ∫ e 4t ∣x∣ e− 2t ∣x∣ dx

1 2 1 2 1 2
⩽ c̃ ∫
∣x∣⩽R ∣x∣>R
− 4t
1
∣x∣2
⩽ ecR + c̃ ∫ e dx < ∞,
∣x∣>R
i.e., E ec∣Bt ∣ < ∞ for all c, t. Furthermore

2 (c− 1 )
E ec∣Bt ∣ = c̃ ∫ ec∣x∣
2− 1 ∣x∣2
dx = c̃ ∫ e∣x∣
2
2t 2t dx
and this integral is finite if, and only if, c − 1

2t < 0 or equivalently c < 1
2t .
∎∎

∣x∣2
(a) We have p(t, x) = (2πt)− 2 e−
d
2t . By the chain rule we get
∂ d d ∣x∣2 ∣x∣2 − ∣x∣2

p(t, x) = − t− 2 −1 (2π)− 2 e− 2t + (2πt)− 2 (−1)t−2 (−1)
d d
e 2t
∂t 2 2
and for all j = 1, . . . , d
∂ 2xj − ∣x∣2
p(t, x) = (2πt)− 2 ( −
d
) e 2t ,
∂xj 2t
∂2 − d2 1 − ∣x∣2 x2 ∣x∣2
− d j − 2t
p(t, x) = (2πt) ( − ) e 2t + (2πt) 2 e .
∂x2j t t2
Adding these terms and noting that ∣x∣2 = ∑dj=1 x2j we get
(2πt)− 2 ∣x∣2 − ∣x∣2

d
1 d ∂2 d ∣x∣2 ∂
− d −1 −
∑ 2 p(t, x) = − (2πt) 2 t e 2t + 2
e 2t = p(t, x).
2 j=1 ∂xj 2 2 t ∂t
(b) A formal calculation yields
1 ∂2
∫ p(t, x) f (t, x) dx
2 ∂x2j
∞
1 ∂ ∂ 1 ∂
= p(t, x) f (t, x)∣ − ∫ p(t, x) ⋅ f (t, x) dx
2 ∂xj −∞ ∂x j 2 ∂xj
45
∞
∂ 1 ∂2 1
=0− p(t, x) ⋅ f (t, x)∣ + ∫ 2
p(t, x) ⋅ f (t, x) dx
∂xj 2 −∞ ∂xj 2
∂2 1
=∫ 2
p(t, x) ⋅ f (t, x) dx.
∂xj 2
By the same arguments as in Exercise 7 we find that all terms are integrable and
vanish as ∣x∣ → ∞. This justifies the above calculation. Furthermore summing over
j = 1, . . . d we obtain the statement.
∎∎
Problem 5.9. Solution: Note that E ∣Xt ∣ < ∞ for all a, b, cf. Problem 5.7. We have
E (eaBt +bt ∣ Fs ) = E (ea(Bt −Bs ) eaBs +bt ∣ Fs )

= eaBs +bt E eaBt−s
2 /2
= eaBs+bt+(t−s)a .
2
Thus, Xt is a martingale if, and only if, bs = bt + (t − s) a2 , i.e., b = − 21 a2 .
∎∎
Problem 5.10. Solution: Measurability (i. e. adaptedness to the Filtration Ft ) and integra-
bility is no issue, see also Problem 5.7.
(a) Ut is only a martingale for c = 0.
Solution 1: see Exercise 9.
Solution 2: if c ≠ 0, E Ut is not constant, i. e. cannot be a martingale. If c = 0, Ut is

trivially a martingale.
(b) Vt is a martingale since

s t
E (Vt ∣ Fs ) = t E(Bt − Bs ) + tBs − E (∫ Br dr ∣ Fs ) − E (∫ Br dr ∣ Fs )
0 s
s t
= tBs − ∫ Br dr − E (∫ (Br − Bs ) + Bs dr ∣ Fs )
0 s
s
= tBs − ∫ Br dr − (t − s)Bs
0
= Vs .
(c) and (e) Let a ∈ R. Then we get
E (aBt3 − tBt ∣ Fs ) = E (a(Bt − Bs + Bs )3 − t(Bt − Bs ) − tBs ∣ Fs )

= aBs3 + 3aBs2 E Bt−s + 3aBs E Bt−s
2
+ a E Bt−s
3
− 0 − tBs
= aBs3 + (3a(t − s) − t)Bs .
This is a martingale if, and only if, −s = 3a(t − s) − t, i.e., a = 1

3. Thus Yt is a
martingale and Wt is not a martingale.
46
d) We have seen in part c) and b) that
E (Bt3 ∣ Fs ) = Bs3 + 3(t − s)Bs
and
t s
3 E (∫ Br dr ∣ Fs ) = 3 ∫ Br dr + 3(t − s)Bs .
0 0
Thus, Xt is a martingale.
(f) Zt is only a martingale for c = 12 , see Exercise 9.
∎∎

d d
(j) 2 Pr. 5 (j) 2
E ( d1 ∣Bt ∣2 − t ∣ Fs ) = −t + 1
d ∑ E ((Bt ) ∣ Fs ) = −t + 1
d ∑ ((Bs ) + t − s) = 1
d ∣Bs ∣2 − s.
j=1 j=1
∎∎
Problem 5.12. Solution: For a)–c) we prove only the statements for τ ○ , the statements for τ
are proved analogously.
(a) The following implications hold:
A ⊂ C Ô⇒ {t ⩾ 0 ∶ Xt ∈ A} ⊂ {t ⩾ 0 ∶ Xt ∈ C} Ô⇒ τA○ ⩾ τC○ .
○
(b) By part a) we have τA∪C ○
⩽ τA○ and τA∪C ⩽ τC○ . Thus,
a)
○
τA∪C ⩽ min{τA○ , τC○ }.
○
To see the converse, min{τA○ , τC○ } ⩽ τA∪C , it is enough to show that
Xt (ω) ∈ A ∪ C Ô⇒ t ⩾ min{τA○ (ω), τC○ (ω)}
○
since this implication shows that τA∪C (ω) ⩾ min{τA○ (ω), τC○ (ω)} holds.
Now observe that
Xt (ω) ∈ A ∪ C Ô⇒ Xt (ω) ∈ A or Xt (ω) ∈ C

Ô⇒ t ⩾ τA○ (ω) or t ⩾ τC○ (ω)
Ô⇒ t ⩾ min{τA○ (ω), τC○ (ω)}.
(c) Part a) implies max{τA○ , τC○ } ⩽ τA∩C

○
.
Remark: we cannot expect “=”. To see this consider a BM1 staring at B0 = 0 and
the set
A = [4, 6] and C = [1, 2] ∪ [5, 7].
Then Bt has to reach first C and A before it hits A ∩ C.
47
(d) as in b) it is clear that τA○ ⩽ τA○ n for all n ⩾ 1, hence
τA○ ⩽ inf τA○ n .

n⩾1
In order to show the converse, τA○ ⩾ inf n⩾1 τA○ n , it is enough to check that
Xt (ω) ∈ A Ô⇒ t ⩾ inf τA○ n (ω)

n⩾1
since, if this is true, this implies that τA○ (ω) ⩾ inf n⩾0 τA○ n (ω).
Now observe that
Xt (ω) ∈ A = ∪n An Ô⇒ Xt (ω) ∈ An for some n ∈ N

Ô⇒ t ⩾ τA○ n (ω) for some n ∈ N
Ô⇒ t ⩾ inf τA○ n (ω).
n⩾0
(e) Note that inf {s ⩾ 0 ∶ Xs+ 1 ∈ A} = inf {s ⩾ 1

n ∶ Xs ∈ A} is monotone decreasing as
n
n → ∞. Thus we get
inf ( n1 + inf{s ⩾ 1
n ∶ Xs ∈ A}) = 0 + inf inf{s ⩾ 1
n ∶ Xs ∈ A}
n n
= inf{s > 0 ∶ Xs ∈ A}
= τA .
○
(f) Let Xt = x0 + t. Then τ{x 0}
= 0 and τ{x0 } = ∞.
More generally, a similar situation may happen if we consider a process with con-
tinuous paths, a closed set F , and if we let the process start on the boundary ∂F .
Then τF○ = 0 a.s. (since the process is in the set) while τF > 0 is possible with positive
probability.
∎∎
Problem 5.13. Solution: We have τU○ ⩽ τU .
Let x0 ∈ U . Then τU○ = 0 and, since U is open and Xt is continuous, there exists an N > 0
such that
X 1 ∈ U for all n ⩾ N.
n
Thus τU = 0.
If x0 ∉ U , then Xt (ω) ∈ U can only happen if t > 0. Thus, τU○ = τU .
∎∎
Problem 5.14. Solution: Suppose d(x, A) ⩾ d(z, A). Then
d(x, A) − d(z, A) = inf ∣x − y∣ − inf ∣z − y∣

y∈A y∈A
⩽ inf (∣x − z∣ + ∣z − y∣) − inf ∣z − y∣

y∈A y∈A
48
= ∣x − z∣
and, with an analogous argument for d(x, A) ⩽ d(z, A), we conclude
∣d(x, A) − d(z, A)∣ ⩽ ∣x − z∣.
Thus x ↦ d(x, A) is globally Lipschitz continuous, hence uniformly continuous.
∎∎
Problem 5.15. Solution: We treat the two cases simultaneously and check the three properties
of a sigma algebra:
i) We have Ω ∈ F∞ and
Ω ∩ {τ ⩽ t} = {τ ⩽ t} ∈ Ft ⊂ Ft+ .
ii) Let A ∈ Fτ (+) . Thus A ∈ F∞ , Ac ∈ F∞ and
Ac ∩ {τ ⩽ t} = Ω ∖ A ∩ {τ ⩽ t} = (Ω ∩ {τ ⩽ t}) ∖ (A ∩ {τ ⩽ t}) ∈ Ft(+) .

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
∈Ft ⊂Ft+ ∈Ft(+) since A∈Fτ (+)
iii) Let An ∈ Fτ (+) . Then An , ⋃n An ∈ F∞ and
⋃ An ∩ {τ ⩽ t} = ⋃(An ∩ {τ ⩽ t}) ∈ Ft(+) .

n n ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
∈Ft(+)
Therefore Fτ and Fτ + are σ-algebras.
∎∎
(a) Let F ∈ Fτ + , i.e., F ∈ F∞ and for all s we have F ∩ {τ ⩽ s} ∈ Fs+ .
Let t > 0. Then

F ∩ {τ < t} = ⋃ (F ∩ {τ ⩽ s}) ∈ ⋃ Fs+ ⊂ Ft .
s<t s<t
For the converse: If τ < ∞ a.s. then F = ⋃t>0 (F ∩ {τ ⩽ t}) ∈ F∞ and
F ∩ {τ ⩽ s} = ⋂ (F ∩ {τ < t}) ∈ ⋂ Ft = Fs+ .

t>s t>s
If τ = ∞ occurs with strictly positive probability, then we have to assume that F ∈ F∞ .
(b) We have {τ ⩽ t} ∈ Ft ⊂ F∞ and

⎧
⎪
⎪
⎪{τ ⩽ t} ∈ Ft if r ⩾ t;
{τ ⩽ t} ∩ {τ ∧ t ⩽ r} = ⎨
⎪
⎪
⎩{τ ⩽ r} ∈ Fr ⊂ Ft
⎪ if r < t.
∎∎
49
(a) eiξBt + 2 t∣ξ∣ is a martingale for all ξ ∈ R by Example 5.2 d). By optional stopping
1 2
τ ∧t
1 = E e 2 (τ ∧t)c
1 2 +icB
.
Since the left-hand side is real, we get
1 = E (e 2 (τ ∧t)c cos(cBτ ∧t )).

1 2
Set m ∶= a ∨ b. Since ∣Bτ ∧t ∣ ⩽ m, we see that for mc < 1

2 π the cosine is positive. By
Fatou’s lemma we get for all mc < 1
2 π
1 = lim E (e 2 (τ ∧t)c cos(cBτ ∧t ))

1 2
t→∞
⩾ E ( lim e 2 (τ ∧t)c cos(cBτ ∧t ))

1 2
t→∞
1 2
⩾ E (e 2 τ c cos(cBτ ))
1 2
⩾ cos(mc) E e 2 τ c .
Thus, E eγτ < ∞ for any γ < 1 2

2c and all c < π/(2m). Since
∞
tj tj
et = ∑ Ô⇒ ∀t ⩾ 0, j ⩾ 0 ∶ et ⩾
j=0 j! j!
we see that E τ j ⩽ j! γ −j E eγτ < ∞ for any j ⩾ 0.

t
(b) By Exercise 10 d) the process Bt3 − 3 ∫0 Bs ds is a martingale. By optional stopping
we get
τ ∧t
E (Bτ3∧t − 3 ∫ Bs ds) = 0 for all t ⩾ 0. (*)
0
Set m = max{a, b}. By the definition of τ we see that ∣Bτ ∧t ∣ ⩽ m; since τ is integrable
we get
τ ∧t
∣Bτ3∧t ∣ ⩽ m3 and ∣∫ Bs ds∣ ⩽ τ ⋅ m.
0
Therefore, we can use in (*) the dominated convergence theorem and let t → ∞:
τ 1
E (∫ Bs ds) = E(Bτ3 )
0 3
1 1
= (−a)3 P(Bτ = −a) + b3 P(Bτ = b)
3 3
(5.12) 1 −a b + b a
3 3
=
3 a+b
1
= ab(b − a).
3
∎∎
Problem 5.18. Solution: By Example 5.2 c) ∣Bt ∣2 − d ⋅ t is a martingale. Thus we get by

optional stopping
1
E(t ∧ τR ) = E ∣Bt∧τR ∣2 for all t ⩾ 0.
d
50
Since ∣Bt∧τR ∣ ⩽ R, we can use monotone convergence on the left and dominated convergence
on the right-hand side to get
1 1 1
E τR = sup E(t ∧ τR ) = lim E ∣Bt∧τR ∣2 = E ∣BτR ∣2 = R2 .
t⩾0 t→∞ d d d
∎∎
(a) For all t we have

{σ ∧ τ ⩽ t} = {σ ⩽ t} ∪ {τ ⩽ t} ∈ Ft .
´¹¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¶
∈Ft ∈Ft
(b) For all t we have
{σ < τ } ∩ {σ ∧ τ ⩽ t} = ⋃ ({σ ⩽ r < τ } ∩ {σ ∧ τ ⩽ t})

0⩽r∈Q
= ⋃ (({σ ⩽ r} ∩ {τ ⩽ r}c ) ∩ {σ ∧ τ ⩽ t}) ∈ Ft .

r∈Q∩[0,t]
This shows that {σ < τ }, {σ ⩾ τ } = {σ < τ }c ∈ Fσ∧τ . Since σ and τ play symmetric
roles, we get with a similar argument that {σ > τ }, {σ ⩽ τ } = {σ > τ }c ∈ Fσ∧τ , and
the claim follows.
(c) Since τ ∧σ is an integrable stopping time, we get from Wald’s identities, Theorem 5.10,
that
E Bτ2∧σ = E(τ ∧ σ) < ∞.
Following the hint we get
E(Bσ Bτ 1{σ⩽τ } ) = E(Bσ∧τ Bτ 1{σ⩽τ } )

= E ( E(Bσ∧τ Bτ 1{σ⩽τ } ∣ Fτ ∧σ ))
b)
= E (Bσ∧τ 1{σ⩽τ } E (Bτ ∣ Fτ ∧σ ))
(*)
= E(Bσ∧τ
2
1{σ⩽τ } ).
(We will discuss the step marked by (*) below.)
With an analogous calculation for τ ⩽ σ we conclude
E(Bσ Bτ ) = E(Bσ∧τ Bτ 1{σ<τ } ) + E(Bσ∧τ Bτ 1{τ ⩽σ} ) = E(Bσ∧τ

2
) = E σ ∧ τ.
In the step marked with (*) we used that for integrable stopping times σ, τ we have
E(Bτ ∣ Fσ∧τ ) = Bσ∧τ .
To see this we use optional stopping which gives
E(Bτ ∧k ∣ Fσ∧τ ∧k ) = Bσ∧τ ∧k for all k ⩾ 1.
51
This is the same as to say that
∫ Bτ ∧k d P = ∫ Bσ∧τ ∧k d P for all k ⩾ 1, F ∈ Fσ∧τ ∧k .

F F
Since Bτ ∧k ÐÐÐ→ Bτ in L2 (P), see the proof of Theorem 5.10, we get for some fixed
k→∞
i < k because of Fσ∧τ ∧i ⊂ Fσ∧τ ∧k that
∫ Bτ d P = k→∞
lim ∫ Bτ ∧k d P = lim ∫ Bσ∧τ ∧k d P = ∫ Bσ∧τ d P
k→∞
for all F ∈ Fσ∧τ ∧i .
F F F F
Let ρ = σ ∧ τ (or any other stopping time). Since Fρ∧k = Fρ ∩ Fk we see that Fρ is
generated by the ∩-stable generator ⋃i Fρ∧i , and (*) follows.
(d) From the above and Wald’s identity we get
E(∣Bτ − Bσ ∣2 ) = E(Bτ2 − 2Bτ Bσ + Bσ2 )

= Eτ − 2Eτ ∧ σ + Eσ
= E(τ − 2(τ ∧ σ) + σ)
= E ∣τ − σ∣.
In the last step we used the elementary relation
(a + b) − 2(a ∧ b) = a ∧ b + a ∨ b − 2(a ∧ b) = a ∨ b − a ∧ b = ∣a − b∣.
∎∎
52
6 Brownian motion as a Markov process
We write gt (x) = (2πt)−1/2 e−x

2 /(2t)
Problem 6.1. Solution: for the one-dimensional normal
density.
(a) This follows immediately from our proof of b).
(b) Let u ∈ Bb (R) and s, t ⩾ 0. Then, by the independent and stationary increments
property of a Brownian motion
E u(∣Bt+s ∣ ∣ Fs ) = E u(∣(Bt+s − Bs ) + Bs ∣ ∣ Fs )
= E u(∣(Bt+s − Bs ) + y∣)∣
y=Bs
= E u(∣Bt + y∣)∣ .
y=Bs
Since B ∼ −B we also get
E u(∣Bt+s ∣ ∣ Fs ) = E u(∣Bt + y∣)∣ = E u(∣Bt − y∣)∣

y=−Bs y=Bs
and, therefore,
1
E u(∣Bt+s ∣ ∣ Fs ) = [E u(∣Bt + y∣) + E u(∣Bt − y∣)]y=B
2 s
1 ∞
= [∫ (u(∣z + y∣) + u(∣z − y∣)) gt (z) dz]
2 −∞ y=Bs
1 ∞
= [∫ u(∣z∣) (gt (z + y) + gt (z − y)) dz]
2 −∞ y=Bs
∞
=∫ u(∣z∣) (gt (z + y) + gt (z − y)) dz∣
0 y=Bs
here we use that the integrand is even in z
∞
= ∫ u(∣z∣) (gt (z + ∣y∣) + gt (z − ∣y∣)) dz ∣
0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ y=Bs
=∶gu,s,t+s (y)—it is independent of s!
since the integrand is also even in y! This shows that
• E u(∣Bt+s ∣ ∣ Fs ) is a function of ∣Bs ∣, i. e. Markovianity.
• Py (∣Bt ∣ ∈ dz) = gt (z − y) + gt (z + y) for z, y ⩾ 0, i. e. the form of the transition

function.
Remark: ∣Bt ∣ is called reflecting (also: reflected ) Brownian motion.
53
(c) Set Mt ∶= sups⩽t Bs for the running maximum, i. e. Yt = Mt − Bt . From the reflection
principle, Theorem 6.9 we know that Yt ∼ ∣Bt ∣. So the guess is that Y and ∣B∣ are
two Markov processes with the same transition function!
Let s, t ⩾ 0 and u ∈ Bb (R). We have by the independent and stationary increments

property of Brownian motion
E (u(Yt+s ) ∣ Fs ) = E (u(Mt+s − Bt+s ) ∣ Fs )
= E (u( max { sup Br , sup Bs+u } − Bt+s ) ∣ Fs )

u⩽s 0⩽u⩽t
= E (u( max { sup(Br − Bs ) + (Bs − Bt+s ), sup (Bs+u − Bs+t )}) ∣ Fs )

u⩽s 0⩽u⩽t
and, as supu⩽s (Br − Bs ) is Fs measurable and (Bs − Bt+s ), sup0⩽u⩽t (Bs+u − Bs+t ) á Fs ,
we get
= E (u( max {y + (Bs − Bt+s ), sup (Bs+u − Bs+t )})) ∣

0⩽u⩽t y=supu⩽s (Br −Bs )
= E (u( max {y − Bt , sup (Bu − Bt )})) ∣

0⩽u⩽t y=Ys
Using time inversion (cf. 2.15) we see that W = (Wu )u∈[0,t] = (Bt−u − Bt )u∈[0,t]
is again a BM1 , and we get (Bt , sup0⩽u⩽t (Bu − Bt )) ∼ (Wt , sup0⩽u⩽t (Wu − Wt ))) =
(−Bt , sup0⩽u⩽t Bu )) (we understand the vector as a function of the whole process B
resp. W and use B ∼ W )
= E (u( max {y + Bt , sup Bu )})) ∣ .

0⩽u⩽t y=Ys
Using Solution 2 of Problem 6.8 we know the joint distribution of (Bt , supu⩽t Bu ):
E (u( max {y + Bt , sup Bu )}))

0⩽u⩽t
∞ z 2 2z − x −(2z−x)2 /2t
=∫ ∫ u(max{y + x, z}) √ e dx dz.
z=0 x=−∞ 2πt t
z z z
Splitting the integral ∫x=−∞ into two parts ∫x=−∞,y+x⩽z + ∫x=−∞,y+x>z we get
∞ 2 z−y 2z − x 2 ∞
e−(2z−x) /2t dx dz = √ −(z+y)2 /2t
2
I =∫ u(z) √ ∫ ∫ u(z) e dz
z=0 2πt x=−∞ t 2πt z=0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
z−y
=e−(2z−x) /2t ∣
2
−∞
and
2 ∞ z 2z − x −(2z−x)2 /2t
II = √ ∫ ∫ u(y + x) e dx dz
2πt z=0 x=−z−y t
2 ∞ x+y 2z − x
e−(2z−x) /2t dz
2
=√ ∫ u(y + x) ∫
2πt x=−y z=x t
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
x+y
=− 21 e−(2z−x) /2t ∣
2
z=x
54
1 ∞
u(y + x) [e−x /2t − e−(x+2y) /2t ] dx
2 2
dx = √ ∫
2πt x=−y
1 ∞
u(ξ) [e−(ξ−y) /2t − e−(ξ+y) /2t ] dξ.
2 2
=√ ∫
2πt x=−y
Finally, adding I and II we end up with
∞
E (u( max {y + Bt , sup Bu )})) = ∫ u(z)(gt (z + y) + gt (z − y)) dz, y⩾0
0⩽u⩽t 0
which is the same transition function as in part b).
(d) See part c).
∎∎
Problem 6.2. Solution: Let s, t ⩾ 0. We use the following abbreviations:

s
Is = ∫ Br dr and Ms = sup Bu and Fs = FsB .
0 u⩽s
(a) Let f ∶ R2 → R measurable and bounded. Then
E (f (Ms+t , Bs+t ) ∣ Fs )
= E (f ( sup Bu ∨ Ms , (Bs+t − Bs ) + Bs ) ∣ Fs )
s⩽u⩽s+t
= E (f ([Bs + sup (Bu − Bs )] ∨ Ms , (Bs+t − Bs ) + Bs ) ∣ Fs ) .

s⩽u⩽s+t
By the independent increments property of BM we get that the random variables

sups⩽u⩽s+t (Bu − Bs ), Bs+t − Bs á Fs while Ms and Bs are Fs measurable. Thus, we
can treat these groups of random variables separately (see, e.g., Lemma A.3:
E (f (Ms+t , Bs+t ) ∣ Fs )
= E (f ([z + sup (Bu − Bs )] ∨ y, (Bs+t − Bs ) + z) ∣ Fs ) ∣

s⩽u⩽s+t y=Ms ,z=Bs
= φ(Ms , Bs )
where
φ(y, z) = E (f ([z + sup (Bu − Bs )] ∨ y, (Bs+t − Bs ) + z) ∣ Fs ) .

s⩽u⩽s+t
(b) Let f ∶ R2 → R measurable and bounded. Then
E (f (Is+t , Bs+t ) ∣ Fs )
s+t
= E (f ( ∫ Bu du + Is , (Bs+t − Bs ) + Bs ) ∣ Fs )
s
s+t
= E (f ( ∫ (Bu − Bs ) du + Is + tBs , (Bs+t − Bs ) + Bs ) ∣ Fs ) .
s
55
By the independent increments property of BM we get that the random variables

s+t
∫s (Bu − Bs ) du, Bs+t − Bs á Fs while Is + tBs and Bs are Fs measurable. Thus, we
can treat these groups of random variables separately (see, e.g., Lemma A.3:
E (f (Is+t , Bs+t ) ∣ Fs )
s+t
= E (f ( ∫ (Bu − Bs ) du + y + tz, (Bs+t − Bs ) + z)) ∣
s y=Is ,z=Bs
= φ(Is , Bs )
for the function

s+t
φ(y, z) = E (f ( ∫ (Bu − Bs ) du + y + tz, (Bs+t − Bs ) + z)) .
s
(c) No! If we use the calculation of a) and b) for the function f (y, z) = g(y), i. e. only
depending on M or I, respectively, we see that we still get
E (g(It+s ) ∣ Fs ) = ψ(Bs , Is ),
i. e. (It , Ft )t cannot be a Markov process. The same argument applies to (Mt , Ft )t .
∎∎
Problem 6.3. Solution: We follow the hint.
First, if f ∶ Rd×n → R, f = f (x1 , . . . , xn ), x1 , . . . , xn ∈ Rd , we see that
Ex f (B(t1 )), . . . , B(tn ))

= E f (B(t1 )) + x, . . . , B(tn ) + x)
= ∫ ⋯ ∫ f (y1 + x, . . . , yn + x) P(B(t1 ) ∈ dy1 , . . . , B(tn ) ∈ dyn )

Rd Rd
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
n times
and the last expression is clearly measurable. This applies, in particular, to f = ∏nj=1 1Aj
where G ∶= ⋂nj=1 {B(tj ) ∈ Aj }, i. e. Ex 1G is Borel measurable.
Set
⎧
⎪ ⎫
⎪
⎪n ⎪
Γ ∶= ⎨ ⋂ {B(tj ) ∈ Aj } ∶ n ⩾ 0, 0 ⩽ t1 < ⋯tn , A1 , . . . An ∈ Bb (Rd )⎬ .
⎪
⎪ ⎪
⎪
⎩j=1 ⎭
Let us see that Σ is a Dynkin system. Clearly, ∅ ∈ Σ. If A ∈ Σ, then
x ↦ Ex 1Ac = Ex (1 − 1A ) = 1 − Ex 1A ∈ Bb (Rd ) Ô⇒ Ac ∈ Σ.
Finally, if (Aj )j⩾1 ⊂ Σ are disjoint and A ∶= ⊍j Aj we get 1A = ∑j 1Aj . Thus,
x ↦ Ex 1A = ∑ Ex 1Aj ∈ Bb (Rd ).
j
56
This shows that Σ is a Dynkin System. Denote by δ(⋅) the Dynkin system generated by
the argument. Then
Γ ⊂ Σ ⊂ F∞
B
Ô⇒ δ(Γ) ⊂ δ(Σ) = Σ ⊂ F∞
B
.
But δ(Γ) = σ(Γ) since Γ is stable under finite intersections and σ(Γ) = F∞
B
. This proves,
in particular, that Σ = F∞
B
.
Since we can approximate every bounded F∞

B
measurable function Z by step functions
with steps from F∞
B
, the claim follows.
∎∎
Problem 6.4. Solution: Following the hint we set un (x) ∶= (−n)∨x∧n. Then un (x) → u(x) ∶=
x. Using (6.7) we see
E [un (Bt+τ ) ∣ Fτ + ](ω) = EBτ (ω) un (Bt ).
Now take t = 0 to get

E [un (Bτ ) ∣ Fτ + ](ω) = un (Bτ )(ω)
and we get
lim E [un (Bτ ) ∣ Fτ + ](ω) = lim un (Bτ )(ω) = Bτ (ω).
n→∞ n→∞
Since the l.h.S. is Fτ + measurable (as limit of such measurable functions!), the claim
follows.
∎∎
Problem 6.5. Solution: By the reflection principle, Theorem 6.9,
P (sup ∣Bs ∣ ⩾ x) ⩽ P (sup Bs ⩾ x) + P (inf Bs ⩽ −x) = P(∣Bt ∣ ⩾ x) + P(∣Bt ∣ ⩾ x).

s⩽t s⩽t s⩽t
∎∎
(a) Since B(⋅) ∼ −B(⋅), we get
τb = inf{s ⩾ 0 ∶ Bs = b} ∼ inf{s ⩾ 0 ∶ −Bs = b} = inf{s ⩾ 0 ∶ Bs = −b} = τ−b .
(b) Since B(c−2 ⋅) ∼ c−1 B(⋅), we get
τcb = inf{s ⩾ 0 ∶ Bs = cb} = inf{s ⩾ 0 ∶ c−1 Bs = b}

∼ inf{s ⩾ 0 ∶ Bs/c2 = b}
= inf{rc2 ⩾ 0 ∶ Br = b}
= c2 inf{r ⩾ 0 ∶ Br = b} = c2 τb .
57
(c) We have
τb − τa = inf{s ⩾ 0 ∶ Bs+τa = b} = inf{s ⩾ 0 ∶ Bs+τa − Bτa = b − a}
which shows that τb − τa is independent of Fτa by the strong Markov property of

Brownian motion.
Now we find for all s, t ⩾ 0 and c ∈ [0, a]

τ ⩽τ
{τc ⩽ s} ∩ {τa ⩽ t} c= a {τc ⩽ s ∧ t} ∩ {τa ⩽ t} ∈ Ft∧s ∩ Ft ⊂ Ft .
This shows that {τc ⩽ s} ∈ Fτa , i. e. τc is Fτa measurable. Since c is arbitrary,

{τc }c∈[0,a] is Fτa measurable, and the claim follows.
∎∎
Problem 6.7. Solution: We begin with a simpler situation. As usual, we write τb for the first
passage time of the level b: τb = inf{t ⩾ 0 ∶ sups⩽t Bs = b} where b > 0. From Example 5.2
d) we know that (Mtξ ∶= exp(ξBt − 21 tξ 2 ))t⩾0 is a martingale. By optional stopping we get
that (Mt∧τ
ξ
)t⩾0 is also a martingale and has, therefore, constant expectation. Thus, for
b
ξ > 0 (and with E = E0 )
1 = E M0ξ = E ( exp(ξBt∧τb − 21 (t ∧ τb )ξ 2 ))
Since the RV exp(ξBt∧τb ) is bounded (mind: ξ ⩾ 0 and Bt∧τb ⩽ b), we can let t → ∞ and
get
1 = E ( exp(ξBτb − 21 τb ξ 2 )) = E ( exp(ξb − 12 τb ξ 2 ))
√
or, if we take ξ = 2λ,
√
E e−λτb = e− 2λb
.
As B ∼ −B, τb ∼ τ−b , and the above calculation yields

√
E e−λτb = e− 2λ∣b∣
∀b ∈ R.
○
Now let us turn to the situation of the problem. Set τ = τ(a,b) c . Here, Bt∧τ is bounded (it
is in the interval (a, b), and this makes things easier when it comes to optional stopping.
As before, we get by stopping the martingale (Mtξ )t⩾0 that
eξx = lim Ex ( exp(ξBt∧τ − 21 (t ∧ τ )ξ 2 )) = Ex ( exp(ξBτ − 12 τ ξ 2 )) ∀ξ

t→∞
(and not, as before, for positive ξ! Mind also the starting point x ≠ 0, but this does not
change things dramatically.) by, e.g., dominated convergence. The problem is now that
Bτ does not attain a particular value as it may be a or b. We get, therefore, for all ξ ∈ R
eξx = Ex ( exp(ξBτ − 21 τ ξ 2 )1{Bτ =a} ) + Ex ( exp(ξBτ − 12 τ ξ 2 )1{Bτ =b} )
= Ex ( exp(ξa − 12 τ ξ 2 )1{Bτ =a} ) + Ex ( exp(ξb − 12 τ ξ 2 )1{Bτ =b} )
58
√
Now pick ξ = ± 2λ. This yields 2 equations in two unknowns:
√ √ √
e 2λ x
=e 2λ a
Ex (e−λτ 1{Bτ =a} ) + e 2λ b
Ex (e−λτ 1{Bτ =b} )
√ √ √
e− 2λ x
= e− 2λ a
Ex (e−λτ 1{Bτ =a} ) + e− 2λ b
Solving this system of equations gives

√ √
2λ (x−a)
e = Ex (e−λτ 1{Bτ =a} ) + e 2λ (b−a)
√ √
e− 2λ (x−a)
= Ex (e−λτ 1{Bτ =a} ) + e− 2λ (b−a)
and so
√ √
sinh ( 2λ (x − a)) sinh ( 2λ (b − x))
Ex (e−λτ 1{Bτ =b} ) = √ and Ex (e−λτ 1{Bτ =a} ) = √ .
sinh ( 2λ (b − a)) sinh ( 2λ (b − a))
This answers Problem b) .
For the solution of Problem a) we only have to add these two expressions:
√ √
−λτ −λτ −λτ sinh ( 2λ (b − x)) + sinh ( 2λ (x − a))
Ee = E (e 1{Bτ =a} ) + E (e 1{Bτ =b} ) = √ .
sinh ( 2λ (b − a))
∎∎
Problem 6.8. Solution: Solution 1 (direct calculation): Denote by τ = τy = inf{s > 0 ∶ Bs = y}

the first passage time of the level y. Then Bτ = y and we get for y ⩾ x
P(Bt ⩽ x, Mt ⩾ y) = P(Bt ⩽ x, τ ⩽ t)
= P(Bt∨τ ⩽ x, τ ⩽ t)
= E ( E (1{Bt∨τ ⩽x} ∣Fτ + ) ⋅ 1{τ ⩽t} )
by the tower property and pull-out. Now we can use Theorem 6.11
= ∫ PBτ (ω) (Bt−τ (ω) ⩽ x) ⋅ 1{τ ⩽t} (ω) P(dω)
= ∫ Py (Bt−τ (ω) ⩽ x) ⋅ 1{τ ⩽t} (ω) P(dω)
= ∫ P(Bt−τ (ω) ⩽ x − y) ⋅ 1{τ ⩽t} (ω) P(dω)
= ∫ P(Bt−τ (ω) ⩾ y − x) ⋅ 1{τ ⩽t} (ω) P(dω)

B∼−B
= ∫ Py (Bt−τ (ω) ⩾ 2y − x) ⋅ 1{τ ⩽t} (ω) P(dω)
= ∫ PBτ (ω) (Bt−τ (ω) ⩾ 2y − x) ⋅ 1{τ ⩽t} (ω) P(dω)
= . . . = P(Bt ⩾ 2y − x, Mt ⩾ y) = P(Bt ⩾ 2y − x).

y⩾x
This means that

∞
(2πt)−1/2 e−z
2 /(2t)
P(Bt ⩽ x, Mt ⩾ y) = P(Bt ⩾ 2y − x) = ∫ dz
2y−x
59
and differentiating in x and y yields
2(2y − x) −(2y−x)2 /(2t)

P(Bt ∈ dx, Mt ∈ dy) = √ e dx dy.
2πt3
Solution 2 (using Theorem 6.18): We have (with the notation of Theorem 6.18)
dx x2 (x−2y)2
P(Mt < y, Bt ∈ dx) = lim P(mt > a, Mt < y, Bt ∈ dx) = √ [e− 2t − e− 2t ]
(6.19)
a→−∞ 2πt
and if we differentiate this expression in y we get
2(2y − x) −(2y−x)2 /(2t)

P(Bt ∈ dx, Mt ∈ dy) = √ e dx dy.
2πt3
∎∎
Problem 6.9. Solution: This is the so-called absorbed or killed Brownian motion. The result
is
1
(e−(x−z) /(2t) − e−(x+z) /(2t) ) dz,
2 2
Px (Bt ∈ dz, τ0 > t) = (gt (x − z) − gt (x + z)) dz = √
2πt
for x, z > 0 or x, z < 0.
To see this result we assume that x > 0. Write Mt = sups⩽t Bs and mt = inf s⩽t Bs for the
running maximum and minimum, respectively. Then we have for A ⊂ [0, ∞)
Px (Bt ∈ A, τ0 > t) = Px (Bt ∈ A, mt > 0)

= Px (Bt ∈ A, x ⩾ mt > 0)
(we start in x > 0, so the minimum is smaller!)
= P0 (Bt ∈ A − x, 0 ⩾ mt > −x)

= P0 (−Bt ∈ A − x, 0 ⩾ −Mt > −x)
B∼−B
= P0 (Bt ∈ x − A, 0 ⩽ Mt < x)
= ∬ 1A (x − a)1[0,x) (b) P0 (Bt ∈ da, Mt ∈ db)
Now we use the result of Problem 6.8:

2(2b − a) (2b − a)2
P0 (Bt ∈ da, Mt ∈ db) = √ exp (− ) da db
2πt3 2t
and we get
2(2b − a)
x (2b − a)2
Px (Bt ∈ A, τ0 > t) = ∫ 1A (x − a) [∫ √ exp (− ) db] da
0 2πt3 2t
t x 2 ⋅ 2 ⋅ (2b − a) (2b − a)2
= ∫ 1A (x − a) √ [∫ exp (− ) db] da
2πt3 0 2t 2t
1 x 2 ⋅ (2b − a) (2b − a)2
= ∫ 1A (x − a) √ [∫ exp (− ) db] da
2πt 0 t 2t
60
x
1 (2b − a)2
=√ ∫ 1A (x − a) [− exp (− )] da
2πt 2t b=0
1 a2 (2x − a)2
=√ ∫ 1A (x − a) {exp (− ) − exp (− )} da
2πt 2t 2t
1 (x − z)2 (x + z)2
=√ ∫ 1A (z) {exp (− ) − exp (− )} da.
2πt 2t 2t
The calculation for x < 0 is similar (actually easier): Let A ⊂ (−∞, 0]
Px (Bt ∈ A, τ0 > 0) = Px (Bt ∈ A, −x ⩽ Mt < 0)

= P0 (Bt ∈ A − x, 0 ⩽ Mt < −x)
2(2b − a) (2b − a)2
= ∬ 1A (a + x)1[0,−x) (b) √ exp (− ) db da
2πt2 2t
t −x 2 ⋅ (2b − a) (2b − a)2
= ∫ 1A (a + x) √ ∫ exp (− ) db da
2πt3 0 t 2t
−x
1 (2b − a)2
=√ ∫ 1A (a + x) [− exp (− )] da
2πt 2t b=0
1 a 2
(2x + a)2
=√ ∫ 1A (a + x) {exp (− ) − exp (− −)} da
2πt 2t 2t
1 (y − x)2 (x + y)2
=√ ∫ 1A (y) {exp (− ) − exp (− −)} da.
2πt 2t 2t
∎∎
Problem 6.10. Solution: For a compact set K ⊂ Rd the set Un ∶= K + B(0, 1/n) ∶= {x + y ∶ x ∈
K, ∣y∣ < 1/n} is open.
φn (x) ∶= d(x, Unc )/(d(x, K) + d(x, Unc )).
Since for d(x, z) ∶= ∣x − z∣ and all x, z ∈ Rd
d(x, A) ⩽ d(x, z) + d(z, A) Ô⇒ ∣d(x, A) − d(z, A)∣ ⩽ d(x, z),
we see that φn (x) is continuous. Obviously, 1Un (x) ⩾ φn (x) ⩾ φn+1 ⩾ 1K , and 1K = inf n φn
follows.
∎∎
Problem 6.11. Solution: Recall that P = P0 . We have for all a ⩾ t ⩾ 0
P(ξ̃t > a) = P (inf {s ⩾ t ∶ Bs = 0} > a)

= P (inf {h ⩾ 0 ∶ Bt+h = 0} + t > a)
= E [PBt (inf {h ⩾ 0 ∶ Bh = 0} > a − t)]
= E [P0 (inf {h ⩾ 0 ∶ Bh + x = 0} > a − t) ∣ ]

x=Bt
= E [P (inf {h ⩾ 0 ∶ Bh = −x} > a − t) ∣ ]

x=Bt
61
= E [P (τ−x > a − t) ∣ ]
x=Bt
= E [P (τBt > a − t)]

B∼−B
∣Bt ∣ −Bt2 /(2s)

∞
= E [∫ √
(6.13)
e ds]
a−t 2πs3
∞ ∣Bt ∣ −Bt2 /(2s)
= ∫ E [√ e ] ds.
a−t 2πs3
Thus, differentiating with respect to a and using Brownian scaling yields
⎡ ∣Bt ∣ Bt2 ⎤
⎢ ⎥
P(ξ̃t ∈ da) = E ⎢ √ exp (− )⎥
⎢ 2π(a − t)3 2(a − t) ⎥
⎣ ⎦
√
1 t ∣B1 ∣ 1 t
= √ E [√ √ exp (− B12 )]
(a − t) π a−t 2 2 a−t
1
= √ E [∣c B1 ∣ exp (−(c B1 )2 )]
(a − t) π
1
= √ E [∣Bc2 ∣ exp (−Bc22 )]
(a − t) π
where c2 = 1 t
2 a−t .
Now let us calculate for s = c2

∞
E [∣Bs ∣ e−Bs ] = (2πs)−1/2 ∫ ∣x∣ e−x e−x
2 2 2 /(2s)
dx
−∞
∞ 2 (1+(2s)−1 )
= (2πs)−1/2 2 ∫ x e−x dx
0
1 ∞ −1
= (2πs)−1/2 2(1 + (2s)−1 )x e−x (1+(2s) ) dx
2
−1 ∫
(1 + (2s) ) 0
1 2s −1 ∞
[e−x (1+(2s) ) ]
2
=√
2πs 2s + 1 x=0
1 2s
=√ .
2πs 2s + 1
Let (Bt )t⩾0 be a BM1 . Find the distribution of ξ̃t ∶= inf{s ⩾ t ∶ Bs = 0}. This gives
1 1 2c2
P(ξ̃t ∈ da) = √ √
(a − t) π 2π c 2c2 + 1
√
1 a−t t
= √
(a − t)π t (a − t)a/(a − t)
√
1 t
= .
aπ a − t
∎∎
Problem 6.12. Solution: We have seen in Problem 6.1 that M − B is a Markov process with
the same law as ∣B∣. This entails immediately that ξ ∼ η.
Attention: this problem shows that it is not enough to have only Mt − Bt ∼ ∣Bt ∣ for all
t ⩾ 0, we do need that the finite-dimensional distributions coincide. The Markov property
guarantees just this once the one-dimensional distributions coincide!
∎∎
62
(a) We have
P (Bt = 0 for some t ∈ (u, v)) = 1 − P (Bt ≠ 0 for all t ∈ (u, v)).
But the complementary probability is known from Theorem 6.19.

√
2 u
P (Bt ≠ 0 for all t ∈ (u, v)) = arcsin
π v
and so √
2 u
P (Bt = 0 for some t ∈ (u, v)) = 1 − arcsin .
π v
(b) Since (u, v) ⊂ (u, w) we find with the classical conditional probability that
P (Bt ≠ 0 ∀t ∈ (u, w) ∣ Bt ≠ 0 ∀t ∈ (u, v))

P ({Bt ≠ 0 ∀t ∈ (u, w)} ∩ {Bt ≠ 0 ∀t ∈ (u, v)})
=
P (Bt ≠ 0 ∀t ∈ (u, v))
P (Bt ≠ 0 ∀t ∈ (u, w))
=
P (Bt ≠ 0 ∀t ∈ (u, v))
√u
a) arcsin
= √w
arcsin uv
(c) We have
P (Bt ≠ 0 ∀t ∈ (0, w) ∣ Bt ≠ 0 ∀t ∈ (0, v))

= lim P (Bt ≠ 0 ∀t ∈ (u, w) ∣ Bt ≠ 0 ∀t ∈ (u, v))
u→0
√
arcsin wu
= lim √u
b)
u→0 arcsin
v
√ √
v v−u
= lim √ √
a)
l’Hôpital u→0 w w−u

√
v
=√ .
w
∎∎
63
7 Brownian motion and transition semigroups
Problem 7.1. Solution: Banach space: It is obvious that C∞ (Rd ) is a linear space. Let us
show that it is closed. By definition, u ∈ C∞ (Rd ) if
∀ > 0 ∃R > 0 ∀∣x∣ > R ∶ ∣u(x)∣ < . (*)
Let (un )n ⊂ C∞ (Rd ) be a Cauchy sequence for the uniform convergence. It is clear that
the uniform limit u = limn un is again continuous. Fix and pick R as in (*). Then we get
∣u(x)∣ ⩽ ∣un (x) − u(x)∣ + ∣un (x)∣ ⩽ ∥un − u∥∞ + ∣un (x)∣.
By uniform convergence, there is some n() such that
∣u(x)∣ ⩽ + ∣un() (x)∣ for all x ∈ Rd .
Since un() ∈ C∞ , we find with (*) some R = R(n(), ) = R() such that
∣u(x)∣ ⩽ + ∣un() (x)∣. ⩽ + ∀∣x∣ > R().
Density: Fix an and pick R > 0 as in (*), and pick a cut-off function χ = χR ∈ C(Rd )
such that
1B(0,R) ⩽ χR ⩽ 1B(0,2R) .
Clearly, supp χR is compact, χR ↑ 1, χR u ∈ Cc (Rd ) and
sup ∣u(x) − χR (x)u(x)∣ = sup ∣χR (x)u(x)∣ ⩽ sup ∣u(x)∣ < .

x ∣x∣>R ∣x∣>R
This shows that Cc (Rd ) is dense in C∞ (Rd ).
∎∎
Problem 7.2. Solution: Fix (t, y, v) ∈ [0, ∞) × Rd × C∞ (Rd ), > 0, and take any (s, x, u) ∈
[0, ∞) × Rd × C∞ (Rd ). Then we find using the triangle inequality
∣Ps u(x) − Pt v(y)∣ ⩽ ∣Ps u(x) − Ps v(x)∣ + ∣Ps v(x) − Pt v(x)∣ + ∣Pt v(x) − Pt v(y)∣
⩽ sup ∣Ps u(x) − Ps v(x)∣ + sup ∣Ps v(x) − Ps Pt−s v(x)∣ + ∣Pt v(x) − Pt v(y)∣
x x
= ∥Ps (u − v)∥∞ + ∥Ps (v − Pt−s v)∥∞ + ∣Pt v(x) − Pt v(y)∣

⩽ ∥u − v∥∞ + ∥v − Pt−s v∥∞ + ∣Pt v(x) − Pt v(y)∣
where we used the contraction property of Ps .
65
• Since y ↦ Pt v(y) is continuous, there is some δ1 = δ1 (t, y, v, ) such that ∣x − y∣ <

δ Ô⇒ ∣Pt v(x) − Pt v(y)∣ < .
• Using the strong continuity of the semigroup (Proposition 7.3 f) there is some δ2 =
δ2 (t, v, ) such that ∣t − s∣ < δ2 Ô⇒ ∥Pt−s v − v∥∞ ⩽ .
. This proves that for δ ∶= min{, δ1 , δ2 }
∣s − t∣ + ∣x − y∣ + ∥u − v∥∞ ⩽ δ Ô⇒ ∣Ps u(x) − Pt v(y)∣ ⩽ 3.
∎∎
Problem 7.3. Solution: By the tower property we find
Ex (f (Xt )g(Xt+s )) = Ex (Ex (f (Xt )g(Xt+s ) ∣ Ft ))

tower
property
= Ex (f (Xt ) Ex (g(Xt+s ) ∣ Ft ))
pull
out
= Ex (f (Xt ) EXt (g(Xs )))

Markov
property
= Ex (f (Xt )h(Xt ))
where, for every s,

h(y) = Ey g(Xs ) is again in C∞ .
Thus, Ex f (Xt )g(Xt+s ) = Ex φ(Xt ) and φ(y) = f (y)h(y) is in C∞ . This shows that
x ↦ Ex (f (Xt )g(Xt+s )) is in C∞ .
Using semigroups we can write the above calculation in the following form:
Ex (f (Xt )g(Xt+s )) = Ex (f (Xt )Ps g(Xt )) = Pt (f Ps g)(x)
i. e. h = Ps and φ = f ⋅ Ps g, and since Pt preserves C∞ , the claim follows.
∎∎
Set u(t, z) ∶= Pt u(z) = pt ⋆ u(z) = (2πt)d/2 ∫Rd u(y)e∣z−y∣

2 /2t
Problem 7.4. Solution: dy.
u(t, ⋅) is in C∞ for t > 0: Note that the Gauss kernel
pt (z − y) = (2πt)−d/2 e−∣z−y∣
2 /2t
, t>0
can be arbitrarily often differentiated in z and
∂zk pt (z − y) = Qk (z, y, t)pt (z − y)
where the function Qk (z, y, t) grows at most polynomially in z and y. Since pt (z − y)

decays exponentially, we see — as in the proof of Proposition 7.3 g) — that for each z
∣∂zk pt (z − y)∣
⩽ sup ∣Qk (z, y, t)∣ 1B(0,2R) (y) + sup ∣Qk (z, y, t)e−∣y∣
2 /(16t)
∣ e−∣y∣
2 /(16t)
1Bc (0,2R) (y).
∣y∣⩽2R ∣y∣⩾2R
66
This inequality holds uniformly in a small neighbourhood of z, i. e. we can use the differ-
entiation lemma from measure and integration to conclude that ∂ k Pt u ∈ Cb .
x ↦ ∂t u(t, x) is in C∞ for t > 0: This follows from the first part and the fact that
d ∣z − y∣2
∂t pt (z − y) = − (2πt)−d/2−1 e−∣z−y∣ /2t + (2πt)−d/2 e−∣z−y∣ /2t
2 2
2 2t2
1 ∣z − y∣2 d
= ( − ) pt (z − y).
2 t2 t
Again with the domination argument of the first part we see that ∂t ∂xk u(t, x) is continuous
on (0, ∞) × Rd .
∎∎
(a) Note that ∣un ∣ ⩽ ∣u∣ ∈ Lp . Since ∣un − u∣p ⩽ (∣un ∣ + ∣u∣)p ⩽ (∣u∣ + ∣u∣)p = 2p ∣u∣p ∈ L1
and since ∣un (x) − u(x)∣ → 0 for every x as n → ∞, the claim follows by dominated
convergence.
(b) Let u ∈ Lp and m < n. We have

Young
∥Pt un − Pt um ∥Lp = ∥pt ⋆ (un − um )∥Lp ⩽ ∥pt ∥L1 ∥un − um ∥Lp = ∥un − um ∥Lp .
Since (un )n is an Lp Cauchy sequence (it converges in Lp towards u ∈ Lp ), so is

(Pt un )n , and therefore P̃t u ∶= limn Pt un exists in Lp .
If vn is any other sequence in Lp with limit u, the above argument shows that
limn Pt vn also exists. ‘Mixing’ the sequences (wn ) ∶= (u1 , v1 , u2 , v2 , u3 , v3 , . . .) pro-
duces yet another convergent sequence with limit u, and we conclude that
lim Pt un = lim Pt wn = lim Pt vn ,

n n n
i. e. P̃t is well-defined.
(c) Any u ∈ Lp with 0 ⩽ u ⩽ 1 has a representative u ∈ Bb . And then the claim follows
since Pt is sub-Markovian.
(d) Recall that y ↦ ∥u(⋅ + y) − u∥Lp is for u ∈ Lp (dx) a continuous function. By Fubini’s
theorem and the Hölder inequality
∥Pt u − u∥pLp = ∫ ∣E u(x + Bt ) − u(x)∣p dx
⩽ E (∫ ∣u(x + Bt ) − u(x)∣p dx)
= E (∥u(⋅ + Bt ) − u∥pLp ) .
The integrand is bounded by 2p ∥u∥pLp , and continuous as a function of t; therefore we

can use the dominated convergence theorem to conclude that limt→0 ∥Pt u − u∥Lp = 0.
∎∎
67
Problem 7.6. Solution: Let u ∈ Cb . Then we have, by definition
Tt+s u(x) = ∫ u(z) pt+s (x, dz)

Rd
Tt (Ts u)(x) = ∫ Ts u(y) pt (x, dy)

Rd
=∫ ∫ u(z) ps (y, dz) pt (x, dy)

Rd Rd
=∫ u(z) ∫ ps (y, dz) pt (x, dy)

Rd Rd
By the semigroup property, Tt+s = Tt Ts , and we see that
pt+s (x, dz) = ∫ ps (y, dz) pt (x, dy).

Rd
If we pick u = 1C , this formal equality becomes
pt+s (x, C) = ∫ ps (y, C) pt (x, dy).

Rd
∎∎
Problem 7.7. Solution: Using Tt 1C (x) = pt (x, C) = ∫ 1C (y) pt (x, dy) we get
pxt1 ,...,tn (C1 × . . . × Cn )
= Tt1 (1C1 [Tt2 −t1 1C2 {⋯Ttn−1 −tn−2 ∫ 1Cn (xn ) ptn −tn−1 (⋅, dxn )⋯}])(x)
= Tt1 (1C1 [Tt2 −t1 1C2 {⋯ ∫ 1Cn−1 (xn−1 ) ∫ 1Cn (xn ) ptn −tn−1 (xn−1 , dxn )×
× ptn−1 −tn−2 (⋅, dxn−1 )⋯}])(x)
= ∫ . . . ∫ 1C1 (x1 )1C2 (x2 )⋯1Cn (xn )ptn −tn−1 (xn−1 , dxn )ptn−1 −tn−2 (xn−2 , dxn−1 )×
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
n integrals
⋯ × pt2 −t1 (x2 , dx2 )pt1 (x, dx1 )

n
= ∫ . . . ∫ 1C1 ×⋯×Cn (x1 , . . . , xn ) ∏ ptj −tj−1 (xj−1 , dxj )
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ j=1
n integrals
(we set t0 ∶= 0 and x0 ∶= x).
This shows that pxt1 ,...,tn (C1 × . . . × Cn ) is the restriction of

n
pxt1 ,...,tn (Γ) = ∫ . . . ∫ 1Γ (x1 , . . . , xn ) ∏ ptj −tj−1 (xj−1 , dxj ), Γ ∈ B(Rd⋅n )
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ j=1
n integrals
and the right-hand side clearly defines a probability measure. By the uniqueness theorem
for measures, each measure is uniquely defined by its values on the rectangles, so we are
done.
∎∎
68
(a) Let x, y ∈ Rd and a ∈ A. Then
inf ∣x − α∣ ⩽ ∣x − a∣ ⩽ ∣x − y∣ + ∣a − y∣
α∈A
Since this holds for all a ∈ A, we get
inf ∣x − α∣ ⩽ ∣x − y∣ + inf ∣a − y∣
α∈A a∈A
and, since x, y play symmetric roles,
∣d(x, A) − d(y, A)∣ = ∣ inf ∣x − α∣ − inf ∣a − y∣∣ ⩽ ∣x − y∣.

α∈A a∈A
d(x,Unc )
(b) By definition, Un = K + B(0, 1/n) and un (x) ∶= d(x,K)+d(x,Unc ) . Being a combination
of continuous functions, see Part (a), un is clearly continuous. Moreover,
un ∣K ≡ 1 and un ∣Unc ≡ 0.
n→∞
This shows that 1K ⩽ un ⩽ 1Unc ÐÐÐ→ 1K .
Picture: un is piecewise linear.
(c) Assume, without loss of generality, that supp χn ⊂ B(0, 1/n2 ). Since 0 ⩽ un ⩽ 1, we
find
χn ⋆ un (x) = ∫ χn (x − y)un (y) dy ⩽ ∫ χn (x − y) dy = 1 ∀x.
Now we observe that for γ ∈ (0, 1)
d(y, Unc ) (1 − γ)/n

un (y) = ⩾ = 1 − γ. ∀y ∈ K + B(0, γ/n)
d(y, K) + d(y, Unc ) 1/n
(Essentially this means that un is ‘linear’ for x ∈ Un ∖ K!). Thus, if γ > 1/n,
χn ⋆ un (x) = ∫ χn (x − y)un (y) dy
⩾ (1 − γ) ∫ χn (x − y)1K+B(0,γ/n) (y) dy
= (1 − γ) ∫ χn (x − y)1B(0,1/n2 ) (x − y)1K+B(0,γ/n) (y) dy
= (1 − γ) ∫ χn (x − y)1x+B(0,1/n2 ) (y)1K+B(0,γ/n) (y) dy
⩾ (1 − γ) ∫ χn (x − y)1x+B(0,1/n2 ) (y) dy
=1−γ ∀x ∈ K.
This shows that
1 − γ ⩽ lim inf χn ⋆ un (x) ⩽ lim sup χn ⋆ un (x) ⩽ 1 ∀x ∈ K,

n n
hence,
lim χn ⋆ un (x) = x for all x ∈ K.
n→∞
69
On the other hand, if x ∈ K c , there is some n ⩾ 1 such that d(x, K) > 1

n + 1
n2
. Since
1 1 1 1
+ 2 < d(x, K) ⩽ d(x, y) + d(y, K) Ô⇒ d(x, y) > 2 or d(y, K) > ,
n n n n
and so, using that supp χn ⊂ B(0, 1/n2 ) and supp un ⊂ K + B(0, 1/n),
1 1
χn ⋆ un (x) = ∫ χn (x − y)un (y) dy = 0 ∀x ∶ d(x, K) > + 2.
n n
It follows that limn χn ⋆ un (x) = 0 for x ∈ K c .
Remark 1: If we are just interested in a smooth function approximating 1K we could
use vn ∶= χn ⋆ 1K+supp un where (χn )n is any sequence of type δ. Indeed, as before,
χn ⋆ 1K+supp un (x) = ∫ χn (x − y)1K+supp un (y) dy ⩽ ∫ χn (x − y) dy = 1 ∀x.
For x ∈ K we find
χn ⋆ 1K+supp un (x) = ∫ χn (x − y)1K+supp un (y) dy
= ∫ χn (y)1K+supp un (x − y) dy
= ∫ χn (y) dy
=1 ∀x ∈ K.
As before we get χn ⋆ 1K+supp un (x) = 0 if d(x, K) > 2/n.

Thus, limn χn ⋆ 1K+supp un (x) = 0 if x ∈ K c .
Remark 2: The naive approach χn ⋆ 1K will, in general, not lead to a (pointwise
everywhere) approximation of 1K : consider K = {0}, then χn ⋆ 1K ≡ 0. In fact, since
1K ∈ L1 we get χn ⋆ 1K → 1K in L1 hence, for a subsequence, a.e. ...
∎∎

(a) This follows from part (c).
(b) This follows by approximating 1K from above by a decreasing sequence of C∞ func-
tions. Such a sequence exists, see Problem 7.8 above.
Remark: If we know that the kernel pt (x, K) ∶= Tt 1K (x) is inner (compact) regular —
or outer (open) regular, which is the same and does always hold in this topologically
nice situation with bounded measures, see Schilling [15, 15.18, 15.19, pp. 159–160]
— then we get that (t, x) ↦ pt (x, C) is measurable for all Borel measures C ⊂ Rd .
Just observe that pt (x, C) = sup{pt (x, K) ∶ K ⊂ C, K compact} and use the fact the
the supremum is attained by a sequence (which may depend on C, of course).
(c) This is a standard 3--trick. Fix > 0, fix (s, x, u) and consider another point (t, y, w).
Without loss of generality we assume that s ⩽ t. Then, by the triangle inequality,
the semigroup property and the contractivity (in the last step), we get
∣Ts u(x) − Tt w(y)∣ ⩽ ∣Ts u(x) − Ts u(y)∣ + ∣Ts u(y) − Tt u(y)∣ + ∣Tt u(y) − Tt w(y)∣
70
⩽ ∣Ts u(x) − Ts u(y)∣ + ∥Ts u − Tt u∥∞ + ∥Tt (u − w)∥∞

⩽ ∣Ts u(x) − Ts u(y)∣ + ∥Ts (u − Tt−s u)∥∞ + ∥Tt (u − w)∥∞
⩽ ∣Ts u(x) − Ts u(y)∣ + ∥u − Tt−s u∥∞ + ∥u − w∥∞ .
Now we know that for given there are δ1 , δ2 , δ3 such that
∥u − w∥∞ < δ1 Ô⇒ ∥u − w∥∞ < (pick δ1 = )

∣t − s∣ < δ2 Ô⇒ ∥Tt−s u − u∥∞ < (by strong continuity)
∣x − y∣ < δ3 Ô⇒ ∣Ts u(x) − Ts u(y)∣ < (by the Feller property).
This proves continuity with δ ∶= min(δ1 , δ2 , δ3 ); note that δ may (and will) depend
on as well as on the fixed point (s, x, u), as we require continuity at this point only.
Mind that there are minor, but obvious, changes necessary if s = 0.
Remark: A full account on Feller semigroups can be found in Böttcher–Schilling–
Wang [1, Chapter 1].
∎∎

(a) Existence, contractivity: Let us, first of all, check that the series converges. Denote
by ∥A∥ any matrix norm in Rd . Then we see
X
X X
X ∞
(tA)j XX ∞ tj ∥Aj ∥ ∞ j
t ∥A∥j
∥Pt ∥ = X
X
X
X
X ∑
X
X
X
X
X ⩽ ∑ ⩽ ∑ = et∥A∥ .
X
X
Xj=0
j! X
X
X j=0
j! j=0 j!
This shows that, in general, Pt is not a contraction. We can make it into a contraction
by setting Qt ∶= e−t∥A∥ Pt . It is clear that Qt is again a semigroup, if Pt is a semigroup.
Semigroup property: This is shown using as for the one-dimensional exponential se-
ries. Indeed,
∞
(t + s)k Ak
e(t+s)A = ∑
k=0 k!
∞ k
1 k j k−j k
=∑∑ ( )t s A
k=0 j=0 k! j
∞ k
tj Aj sk−j Ak−j
=∑∑
k=0 j=0 j! (k − j)!
∞
tj Aj ∞ sk−j Ak−j
=∑ ∑
j=0 j! k=j (k − j)!
∞
tj Aj ∞ sl Al
=∑ ∑
j=0 j! l=0 l!
= etA esA .
Strong continuity: We have

X
X ∞ j jX X ∞ j−1 j X
X t A XX X
X t A XX
∥etA − id∥ = X
X
X
X ∑
X
X
X
X = t X
X
X
X ∑
X
X
X
X
X
X
X j! X
X
X X
X
X j! X
X
X
Xj=1 X Xj=1 X
71
and, as in the first calculation, we see that the series converges absolutely. Letting
t → 0 shows strong continuity, even continuity in the operator norm.
(Strictly speaking, strong continuity means that for each vector v ∈ Rd
lim ∣etA v − v∣ = 0.
t→0
Since
∣etA v − v∣ ⩽ ∥etA − id ∥ ⋅ ∣v∣
strong continuity is implied by uniform continuity. One can show that the generator
of a norm-continuous semigroup is already a bounded operator, see e.g. Pazy.)
(b) Let s, t > 0. Then
∞ ∞
tj Aj sj Aj (tj − sj )Aj
etA − esA = ∑ ( − )=∑
j=0 j! j! j=1 j!
Since the sum converges absolutely, we get

etA − esA ∞ (tj − sj ) Aj s→t ∞ j−1 Aj
=∑ ÐÐ→ ∑ jt .
t−s j=1 t − s j! j=1 j!
The last expression, however, is

∞ j ∞
j−1 A Aj−1
∑ jt = A ∑ tj−1 = AetA .
j=1 j! j=1 (j − 1)!
A similar calculation, pulling out A to the back, yields that the sum is also etA A.
(c) Assume first that AB = BA. Repeated applications of this rule show Aj B k = B k Aj
for all j, k ⩾ 0. Thus,
∞ ∞
tj Aj tk B k ∞ ∞ tj tk Aj B k ∞ ∞ tk tj B k Aj
etA etB = ∑ ∑ =∑∑ =∑∑ = etB etA .
j=0 k=0 j! k! j=0 k=0 j!k! k=0 j=0 k!j!
Conversely, if etA etB = etB etA for all t > 0, we get

etA − id etB − id etB − id etA − id
lim = lim
t→0 t t t→0 t t
and this proves AB = BA.
Alternative solution for the converse: If s = j/n and t = k/n for some common de-
nominator n, we get from etA etB = etB etA that
1 1 1 1 1 1 1 1
etA esB = e n A ⋯e n A e n B ⋯e n B = e n B ⋯e n B e n A ⋯e n A = esB etA .
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
k j j k
Thus, if s, t > 0 are dyadic numbers, we get

etA − id sB etA − id
AesB = lim e = esB lim = esB A
t→0 t t→0 t
and,
esB − id esB − id
AB = A lim = lim A = BA.
s→0 s s→0 s
72
(d) We have
∞
Aj 1
eA/k = id + k1 A + ρk and k 2 ρk = ∑ j−2
.
j=2 j! k
Note that k 2 ρk is bounded. Do the same for B (with the remainder term ρ′k ) and
multiply these expansions to get
eA/k eB/k = id + k1 A + k1 B + σk
where k 2 σk is again bounded. In particular, if k ≫ 1,
∥ k1 A + k1 B + σk ∥ < 1.
This allows us to (formally) apply the logarithm series
log(eA/k eB/k ) = 1
k A + k1 B + σk + σk′
where k 2 σk′ is bounded. Multiply with k to get
k log(eA/k eB/k ) = A + B + τk
with kτk bounded. Then we get
eA+B = lim eA+B+τk

k→∞
A/k eB/k )
= lim ek log(e
k→∞
A/k eB/k ) k
= lim (elog(e )
k→∞
k
= lim (eA/k eB/k )
k→∞
Alternative Solution: Set Sk = e(A+B)/k and Tk = eA/k eB/k . Then

k−1
Skk − Tkk = ∑ Skj (Sk − Tk )Tkk−1−j .
j=0
This shows that

k−1
∥Skk − Tkk ∥ ⩽ ∑ ∥Skj (Sk − Tk )Tkk−1−j ∥
j=0
k−1
⩽ ∑ ∥Skj ∥ ⋅ ∥Sk − Tk ∥ ⋅ ∥Tkk−1−j ∥
j=0
⩽ k ∥Sk − Tk ∥ ⋅ max{∥Sk ∥, ∥Tk ∥}k−1

⩽ k ∥Sk − Tk ∥ ⋅ e∥A∥+∥B∥ .
Observe that X X
X
X ∞
(A + B)j ∞ ∞ Aj B l X X
∥Sk − Tk ∥ = X X
X X C
X
X
X ∑ − ∑ ∑ X
j j! k l l! X
X ⩽ 2
X
X
Xj=0
k j j!
j=0 l=0 k X
X
X
k
with a constant C depending only on ∥A∥ and ∥B∥. This yields Skk − Tkk → 0.
73
∎∎
Problem 7.11. Solution: Add “D(B) ⊂ D(A)” to the statement of the problem.
(a) Add “on D(B)” to the statement of the problem.
Let 0 < s < t and assume throughout that h ∈ R is such that t − s − h > 0. We have
Pt−(s+h) Ts+h − Pt−s Ts

= Pt−(s+h) Ts+h − Pt−(s+h) Ts + Pt−(s+h) Ts − Pt−s Ts
= Pt−(s+h) (Ts+h − Ts ) + (Pt−(s+h) − Pt−s )Ts
= (Pt−(s+h) − Pt−s )(Ts+h − Ts ) + Pt−s (Ts+h − Ts ) + (Pt−(s+h) − Pt−s )Ts .
Divide by h ≠ 0 to get for all u ∈ D(B) ⊂ D(A)
1
(Pt−(s+h) Ts+h u − Pt−s Ts u)
h
Ts+h u − Ts u Ts+h u − Ts u Pt−(s+h) − Pt−s
= (Pt−(s+h) − Pt−s ) + Pt−s + Ts u
h h h
= I + II + III.
Letting h → 0 gives for all u ∈ D(B) ⊂ D(A)
II → Pt−s BTs and III → −Pt−s ATs
Lemma
(we use for the last asserting that Ts (D(B)) ⊂ D(B) ⊂ D(A)). Let us show that
7.10.a)
I → 0. We have
Ts+h u − Ts u
I = (Pt−(s+h) − Pt−s ) ( − Ts Bu) + (Pt−(s+h) − Pt−s )Ts Bu = I1 + I2 .
h
By the strong continuity of the semigroup (Pt )t , we see that I2 → 0 as h → 0.
Furthermore, by contractivity,
Ts+h u − Ts u Ts+h u − Ts u
∥I1 ∥ ⩽ (∥Pt−(s+h) ∥ + ∥Pt−s ∥) ⋅ ∥ − Ts Bu∥ ⩽ 2 ∥ − Ts Bu∥ → 0
h h
since u ∈ D(B).
Remark. Usually this identity is used if D(A) = D(B), for example we have used it
in this way in the proof of Corollary 7.11.c) on page 90. Another, typical application
is the situation where B−A is a bounded operator (hence, D(A) = D(B)). Integrating
the identity of part a) yields
t d t
Tt u − Pt u = ∫ (Pt−s Ts )u ds = ∫ Pt−s (B − A)Ts u ds
0 ds 0
which is often referred to as Duhamel’s formula. This formula holds first for u ∈ D(B)
and then, by extension of bounded linear operators defined on a dense set, for all u
in the closure D(B).
74
(b) In general, no. The problem is the semigroup property (unless Tt and Ps commute
for all s, t ⩾ 0):
Ut Us = Tt Pt Ts Ps ≠ Tt Ts Pt Ps = Tt+s Pt+s = Ut+s .
In (c) we see how this can be ‘remedied’.
It is interesting to note (and helpful for the proof of (c)) that Ut is an operator on
C∞ :
Pt Tt
Ut ∶ C∞ ÐÐÐÐ→ C∞ ÐÐÐÐ→ C∞
and that Ut is strongly continuous: for all s, t ⩾ 0 and f ∈ C∞
∥Ut f − Us f ∥ = ∥Tt Pt f − Ts Pt f + Ts Pt f − Ts Ps f ∥
⩽ ∥(Tt − Ts )Pt f ∥ + ∥Ts (Pt − Ps )f ∥
⩽ ∥(Tt − Ts )Pt f ∥ + ∥(Pt − Ps )f ∥
and, as s → t, both expressions tend to 0 since f, Pt f ∈ C∞ .

n
(c) Set Ut,n ∶= (Tt/n Pt/n ) .
Ut is a contraction on C∞ : By assumption, Pt/n and Tt/n map C∞ into itself and,

therefore, Tt/n Pt/n ∶ C∞ → C∞ as well as Ut,n .
We have ∥Ut,n f ∥ = ∥Tt/n Pt/n ⋯Tt/n Pt/n f ∥ ⩽ ∏nj=1 ∥Tt/n ∥∥Pt/n ∥∥f ∥ ⩽ ∥f ∥. So, by the
continuity of the norm
∥Ut f ∥ = ∥lim Ut,n f ∥ = lim ∥Ut,n f ∥ ⩽ ∥f ∥.

n n
Strong continuity: Since the limit defining Ut is locally uniform in t, it is enough to

show that Ut,n is strongly continuous. Let X, Y be contractions in C∞ . Then we get
X n − Y n = X n−1 X − X n−1 Y + X n−1 Y − Y n−1 Y

= X n−1 (X − Y ) + (X n−1 − Y n−1 )Y
hence, by the contraction property,
∥X n f − Y n f ∥ ⩽ ∥(X − Y )f ∥ + ∥(X n−1 − Y n−1 )Y f ∥.
By iteration, we get
n−1
∥X n f − Y n f ∥ ⩽ ∑ ∥(X − Y )Y k f ∥.
k=0
Take Y = Tt/n Pt/n , X = Ts/n Ps/n where n is fixed. Then letting s → t shows the
strong continuity of each t ↦ Ut,n .
Semigroup property: Let s, t ∈ Q and write s = j/m and t = k/m for the same m.
Then we take n = l(j + k) and get
n l(j+k)
(T s+t P s+t ) = (T 1 P 1 )
n n lm lm
75
lj lk
= (T 1 P 1 ) (T 1 P 1 )
lm lm lm lm
lj lk
= (T j P j ) (T k P k )
ljm ljm lkm lkm
lj lk
= (T ljs P ljs ) (T t P t )
lk lk
Since n → ∞ ⇐⇒ l → ∞ ⇐⇒ lk, lj → ∞, we see that Us+t = Us Ut for rational s, t.
For arbitrary s, t the semigroup property follows by approximation and the strong
continuity of Ut : let Q ∋ sn → s and Q ∋ tn → t. Then, by the contraction property,
∥Us Ut f − Usn Utn f ∥ ⩽ ∥Us Ut f − Us Utn f ∥ + ∥Us Utn f − Usn Utn f ∥

⩽ ∥Ut f − Utn f ∥ + ∥(Us − Usn )(Utn − Ut )f ∥ + ∥(Us − Usn )Ut f ∥
⩽ ∥Ut f − Utn f ∥ + 2∥(Utn − Ut )f ∥ + ∥(Us − Usn )Ut f ∥
and the last expression tends to 0. The limit limn Usn +tn u = Us+t u is obvious.
Generator: Let us begin with a heuristic argument (by ? and ?? indicate the steps
which are questionable!). By the chain rule
d d
∣ Ut g = ∣ lim(Tt/n Pt/n )n g
dt t=0 dt t=0 n
d
= lim ∣ (Tt/n Pt/n )n g
?
n dt t=0
n−1
= lim [n(Tt/n Pt/n ) (Tt/n n1 BPt/n + Tt/n n1 APt/n )g∣ ]
??
n t=0
= Bg + Ag.
So it is sensible to assume that D(A)∩D(B) is not empty. For the rigorous argument
we have to justify the steps marked by question marks.
Since D(B) ⊂ D(A), we can argue as follows: ?? We have to show that d

ds Ts Ps f
exists and is Ts Af + BPs f for f ∈ D(A) ∩ D(B). This follows similar to (a) since we
have for s, h > 0
Ts+h Ps+h f − Ts Ps f = Ts+h (Ps+h − Ps )f + (Ts+h − Ts )Ps f

= (Ts+h − Ts )(Ps+h − Ps )f + Ts (Ps+h − Ps )f + (Ts+h − Ts )Ps f.
Divide by h. Then the first term converges to 0 as h → 0, while the other two terms
tend to Ts Af and BPs f , respectively.
? This is a matter of interchanging limit and differentiation. Recall the following

theorem from calculus, e.g. Rudin [13, Theorem 7.17].
Theorem. Let (fn )n be a sequence of differentiable functions on [0, ∞) which con-

verges for some t0 > 0. If (fn′ )n converges [locally] uniformly, then (fn )n converges
[locally] uniformly to a differentiable function f and we have f ′ = limn fn′ .
76
This theorem holds for functions with values in any Banach space space and, there-
fore, we can apply it to the situation at hand: Fix g ∈ D(A) ∩ D(B); we know
that fn (t) ∶= Ut,n g converges (even locally uniformly) and, because of ?? , that
fn′ (t) = (Tt/n Pt/n )n−1 (Tt/n A + BPt/n )g.
Since limn (Tt/n Pt/n )n u converges locally uniformly, so does limn (Tt/n Pt/n )n−1 u; more-
over, by the strong continuity, Tt/n A + BPt/n → (A + B)g locally uniformly for
g ∈ D(A) ∩ D(B). Therefore, the assumptions of the theorem are satisfied and
we may interchange the limits in the calculation above.
Remark. It is surprisingly difficult to verify that A + B is the generator of the
semigroup Ut – even if one already knows that the Trotter Formula converges. The
obvious failure is that the canonical (pre-)domain of A + B, the set D(A) ∩ D(B) is
too small, i.e. not dense or even empty!. The assumption that D(B) ⊂ D(A) is a
strong, but still reasonable assumption. Alternatively one can require that A and B
commute.
The usual statements of Trotter’s formula, see e.g. the excellent monograph by En-
gel & Nagel [5, Chapter III.5], is such that one has a condition on D(A) ∩ D(B)
which also ensures that the limit defining Ut exists. In an L2 -context one can find a
counterexample on p. 229 of [5].
∎∎
Problem 7.12. Solution: The idea is to show that A = − 12 ∆ is closed when defined on C2∞ (R).
Since C2∞ (R) ⊂ D(A) and since (A, D(A)) is the smallest closed extension, we are done.
So let (un )n ⊂ C2∞ (R) be a sequence such that un → u uniformly and (Aun )n is a C∞
Cauchy sequence. Since C∞ (R) is complete, we can assume that u′′n → 2g uniformly for
some g ∈ C∞ (Rd ). The aim is to show that u ∈ C2∞ .
(a) By the fundamental theorem of differential and integral calculus we get
x x y
un (x) − un (0) − xu′n (0) = ∫ (u′n (y) − u′n (0)) dy = ∫ ∫ u′′n (z) dz.
0 0 0
Since u′′n → 2g uniformly, we get

x y x y
un (x) − un (0) − xu′n (0) = ∫ ∫ u′′n (z) dz → ∫ ∫ 2g(z) dz.
0 0 0 0
Since un (x) → u(x) and un (0) → u(0), we conclude that u′n (0) → c converges.
(b) Recall the following theorem from calculus, e.g. Rudin [13, Theorem 7.17].
Theorem. Let (fn )n be a sequence of differentiable functions on [0, ∞) which con-

verges for some t0 > 0. If (fn′ )n converges uniformly, then (fn )n converges uniformly
to a differentiable function f and we have f ′ = limn fn′ .
If we apply this with fn′ = u′′n → 2g and fn (0) = u′n (0) → c, we get that u′n (x)−u′n (0) →
x
∫0 2g(z) dt.
77
Let us determine the constant c′ ∶= limn u′n (0). Since u′n converges uniformly, the
limit as n → ∞ is in C∞ , and so we get
x
− lim u′n (0)) = lim lim (u′n (x) − u′n (0)) = lim ∫ 2g(z) dz
n→∞ x→−∞ n→∞ x→−∞ 0
i. e. c′ = ∫−∞ g(z) dz. We conclude that u′n (x) → ∫−∞ g(z) dt uniformly.
0 x
x y
(c) Again by the Theorem quoted in (b) we get un (x) − un (0) → ∫0 ∫−∞ 2g(z) dz uni-
0 y
formly, and with the same argument as in (b) we get un (0) = ∫−∞ ∫−∞ 2g(z) dz.
∎∎
Problem 7.13. Solution: By definition, (for all α > 0 and formally but justifiable via monotone
convergence also for α = 0)
∞
Uα 1C (x) = ∫ e−αt Pt 1C (x) dt
0
∞
=∫ e−αt E 1C (Bt + x) dt
0
∞
= E∫ e−αt 1C−x (Bt ) dt.
0
This is the ‘discounted’ (with ‘interest rate’ α) total amount of time a Brownian motion
spends in the set C − x.
∎∎
Problem 7.14. Solution: Let u ∈ Bb (Rd ) and α, β > 0. By the definition of the potential
operator we get
∞ ∞
(β − α)Uα Uβ u(x) = (β − α) ∫ ∫ e−sα e−tβ Ps+t u(x) ds dt
0 0
∞ ∞
= (β − α) ∫ e−(r−t)α e−tβ Pr u(x) dr dt
r=s+t
∫
0 t
∞ r
= (β − α) ∫ e−t(β−α) e−rα Pr u(x) dt dr
Fubini
∫
0 0
∞ t=r
= ∫ [ − e−t(β−α) ] e−rα Pr u(x) dr
0 t=0
∞
= ∫ (e−rα − e−rβ )Pr u(x) dr
0
= Uα u(x) − Uβ u(x).
∎∎
Problem 7.15. Solution: First formula: We use induction. The induction start with n = 0 is
clearly correct. Let us assume that the formula holds for some n and we do the induction
step n ↝ n + 1. We have for β ≠ α
dn dn
dn+1 dαn Uα f (x) − dβ n Uβ f (x)
Uα f (x) = lim
dαn+1 β→α β−α
n!(−1)n Uαn+1 f (x) − n!(−1)n Uβn+1 f (x)
= lim
β→α β−α
78
Uαn+1 f (x) − Uβn+1 f (x)

= n!(−1)n lim
β→α β−α
Using the identity an+1 − bn+1 = (a − b) ∑nj=0 an−j bj we get, since the resolvents commute,
Uαn+1 f (x) − Uβn+1 f (x) Uα − Uβ n n−j j n

= ∑ Uα Uβ f (x) = −Uα Uβ ∑ Uα Uβ f (x)
n−j j
β−α β − α j=0 j=0
In the last line we used the resolvent identity. Now we can let β → α to get
β→α n
ÐÐ→ −Uα Uα ∑ Uαn−j Uαj f (x) = −(n + 1)Uαn+2 f (x).
j=0
This finishes the induction step.
Second formula: We use Leibniz’ formula for the derivative of a product:

n
n
∂ n (f g) = ∑ ( )∂ j f ∂ n−j g
j=0 j
and we get, using the first formula

n n
∂ n (αUα f (x)) = ( )α∂ n Uα f (x) + ( )∂ n−1 Uα f (x)
0 1
= αn!(−1)n Uαn+1 f (x) + n(n − 1)!(−1)n−1 Uαn f (x)
= n!(−1)n+1 (id −αUα )Uαn f (x).
∎∎
Problem 7.16. Solution: Using Dini’s Theorem (e. g.: Rudin, p. 150) we see that
lim sup ∣f (x) − fn (x)∣ = 0

n→∞ x∈K
for any compact set K ⊂ Rd . Fix > 0 and pick a compact set K = K ⊂ Rd such that
0 ⩽ fn (x) ⩽ f (x) ⩽ on Rd ∖ K. Then
lim sup ∣fn (x) − f (x)∣ ⩽ lim sup ∣fn (x) − f (x)∣ + lim sup ∣fn (x)∣ + sup ∣f (x)∣
n→∞ x n→∞ x∈K n→∞ x∉K x∉K
⩽ 3.
Remark: Positivity is, in fact, not needed. Here is the argument: Let f1 ⩽ f2 ⩽ . . . ⩽ fn ⩽ . . .
and fn ∈ C∞ (Rd ) (any sign is now allowed!) and set f = supn fn . Using Dini’s Theorem
(e. g.: Rudin, p. 150) we see that
lim sup ∣f (x) − fn (x)∣ = 0

n→∞ x∈K
for any compact set K ⊂ Rd . Fix > 0 and pick a compact set K = K ⊂ Rd such that
− ⩽ f1 (x) ⩽ fn (x) ⩽ f (x) ⩽ on Rd ∖ K. Then
lim sup ∣fn (x) − f (x)∣ ⩽ lim sup ∣fn (x) − f (x)∣ + lim sup(f (x) − fn (x))
n→∞ x n→∞ x∈K n→∞ x∉K
79
⩽ lim sup ∣fn (x) − f (x)∣ + 2

n→∞ x∈K
⩽ 3.
Alternative Solution (non-positive case): Apply the solution of the positive case to the
sequence gn ∶= fn − f1 . This is possible since 0 ⩽ gn ⩽ gn+1 and supn gn = f − f1 .
∎∎
Problem 7.17. Solution: Because of Lemma 7.24 c),d) it is enough to show that there are
positive functions u, v ∈ C∞ (Rd ) such that
sup Uα u, sup Uα v ∈ C∞ (Rd ) but U0 (u − v)(x) = sup Uα u(x) − sup Uα v(x) ∉ C2 (Rd ).
α>0 α>0 α>0 α>0
As in Example 7.25 we have
U0 u(x) = αd ∫ ∣y − x∣2−d u(y) dy

Rd
for any u ∈ C+∞ (Rd ) with αd = π −d/2 Γ( d2 )/(d − 2). Since ∫∣y∣⩽1 ∣y∣2−d dy < ∞, we see with
dominated convergence that the function U0 w is in C∞ (Rd ) for all u ∈ C+∞ (Rd ) ∩ L1 (dy).
Pick any f ∈ Cc ([0, 1)) such that f (0) = 0. We denote by (x1 , . . . , xd ) points in Rd and
set r2 ∶= x21 + . . . + x2d . Then let
x2d
u(x1 , . . . , xd ) ∶= γ f (r), v(x1 , . . . , xd ) ∶= f (r)
r2
and
w(x1 , . . . , xd ) ∶= u(x1 , . . . , xd ) − v(x1 , . . . , xd ).
We will show that there is some f and a constant γ > 0 such that w ∈ D(U0 ) and
U0 w ∉ C2 (Rd ). The first assertion follows directly from Lemma 7.24 d). Introducing polar
coordinates
yd = r cos θd−2 ,
yd−1 = r cos θd−3 ⋅ sin θd−2 ,
yd−2 = r cos θd−4 ⋅ sin θd−3 ⋅ sin θd−2 ,
⋮
y2 = r cos φ sin θ1 ⋅ . . . ⋅ sin θd−2 ,
y1 = r sin φ sin θ1 ⋅ . . . ⋅ sin θd−2 ,
and using the integral formula for U0 u and U0 v, we get for xd ∈ (0, 1/2)
U0 w(0, . . . , 0, xd )
1 π (γ cos2 θd−2 − 1)(sin θd−2 )d−2 d−2 π
= αd [ ∫ rd−1 f (r)( ∫ √ d−2
dθ d−2 ) dr] ∏ (∫ (sin θj ) dθj ) .
j
0 0 0
r2 + x2d − 2xd r cos θd−2 j=1
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
=∶βd
Note that βd > 0. For brevity we write x = xd and θ = θd−2 . From
γ cos2 θ − 1 = −γ sin2 θ − (1 − γ)
80
we conclude
π (γ cos2 θ − 1)(sin θ)d−2
∫ √ d−2
dθ
0
r2 + x2 − 2xr cos θ
π (sin θ)d π (sin θ)d−2
= −γ ∫ √ d−2
dθ − (1 − γ) ∫ √ d−2
dθ
0
r2 + x2 − 2xr cos θ 0
=∶ −γI1 (r, x) − (1 − γ)I2 (r, x).
By (??), there exist constants bd , cd ∈ R such that

1 π (sin θ)d 1 r 2
I1 (r, x) = ∫ √ d−2
dθ = (bd ( ) + cd )
xd−2 0 2 xd−2 x
( xr ) + 1 − 2 xr cos θ
for any 0 < r < x. Similarly,
1 π (sin θ)d 1 x 2
I1 (r, x) = ∫ √ d−2
dθ = (bd ( ) + cd )
rd−2 0 2 rd−2 r
1 + ( xr ) − 2 xr cos θ
for x < r < 1. Analogously, we find by (??)
⎧
⎪ ad
⎪
⎪ 0<r<x
⎪ d−2
I2 (r, x) = ⎨ xa
⎪
⎪
⎪
d
x<r<1
⎪
⎩ rd−2
for some ad ∈ R. It is not difficult to see that 0 < cd < ad . Therefore, we may choose
γ = γd = ad /(ad − cd ). Then,
⎧
⎪ r2
⎪
⎪
⎪−γ b 0<r<x
π (γd cos θ − 1)(sin θ)
2 d−2
⎪ d d xd
∫ √ dθ = ⎨
0
d−2 ⎪
⎪
⎪ x2
⎪
⎪−γ b x < r < 1.
⎩ d d rd
Hence,
1 x 1 f (r)
U0 w(0, . . . , 0, x) = −αd βd γd bd ( d ∫ rd+1 f (r) dr + x2 ∫ dr) .
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ x 0 x r
=∶Cd
The remaining part of the proof follows as in Example 7.25. Differentiating in x yields
d d x 1 f (r)
U0 w(0, . . . , 0, x) = Cd (− d+1 ∫ rd+1 f (r) dr + 2x ∫ dr) .
dx x 0 x r
It is not hard to show that limx→0+ d
dx U0 w(0, . . . , 0, x) = 0. Thus,
dx U0 w(0, . . . , 0, x) − dx U0 w(0, . . . , 0)
d d
d x 1 f (r)
= Cd (− ∫ rd+1 f (r) dr + 2 ∫ dr) .
x xd+2 0 x r
Applying l’Hôpital’s rule we obtain
dx U0 w(0, . . . , 0, x) − dx U0 (0, . . . , 0)
d d 1 f (r)
d
lim = Cd (− f (0) + 2 lim ∫ dr) .
x→0+ x d+2 x→0+ x r
This means that the second derivative of U0 w at x = 0 in xd -direction does not exist if
1 f (r) −1
∫0+ r dr diverges. A canonical candidate is f (r) = ∣ log r∣ χ(r) with a suitable cut-off
function χ ∈ C+c ([0, 1)) and χ∣[0,1/2] ≡ 1.
∎∎
81
(a) The process (t, Bt ) starts at (0, B0 ) = 0, and if we start at (s, x) we consider the
process (s + t, x + Bt ) = (s, x) + (t, Bt ). Let f ∈ Bb ([0, ∞) × R). Since the motion in
t is deterministic, we can use the probability space (Ω, A, P = P) generated by the
Brownian motion (Bt )t⩾0 . Then
Tt f (s, x) ∶= E(s,x) f (t, Bt ) ∶= E f (s + t, x + Bt ).
Tt preserves C∞ ([0, ∞) × R): If f ∈ C∞ ([0, ∞) × R), we see with dominated conver-

gence that
lim Tt f (σ, ξ) = lim E f (σ + t, ξ + Bt )

(σ,ξ)→(s,x) (σ,ξ)→(s,x)
=E lim f (σ + t, ξ + Bt )
(σ,ξ)→(s,x)
= E f (s + t, x + Bt )
= Tt f (s, x)
which shows that Tt preserves f ∈ Cb ([0, ∞) × R). In a similar way we see that
lim Tt f (σ, ξ) = E lim f (σ + t, ξ + Bt ) = 0,

∣(σ,ξ)∣→∞ ∣(σ,ξ)∣→∞
i. e. Tt maps C∞ ([0, ∞) × R) into itself.
Tt is a semigroup: Let f ∈ C∞ ([0, ∞)×R). Then, by the independence and stationary

increments property of Brownian motion,
Tt+τ f (s, x) = E f (s + t + τ, x + Bt+τ )

= E f (s + t + τ, x + (Bt+τ − Bt ) + Bt )
= E E(t,Bt ) f (s + τ, x + (Bt+τ − Bt ))
= E E(t,Bt ) f (s + τ, x + Bτ )
= E Tτ f (s + t, x + Bt )
= Tt Tτ f (s, x).
Tt is strongly continuous: Since f ∈ C∞ ([0, ∞) × R) is uniformly continuous, we see

that for every > 0 there is some δ > 0 such that
∣f (s + h, x + y) − f (s, x)∣ ⩽ ∀h + ∣y∣ ⩽ 2δ.
So, let t < h < δ, then
∣Tt f (s, x) − f (s, x)∣ = ∣ E (f (s + t, x + Bt ) − f (s, x))∣
⩽∫ ∣f (s + t, x + Bt ) − f (s, x)∣ d P +2∥f ∥∞ P(∣Bt ∣ > δ)

∣Bt ∣⩽δ
82
1
⩽ + 2∥f ∥∞ E(Bt2 )
δ2
t
= + 2∥f ∥∞ .
δ2
Since the estimate is uniform in (s, x), this proves strong continuity.
Markov property: this is trivial.
(b) The transition semigroup is
Tt f (s, x) = E f (s + t, x + Bt ) = (2πt)−1/2 ∫ f (s + t, x + y) e−y

2 /(2t)
dy.
R
The resolvent is given by

∞
Uα f (s, x) = ∫ e−tα Tt f (s, x) dt
0
and the generator is, for all f ∈ C 1,2

([0, ∞) × R)
Tt f (s, x) − f (s, x) E f (s + t, x + Bt ) − f (s, x)
=
t t
E [f (s + t, x + Bt ) − f (s, x + Bt )] E f (s, x + Bt ) − f (s, x)
= +
t t
t→0
ÐÐ→ E ∂t f (s, x + B0 ) + 12 ∆x f (s, x)
= (∂t + 12 ∆x )f (s, x).
Note that, in view of Theorem 7.22, pointwise convergence is enough (provided the
pointwise limit is a C∞ -function).
(s,x)
(c) We get for u ∈ C1,2
∞ that under P
t
Mtu ∶= u(s + t, x + Bt ) − u(s, x) − ∫ (∂r + 21 ∆x )u(s + r, x + Br ) dr
0
is an Ft -martingale. This is the same assertion as in Theorem 5.6 (up to the choice
of u which is restricted here as we need it in the domain of the generator...).
∎∎
Problem 7.19. Solution: Let u ∈ D(A) and σ a stopping time with Ex σ < ∞. Use optional
stopping (Theorem A.18 in combination with remark A.21) to see that
σ∧t
u
Mσ∧t ∶= u(Xσ∧t ) − u(x) − ∫ Au(Xr ) dr
0
is a martingale (for either Ft or Fσ∧t ). If we take expectations we get

σ∧t
Ex u(Xσ∧t ) − u(x) = Ex (∫ Au(Xr ) dr) .
0
Since u, Au ∈ C∞ we see
σ∧t σ∧t
∣Ex (∫ Au(Xr ) dr)∣ ⩽ Ex (∫ ∥Au∥∞ dr) ⩽ ∥Au∥∞ ⋅ Ex σ < ∞,
0 0
i. e. we can use dominated convergence and let t → ∞. Because of the right-continuity of

the paths of a Feller process we get Dynkin’s formula (7.28).
∎∎
83
Problem 7.20. Solution: Clearly,
P(Xt ∈ F ∀t ∈ R+ ) ⩽ P(Xq ∈ F ∀q ∈ Q+ ).
On the other hand, since F is closed and Xt has continuous paths,
Xq ∈ F ∀q ∈ Q+ Ô⇒ Xt = +lim Xq ∈ F ∀t ⩾ 0
Q ∋q→t
and the converse inequality follows.
∎∎
84
8 The PDE connection
Write gt (x) = (2πt)−d/2 e−∣x∣

2 /2t
Problem 8.1. Solution: for the heat kernel. Since convolutions
are smoothing, one finds easily that P f = g ⋆ f ∈ C∞
∞ ⊂ D(∆). (There is a more general
concept behind it: whenever the semigroup is analytic—i. e. z ↦ Pz has an extension to,
say, a sector in the complex plane and it is holomorphic there—one has that Tt maps the
underlying Banach space into the domain of the generator; cf. e.g. Pazy [10, pp. 60–63].)
Thus, if we set f ∶= P f , we can apply Lemma 8.1 and find that
u (t, x) = Pt f (x) = Pt P f (x) = Pt+ f (x).

Lemma 8.1 def semi-
group
By the strong continuity of the heat semigroup, we find that

uniformly
u (t, x) ÐÐÐÐÐ→ Pt f (x).
→0
Moreover,
∂ 1
u (t, ⋅) = ∆x Pt P f
∂t 2
1 uniformly 1
= P ( ∆x Pt f ) ÐÐÐÐÐ→ ∆x Pt f.
2 →0 2
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
∈C∞
Since both the sequence and the differentiated sequence converge uniformly, we can inter-
change differentiation and the limit, cf. [13, Theorem 7.17, p. 152], and we get
∂ ∂ 1
u(t, x) = lim u (t, x) = ∆x u(t, x)
∂t →0 ∂t 2
and
u (0, ⋅) = P f ÐÐ→ f = u(0, ⋅)
→0
and we get a solution for the initial value f . The proof of the uniqueness part in Lemma 8.1
stays valid.
∎∎
t
dt ∫0 f (Bs ) ds = f (Bt ) so that f (Bt ) = 0.
d
Problem 8.2. Solution: By differentiation we get
We can assume that f is positive and bounded, otherwise we could consider f ± (Bt ) ∧ c
for some constant c > 0. Now E f (Bt ) = 0 and we conclude from this that f = 0.
∎∎
85
(a) Note that

t t
∣χn (Bt )e−α ∫0 gn (Bs ) ds ∣ ⩽ ∣e−α ∫0 ds ∣ = e−αt ⩽ 1
is uniformly bounded. Moreover,

t t
lim χn (Bt )e−α ∫0 gn (Bs ) ds = 1R (Bt )e−α ∫0 1(0,∞) (Bs ) ds
n→∞
which means that, by dominated convergence,

∞ t
vn,λ (x) = ∫ e−λt E (χn (Bt )e−α ∫0 gn (Bs ) ds ) dt ÐÐÐ→ vλ (x).
0 n→∞
Moreover, we get that ∣vλ (x)∣ ⩽ λ−1 .
If we rearrange (8.12) we see that
′′
vn,λ (x) = 2(αχn (x) + λ)vn,λ (x) − gn (x), (*)
′′
and since the expression on the right has a limit as n → ∞, we get that limn→∞ vn,λ (x)
exists.
(b) Integrating (*) we find

x x
′ ′
vn,λ (x) − vn,λ (0) = 2 ∫ (αχn (y) + λ)vn,λ (y) dy − ∫ gn (y) dy, (**)
0 0
′ ′
and, again by dominated convergence, we conclude that limn→∞ [vn,λ (x) − vn,λ (0)]
exists. In addition, the right-hand side is uniformly bounded (for all ∣x∣ ⩽ R):
x x R R
∣2 ∫ (αχn (y) + λ)vn,λ (y) dy − ∫ gn (y) dy∣ ⩽ 2 ∫ (α + λ) dy + ∫ dy
0 0 0 0
⩽ 2(α + λ + 1)R.
Integrating (**) reveals

x
′ ′ ′
vn,λ (x) − vn,λ (0) − x vn,λ (0) = ∫ [vn,λ (z) − vn,λ (0)] dz.
0
Since the expression under the integral converges boundedly and since limn→∞ vn,λ (x)
′ ′
exists, we conclude that limn→∞ vn,λ (0) exists. Consequently, limn→∞ vn,λ (x) exists.
(c) The above considerations show that
vλ (x) = lim vn,λ (x)

n→∞
vλ′ (x) ′
= lim vn,λ (x)
n→∞
vλ′′ (x) ′′
= lim vn,λ (x)
n→∞
∎∎
t
Problem 8.4. Solution: We have to show that v(t, x) ∶= ∫0 Ps g(x) ds is the unique solution
of the initial value problem (8.7) with g = g(x) satisfying ∣v(t, x)∣ ⩽ C t.
86
Existence: The linear growth bound is obvious from ∣Ps g(x)∣ ⩽ ∥Ps g∥∞ ⩽ ∥g∥∞ < ∞. The
rest follows from the hint if we take A = 1
2 ∆ and Lemma 7.10.
∞
Uniqueness: We proceed as in the proof of Lemma 8.1. Set vλ (x) ∶= ∫0 e−λt v(t, x) dt.
This integral is, for λ > 0, convergent and it is the Laplace transform of v(⋅, x). Under the
Laplace transform the initial value problem (8.7) with g = g(x) becomes
λvλ (x) − Avλ (x) = λ−1 g(x)
and this problem has a unique solution, cf. Proposition 7.13 f). Since the Laplace trans-
form is invertible, we see that v is unique.
∎∎
Problem 8.5. Solution: Integrating u′′ (x) = 0 twice yields
u′ (x) = c and u(x) = cx + d
with two integration constants c, d ∈ R. The boundary conditions u(0) = a and u(1) = b
show that
d = a and c = b − a
so that
u(x) = (b − a)x + a.
On the other hand, by Corollary 5.11 (Wald’s identities), Brownian motion started in
x ∈ (0, 1) has the probability to exit (at the exit time τ ) the interval (0, 1) in the following
way:
Px (Bτ = 1) = x and Px (Bτ = 0) = 1 − x.
Therefore, if f ∶ {0, 1} → R is a function on the boundary of the interval (0, 1) such that
f (0) = a and f (1) = b,then
Ex f (Bτ ) = (1 − x)f (0) + xf (1) = (b − a)x + a.
This means that u(x) = Ex f (Bτ ), a result which we will see later in Section 8.4 in much
greater generality.
∎∎
Problem 8.6. Solution: The key is to show that all points in the open and bounded, hence
relatively compact, set D are non-absorbing. Thus the closure of D has an neighbourhood,
say V ⊃ D̄ such that E τDc ⩽ E τV c . Let us show that E τV c < ∞.
Since D is bounded, there is some R > 0 such that B(0, R) ⊃ D̄. Pick some test function
χ = χR such that χ∣Bc (0,R) ≡ 0 and χ ∈ Cc∞ (Rd ). Pick further some function u ∈ C2 (Rd )
such that ∆u > 0 in B(0, 2R). Here are two possibilities to get such a function:
d
u(x) = ∣x∣2 = ∑ x2j Ô⇒ 1
2 ∆u(x) = 1
j=1
87
or, if f ∈ Cb (Rd ), f ⩾ 0 and f = f (x1 ) we set

x1
F (x) = F (x1 ) ∶= ∫ f (z1 ) dz1
0
and
x1 x1 y1
U (x) = U (x1 ) ∶= ∫ F (y1 ) dy1 = ∫ ∫ f (z1 ) dz1 .
0 0 0
Clearly, 1
2 ∆U (x) = 1
2 ∂x21 U (x1 ) = f (x1 ), and we can arrange things by picking the correct
f.
Problem: neither u nor U will be in D(∆) (unless you are so lucky as in the proof of
Lemma 8.8 to pick instantly the right function).
Now observe that
χ ⋅ u, χ ⋅ U ∈ C2c (Rd ) ⊂ D(∆)

∆(χ ⋅ U ) = χ ⋅ ∆U + U ⋅ ∆χ + 2⟨∇χ, ∇U ⟩
which means that

∆(χ ⋅ U )∣B(0,R) = ∆U ∣B(0,R) .
The rest of the proof follows now either as in Lemma 7.33 or Lemma 8.8 (both employ,
anyway, the same argument based on Dynkin’s formula).
∎∎
Problem 8.7. Solution: We are following the hint. Let L = ∑dj,k=1 ajk (x) ∂j ∂k + ∑dj=1 bj (x) ∂j .
Then
L(χf ) = ∑ ajk ∂j ∂k (χf ) + ∑ bj ∂j (χf )

j,k j
= ∑ ajk (∂j ∂k χ + ∂j ∂k f + ∂k χ∂j f + ∂j χ∂k f ) + ∑ bj (f ∂j χ + χ∂j f )

j,k j
= χLf + f Lχ + ∑(ajk + akj )∂j χ∂k f.

j,k
If ∣x∣ < R and χ∣B(0,R) = 1, then L(uχ)(x) = Lu(x). Set u(x) = e−x1 /γr . Then only the
2 2
derivatives in x1 -direction give any contribution and we get

2 2
x1
2x1 − γr 2 2x21 x
− 12
∂1 u(x) = − e 2
and ∂12 u(x) = ( − 1) e γr
γr2 γr2 γr2
Thus we get for L(−u) = −Lu and any ∣x∣ < r

2 2
2a11 (x) 2x21 − γr
x1
2b1 (x)x1 − γr
x1
−Lu(x) = (1 − ) e 2
+ e 2
γr2 γr2 γr2
2
2a11 (x) 2x21 2b1 (x)x1 − γr
x1
=[ (1 − ) + ] e 2
γr2 γr2 γr2
2a0 2 2b0 − γr r2
⩾ [ 2 (1 − ) − ]e 2
γr γ γr
88
This shows that the drift b1 (x) can make the expression in the bracket negative!
Let us modify the Ansatz. Observe that for f (x) = f (x1 ) we have
Lf (x) = a11 (x)∂12 f (x) − b1 (x)∂1 f (x)
and if we know that ∂12 f, ∂1 f ⩾ 0 we get

!!
Lf (x) ⩾ a0 ∂12 f (x) − b0 ∂1 f (x) > 0.
This means that ∂12 f /∂1 f > b0 /a0 seems to be natural and a reasonable Ansatz would be
x1 2b0
y
f (x) = ∫ e a0 dy.
0
Then
2b0
x1 2b0 2b 0 x
∂1 f (x) = e a0 and ∂12 f (x) = e a0 1
a0
and we get
2b0 2b 0 x 2b0
x
Lf (x) = a11 (x) e a0 1 − b1 (x)e a0 1
a0
2b0 2b 0 x 2b0
x
⩾ a0 e a0 1 − b0 e a0 1
a0
2b0
x1
⩾ (2b0 − b0 ) e a0 > 0.
With the above localization trick on balls, we are done.
∎∎
Problem 8.8. Solution: Assume that B0 = 0. Any other starting point can be reduced to this
situation by shifting Brownian motion to B0 = 0. The LIL shows that a Brownian motion
satisfies
B(t) B(t)
−1 = lim √ < lim √ =1
t→0
t→0 2t log log 1t 2t log log 1t
√
i. e. B(t) oscillates for t → 0 between the curves ± 2t log log 1t . Since a Brownian motion
has continuous sample paths, this means that it has to cross the level 0 infinitely often.
∎∎
Problem 8.9. Solution: The idea is to proceed as in Example 8.12 e) where Zaremba’s needle
plays the role a truncated flat cone in dimension d = 2 (but in dimension d ⩾ 3 it has
too small dimension). The set-up is as follows: without loss of generality we take x0 = 0
(otherwise we shift Brownian motion) and we assume that the cone lies in the hyperplane
{x ∈ Rd ∶ x1 = 0} (otherwise we rotate things).
Let B(t) = (b(t), β(t)), t ⩾ 0, be a BMd where b(t) is a BM1 and β(t) is a (d − 1)-
dimensional Brownian motion. Since B is a BMd , we know that the coordinate processes
b = (b(t))t⩾0 and β = (β(t))t⩾0 are independent processes. Set σn = inf{t > 1/n ∶ b(t) = 0}.
89
Since 0 ∈ R is regular for {0} ⊂ R, see Example 8.12 e), we get that limn→∞ σn = τ{0} = 0
almost surely with respect to P0 . Since β á b, the random variable β(σn ) is rotationally
symmetric (see, e.g., the solution to Problem 8.10).
Let C be a flat (i. e. in the hyperplane {x ∈ Rd ∶ x1 = 0}) cone such that some truncation
C ′ of it lies in Dc . By rotational symmetry, we get
opening angle of C
P0 (β(σn ) ∈ C) = γ = .
full angle
By continuity of BM, β(σn ) → β(0) = 0, and this gives
P0 (β(σn ) ∈ C ′ ) = γ.
Clearly, B(σn ) = (b(σn ), β(σn )) = (0, β(σn )) and {β(σn ) ∈ C ′ } ⊂ {τDc ⩽ σn }, so
P0 (τDc = 0) = lim P0 (τDc ⩽ σn ) ⩾ lim P0 (β(σn ) ∈ C ′ ) ⩾ γ > 0.

n→∞ n→∞
Now Blumenthal’s 0-1–law, Corollary 6.22, applies and gives P0 (τDc = 0) = 1.
∎∎
Problem 8.10. Solution: Proving that the random variable β(σn ) is absolutely continuous
with respect to Lebesgue measure is relatively easy: note that, because of the independence
of b and β, hence σn and β,
d 0 d
− P (β(σn ) ⩾ x) = − ∫ P0 (βt ⩾ x) P(σn ∈ dt)
dx dx R
d
= ∫ − P0 (βt ⩾ x) P(σn ∈ dt)
R dx
1
e−x /(2t) P(σn ∈ dt)
2
=∫ √
R 2πt
∞ 1
e−x /(2t) P(σn ∈ dt).
2
=∫ √
1/n 2πt
(observe, for the last equality, that σn takes values in [1/n, ∞).) Since the integrand
is bounded (even as t → 0), the interchange of integration and differentiation is clearly
satisfied.
(d − 1)-dimensional version: Let β be a (d − 1)-dimensional version as in Problem 8.9

Proving that the random variable β(σn ) is rotationally symmetric is easy: note that,
because of the independence of b and β, hence σn and β, we have for all Borel sets
A ⊂ Rd−1
∞
P0 (β(σn ) ∈ A) = ∫ P0 (βt ∈ A) P(σn ∈ dt)
1/n
and this shows that the rotational symmetry of β is inherited by β(σn ).
90
We even get a density by formally replacing A by dx:
β(σn ) ∼ ∫ P0 (βt ∈ dx) P(σn ∈ dt)

R
∞ 1 −∣x∣2 /(2t)
=∫ e P(σn ∈ dt) dx.
1/n (2πt)(d−1)/2
(here x ∈ Rd−1 ).
It is a bit more difficult to work out the exact shape of the density. Let us first determine
the distribution of σn . Clearly,
{σn > t} = { inf 1/n⩽s⩽t ∣b(s)∣ > 0}.
By the Markov property of Brownian motion we get
P0 (σn > t) = P0 ( inf 1/n⩽s⩽t ∣b(s)∣ > 0)
= E0 Pb(1/n) ( inf s⩽t−1/n ∣b(s)∣ > 0)
= E0 (1{b(1/n)>0} Pb(1/n) ( inf s⩽t−1/n b(s) > 0)
+ 1{b(1/n)<0} Pb(1/n) ( sups⩽t−1/n b(s) < 0))
= E0 (1{b(1/n)>0} P0 ( inf s⩽t−1/n b(s) > −y)
+ 1{b(1/n)<0} P0 ( sups⩽t−1/n b(s) < −y)∣ )

y=b(1/n)
= E0 (1{b(1/n)>0} P0 ( sups⩽t−1/n b(s) < y)

b∼−b
+ 1{b(1/n)<0} P0 ( sups⩽t−1/n b(s) < −y)∣ )

y=b(1/n)
= E0 (1{b(1/n)>0} P0 ( sups⩽t−1/n b(s) < y)

b∼−b
+ 1{b(1/n)>0} P0 ( sups⩽t−1/n b(s) < −y)∣ )

−y=b(1/n)
= 2 E0 (1{b(1/n)>0} P0 ( sups⩽t−1/n b(s) < y)∣ )

y=b(1/n)
= 2 E0 (1{b(1/n)>0} P0 (∣b(t − 1/n)∣ < y)∣ )

(6.12)
y=b(1/n)
∞
=4∫ P0 (b(t − 1/n) < y) P0 (b(1/n) ∈ dy)
0
2 1 ∞ y
√ ∫ ∫ e−z /2(t−1/n) dz e−ny /2 dy
2 2
= √
π t− 1 1 0 0
n n
√
change of variables: ζ = z/ t − 1
n
√ √
2 n ∞ 1
y/ t− n
e−ζ /2 dζ e−ny /2 dy.
2 2
= ∫ ∫
π 0 0
91
For the density we differentiate in t:

√ √
d 0 2 n d ∞ 1
y/ t− n
e−ζ /2 dζ eny /2 dy
2 2
− P (σn > t) = − ∫ ∫
dt √ π dt 0 0
n ∞
1 −3/2 −y 2 /2(t− n1
) −ny 2 /2
= (t − n ) ∫ ye e dy
√π 0
∞ 2
n −3/2 −y nt
= (t − n1 ) ∫ y e 2 t−1/n dy
π 0
√ ∞
n −3/2 t − n1 −y
2
nt
= (t − n1 ) [−e 2 t−1/n ]
π nt y=0
√
n −1/2 1
= (t − n1 )
π nt
1 1
= √ .
π t nt − 1
Now we proceed with the d-dimensional case. We have for all x ∈ Rd−1
∞ 1
e−∣x∣ /(2t) P(σn ∈ dt) dx
2
β(σn ) ∼ ∫ (d−1)/2
1/n (2πt)
1 ∞ 1
e−∣x∣ /(2t) dt
2
= (d+1)/2 (d−1)/2 ∫ √
π 2 1/n t(d+1)/2 nt − 1
n(d−1)/2 ∞ 1 −n∣x∣2 /(2s)

= ∫ √ e ds
π (d+1)/2 2(d−1)/2 1 s(d+1)/2 s − 1
(∗) n(d−1)/2
= (d+1)/2 (d−1)/2 B( d2 , 12 ) 1F1 ( d2 , d+1
2 ; − 2 ∣x∣
n 2
)
π 2
where B(⋅, ⋅) is Euler’s Beta-function and 1F1 is the degenerate hypergeometric function,
cf. Gradshteyn–Ryzhik [8, Section 9.20, 9.21] and, for (∗), [8, Entry 3.471.5, p. 340].
∎∎
92
9 The variation of Brownian paths
Problem 9.1. Solution: Let > 0 and Π = {t0 = 0 < t1 < . . . < tm = 1} be any partition of
[0, 1]. As a continuous function on a compact space, f is uniformly continuous, i. e. there
exists δ > 0 such that ∣f (x) − f (y)∣ <
2m for all x, y ∈ [0, 1] with ∣x − y∣ < δ. Pick n0 ∈ N so
′ ∣Π∣
that ∣Πn ∣ < δ ∶= δ ∧ 2 for all n ⩾ n0 .
∣Π∣
Now, the balls B(tj , δ ′ ) for 0 ⩽ j ⩽ m are disjoint as δ ′ ⩽ 2 . Therefore the sets B(tj , δ ′ ) ∩
Πn0 for 0 ⩽ j ⩽ m are also disjoint, and non-empty as ∣Πn0 ∣ < δ ′ . In particular, there exists
a subpartition Π′ = {q0 = 0 < q1 < . . . < qm = 1} of Πn0 such that ∣tj − qj ∣ < δ ′ ⩽ δ for all
0 ⩽ j ⩽ m. This implies
RRR m m RRR m
RRR ∣f (t ) − f (t )∣ − RR
RRR ∑ j j−1 ∑ ∣f (q j ) − f (q j−1 RRR ⩽ ∑ ∣∣f (tj ) − f (tj−1 )∣ − ∣f (qj ) − f (qj−1 )∣∣
)∣
RRj=1 j=1 RRR j=1
m
⩽ ∑ ∣f (tj ) − f (qj ) + f (tj−1 ) − f (qj−1 )∣
j=1
m
⩽ 2 ⋅ ∑ ∣f (tj ) − f (qj )∣
j=0
⩽ .
Because adding points to a partition increases the corresponding variation sum, we have
′ Πn0
S1Π (f, 1) ⩽ S1Π (f, 1) + ⩽ S1 (f, 1) + ⩽ lim S1Πn (f, 1) + ⩽ VAR1 (f, 1) +
n→∞
and since Π was arbitrarily chosen, we deduce
VAR1 (f, 1) ⩽ lim S1Πn (f, 1) + ⩽ VAR1 (f, 1) +

n→∞
for every > 0. Letting tend to zero completes the proof.
Remark: The continuity of the function f is essential. A counterexample would be Dirich-

let’s discontinuous function f = 1Q∩[0,1] and Πn a refining sequence of partitions made up
of rational points.
∎∎
Problem 9.2. Solution: Note that the problem is straightforward if ∥x∥ stands for the maxi-
mum norm: ∥x∥ = max1⩽j⩽d ∣xj ∣.
Remember that all norms on Rd are equivalent. One quick way of showing this is the
following: Denote by ej with j ∈ {1, . . . , d} the usual basis of Rd . Then
∥x∥ ⩽ (d ⋅ max ∥ej ∥) ⋅ max ∣xj ∣ = (d ⋅ max ∥ej ∥) ⋅ ∥x∥∞ =∶ B ⋅ ∥x∥∞

1⩽j⩽d 1⩽j⩽d 1⩽j⩽d
93
for every x = ∑dj=1 xj ej in Rd using the triangle inequality and the positive homogeneity
of norms. In particular, x ↦ ∥x∥ is a continuous mapping from Rd equipped with the
supremum-norm ∥ ⋅ ∥∞ to R, since
∣∥x∥ − ∥y∥∣ ⩽ ∥x − y∥ ⩽ B ⋅ ∥x − y∥∞
holds for every x, y in Rd . Hence, the extreme value theorem claims that x ↦ ∥x∥ attains
its minimum on the compact set {x ∈ Rd ∶ ∥x∥∞ = 1}. Finally, this implies A ∶= min{∥x∥ ∶
∥x∥∞ = 1} > 0 and hence
x
∥x∥ = ∥ ∥ ⋅ ∥x∥∞ ⩾ A ⋅ ∥x∥∞
∥x∥∞
for every x ≠ 0 in Rd as required.
As a result of the equivalence of norms on Rd , it suffices to consider the supremum-norm

to determine the finiteness of variations. In particular, VARp (f ; t) < ∞ if, and only if,
⎧
⎪ ⎫
⎪
⎪ ⎪
sup ⎨ ∑ ∣g(tj ) − g(tj−1 )∣p ∨ ∣h(tj ) − h(tj−1 )∣p ∶ Π finite partition of [0, 1]⎬
⎪
⎪ ⎪
⎪
⎩tj−1 ,tj ∈Π ⎭
is finite. But this term is bounded from below by VARp (g; t) ∨ VARp (h; t) and from above
by VARp (g; t) + VARp (h; t), which proves the desired result.
∎∎
Problem 9.3. Solution: Let p > 0, > 0 and Π = {t0 = 0 < t1 < . . . < tn = 1} a partition
of [0, 1]. Since f is continuous and the rational numbers are dense in R, there exist
0 < q1 < . . . < qn−1 < 1 such that qj is rational and ∣f (tj ) − f (qj )∣ < n−1/p 1/p for every
1 ⩽ j ⩽ n − 1. In particular, Π′ = {q0 = 0 < q1 < . . . < qn = 1} is a rational partition of [0, 1]
such that ∑nj=0 ∣f (tj ) − f (qj )∣p ⩽ .
Some preliminary considerations: If φ ∶ [0, ∞) → R is concave and φ(0) ⩾ 0 then φ(ta) =

φ(ta + (1 − t)0) ⩾ tφ(a) + (1 − t)φ(0) ⩾ tφ(a) for all a ⩾ 0 and t ∈ [0, 1]. Hence
a b
φ(a + b) = φ(a + b) + φ(a + b) ⩽ φ(a) + φ(b)
a+b a+b
for all a, b ⩾ 0, i. e. φ is subadditive. In particular, we have ∣x + y∣p ⩽ (∣x∣ + ∣y∣)p ⩽ ∣x∣p + ∣y∣p
and thus
∣∣x∣p − ∣y∣p ∣ ⩽ ∣x − y∣p for all p ⩽ 1 and x, y ∈ R. (*)
For p > 1, on the other hand, and x, y ∈ R such that ∣x∣ < ∣y∣ we find
∣y∣
∣∣y∣p − ∣x∣p ∣ = ∫ ptp−1 dt ⩽ p ⋅ (∣x∣ ∨ ∣y∣)p−1 ⋅ (∣y∣ − ∣x∣) ⩽ p ⋅ (∣x∣ ∨ ∣y∣)p−1 ⋅ ∣y − x∣
∣x∣
and hence
∣∣y∣p − ∣x∣p ∣ ⩽ p ⋅ (∣x∣ ∨ ∣y∣)p−1 ⋅ ∣y − x∣ for all p > 1 and x, y ∈ R (**)
94
using the symmetry of the inequality.
Let p > 0 and > 0. For every partition Π = {t0 = 0 < t1 < . . . < tn = 1} there exists a
rational partition Π′ = {q0 = 0 < q1 < . . . < qn = 1} such that ∑nj=0 ∣f (tj ) − f (qj )∣1∧p ⩽ and
hence
RRR n RR
p RRR
n
RRR ∣f (t ) − f (t )∣p − ∣f (q ) − (q )∣ RRR
RRR ∑ j j−1 ∑ j f j−1
RRR
RRj=1 j=1
n
⩽ ∑ ∣∣f (tj ) − f (tj−1 )∣p − ∣f (qj ) − f (qj−1 )∣p ∣
j=1
(*) n
⩽ max {1, (p ⋅ 2p−1 ⋅ ∥f ∥p−1
∞ )} ⋅ ∑ ∣f (tj ) − f (qj ) + f (tj−1 ) − f (qj−1 )∣
1∧p
(**)
j=1
n
⩽ C ⋅ ∑ ∣f (tj ) − f (qj )∣1∧p
j=0
⩽C ⋅
with a finite constant C > 0.
In particular, we have VARp (f ; 1) − C ⋅ ⩽ VARQ

p (f ; 1) ⩽ VARp (f ; 1) where
⎧
⎪ ⎫
⎪
⎪ ′ ⎪
VARQ (f ; 1) ∶= sup ⎨ ∑ ∣f (q ) − f (q )∣ p
∶ Π finite, rational partition of [0, 1]⎬
p
⎪
⎪ ′
j j−1
⎪
⎪
⎩ j−1 j
q ,q ∈Π ⎭
and hence the desired result as tends to zero.
Alternative Approach: Note that (ξ1 , . . . , ξn ) ↦ ∑nj=1 ∣f (ξj ) − f (ξj−1 )∣p is a continuous
map since it is the finite sum and composition of continuous maps, and that the rational
numbers are dense in R.
∎∎
Problem 9.4. Solution: Obviously, we have VAR○p (f ; t) ⩽ VARp (f ; t) with

⎧
⎪ ⎫
⎪
⎪n ⎪
VAR○p (f ; t) ∶= sup ⎨ ∑ ∣f (sj ) − f (sj−1 )∣p ∶ n ∈ N and 0 < s0 < s1 < . . . < sn < t⎬
⎪
⎪ ⎪
⎪
⎩j=1 ⎭
because there are less (non-negative) summands in the definition of VAR○p (f ; t).
Let > 0 and Π = {t0 = 0 < t1 < . . . < tn = t} a partition of [0, t]. Set sj = tj for 1 ⩽ j ⩽ n − 1
and note that ξ ↦ ∣f (ξ0 ) − f (ξ)∣p is a continuous map for every ξ0 ∈ [0, t] since it is the
composition of continuous maps. Hence we can pick s0 ∈ (t0 , t1 ) and sn ∈ (tn−1 , tn ) with
ε
∣∣f (s1 ) − f (t0 )∣p − ∣f (s1 ) − f (s0 )∣p ∣ <
2
ε
∣∣f (tn ) − f (tn−1 )∣ − ∣f (sn ) − f (tn−1 )∣ ∣ <
p p
2
and so that 0 < s0 < s1 < . . . < sn < t. This implies
n n−1
∑ ∣f (tj ) − f (tj−1 )∣ = ∣f (s1 ) − f (t0 )∣ + ∑ ∣f (sj ) − f (sj−1 )∣ + ∣f (tn ) − f (sn−1 )∣
p p p p
j=1 j=2
95
ε n ε
⩽ + ∑ ∣f (sj ) − f (sj−1 )∣p +
2 j=1 2
⩽ + VAR○p (f ; t)
and thus VARp (f ; t) ⩽ + VAR○p (f ; t) since the partition Π = {t0 = 0 < t1 < . . . < tn = t} was
arbitrarily chosen. Consequently, VARp (f ; t) ⩽ VAR○p (f ; t) as tends to zero, as required.
The same argument shows that varp (f ; t) does not change its value (if it exists).
∎∎
(a) Use B(t) − B(s) ∼ N(0, ∣t − s∣) to find

n
2
E Yn = ∑ E (B ( nk ) − B ( k−1
n
))
k=1
n
= ∑ V (B ( nk ) − B ( k−1
n
))
k=1
n
= ∑ ( nk − k−1
n
)
k=1
n
=∑ 1
n
k=1
=1
and the independence of increments to get

n
2
V Yn = ∑ V (B ( nk ) − B ( k−1
n
))
k=1
n
4 2 2
= ∑ E (B ( nk ) − B ( k−1
n
)) − (E (B ( nk ) − B ( k−1
n
)) )
k=1
n
k−1 2 k−1 2
= ∑ 3 ⋅ ( nk − n
) − ( nk − n
)
k=1
n
=2⋅ ∑ 1
n2
k=1
=2⋅ 1
n
where we also used that E(X 4 ) = 3 ⋅ σ 4 for X ∼ N(0, σ 2 ).
(b) Note that the increments B( nk ) − B( k−1

n ) ∼ N(0, n ) are iid random variables. By a
1
2
standard result the sum of squares n ∑nk=1 (B( nk ) − B( k−1
n )) has a χn -distribution,
2
i. e. its density is given by

1
2−n/2 s 2 −1 e− 2 1[0,∞) (s).
n s
Γ( 2 )
n
and we get
n
−n/2 n
(ns) 2 −1 e− 2 1[0,∞) (s).
2 n ns
∑ (B( nk ) − B( k−1
n )) ∼ 2
k=1 Γ( 2 )
n
96
Here is the calculation: (in case you do not know this standard result...): If X ∼
N(0, 1) and x > 0, we have
√
√ 1 x t2
P(X ⩽ x) = P(X ⩽ x) = √ ∫ √ exp (− ) dt
2
2π − x 2
√
2 x t2
=√ ∫ exp (− ) dt
2π 0 2
1 x s
= √ ∫ exp (− ) ⋅ s−1/2 ds
2π 0 2
using the change of variable s = t2 . Hence, X 2 has density
1 s
fX 2 (s) = 1(0,∞) (s) ⋅ √ ⋅ exp (− ) ⋅ s−1/2 .
2π 2
Let X1 , X2 , . . . be independent and identically distributed random variables with
X1 ∼ N(0, 1). We want to prove by induction that for n ⩾ 1
s
fX 2 +...+X 2 (s) = Cn ⋅ 1(0,∞) (s) ⋅ exp (− ) ⋅ sn/2−1
1 n 2
with some normalizing constants Cn > 0. Assume that this is true for 1, . . . , n. Since
2
Xn+1 is independent of X12 + . . . + Xn2 and distributed like X12 , we know that the
density of the sum is a convolution. This leads to
∞
fX 2 +...+X 2 (s) = ∫ fX 2 +...+X 2 (t) ⋅ fX 2 (s − t) dt
1 n+1 −∞ 1 n n+1
st s−t
= Cn ⋅ C1 ⋅ ∫ exp (− ) ⋅ tn/2−1 ⋅ exp (− ) ⋅ (s − t)−1/2 dt
0 2 2
s s
= Cn ⋅ C1 ⋅ exp (− ) ⋅ ∫ tn/2−1 ⋅ (s − t)−1/2 mdt
2 0
s s t n/2−1 t −1/2
= Cn ⋅ C1 ⋅ exp (− ) ⋅ sn/2−1 ⋅ s−1/2 ⋅ ∫ ( ) ⋅ (1 − ) dt
2 0 s s
s 1
= Cn ⋅ C1 ⋅ exp (− ) ⋅ s(n+1)/2−1 ⋅ ∫ xn/2−1 ⋅ (1 − x)−1/2 dx
2 0
s (n+1)/2−1
= Cn+1 ⋅ exp (− ) ⋅ s
2
using the change of variable x = t/s. Since probability distribution functions integrate
to one, we find
∞ s ∞
1 = Cn ⋅ ∫ exp (− ) ⋅ sn/2−1 ds = Cn ⋅ 2n/2 ∫ exp (−t) ⋅ tn/2−1 dt
0 2 0
= Cn ⋅ 2n/2 ⋅ Γ(n/2)
and thus
−1
fX 2 +...+Xn2 (s) = (2n/2 ⋅ Γ(n/2)) ⋅ 1(0,∞) (s) ⋅ e−s/2 ⋅ sn/2−1
1
which is usually called chi-squared or χ2 -distribution with n degrees of freedom. Now,

remember that B ( nk ) − B ( k−1
n
) ∼ N(0, 1/n) ∼ n−1/2 ⋅ Xk for 1 ⩽ k ⩽ n. Hence
fYn (s) = n ⋅ fX 2 +...+Xn2 (n ⋅ s)

1
−1
= n ⋅ (2 n/2
⋅ Γ(n/2)) ⋅ 1(0,∞) (s) ⋅ e−n⋅s/2 ⋅ (ns)n/2−1 .
97
(c) For X ∈ N(0, 1) and ξ < 1/2, we find

∞ 2 ∞
E(eξ⋅X ) = (2 ⋅ π)−1/2 ∫ eξ⋅x e−x
2 2 2 /2 −1/2⋅(1−2ξ)⋅x2
dx = √ ∫ e dx
−∞ 2⋅π 0
2 ∞
= (1 − 2ξ)−1/2 √ ∫ e
−y 2 /2
dy
2⋅π 0
= (1 − 2ξ)−1/2
using the change of variable x2 = (1 − 2ξ)y 2 . Since the moment generating function
ξ ↦ (1−2ξ)−1/2 has a unique analytic extension to an open strip around the imaginary
axis, the characteristic function is of the form
E(ei⋅ξ⋅X ) = (1 − 2iξ)−1/2 .
2
Using the independence and B ( nk ) − B ( k−1

n
) ∼ N(0, 1/n), we obtain
n n
E(ei⋅ξ⋅Yn ) = ∏ E(ei⋅ξ⋅(Bk/n −B(k−1)/n ) ) = ∏ E(ei⋅(ξ/n)⋅X ) = (1 − 2i(ξ/n))−n/2
2 2
k=1 k=1
and hence
2iξ n −1/2 −1/2

lim φn (ξ) = lim (1 − 2i(ξ/n))−n/2 = ( lim (1 − ) ) = (e−2iξ ) = eiξ .
n→∞ n→∞ n→∞ n
(d) We have shown in a) that E ((Yn − 1)2 ) = V(Yn ) = 2/n which tends to zero as n → ∞.
∎∎
(a)
√ ∞ y −y2 /2 ∞
1 ∞ 1
e−y
2 /2
dy = ⋅ [ − e−y /2 ] = ⋅ e−x /2
2 2
2π ⋅ P(Z > x) = ∫ dy > ∫
⋅e
x x x x x x
1 e−x /2
2
Ô⇒ P(Z > x) < √

2π x
On the other hand
√ ∞
e−y
2 /2
2π ⋅ P(Z > x) = ∫ dy
x
∞
x2 −y2 /2
<∫ ⋅e dy
x y2
∞ ∞
1
= x2 ⋅ ([− ⋅ e−y /2 ] − ∫ e−y /2 dy)
2 2
y x x
1 ∞ √
= x2 ⋅ ([− ⋅ e−y /2 ] − 2π ⋅ P(Z > x))
2
y x
√
Ô⇒ (1 + x ) ⋅ 2π ⋅ P(Z > x) ⩾ x ⋅ e−x /2
2 2
1 xe−x /2
2
Ô⇒ P(Z > x) > √

2π x2 + 1
98
(b) Using the independence of Ak,n for 1 ⩽ k ⩽ 2n , we find
2n 2n
P ( lim ⋃ Ak,n ) = 1 − P (lim inf ⋂ Ack,n )
n→∞ k=1 n→∞ k=1
2n
⩾ 1 − lim inf P ( ⋂ Ack,n )
n→∞ k=1
2n
= 1 − lim inf ∏ P(Ack,n )
n→∞
k=1
n
and hence it suffices to prove lim inf n→∞ ∏2k=1 P(Ack,n ) = 0.
Since 1 − x ⩽ e−x for x ⩾ 0, we obtain

2n 2n
⩽ e−2
n ⋅P(A
1,n )
∏ P(Ak,n ) = (1 − P(A1,n ))
c
k=1
and a) implies
√ √
2n ⋅ P(A1,n ) = 2n ⋅ P( 2−n ⋅ ∣Z∣ > c n2−n )
√
= 2n+1 ⋅ P(Z > c n)
√
2n+1 c n
⋅ e−c n/2 .
2
⩾√ ⋅ 2
2π c n + 1
Now, (c2 n)/(c2 n + 1) → 1 as n → ∞ and thus there exists some n0 ∈ N such that
√
c2 n 1 c n 1
⩾ ⇐⇒ 2 ⩾ √
c n+1 2
2 c n + 1 2c n
for all n ⩾ n0 . Therefore, we have

2n 1 1 1
2n ⋅ P(A1,n ) ⩾ √ ⋅ √ ⋅ e−c n/2 = √ ⋅ √ ⋅ e(log(2)−c /2)n
2 2
2π c n 2πc n
√
for n ⩾ n0 . Since ln(2)−c2 /2 > 0 if, and only if, c < 2 log(2), we have 2n ⋅P(A1,n ) → ∞
n √
and thus lim inf n→∞ ∏2k=1 P(AC k,n ) = 0 if c < 2 log(2).
√
(c) With c < 2 log(2) we deduce
2n
1 = P (lim sup ⋃ Ak,n )
n→∞ k=1
= P ({ω ∈ Ω ∶ for infinitely many n ∈ N ∃k ∈ {1, . . . , 2n }

√
such that ∣B(k2−n )(ω) − B((k − 1)2−n )(ω)∣ > c n 2−n })
= P ({ω ∈ Ω ∶ for infinitely many n ∈ N ∃k ∈ {1, . . . , 2n }

∣B(k2−n )(ω) − B((k − 1)2−n )(ω)∣ √
such that √ > c n})
2−n
⩽ P ({ω ∈ Ω ∶ t ↦ Bt (ω) is NOT 1/2-Hölder continuous}).
∎∎
99
Problem 9.7. Solution: From Problem 9.5 we know that

2 −1)
) = e−λ E(eλX ) = e−λ (1 − 2λ)−1/2
2
Φ(λ) = E(eλ(X for all 0 < λ < 1/2.
Using (a − b)2 ⩽ 2 (a2 + b2 ), we get

2 −1) 2 2
∣(X 2 − 1)2 eλ(X ∣ ⩽ ∣X 2 − 1∣2 ⋅ eλX ⩽ 2(X 4 + 1) ⋅ eλ0 X .
Since λ < λ0 < 1/2 there is some > 0 such that λ < λ0 < λ0 + < 1/2. Thus,
2 −1)
∣ ⩽ 2(X 4 + 1)e−X ⋅ e(λ0 +)X .
2 2
∣(X 2 − 1)2 eλ(X
It is straightforward to see that
2(X 4 + 1)e−X ⩽ C = C(λ0 ) < ∞,

2
and the claim follows.
∎∎
Problem 9.8. Solution: We follow the hint. Note that

⎧
⎪ ⎫
⎪ iη
⎪e , x ∈ F⎪
⎪
⎪
e iη1F
=⎨ ⎬ = eiη 1F + 1F c .
⎪
⎪
⎪ x ∉ F⎪
⎪
⎪
⎩ 1, ⎭
Since F, F c ∈ F, we get ∀ξ, η ∈ R
E (eiξX eiη1F ) = E (eiξX eiη 1F ) + E (eiξX 1F c )
= E (eiξX ) E (eiη 1F ) + E (eiξX ) E (1F c )
= E (eiξX ) E (eiη1F ).
This is, however, M. Kac’s characterization of independence and we conclude that X á

1F , hence, X á F for all sets F ∈ F.
The converse is obvious.
∎∎
100
10 Regularity of Brownian paths
(a) Note that for t, h ⩾ 0 and any integer k = 0, 1, 2, . . .
(λh)k −λh
P(Nt+h − Nt = k) = P(Nh = k) = e .
k!
This shows that we have for any α > 0
∞
(λh)k −λh
E (∣Nt+h − Nt ∣α ) = ∑ k α e
k=0 k!
∞
(λh)k −λh
= λh e−λh + ∑ k α e
k=2 k!
∞
(λh)k−1 −λh
= λh e−λh + λh ∑ k α e
k=2 k!
−λh
= λh e + o(h)
and, thus,
E (∣Nt+h − Nt ∣α )
lim =λ
h→0 h
which means that (10.1) cannot hold for any α > 0 and β > 0.
(b) Part a) shows also E (∣Nt+h − Nt ∣α ) ⩽ c h, i. e. condition (10.1) holds for α > 0 and
β = 0.
The fact that β = 0 is needed for the convergence of the dyadic series (with the power
γ < β/α) in the proof of Theorem 10.1.
(c) We have
∞
tk −t ∞ tk −t ∞
tk−1 ∞ j
t
E(Nt ) = ∑ k e = ∑k e =t∑ e−t = t ∑ e−t = t
k=0 k! k=1 k! k=1 (k − 1)! j=0 j!
∞
tk −t ∞ 2 tk −t ∞
tk−1
E(Nt2 ) = ∑ k 2 e = ∑k e =t∑k e−t
k=0 k! k=1 k! k=1 (k − 1)!
∞ ∞
tk−1 tk−1
= t ∑ (k − 1) e−t + t ∑ e−t
k=1 (k − 1)! k=1 (k − 1)!
∞ ∞
tk−2 tk−1
=t ∑
2
e−t + t ∑ e−t = t2 + t
k=2 (k − 2)! k=1 (k − 1)!
and this shows that
E(Nt − t) = E Nt − t = 0
101
E ((Nt − t)2 ) = E(Nt2 ) − 2t E Nt + t2 = t
and, finally, if s ⩽ t
Cov ((Nt − t)(Ns − s)) = E ((Nt − t)(Ns − s))

= E ((Nt − Ns − t + s)(Ns − s)) + E ((Ns − s)2 )
= E ((Nt − Ns − t + s)) E ((Ns − s)) + s
=s=s∧t
where we used the independence of Nt − Ns á Ns .
Alternative Solution: One can show, as for a Brownian motion (Example 5.2 a)), that
Nt is a martingale for the canonical filtration FtN = σ(Ns ∶ s ⩽ t). The proof only
uses stationary and independent increments. Thus, by the tower property, pull out
and the martingale property,
E ((Nt − t)(Ns − s)) = E ( E ((Nt − t)(Ns − s) ∣ FsN ))

= E ((Ns − s) E ((Nt − t) ∣ FsN ))
= E ((Ns − s)2 )
= s = s ∧ t.
∎∎

n n
max ∣xj ∣p ⩽ max (∣x1 ∣p + ⋯ + ∣xn ∣p ) = ∑ ∣xj ∣p ⩽ ∑ max ∣xk ∣p = n max ∣xk ∣p .
1⩽j⩽n 1⩽j⩽n j=1 j=1 1⩽k⩽n 1⩽k⩽n
p
Since max1⩽j⩽n ∣xj ∣p = ( max1⩽j⩽n ∣xj ∣) the claim follows (actually with n1/p which is
smaller than n....)
∎∎
Problem 10.3. Solution: Let α ∈ (0, 1). Since
∣x + y∣α ⩽ (∣x∣ + ∣y∣)α
it is enough to show that

(∣x∣ + ∣y∣)α ⩽ ∣x∣α + ∣y∣α
and, without loss of generality
(s + t)α ⩽ sα + tα ∀s, t > 0.
This follows from
sα + tα = s ⋅ sα−1 + t ⋅ tα−1 ⩾ s ⋅ (s + t)α−1 + t ⋅ (s + t)α−1 = (s + t)(s + t)α−1 = (s + t)α .
102
Since the expectation is linear, this proves that
E(∣X + Y ∣α ) ⩽ E(∣X∣α ) + E(∣Y ∣α ).
In the proof of Theorem 10.1 (page 154, line 1 from above and onwards) we get:
This entails for α ∈ (0, 1) because of the subadditivity of x ↦ ∣x∣α

α
∣ξ(x) − ξ(y)∣ ∣ξ(x) − ξ(y)∣α
( sup ) = sup sup
x,y∈D, x≠y ∣x − y∣γ m⩾0 x,y∈D 2−(m+1)γα
2−m−1 ⩽∣x−y∣<2−m
⩽ sup (2α ⋅ 2(m+1)γα ∑ σjα )

m⩾0 j⩾m
(1+γ)α
=2 sup ∑ 2mγα σjα
m⩾0 j⩾m
∞
⩽ 2(1+γ)α ∑ 2jγα σjα .
j=0
For α ∈ (0, 1) and αγ < β we get

α ∞
∣ξ(x) − ξ(y)∣ (1+γ)α
E [( sup ) ] ⩽ 2 ∑ 2jγα E [σjα ]
x≠y, x,y∈D ∣x − y∣γ
j=0
∞
⩽ c 2(1+γ)α ∑ 2jγα 3n 2−jβ
j=0
∞
= c 2(1+γ)α 3n ∑ 2j(γα−β) < ∞.
j=0
The rest of the proof continues literally as on page 154, line 10 onwards.
Alternative Solution: use the subadditivity of Z ↦ E(∣Z∣α ) directly in the second part of
the calculation, replacing ∥Z∥Lα by E(∣Z∣α ).
∎∎
Problem 10.4. Solution: We show the following
Theorem. Let (Bt )t⩾0 be a BM1 . Then t ↦ Bt (ω) is for almost all ω ∈ Ω nowhere Hölder
continuous of any order α > 1/2.
Proof. Set for every n ⩾ 1
An ∶= An,α = {ω ∈ Ω ∶ B(⋅, ω) is in [0, n] nowhere Hölder continuous of order α > 21 }.
It is not clear if the set An,α is measurable. We will show that Ω ∖ An,α ⊂ Nn,α for a
measurable null set Nn,α .
Assume that the function f is α-Hölder continuous of order α at the point t0 ∈ [0, n].
Then
∃ δ > 0 ∃ L > 0 ∀ t ∈ B(t0 , δ) ∶ ∣f (t) − f (t0 )∣ ⩽ L ∣t − t0 ∣α .
103
Since [0, n] is compact, we can use a covering argument to get a uniform Hölder constant.
Consider for sufficiently large values of k ⩾ 1 the grid { kj ∶ j = 1, . . . , nk}. Then there
exists a smallest index j = j(k) such that for ν ⩾ 3 and, actually, 1 − να + ν/2 < 0
j j j+ν
t0 ⩽ and ,..., ∈ B(t0 , δ).
k k k
For i = j + 1, j + 2, . . . , j + ν we get therefore
∣f ( ki ) − f ( i−1
k
)∣ ⩽ ∣f ( ki ) − f (t0 )∣ + ∣f (t0 ) − f ( i−1
k
)∣
α α
⩽ L(∣ ki − t0 ∣ + ∣ i−1
k − t0
∣ )
(ν+1)α να 2L(ν+1)α
⩽ L( kα + kα
) = kα .
If f is a Brownian path, this implies that for the sets

∞ kn j+ν
2L(ν+1)α
L,ν,α
Cm ∶= ⋂ ⋃ ⋂ {∣B( ki ) − B( i−1
k
)∣ ⩽ kα
}
k=m j=1 i=j+1
we have
∞ ∞
Ω ∖ An,α ⊂ ⋃ ⋃ Cm
L,ν,α
.
L=1 m=1
Our assertion follows if we can show that P(Cm L,ν,α

) = 0 for all m, L ⩾ 1 and all rational
α > 1/2. If k ⩾ m,
kn j+ν
2L(ν+1)α
L,ν,α
P(Cm ) ⩽ P ( ⋃ ⋂ {∣B( ki ) − B( i−1
k
)∣ ⩽ kα
})
j=1 i=j+1
kn j+ν
2L(ν+1)α
⩽ ∑ P ( ⋂ {∣B( ki ) − B( i−1
k
)∣ ⩽ kα
})
j=1 i=j+1
kn
2L(ν+1)α ν
= ∑ P ({∣B( ki ) − B( i−1 )∣ ⩽ })
(B1)
k kα
j=1
2L(ν+1)α ν
= kn P ({∣B( k1 )∣ ⩽ })
(B2)
kα
c ν 1−να+ν/2<0
⩽ kn ( α−1/2
) = cν n k 1−να+ν/2 ÐÐÐÐÐÐÐÐÐÐ→ 0.
k k→∞
−1/2
For the last estimate we use B( k1 ) ∼k B(1), cf. 2.16, and therefore
√
1 √
−y 2 /2
√ x k
P (∣B( k1 )∣ ⩽ x) = P (∣B(1)∣ ⩽ x k) = √ ∫ e dy ⩽ c x k.
2π √ ²
−x k ⩽1
This proves that a Brownian path is almost surely nowhere not Hölder continuous of a
fixed order α > 1/2. Call the set where this holds Ωα . Then Ω0 ∶= ⋂Q∋α>1/2 Ωα is a set
with P(Ω0 ) = 1 and for all ω ∈ Ω0 we know that BM is nowhere Hölder continuous of any
order α > 1/2.
The last conclusion uses the following simple remark. Let 0 < α < q < ∞. Then we have
for f ∶ [0, n] → R and x, y ∈ [0, n] with ∣x − y∣ < 1 that
∣f (x) − f (y)∣ ⩽ L∣x − y∣q ⩽ L∣x − y∣α .
Thus q-Hölder continuity implies α-Hölder continuity.
104
∎∎
Problem 10.5. Solution: Fix > 0, fix a set Ω0 ⊂ Ω with P(Ω0 ) = 1 and h0 = h0 (2, ω) such
that (10.6) holds for all ω ∈ Ω0 , i. e. for all h ⩽ h0 we have
√
sup ∣B(t + h, ω) − B(t, ω)∣ ⩽ 2 2h log h1 .
0⩽t⩽1−h
Pick a partition Π = {t0 = 0 < t1 < . . . < tn } of [0, 1] with mesh size h = maxj (tj − tj−1 ) ⩽ h0
and assume that h0 /2 ⩽ h ⩽ h0 . Then we get
n n 1+
∑ ∣B(tj , ω) − B(tj−1 , ω)∣ ⩽ 22+2 ⋅ 21+ ∑ ((tj − tj−1 ) log tj −t1j−1 )
2+2
j=1 j=1
n
⩽ c ∑ (tj − tj−1 ) = c .
j=1
This shows that

n
sup ∑ ∣B(tj , ω) − B(tj−1 , ω)∣2+2 ⩽ c .
∣Π∣⩽h0 j=1
Since we have ∣x − y∣p ⩽ 2p−1 (∣x − z∣p + ∣z − y∣p ) and since we can refine any partition Π of
[0, 1] in finitely many steps to a partition of mesh < h0 , we get
n
VAR2+2 (B; 1) = sup ∑ ∣B(tj , ω) − B(tj−1 , ω)∣2+2 < ∞
Π⊂[0,1] j=1
for all ω ∈ Ω0 .
∎∎
105
11 Brownian motion as a random fractal
Problem 11.1. Solution: The idea is to show that Hδs for every δ > 0 is an outer measure.
This solves the problem, since these properties are retained by taking the supremum over
δ > 0.
It is easy to see that Ej ∶= ∅ for j ∈ N is a δ-cover of E = ∅ and hence Hδs (∅) = 0.

Moreover, if (Ej )j∈N is a δ-cover of F ⊂ Rd and E ⊂ F , then it is also a δ-cover of E and
therefore Hδs (E) ⩽ Hδs (F ).
Let > 0 and suppose that (E k )k∈N is a sequence of subsets of Rd . Due to the definition
of Hδs , there exists a δ-cover (Ejk )j∈N for every k ∈ N such that
−k
∑ ∣Ej ∣ ⩽ Hδ (E ) + 2
ks s k
j∈N
holds. Since the double sequence (Ejk )j,k∈N is obviously a δ-cover of ⋃k∈N E k , we find that
Hδs ( ⋃ E k ) ⩽ ∑ ∑ ∣Ejk ∣s ⩽ ∑ (Hδs (E k ) + 2−k ) ⩽ ∑ Hδs (E k ) + .

k∈N k∈N j∈N k∈N k∈N
holds. This implies the σ-subadditivity as → 0.
∎∎
Problem 11.2. Solution: Let U be open. Then U contains an open ball B ⊂ U and B contains
a cube Q ⊂ B ⊂ U . On the other hand, since U is bounded, it is contained in a large cube
Q′ ⊃ U . Since Hausdorff measure is monotone, we have Hd (Q) ⩽ Hd (U ) ⩽ Hd (Q′ ) and it
is, thus, enough to show the claim for cubes.
The following argument is easily adapted to a general cube. Assume that Q = [0, 1]d
and cover Q by nd non-overlapping cubes which are shifted copies of [0, 1/n]d . Clearly, if
n > 1/δ,
nd
d √ d √
Hδd (Q) ⩽ ∑ ∣[0, 1/n]d ∣ = nd ( dn−1 ) = ( d)d .
j=1
√
This shows that Hd (Q) ⩽ ( d)d < ∞.
For the lower bound we take any δ-cover (Ej )j⩾1 of Q. For each j there is a closed cube
Cj such that Ej ⊂ Cj and the lengths of the edges of Cj are less or equal to 2∣Ej ∣ ⩽ 2δ. If
λd is Lebesgue measure, we get
1 = λd (Q) ⩽ λd (⋃j Ej ) ⩽ λd (⋃j Cj ) ⩽ ∑ λd (Cj ) = ∑ (2∣Ej ∣)d .

j j
107
This gives
∞
−d −d
∑ ∣Ej ∣ ⩾ 2 > 0 Ô⇒ H (Q) ⩾ 2 .
d d
j=1
∎∎
Problem 11.3. Solution: It is enough to show the following two assertions: Let 0 ⩽ α < β < ∞
and E ⊂ R . Then
d
a) Hα (E) < ∞ Ô⇒ Hβ (E) = 0.
b) Hβ (E) > 0 Ô⇒ Hα (E) = ∞.
Claim a) is just Lemma 11.4. Part b) is just the contraposition of a).
∎∎
Problem 11.4. Solution: Since Ej ⊂ E, we have dim Ej ⩽ dim E and supj⩾1 dim Ej ⩽ dim E.
Conversely, if α < dim E, then Hα (E) = ∞ and by the σ-subadditivity of the outer measure
Hα , we get Hα (Ej0 ) > 0 for at least one index j0 . Thus, α ⩽ dim Ej0 ⩽ supj⩾1 dim Ej . This
proves dim E ⩽ supj⩾1 dim Ej . (Indeed, if we had dim E > supj⩾1 dim Ej , we could find
some λ such that dim E > λ > supj⩾1 dim Ej contradicting our previous calculation.)
∎∎
Problem 11.5. Solution: It is possible to show that dim(E × F ) ⩾ dim(E) + dim(F ) holds for
arbitrary E ⊂ Rd and F ⊂ Rn , cf. [6, Theorem 5.12]. Unfortunately, the opposite direction
only holds under certain restriction on the sets E and F , cf. for example [7, Corollary 7.4].
In fact, one can show that there exist Borel sets E, F ⊂ R with dim(E) = dim(F ) = 0 and
dim(E × F ) ⩾ 1, cf. [6, Theorem 5.11].
We are going to prove the other direction (that does not hold in general) for this special
case: Let t > dim(E) and δ > 0. According to the Definition 11.5, there exists a δ-cover
√
(Ej )j∈N of E ⊂ Rd with ∑j∈N ∣Ej ∣t ⩽ δ. Let m ∈ N so that n/m ⩽ δ, and (Fk )k be a
disjoint tessellation of [0, 1)n by mn -many cubes with side-length 1/m. Now, (Ej × Fk )j,k
is a δ 2 -cover of E × [0, 1)n and hence
mn mn
2 (E
Hδt+n × [0, 1) ) ⩽ ∑ ∑ ∣Ej × Fk ∣
n t+n
⩽ ∑ ∑ ∣Ej ∣t nn/2 m−n ⩽ nn/2 δ
k=1 j∈N k=1 j∈N
holds. In particular, Ht+n (E × [0, 1)n ) = 0 as δ → 0 and thus dim(E × [0, 1)n ) ⩽ t + n. Since
Rn can be represented as countable union of cubes with unit side-length, Problem 11.4
tells us that we also have dim(E × Rn ) = dim(E × [0, 1)n ) ⩽ t + n. This proves that
dim(E × Rn ) ⩽ dim(E) + n, as required.
∎∎
Problem 11.6. Solution: Remark 11.6.3 says that dim f (E) ⩽ dim E holds for a Lipschitz
map f ∶ Rd → Rn . Therefore, we also have dim E = dim f −1 (f (E)) ⩽ dim f (E) for a
bi-Lipschitz map f and hence the desired result.
108
Moreover, Remark 11.6.3 tells us that dim f (E) ⩽ γ −1 dim E holds for a Hölder continuous
map f ∶ Rd → Rn with index γ ∈ (0, 1]. Note that this inequality can be strict, e.g. take
f ≡ 0 and any E ⊂ Rd with dim E > 0.
Note that there is no bi-Lipschitz f ∶ Rd → Rn that is also Hölder continuous with index
γ ∈ (0, 1): Suppose f had these properties, then there would exist a constant C > 0 such
that
∣x − y∣ = ∣f (f −1 (x)) − f (f −1 (y))∣ ⩽ C∣x − y∣γ
holds for all x, y ∈ Rd . This leads to a contradiction to the boundedness of C > 0. Hence,
there is no bi-Lipschitz map that is also Hölder continuous with index γ ∈ (0, 1).
∎∎
Problem 11.7. Solution: Let C0 ∶= [0, 1]. It is easy to see that Cn = f1 (Cn−1 ) ∪ f2 (Cn−1 ) for
n ∈ N and C ∶= ⋂n∈N Cn models the recursive definition of Cantor’s discontinuum in the
description of the problem. Now, note that
⎛∞ ⎞ ∞ ∞
f1 ∑ tj 3−j = ∑ tj 3−(j+1) = 0 ⋅ 3−1 + ∑ tj−1 3−j
⎝j=1 ⎠ j=1 j=2
⎛∞ ⎞ ∞ ∞
f2 ∑ tj 3−j = ∑ tj 3−(j+1) + 2/3 = 2 ⋅ 3−1 + ∑ tj−1 3−j
⎝j=1 ⎠ j=1 j=2
holds for sequences (tj )j∈N with tj ∈ {0, 1, 2} and that

⎧
⎪ ⎫
⎪
⎪∞ ⎪
C0 = ⎨ ∑ tj 3−j ∶ tj ∈ {0, 1, 2} for j ∈ N⎬
⎪
⎪ ⎪
⎪
⎩j=1 ⎭
reflects the triadic representation of real numbers. This representation implies that
⎧
⎪ ⎫
⎪
⎪∞ −j ⎪
Cn = ⎨ ∑ tj 3 ∶ tj ∈ {0, 2} for j ⩽ n and tj ∈ {0, 1, 2} for j > n⎬
⎪
⎪ ⎪
⎪
⎩j=1 ⎭
holds for every n ∈ N using mathematical induction. Hence,
⎧
⎪ ⎫
⎪
⎪∞ ⎪
C = ⎨ ∑ tj 3−j ∶ tj ∈ {0, 2} for j ∈ N⎬
⎪
⎪ ⎪
⎪
⎩j=1 ⎭
and therefore C = f1 (C) ∪ f2 (C). The results from Remark 11.6.5 solve the rest of the
problem.
∎∎
Problem 11.8. Solution: Denote by σd = 2π d/2 /Γ(d/2) the surface volume of the (d − 1)-
dimensional unit sphere in Rd . Using polar coordinates, we find
E (∣B1 ∣−λ ) = ∫ ∣x∣−λ P(B1 ∈ dx) = (2π)− 2 ∫ ∣x∣−λ e− 2 ∣x∣ dx

d 1 2
Rd Rd
∞
= σd (2π)− 2 ∫ rd−λ−1 e− 2 r dr
d 1 2
109
∞
= σd (2π)− 2 ∫ e−u du
d d−λ−2
(2u) 2
0
= σd (2π)− 2 2
d d−λ−2
2 Γ ( d−λ
2
)
Γ ( d−λ )
= λ
2
.
2 2 Γ ( d2 )
∎∎
Problem 11.9. Solution: Following the hint we have

1 1
B −1 (A) = W −1 (A × R) ⩽ dim(A × R) ⩽ (1 + dim A),
2 2
where we used the result of Problem 11.5.
∎∎
Problem 11.10. Solution: (F. Hausdorff) We show that a perfect set contains a Cantor-type
set. Since Cantor sets are uncountable, we are done.
Pick a1 , a2 ∈ F and disjoint closed balls Fj , j = 1, 2 with centre aj . Now take open balls such
that Uj ⊂ Fj . Since Ūj ∩ F , j = 1, 2, are again perfect sets, we can repeat this construction,
i. e. pick aj1 , aj2 ∈ Uj ∩ F and disjoint closed balls Fjk ⊂ Uj with centre ajk and open balls
Ujk ⊂ Fjk , k = 1, 2. Each of the four sets A ∩ Ūjk , j, k = 1, 2, is perfect. Again we find
points ajk1 , ajk2 ∈ Ujk etc. Without loss of generality we can arrange things such that the
diameters of the balls Fj , Fjk , Fjkl , . . . are smaller than 1, 12 , 13 , . . .. This construction yields
a discontinuum set D ⊂ F : Any x ∈ D which is contained in Fj , Fjk , Fjkl , . . . is the limit of
the centres aj , ajk , ajkl , . . . It is now obvious how to make a correspondence between the
points ajkl... ∈ F and the Cantor ternary set.
∎∎

(a) We have
∞ ds ∞
y e−y /(2(t−s)) dy
2
P(ξt ∈ ds) = ∫ P(ξt ∈ ds, ∣Bt ∣ ∈ dy) = √ ∫
0 π s(t − s)3 0
ds ∞ y
e−y /(2(t−s)) dy
2
= √ ∫
π s(t − s) 0 (t − s)
ds ∞ ds
[−e−y /(2(t−s)) ]
2
= √ = √ .
π s(t − s) y=0 π s(t − s)
Using sin2 φ + cos2 φ = 1 we see

s du
P(ξt < s) = ∫ √
0 π u(t − u)
s/t dv
= √
u=tv
∫
0 π v(1 − v)
√ √
v=sin2 φ
arcsin s/t 2 sin φ cos φ 2 s
= ∫ dφ = arcsin .
0 π sin φ cos φ π t
110
(b) From part a) we know that the density of ξt is
1
fξt (s) = √ , 0 < s < t.
π s(t − s)
The joint density of (∣Bt ∣, ξt ) is, by Theorem 11.25
y
e−y /(2(t−s)) ,
2
f(∣Bt ∣, ξt ) (y, s) = √ 0 < s < t, y > 0.
π s(t − s)3
Using standard formulae for the conditional density, we find
f(∣Bt ∣, ξt ) (y, s) f(∣Bt ∣, ξt ) (y, s) y −y2 /(2(t−s))

f (y ∣ s) = ∞ = = e .
∣Bt ∣∣ξt fξt (s) t−s
∫0 f(∣Bt ∣, ξt ) (y, s) dy
(c) Write
1
e−(x−y) /(2t)
2
pt (x, y) = fBt (x − y) = √
2πt
for the law of Bt and set
∣x − y∣ −(x−y)2 /(2t) y>x ∂

gt (x, y) ∶= √ e = pt (x, y).
2πt3 ∂x
As a function of t, this is the density of τx−y , see (6.13). Then the identity reads
(after cancelling the factor 2)
t t
pt (0, y) = ∫ ps (0, 0)gt−s (0, y) ds = ∫ gt−s (0, y)ps (y, y) ds.
0 0
The first identity is a “last exit decomposition” of the density pt (0, y) while the last
identity is a “first entrance decomposition”.
∎∎
Problem 11.12. Solution: Following the hint we find

√
∂ 2 s 2 1 1 1 1 1 1 1 1 1
(1 − arccos )= √ √s = √ √ √ = √ √ .
∂s π u π 1− u 2 u u π 1−
s s
u
s u π s u−s
and so
√
∂ ∂ 2 s ∂ 1 1 1 −1
(1 − arccos )= ( √ √ )= √ .
∂u ∂s π u ∂u π s u − s 2π s(u − s)3
This is (up to the minus sign) just the density from Corollary 11.26. From Lemma 11.23
we know, however, that for s < t < u
√
2 s
P (ξt ⩽ s, ηt ⩾ u) = P (B● has no zero in (s, u)) = 1 − arccos .
π u
This proves the claim.
∎∎
111
Problem 11.13. Solution: Notation: We write fX for the density of the random variable X.
Using Corollary 11.26 we have, with some obvious changes in the integration variables,
E φ(L−t , Lt ) = E φ(t − ξt , ηt − ξt )
= ∬ φ(t − s, u − s) P(ξt ∈ ds, ηt ∈ du)

t ∞ du ds
=∫ ∫ √
φ(t − s, u − s)
s=0 u=t 2π s(u − s)3
t ∞ du dr
= ∫ ∫ φ(r, u − t + r) √
r=0 u=t 2π (t − r)(u − t + r)3
t ∞ dl dr
= ∫ ∫ φ(r, l) √
r=0 l=r 2π (t − r)l3
which shows that the joint density satisfies

1
f(L−t ,Lt ) (r, l) = √ , 0 < r < t, l > r ( ⇐⇒ 0 < r < t ∧ l)
2π (t − r)l3
Integrating out Lt ∈ dl now yields

1
fL−t (r) = √ , 0<r<t
π r(t − r)
(this is just the arc-sine density, rewritten for L−t = t − ξt ).
Integrating out L−t ∈ dr now yields

√
⎧
⎪ t
⎪
⎪
⎪ √ , l ∈ [t, ∞)
1 t∧l dr ⎪
⎪ π l3
fLt (l) = √ ∫ √ = ⎨√ √
2π l3 0 t−r ⎪⎪
⎪ t− t−l
⎪
⎪
⎪ √ , l ∈ (0, t).
⎩ π l3
Using standard formulae for the conditional densities, we get
f(L−t ,Lt ) (r, l) √

1 r
fLt ∣L−t (l ∣ r) = = , 0 < r < t ∧ l.
fL−t (r) 2 l3
From this we get

√ √
∞ 1 r 1√ 1 ∞ r
P(Lt > r + s ∣ L−t = r) = ∫ dl = r [−2l− 2 ] = .
r+s 2 l 3 2 l=r+s r+s
∎∎
112
12 The growth of Brownian paths
√
Problem 12.1. Solution: Fix C > 2 and define An ∶= {Mn > C n log n}. By the reflection
principle we find
√
P(An ) = P (sup Bs > C n log n)
s⩽n
√
= 2 P (Bn > C n log n)
√ √
= 2 P ( n B1 > C n log n)
scaling
√
= 2 P (B1 > C log n)
2
(12.1) 1 2
⩽ √ √ exp (− C2 log n)
2π C log n
2 1 1
=√ √ .
2π C log n n /2
C 2
Since C 2 /2 > 2, the series ∑n P(An ) converges and, by the Borel–Cantelli lemma we see
that
√
∃ΩC ⊂ Ω, P(ΩC ) = 1, ∀ω ∈ ΩC ∃n0 (ω) ∀n ⩾ n0 (ω) ∶ Mn (ω) ⩽ C n log n.
This shows that

Mn
∀ω ∈ ΩC ∶ lim √ ⩽ C.
n→∞ n log n
√
Since every t is in some interval [n − 1, n] and since t ↦ t log t is increasing, we see that
√
Mt Mn Mn n log n
√ ⩽√ =√ √
t log t (n − 1) log(n − 1) n log n (n − 1) log(n − 1)
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
→1 as n→∞
Remark: We can get the exceptional set in a uniform way: On the set Ω0 ∶= ⋂Q∋C>2 ΩC
we have P(Ω0 ) = 1 and
Mn
∀ω ∈ Ω0 ∶ lim √ ⩽ 2.
n→∞ n log n
∎∎
Problem 12.2. Solution: One should assume that ξ > 0. Since y ↦ exp(ξy) is monotone
increasing, we see
1
P (sup(Bs − 12 ξs) > x) = P (esups⩽t (ξBs − 2 ξ
2 s)
> eξx )
s⩽t
113
Doob
⩽ e−xξ E eξBt − 2 ξ t = e−xξ .
1 2
(A.13)
(Remark: we have shown (A.13) only for supD∋s⩽t Msξ where D is a dense subset of [0, ∞).
Since s ↦ Msξ has continuous paths, it is easy to see that supD∋s⩽t Msξ = sups⩽t Msξ almost
surely.)
Usage in step 1o of the Proof of Theorem 12.1: With the notation of the proof we set
√ 1√ n
t = qn and ξ = q −n (1 + ) 2q n log log q n and x = 2q log log q n .
2
Since sups⩽t (Bs − 12 ξs) ⩾ sups⩽t Bs − 21 ξt the above inequality becomes
P (sup Bs > x + 21 ξt) ⩽ e−xξ

s⩽t
and if we plug in t, x, ξ we see

√ √
P (sup Bs > x + 12 ξt) = P (sup Bs > 1
2 2q n log log q n + 21 (1 + ) 2q n log log q n )
s⩽t s⩽q n
√
= P (sup Bs > (1 + 2 ) 2q n log log q n )
s⩽q n
√ √
⩽ exp (− 12 2q n log log q n q −n (1 + ) 2q n log log q n )
= exp (−(1 + ) log log q n )

1
=
(log q n )1+
1 1
= .
(log q) 1+ n 1+
Now we can argue as in the proof of Theorem 12.1.

∎∎
Problem 12.3. Solution: Actually, the hint is not needed, the present proof can be adapted in
an easier way. We perform the following changes at the beginning of page 166: Since every
√
t > 1 is in some interval of the form [q n−1 , q n ] and since the function Λ(t) = 2t log log t
is increasing for t > 3, we find for all t ⩾ q n−1 > 3
√ n
∣B(t)∣ sups⩽qn ∣B(s)∣ 2q log log q n
√ ⩽√ n √ .
2t log log t 2q log log q n 2q n−1 log log q n−1
Therefore
∣B(t)∣ √
lim √ ⩽ (1 + ) q a.s.
2t log log t
t→∞
Letting → 0 and q → 1 along countable sequences, we find the upper bound.
Remark: The interesting paper by Dupuis [4] shows LILs for processes (Xt )t⩾0 with sta-
tionary and independent increments. It is shown there that the important ingredient are
estimates of the type P(Xt > x). Thus, if we know that P(Xt > x) ≍ P ( sups⩽t Xs > x),
we get a LIL for Xt if, and only if, we have a LIL for sups⩽t Xs .
∎∎
114
Problem 12.4. Solution: Direct calculation using (12.6).
∎∎
Problem 12.5. Solution: Denote by W (t) = tB(1/t) and W (0) = 0 the projective reflection
of (Bt )t⩾0 . This is again a BM1 . Thus
P (B(t) < κ(t) as t → ∞) = P (W (t) < κ(t) as t → ∞)

= P (tB(1/t) < κ(t) as t → ∞)
= P (B(1/t) < κ(t)/t as t → ∞)
= P (B(s) < sκ(1/s) as s → 0) .
Set K(s) ∶= sκ(1/s). In order to apply Kolmogorov’s test we need (always s → 0, t = 1/s →
∞, ↑=increasing, ↓=decreasing) that
K(s) ↑ ⇐⇒ sκ(1/s) ↑ ⇐⇒ κ(t)/t ↓
and
√ √ √
K(s)/ s ↓ ⇐⇒ sκ(1/s) ↓ ⇐⇒ κ(t)/ t ↑ .
Finally, by a change of variables and using the integral test,

1 ∞
s−3/2 K(s)e−K
2 (s)/2s
t−3/2 κ(t)e−κ
2 (t)/2t
ds = ∫
s=1/t
∫ dt,
0 1
∎∎
(a) By the LIL for Brownian motion we find

√
Bt Bt 2t log log t
√ =√ ⋅ √
b a+t 2t log log t b a+t
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
limt→∞ (⋯)=1 limt→∞ (⋯)=∞
which shows that

Bt
lim √ =∞
t→∞ b a+t
almost surely. Therefore, P(τ < ∞) = 1.
(b) Let b ⩾ 1 and assume, to the contrary, that E τ < ∞. Then we can use the second
Wald identity, cf. Theorem 5.10, and get
E τ = E B 2 (τ ) = E(b2 (a + τ )) = ab2 + b2 E τ > b2 E τ ⩾ E τ,
leading to a contradiction. Thus, E τ = ∞.
115
(c) Consider the stopping time τ ∧ n. As in b) we get for all b > 0
E (τ ∧ n) = E B 2 (τ ∧ n) ⩽ E(b2 (a + τ ∧ n)).
This gives, if b < 1,
b2 <1 ab2 monotone ab2

(1 − b2 ) E (τ ∧ n) ⩽ ab2 Ô⇒ E (τ ∧ n) ⩽ Ô⇒ Eτ ⩽ < ∞.
1 − b2 convergence 1 − b2
∎∎
116
13 Strassen’s functional law of the iterated
logarithm
Problem 13.1. Solution: We construct a counterexample.

√
The function w(t) = t, 0 ⩽ t ⩽ 1, is a limit point of the family
B(st)
Zs (t) = √
2s log log s
where t > 0 is fixed and for s → ∞.
By the Khintchine’s LIL (cf. Theorem 11.1) we obtain
B(st)
lim √ =1 (almost surely P)
s→∞ 2st log log(st)
and so
B(st) √
lim √ = t (almost surely P)
s→∞ 2s log log(st)
which implies
√
B(st) B(st) log log(st) √
lim √ = lim √ ⋅ = t.
s→∞ 2s log log s s→∞ 2s log log(st) log log s
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
→ 1 for s → ∞
√
On the other hand, the function w(t) = t cannot be a limit point of Zs (⋅) in C(o) [0, 1] for
s → ∞. We prove this indirectly: Let sn = sn (ω) be a sequence, such that limn→∞ sn = ∞.
Then
∥Zsn (⋅) − w(⋅)∥∞ ÐÐÐ→ 0
n→∞
implies that for every > 0 the inequality
√ √ √ √
( t − ) ⋅ 2sn log log sn ⩽ B(sn ⋅ t) ⩽ ( t + ) 2sn log log sn (*)
holds for all sufficiently large n and every t ∈ [0, 1]. This, however, contradicts
√ √
1 1
(1 − ) 2tk log (log ) ⩽ B(tk ) ⩽ (1 + ) 2tk log (log ), (**)
tk tk
for a sequence tk = tk (ω) → 0, k → ∞, cf. Corollary 12.2.
Indeed: fix some n, then the right side of (*) is in contradiction with the left side of (**).
Remark: Note that
1 1 1 ds
∫ w′ (s)2 ds = ∫ = +∞.
0 4 0 s
∎∎
117
Problem 13.2. Solution: For any w ∈ K we have

t 2 t t 1
∣w(t)∣2 = ∣∫ w′ (s) ds∣ ⩽ ∫ w′ (s)2 ds ⋅ ∫ 1 ds ⩽ ∫ w′ (s)2 ds ⋅ t ⩽ t.
0 0 0 0
∎∎
Problem 13.3. Solution: Since u is absolutely continuous (w.r.t. Lebesgue measure), for
almost all t ∈ [0, 1], the derivative u′ (t) exists almost everywhere.
Let t be a point where u′ exists and let (Πn )n⩾1 be a sequence of partitions of [0, 1] such
(n)
that ∣Πn ∣ → 0 as n → ∞. We denote the points in Πn by tk . Clearly, there exists a
(n) (n) (n) (n) (n) (n)
sequence (tjn )n⩾1 such that tjn ∈ Πn and tjn −1 ⩽ t ⩽ tjn for all n ∈ N and tjn − tjn −1 → 0
as n → ∞. We obtain
⎡ (n) ⎤2
⎢ 1 tjn ⎥
fn (t) = ⎢ (n) (n) ∫ (n) u (s) ds⎥⎥
⎢ ′
⎢t − t tjn −1 ⎥
⎣ jn jn −1 ⎦
(n) (n)
to simplify notation, we set tj ∶= tjn and tj−1 ∶= tjn −1 , then
2
1
=[ ⋅ (u(tj ) − u(tj−1 ))]
tj − tj−1
2
1
=[ ⋅ (u(tj ) − u(t) + u(t) − u(tj−1 ))]
tj − tj−1
⎡ ⎤2
⎢ tj − t u(tj ) − u(t) t − tj−1 u(t) − u(tj−1 ) ⎥⎥
= ⎢⎢ ⋅ + ⋅ ⎥
⎢ tj − tj−1 tj − t tj − tj−1 t − tj−1 ⎥
⎣ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ⎦
→ u′ (t) → u′ (t)
ÐÐÐ→ [u′ (t)] .
2
n→∞
∎∎
Problem 13.4. Solution: We use the notation of Chapter 4: Ω = C(o) [0, 1], w = ω, A =
B(C(o) [0, 1]), P = µ, B(t, ω) = Bt (ω) = w(t), t ∈ [0, ∞).
Linearity of Gφ is clear. Let Πn , n ⩾ 1, be a sequence of partitions of [0, 1] such that
limn→∞ ∣Πn ∣ = 0,
(n) (n) (n) (n)
Πn = {sk ∶ 0 = s0 < s1 < . . . < sln = 1} ;
(n) (n) (n) (n)
by s̃k , k = 1, . . . , ln we denote arbitrary intermediate points, i. e. sk−1 ⩽ s̃k ⩽ sk for all
k. Then we have
1
Gφ (ω) = φ(1)B1 (ω) − ∫ Bs (ω) dφ(s)
0
ln
(n) (n)
= φ(1)B1 (ω) − lim ∑ Bs̃(n) (ω)(φ(sk ) − φ(sk−1 )).
∣Πn ∣→0 k=1 k
Write
ln
(n) (n)
Gφn ∶=φ(1)B1 − ∑ Bs̃(n) (φ(sk ) − φ(sk−1 ))
k
k=1
118
ln
(n) (n)
= ∑ (B1 − Bs̃(n) )(φ(sk ) − φ(sk−1 )) + B1 φ(0).
k
k=1
Then Gφ (ω) = limn→∞ Gφn (ω) for all ω ∈ Ω. Moreover, the elementary identity
l l−1
∑ ak (bk − bk−1 ) = ∑ (ak − ak+1 )bk + al bl − a1 b0
k=1 k=1
implies
ln −1
(n)
Gφn = ∑ (Bs̃(n) − Bs̃(n) )φ(sk ) + (B1 − Bs̃(n) )φ(1) − (B1 − Bs̃(n) )φ(0) + B1 φ(0)
k+1 k ln 1
k=1
ln
(n)
= ∑ (Bs̃(n) − Bs̃(n) )φ(sk ) + Bs̃(n) φ(0),
k+1 k 1
k=0
(n) (n)
where s̃ln +1 ∶= 1, s̃0 ∶= 0.
(a) Gφn is a Gaussian random variable with mean E Gφn = 0 and variance
ln
(n)
V Gφn = ∑ φ2 (sk ) V(Bs̃(n) − Bs̃(n) ) + φ2 (0) V Bs̃(n)
k+1 k 1
k=0
ln
(n) (n) (n) (n)
= ∑ φ2 (sk )(s̃k+1 − s̃k ) + φ2 (0)s̃1
k=0
1
ÐÐÐ→ ∫ φ2 (s) ds.
n→∞ 0
This and limn→∞ Gφn = Gφ (P-a.s.) imply that Gφ is a Gaussian random variable
1
with E Gφ = 0 and V Gφ = ∫0 φ2 (s) ds.
(b) Without loss of generality we use for φ and ψ the same sequence of partitions.
Clearly, Gφn ⋅ Gψ
n → G ⋅ G for n → ∞ (P-a.s.) Using the elementary inequality
φ ψ
2ab ⩽ a2 + b2 and the fact that for a Gaussian random variable E(G4 ) = 3(E(G2 ))2 ,
we get
1
E ((Gφn Gψ
n) ) ⩽
2
n ) )]
[ E ((Gφn )4 ) + E ((Gψ 4
2
3 2 2 2
= [( E(Gφn )2 ) + ( E(Gψ n) ) ]
2
3 1 2 1 2
⩽ [( ∫ φ2 (s) ds) + ( ∫ ψ 2 (s) ds) ] + (n ⩾ n ).
2 0 0
This implies
n) Ð
E(Gφn Gψ ÐÐ→ E(Gφ Gψ ).
n→∞
Moreover,
ln ln
(n) (n)
n ) = E [( ∑ (Bs̃(n) − Bs̃(n) )φ(sk )) ⋅ ( ∑ (Bs̃(n) − Bs̃(n) )ψ(sj ))]
E(Gφn Gψ
k+1 k j=0 j+1 j
k=0
ln
(n)
+ φ(0)ψ(0) E(B 2(n) ) + φ(0) E [Bs̃(n) ∑ (Bs̃(n) − Bs̃(n) )ψ(sj )]
s̃1 1 j+1 j
j=0
119
ln
(n)
+ ψ(0) E [Bs̃(n) ∑ (Bs̃(n) − Bs̃(n) )ψ(sk )]
1 k+1 k
k=0
ln
(n) (n)
= ∑ E ((Bs̃(n) − Bs̃(n) )2 ) φ(sk )ψ(sk ) + ⋯
k+1 k
k=0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
(n) (n)
=s̃k+1 −s̃k
1
ÐÐÐ→ ∫ φ(s)ψ(s) ds.
n→∞ 0
This proves
1
E (Gφ Gψ ) = ∫ φ(s)ψ(s) ds.
0
(c) Using a) and b) we see
E [(Gφn − Gψ
n ) ] = E [(Gn ) ] − 2 E [Gn Gn ] + E [(Gn ) ]
2 φ 2 φ ψ ψ 2
1 1 1
=∫ φ2n (s) ds − 2 ∫ φn (s)ψn (s) ds + ∫ ψn2 (s) ds
0 0 0
1
=∫ (φn (s) − ψn (s))2 ds.
0
This and φn → φ in L2 imply that (Gφn )n⩾1 is a Cauchy sequence in L2 (Ω, A, P).
Consequently, the limit X = limn→∞ Gφn exists in L2 . Moreover, as φn → φ in L2 , we
1 1
also obtain that ∫0 φ2n (s) ds → ∫0 φ2 (s) ds.
1
Since Gφn is a Gaussian random variable with mean 0 and variance ∫0 φ2n (s) ds, we
1
see that Gφ is Gaussian with mean 0 and variance ∫0 φ2 (s) ds.
Finally, we have φn → φ and ψn → ψ in L2 ([0, 1]) implying
E(Gφn Gψn ) → E(Gφ Gψ )
—see part b)—and

1 1
∫ φn (s)ψn (s) ds → ∫ φ(s)ψ(s) ds.
0 0
Thus,
1
E(Gφ Gψ ) = ∫ φ(s)ψ(s) ds.
0
∎∎

1. It is clear that H1 is a normed vector space with a scalar product. (The definiteness
of the norm in H1 follows from the absolute continuity! Note that h′ (s) ds = dh(s)
in the scalar product since h ∈ H1 is absolutely continuous.) Let us show that H1 is
L2 (ds)
closed. Assume that (un )n⩾1 ⊂ H1 converges. This means that u′n ÐÐÐ→ w where
n→∞
w ∈ L2 (ds). Thus,
t
W (t) ∶= ∫ w(s) ds exists and W ∈ H1 , W ′ (t) = w(t).
0
H1
Moreover, by construction un ÐÐÐ→ W .
n→∞
120
2. See the last part of Paragraph 13.5.
3. With the Cauchy–Schwarz inequality we see
⟨φ, h⟩H1 ⩽ ∥φ∥H1 ∥h∥H1

φ
Ô⇒ ⟨ , h⟩ ⩽ ∥h∥H1
∥φ∥H1 H1
Ô⇒ sup ⟨φ, h⟩H1 ⩽ sup ⟨φ, h⟩H1 ⩽ ∥h∥H1 .

φ∈H○1 , ∥φ∥H1 =1 φ∈H1 , ∥φ∥H1 =1
Conversely, the supremum is attained if we take φ = h/ ∥h∥H1 . Thus,
h φ
∥h∥H1 = ⟨ , h⟩ ⩽ sup ⟨ , h⟩ .
∥h∥H1 H1 φ∈H1 ∥φ∥H1 H1
If we approximate h ∈ H1 by a sequence (φn )n⩾1 ⊂ H○1 , we get
h φn φ
∥h∥H1 = ⟨ , h⟩ = lim ⟨ , h⟩ ⩽ sup ⟨ , h⟩ .
∥h∥H1 H1
n→∞ ∥φn ∥ 1
H H1 φ∈H○1 ∥φ∥H1 H1
4. Assume that there is some (φn )n⩾1 ⊂ H○1 with ∥φn ∥H1 = 1 and ⟨φn , h⟩H1 ⩾ 2n. Then
∥h∥H1 = sup ⟨h, φ⟩H1 ⩾ sup ⟨φn , h⟩H1 = ∞

∥φ∥H1 =1 n⩾1
which means that h ∉ H1 .
Conversely, assume that for every sequence (φn )n⩾1 ⊂ H○1 with ∥φn ∥H1 = 1 we have
⟨φn , h⟩H1 ⩽ C. (Think! Why this is the proper negation of the condition in the
problem?) Since the supremum can be realized by a sequence, we get for a suitable
sequence of φn ’s
∥h∥H1 = sup ⟨h, φ⟩H1 = lim ⟨φn , h⟩H1 ⩽ C.

∥φ∥H1 =1, φ∈H○1 n→∞
This means that h ∈ H1 .
Remark. An alternative, and more elementary argument for part (d) can be based
on step functions and Lemmas 13.2 and 13.3.
∎∎
Problem 13.6. Solution: The vectors (X, Y ) in a) – d) are a.s. limits of two-dimensional
Gaussian distributions. Therefore, they are also Gaussian. Their mean is clearly 0. The
general density of a two-dimensional Gaussian law (with mean zero) is given by
1 1 x2 y 2 2ρ xy
f (x, y) = √ exp {− ( + − )} .
2πσ1 σ2 1 − ρ2 2(1 − ρ2 ) σ12 σ22 σ1 σ2
In order to solve the problems we have to determine the variances σ12 = V X, σ22 = V Y
E XY
and the correlation coefficient ρ = . We will use the results of Problem 13.4.
σ1 σ2
121
t 1 1 5 1
(a) σ12 = V (∫ s2 dw(s)) = ∫ 1[1/2,t] (s)s4 ds = (t − ),
1/2 0 5 32
σ22 = V w(1/2) = 1/2 (= V B1/2 cf. canonical model),
t 1
E (∫ s2 dw(s) ⋅ w(1/2)) = ∫ 1[1/2,t] (s)s2 ⋅ 1[0,1/2] (s) ds = 0
1/2 0
Ô⇒ ρ = 0.
1 5 1
(b) σ12 = (t − )
5 32
σ22 = V w(u + 1/2) = u + 1/2
t
E (∫ s2 dw(s) ⋅ w(u + 1/2))
1/2
1
=∫ 1[1/2,t] (s)s2 ⋅ 1[0,u+1/2] (s) ds
0
(1/2+u)∧t
=∫ s2 ds
1/2
3
1 1 1
= ((( + u) ∧ t) − ) .
3 2 8
3
1
3 ((( 12 + u) ∧ t) − 81 )
Ô⇒ ρ = 1/2
.
[ 15 (t5 − ) (u + 21 )]
32 ⋅
1
1 5 1
t
(c) σ12 = V (∫ s2 dw(s)) =
(t − ),
1/2 5 32
t 1 1
σ22 = V (∫ s dw(s)) = (t3 − )
1/2 3 8
t t t 1 4 1
E (∫ s2 dw(s) ⋅ ∫ s dw(s)) = ∫ s3 ds = (t − )
1/2 1/2 1/2 4 16
1
(t4 − 1
)
Ô⇒ ρ = 4 16
1/2
.
[ 15 (t5 − ) 1
32 ⋅ 3
1
(t3 − 81 )]
1 1 1
(d) σ12 = V (∫ es dw(s)) = ∫ e2s ds = (e2 − e),
1/2 1/2 2
σ22 = V(w(1) − w(1/2)) = 1/2,
1 1
E (∫ es dw(s) ⋅ (w(1) − w(1/2))) = ∫ es ⋅ 1 ds = e − e1/2 .
1/2 1/2
e − e1/2
Ô⇒ ρ = 1/2
.
( 14 (e2 − e))
∎∎
Problem 13.7. Solution: Let wn ∈ F , n ⩾ 1, and wn → v in C(o) [0, 1]. We have to show that
v ∈ F.
Now:
wn ∈ F Ô⇒ ∃(cn , rn ) ∈ [q −1 , 1] × [0, 1] ∶ ∣wn (cn rn ) − wn (rn )∣ ⩾ 1.
122
Observe that the function (c, r) ↦ w(cr) − w(r) with (c, r) ∈ [q −1 , 1] × [0, 1] is continuous
for every w ∈ C(o) [0, 1].
Since [q −1 , 1] × [0, 1] is compact, there exists a subsequence (nk )k⩾1 such that cnk → c̃ and
rnk → r̃ as k → ∞ and (c̃, r̃) ∈ [q −1 , 1] × [0, 1].
By assumption, wnk → v uniformly and this implies
wnk (cnk rnk ) → v(c̃r̃) and wnk (rnk ) → v(r̃).
Finally,
∣v(c̃r̃) − v(r̃)∣ = lim ∣wnk (cnk rnk ) − wnk (rnk )∣ ⩾ 1,
k→∞
and v ∈ F follows.
∎∎
√
Problem 13.8. Solution: Set L(t) = 2t log log t, t ⩾ e and sn = q n , n ∈ N, q > 1. Then:
(a) for the first inequality:
∣B(sn−1 )∣ ⎛ B(sn−1 ) 1 ⎞
P( > )=P ∣ √ ∣ ⋅√ >
L(sn ) 4 ⎝ sn−1 2q log log sn 4 ⎠
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
∼N (0,1)
√
= P (∣B(1)∣ > 2q log log q n )
4
using Problem 9.6 and P(∣Z∣ > x) = 2 P(Z > x) for x ⩾ 0

√
2 4 2
⩽ √ ⋅ exp {− ⋅ 2q log log q n }
π 2q log log q n 32
C
⩽
n2
if q is sufficiently large.
(b) for the second inequality:

t
sup ∣w(t)∣ = sup ∣∫ w′ (s) ds∣
t⩽q −1 t⩽q −1 0
1/q
⩽∫ ∣w′ (s)∣ ds
0
1/q 1/q 1/2
′
⩽ [∫ w (s) ds ⋅ ∫
2
ds]
0 0
1 1 1/2
⩽ [∫ w′ (s)2 ds ⋅ ]
0 q
√
2r
⩽ <
q 4
for all sufficiently large q.
123
B(⋅ sn )
(c) for the third inequality: Brownian scaling √
sn
∼ B(⋅) yields
⎛ ∣B(tsn )∣ ⎞ ⎛ ∣B(t)∣ ⎞
P sup √ > = P sup √ >
⎝0⩽t⩽q−1 2sn log log sn 4 ⎠ ⎝0⩽t⩽q−1 2 log log sn 4 ⎠
⎛ √ ⎞
=P sup ∣B(t)∣ > 2 log log sn
⎝0⩽t⩽q−1 4 ⎠
(*) √
⩽ 2 P (∣B(1/q)∣ > 2 log log sn )
4
⎛ ∣B(1/q)∣ √ ⎞
= 2P √ > 2q log log q n
⎝ 1/q 4 ⎠
C
⩽
n2
for all q sufficiently large. In the estimate marked with (*) we used
P ( sup ∣B(t)∣ > x) ⩽ 2 P ( sup B(t) > x) = 2 P(M (t0 ) > x) = 2 P (∣B(t0 )∣ > x).
Thm.
0⩽t⩽t0 0⩽t⩽t0 6.9
(d) for the last inequality:
⎛ ∣B(sn−1 )∣ ∣B(tsn )∣ 3 ⎞
P + sup ∣w(t)∣ + sup >
⎝ L(sn ) t⩽q −1 0⩽t⩽q −1 L(sn ) 4⎠
⎛ ∣B(sn−1 )∣ ∣B(tsn )∣ ⎞
⩽P > or sup ∣w(t)∣ > or sup >
⎝ L(sn ) 4 t⩽q −1 4 0⩽t⩽q −1 L(sn ) 4⎠
∣B(sn−1 )∣ ⎛ ⎞ ⎛ ∣B(tsn )∣ ⎞
⩽ P( > ) + P sup ∣w(t)∣ > + P sup >
L(sn ) 4 ⎝t⩽q−1 4⎠ ⎝0⩽t⩽q−1 L(sn ) 4⎠
C C
⩽ 2
+0+ 2
n n
for all sufficiently large q. Using the Borel–Cantelli lemma we see that
⎛ ∣B(sn−1 )∣ ∣B(tsn )∣ ⎞ 3
lim + sup ∣w(t)∣ + sup ⩽ .
n→∞ ⎝ L(sn ) t⩽q −1 0⩽t⩽q −1 L(sn ) ⎠ 4
∎∎
124
14 Skorokhod representation
Problem 14.1. Solution: Clearly, FtB ∶= σ(Br ∶ r ⩽ t) ⊂ σ(Br ∶ r ⩽ t, U, V ) = Ft . It remains

to show that Bt − Bs á Fs for all s ⩽ t. Let A, A′′ , C be Borel sets in Rd . Then we find for
F ∈ FsB
P ({Bt − Bs ∈ C} ∩ F ∩ {U ∈ A} ∩ {V ∈ A′ })
= P ({Bt − Bs ∈ C} ∩ F ) ⋅ P ({U ∈ A} ∩ {V ∈ A′ }) (since U, V á F∞
B
)
= P ({Bt − Bs ∈ C}) ⋅ P (F ) ⋅ P ({U ∈ A} ∩ {V ∈ A′ }) (since Bt − Bs á F∞
B
)
= P ({Bt − Bs ∈ C}) ⋅ P (F ∩ {U ∈ A} ∩ {V ∈ A′ }) (since U, V á F∞
B
)
and this shows that Bt −Bs is independent of the family Es = {F ∩G ∶ F ∈ FsB , G ∈ σ(U, V )}.
This family is stable under finite intersections, so Bt − Bs á σ(Es ) = Fs .
∎∎
125
15 Stochastic integrals: L2–theory
Problem 15.1. Solution: By definition of the angle bracket,
M 2 − ⟨M ⟩ and N 2 − ⟨N ⟩
are martingales. Moreover, M ± N are L2 -martingales, i. e.
(M + N )2 − ⟨M + N ⟩ and (M − N )2 − ⟨M − N ⟩
are martingales. So, we subtract them to get a new martingale:
(M + N )2 − (M − N )2 = 4M N and ⟨M + N ⟩ − ⟨M − N ⟩ = 4⟨M, N ⟩
def
which shows that 4M N − 4⟨M N ⟩ is a martingale.
∎∎
Problem 15.2. Solution: Note that
[a, b) ∩ [c, d) = [a ∨ c, b ∧ d) (with the convention [M, m) = ∅ if M ⩾ m).
Then assume that we have any two representations for a simple process
f = ∑ φj−1 1[sj−1 ,sj ) = ∑ ψk−1 1[tk−1 ,tk )

j k
Then
f = ∑ φj−1 1[sj−1 ,sj ) 1[0,T ) = ∑ φj−1 1[sj−1 ,sj ) 1[tk−1 ,tk )
j j,k
and, similarly,
f = ∑ ψk−1 1[sj−1 ,sj ) 1[tk−1 ,tk ) .
k,j
Then we get, since φj−1 = ψk−1 whenever [sj−1 , sj ) ∩ [tk−1 , tk ) ≠ ∅
∑ φj−1 (B(sj ) − B(sj−1 )) = ∑∑ φj−1 (B(sj ∧ tk ) − B(sj−1 ∨ tk−1 ))

j (j,k) ∶ [sj−1 ,sj )∩[tk−1 ,tk )≠∅
= ∑∑ ψk−1 (B(sj ∧ tk ) − B(sj−1 ∨ tk−1 ))

(j,k) ∶ [sj−1 ,sj )∩[tk−1 ,tk )≠∅
= ∑∑ ψk−1 (B(sj ∧ tk ) − B(sj−1 ∨ tk−1 ))

(k,j) ∶ [sj−1 ,sj )∩[tk−1 ,tk )≠∅
= ∑ ψk−1 (B(tk ) − B(tk−1 ))

k
∎∎
127
• Positivity is clear, finiteness follows with Doob’s maximal inequality
E [sup ∣Ms ∣2 ] ⩽ 4 sup E [∣Ms ∣2 ] = 4 E [∣MT ∣2 ] .

s⩽T s⩽T
• Triangle inequality:
1
2
∥M + N ∥M2 = (E [sup ∣Ms + Ns ∣ ]) 2
T
s⩽T
1
2
⩽ (E [(sup ∣Ms ∣ + sup ∣Ns ∣) ]) 2
s⩽T s⩽T
1 1
2 2
⩽ (E [sup ∣Ms ∣ ]) + (E [sup ∣Ns ∣ ])
2 2
s⩽T s⩽T
where we used in the first estimate the subadditivity of the supremum and in the
second inequality the Minkowski inequality (triangle inequality) in L2 .
• Positive homogeneity
1 1
2 2
∥λM ∥M2 = (E [sup ∣λMs ∣ ]) = ∣λ∣ (E [sup ∣Ms ∣ ]) = ∣λ∣ ⋅ ∥M ∥M2 .
2 2
T T
s⩽T s⩽T
• Definiteness
∥M ∥M2 = 0 ⇐⇒ sup ∣Ms ∣2 = 0 (almost surely).
T
s⩽T
∎∎
Problem 15.4. Solution: Let fn → f and gn → f be two sequences which approximate f in

the norm of L2 (λT ⊗ P). Then we have
2 2
E (∣fn ● BT − gn ● BT ∣ ) = E (∣(fn − gn ) ● BT ∣ )
T
= E (∫ ∣fn (s) − gn (s)∣2 ds)
0
= ∥fn − gn ∥2L2 (λT ⊗P)

ÐÐÐ→ 0.
n→∞
This means that

L2 (P)- lim fn ● BT = L2 (P)- lim gn ● BT .
n→∞ n→∞
∎∎
Problem 15.5. Solution: Solution 1: Let τ be a stopping time and consider the sequence of
discrete stopping times
⌊2m τ ⌋ + 1
τm ∶= ∧ T.
2m
128
Let t0 = 0 < t1 < t2 < . . . < tn = T and, without loss of generality, τm (Ω) ⊂ {t0 , . . . , tn }.
Then (Bt2j − tj )j is again a discrete martingale and by optional stopping we get that
(Bτ2m ∧tj − τm ∧ tj )j is a discrete martingale. This means that for each m ⩾ 1
⟨B τm ⟩tj = τm ∧ tj for all j
and this indicates that we can set ⟨B τ ⟩t = t ∧ τ . This process makes Bt∧τ
2
− t ∧ τ into a
martingale. Indeed: fix 0 ⩽ s ⩽ t ⩽ T and add them to the partition, if necessary. Then
a.e. L1 (P)
Bτ2m ∧t ÐÐÐ→ Bτ2∧t and Bτ2m ∧t ÐÐÐ→ Bτ2∧t
m→∞ m→∞
by dominated convergence, since supr⩽T Br2 is an integrable majorant. Thus,
∫ (Bτ ∧s − τ ∧ s) d P = m→∞
lim ∫ (Bτ2m ∧s − τm ∧ s) d P
2
F F
= lim ∫ (Bτ2m ∧t − τm ∧ t) d P
m→∞ F
=∫ (Bτ2∧t − τ ∧ t) d P for all F ∈ Fs

F
and we conclude that (Bτ2∧t − τ ∧ t)t is a martingale.
Solution 2: Observe that

t
Btτ = ∫ 1[0,τ ) dBs
0
and by Theorem 15.13 b) we get
● t t
⟨∫ 1[0,τ ) dBs ⟩ = ∫ 12[0,τ ) ds = ∫ 1[0,τ ) ds = τ ∧ t.
0 t 0 0
(Of course, one should make sure that 1[0,τ ) ∈ L2T , see e.g. Problem 15.16 below or Prob-
lem 16.2 in combination with Theorem 15.20.)
∎∎
Problem 15.6. Solution: We begin with a general remark: if f = 0 on [0, s] × Ω, we can use
Theorem 15.13 f) and deduce f ● Bs = 0.
(a) We have
t
E [(f ● Bt )2 ∣ Fs ] = E [(f ● Bt − f ● Bs )2 ∣ Fs ] = E [∫ f 2 (r) dr ∣ Fs ] .
15.13 b)
(15.20) s
If both f and g vanish on [0, s], the same is true for f ± g. We get
2 t
E [((f ± g) ● Bt ) ∣ Fs ] = E [∫ (f ± g)2 (r) dr ∣ Fs ] .
s
Subtracting the ‘minus’ version from the ‘plus’ version and gives
2 2 t
E [((f + g) ● Bt ) − ((f − g) ● Bt ) ∣ Fs ] = E [∫ (f + g)2 (r) − (f − g)2 (r) dr ∣ Fs ] .
s
or
t
4 E [(f ● Bt ) ⋅ (g ● Bt ) ∣ Fs ] = 4 E [∫ (f ⋅ g)(r) dr ∣ Fs ] .
s
129
(b) Since f ● Bt is a martingale, we get for t ⩾ s
E (f ● Bt ∣ Fs ) = f ● Bs =
martingale see above
0
since f vanishes on [0, s].
(c) By Theorem 15.13 f) we have for all t ⩽ T
f ● Bt (ω)1A (ω) = 0 ● Bt (ω)1A (ω) = 0.
∎∎
n→∞
Problem 15.7. Solution: Because of Lemma 15.10 it is enough to show that fn ● BT ÐÐÐ→
f ● BT in L2 (P). This follows immediately from Theorem 15.13 c):
2 2
E [∣fn ● BT − f ● BT ∣ ] = E [∣(fn − f ) ● BT ∣ ]
T n→∞
= E [∫ ∣fn (s) − f (s)∣2 ds] ÐÐÐ→ 0.
0
∎∎
Problem 15.8. Solution: Without loss of generality we assume that f (0) = 0. Fix c > 0.
Then we have for all > 0, using the Markov inequality and the Hölder inequality with
p = 4 and q = 4/3
1/2 √
1 1
P (∣ ∫ f (s) dBs ∣ > c) = P (∣ ∫ f (s) dBs ∣ > c)
B 0 B 0
⎡ 1 1/2 ⎤
⎢
⎥
⩽ c−1/2 E ⎢ ∣ (s) ∣ ⎥
⎢ ∣B ∣1/2 ∫0
f dBs
⎥
⎣ ⎦
⎡ 1 ⎤⎞3/4 2 1/4
−1/2 ⎛ ⎢ ⎥
⩽c E⎢ ⎥ (E [( ∫ f (s) dB ) ]) .
⎝ ⎢⎣ ∣B ∣2/3 ⎥⎦⎠
s
0
Using Itô’s isometry and Brownian scaling yields
1 3/4 1/4
P (∣ ∫ f (s) dBs ∣ > c) ⩽ c
−1/2
(E [∣B1 ∣−2/3 ]) −1/4 (∫ E [∣f (s)∣2 ] ds)
B 0 0
1/4
3/4 1
= c−1/2 (E [∣B1 ∣−2/3 ]) ( ∫ E [∣f (s)∣2 ] ds)
0
1/4
3/4
⩽ c−1/2 (E [∣B1 ∣−2/3 ]) (sup E [∣f (s)∣2 ]) .
s⩽
Since E [∣B1 ∣−2/3 ] < ∞, see (the solution of) Problem 11.8, we find
1/4
1
lim P (∣ ∫ f (s) dBs ∣ > c) ⩽ C ( lim sup E [∣f (s)∣2 ])
→0 B 0 →0 s⩽
´¹¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¶
=lim sup→0
1/4
= C (lim E [∣f ()∣ ]) 2
= 0.
→0
130
Alternative solution: We fix c, K > 0 and observe

√ √
1 1
P (∣ ∫ f (s) dBs ∣ > c) ⩽ P (∣ ∫ f (s) dBs ∣ > c, ∣ ∣ ⩽ K) + P (∣ ∣ > K)
B 0 B 0 B B
1 c ∣B ∣ 1
⩽ P (∣ √ ∫ f (s) dBs ∣ > ) + P ( √ < )
0 K K
The first expression can now be estimated using the Markov inequality and Itô’s isometry,
for the second expression we use Brownian scaling.
∎∎
Problem 15.9. Solution: Assume that (f ● B)2 − A is a martingale where At is continuous

and increasing. Since (f ● B)2 − f 2 ● ⟨B⟩ is a martingale, we conclude that
((f ● B)2 − f 2 ● ⟨B⟩) − ((f ● B)2 − A) = f 2 ● ⟨B⟩ − A
is a continuous martingale with BV paths. Hence, it is a.s. constant.
∎∎
Problem 15.10. Solution: By the Cauchy–Schwarz inequality we see that

√ √
t t t
E [∫ f (s)g(s) ds] ⩽ E [∫ ∣f (s)∣2 ds] E [∫ ∣g(s)∣2 ds] < ∞
0 0 0
t
which means that ∫0 f (s)g(s) ds is well-defined. Since f ± g ∈ L2t , we get by polarization
4⟨f ● B, g ● B⟩t = ⟨(f + g) ● B⟩t − ⟨(f − g) ● B⟩t

t t
= ∫ (f + g) (s) ds − ∫ (f − g) (s) ds
Thm. 15.13b) 2 2
0 0
t
= ∫ [(f + g)2 (s) − (f − g)2 (s)] ds
0
t
= ∫ 4f (s)g(s) ds,
0
∎∎
L2
Problem 15.11. Solution: If Xn Ð→ X then supn E(Xn2 ) < ∞ and the claim follows from the
fact that
E ∣Xn2 − Xm
2
∣ = E [∣Xn − Xm ∣∣Xn + Xm ∣]
√ √
⩽ E ∣Xn + Xm ∣2 E ∣Xn − Xm ∣2
√ √ √
⩽ ( E ∣Xn ∣2 + E ∣Xm ∣2 ) E ∣Xn − Xm ∣2 .
∎∎
131
Problem 15.12. Solution: Let Π = {t0 = 0 < t1 < . . . < tn = T } be a partition of [0, T ]. Then
we get
n
BT3 = ∑ (Bt3j − Bt3j−1 )
j=1
n
= ∑ (Btj − Btj−1 )[Bt2j + Btj Btj−1 + Bt2j−1 ]
j=1
n
= ∑ (Btj − Btj−1 )[Bt2j − 2Btj Btj−1 + Bt2j−1 + 3Btj Btj−1 ]
j=1
n
= ∑ (Btj − Btj−1 )[(Btj − Btj−1 )2 + 3Btj Btj−1 ]
j=1
n
= ∑ (Btj − Btj−1 )[(Btj − Btj−1 )2 + 3Bt2j−1 + 3Btj−1 (Btj − Btj−1 )]
j=1
n n n
3 2
= ∑ (Btj − Btj−1 ) + 3 ∑ Bt2j−1 (Btj − Btj−1 ) + 3 ∑ Btj−1 (Btj − Btj−1 )
j=1 j=1 j=1
n n n
3
= ∑ (Btj − Btj−1 ) + 3 ∑ Bt2j−1 (Btj − Btj−1 ) + 3 ∑ Btj−1 (tj − tj−1 )
j=1 j=1 j=1
n
2
+ 3 ∑ Btj−1 [(Btj − Btj−1 ) − (tj − tj−1 )]
j=1
= I1 + I2 + I3 + I4 .
Clearly,
T T
I2 ÐÐÐ→ 3 ∫ Bs2 dBs and I3 ÐÐÐ→ 3 ∫ Bs ds
∣Π∣→0 0 ∣Π∣→0 0
by Proposition 15.16 and by the construction of the stochastic resp. Riemann-Stieltjes

integral. The latter also converges in L2 since I2 and, as we will see in a moment, I1 and
I4 converge in L2 -sense.
Let us show that I1 , I4 → 0.
⎛n 3 ⎞ (B1)
n
3
V I1 = V ∑ (Btj − Btj−1 ) = ∑ V ((Btj − Btj−1 ) )
⎝j=1 ⎠ j=1
n
= ∑ V (Bt3j −tj−1 )
(B2)
j=1
n
= ∑ (tj − tj−1 ) V (B1 )
scaling 3 3
j=1
n
⩽ ∣Π∣ ∑ (tj − tj−1 ) V (B13 )
2
j=1
= ∣Π∣2 T V(B13 ) ÐÐÐ→ 0.

∣Π∣→0
Moreover,
⎛⎛ n 2 ⎞ ⎞
2
E(I42 ) = E ⎜ 3 ∑ Btj−1 [(Btj − Btj−1 ) − (tj − tj−1 )] ⎟

⎝⎝ j=1 ⎠ ⎠
132
⎛n n 2 2 ⎞
= 9 E ∑ ∑ Btj−1 [(Btj − Btj−1 ) − (tj − tj−1 )]Btk−1 [(Btk − Btk−1 ) − (tk − tk−1 )]
⎝j=1 k=1 ⎠
⎛n 2 2⎞
= 9 E ∑ Bt2j−1 [(Btj − Btj−1 ) − (tj − tj−1 )]
⎝j=1 ⎠
since the mixed terms break away, see below.

n
2 2
= 9 ∑ E (Bt2j−1 [(Btj − Btj−1 ) − (tj − tj−1 )] )
j=1
n
2 2
= 9 ∑ E (Bt2j−1 ) E ([(Btj − Btj−1 ) − (tj − tj−1 )] )
(B1)
j=1
n
2
= 9 ∑ E (Bt2j−1 ) E ([Bt2j −tj−1 − (tj − tj−1 )] )
(B2)
j=1
n
2
= 9 ∑ tj−1 E (B12 )(tj − tj−1 )2 E ([B12 − 1] )
scaling
j=1
n
= 9 ∑ tj−1 (tj − tj−1 )2 V(B12 )
j=1
n
⩽ 9T ∣Π∣ ∑ (tj − tj−1 ) V(B12 )
j=1
⩽ 9T 2 ∣Π∣ V(B12 ) ÐÐÐ→ 0.

∣Π∣→0
Now for the argument with the mixed terms. Let j < k; then tj−1 < tj ⩽ tk−1 < tk , and by
the tower property,
2 2
E (Btj−1 [(Btj − Btj−1 ) − (tj − tj−1 )]Btk−1 [(Btk − Btk−1 ) − (tk − tk−1 )])
2 2
= E (E [Btj−1 [(Btj − Btj−1 ) − (tj − tj−1 )]Btk−1 [(Btk − Btk−1 ) − (tk − tk−1 )] ∣ Ftk−1 ])
tower
2 2
= E (Btj−1 [(Btj − Btj−1 ) − (tj − tj−1 )]Btk−1 E [[(Btk − Btk−1 ) − (tk − tk−1 )] ∣ Ftk−1 ])
pull
out
2 2
= E (Btj−1 [(Btj − Btj−1 ) − (tj − tj−1 )]Btk−1 E [(Btk − Btk−1 ) − (tk − tk−1 )] )
(B1)
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
=0
= 0.
∎∎
Problem 15.13. Solution: Let Π = {t0 = 0 < t1 < . . . < tn = T } be a partition of [0, T ]. Then
we get
f (tj )Btj − f (tj−1 )Btj−1

= f (tj−1 )(Btj − Btj−1 ) + Btj−1 (f (tj ) − f (tj−1 )) + (Btj − Btj−1 )(f (tj ) − f (tj−1 )).
If we sum over j = 1, . . . , n we get
f (T )BT − f (0)B0
n n n
= ∑ f (tj−1 )(Btj − Btj−1 ) + ∑ Btj−1 (f (tj ) − f (tj−1 )) + ∑ (Btj − Btj−1 )(f (tj ) − f (tj−1 ))
j=1 j=1 j=1
133
= I1 + I2 + I3 .
Clearly,
L2 T
I1 Ð→ ∫ f (s) dBs (stochastic integral)
0
a.s. T
I2 ÐÐ→ ∫ Bs df (x) (Riemann-Stieltjes integral)
0
and if we can show that I3 → 0 in L2 , then we are done (as this also implies the L2 -
convergence of I2 ). Now we have
⎡ n 2⎤
⎢⎛ ⎞ ⎥⎥
⎢
E ⎢ ∑ (Btj − Btj−1 )(f (tj ) − f (tj−1 )) ⎥
⎢⎝j=1 ⎠ ⎥
⎣ ⎦
⎡n n ⎤
⎢ ⎥
= E ⎢ ∑ ∑ (Btj − Btj−1 )(f (tj ) − f (tj−1 ))(Btk − Btk−1 )(f (tk ) − f (tk−1 ))⎥⎥
⎢
⎢j=1 k=1 ⎥
⎣ ⎦
the mixed terms break away because of the independent increments property of Brownian
motion
n
= ∑ E [(Btj − Btj−1 )2 (f (tj ) − f (tj−1 ))2 ]
j=1
n
= ∑ (f (tj ) − f (tj−1 ))2 E [(Btj − Btj−1 )2 ]
j=1
n
= ∑ (tj − tj−1 )(f (tj ) − f (tj−1 ))2
j=1
n
⩽ 2 ∣Π∣ ⋅ ∥f ∥∞ ∑ ∣f (tj ) − f (tj−1 )∣
j=1
⩽ 2 ∣Π∣ ⋅ ∥f ∥∞ VAR1 (f ; [0, T ]) ÐÐÐ→ 0

∣Π∣→0
where we used the fact that a BV-function is necessarily bounded:
∣f (t)∣ ⩽ ∣f (t) − f (0)∣ + ∣f (0)∣ ⩽ VAR1 (f ; [0, t]) + VAR1 (f ; {0}) ⩽ 2VAR1 (f ; [0, T ])
for all t ∈ [0, T ].
∎∎
Problem 15.14. Solution: Replace, starting in the fourth line of the proof of Proposi-
tion 15.16, the argument as follows:
By the maximal inequalities (15.22) for Itô integrals we get

t 2
E [sup ∣∫ [f (s) − f Π (s)] dBs ∣ ]
t⩽T 0
T 2
⩽ 4∫ E [∣f (s) − f Π (s)∣ ] ds
0
n sj
=4∑∫ E [∣f (s) − f (sj−1 )∣2 ] ds
j=1 sj−1
134
n sj
⩽4∑∫ sup E [∣f (u) − f (v)∣2 ] ds ÐÐÐ→ 0.
j=1 sj−1 u,v∈[sj−1 ,sj ] ∣Π∣→0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
→0, ∣Π∣→0
∎∎
Problem 15.15. Solution: To simplify notation, we drop the n in Πn and write only 0 = t0 <
t1 < . . . < tk = T and
α
θn,j = θj = αtj + (1 − α)tj−1 .
We get
k T
LT (α) ∶= L2 (P)- lim ∑ Bθj (Btj − Btj−1 ) = ∫ Bs dBs + αT.
∣Π∣→0 j=1 0
Indeed, we have
k
∑ Bθj (Btj − Btj−1 )
j=1
k k
= ∑ Btj−1 (Btj − Btj−1 ) + ∑ (Bθj − Btj−1 )(Btj − Btj−1 )
j=1 j=1
k k k
= ∑ Btj−1 (Btj − Btj−1 ) + ∑ (Bθj − Btj−1 )2 + ∑ (Btj − Bθj )(Bθj − Btj−1 )
j=1 j=1 j=1
= X + Y + Z.
L2 T
We know already that X ÐÐÐ→ ∫0 Bs dBs . Moreover,
∣Π∣→0
⎛k ⎞
V Z = V ∑ (Btj − Bθj )(Bθj − Btj−1 )
⎝j=1 ⎠
k
= ∑ V [(Btj − Bθj )(Bθj − Btj−1 )]
j=1
k
= ∑ E [(Btj − Bθj )2 (Bθj − Btj−1 )2 ]
j=1
k
= ∑ E [(Btj − Bθj )2 ] E [(Bθj − Btj−1 )2 ]
j=1
k
= ∑ (tj − θj )(θj − tj−1 )
j=1
k
as in Theorem 9.1
= α(1 − α) ∑ (tj − tj−1 )(tj − tj−1 ) ÐÐÐÐÐÐÐÐÐ→ 0.
j=1
Finally,
⎛k ⎞ k
E Y = E ∑ (Bθj − Btj−1 )2 = ∑ E(Bθj − Btj−1 )2
⎝j=1 ⎠ j=1
k k
= ∑ (θj − tj−1 ) = α ∑ (tj − tj−1 ) = αT.
j=1 j=1
135
The L2 -convergence follows now literally as in the proof of Theorem 9.1.
Consequence: LT (α) = 1
2
(BT2 + (2α − 1)T ), and this stochastic integral is a martingale if,
and only if, α = 0, i. e. if θj = tj−1 is the left endpoint of the interval.
For α = 1
2 we get the so-called Stratonovich or mid-point stochastic integral. This will
obey the usual calculus rules (instead of Itô’s rule). A first sign is the fact that
LT ( 12 ) = 21 BT2
and we usually write

T
LT ( 12 ) = ∫ Bs ○ dBs
0
with the Stratonovich-circle ○ to indicate the mid-point rule.
∎∎
(a) Let τk be a sequence of stopping times with countably many, discrete values such
that τk ↓ τ . For example, τk ∶= (⌊2k τ ⌋ + 1)/2k , see Lemma A.15 in the appendix.
Write s1 < . . . < sK for the values of τk . In particular,
1[0,T ∧τk ) = ∑ 1{T ∧τk =T ∧sj } 1[0,T ∧sj )

j
And so
{(s, ω) ∶ 1[0,T ∧τk (ω)) (s) = 1} = ⋃[0, T ∧ sj ) × {T ∧ τk = T ∧ sj }.

j
Since {T ∧ τk = T ∧ sj } ∈ FT ∧sj , it is clear that
{(s, ω) ∶ 1[0,T ∧τk (ω)) (s) = 1} ∩ ([0, t] × Ω) ∈ B[0, t] × Ft for all t ⩾ 0
and progressive measurability of 1[0,T ∧τk ) follows.
(b) Since T ∧ τk ↓ T ∧ τ and T ∧ τk has only finitely many values, and we find
lim 1[0,T ∧τk ) = 1[0,T ∧τ ]

k→∞
almost surely. Consequently, 1[0,T ∧τ (ω)] (s) is also P-measurable.
In fact, we do not need to prove the progressive measurability of 1[0,T ∧τ ) to evaluate

the integral. If you want to show it nevertheless, have a look at Problem 16.2 below.
(c) Fix k and write 0 ⩽ s1 < . . . < sK for the values of T ∧ τk . Following the proof of
Theorem 15.9.c)
∫ 1[0,T ∧τk ) (s) dBs = ∫ ∑ 1[T ∧sj−1 ,T ∧sj ) (s)1[0,T ∧τk ) (s) dBs
j
= ∑ ∫ 1[T ∧sj−1 ,T ∧sj ∧τk ) (s) 1{T ∧τk >T ∧sj−1 } dBs
j ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
T ∧sj ∧τk =T ∧sj FT ∧sj−1 -mble
136
= ∑ 1{T ∧τk >T ∧sj−1 } ∫ 1[T ∧sj−1 ,T ∧sj ) (s) dBs

j
= ∑ 1{T ∧τk >T ∧sj−1 } (BT ∧sj − BT ∧sj−1 )

j
= BT ∧τk .
(d) 1[0,T ∧τ ) = L2 - limk 1[0,T ∧τk ) : This follows from
E ∫ ∣1[0,T ∧τk ) (s) − 1[0,T ∧τ ) (s)∣2 ds = E ∫ ∣1[T ∧τ, T ∧τk ) (s)∣2 ds
= E ∫ 1[T ∧τ, T ∧τk ) (s) ds
= E(T ∧ τk − T ∧ τ ) ÐÐÐ→ 0
k→∞
by dominated convergence.
(e) By the very definition of the stochastic integral we find now
∫ 1[0,T ∧τ ) (s) dBs = L - lim ∫ 1[0,T ∧τk ) (s) dBs = L - lim BT ∧τk = BT ∧τ
2 d) 2 c)
k k
by the continuity of Brownian motion and dominated convergence: sups⩽T ∣Bs ∣ is

integrable.
(f) The result is, in the light of the localization principle of Theorem 15.13 not unex-
pected.
∎∎
Problem 15.17. Solution: Throughout the proof t ⩾ 0 is arbitrary but fixed.
• Clearly, ∅, [0, T ] × Ω ∈ P.
• Let Γ ∈ P. Then
Γc ∩ ([0, t] × Ω) = ([0, t] × Ω) ∖ (Γ ∩ ([0, t] × Ω)) ∈ B[0, t] ⊗ Ft ,

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
∈B[0,t]⊗Ft ∈B[0,t]⊗Ft
thus Γc ∈ P.
• Let Γn ∈ P. By definition
Γn ∩ ([0, t] × Ω) ∈ B[0, t] ⊗ Ft
and we can take the union over n to get
(⋃ Γn ) ∩ ([0, t] × Ω) = ⋃ (Γn ∩ ([0, t] × Ω)) ∈ B[0, t] ⊗ Ft

n n
i. e. ⋃n Γn ∈ P.
∎∎
137
Problem 15.18. Solution: Let f (t, ω) be right-continuous on the interval [0, T ]. (We consider
only T < ∞ since the case of the infinite interval [0, ∞) is actually easier.)
Set
⌊2n s⌋+1
fnT (s, ω) ∶= f ( 2n ∧ T, ω)
then
fnT (s, ω) = ∑ f ( k+1
2n ∧ T, ω) 1[k2−n ,(k+1)2−n ) (s) (s ⩽ T )
k
and, since (⌊2n s⌋ + 1)/2n ↓ s, we find by right-continuity that fn → f as n → ∞. This

means that it is enough to consider the P-measurability of the step-function fn .
Fix n ⩾ 0, write tj = j2−n . Then t0 = 0 < t1 < . . . tN ⩽ T for some suitable N . Observe that
for any x ∈ R
N
{(s, ω) ∶ f (s, ω) ⩽ x} = {T } × {ω ∶ f (T, ω) ⩽ x} ∪ ⋃ [tj−1 , tj ) × {ω ∶ f (tj , ω) ⩽ x}
j=1
and each set appearing in the union set on the right is in B[0, T ] ⊗ FT .
This shows that fnT and f are B[0, T ] ⊗ FT measurable.
Now consider fnt and f (t)1[0,t] . We conclude, with the same reasoning, that both are
B[0, t] ⊗ Ft measurable.
This shows that a right-continuous f is progressive.
If f is left-continuous, we use ⌊2n s⌋/2n ↑ s and define the approximating function as
gnT (s, ω) = ∑ f ( 2kn ∧ T, ω) 1[k2−n ,(k+1)2−n ) (s) (s ⩽ T ).

k
The rest of the proof is similar.
∎∎
Problem 15.19. Solution: By definition, there is a sequence fn of elementary processes, i. e. of

processes of the form
fn (s, ω) = ∑ φj−1 (s)1[tj−1 ,tj ) (s)
j
where φj−1 is Ftj−1 measurable such that fn → f in L2 (µT ⊗ P). In particular, there is a
subsequence such that
t t
lim ∫ ∣fn(k) (s)∣2 dAs = ∫ ∣f (s)∣2 dAs a.s.
k→∞ 0 0
t
so that it is enough to check that the integrals ∫0 ∣fn(j) (s)∣2 dAs are adapted. By defintion
t
∫ ∣fn(j) (s)∣2 dAs = ∑ φ2j−1 (Atj ∧t − Atj−1 ∧t )
0 j
and from this it is clear that the integral is Ft measurable for each t.
∎∎
138
16 Stochastic integrals: Beyond L2T
Problem 16.1. Solution: Yes. In view of Lemma 16.3 we have to show that L0T ⊃ L2T,loc . Let
f ∈ L2T,loc and take some localizing sequence (σn )n⩾1 such that
a.s. T ∧σn
σn ÐÐÐ→ ∞ and ∫ ∣f (s, ⋅)∣2 ds < ∞
n→∞ 0
(the finiteness of the integral latter follows from the fact that f 1[0,σn ) is in L2T for each
n. Moreover, f 1[0,σn ) is P-measurable, hence f 1[0,T ∧σn ) → f 1[0,T ) is P-measurable. Note
that the completeness of the filtration allows that σn → ∞ holds only a.s.). Now observe
that for every fixed ω there is some n(T, ω) ⩾ 1 such that for all n ⩾ n(T, ω) we have
σn (ω) ⩾ T . Thus,
T T ∧σn (ω)
∫ ∣f (s, ω)∣2 ds = ∫ ∣f (s, ω)∣2 ds < ∞.
0 0
∎∎
Problem 16.2. Solution: Solution 1: We have that the process t ↦ 1[0,τ (ω)) (t) is adapted
{ω ∶ 1[0,τ (ω)) (t) = 0} = {τ ⩽ t} ∈ Ft
since τ is a stopping time. By Problem 15.18 we conclude that 1[0,τ ) is progressive.
Solution 2: Set tj = j2−n and define
⌊2n s⌋
Int (s, ω) ∶= 1[0,τ (ω)) ( 2n ∧ t) = ∑ 1[0,τ (ω)) (tj+1 ∧ t)1[tj ,tj+1 ) (s ∧ t).
j
Since ⌊2n s⌋/2n ↓ s we find, by right-continuity, Int → 1[0,τ ) . Therefore, it is enough to

check that Int is B[0, t] ⊗ Ft -measurable. But this is obvious from the form of Int .
∎∎
Problem 16.3. Solution: Assume that σn are stopping times such that (Mtσn 1{σn >0} )t is a
martingale. Clearly,
• τn ∶= σn ∧ n ↑ ∞ almost surely as n → ∞;
• {σn > 0} = {σn ∧ n > 0} = {τn > 0};
• by optional stopping, the following process is a martingale for each n:
1{σn >0} = Mtσn ∧n 1{σn >0} = Mtσn ∧n 1{σn ∧n>0} = Mt n 1{τn >0} .
σn τ
Mt∧n
139
Remark: This has an interesting consequence:

Doob
E [sup ∣M (s ∧ τn )∣2 ] ⩽ 4 E [∣M (τn )∣2 ] ⩽ 4 E [∣M (n)∣2 ].
s⩽T
∎∎
(a) The picture below show that Iσu = Iτu = u since t ↦ It is continuous.
?tn
11
3-n
?-D_, nA ?a nrb
rl rl
-) )a .r
Q- ?+l,r3
Thus,
ω ∈ {σu ⩾ t} ⇐⇒ σu (ω) ⩾ t
⇐⇒ inf{s ⩾ 0 ∶ Is (ω) > u} ⩾ t

It cts.
⇐⇒ It (ω) ⩽ u
⇐⇒ ω ∈ {It ⩽ u}
and
ω ∈ {τu > t} ⇐⇒ τu (ω) > t
⇐⇒ inf{s ⩾ 0 ∶ Is (ω) ⩾ u} > t

It cts.
⇐⇒ It (ω) < u
⇐⇒ ω ∈ {It < u}.
(b) We have
{τu ⩽ t} = {τu > t}c = {It < u} ∈ Ft
(a)
140
and
c c
{σu ⩽ t} = ⋂ {σu < t + k1 } = ⋂ {σu ⩾ t + k1 } = ⋂ {It+ 1 ⩽ u} ∈ ⋂ Ft+ 1 = Ft+ .
k k
k k k k
(c) Proof for σ: Clearly, σu ⩽ σu+ for all ⩾ 0. Thus, σu ⩽ lim↓0 σu+ .
In order to show that σu ⩾ lim↓0 σu+ , it is enough to check that
lim σu+ ⩾ t Ô⇒ σu ⩾ t. (*)

↓0
Indeed: if lim↓0 σu+ > σu , then there is some q such that lim↓0 σu+ > q > σu , and
this contradicts (*).
Let us show (*):
(a)
lim σu+ ⩾ t Ô⇒ ∀ < 0 ∶ It ⩽ u + Ô⇒ It ⩽ u Ô⇒ σu ⩾ t.
↓0
Proof for τ : Clearly, τu− ⩽ τu for all ⩾ 0. Thus, τu ⩾ lim↓0 τu− .

In order to show that τu ⩽ lim↓0 τu− , it is enough to check that
lim τu− ⩽ t Ô⇒ τu ⩽ t. (**)

↓0
Indeed: if lim↓0 τu− < τu , then there is some q such that lim↓0 τu− < q < τu and this
contradicts (**).
Let us show (**):
(a)
lim τu− ⩽ t Ô⇒ ∀ < 0 ∶ It ⩾ u − Ô⇒ It ⩾ u Ô⇒ τu ⩽ t.
↓0
The following picture motivates why we should expect that τu = σu− :
?tn
11
3-n
?-D_, nA ?a nrb
rl rl
-) )a .r
Q- ?+l,r3
141
(d) Clearly, σu− ⩽ τu for all > 0, i. e. σu− ⩽ τu .
We show now σu− ⩾ τu . For this it is enough to check that
σu− < t Ô⇒ τu ⩽ t. (***)
Indeed: if σu− < τu , then there is some q with σu− < q < τu contradicting (***).
Let us verify (***). We have
σu− < t ⇐⇒ lim σu− < t

↓0
Ô⇒ ∀ < 0 ∶ σu− < t

(a)
Ô⇒ ∀ < 0 ∶ It > u −
Ô⇒ It ⩾ u
Ô⇒ τu ⩽ t.
(e) Clear, since in this case It is continuously invertible and σ, τ are the left- and right-
continuous inverses.
∎∎
142
17 Itô’s formula
Problem 17.1. Solution: We try to identify the bits and pieces as parts of Itô’s formula. For
f (x) = e we get f (x) = f ′′ (x) = ex and so
x ′
t 1 t Bs
eBt − 1 = ∫ eBs dBs + ∫ e ds.
0 2 0
Thus,
1 t Bs
Xt = eBt − 1 − ∫ e ds.
2 0
With the same trick we try to find f (x) such that f ′ (x) = xex . A moment’s thought
2
1 x2
will do. Moreover f ′′ (x) = ex + 2x2 ex . This then gives
2 2
reveals that f (x) = 2e
1 Bt2 1 t 2 1 t 2 2
e − = ∫ Bs eBs dBs + ∫ (eBs + 2Bs2 eBs ) ds
2 2 0 2 0
and we see that
1 Bt2 t 2 2
Yt = (e − 1 − ∫ (eBs + 2Bs2 eBs ) ds) .
2 0
2
Note: the integrand Bs2 eBs is not of class L2T , thus we have to use a stopping technique
(as in step 4o of the proof of Itô’s formula or as in Chapter 16).
∎∎
Problem 17.2. Solution: For γ = 1 we get a telescoping sum

N N
T = tN − t0 = ∑ (tj − tj−1 ) = ∑ (tj − tj−1 )γ .
j=1 j=1
If γ = 1 + > 1 we get
N N
∑ (tj − tj−1 ) ⩽ ∣Π∣ ∑ (tj − tj−1 ) = ∣Π∣ T ÐÐÐ→ 0,
1+
j=1 j=1 ∣Π∣→0
and if γ = 1 − < 1 we have

N N
∑ (tj − tj−1 )
1−
⩾ ∣Π∣− ∑ (tj − tj−1 ) = ∣Π∣− T ÐÐÐ→ ∞.
j=1 j=1 ∣Π∣→0
∎∎
Problem 17.3. Solution: Let 0 = t0 < t1 < . . . < tN = T be a generic partition of [0, T ] and
write ∆j = Btj − Btj−1 . Then we get
4
( ∑∆2j ) = ∑ ∑ ∑ ∑ ∆2j ∆2k ∆2l ∆2m
j j k l m
143
= c1,1,1,1 ∑ ∑ ∑ ∑ ∆2j ∆2k ∆2l ∆2m

j<k<l<m
+ c1,1,2 ∑ ∑ ∑ ∆2j ∆2k (∆2l )2 + c1,2,1 ∑ ∑ ∑ ∆2j (∆2k )2 ∆2l + c2,1,1 ∑ ∑ ∑(∆2j )2 ∆2k ∆2l
j<k<l j<k<l j<k<l
+ c2,2 ∑ ∑(∆2j )2 (∆2k )2 + c1,3 ∑ ∑ ∆2j (∆2k )3 + c3,1 ∑ ∑(∆2j )3 ∆2k

j<k j<k j<k
+ c4 ∑(∆2j )4 .
j
By the scaling property E(∆2j )n = (tj − tj−1 )n E B12n = δjn E B12n where δj = tj − tj−1 . Using
the independent increments property we get
4
E [( ∑∆2j ) ] = ∑ ∑ ∑ ∑ E (∆2j ∆2k ∆2l ∆2m )
j j k l m
= c′1,1,1,1 ∑ ∑ ∑ ∑ δj δk δl δm
j<k<l<m
+ c′1,1,2 ∑ ∑ ∑ δj δk δl2 + c′1,2,1 ∑ ∑ ∑ δj δk2 δl + c′2,1,1 ∑ ∑ ∑ δj2 δk δl

j<k<l j<k<l j<k<l
+ c′2,2 ∑ ∑ δj2 δk2 + c′1,3 ∑ ∑ δj δk3 + c′3,1 ∑ ∑ δj3 δk

j<k j<k j<k
+ c′4 ∑ δj4
j
4
′
= c1,1,1,1 ( ∑ δj )
j
+ c′′1,1,2 ∑ ∑ ∑ δj δk δl2 + c′′1,2,1 ∑ ∑ ∑ δj δk2 δl + c′′2,1,1 ∑ ∑ ∑ δj2 δk δl

j<k<l j<k<l j<k<l
+ c′′2,2 ∑ ∑ δj2 δk2 + c′′1,3 ∑ ∑ δj δk3 + c′′3,1 ∑ ∑ δj3 δk

j<k j<k j<k
+ c′′4 ∑ δj4 .
j
Since ∑j δj = T and since we can estimate the terms containing powers of δj by, for
example,
∑ ∑ ∑ δj δk δl ⩽ ∣Π∣ ∑ ∑ ∑ δj δk δl ⩽ ∣Π∣ ∑ ∑ ∑ δj δk δl = ∣Π∣ T ÐÐÐ→ 0

2 3
j<k<l j<k<l j k l ∣Π∣→0
we get
4
E [( ∑ (Btj − Btj−1 ) ) ] ÐÐÐÐÐÐ→ c′1,1,1,1 T 4 .
2
j ∣Π∣→0
We will use this on page 252 (of Brownian Motion) when we estimate ∣J2 ∣:
N 2
2 2
∣J2 ∣ ⩽ max ∣g(ξl ) − g(Btl−1 )∣ [∑ (Btl − Btl−1 ) ] ,
2
1⩽l⩽N l=1
and taking now the Cauchy-Schwarz inequality gives

√ √
4
E [J2 ] ⩽ E ( max ∣g(ξl ) − g(Btl−1 )∣ )
2
E [(S2Π (B; t))4 ].
1⩽l⩽n
144
The second factor is, however, bounded by CT 2 , see the considerations from above, and
the L2 -convergence follows.
Alternative Solution: Let 0 = t0 < t1 < . . . < tn = T be a generic partition of [0, T ] and
write ∆j = Btj − Btj−1 . By the independence and stationarity of the increments, we have
for any ξ ∈ R
⎛ n ⎞ n n
1 n
E exp i ξ ∑ ∆2j = ∏ E exp (i ξ (tj − tj−1 )B12 ) = ∏ √ =∶ ∏ gj (ξ)
⎝ j=1 ⎠ j=1 j=1 1 − 2i ξ(tj − tj−1 ) j=1
using that B12 is χ21 -distributed. Obviously,
dk (tj − tj−1 )k
gj (ξ) = ck
dξ k (1 − 2i ξ(tj − tj−1 ))1/2+k
for some constants ck , k ⩾ 1, which do not depend on ξ, j. In particular,

RRR
dk RRR = ∣c ⋅ (t − t )k ∣ ⩽ ∣c ∣ ⋅ ∣Π∣k .
∣ k gj (ξ)∣ RRR k j j−1 k
dξ RRξ=0
From
⎡ n 4⎤ ⎡ ⎤
⎢⎛ 2⎞ ⎥ d4 ⎢⎢ ⎛ n 2 ⎞⎥⎥
⎢ ⎥
E ⎢ ∑ ∆j ⎥ = 4 ⎢E exp i ξ ∑ ∆j ⎥ ∣
⎢⎝j=1 ⎠ ⎥ dξ ⎢ ⎝ j=1 ⎠⎥ ξ=0
⎣ ⎦ ⎣ ⎦
we conclude, by applying Leibniz’ product rule,
⎡ n 4⎤
⎢⎛ ⎞ ⎥ ⎛ n
dαj ⎞ n
E ⎢⎢ ∑ ∆j ⎥⎥ = ∑ Cα ∏ αj gj (ξ) ∣ ⩽ C ∑ (Cα ∏(tj − tj−1 )αj ) ⩽ C∣Π∣4 n4 .
2
⎢⎝j=1 ⎠ ⎥ α∈Nn0 ⎝ j=1 dξ ⎠ ξ=0
⎣ ⎦ α∈Nn 0 j=1
∣α∣=4 ∣α∣=4 ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
⩽∣Π∣∣α∣ =∣Π∣4
By the definition of the mesh size we have n ⩽ T /∣Π∣; thus,

⎡ n 4⎤
⎢⎛ 2⎞ ⎥
E ⎢ ∑ ∆j ⎥⎥ ⩽ CT 4 .
⎢
⎢⎝j=1 ⎠ ⎥
⎣ ⎦
Note that the constant C does not depend on the partition Π. The rest of the proof
follows as in the preceding solution.
∎∎
(a) Assume first that f ∈ C1b . Let Π = {0 = t0 < t1 < . . . < tn = t} be any partition. We
have
n
Bt f (Bt ) = ∑ (Btl f (Btl ) − Btl−1 f (Btl−1 )).
l=1
Using
Bt f (Bt ) − Bs f (Bs )
145
= Bs (f (Bt ) − f (Bs )) + f (Bs )(Bt − Bs ) + (f (Bt ) − f (Bs ))(Bt − Bs )

= Bs (f (Bt ) − f (Bs )) + f (Bs )(Bt − Bs ) + f ′ (ξ)(Bt − Bs )2
with some intermediate value ξ between Bs and Bt , the identity follows.
Letting ∣Π∣ → 0 in this identity, we see that the left-hand side converges (it is con-
stant!) and the second and third term on the right converge, in probability, to
t t
∫ f (Bs ) dBs and ∫ f ′ (Bs ) ds,
0 0
respectively, cf. Lemma 17.4. Therefore, the first term has to converge, i. e.
t
∫ Bs df (Bs )
0
makes sense (and this is all we need!).
If f ′ is not bounded, we can use a stopping and cutting technique as in the proof of
Theorem 17.1 (step 4o ).
(b) This follows from (a) after having taken the limit.
(c) Applying (b) to f (x) = xn−1 gives
dBtn = d(Btn−1 Bt ) = Bt dBtn−1 + Btn−1 dBt + (n − 1)Btn−2 dt
and iterating this yields

1 1
dBtn = nBtn−1 dBt + n(n − 1)Btn−2 dt = npn−1 (Bt ) dBt + p′′n (Bt ) dt
2 2
where we use pn (x) = xn for the monomial of order n. Since the Itô integral is linear,
we get the claim for all polynomials of any order.
(d) This follows directly from Itô’s isometry:

T 2 T
E [∣∫ (g(s) − gn (s)) dBs ∣ ] = ∫ E [∣g(s) − gn (s)∣2 ] ds.
0 0
(e) We can assume that f has compact support, supp f ⊂ [−K, K], say. Otherwise, we
use the stopping and cutting technique from the proof (step 4o ) of Theorem 17.1,
to remove this assumption.
The classical version of Weierstraß’ approximation theorem tells us that f can be

uniformly approximated on [−K, K] by a sequence of polynomials (pfn )n , see e.g. [15,
Theorem 24.6]. We apply this theorem to f ′′ and observe that f ′ and f are still
uniformly approximated by the primitives Pnf ∶= ∫ pfn and Qfn ∶= ∫ Pnf = ∬ pfn .
The rest follows from the previous step (d) which allows us to interchange (stochastic)
integration and limits. (The Riemann part in Itô’s formula is clear, since we have
uniform convergence!).
∎∎
146
(a) Set F (x, y) = xy and G(t) = (f (t), g(t)).
Then f (t)g(t) = F ○ G(t). If we differentiate this using the chain rule we get
d
(F ○ G) = ∂x F ○ G(t) ⋅ f ′ (t) + ∂y F ○ G(t) ⋅ g ′ (t) = g(t) ⋅ f ′ (t) + f (t) ⋅ g ′ (t)
dt
(surprised?) and if we integrate this up we see
t t
F ○ G(t) − F ○ G(0) = ∫ f (s)g ′ (s) ds + ∫ g(s)f ′ (s) ds
0 0
t t
=∫ f (s) dg(s) + ∫ g(s) df (s).
0 0
Note: For the first equality we have to assume that f ′ , g ′ exist Lebesgue a.e. and
that their primitives are f and g, respectively. This is tantamount to saying that f, g
are absolutely continuous with respect to Lebesgue measure.
(b) f (x, y) = xy. Then ∂x f (x, y) = y, ∂y f (x, y) = x and ∂x ∂y f (x, y) = ∂y ∂x f (x, y) = 1

and ∂x2 f (x, y) = ∂y2 f (x, y) = 0. Thus, the two-dimensional Itô formula yields
t t
bt βt = ∫ bs dβs + ∫ βs dbs +
0 0
1 t 2 1 t 2 t
+ ∫ ∂x f (bs , βs ) ds + ∫ ∂y f (bs , βs ) ds + ∫ ∂x ∂y f (bs , βs ) d⟨b, β⟩s
2 0 2 0 0
t t
=∫ bs dβs + ∫ βs dbs + ⟨b, β⟩t .
0 0
If b á β we have ⟨b, β⟩ ≡ 0 (note our Itô formula has no mixed second derivatives!)
and we get the formula as in the statement. Otherwise we have to take care of
⟨b, β⟩. This is not so easy to calculate since we need more information on the joint
distribution. In general, we have
⟨b, β⟩t = lim ∑ (b(tj ) − b(tj−1 ))(β(tj ) − β(tj−1 )).

∣Π∣→0 tj ,tj−1
Where Π stands for a partition of the interval [0, t].
∎∎
Problem 17.6. Solution: Consider the two-dimensional Itô process Xt = (t, Bt ) with param-
eters
⎛0⎞ ⎛1⎞
σ≡ and b ≡ .
⎝1⎠ ⎝0⎠
Applying the Itô formula (17.14) we get
f (t, Bt ) − f (0, 0) = f (Xt ) − f (X0 )

t
=∫ (∂1 f (Xs )σ11 + ∂2 f (Xs )σ21 ) dBs
0
t 1
+∫ (∂1 f (Xs )b1 + ∂2 f (Xs )b2 + ∂2 ∂2 f (Xs )σ21
2
) ds
0 2
147
t t 1
=∫ ∂2 f (Xs ) dBs + ∫ (∂1 f (Xs )b1 + ∂2 ∂2 f (Xs )) ds
0 0 2
t ∂f t ∂f 1 ∂2f
=∫ (s, Bs ) dBs + ∫ ( (s, Bs ) + (s, Bs )) ds.
0 ∂x 0 ∂t 2 ∂x2
In the same way we obtain the d-dimensional counterpart:

Let (Bt1 , . . . , Btd )t⩾0 be a BMd and f ∶ [0, ∞)×Rd → R be a function of class C1,2 . Consider
the (d + 1)-dimensional Itô process Xt = (t, Bt1 , . . . , Btd ) with parameters
⎛1⎞
⎧
⎪ ⎜ ⎟
⎪
⎪1, if i = k + 1; ⎜0⎟
σ ∈ Rd+1×d , σik = ⎨ and b = ⎜ ⎟
⎜ ⎟.
⎪
⎪
⎪ ⎜⋮⎟
⎩0, else; ⎜ ⎟
⎝0⎠
The multidimensional Itô formula (17.14) yields
f (t, Bt1 , . . . , Btd ) − f (0, 0, . . . , 0)

= f (Xt ) − f (X0 )
⎡
t ⎢d+1
⎤
d ⎥ d+1 t 1 d+1 t d
= ∑ ∫ ⎢⎢ ∑ ∂j f (Xs )σjk ⎥⎥ dBsk + ∑ ∫ ∂j f (Xs )bj ds + ∑ ∫ ∂i ∂j f (Xs ) ∑ σik σjk ds
k=1 0 ⎢ ⎣ j=1
⎥
⎦ j=1 0 2 i,j=1 0 k=1
d d+1
t t 1 t
= ∑ ∫ ∂k+1 f (Xs ) dBsk + ∫ ∂1 f (Xs ) ds + ∑ ∫ ∂j ∂j f (Xs ) ds
k=1 0 0 2 j=2 0
d t ∂f t ∂f 1 d ∂2f
= ∑∫ (s, Bs1 , . . . , Bsd ) dBsk + ∫ ( (s, Bs1 , . . . , Bsd ) + ∑ (s, Bs1 , . . . , Bsd )) ds.
k=1 0 ∂xk 0 ∂t 2 k=1 ∂x2k
∎∎
Problem 17.7. Solution: Let Bt = (Bt1 , . . . , Btd ) be a BMd and f ∈ C1,2 ((0, ∞) × Rd , R) as in
Theorem 5.6. Then the multidimensional time-dependent Itô’s formula shown in Problem
17.6 yields
t
Mtf = f (t, Bt ) − f (0, B0 ) − ∫ Lf (s, Bs ) ds
0
t ∂ 1
= f (t, Bt ) − f (0, B0 ) − ∫ ( f (s, Bs ) + ∆x f (s, Bs )) ds
0 ∂t 2
d t ∂f
= ∑∫ (s, Bs1 , . . . , Bsd ) dBsk .
k=1 0 ∂xk
By Theorem 15.13 it follows that Mtf is a martingale (note that the assumption (5.5)
guarantees that the integrand is of class L2T !)
∎∎
Problem 17.8. Solution: First we show that Xt = et/2 cos Bt is a martingale. We use the
time-dependent Itô’s formula from Problem 17.6. Therefore, we set f (t, x) = et/2 cos x.
Then
∂f 1 ∂f ∂2f
(t, x) = et/2 cos x, (t, x) = −et/2 sin x, (t, x) = −et/2 cos x.
∂t 2 ∂x ∂x2
148
Hence we obtain
Xt = et/2 cos Bt = f (t, Bt ) − f (0, 0) + 1

t∂f t ∂f 1 ∂2f
=∫ (s, Bs ) dBs + ∫ ( (s, Bs ) + (s, Bs )) ds + 1
0 ∂x 0 ∂t 2 ∂x2
t t 1 1
= − ∫ es/2 sin Bs dBs + ∫ ( es/2 cos Bs − es/2 cos Bs ) ds + 1
0 0 2 2
t
= −∫ es/2 sin Bs dBs + 1,
0
and the claim follows from Theorem 15.13.
Analogously, we show that Yt = (Bt + t)e−Bt −t/2 is a martingale. We set f (t, x) = (x +

t)e−x−t/2 . Then
∂f 1
(t, x) = e−x−t/2 − (x + t)e−x−t/2 ,
∂t 2
∂f −x−t/2
(t, x) = e − (x + t)e−x−t/2 ,
∂x
∂f
(t, x) = −2e−x−t/2 + (x + t)e−x−t/2 .
∂x2
By the time-dependent Itô’s formula we have
Yt = (Bt + t)e−Bt −t/2

= f (t, Bt ) − f (0, 0)
t
=∫ (e−Bs −s/2 − (Bs + s)e−Bs −s/2 ) dBs +
0
t 1 1
+∫ (e−Bs −s/2 − (Bs + s)e−Bs −s/2 + (−2e−Bs −s/2 + (Bs + s)e−Bs −s/2 )) ds
0 2 2
t
=∫ (e−Bs −s/2 − (Bs + s)e−Bs −s/2 ) dBs .
0
Again, from Theorem 15.13 we deduce that Yt is a martingale.
∎∎
(a) The stochastic integrals exist if bs /rs and βs /rs are in L2T . As ∣bs /rs ∣ ⩽ 1 we get
T T
∥b/r∥2L2 (λT ⊗P) = ∫ [E (∣bs /rs ∣2 )] ds ⩽ ∫ 1ds = T < ∞.
0 0
Since bs /rs is adapted and has continuous sample paths, it is progressive and so an
element of L2T . Analogously, ∣βs /rs ∣ ⩽ 1 implies βs /rs ∈ L2T .
(b) We use Lévy’s characterization of a BM1 , Theorem 9.12 or 18.5. From Theorem
15.13 it follows that
t t
• t ↦ ∫0 bs /rs dbs , t ↦ ∫0 βs /rs dβs are continuous; thus t ↦ Wt is a continuous
process.
t t
• ∫0 bs /rs dbs , ∫0 βs /rs dβs are square integrable martingales, and so is Wt .
149
• the quadratic variation is given by
⟨W ⟩t = ⟨b/r ● b⟩t + ⟨β/r ● β⟩t

t t
=∫ b2s /rs2 ds + ∫ βs2 /rs2 ds
0 0
t b2s + βs2
=∫ ds
0 rs2
t
=∫ ds = t,
0
i. e. (Wt2 − t)t⩾0 is a martingale.

Therefore, Wt is a BM1 .
Note, that the above processes can be used to calculate Lévy’s stochastic area formula,
see Protter [11, Chapter II, Theorem 43]
∎∎
Problem 17.10. Solution: The function f = u + iv is analytic, and as such it satisfies the
Cauchy–Riemann equations, see e.g. Rudin [14, Theorem 11.2],
ux = vy and uy = −vx .
First, we show that u(bt , βt ) is a BM1 . Therefore we apply Itô’s formula
u(bt , βt ) − u(b0 , β0 )
t t 1 t
=∫ ux (bs , βs ) dbs + ∫ uy (bs , βs ) dβs + ∫ (uxx (bs , βs ) + uyy (bs , βs )) ds
0 0 2 0
t t
=∫ ux (bs , βs ) dbs + ∫ uy (bs , βs ) dβs ,
0 0
where the last term cancels as uxx = vyx and uyy = −vxy . Theorem 15.13 implies
t t
• t ↦ u(bt , βt ) = ∫0 ux (bs , βs ) dbs + ∫0 uy (bs , βs ) dβs is a continuous process.
t t
• ∫0 ux (bs , βs ) dbs , ∫0 uy (bs , βs ) dβs are square integrable martingales, and so u(bt , βt )
is a square integrable martingale.
• the quadratic variation is given by
⟨u(b, β)⟩t = ⟨ux (b, β) ● b⟩t + ⟨uy (b, β) ● β⟩t

t t t
=∫ u2x (bs , βs ) ds + ∫ u2y (bs , βs ) ds = ∫ 1 ds = t,
0 0 0
i. e. (u2 (bt , βt ) − t)t⩾0 is a martingale.

Due to Lévy’s characterization of a BM1 , Theorem 9.12 or 18.5, we know that u(bt , βt )
is a BM1 . Analogously, we see that v(bt , βt ) is also a BM1 . Just note that, due to the
Cauchy–Riemann equations we get from u2x + u2y = 1 also vy2 + vx2 = 1.
The quadratic covariation is (we drop the arguments, for brevity):
1
⟨u, v⟩t = (⟨u + v⟩t − ⟨u − v⟩t )
4
150
1 t t t t
= (∫ (ux + vx )2 ds + ∫ (uy + vy )2 ds − ∫ (ux − vx )2 ds − ∫ (uy − vy )2 ds)
4 0 0 0 0
t
= ∫ (ux vx + uy vy ) ds
0
t
= ∫ (−vy uy + uy vy ) ds = 0.
0
As an abbreviation we write ut = u(bt , βt ) and vt = v(bt , βt ). Applying Itô’s formula to

the function g(ut , vt ) = ei(ξut +ηvt ) and s < t yields
t t 1 2 t
g(ut , vt ) − g(us , vs ) = iξ ∫ g(ur , vr ) dur + iη ∫ g(ur , vr ) dvr − (ξ + η 2 ) ∫ g(ur , vr ) dr,
s s 2 s
as the quadratic covariation ⟨u, v⟩t = 0. Since ∣g∣ ⩽ 1 and since g(ut , vt ) is progressive,
the integrand is in L2T and the above stochastic integrals exist. From Theorem 15.13 we
deduce that
t t
E (∫ g(ur , vr ) dur 1F ) = 0 and E (∫ g(ur , vr ) dvr 1F ) = 0.
s s
for all F ∈ σ(ur , vr ∶ r ⩽ s) =∶ Fs . If we multiply the above equality by e−i(ξus +ηvs ) 1F and
take expectations, we get
1 t
E (g(ut − us , vt − vs )1F ) = P(F ) − (ξ 2 + η 2 ) ∫ E (g(ur − us , vr − vs )1F ) dr.
2 0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
=Φ(t) =Φ(r)
Since this integral equation has a unique solution (use Gronwall’s lemma, Theorem A.47),
we get
E(ei(ξ(ut −us )+η(vt −vs )) 1F ) = P(F ) e− 2 (t−s)(ξ

1 2 +η 2 )
= P(F ) e− 2 (t−s)ξ e− 2 (t−s)η

1 2 1 2
= P(F ) E(eiξ(ut −us ) ) E(eiη(vt −vs ) ).
From this we deduce with Lemma 5.4 that (u(bt , βt ), v(bt , βt )) is a BM2 .
Note that the above calculation is essentially the proof of Lévy’s characterization theorem.
Only a few modifications are necessary for the proof of the multidimensional version, see
e.g. Karatzas, Shreve [9, Theorem 3.3.16].
∎∎
t t
Problem 17.11. Solution: Let Xt = ∫0 σ(s) dBs + ∫0 b(s) ds be an d-dimensional Itô process.
Assuming that f = u + iv and thus u = Re f = 21 f + 12 f¯ and v = Im f = 2i
1 1 ¯
f + 2i f are
C2 -functions, we may apply the real d-dimensional Itô formula (17.14) to the functions
u, v ∶ Rd → R,
f (Xt ) − f (X0 )
= u(Xt ) − u(X0 ) + i(v(Xt ) − v(X0 ))
151
t t 1 t
=∫ ∇u(Xs )⊺ σ(s) dBs + ∫ ∇u(Xs )⊺ b(s) ds + ⊺ 2
∫ trace(σ(s) D u(Xs )σ(s)) ds
0 0 2 0
t t 1 t
+ i (∫ ∇v(Xs )⊺ σ(s) dBs + ∫ ∇v(Xs )⊺ b(s) ds + ∫ trace(σ(s)⊺ D2 v(Xs )σ(s)) ds)
0 0 2 0
t t 1 t
= ∫ ∇f (Xs )⊺ σ(s) dBs + ∫ ∇f (Xs )⊺ b(s) ds + ∫ trace(σ(s)⊺ D2 f (Xs )σ(s)) ds,
0 0 2 0
by the linearity of the differential operators and the (stochastic) integral.
∎∎
(a) By definition we have supp χ ⊂ [−1, 1] hence it is obvious that for χn (x) ∶= nχ(nx)
we have supp χn ⊂ [−1/n, 1/n]. Substituting y = nx we get
1/n 1/n 1
∫ χn (x) dx = ∫ nχ(nx) dx = ∫ χ(y) dy = 1
−1/n −1/n −1
(b) For derivatives of convolutions we know that ∂(f ⋆ χn ) = f ⋆ (∂χn ). Hence we obtain
∣∂ k fn (x)∣ = ∣f ⋆ (∂ k χn )(x)∣
= ∣∫ f (y)∂ k χn (x − y) dy∣
B(x,1/n)
⩽ sup ∣f (y)∣ ∫ n ∣∂ k χ(n(x − y))∣ dy

y∈B(x,1/n) R
= sup ∣f (y)∣ ∫ nk ∣∂ k χ(z)∣ dz

y∈B(x,1/n) R
= sup ∣f (y)∣ nk ∥∂ k χ∥L1 ,

y∈B(x,1/n)
where we substituted z = n(y − x) in the penultimate step.
(c) For x ∈ R we have
∣f ⋆ χn (x) − f (x)∣ = ∣∫ (f (y) − f (x))χn (x − y) dy∣

R
⩽∣ sup ∣f (y) − f (x)∣ ⋅ ∥χ∥L1

y∈B(x,1/n)
= sup ∣f (y) − f (x)∣.

y∈B(x,1/n)
This shows that limn→∞ ∣f ⋆ χn (x) − f (x)∣ = 0, i. e. limn→∞ f ⋆ χn (x) = f (x), at all x
where f is continuous.
(d) Using the above result and taking the supremum over all x ∈ R we get
sup ∣f ⋆ χn (x) − f (x)∣ ⩽ sup sup ∣f (y) − f (x)∣.

x∈R x∈R y∈B(x,1/n)
Thus limn→∞ ∥f ⋆ χn − f ∥∞ = 0 whenever the function f is uniformly continuous.
∎∎
152
Problem 17.13. Solution: We follow the hint and use Lévy’s characterization of a BM1 ,
Theorem 9.12 or 18.5.
• t ↦ βt is a continuous process.
• the integrand sgn Bs is bounded, hence it is in L2T for any T > 0.
• by Theorem 15.13 βt is a square integrable martingale
• by Theorem 15.13 the quadratic variation is given by

● t t
⟨β⟩t = ⟨∫ sgn(Bs ) dBs ⟩ = ∫ (sgn(Bs ))2 ds = ∫ ds = t,
0 t 0 0
i. e. (βt2 − t)t⩾0 is also a martingale.
Thus, β is a BM1 .
∎∎
Problem 17.14. Solution: 1o — Consider the Itô processes
dXj(t) = σj (t) dBt + bj (t) dt, Xj (0) = 0, (j = 1, 2)
where σj ∈ L2T and bj is bounded. Then we get from the two-dimensional Itô’s formula
t t t
X1 (t)X2 (t) = ∫ σ1 (s)σ2 (s) ds + ∫ X1 (s) dX2 (s) + ∫ X2 (s) dX1 (s).
0 0 0
Taking expectations, the martingale parts containing dB(s) vanish, so

t t t
E (X1 (t)X2 (t)) = E ∫ σ1 (s)σ2 (s) ds + E ∫ X1 (s) b2 (s) ds + E ∫ X2 (s) b1 (s) ds.
0 0 0
2o — Now let X1 = f ● B and X2 = Φ(g ● B) with Φ ∈ C2b (R). Then, by Itô’s formula
(17.1),
dX1 (t) = f (t) dBt ,

1 ′′
dX2 (t) = dΦ(g ● Bt ) = Φ′ (g ● Bt ) g(t) dBt + Φ (g ● Bt ) g 2 (t) dt.
2
3o — Combining steps 1o and 2o gives

t
f (s)g(s) Φ′ (∫0 g(r) dBr ) ds
t t s
E (∫0 f (r) dBr ⋅ Φ (∫0 g(r) dBr ) ) = E ( ∫
0
1 t s ′′ s
+ ∫ ∫ f (r) dBr Φ (∫0 g(r) dBr ) g (s) ds)
2
2 0 0
∎∎
L2 (λT ⊗P) L2 (λT ⊗P)
Problem 17.15. Solution: Let σΠ , bΠ ∈ ET such that σΠ ÐÐÐÐÐÐ→ σ, bΠ ÐÐÐÐÐÐ→ b. By the
∣Π∣→0 ∣Π∣→0
Chebyshev inequality, Doob’s maximal inequality, and Itô’s isometry, we have
t t
P ( sup ∣ ∫ g(XsΠ )σΠ (s) dBs − ∫ g(Xs )σ(s) dBs ∣ > )
t⩽T 0 0
153
2
1 t
⩽ E (sup ∣∫ (g(Xs
Π
)σΠ
(s) − g(Xs )σ(s)) dBs ∣ )
2 t⩽T 0
2
4 T
⩽ E (∣ ∫ (g(Xs
Π
)σ Π
(s) − g(Xs )σ(s)) dB s ∣ )
2 0
4 T
= 2 E (∫ ∣g(XsΠ )σΠ (s) − g(Xs )σ(s)∣2 ds) .
0
From
g(XsΠ )σΠ (s) − g(Xs )σ(s) = g(XsΠ )(σΠ (s) − σ(s)) − σ(s)(g(Xs ) − g(XsΠ ))
and the inequality (a + b)2 ⩽ 2a2 + 2b2 we conclude

t t
P (sup ∣∫ g(XsΠ )σΠ (s) dBs − ∫ g(Xs )σ(s) dBs ∣ > )
t⩽T 0 0
8 T 8 T
⩽ E (∫ g(Xs )
Π 2
∣σ Π
(s) − σ(s)∣ 2
ds) + E (∫ σ(s)2 ∣g(XsΠ ) − g(Xs )∣2 ds)
2 0 2 0
=∶ I1 + I2 .
L2 (λT ⊗P)
Since g is bounded and σΠ ÐÐÐÐÐÐ→ σ, it follows that I1 → 0 as ∣Π∣ → 0. For the second
∣Π∣→0
term we note that, by Lemma 17.5,
P
sup ∣g(XsΠ ) − g(Xs )∣ ÐÐÐ→ 0.
s⩽T ∣Π∣→0
Hence, by Vitali’s convergence theorem, I2 → 0 as ∣Π∣ → 0.
A similar, but simpler, calculation shows

t t
P (sup ∣∫ g(XsΠ )bΠ (s) ds − ∫ g(Xs )b(s) ds∣ > ) ÐÐÐ→ 0.
t⩽T 0 0 ∣Π∣→0
Consequently,
t t
P (sup ∣∫ g(XsΠ )dXsΠ − ∫ g(Xs ) dXs ∣ > ) ÐÐÐ→ 0.
t⩽T 0 0 ∣Π∣→0
∎∎
154
18 Applications of Itô’s formula
Problem 18.1. Solution: Lemma. Let (Bt , Ft )t⩾0 be a BMd , f = (f1 , . . . , fd ), fj ∈ L2P (λT ⊗P)
for all T > 0, and assume that ∣fj (s, ω)∣ ⩽ C for some C > 0 and all s ⩾ 0, 1 ⩽ j ⩽ d, and
ω ∈ Ω. Then
⎛d t 1 d t ⎞
exp ∑ ∫ fj (s) dBsj − ∑ ∫ fj2 (s) ds , t ⩾ 0, (18.1)
⎝j=1 0 2 j=1 0 ⎠
is a martingale for the filtration (Ft )t⩾0 .
t t
Proof. Set Xt = ∑dj=1 ∫0 fj (s) dBsj − 21 ∑dj=1 ∫0 fj2 (s) ds. Itô’s formula, Theorem 17.7, yields
d t 1 d t 1 d t
eXt − 1 = ∑ ∫ eXs fj (s) dBsj − ∑ ∫ e s fj (s) ds + ∑ ∫ e s fj (s) ds
X 2 X 2
j=1 0 2 j=1 0 2 j=1 0
d d
t s 1 d s
= ∑∫ exp ( ∑ ∫ fk (r) dBrk − ∑ ∫ fk (r) dr) fj (s) dBs
2 j
j=1 0 k=1 0 2 k=1 0
d t d s 1 s 2
= ∑∫ ∏ exp (∫ fk (r) dBrk − ∫ f (r) dr) fj (s) dBs .
j
j=1 0 k=1 0 2 0 k
If we can show that the integrand is in L2P (λT ⊗ P) for every T > 0, then Theorem 15.13
applies and shows that the stochastic integral, hence eXt , is a martingale.
We will see that we can reduce the d-dimensional setting to a one-dimensional setting.
The essential step in the proof is the analogue of the estimate on page 250, line 6 from
above. In the d-dimensional setting we have for each k = 1, . . . , d
T 2
fj (r) dBrj − 12 ∑dj=1 ∫0T fj2 (r) dr T
fj (r) dBrj
E [∣e∑j=1 ∫0
d
fk (T )∣ ] ⩽ C 2 E [e2 ∑j=1 ∫0
d
]
⎡d ⎤
⎢ 2 ∫0T fj (r) dBrj ⎥
⎢
= C E ⎢∏ e
2 ⎥
⎥
⎢j=1 ⎥
⎣ ⎦
d T 1/d
fj (r) dBrj
⩽ C 2 ∏ (E [e2d ∫0 ]) .
j=1
In the last step we used the generalized Hölder inequality

n n 1/pk
∫ ∏ φk dµ ⩽ ∏ (∫ ∣φk ∣ dµ) ∀(p1 , . . . , pn ) ∈ [1, ∞)n ∶ ∑nk=1 =1
pk 1
pk
k=1 k=1
with n = d and p1 = . . . = pd = d. Now the one-dimensional argument with dfj playing the
role of f shows (cf. page 250, line 9 from above)
T 2 d 1/d
fj (r) dBrj − 12 ∑dj=1 ∫0T fj2 (r) dr T
fj (r) dBrj
E [∣e∑j=1 ∫0
d
fk (T )∣ ] ⩽ C 2 ∏ (E [e2d ∫0 ])
j=1
155
2T
⩽ C 2 e2dC < ∞.
∎∎
Problem 18.2. Solution: As for a Brownian motion one can see that the independent incre-
ments property of a Poisson process is equivalent to saying that Nt − Ns á FsN for all s ⩽ t,
cf. Lemma 2.14 or Section 5.1. Thus, we have for s ⩽ t
E(Nt − t ∣ FsN ) = E(Nt − Ns − (t − s) ∣ FsN ) + E(Ns − s ∣ FsN )

Nt −Ns áFs
N
= E(Nt − Ns − (t − s)) + Ns − s
pull out
Nt −Ns ∼Nt−s
= E(Nt − Ns ) − (t − s) + Ns − s
= E(Nt−s ) − (t − s) + Ns − s
= Ns − s.
Observe that
2
(Nt − t)2 − t = (Nt − Ns − (t − s) + (Ns − s)) − t
2
= (Nt − Ns − (t − s)) + (Ns − s)2 + 2(Ns − s)(Nt − Ns − t + s) − t.
Thus,
((Nt − t)2 − t) − ((Ns − s)2 − s)

2
= (Nt − Ns − (t − s)) + 2(Ns − s)(Nt − Ns − t + s) − (t − s).
Now take E(⋯ ∣ FsN ) in the last equality and observe that Nt − Ns á Fs . Then
E [((Nt − t)2 − t) − ((Ns − s)2 − s) ∣ FsN ]

Nt −Ns áFs
N 2
= E [(Nt − Ns − (t − s)) ] + 2 E [(Ns − s)(Nt − Ns − t + s) ∣ FsN ] − (t − s)
Nt −Ns ∼Nt−s 2
= E [(Nt−s − (t − s)) ] + 2(Ns − s) E [(Nt − Ns − t + s) ∣ FsN ] − (t − s)
pull out
Nt −Ns áFs
N
= V Nt−s + 2(Ns − s) E(Nt − Ns − t + s) − (t − s)
= t − s + 2(Ns − s) ⋅ 0 − (t − s) = 0.
Since t ↦ Nt is not continuous, this does not contradict Theorem 18.5.
∎∎
Problem 18.3. Solution: We want to use Lévy’s characterization, Theorem 18.5. Clearly,
t ↦ Wt is continuous and W0 = 0. Set Ftb = σ(br ∶ r ⩽ t), Ftβ = σ(βr ∶ r ⩽ t) and
FtW = σ(br , βr ∶ r ⩽ t) = σ(Ftb , Ftβ ), and
√
λ = σ1 / σ12 + σ22 ,
√
µ = σ2 / σ12 + σ22 .
156
We have
E(Wt ∣ FsW ) = E(λbt ∣ FsW ) + E(µβt ∣ FsW )

Ftb áF
β
= t
λ E(bt ∣ Fsb ) + µ E(βt ∣ Fsβ ) = λbs + µβs = Ws .
proving that (Wt , FtW ) is a martingale. Similarly one shows that (Wt2 − t, FtW )t⩾0 is a
martingale. Now Theorem 18.5 applies.
∎∎
Problem 18.4. Solution: Solution 1: Note that

n
Q(W (tj ) ∈ Aj , ∀j = 1, . . . , n) = ∫ ∏ 1Aj (W (tj )) dQ
j=1
n
= ∫ ∏ 1Aj (B(tj ) − ξtj ) eξB(T )− 2 ξ
1 2
T
dP.
j=1
1 2
By the tower property and the fact that eξB(t)− 2 ξ t
is a martingale we get
n
ξB(T )− 12 ξ 2 T
∫ ∏1Aj (B(tj ) − ξtj ) e dP
j=1
⎡ ⎤
⎢ ⎛n ⎞⎥⎥
⎢
= E ⎢E ∏ 1Aj (B(tj ) − ξtj ) e ξB(T )− 12 ξ 2 T
∣ Ftn ⎥
⎢ ⎝j=1 ⎠⎥
⎣ ⎦
⎡n ⎤
⎢ ⎥
= E ⎢⎢∏ 1Aj (B(tj ) − ξtj ) E (eξB(T )− 2 ξ T ∣ Ftn )⎥⎥
1 2
⎢j=1 ⎥
⎣ ⎦
⎡n ⎤
⎢ ⎥
= E ⎢⎢∏ 1Aj (B(tj ) − ξtj ) eξB(tn )− 2 ξ tn ⎥⎥
1 2
⎢j=1 ⎥
⎣ ⎦
⎡ ⎤
⎢ ⎛n ⎞⎥
= E ⎢⎢E ∏ 1Aj (B(tj ) − ξtj ) eξB(tn )− 2 ξ tn ∣ Ftn−1 ⎥⎥
1 2
⎢ ⎝j=1 ⎠⎥
⎣ ⎦
⎡ n−1
⎢
= E ⎢⎢ ∏ 1Aj (B(tj ) − ξtj ) eξB(tn−1 )− 2 ξ tn−1 ×
1 2
⎢ j=1
⎣
⎤
⎥
× E (1An (B(tn ) − ξtn ) e ξ(B(tn )−B(tn−1 ))− 12 ξ 2 (tn −tn−1 )
∣ Ftn−1 ) ⎥⎥
⎥
⎦
Now, since B(tn ) − B(tn−1 ) á Ftn−1 we get
E (1An (B(tn ) − ξtn ) eξ(B(tn )−B(tn−1 ))− 2 ξ (tn −tn−1 )

1 2
∣ Ftn−1 )
= E (1An ((B(tn ) − B(tn−1 )) − ξ(tn − tn−1 ) + B(tn−1 ) − ξtn−1 )×
× eξ(B(tn )−B(tn−1 ))− 2 ξ (tn −tn−1 )

1 2
∣ Ftn−1 )
= E (1An ((B(tn ) − B(tn−1 )) − ξ(tn − tn−1 ) + y)×
157
× eξ(B(tn )−B(tn−1 ))− 2 ξ (tn −tn−1 )

1 2
)∣
y=B(tn−1 )−ξtn−1
A direct calculation now gives
E (1An ((B(tn ) − B(tn−1 )) − ξ(tn − tn−1 ) + y)eξ(B(tn )−B(tn−1 ))− 2 ξ (tn −tn−1 )
1 2
)
= E (1An (B(tn − tn−1 ) − ξ(tn − tn−1 ) + y)eξB(tn −tn−1 )− 2 ξ (tn −tn−1 )

1 2
)
1 1 2
(tn −tn−1 ) − 2(t 1
n −tn−1 )
x2
=√ ∫ 1An (x − ξ(tn − tn−1 ) + y)e 2
ξx− ξ
e dx
2π(tn − tn−1 )
1 − 2(t 1
n −tn−1 )
(x−ξ(tn −tn−1 ))2
=√ ∫ 1An (x − ξ(tn − tn−1 ) + y)e dx
2π(tn − tn−1 )
1 − 2(t 1
n −tn−1 )
z2
=√ ∫ 1An (z + y)e dz
2π(tn − tn−1 )
= E 1An (B(tn ) − B(tn−1 ) + y)
In the next iteration we get
E 1An ((B(tn ) − B(tn−1 )) + (B(tn−1 ) − B(tn−2 ) + y))1An−1 ((B(tn−1 ) − B(tn−2 ) + y))
= E 1An ((B(tn ) − B(tn−2 ) + y))1An−1 ((B(tn−1 ) − B(tn−2 ) + y))
etc. and we finally arrive at

n
j
Q(W (tj ) ∈ Aj , ∀j = 1, . . . , n) = E ∏ 1Aj ( ∑k=1 (B(tk ) − B(tk−1 ))).
j=1
Solution 2: As in the first part of Solution 1 we see that we can assume that T = tn . Since
we know the joint distribution of (B(t1 ), . . . , B(tn )), cf. (2.10b), we get (using x0 = t0 = 0)
Q(W (t1 ) ∈ A1 , . . . , W (tn ) ∈ An )

n
= ∫ ∏ 1Aj (B(tj ) − ξtj ) eξB(tn )− 2 ξ
1 2
tn
dP
j=1
n (xj −xj−1 )2
ξxn − 21 ξ 2 tn − 2 ∑j=1
1 n
tj −tj−1
dx1 . . . dxn
= ∫…∫ ∏ 1Aj (xj − ξtj ) e e √
j=1 (2π)n/2 ∏nj=1 tj − tj−1
⎡ (x −x )2 ⎤
⎛ n ⎢⎢ − 12 tj −tj−1 ⎥⎞ ∑n (ξ(xj −xj−1 )− 1 ξ 2 (tj −tj−1 ))
⎥ e j=1 dx1 . . . dxn
= ∫…∫ ∏ ⎢1Aj (xj − ξtj ) e j j−1
⎥ 2 √
⎝j=1 ⎢ ⎥⎠ (2π) ∏nj=1 tj − tj−1
n/2
⎣ ⎦
n ⎡⎢ (x −x )2 ⎤
− 12 tj −tj−1 +ξ(xj −xj−1 )− 21 ξ 2 (tj −tj−1 ) ⎥
⎢ ⎥ dx1 . . . dxn
= ∫…∫ ∏ ⎢1Aj (xj − ξtj ) e j j−1
⎥ √
j=1 ⎢ ⎥ (2π) ∏nj=1 tj − tj−1
n/2
⎣ ⎦
n ⎡⎢ − 1
2⎤
((xj −xj−1 )+ξ(tj −tj−1 )) ⎥⎥ dx1 . . . dxn
= ∫…∫ ∏ ⎢⎢1Aj (xj − ξtj ) e 2(tj −tj−1 ) ⎥ √
j=1 ⎢ ⎥ (2π) ∏nj=1 tj − tj−1
n/2
⎣ ⎦
n − 1
(zj −zj−1 )2 dz1 . . . dzn
= ∫…∫ ∏ [1Aj (zj ) e 2(tj −tj−1 ) ] √
j=1 (2π)n/2 ∏nj=1 tj − tj−1
= P (B(t1 ) ∈ A1 , . . . , B(tn ) ∈ An ).
∎∎
158
P(Bt + αt ⩽ x, sup(Bs + αs) ⩽ y)

s⩽t
= ∫ 1(−∞,x] (Bt + αt)1(−∞,y] ( sups⩽t (Bs + αs)) d P

1
= ∫ 1(−∞,x] (Bt + αt)1(−∞,y] ( sups⩽t (Bs + αs)) dQ
βt
where Q = βt ⋅ P with βt = exp ( − αBt − 12 α2 t)
= ∫ 1(−∞,x] (Bt + αt)1(−∞,y] ( sups⩽t (Bs + αs)) eαBt + 2 α t dQ

1 2
= ∫ 1(−∞,x] (Bt + αt)1(−∞,y] ( sups⩽t (Bs + αs)) eα(Bt +αt) e− 2 α t dQ

1 2
e− 2 α t ∫ 1(−∞,x] (Wt )1(−∞,y] ( sups⩽t Ws ) eαWt dQ

1 2
=
Girsanov
= e− 2 α t ∫
1 2
1(−∞,x] (ξ) eαξ Q(Wt ∈ dξ, sups⩽t Ws ⩽ y).
Rd
where (Ws )s⩽t is a Brownian motion for the probability measure Q.

From Solution 2 of Problem 6.8 (or with Theorem 6.18) we have
Q( sups⩽t Wt < y, Wt ∈ dξ) = lim Q( inf s⩽t Ws > a, sups⩽t Wt < y, Wt ∈ dξ)
a→−∞
dξ ξ2 (ξ−2y)2
= √ [e− 2t − e− 2t ]
(6.19)
2πt
and we get the same result for Q( sups⩽t Wt ⩽ y, Wt ∈ dξ). Thus,
P(Bt + αt ⩽ x, sup(Bs + αs) ⩽ y)

s⩽t
x 1 ξ2 (ξ−2y)2
− 21 t α2
=∫ eαξ e √ (e− 2t − e− 2t ) dξ
−∞ 2πt
1 x (ξ−αt)2 (ξ−2y−αt)2
− 2t 2αy −
=√ ∫ (e − e e 2t ) dξ
2πt −∞
√ √
x−αt x−2y−αt
1 − z2
2
e2αy z2
=√ − √ e− 2 dz
t t
∫ e dz ∫
2πt −∞ 2πt −∞
√ ) − e2αy Φ( x−2y−αt
= Φ( x−αt √ ).
t t
∎∎

(a) Since Xt has continuous sample paths we find that
̂
τ b = inf {t ⩾ 0 ∶ Xt ⩾ b}.
Moreover, we have
{̂
τ b ⩽ t} = { sups⩽t Xs ⩾ b}.
Indeed,
ω ∈ { sups⩽t Xs ⩾ b} Ô⇒ ∃s ⩽ t ∶ Xs (ω) ⩾ b (continuous paths!)
159
Ô⇒ ̂
τ b (ω) ⩽ t
Ô⇒ ω ∈ {̂
τ b ⩽ t},
and so {̂
τ b ⩽ t} ⊃ { sups⩽t Xs ⩾ b}. Conversely,
ω ∈ {̂
τ b ⩽ t} Ô⇒ ̂
τ b (ω) ⩽ t
Ô⇒ Xτ̂ (ω) ⩾ b, ̂
τ b (ω) ⩽ t
b (ω)
Ô⇒ sup Xs (ω) ⩾ b
s⩽t
Ô⇒ ω ∈ { sups⩽t Xs ⩾ b},
and so {̂
τ b ⩽ t} ⊂ { sups⩽t Xs ⩾ b}.
By the previous problem, Problem 18.5, P(sups⩽t Xs = b) = 0. This means that
P (̂
τ b > t) = P ( sups⩽t Xs < b)
= P ( sups⩽t Xs ⩽ b)
= P (Xt ⩽ b, sups⩽t Xs ⩽ b)
√ ) − e2αb Φ( −b−αt
= Φ( b−αt √ )
Prob.
5 t t
√ √
= Φ( √bt − α t) − e2αb Φ( − √bt − α t).
Differentiating in t yields
d √ √
− P (̂
τ b > t) = e2αb ( 2tb√t − 2√
α
t
) Φ ′
( − √b − α t) + ( b√ + √
t 2t t
α
2 t
) Φ ′ √b
( t
− α t)
dt
1 (b+αt)2 (b−αt)2
− 2t − 2t
= √ (e2αb ( 2tb√t − 2√ α
t
) e + ( b√
2t t
+ α
√
2 t
) e )
2π
1 (b−αt)2 (b−αt)2
− 2t − 2t
= √ (( 2tb√t − 2√ α
t
) e + ( b√
2t t
+ α
√
2 t
) e )
2π
1 2b − (b−αt)2
=√ √ e 2t
2π 2t t
b (b−αt)2
= √ e− 2t
t 2πt
(b) We have seen in part a) that
P (̂ √ ) − e2αb Φ( −b−αt
τ b > t) = Φ( b−αt
t
√ )
t
⎧
⎪
⎪
⎪
⎪Φ(−∞) − e2αb Φ(−∞) = 0 if α > 0
⎪
⎪
⎪
ÐÐ→ ⎨Φ(0) − e0 Φ(0) = 0 if α = 0
t→∞ ⎪
⎪
⎪
⎪
⎪
⎪
⎩Φ(∞) − e Φ(∞) = 1 − e
⎪ if α < 0
2αb 2αb
Therefore, we get
⎧
⎪
⎪
⎪1 if α ⩾ 0
̂
P (τ b < ∞) = ⎨
⎪
⎪
⎪ 2αb
if α < 0.
⎩e
160
∎∎
Problem 18.7. Solution: Basically, the claim follows from Lemma 18.10. Indeed, if we set
n
g(s) ∶= ∑ (ξj + . . . + ξn )1[tj−1 ,tj ) (s),
j=1
then
T n
∫ g(s) dBs = ∑ (ξj + . . . + ξn )(Btj − Btj−1 )
0 j=1
n
= ∑ ((ξj + . . . + ξn ) − (ξj+1 + . . . + ξn ))Btj
j=1
n
= ∑ ξj Btj .
j=1
= ei ∑j=1 ξj Btj is in HT2 ⊕ iHT2 .

T n
Lemma 18.10 shows that ei ∫0 g(s) dBs
If you want to be a bit more careful, you should treat the real and imaginary parts of
T T T
exp (i ∫0 g(s) dBs ) = cos (∫0 g(s) dBs ) + i sin (∫0 g(s) dBs ) separately. Let us do this for
the real part.
We apply the two-dimensional Itô-formula (17.14) to f (x, y) = cos(x)ey/2 and the process
t t
(Xt , Yt ) = (∫0 g(s) dBs , ∫0 g 2 (s) ds): Since
∂x f (x, y) = − sin(x)ey/2
∂x2 f (x, y) = − cos(x)ey/2
1
∂y f (x, y) = cos(x)ey/2
2
we get
cos(XT )eYT /2 − 1
T 1 T 1 T
= −∫ sin(Xs )eYs /2 dXs + ∫ cos(Xs )e
Ys /2
dYs − ∫ cos(Xs )eYs /2 g 2 (s) ds
0 2 0 2 0
T
= −∫ sin(Xs )eYs /2 g(s) dBs .
0
Thus, by the definition of g, X and Y ,
⎛n ⎞ T
cos ∑ ξj Btj = cos (∫ g(s) dBs )
⎝j=1 ⎠ 0
T T s T
= e− ∫0 g 2 (s) ds
g(r) dBr ) e− 2 ∫s g 2 (r) dr
1
−∫ sin (∫ g(s) dBs .
0 0
Since the integrand of the stochastic integral is continuous and bounded, it is clear that
it is in L2P (λT ⊗ P). Hence cos (∑nj=1 ξj Btj ) ∈ HT2 .
The imaginary part can be treated in a similar way.
∎∎
161
Problem 18.8. Solution: Set Σt1 ,...,tn ∶= σ(Bt1 , . . . , Btn ). There are several possibilities to
prove this result.
Possibility 1: Set tn = (t1 , . . . , tn ) and Σ(tn ) = Σt1 ,...,tn . Then the family of σ-algebras
Σ(tn ) is upwards filtering, i. e. whenever we have tn and sm there is some un+m such that
Σ(sm )∪Σ(tn ) ⊂ Σ(un+m ). Therefore we can use Lévy’s (upwards) martingale convergence
theorem and conclude that
E(Y ∣ FTB ) = L1 - lim E(Y ∣ Σ(tn )) = 0.
Since FTB and F̄TB differ only by trivial sets (with probability zero or one), we get a. s.
Y = E(Y ∣ F̄TB ) = E(Y ∣ FTB ) = 0.
Possibility 2: Set ΣT ∶= ⋃0⩽t1 <⋯<tn =T Σt1 ,...,tn . Then σ(ΣT ) = FTB and ΣT is stable under
n⩾1
intersections. Consider the measures
µ± (F ) ∶= ∫ Y ± d P ∀F ∈ ΣT .
F
By assumption, µ+ (F ) = µ− (F ) on ΣT , and by the uniqueness theorem for measures we

get µ+ = µ− on FT . But then we get ∫F Y d P = 0 for all F ∈ FTB .
If we add to ΣT all P null set, the above considerations remain valid (without changes!)
and we get ∫F Y d P = 0 for all F ∈ F̄TB , hence Y = 0 as Y is F̄T measurable.
∎∎
Problem 18.9. Solution: Because of the properties of conditional expectations we have for
s⩽t
M áG∞
E (Mt ∣ Hs ) = E (Mt ∣ σ(Fs , Gs )) = E (Mt ∣ Fs ) = Ms .
Thus, (Mt , Ht )t⩾0 is still a martingale; (Bt , Ht )t⩾0 is treated in a similar way.
∎∎
Problem 18.10. Solution: Recall that
τ (s) = inf{t ⩾ 0 ∶ a(t) > s}.
Since for any > 0
{t ∶ a(t) ⩾ s} ⊂ {t ∶ a(t) > s − } ⊂ {t ∶ a(t) ⩾ s − }
we get
inf{t ∶ a(t) ⩾ s} ⩾ inf{t ∶ a(t) > s − } ⩾ inf{t ∶ a(t) ⩾ s − }
and
inf{t ∶ a(t) ⩾ s} ⩾ lim inf{t ∶ a(t) > s − } ⩾ lim inf{t ∶ a(t) ⩾ s − }.
↑0 ↑0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
=lim↑0 τ (s−)=τ (s−)
162
Thus, inf{t ∶ a(t) ⩾ s} ⩾ τ (s−). Assume that inf{t ∶ a(t) ⩾ s} > τ (s−). Then
a(τ (s−)) < s.
On the other hand, by Lemma 18.15 b)
s − ⩽ a(τ (s − )) ⩽ a(τ (s−)) < s ∀ > 0.
This leads to a contradiction, and so inf{t ∶ a(t) ⩾ s} ⩽ τ (s−).

The proof for a(s−) is similar.
Assume that τ (s) ⩾ t. Then a(t−) = inf{s ⩾ 0 ∶ τ (s) ⩾ t} ⩽ s. On the other hand,
18.15 d)
a(t−) ⩽ s Ô⇒ ∀ > 0 ∶ a(t − ) ⩽ s Ô⇒ ∀ > 0 ∶ τ (s) > t − Ô⇒ τ (s) ⩾ t.
∎∎
{⟨M ⟩t ⩽ s} = ⋂ {⟨M ⟩t < s + 1/n} = ⋂ {⟨M ⟩t ⩾ s + 1/n}c

n⩾1 n⩾1
= ⋂ {τs+1/n− ⩽ s}c ∈ ⋂ Fτ = Fτs + .

18.15 A.15
c) s+1/n
n⩾1 n⩾1
As Ft is right-continuous, Fτs + = Fτs = Gs and we conclude that ⟨M ⟩t is a Gt stopping

time.
∎∎
Problem 18.12. Solution: Solution 1: Assume that f ∈ C2 . Then we can apply Itô’s formula.
Use Itô’s formula for the deterministic process Xt = f (t) and apply it to the function xa
(we assume that f ⩾ 0 to make sure that f a is defined for all a > 0):
t d a t
f a (t) − f a (0) = ∫ [ x ] df (s) = ∫ af a−1 (s) df (s).
0 dx x=f (s) 0
This proves that the primitive ∫ f a−1 df = f a /a. The rest is an approximation argument
(f ∈ C1 is pretty immediate).
Solution 2: Any absolutely continuous function has an Lebesgue a.e. defined derivative f ′
and f = ∫ f ′ ds. Thus,
t
t t
′
t 1 d a f a (s) f a (t) − f a (0)
∫ f a−1
(s) df (s) = ∫ f a−1
(s)f (s) ds = ∫ f (s) ds = [ ] = .
0 0 0 a ds a 0 a
∎∎
Problem 18.13. Solution: Theorem. Let Bt = (Bt1 , . . . , Btd ) be a d-dimensional Brownian

motion and f1 , . . . , fd ∈ L2P (λT ⊗ P) for all T > 0. Then, we have for 2 ⩽ p < ∞
⎡ p/2 ⎤
⎢ T d ⎥ t p
E ⎢⎢(∫ ∑ ∣fk (s)∣ ds) ⎥⎥ ≍ E [sup ∣∑ ∫ fk (s) dBsk ∣ ]
2
(18.2)
⎢ 0 k=1 ⎥ 0
⎣ ⎦ t⩽T k
with finite comparison constants which depend only on p.
163
t
Proof. Let Xt = ∑k ∫0 fk (s) dBsk . Then we have
t t
⟨X⟩t = ⟨∑ ∫ fk (s) dBsk , ∑ ∫ fl (s) dBsl ⟩
k 0 l 0
t t
= ∑ ⟨∫ fk (s) dBsk , ∫ fl (s) dBsl ⟩
k,l 0 0
t
= ∑∫ fk (s)fl (s) d ⟨B k , B l ⟩s
k,l 0
t
= ∑∫ fk2 (s) ds
k 0
since dBsk dBsl = d⟨B k , B l ⟩s = δkl ds.
With these notations, the proof of Theorem 17.16 goes through almost unchanged and we
get the inequalities for p ⩾ 2.
Remark: Often one needs only one direction (as we do later in the book) and one can use
18.20 directly, without going through the proof again. Note that
d p d p
t t
∣∑ ∫ fk (s) dBs ∣
k
⩽ ( ∑ ∣∫ fk (s) dBs ∣)
k
k=1 0 k=1 0
d t p
⩽ cd,p ∑ ∣∫ fk (s) dBsk ∣ .
k=1 0
Thus, by (18.20)
⎡ d p⎤ d p
⎢ t
k ⎥
t
E ⎢sup ∣ ∑ ∫ fk (s) dBs ∣ ⎥ ⩽ cd,p ∑ E [sup ∣∫ fk (s) dBsk ∣ ]
⎢ t⩽T k=1 0 ⎥
⎣ ⎦ k=1 t⩽T 0
d ⎡ p/2 ⎤
⎢ T
⎥
≍ cd,p ∑ E ⎢(∫ ∣fk (s)∣2 ds) ⎥
k=1 ⎣
⎢ ⎥
0 ⎦
⎡ p/2 ⎤
⎢ T d ⎥
≍ cd,p E ⎢(∫ ∑ ∣fk (s)∣ ds) ⎥⎥ .
⎢ 2
⎢ 0 k=1 ⎥
⎣ ⎦
∎∎
164
19 Stochastic differential equations
dXt = b(t) dt + σ(t) dBt
where b, σ are non-random coefficients such that the corresponding (stochastic) integrals
exist. Obviously,
(dXt )2 = σ 2 (t) (dBt )2 = σ 2 (t) dt
and we get for 0 ⩽ s ⩽ t < ∞, using Itô’s formula,

t t
eiξXt − eiξXs = ∫ iξeiξXr b(r) dr + ∫ iξeiξXr σ(r) dBr
s s
1 t 2 iξXr 2
− ∫ ξ e σ (r) dr.
2 s
Now take any F ∈ Fs and multiply both sides of the above formula by e−ξXs 1F . We get
t t
eiξ(Xt −Xs ) 1F − 1F = ∫ iξeiξ(Xr −Xs ) 1F b(r) dr + ∫ iξeiξ(Xr −Xs ) 1F σ(r) dBr
s s
1 t 2 iξ(Xr −Xs )
− ∫ ξ e 1F σ 2 (r) dr.
2 s
Taking expectations gives
t
E (eiξ(Xt −Xs ) 1F ) = P(F ) + ∫ iξ E (eiξ(Xr −Xs ) 1F ) b(r) dr
s
1 t 2 iξ(Xr −Xs )
− ∫ ξ E (e 1F ) σ 2 (r) dr
2 s
t 1
= P(F ) + ∫ (iξb(r) − ξ 2 σ 2 (r)) E (eiξ(Xr −Xs ) 1F ) dr.
s 2
Define φs,t (ξ) ∶= E (eiξ(Xt −Xs ) 1F ). Then the integral equation
t 1 2 2
φs,t (ξ) = P(F ) + ∫ (iξb(r) − ξ σ (r))φr,s (ξ) dr
s 2
has the unique solution (use Gronwall’s lemma, cf. also the proof of Theorem 18.5)
t
b(s) ds− 12 ξ 2 ∫st σ 2 (r) dr
φs,t (ξ) = P(F ) eiξ ∫s
and so
t
b(r) dr− 12 ξ 2 ∫st σ 2 (r) dr
E (eiξ(Xt −Xs ) 1F ) = P(F ) eiξ ∫s . (*)
If we take in (*) F = Ω and s = 0, we see that

t 1 t 2
Xt ∼ N(µt , σt2 ), µt = ∫ b(r) dr, σt2 = ∫ σ (r) dr.
0 2 0
165
If we take in (*) F = Ω then the increment satisfies Xt − Xs ∼ N(µt − µs , σt2 − σs2 ). If F is

arbitrary, (*) shows that
Xt − Xs á Fs ,
see the Lemma at the end of this section.
The above considerations show that

n tj 1 tj
E e∑j=1 ξj (Xtj −Xtj−1 ) = ∏ exp (iξ ∫
n
b(r) dr − ξ 2 ∫ σ 2 (r) dr) ,
j=1 tj−1 2 tj−1
i. e. (Xt1 , Xt2 − Xt1 , . . . , Xtn − Xtn−1 ) is a Gaussian random vector with independent com-
ponents. Since Xtk = ∑kj=1 (Xtj − Xtj−1 ) we see that (Xt1 , . . . , Xtn ) is a Gaussian random
variable.
Let us, finally, compute E(Xs Xt ). By independence, we have
E(Xs Xt ) = E(Xs2 ) + E Xs (Xt − Xs )

= E(Xs2 ) + E Xs E(Xt − Xs )
= E(Xs2 ) + E Xs E Xt − (E Xs )2
= V Xs + E Xs E Xt
s s t
=∫ σ 2 (r) dr + ∫ b(r) dr ∫ b(r) dr.
0 0 0
In fact, since the mean is not zero, it would have been more elegant to compute the
covariance
s
Cov(Xs , Xt ) = E(Xs − µs )(Xt − µt ) = E(Xs Xt ) − E Xs E Xt = V Xs = ∫ σ 2 (r) dr.
0
Lemma. Let X be a random variable and F a σ field. Then
E (eiξX 1F ) = E eiξX ⋅ P(F ) ∀ξ ∈ R Ô⇒ X á F.
Proof. Note that eiη1F = eiη 1F + 1F c . Thus,
E (eiξX 1F c ) = E (eiξX ) − E (eiξX 1F )

= E (eiξX ) − E (eiξX ) P(F )
= E (eiξX ) P(F c )
and this implies
E (eiξX eiη1F ) = E (eiξX ) E (eiη1F ) ∀ξ, η ∈ R.
This shows that X á 1F and X á F for all F ∈ F.
166
∎∎
(a) We have ∆t = 2−n and
∆Xn (tk−1 ) = Xn (tk ) − Xn (tk−1 ) = − 21 Xn (tk−1 ) 2−n + B(tk ) − B(tk−1 )
and this shows
Xn (tk ) = Xn (tk−1 ) − 21 Xn (tk−1 ) 2−n + B(tk ) − B(tk−1 )

= (1 − 2−n−1 ) Xn (tk−1 ) + B(tk ) − B(tk−1 )
= (1 − 2−n−1 )[(1 − 2−n−1 ) Xn (tk−2 ) + B(tk−1 ) − B(tk−2 )] + [B(tk ) − B(tk−1 )]
⋮
= (1 − 2−n−1 )k Xn (t0 ) + (1 − 2−n−1 )k−1 [B(t1 ) − B(t0 )] + . . . +
+ (1 − 2−n−1 )[B(tk−1 ) − B(tk−2 )] + [B(tk ) − B(tk−1 )]
k−1
= (1 − 2−n−1 )k A + ∑ (1 − 2−n−1 )j [B(tk−j ) − B(tk−j−1 )]
j=1
Observe that B(tj ) − B(tj−1 ) ∼ N(0, 2−n ) for all j and A ∼ N(0, 1). Because of the
independence we get
Xn (tn ) = Xn (k2−n ) ∼ N(0, (1 − 2−n−1 )2k + ∑j=1 (1 − 2−n−1 )2j ⋅ 2−n )

k−1
For k = 2n−1 we get tk = 1

2 and so
2n −1
Xn ( 12 ) ∼ N(0, (1 − 2−n−1 )2 + ∑j=1 (1 − 2−n−1 )2j ⋅ 2−n ).
n
Using
lim (1 − 2−n−1 )2 = e− 2
n 1
n→∞
and
2n −1
1 − (1 − 2−n−1 )2 1 − (1 − 2−n−1 )2
n n
−n−1 2j
∑ (1 − 2 ) ⋅ 2−n = ⋅ 2−n
= ÐÐÐ→ 1 − e − 12
j=1 1 − (1 − 2−n−1 )2 1 − 2−n−2 n→∞
d
finally shows that Xn ( 12 ) ÐÐÐ→ X ∼ N(0, 1).
n→∞
(b) The solution of this SDE follows along the lines of Example 19.7 where α(t) ≡ 0,
β(t) ≡ − 12 , δ(t) ≡ 0 and γ(t) ≡ 1:
dXt○ = 1 ○
2 Xt dt Ô⇒ Xt○ = et/2
Zt = et/2 Xt , Z0 = X0
t
dZt = et/2 dBt Ô⇒ Zt = Z0 + ∫ es/2 dBs
0
t
Xt = e−t/2 A + e−t/2 ∫ es/2 dBs .
0
167
For t = 1
2 we get
1/2
X1/2 = A e−1/4 + e−1/4 ∫ es/2 dBs
0
Ô⇒ X1/2 ∼ N(0, e−1/2 + e−1/2 ∫0

1/2 s
e ds) = N(0, 1).
So, we find for all s ⩽ t

s t
C(s, t) = E Xs Xt = e−s/2 e−t/2 E A2 + e−s/2 e−t/2 E (∫ er/2 dBr ∫ eu/2 dBu )
0 0
s
= e−(s+t)/2 + e−(s+t)/2 ∫ er dr
0
= e−(t−s)/2 .
This finally shows that C(s, t) = e−∣t−s∣/2 .
∎∎
Problem 19.3. Solution: We can rewrite the SDE as

t t
Xt = x + b ∫ Xs ds + ∫ Xs d(σ1 bs + σ2 βs )
0 0
t √ t
= x + b ∫ Xs ds + σ12 + σ22 ∫ Xs dWs
0 0
where
σ1 σ2
Ws ∶= √ bt + √ βt
σ1 + σ2
2 2 σ12 + σ22
is, by Problem 18.3, a BM1 . This reduces the problem to a geometric Brownian motion
as in Example a:
1 2 √
Xt = x exp ([b − (σ1 + σ22 )] t + σ12 + σ22 Wt )
2
1
= x exp ([b − (σ12 + σ22 )] t + σ1 bt + σ2 βt ) .
2
Alternative Solution: As in Example 19.6, we assume that the initial condition X0 = x is

positive and apply Itô’s formula (17.15) to Zt ∶= log Xt :
t 1 1 t 1
Zt − Z0 = ∫ dXs + ∫ (− 2 ) (dXs )2
0 Xs 2 0 Xs
t t t 1 t
= ∫ b ds + ∫ σ1 dbs + ∫ σ2 dβ2 − ∫ (σ12 + σ22 ) ds
0 0 0 2 0
1 2
= (b − (σ1 + σ22 )) ⋅ t + σ1 bt + σ2 βt .
2
Since, by assumption,
dXs = b Xs ds + σ1 Xs dβs + σ2 Xs dbs

Ô⇒ (dXs )2 = (σ1 Xs dβs )2 + (σ2 Xs dbs )2 = (σ12 + σ22 ) Xs2 ds.
168
Consequently,
1
Xt = x exp (σ1 bt + σ2 βt + (b − (σ12 + σ22 )) t) .
2
A direct calculation shows that Xt is indeed a solution of the given SDE.
∎∎
Problem 19.4. Solution: Since Xt○ is such that 1/Xt○ solves the homogeneous SDE from
Example 19.6, we see that
t t
Xt○ = exp (− ∫ (β(s) − 21 δ 2 (s)) ds) exp (− ∫ δ(s) dBs )
0 0
(mind that the ‘minus’ sign comes from 1/Xt○ ).
Observe that Xt○ = f (It1 , It2 ) where It is an Itô process with

t
It1 = − ∫ (β(s) − 12 δ 2 (s)) ds
0
t
It2 = − ∫ δ(s) dBs .
0
Now we get from Itô’s multiplication table
dIt1 dIt1 = dIt1 dIt2 = 0 and dIt2 dIt2 = δ 2 (t) dt
and, by Itô’s formula

2
dXt○ = ∂1 f (It1 , It2 ) dIt1 + ∂2 f (It1 , It2 ) dIt2 + 1
2
j k
∑ ∂j ∂k dIt dIt
j,k=1
= Xt○ (dIt1 + dIt2 + 21 dIt2 dIt2 )

= Xt○ (−β(t) dt + 12 δ 2 (t) dt − δ(t) dBt + 21 δ 2 (t) dt)
= Xt○ (−β(t) + δ 2 (t)) dt − Xt○ δ(t) dBt .
Remark:
1. We used here the two-dimensional Itô formula (17.14) but we could have equally well
used the one-dimensional version (17.13) with the Itô process It1 + It2 .
2. Observe that Itô’s multiplication table gives us exactly the second-order term in
(17.14).
Since
dZt = (α(t) − γ(t)δ(t))Xt○ dt + γ(t)Xt○ dBt and Xt = Zt /Xt○
we get
1 t t
Xt = ○
(X0 + ∫ (α(s) − γ(s)δ(s))Xs○ ds + ∫ γ(s)Xs○ dBs ) .
Xt 0 0
∎∎
169
(a) We have Xt = e−βt X0 + ∫0 σe−β(t−s) dBs . This can be shown in four ways:
t
Solution 1: you guess the right result and use Itô’s formula (17.7) to verify that the
above Xt is indeed a solution to the SDE. For this rewrite the above solution as
t
eβt Xt = X0 + ∫ σeβs dBs Ô⇒ d(eβt Xt ) = σeβt dBt .
0
Now with the two-dimensional Itô formula for f (x, y) = xy and the two-dimensional
Itô-process (eβt , Xt ) we get
d(eβt Xt ) = βXt eβt dt + eβt dXt
so that
βXt eβt dt + eβt dXt = σeβt dBt ⇐⇒ dXt = −βXt dt + σ dBt .
Admittedly, this is unfair as one has to know the solution beforehand. On the other
hand, this is exactly the way one verifies that the solution one has found is the
correct one.
Solution 2: you apply the time-dependent Itô formula from Problem 17.6 or the two-
dimensional Itô formula, Theorem 17.8 to
t
Xt = u(t, It ) and It = ∫ eβs dBs and u(t, x) = eβt X0 + σeβt x
0
to get—as dt dBt = 0—
dXt = ∂t u(t, It ) dt + ∂x u(t, It ) dIt + 21 ∂x2 u(t, Bt ) dt.
Again, this is best for the verification of the solution since you need to know its form
beforehand.
Solution 3: you use Example 19.7 with α(t) ≡ 0, β(t) ≡ −β, γ(t) ≡ σ and δ(t) ≡ 0.
But, honestly, you will have to look up the formula in the book. We get
dXt○ = βXt○ dt, X0○ = 1 Ô⇒ Xt○ = eβt ;

Zt = eβt Xt , Z0 = X0 = ξ = const.;
dZt = σe dBt ;
βt
t
Zt = σ ∫ eβs dBs + Z0 ;
0
t
Xt = e−βt ξ + e−βt σ ∫ eβs dBs , t ⩾ 0.
0
Solution 4: by bare hands and with Itô’s formula! Consider first the deterministic
ODE
t
xt = x0 − β ∫ xs ds
0
170
which has the solution xt = x0 e−βt , i. e. eβt xt = x0 = const. This indicates that the
transformation
Yt ∶= eβt Xt
might be sensible. Thus, Yt = f (t, Xt ) where f (t, x) = eβt x. Thus,
∂t f (t, x) = βf (t, x) = βxeβt , ∂x f (t, x) = eβt , ∂x2 fxx (t, x) = 0.
By assumption,
dXt = −βXt dt + σ dBt Ô⇒ (dXt )2 = σ 2 (dBt )2 = σ 2 dt,
and by Itô’s formula (17.8) we get
Yt − Y0
t t
=∫ ( ft (s, Xs ) − βXs fx (s, Xs ) + 21 σ 2 fxx (s, Xs ) ) ds + ∫ σfx (s, Xs ) dBs
0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶ 0
=0 =0
t
=∫ σfx (s, Xs ) dBs .
0
So we have the solution, but we still have to go through the procedure in Solution 1
or 2 in order to verify our result.
(b) Since Xt is the limit of normally distributed random variables, it is itself Gaussian (see
also part d))—if ξ is non-random or itself Gaussian and independent of everything
else. In particular, if X0 = ξ = const.,
Xt ∼ N (e−βt ξ, σ 2 e−2βt ∫0 e2βs ds) = N (e−βt ξ, σ2 −2βt

t
2β (1 − e )) .
Now
σ 2 −β(t+s) 2βs
C(s, t) = E Xs Xt = e−β(t+s) ξ 2 + e (e − 1), t ⩾ s ⩾ 0,
2β
and, therefore
σ 2 −β∣t−s∣
C(s, t) = e−β(t+s) ξ 2 + (e − e−β(t+s) ) for all s, t ⩾ 0.
2β
(c) The asymptotic distribution, as t → ∞, is X∞ ∼ N(0, σ 2 (2β)−1 ).
(d) We have
⎡ n ⎤
⎛ ⎢ ⎥⎞
E exp ⎢i ∑ λj Xtj ⎥⎥
⎢
⎝ ⎢ j=1 ⎥⎠
⎣ ⎦
⎡ n ⎤
⎛ ⎢ n tj ⎥⎞
= E exp ⎢⎢i ∑ λj e−βtj ξ + iσ ∑ λj e−βtj ∫ eβs dBs ⎥⎥
⎝ ⎢ j=1 0 ⎥⎠
⎣ j=1 ⎦
⎛ σ 2 ⎡⎢ n ⎤2 ⎞
⎥ ⎛
⎡ n
⎢
⎤
⎥⎞
= exp ⎜− ⎢⎢ ∑ λj e−βtj ⎥⎥ ⎟ E exp ⎢⎢iσ ∑ ηj Yj ⎥⎥
⎝ 4β ⎢⎣j=1 ⎥ ⎠ ⎝ ⎢ j=1 ⎥⎠
⎦ ⎣ ⎦
171
where
tj
ηj = λj e−βtj , Yj = ∫ eβs dBs , t0 = 0, Y0 = 0.
0
Moreover,
n n n
∑ ηj Yj = ∑ (Yk − Yk−1 ) ∑ ηj
j=1 k=1 j=k
and
tk
Yk − Yk−1 = ∫ eβs dBs ∼ N(0, (2β)−1 (e2βtk − e2βtk−1 )) are independent.
tk−1
Consequently, we see that

⎡ n ⎤
⎛ ⎢ ⎥⎞
⎢
E exp ⎢i ∑ λj Xtj ⎥⎥
⎝ ⎢ j=1 ⎥⎠
⎣ ⎦
2
σ 2 n σ2 2
= exp [− ( ∑j=1 λj e−βtj ) ] ∏ exp [− (e2βtk − e2βtk−1 )( ∑j=k λj e−βtj ) ]
n n
4β k=1 4β
σ2 2
( ∑j=1 λj e−βtj ) {1 + e2βt1 − 1}] ×
n
= exp [−
4β
n
σ2 2
(1 − e−2β(tk −tk−1 ) ) ⋅ ( ∑j=k λj e−β(tj −tk ) ) ]
n
× ∏ exp [−
k=2 4β
σ2 2
( ∑j=1 λj e−β(tj −t1 ) ) ] ×
n
= exp [−
4β
n
σ2 2
(1 − e−2β(tk −tk−1 ) ) ⋅ ( ∑j=k λj e−β(tj −tk ) ) ] .
n
× ∏ exp [−
k=2 4β
Note: the distribution of (Xt1 , . . . , Xtn ) depends on the difference of the consecutive
epochs t1 < . . . < tn .
(e) We write for all t ⩾ 0

X̃t = eβt Xt and Ũt = eβt Ut
and we show that both processes have the same finite-dimensional distributions.
Clearly, both processes are Gaussian and both have independent increments. From
X̃0 = X0 = 0 and Ũ0 = U0 = 0
and for s ⩽ t
t
X̃t − X̃s = σ ∫ eβr dBr
s
2
∼ N(0, 2β
σ
(e2βt − e2βs )),
σ
Ũt − Ũs = √ (B(e2βt − 1) − B(e2βs − 1))
2β
σ
∼ B(e2βt − e2βs )
2β
σ 2 2βt
∼ N(0, 2β
(e − e2βs ))
we see that the claim is true.
172
∎∎
Problem 19.6. Solution: We use the time-dependent Itô formula from Problem 17.6 (or the
x dy
two-dimensional Itô-formula for the process (t, Xt )) with f (t, x) = ect ∫0 σ(y) . Note that
the parameter c is still a free parameter.
Using Itô’s multiplication rule—(dt)2 = dt dBt = 0 and (dBt )2 = dt we get
dXt = b(Xt ) dt + σ(Xt ) dBt Ô⇒ (dXt )2 = d⟨X⟩t = σ 2 (Xt ) dt.
Thus,
dZt = df (t, Xt ) = ∂t f (t, Xt ) dt + ∂x f (t, Xt ) dXt + 21 ∂x2 f (t, Xt ) (dXt )2

Xt dy 1 1 σ ′ (Xt ) 2
= cect ∫ dt + ect dXt − ect 2 σ (Xt ) dt
0 σ(y) σ(Xt ) 2 σ (Xt )
Xt dy b(Xt ) 1
= cect ∫ dt + ect dt + ect dBt − ect σ ′ (Xt ) dt
0 σ(y) σ(Xt ) 2
Xt dy 1 b(Xt )
= ect [c ∫ − σ ′ (Xt ) + ] dt + ect dBt .
0 σ(dy) 2 σ(Xt )
Let us show that the expression in the brackets [⋯] is constant if we choose c appropriately.
For this we differentiate this expression:
d x dy 1 b(x) c d 1 ′ b(x)
[c ∫ − σ ′ (x) + ]= − [ σ (x) − ]
dx 0 σ(dy) 2 σ(x) σ(x) dx 2 σ(x)
c 1 d b(x)
= − [ σ ′′ (x) − ]
σ(x) 2 dx σ(x)
1 1 d b(x)
= (c − σ(x) [ σ ′′ (x) − ])
σ(x) 2 dx σ(x)
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
=const. by assumption
This shows that we should choose c in such a way that the expression c − σ ⋅ [⋯] becomes
zero, i. e.
1 d b(x)
c = σ(x) [ σ ′′ (x) − ].
2 dx σ(x)
∎∎
Problem 19.7. Solution: Set f (t, x) = tx. Then
∂t f (t, x) = x, ∂x f (t, x) = t, ∂x2 f (t, x) = 0.
Using the time-dependent Itô formula (cf. Problem 17.6) or the two-dimensional Itô for-
mula (cf. Theorem 17.8) for the process (t, Bt ) we get
dXt = ∂t f (t, Bt ) dt + ∂x f (t, Bt ) dBt + 12 ∂x2 f (t, Bt ) dt

= Bt dt + t dBt
Xt
= dt + t dBt .
t
173
Together with the initial condition X0 = 0 this is the SDE which has Xt = tBt as solution.
The trouble is, that the solution is not unique! To see this, assume that Xt and Yt are
any two solutions. Then
Xt Yt Zt
dZt ∶= d(Xt − Yt ) = dXt − dYt = ( − ) dt = dt, Z0 = 0.
t t t
This is an ODE and all (deterministic) processes Zt = ct are solutions with initial condition
Z0 = 0. If we want to enforce uniqueness, we need a condition on Z0′ . So
Xt d
dXt = dt + t dBt and Xt ∣ = x′0
t dt t=0
will do. (Note that tBt is differentiable at t = 0!).

∎∎

(a) With the argument from Problem 19.7, i. e. Itô’s formula, we get for f (t, x) = x/(1+t)
x 1
∂t f (t, x) = − , ∂x f (t, x) = , ∂x2 f (t, x) = 0.
(1 + t)2 1+t
And so
Bt 1
dUt = − dt + dBt
(1 + t) 2 1+t
Ut 1
=− dt + dBt .
1+t 1+t
The initial condition is U0 = 0.
(b) Using Itô’s formula for f (x) = sin x we get, because of sin2 x + cos2 x = 1, that
dVt = cos Bt dBt − 12 sin Bt dt

√
= 1 − sin2 Bt dBt − 21 sin Bt dt
√
= 1 − Vt2 dBt − 12 Vt dt
and the initial condition is V0 = 0.

Attention: We loose all information on the sign of cos Bt when taking the square root
√
1 − sin Bt . This means that the SDE corresponds to Vt = sin Bt only when cos Bt is
positive, i. e. for t < inf {s > 0 ∶ Bs ∉ [− 12 π, 12 π]}.
(c) Using Itô’s formula in each coordinate we get
Xt −a sin Bt 1 −a cos Bt
d( ) = ( ) dBt + ( ) dt
Yt b cos Bt 2 −b sin Bt
− a b sin Bt 1 a cos Bt
= ( bb ) dBt − ( ) dt
a a cos Bt
2 b sin Bt
− a Yt 1 Xt
= ( bb ) dBt − ( ) dt.
a Xt
2 Yt
The initial condition is (X0 , Y0 ) = (a, 0).
174
∎∎

(a) We use Example 19.7 (and 19.6) where we set
α(t) ≡ b, β(t) ≡ 0, γ(t) ≡ 0, δ(t) ≡ σ.
Then we get
dXt○ = σ 2 Xt○ dt − σXt○ dBt

dZt = bXt○ dt
and, by Example 19.6 we see

t t
Xt○ = X0○ exp (∫ (σ 2 − 21 σ 2 ) ds − ∫ σ dBs )
0 0
= X0○ exp ( 12 σ 2 t − σ Bt )
t
Zt = ∫ bXs○ ds
0
Thus,
t 1 2 s−σB
Zt = ∫ b e2 σ s
ds
0
Zt t 1 2
= be− 2 σ t+σBt ∫ e 2 σ s−σBs ds.
1 2
Xt = ○
Xt 0
We finally have to adjust the initial condition by adding X0 = x0 to the Xt we have

just found:
t
Ô⇒ Xt = X0 + be− 2 σ
1 2 t+σB 1 2 s−σB
t
∫ e2 σ s
ds.
0
Alternative Solution (by R. Baumgarth, TU Dresden): This solution does

not use Example 19.7 First, we solve the homogeneous SDE, i. e. b = 0:
dXt = σXt dBt .
Using Zt ∶= log Xt and Itô’s formula (or simply Example we see that this equation
has the unique solution
Xt = x0 e− 2 σ
1 2 t+σB
t
,
which is of the form

1 2 t−σB
constant = x0 = Xt e 2 σ t
= Xt Xto .
Because of the form of the homogeneous solution, we now use stochastic integration
by parts1
d(Xt Xto ) = Xt dXt0 + Xt0 dXt + d⟨X, X 0 ⟩t . (*)

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
= dXt dXt0
1
Note this formula can be shown by applying Itô’s formula on f (x, x0 ) = xx0 .
175
So we need to find dXt0 . Using f (y) = ey for the process Yt = 12 σ 2 t − σBt gives
1
dXto = df (Yt ) = f ′ (Yt ) dYt + f ′′ (Yt )(dYt )2
2
o 1 2 1
= Xt ( σ dt − σ dBt + σ 2 dt)
2 2
= σ 2 Xt0 dt − σXto dBt .
Inserting everything in (*) yields
d(Xt Xto ) = Xt Xto (σ 2 dt − σ dBt ) + Xto (b dt + σXt dBt )

= Xto b dt,
and so the solution is

1 t
Xt − X0 = b ∫ Xso ds
Xto 0
t
Xt = X0 + b e− 2 σ
1 2 t+σB 1 2 s−σB
t
∫ e2σ s
ds.
0
(b) We use Example 19.7 (and 19.6) where we set
α(t) ≡ m, β(t) ≡ −1, γ(t) ≡ σ, δ(t) ≡ 0.
Then we get
dXt○ = Xt○ dt
dZt = mXt○ dt + σXt○ dBt
Thus,
Xt○ = X0○ et
t t
Zt = ∫ m es ds + σ ∫ es dBs
0 0
t
= m (et − 1) + σ ∫ es dBs
0
Zt t
Xt = ○ = m (1 − e−t ) + σ ∫ es−t dBs
Xt 0
and, if we take care of the initial condition X0 = x0 , we get

t
Ô⇒ Xt = x0 + m (1 − e−t ) + σ ∫ es−t dBs .
0
Alternative Solution (by R. Baumgarth, TU Dresden): This does not use

Example 19.7. Consider the deterministic ODE σ = 0 and, to simplify the problem,
m = 0.
ẋ(t) = −x(t)
176
has the unique solution
x(t) = x0 e−t .
Thus, the solution is of the form
constant = x0 = x(t)et = f (t, xt ).
Now use Itô’s formula for Yt = f (t, Xt ):
1
df (t, Xt ) = ∂t f (t, Xt ) dt + ∂x f (t, Xt ) dXt + ∂x2 f (t, Xt )(dXt )2
2
= Xt e dt + e ((m − Xt ) dt + σ dBt ) + 0
t t
= et (m dt + σ dBt ) .
Hence,
t t
Xt et − X0 = m ∫ es ds +σ ∫ es dBs
0 0
´¹¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
= (et −1)
or
t
Xt = X0 + m (1 − e−t ) + σ ∫ es−t dBs .
0
∎∎
Problem 19.10. Solution: Set

√ √
b(x) = 1 + x2 + 21 x and σ(x) = 1 + x2 .
Then we get (using the notation of Lemma 19.10)
x b(x) 1 ′
σ ′ (x) = √ and κ(x) = − σ (x) = 1.
1 + x2 σ(x) 2
Using the Ansatz of Lemma 19.10 we set

x dy
d(x) = ∫ = arsinh x and Zt = f (Xt ) = d(Xt ).
0 σ(y)
Using Itô’s formula gives
dZt = ∂x f (Xt ) dXt + 21 ∂x2 f (Xt ) σ 2 (Xt ) dt

1 ′
= dXt + 21 ( σ1 ) (Xt ) σ 2 (Xt ) dt
σ(Xt )
Xt Xt
= (1 + √ ) dt + dBt + 12 (− ) (1 + Xt2 ) dt
2 1 + Xt 2 (1 + Xt
2 )3/2
= dt + dBt ,
177
and so Zt = Z0 + t + Bt . Finally,
Xt = sinh(Z0 + t + Bt ) where Z0 = arsinh X0 .
Alternative Solution (by R. Baumgarth, TU Dresden): In order to guess the

correct Ansatz, consider first the ODE
√ 1
dx(t) = ( 1 + x(t)2 + ///////
x(t)) dt.
2
If we do not get rid of the second tern, things get messy when integrating. Hence, this
ODE is very easy to solve by separation of variables and we see
arsinh x(t) = t + c,
so a suitable Ansatz for Itô’s formula is Yt ∶= f (t, Xt ) = arsinh Xt − t.

1
dYt = ft (t, Xt ) dt + fx (t, Xt ) dXt + fxx (t, Xt ) (dXt )2
2
1 1⎛ Xt ⎞
= −dt + √ dXt + −√ (1 + Xt2 ) dt
1 + Xt2 2⎝ (1 + Xt ) ⎠
2 3
1 √ 1 √ Xt
= −dt + √ [( 1 + Xt2 + Xt ) dt + 1 + Xt2 dBt ] − √ dt
1 + Xt2 2 2 1 + Xt2
Xt Xt
= −dt + dt + √ dt + dBt − √ dt
2 1 + Xt 2 2 1 + Xt2
= dBt ,
hence Yt = Y0 + Bt and Xt = sinh(X0 + Bt + t).
∎∎
1/2
Problem 19.11. Solution: Set b = b(t, x), b0 = b(t, 0) etc. Observe that ∥b∥ = (∑j ∣bj (t, x)∣2 )
1/2
and ∥σ∥ = (∑j,k ∣σjk (t, x)∣2 ) are norms; therefore, we get using the triangle estimate
and the elementary inequality (a + b)2 ⩽ 2(a2 + b2 )
∥b∥2 + ∥σ∥2 = ∥b − b0 + b0 ∥2 + ∥σ − σ0 + σ0 ∥2
⩽ 2∥b − b0 ∥2 + 2∥σ − σ0 ∥2 + 2∥b0 ∥2 + 2∥σ0 ∥2
⩽ 2L2 ∣x∣2 + 2∥b0 ∥2 + 2∥σ0 ∥2
⩽ 2L2 (1 + ∣x∣)2 + 2(∥b0 ∥2 + ∥σ0 ∥2 )(1 + ∣x∣)2
⩽ 2(L2 + ∥b0 ∥2 + ∥σ0 ∥2 )(1 + ∣x∣)2 .
∎∎
(a) If b(x) = −ex and X0x = x we have to solve the following ODE/integral equation
t x
Xtx = x − ∫ eXs ds
0
178
and it is not hard to see that the solution is

1
Xtx = log ( ).
t + e−x
This shows that
1 1
lim Xtx = lim log ( −x
) = log = − log t.
x→∞ x→∞ t+e t
This means that Corollary 19.31 fails in this case since the coefficient of the ODE
grows too fast.
(b) Now assume that ∣b(x)∣ + ∣σ(x)∣ ⩽ M for all x. Then we have
t
∣∫ b(Xs ) ds∣ ⩽ M t.
0
By Itô’s isometry we get

t 2 t
E [∣∫ σ(Xsx ) dBs ∣ ] = E [∫ ∣σ 2 (Xsx )∣ ds] ⩽ M 2 t.
0 0
Using (a + b)2 ⩽ 2a2 + 2b2 we see

t 2 t 2
E(∣Xtx − x∣2 ) ⩽ 2 E [∣∫ b(Xs ) ds∣ ] + 2 E [∣∫ σ(Xsx ) dBs ∣ ]
0 0
⩽ 2(M t)2 + 2M 2 t
= 2M 2 t(t + 1).
By Fatou’s lemma
⎛ ⎞
E lim ∣Xtx − x∣2 ⩽ lim E(∣Xtx − x∣2 ) ⩽ 2M 2 t(t + 1)
⎝∣x∣→∞ ⎠ ∣x∣→∞
which shows that ∣Xtx ∣ cannot be bounded as ∣x∣ → ∞.
(c) Assume now that b(x) and σ(x) grow like ∣x∣p/2 for some p ∈ (0, 2). A calculation as
above yields
t 2 Cauchy t t
∣∫ b(Xs ) ds∣ ⩽ t∫ ∣b(Xs )∣2 ds ⩽ cp t ∫ (1 + ∣Xs ∣p ) ds
0 Schwarz 0 0
and, by Itô’s isometry

t 2 t t
E [∣∫ σ(Xsx ) dBs ∣ ] = E [∫ ∣σ 2 (Xsx )∣ ds] ⩽ c′ ∫ E(1 + ∣Xs ∣p ) ds.
0 0 0
Using (a + b)2 ⩽ 2a2 + 2b2 and Theorem 19.28 we get

t t
E ∣Xtx − x∣2 ⩽ 2cp t ∫ (1 + E(∣Xs ∣p )) ds + 2c′ ∫ (1 + E(∣Xs ∣p )) ds
0 0
t
⩽ ct,p + c′t,p ∫ ∣x∣p dt
0
′
= ct,p + t ct,p ∣x∣p .
179
Again by Fatou’s theorem we see that the left-hand side grows like ∣x∣2 (if Xtx is
unbounded) while the (larger!) right-hand side grows like ∣x∣p , p < 2, and this is
impossible.
Thus, (Xtx )x is unbounded as ∣x∣ → ∞.
∎∎
Problem 19.13. Solution: We have to show

∣x − y∣ ! x y
⩽ ∣ 2 − 2∣
(1 + ∣x∣)(1 + ∣y∣) ∣x∣ ∣y∣
2
∣x − y∣2 x y
⇐⇒ ⩽ ∣ 2 − 2∣
(1 + ∣x∣) (1 + ∣y∣)
2 2 ∣x∣ ∣y∣
∣x∣2 − 2⟨x, y⟩ + ∣y∣2 ∣x∣2 2⟨x, y⟩ ∣y∣2
⇐⇒ ⩽ − +
(1 + ∣x∣)2 (1 + ∣y∣)2 ∣x∣4 ∣x∣2 ∣y∣2 ∣y∣4
1 1 1 1 ∣x∣2 + ∣y∣2
⇐⇒ 2⟨x, y⟩ ( 2 2 − )⩽ + −
∣x∣ ∣y∣ (1 + ∣x∣)2 (1 + ∣y∣)2 ∣x∣2 ∣y∣2 (1 + ∣x∣)2 (1 + ∣y∣)2
1 1 1 1
⇐⇒ 2⟨x, y⟩ ( − ) ⩽ (∣x∣2 + ∣y∣2 ) ( 2 2 − ).
∣x∣2 ∣y∣2 (1 + ∣x∣)2 (1 + ∣y∣)2 ∣x∣ ∣y∣ (1 + ∣x∣)2 (1 + ∣y∣)2
By the Cauchy-Schwarz inequality we get 2⟨x, y⟩ ⩽ 2∣x∣ ⋅ ∣y∣ ⩽ ∣x∣2 + ∣y∣2 , and this shows that
the last estimate is correct.
Alternative Solution (by R. Baumgarth, TU Dresden): We can also use the
following direct calculation:
∣x − y∣2 ∣x∣2 − 2⟨x, y⟩ + ∣y∣2
=
(1 + ∣x∣) (1 + ∣y∣)
2 2
(1 + ∣x∣)2 (1 + ∣y∣)2
∣x∣2 − 2⟨x, y⟩ + ∣y∣2
⩽
∣x∣2 ∣y∣2
1 2⟨x, y⟩ 1
= 2− 2 2+ 2
∣y∣ ∣x∣ ∣y∣ ∣x∣
2
x y
=∣ − ∣ .
∣x∣2 ∣y∣2
∎∎

(a) We have seen in Corollary 19.24 that the transition function is given by
p(t, x; B) = P(Xtx ∈ B), t ⩾ 0, x ∈ R, B ∈ B(R).
Consequently,
Tt f (x) = ∫ f (y) p(t, x; dy) = E(f (Xtx )).
By Theorem 19.27 we know that x ↦ Tt f (x) = E(f (Xtx )) is continuous for each
f ∈ Cb (R). Since
∣Tt f (x)∣ ⩽ E ∣f (Xtx )∣ ⩽ ∥f ∥∞
we conclude that Tt maps Cb (R) into itself.
180
(b) Let f ∈ C∞ (R), t ⩾ 0. By part (a), Tt f ∈ Cb (R). Therefore, it suffices to show
lim ∣Tt f (x)∣ = lim ∣ E(f (Xtx ))∣ = 0.

∣x∣→∞ ∣x∣→∞
Since f ∈ C∞ (R) we obtain by applying Corollary 19.31,
lim ∣f (Xtx )∣ = 0 almost surely.

∣x∣→∞
The claim follows from the dominated convergence theorem.

(c) Let f ∈ C2c (R), x ∈ R. By Itô’s formula,
t 1 t ′′ x 2 x
f (Xtx ) − f (x) = ∫ f ′ (Xsx ) dXsx + ∫ f (Xs )σ (Xs ) ds
0 2 0
t t 1
=∫ f ′ (Xsx )σ(Xsx ) dBs + ∫ (f ′ (Xsx )b(Xsx ) + f ′′ (Xsx )σ 2 (Xsx )) ds.
0 0 2
The first term on the right is a martingale, its expectation equals 0. Thus,
E(f (Xtx )) − f (x) 1 t 1
= E [∫ (f ′ (Xsx )b(Xsx ) + f ′′ (Xsx )σ 2 (Xsx )) ds]
t t 0 2
1 t 1
= ∫ E (f ′ (Xsx )b(Xsx ) + f ′′ (Xsx )σ 2 (Xsx )) ds.
t 0 2
Using Theorem 7.22, we get
Tt f (x) − f (x) 1
Af (x) = lim = f ′ (x)b(x) + f ′′ (x)σ 2 (x).
t→0 t 2
Since the right-hand side is in C∞ , Theorem 7.22 applies and shows that C2c (R) ⊂
D(A). Moreover, the same calculation shows that
D(A) ⊃ {u ∈ C∞ (R) ∶ u′ , u′′ ∈ C(R), bu′ + 12 σ 2 u′′ ∈ C∞ }
i. e. the domain of A takes into account the growth of σ and b.

Since A is a second-order differential operator, it clearly has an extension onto Cb2 (R).
(d) Let u ∈ C 1,2 ([0, ∞) × R). By the time-dependent Itô formula we have
u(t, Xtx ) − u(0, x)

t t 1
=∫ ∂x u(s, Xsx ) dXsx + ∫ (∂t u(s, Xsx ) + ∂x2 u(s, Xsx )σ 2 (Xsx )) ds
0 0 2
t t
=∫ ∂x u(s, Xsx )σ(Xsx ) dBs + ∫ (∂t u(s, Xsx ) + ∂x u(s, Xsx )b(Xsx ) +
0 0
1
+ ∂x2 u(s, Xsx )σ 2 (Xsx )) ds.
2
We have shown in part (c) that (for an extension of A)
1
Ax u(s, Xsx ) = b(Xsx )∂x u(s, Xsx ) + σ 2 (Xsx )∂x2 u(s, Xsx ) for all fixed s ⩾ 0.
2
Consequently,
t t
u(t, Xtx ) − u(0, x) = ∫ ∂x u(s, Xsx )σ(Xsx ) dBs + ∫ (∂t u(s, Xsx ) + Ax u(s, Xsx )) ds.
0 0
181
In particular we find that

t
Mtu,x ∶= u(t, Xtx ) − u(0, x) − ∫ (∂t u(s, Xsx ) + Ax u(s, Xsx )) ds
0
is a martingale. By our assumptions, it is a bounded martingale and so we can use

Doob’s optional stopping theorem for the stopping time τ ∧ n
τ ∧n
E u(t, Xτx∧n ) = u(0, x) + E (∫ (∂t u(s, Xsx ) + Ax u(s, Xsx )) ds)
0
and, since everything is bounded and since E τ < ∞, dominated convergence proves
the claim.
Remark: A close inspection of our argument reveals that we do not need boundedness
of b, σ if we replace E τ < ∞ by
τ τ
E (∫ σ 2 (Xsx ) ds) + E (∫ ∣b(Xσx )∣ ds) < ∞.
0 0
∎∎
182
20 Stratonovich’s stochastic calculus
Problem 20.1. Solution: We have, using the time-dependent Itô formula (17.18) with d =
m = 1,
df (t, Xt ) = ∂t f (t, Xt ) dt + ∂x f (t, Xt )b(t) dt + ∂x f (t, Xt )σ(t) dBt + 21 ∂x2 f (t, Xt )σ 2 (t) dt
= ∂t f (t, Xt ) dt + ∂x f (t, Xt )b(t) dt + ∂x f (t, Xt )σ(t) ○dBt

(17.18)
and this is exactly what we would get if we would use normal calculus rules and if dBt =
β̇t dt: By the usual chain rule
f (t, ξt ) = ∂t f (t, ξt ) + ∂x f (t, ξt )ξ˙t

= ∂t f (t, ξt ) + ∂x f (t, ξt )σ(t)β̇t + ∂x f (t, ξt )b(t)
t t
where we used that ξt = ξ0 + ∫0 σ(s)β̇(s) ds + ∫0 b(s) ds.
∎∎
Problem 20.2. Solution: Then we have

⎧
⎪
x dξ
⎪
⎪∫ , if 0 < x0 < x,
⎪
⎪
⎪ x0 σ(ξ)
⎪
⎪
⎪
g(x) = ⎨
⎪
⎪
⎪
⎪
⎪
⎪ x0 dξ
⎪
⎪
⎪− ∫ , if 0 < x ⩽ x0 .
⎩ x σ(ξ)
The Function g and its inverse u = g −1 are shown in the pictures below. Note the difference
between the cases ∫0+ dξ
σ(ξ) = ∞ and < ∞.
Yo
So. *'*
The solution to the SDE is now given by
τ ∶= inf {t ⩾ 0 ∶ u(Bt ) = 0} ,
183
Xt ∶=u(Bt )1{τ >t} .
Note that the fact that σ(0) = 0 means that in the SDE Xt cannot move once it reaches
0. The pictures above illustrate that
dξ
∫ = ∞ Ô⇒ τ = ∞
0+ σ(ξ)
dξ
∫ < ∞ Ô⇒ τ < ∞.
0+ σ(ξ)
∎∎
Problem 20.3. Solution: Denote by Lf , Lg the global Lipschitz constants and observe that
the global Lipschitz property entails linear growth:
∣g(x)∣ ⩽ ∣g(0)∣ + ∣g(x) − g(0)∣ ⩽ ∣g(0)∣ + Lg ∣x∣.
Now let −r ⩽ x, y ⩽ r. Then
∣h(x) − h(y)∣ = ∣f (x)g(x) − f (y)g(y)∣

⩽ ∣f (x)g(x) − f (y)g(x)∣ + ∣f (y)g(x) − f (y)g(y)∣
⩽ ∣f (x) − f (y)∣∣g(x)∣ + ∥f ∥∞ ∣g(x) − g(y)∣
⩽ Lf sup ∣g(x)∣∣x − y∣ + ∥f ∥∞ Lg ∣x − y∣,
∣x∣⩽r
and the local Lipschitz property follows. Finally,
∣h(x)∣ = ∣f (x)∣∣g(x)∣ ⩽ ∥f ∥∞ (∣g(0)∣ + Lg ∣x∣),
and we are done.
∎∎
184
21 On diffusions
1 d d
Au = Lu = ∑ aij ∂i ∂j u + ∑ bi ∂i u
2 i,j=1 i=1
and we know that L ∶ C∞

c → C. Fix R > 0 and i, j ∈ {1, . . . , d} where x = (x1 , . . . , xd ) ∈ R
d
and χ ∈ C∞
c (R ) such that χ∣B(0,R) ≡ 1.
d
For all u, χ ∈ C2 we get

1
L(φu) = ∑ aij ∂i ∂j (φu) + ∑ bi ∂i (φu)
2 i,j i
1
= ∑ aij (∂i ∂j φ + ∂i ∂j u + ∂i φ∂j u + ∂i u∂j φ) + ∑ bi (u∂i φ + φ∂i u)
2 i,j i
= φLu + uLφ + ∑ aij ∂i φ∂j u

i,j
where we used the symmetry aij = aji in the last step.
Now use u(x) = xi and φ(x) = χ(x). Then uχ ∈ C∞

c , L(uχ) ∈ C and so
L(uχ)(x) = bi (x) for all ∣x∣ < R Ô⇒ bi ∣B(0,R) continuous.
Now use u(x) = xi xj and φ(x) = χ(x). Then uχ ∈ C∞

c , L(uχ) ∈ C and so
L(uχ)(x) = aij + xj bi (x) + xi bj (x) for all ∣x∣ < R Ô⇒ aij ∣B(0,R) continuous.
Since R > 0 is arbitrary, the claim follows.
∎∎
Problem 21.2. Solution: This is a straightforward application of the differentiation Lemma

which is familiar from measure and integration theory, cf. Schilling [15, Theorem 11.5, pp.
92–93]: observe that by our assumptions
∂ 2 p(t, x, y)
∣ ∣ ⩽ C(t) for all x, y ∈ Rd
∂xj ∂xk
which shows that for u ∈ Cc∞ (Rd )
∂ 2 p(t, x, y)
∣ u(y)∣ ⩽ C(t) ∣u(y)∣ ∈ L1 (Rd ) (*)
∂xj ∂xk
185
for each t > 0. Thus we get
∂2 ∂2
∫ p(t, x, y) u(y) dy = ∫ p(t, x, y) u(y) dy.
∂xj ∂xk ∂xj ∂xk
Moreover, (*) and the fact that p(t, ⋅, y) ∈ C∞ (Rd ) allow us to change limits and integrals
to get for x → x0 and ∣x∣ → ∞
∂2 ∂2
lim ∫ p(t, x, y) u(y) dy = ∫ lim p(t, x, y) u(y) dy
x→x0 ∂xj ∂xk x→x0 ∂xj ∂xk
∂2
=∫ p(t, x0 , y) u(y) dy
∂xj ∂xk
Ô⇒ Tt maps C∞
c (R ) into C(R );
d d
∂2 ∂2
lim ∫ p(t, x, y) u(y) dy = ∫ lim p(t, x, y) u(y) dy = 0
∣x∣→∞ ∂xj ∂xk ∣x∣→∞ ∂xj ∂xk
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
=0
Ô⇒ Tt maps C∞
c (R ) into C∞ (R ).
d d
Addition: With a standard uniform boundedness and density argument we can show that
Tt maps C∞ into C∞ : fix u ∈ C∞ (Rd ) and pick a sequence (un )n ⊂ C∞
c (R ) such that
d
lim ∥u − un ∥∞ = 0.
n→∞
Then we get
∥Tt u − Tt un ∥∞ = ∥Tt (u − un )∥∞ ⩽ ∥u − un ∥∞ ÐÐÐ→ 0

n→∞
which means that Tt un → Tt u uniformly, i. e. Tt u ∈ C∞ as Tt un ∈ C∞ .
∎∎
Problem 21.3. Solution: Let u ∈ C2∞ . Then there is a sequence of test functions (un )n ⊂ C∞
c
such that ∥un − u∥(2) → 0. Thus, un → u uniformly and A(un − um ) → 0 uniformly. The
closedness now gives u ∈ D(A).
∎∎
Problem 21.4. Solution: Let u, φ ∈ C∞

c (R ). Then
d
⟨Lu, φ⟩L2 = ∑ ∫ aij ∂i ∂j u ⋅ φ dx + ∑ ∫ bj ∂j u ⋅ φ dx + ∫ cu ⋅ φ dx

i,j Rd j Rd Rd
= ∑∫ u ⋅ ∂i ∂j (aij φ) dx − ∑ ∫ u ⋅ ∂j (bj φ) dx + ∫ u ⋅ cφ dx
int by
parts
i,j Rd j Rd Rd
= ⟨u, L∗ φ⟩L2
where
L∗ (x, Dx )φ(x) = ∑ ∂i ∂j (aij (x)φ(x)) − ∑ ∂j (bj (x)φ(x)) + c(x)φ(x).
ij j
186
Now assume that we are in (t, x) ∈ [0, ∞) × Rd —the case R × Rd is easier, as we have no
boundary term. Consider L + ∂t = L(x, Dx ) + ∂t for sufficiently smooth u = u(t, x) and
φ = φ(t, x) with compact support in [0, ∞) × Rd . We find
∞
∫ ∫ (L + ∂t )u(t, x) ⋅ φ(t, x) dx dt
0 Rd
∞ ∞
=∫ ∫ Lu(t, x) ⋅ φ(t, x) dx dt + ∫ ∫ ∂t u(t, x) ⋅ φ(t, x) dx dt
0 Rd 0 Rd
∞ ∞
=∫ ∫ Lu(t, x) ⋅ φ(t, x) dx dt + ∫ ∫ ∂t u(t, x) ⋅ φ(t, x) dt dx
0 Rd Rd 0
∞ ∞ ∞
=∫ ∫ u(t, x) ⋅ L∗ φ(t, x) dx dt + ∫ (u(t, x)φ(t, x)∣ −∫ u(t, x) ⋅ ∂t φ(t, x) dt) dx
0 Rd Rd t=0 0
∞ ∞
=∫ ∫ u(t, x) ⋅ L∗ φ(t, x) dx dt − ∫ (u(0, x)φ(0, x) + ∫ u(t, x) ⋅ ∂t φ(t, x) dt) dx.
0 Rd Rd 0
This shows that (L(x, Dx ) + ∂t )∗ = L∗ (x, Dx ) − ∂t − δ(0,x) .

∎∎
Problem 21.5. Solution: Using Lemma 7.10 we get for all u ∈ C∞

c (R )
d
d
Tt u(x) = Tt L(⋅, D)u(x)
dt
d
Ô⇒ ∫ p(t, x, y) u(y) dy = ∫ p(t, x, y) L(y, Dy )u(y) dy
dt
d
Ô⇒ ∫ p(t, x, y) u(y) dy = ∫ p(t, x, y) L(y, Dy )u(y) dy.
dt
The change of differentiation and integration can easily be justified by a routine application
of the differentiation lemma (e.g. Schilling [15, Theorem 11.5, pp. 92–93]): under our
assumptions we have for all ∈ (0, 1) and R > 0
d
sup sup ∣ p(t, x, y) u(y)∣ ⩽ C(, R) ∣u(y)∣ ∈ L1 (Rd ).
t∈[,1/] ∣x∣⩽R dt
Inserting the expression for the differential operator L(y, Dy ), we find for the right-hand
side
∫ p(t, x, y) L(y, Dy )u(y) dy

1 d ∂ 2 u(y) d
∂u(y)
= ∑ ∫ p(t, x, y) ⋅ ajk (y) dy + ∑ ∫ p(t, x, y) ⋅ bj (y) dy
2 j,k=1 ∂yj ∂yk j=1 ∂yj
1 d ∂2 d
∂
= (ajk (y) ⋅ p(t, x, y))u(y) dy + ∑ ∫ (bj (y) ⋅ p(t, x, y))u(y) dy
int. by
∑ ∫
parts 2 j,k=1 ∂yj ∂yk j=1 ∂yj
= ∫ L∗ (y, Dy )p(t, x, y) u(y) dy
and the claim follows since u ∈ C∞

c (R ) is arbitrary.
d
∎∎
Problem 21.6. Solution: Problem 6.2 shows that Xt is a Markov process. The continuity of
the sample paths is obvious and so is the Feller property (using the form of the transition
function found in the solution of Problem 6.2).
187
t
Let us calculate the generator. Set It = ∫0 Bs ds. The semigroup is given by
t
Tt u(x, y) = Ex,y u(Bt , It ) = E u (Bt + x, ∫0 (Bs + x) ds + y) = E u(Bt + x, It + tx + y).
If we differentiate the expression under the expectation with respect to t, we get with the
help of Itô’s formula
du(Bt + x, It + tx + y) = ∂x u(Bt + x, It + tx + y) dBt

+ ∂y u(Bt + x, It + tx + y) d(It + tx)
1
+ ∂x2 u(Bt + x, It + tx + y) dt
2
= ∂x u(Bt + x, It + tx + y) dBt
+ ∂y u(Bt + x, It + tx + y)(Bt + x) dt
1
+ ∂x2 u(Bt + x, It + tx + y) dt
2
since dBs dIs = 0. So,
t
E u(Bt + x, It + tx + y) − u(x, y) = ∫ E [∂y u(Bs + x, Is + sx + y)(Bs + x)] ds
0
1 t
+ ∫ E [∂x u(Bs + x, Is + sx + y)] ds.
2
2 0
Dividing by t and letting t → 0 we get
1 2
Lu(x, y) = x ∂y u(x, y) + ∂ u(x, y).
2 x
∎∎
Problem 21.7. Solution: We assume for a) and b) that the operator L is more general than
written in (21.1), namely
1 d ∂ 2 u(x) d ∂u(x)
Lu(x) = ∑ aij (x) + ∑ bj (x) + c(x)u(x)
2 i,j=1 ∂xi ∂xj j=1 ∂xj
where all coefficients are continuous functions.
(a) If u has compact support, then Lu has compact support. Since, by assumption, the
coefficients of L are continuous, Lu is bounded, hence Mtu is square integrable.
Obviously, Mtu is Ft measurable. Let us establish the martingale property. For this
we fix s ⩽ t. Then
t
Ex (Mtu ∣ Fs ) = Ex (u(Xt ) − u(X0 ) − ∫ Lu(Xr ) dr ∣ Fs )
0
t
= Ex (u(Xt ) − u(Xs ) − ∫ Lu(Xr ) dr ∣ Fs )
s
s
+ u(Xs ) − u(X0 ) − ∫ Lu(Xr ) dr
0
t−s
= Ex (u(Xt ) − u(Xs ) − ∫ Lu(Xr+s ) dr ∣ Fs ) + Msu
0
188
t−s
= EXs (u(Xt−s ) − u(X0 ) − ∫ Lu(Xr ) dr) + Msu .
Markov
property 0
Observe that Tt u(y) = Ey u(Xt ) is the semigroup associated with the Markov process.
Then
t−s
Ey (u(Xt−s ) − u(X0 ) − ∫ Lu(Xr ) dr)
0
t−s
= Tt−s u(y) − u(y) − ∫ Ey (Lu(Xr )) dr = 0
0
by Lemma 7.10, see also Theorem 7.30. This shows that Ex (Mtu ∣ Fs ) = Msu , and we
are done.
(b) Fix R > 0, x ∈ Rd , and pick a smooth cut-off function χ = χR ∈ C∞

c (R ) such that
d
χ∣B(x, R) ≡ 1. Then for all f ∈ C2 (Rd ) we have χf ∈ C2c (Rd ) and it is not hard to see
that the calculation in part a) still holds for such functions.
Set τ = τRx = inf{t > 0 ∶ ∣Xt − x∣ ⩾ R}. This is a stopping time and we have
f (Xtτ ) = χ(Xtτ )f (Xtτ ) = (χf )(Xtτ ).
Moreover,
1
L(χf ) = ∑ aij ∂i ∂j (χf ) + ∑ bi ∂i (χf ) + cχf
2 i,j i
1
= ∑ aij (f ∂i ∂j χ + χ∂i ∂j f + ∂i χ∂j f + ∂i f ∂j χ) + ∑ bi (f ∂i χ + χ∂i f ) + cχf
2 i,j i
= χLf + f Lχ + ∑ aij ∂i χ∂j f − cχf

i,j
where we used the symmetry aij = aji in the last step.
This calculation shows that L(χf ) = Lf on B(x, R).
By optional stopping and part a) we know that (Mt∧τ

χf
R
, Ft )t⩾0 is a martingale. More-
over, we get for s ⩽ t
Ex (Mt∧τ
f
R
∣ Fs ) = Ex (Mt∧τ
χf
R
∣ Fs )
= Ms∧τ
χf
R
= Ms∧τ
f
R
.
Since (τR )R is a localizing sequence, we are done.
(c) A diffusion operator L satisfies that c = 0. Thus, the calculation for L(χf ) in part
b) shows that
L(uφ) − uLφ − φLu = ∑ aij ∂i u∂j φ = ∇u(x) ⋅ a(x)∇φ(x).

ij
This proves the second equality in the formula of the problem.
189
For the first we note that d⟨M u , M φ ⟩t = dMtu dMtφ (by the definition of the bracket
process) and the latter we can calculate with the rules for Itô differentials. We have
dXtj = ∑ σjk (Xt )dBtk + bj (Xt ) dt

k
and, by Itô’s formula,
du(Xt ) = ∑ ∂j u(Xt ) dXtj + dt-terms = ∑ ∂j u(Xt )σjk (Xt ) dBtk + dt-terms.

j j,k
By definition,
dMtu = du(Xt ) − Lu(Xt ) dt = ∑ ∂j u(Xt )σjk (Xt ) dBtk + dt-terms.

j,k
Thus, using that all terms containing (dt)2 and dBtk dt are zero, we get
dMtu dMtφ = ∑ ∑ ∂j u(Xt )∂l φ(Xt )σjk (Xt )σlm (Xt ) dBtk dBtm
j,k l,m
= ∑ ∑ ∂j u(Xt )∂l φ(Xt )σjk (Xt )σlm (Xt ) δkm dt

j,k l,m
= ∑ ∂j u(Xt )∂l φ(Xt ) ∑ σjk (Xt )σlk (Xt ) dt

j,l k
= ∑ ∂j u(Xt )∂l φ(Xt )ajl dt

j,l
= ∇u(Xt ) ⋅ a(Xt )∇φ(Xt )
where ajl = ∑k σjk (Xt )σlk (Xt ) = (σσ ⊺ )jl . (x ⋅ y denotes the Euclidean scalar product
and ∇ = (∂1 , . . . , ∂d )⊺ .)
Alternative proof of the first equality: Let u ∈ C∞

c (R ), x ∈ R . Without loss of
d d
generality we may assume u(x) = 0. The equality

t 2
u2 (Xt ) = (Mtu + ∫ Lu(Xr ) dr)
0
implies
t t 2
(Mtu )2 = u2 (Xt ) − 2Mtu ∫ Lu(Xr ) dr − (∫ Lu(Xr ) dr) .
0 0
Part a) shows that
t
u2 (Xt ) − ∫ L(u2 )(Xr ) dr
0
is a martingale. Moreover, since (Mtu )t⩾0 is a martingale, we obtain by the tower
property
t
Ex (2Mtu ∫ Lu(Xr ) dr ∣ Fs )
0
s t
= 2 Ex (Mtu ∣ Fs ) ∫ Lu(Xr ) dr + 2 ∫ Ex ( Ex (Mtu Lu(Xr ) ∣ Fr ) ∣ Fs ) dr
0 s
s t
= 2Msu ∫ Lu(Xr ) dr + 2 Ex (∫ Mru Lu(Xr ) dr ∣ Fs ) . (*)
0 s
190
By the definition of Mtu ,

t
2 ∫ Mru Lu(Xr ) dr
s
t t r
= 2∫ u(Xr )Lu(Xr ) dr − 2 ∫ ∫ Lu(Xv ) dv Lu(Xr ) dr
s s 0
t t s
= 2∫ u(Xr )Lu(Xr ) dr − 2 ∫ ∫ Lu(Xv )Lu(Xr ) dv dr
s s 0
t r
− 2∫ ∫ Lu(Xv )Lu(Xr ) dv dr
s s
t t 2 s 2
= 2∫ u(Xr )Lu(Xr ) dr − (∫ Lu(Xr ) dr) + (∫ Lu(Xr ) dr) (**)
s 0 0
using that
t r t 2
2∫ ∫ Lu(Xv )Lu(Xr ) dv dr = (∫ Lu(Xr ) dr) .
s s s
Combining (*) and (**), we see that

t t 2 t
2Mtu ∫ Lu(Xr ) dr + (∫ Lu(Xr ) dr) − 2 ∫ u(Xr )Lu(Xr ) dr
0 0 0
is a martingale. Consequently,
t
⟨M u ⟩t = ∫ (L(u2 ) − 2uLu)(Xr ) dr.
0
This proves the first equality for u = φ. The formula for the quadratic covariation
⟨M u , M φ ⟩ follows by using polarization, i. e.
1 1
⟨M u , M φ ⟩t = (⟨M u + M φ ⟩t − ⟨M u − M φ ⟩t ) = (⟨M u+ϕ ⟩t − ⟨M u−ϕ ⟩t ).
4 4
∎∎
191
Bibliography
[1] Böttcher, B., Schilling, R.L., Wang, J.: Lévy-Type Processes: Construction, Approxi-
mation and Sample Path Properties. Springer Lecture Notes in Mathematics vol. 2099,
(vol. III of the “Lévy Matters” subseries). Springer, Berlin 2014.
[2] Chung, K.L.: Lectures from Markov Processes to Brownian Motion. Springer, New York
1982. Enlarged and reprinted as [3].
[3] Chung, K.L., Walsh, J.B.: Markov Processes, Brownian Motion, and Time Symmetry
(Second Edition). Springer, Berlin 2005. This is the 2nd edn. of [2].
[4] Dupuis, C.: Mesure de Hausdorff de la trajectoire de certains processus à accroissements

indépendants et stationnaires. In: Séminaire de Prohabilités, Springer, Lecture Notes in
Mathematics 381, Berlin 1974, 40–77.
[5] Engel, K.-J., Nagel, R.: One-Parameter Semigroups for Linear Evolution Equations.
Springer, New York 2000.
[6] Falconer, K.: The Geometry of Fractal Sets. Cambridge University Press, Cambridge
1985.
[7] Falconer, K.: Fractal Geometry. Mathematical Foundations and Applications. Second Edi-
tion. Wiley, Chichester 2003.
[8] Gradshteyn, I.S., Ryzhik, I.M.: Table of Integrals, Series, and Products (4th corrected and
enlarged edn). Academic Press, San Diego 1980.
[9] Karatzas, I., Shreve, S.E.: Brownian Motion and Stochastic Calculus. Springer, New York
1988.
[10] Pazy, A.: Semigroups of Linear Operators and Applications to Partial Differential Equa-
tions. Springer, New York 1983.
[11] Protter, P.E.: Stochastic Integration and Differential Equations (Second Edn.). Springer,
Berlin 2004.
[12] A. Rényi: Probability Theory. Dover, Mineola (NY) 2007. (Reprint of the 1970 edition
published by North-Holland and Akadémiai Kiadó.)
193
[13] Rudin, W.: Principles of Mathematical Analysis. McGraw-Hill, Auckland 1976 (3rd edn).
German Translation: Analysis. 4. Auflage. Oldenbourg, München 2009.
[14] Rudin, W.: Real and Complex Analysis. McGraw-Hill, New York 1986 (3rd edn).
German Translation: Reelle und Komplexe Analysis. Oldenbourg, München 2009.
[15] Schilling, R.L.: Measures, Integrals and Martingales. Cambridge University Press, Cam-
bridge 2011 (3rd printing with corrections).
[16] Schilling, R.L. and Partzsch, L.: Brownian Motion. An Introduction to Stochastic Pro-
cesses. De Gruyter, Berlin 2012.
[17] Trèves, F.: Topological Vector Spaces, Distributions and Kernels. Academic Press, San
Diego (CA) 1990.
[18] Wentzell, A. D.: Theorie zufälliger Prozesse. Birkhäuser, Basel 1979. Russian original.
English translation: Course in the Theory of Stochastic Processes, McGraw-Hill 1981.
194

Brownian Motion

Uploaded by

Copyright:

Brownian Motion

Uploaded by

Document Information

Original Description:

Copyright

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Brownian Motion

Uploaded by

Copyright:

Brownian Motion (2nd edition)

An Introduction to Stochastic Processes

René L. Schilling & Lothar Partzsch

Dresden, June 2014. Last update December 19, 2017

Acknowledgement. We are grateful to Björn Böttcher, Katharina Fischer, Julian Hollender,

Dresden, June 2014 René Schilling

1 Robert Brown’s new thing 5

2 Brownian motion as a Gaussian process 13

3 Constructions of Brownian motion 25

4 The canonical model 35

5 Brownian motion as a martingale 41

6 Brownian motion as a Markov process 53

7 Brownian motion and transition semigroups 65

8 The PDE connection 85

9 The variation of Brownian paths 93

10 Regularity of Brownian paths 101

11 Brownian motion as a random fractal 107

12 The growth of Brownian paths 113

13 Strassen’s functional law of the iterated logarithm 117

14 Skorokhod representation 125

15 Stochastic integrals: L2 –theory 127

16 Stochastic integrals: Beyond L2T 139

17 Itô’s formula 143

18 Applications of Itô’s formula 155

19 Stochastic differential equations 165

20 Stratonovich’s stochastic calculus 183

Problem 1.1. Solution:

If we take ξ = 0 and η = 0, respectively, we see that

E exp [i⟨ξ, X⟩ + i⟨η, Y ⟩] = lim E exp [i⟨ξ, Xn ⟩ + i⟨η, Yn ⟩]

= lim E exp [i⟨ξ, Xn ⟩] E exp [i⟨η, Yn ⟩]

= lim E exp [i⟨ξ, Xn ⟩] lim E exp [i⟨η, Yn ⟩]

= E exp [i⟨ξ, X⟩] E exp [i⟨η, Y ⟩]

and this shows that X á Y .

A simple direct calculation shows that 1 − X ∼ 21 (δ0 + δ1 ) ∼ Y . Thus,

X + Y ∼ 12 (δ0 + δ1 ) ∗ 21 (δ0 + δ1 ) = 41 (δ0 ∗ δ0 + 2δ1 ∗ δ0 + δ1 ∗ δ1 ) = 41 (δ0 + 2δ1 + δ2 ).

lim E eiξ(Xn +Yn ) = lim E eiξXn E eiξYn

= lim E eiξXn lim E eiξYn

is bounded and continuous, i. e. we get directly

lim E eiξ(Xn +Yn ) lim E f (Xn , Yn ) = E f (X, Y ) = E eiξ(X+Y ) .

Notice that the independence and d-convergence of the sequences Xn , Yn already

lim φXn (ξ) = lim E eiξXn = E eiξX = φX (ξ).

lim φXn ,Yn (ξ, η) = lim E eiξXn +iηYn

= lim E eiξXn E eiηYn

= lim E eiξXn lim E eiηYn

h(ξ, η) = φX,Y (ξ, η) = φX (ξ)φY (η)

which shows that X á Y .

Problem 1.2. Solution: Using the elementary estimate

we see that the function t ↦ ei⟨ξ,t⟩ , ξ, t ∈ Rd is locally Lipschitz continuous:

∣ei⟨ξ,t⟩ − ei⟨ξ,s⟩ ∣ = ∣ei⟨ξ,t−s⟩ − 1∣ ⩽ ∣⟨ξ, t − s⟩∣ ⩽ ∣ξ∣ ⋅ ∣t − s∣ for all ξ, t, s ∈ Rd ,

E ei⟨ξ,Yn ⟩ = E [ei⟨ξ,Yn −Xn ⟩ ei⟨ξ,Xn ⟩ ]

= E [(ei⟨ξ,Yn −Xn ⟩ − 1)ei⟨ξ,Xn ⟩ ] + E ei⟨ξ,Xn ⟩ .

∣ E [(ei⟨ξ,Yn −Xn ⟩ − 1)ei⟨ξ,Xn ⟩ ]∣

⩽ E ∣(ei⟨ξ,Yn −Xn ⟩ − 1)ei⟨ξ,Xn ⟩ ∣

=∫ ∣ei⟨ξ,Yn −Xn ⟩ − 1∣ d P + ∫ ∣ei⟨ξ,Yn −Xn ⟩ − 1∣ d P

= δ ∣ξ∣ + 2 P (∣Yn − Xn ∣ > δ)

This proves (a) and (b).

In order to show that (Xn , Yn ) converges in distribution, we use Lévy’s characterization of

⩽ E ∣ei(ξXn +ηYn ) − E ei(ξXn +ηc) ∣ + ∣E ei(ξXn +ηc) − E ei(ξX+ηc) ∣

⩽ E ∣eiηYn − eiηc ∣ + ∣E eiξXn − E eiξX ∣ .

E eiξ(Xn +Yn ) − E eiξX = ( E eiξ(Xn +Yn ) − E eiξXn ) + ( E eiξXn − E eiξX ) .

∣ E eiξ(Xn +Yn ) − E eiξXn ∣ ⩽ E ∣eiξXn (eiξYn − 1)∣

where we use P-convergence in the penultimate step.

= lim E eiξXn eiηY

⩽ P(∣f (Xn ) − f (X)∣ > ) + P(∣X∣ ⩾ R) + P(∣f (Xn )∣ ⩾ R)