Beruflich Dokumente
Kultur Dokumente
Lecturer: Xue-Mei Li
1 Problem Sheet 1
In these solution I avoid using conditional expectations. But do try to give alterna-
tive proofs once we learnt conditional expectations.
d |x−y|2
Exercise 1 For x, y ∈ Rd define pt (x, y) = (2πt)− 2 e− 2t . Prove that Pt (x, dy) =
pt (x, y)dy satisfies the Chapman-Kolmogorov equation: for Γ Borel subset of Rd ,
Z
Pt+s (x, Γ) = Ps (y, Γ)Pt (x, dy).
Rd
1
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 2
tz+sx
next we change the variable y − t+s to ỹ, then
Z
Ps (y, Γ)Pt (x, dy)
Rd
2
Z Z
1 t+s 1 |x−z|
− d2 − d2 2
= (2πt) (2πs) e− 2 st |ỹ| e− 2 t−s dzdỹ
d
ZR Γ Z 2
d d 1 t+s 2 1 |x−z|
= (2πt)− 2 (2πs)− 2 e− 2 st |ỹ| dỹ e− 2 t−s dz
d
Z R |x−z|2 Γ
−
= (2π(t − s))−f d2 e 2(t−s) dz.
Γ
(3) For every number p > 0 there exists a constant c(p) such that
p
E|Xt − Xs |p = c(p)|t − s| 2 .
(4) State Kolomogorov’s continuity theorem and conclude that for almost surely
all ω, Xt (ω) is locally Hölder continuous with exponent α for any number
α < 1/2.
(2) Let us fix 0 = t0 < t1 < t2 < . . . < tn and Borel sets Ai ∈ B(R),
i = 1, . . . , n. Let fi (x) = 1x∈Ai where Ai are Borel measurable set. Then we
obtain
where in the last line we have used the identity (ii). Introducing new variables:
y1 = x1 , y2 = x2 − x1 , . . ., yn = xn−1 − xn , we obtain
Z Z
P(Xt1 ∈ A1 , . . . , Xtn − Xtn−1 ∈ An ) = ... f1 (y1 )f2 (y2 ) . . . fn (yn )
R R
× pt1 (0, y1 )pt2 −t1 (0, y2 ) . . . ptn −tn−1 (0, yn ) dyn . . . dy1
n Z
Y
= fi (yi )pti −ti1 (0, yi ) dyi .
i=1 R
Exercise 3 If (Bt ) is a Brownian motion prove that (Bt ) is a Markov process with
transition function pt (x, y)dy.
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 4
Solution: Let us denote for simplicity fi (x) = 1x∈Ai . Furthermore, we define the
random variables Yi = Xti − Xti−1 , for i = 1, . . . , n, where we have postulated
t0 = 0. From the properties of the Brownian motion we obtain that the random
variables {Yi } are independent and moreover Yi ∼ N (0, ti − ti−1 ). Thus, we have
n
Y
P[Xt1 ∈ A1 , . . . , Xtn ∈ An ] = E fi (Xti )
i=1
= E[f1 (Y1 )f2 (Y2 + Y1 ) . . . fn (Yn + . . . + Y1 )]
Z Z
= ... f1 (y1 )f2 (y2 + y1 ) . . . fn (yn + . . . + y1 )
R R
× pt1 (0, y1 )pt2 −t1 (0, y2 ) . . . ptn −tn−1 (0, yn ) dyn . . . dy1 .
Noticing that pt (0, y − x) = pt (x, y) and recalling the definition of the functions
fi , so the finite dimensional distribution agrees with that of the Markov process
determined by the heat kernels. The two processes must agree.
(ii) For any number n ∈ N, any sets Ai ∈ B(R), i = 1, . . . , n, and any 0 <
t1 < t2 < . . . < tn ,
P[Xt1 ∈ A1 , . . . , Xtn ∈ An ]
Z Z
= ... pt1 (0, y1 )pt2 −t1 (y1 , y2 ) . . . ptn −tn−1 (yn−1 , yn ) dyn . . . dy1 .
A1 Ak
We observe that
t 2 2t t2
|x| = x − y + hx, yi − 2 |y|2
2
T T T
and that
2
2t(T − t) 2 (T − t)2 2
2
t (T − t)
|x − y| = x − y − 2 hx, yi + |y| + |y| .
T T T2 T2
Thus,
|x|2 |y − x|2 t 2 t 2 1 2
1 1
+ = x − y + x − y + |y| .
t (T − t) t T T −t T T
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 6
Finally
t
Ef (Bt − BT )
T Z Z
t 2 2 |y|2
f x − y e− 2t |x− T y| e 2(T −t) | T | e− 2T dxdy
− d2 − d2 1 t − 1 x− t y
= (2πt) (2π(T − t))
d d T
ZR R Z
d d 1 1 2 |y|2
|z|2 − 2(T −t) |z|
= (2πt)− 2 (2π(T − t))− 2 f (z)e − 2t
e dz e− 2T dy
Z Rd Rd
d
= f (z)pt (0, z)pT −t (z, 0)dz · (2πT ) 2 .
Rd
Finally we see
t pt (0, z)pT −t (z, 0)
Bt − BT ∼ dz.
T pT (0, 0)
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 7
Solution: (a) (i) Since the distribution of Bt is N (0, t), we have E[Bt ] = 0 and
E[Bt2 ] = t.
(a) (ii) We fix any t ≥ s ≥ 0. Then we have
Since the Brownian motion has independent increments, the random variables Bt −
Bs and Bs are independent and we have
E[Bs Bt ] = s ,
1 a(t − s)
Wt − Ws = √ (Bat − Bas ) ∼ N (0, ) = N (0, t − s) ,
a a
(c) If we denote Wt = Bt0 +t − Bt0 , then W has continuous sample paths and
independent increments, which follows from the same properties of B. Moreover,
for any t > s ≥ 0, we have
is uniformly integrable.
Solution: We note that, if a set of measurable sets AC satisfies limC→∞ P[AC ] =
0, then limC→∞ E[1AC |X|] = 0 ( since X ∈ L1 , dominated convergence).
Let us take any G ⊂ F and consider the family of events
A(C, G) = {ω : |E[X|G](ω)| ≥ C} .
For any ε > 0, there exist a δ > 0 such that if P[A] < δ, then E(1A |X|) < ε.
For this δ, take C > E|X|
δ , then P[A(C, G)] < δ for any G ⊂ F, which implies
Finally, we conclude
Exercise 10 Let (Gt , t ≥ 0), (Ft , t ≥ 0) be filtrations with the property that Gt ⊂
Ft for each t ≥ 0. Suppose that (Xt ) is adapted to (Gt ). If (Xt ) is an (Ft )-
martingale prove that (Xt ) is an (Gt )-martingale.
Solution: The fact that (Xt ) is (Gt )-adapted, follows from the inclusion Gt ⊂ Ft
for each t ≥ 0. Furthermore, for any t > s ≥ 0, using the tower property of the
conditional expectation, we obtain
Solution: First, we will prove that the function (1) is Borel-measurable. To this
end, we take any Borel set A ∈ B(R) and we have to show that the set {(t, ω) :
H(t, ω) ∈ A} is measurable in the product space (R+ , B(R+ )) × (Ω, F). We can
rewrite this set in the following way:
{(t, ω) : H(t, ω) ∈ A} = ({0} × {ω : H0 (ω) ∈ A})
∪ni=1 ((ti , ti+1 ] × {ω : Hi (ω) ∈ A}) .
Since the sets {0} and (ti , ti+1 ], i = 1, . . . , n, belong to B(R+ ), and {ω : Hi (ω) ∈
A} ∈ Fti ⊂ F, the claim now follows from the fact that the product of two
measurable sets is measurable in the product space. Rt
Next, we will show the identities (2). Let us denote It = 0 Hs dBs . Then we
have
n
X n
X
E[It ] = E Hi (Bti+1 ∧t − Bti ∧t ) = E [Hi ] E Bti+1 ∧t − Bti ∧t = 0 ,
i=1 i=1
where in the second equality we have used the independence of Bti+1 ∧t − Bti ∧t
from Fti , which follows from the properties of the Brownian motion and the fact
Bti+1 ∧t − Bti ∧t = 0 if ti ≥ t. Furthermore, in the last equality we have used
that
E Bti+1 ∧t − Bti ∧t = 0.
For the variance of the stochastic integral we have
n
X
E[It2 ] = E Hi2 (Bti+1 ∧t − Bti ∧t )2
i=1
X
+ E Hi Hj (Bti+1 ∧t − Bti ∧t )(Btj+1 ∧t − Btj ∧t ) .
i6=j
In the same way as before, using the independence of the increments of the Brow-
nian motion, we obtain that the second sum is 0. Thus, we have
n
X
E[It2 ] = E Hi2 (Bti+1 ∧t − Bti ∧t )2
i=1
n
X
E Hi2 E (Bti+1 ∧t − Bti ∧t )2
=
i=1
n
X Z t
E Hi2 (ti+1 ∧ t − ti ∧ t) = E (Hs )2 ds ,
=
i=1 0
where in the second line we have used the fact that Hi2 and (Bti+1 ∧t − Bti ∧t )2 are
independent.
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 11
which is finite.
where the inequality follows from |Mn (t)|1A ≤ |Mn (t)| and the last equality
follows from the fact that {Mn (t), n ∈ N} are uniformly integrable.
Since, limn→∞ Mn (t)1A = Mt 1A almost surely, the uniform integrability
implies convergence in L1 . Take A = Ω, we see that Mt is integrable. For any
0 ≤ s < t ≤ 1 and any A ∈ Fs we have
E[M (t)1A ] = lim E[Mn (t)1A ] = lim E[Mn (s)1A ] = E[M (s)1A ] ,
n→∞ n→∞
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 13
where the second equality holds, because Mn is a martingale. This implies that
Exercise 17 Let a > 0 be a real number. For 0 = t1 < · · · < tn < tn+1 < . . .
and i = 1, . . . , n, . . . let Hi be bounded Fti measurable functions and, let H0 be a
bounded F0 -measurable function. Define
∞
X
Ht (ω) = H0 (ω)1{0} (t) + Hi (ω)1(ti ,ti+1 ] (t).
i=1
Mti+1 ∧t − Mti ∧t = 0, ∀i ≥ n + 1.
Z t n
X
Hs dMs ≡ Hi (Mti+1 ∧t − Mti ∧t ),
0 i=1
Take t ≥ s ≥ 0 and assume that s ∈ [tk , tk+1 ), for some k ∈ {0, . . . , n}.
Explanation: We only need to consider two cases: (1) i ≤ k in which case
Hi ∈ Fs in which case we can take Hi out of the conditional expectation and (2)
i ≥ k + 1 in which case s < ti and we may use tower property, to condition in
addition w.r.t Fti and take Hi out.
Then we have
n
X
E[It |Fs ] = E[Hi (Mti+1 ∧t − Mti ∧t )|Fs ]
i=1
k
X n
X
= E[Hi (Mti+1 ∧t − Mti ∧t )|Fs ] + E[Hi (Mti+1 ∧t − Mti ∧t )|Fs ] .
i=1 i=k+1
E[It |Fs ] = Is ,
(2) If T is stopping time, then there exists a sequence of stopping times Tn such
that Tn takes only a finite number of values and Tn decreases to T .
{S ∧ T ≤ t} = {S ≤ t} ∪ {T ≤ t} ∈ Ft ,
{S ∨ T ≤ t} = {S ≤ t} ∩ {T ≤ t} ∈ Ft ,
{aS ≤ t} = {S ≤ t/a} ∈ Ft/a ⊂ Ft ,
Exercise 19 Prove that T is a stopping time iff {T < t} ∈ Ft , for any t > 0.
Solution: If T is a stopping time, then {T < t} = ∪n≥1 {T ≤ t − n1 } ∈ Ft ,
because {T ≤ t − n1 } ∈ Ft−1/n ⊂ Ft .
Conversely, if {T < t} ∈ Ft , for any t > 0, then
1
{T ≤ t} = ∩n≥1 {T < t + } ∈ Ft+ = Ft ,
n
because the filtration is right-continuous.
E[Nt |Fs ] = E[Y (Mt − Mt∧τ )1τ >s |Fs ] + E[Y (Mt − Mt∧τ )1τ ≤s |Fs ]
= I1 + I2 .
For the second term we can get, again by the optional stopping theorem,
Exercise 21 Show that for s < t and A ∈ Fs , τ = s1A + t1Ac is a stopping time.
Solution: Indeed, for any r ≥ 0 we have
∅, if r < s
{τ ≤ r} = A, if s ≤ r < t ∈ Fr .
Ω, if r ≥ t
Exercise 22 Let 0 = t1 < · · · < tn+1 < . . . with limn→∞ tn = ∞. For each
i = 0, 1, . . . , let Hi be a real valued Fti measurable random variable and H0 an
F0 -measurable random variable. For t > 0, we define
∞
X
Xt (ω) = H0 (ω)1{0} (t) + Hi (ω)1(ti ,ti+1 ] (t).
i=0
Solution: To this end, we take any Borel set A ∈ B(R) and we have to show that
for any t ≥ 0, the set {(s, ω) : H(s, ω) ∈ A} is measurable in the product space
([0, t], B([0, t])) × (Ω, Ft ). We can rewrite this set in the following way:
Since the sets {0} and (ti , ti+1 ], i = 1, . . . , n, belong to B([0, t]), and {ω :
Hi (ω) ∈ A} ∈ Fti ⊂ F, the claim now follows from the fact that the product
of two measurable sets is measurable in the product space.
Exercise 23 Let s R< t ≤ u < vR and let (HsR) and (Ks ) be two elementary pro-
t t s
cesses. We define: s Hr dBr = 0 Hr dBr − 0 Hr dBr . Prove that
Z t Z v
E Hr dBr Kr dBr = 0.
s u
Rt
Solution: Recall that the stochastic process ( 0 Hr dBr , t ≥ 0) is measurable w.r.t.
Ft and is a martingale. We use the tower property to obtain the following:
Z t Z v Z t Z v
Hr dBr Kr dBr = E E Hr dBr Kr dBr Fu
E
s u s u
Z t Z v
=E Hr dBr E Kr dBr Fu = 0 ,
s u
because the stochastic integral is a martingale, and hence the inner expectation
vanishes.
(n) (n)
Firstly, we have maxj f (tj ) − f (tj−1 ) → 0, as n → ∞, because of the uniform
continuity of f on [0, t]. Secondly, we can estimate
Nn Z t(n) Nn Z t(n)
Nn
j j
X X
X (n) (n) 0
0
f (tj ) − f (tj−1 ) = (n) f (r) dr ≤ f (r) dr
tj−1 (n)
j=1 j=1
t
j=1 j−1
Z t
0
= f (r) dr < ∞ ,
0
which follows from the properties of f . Thus, from these facts the claim follows.
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 19
almost surely, as n → ∞. This shows that E[Mt |Fs ] = Ms , what means that (Mt )
is a martingale.
Exercise 27 Let {Bt }t≥0 be one dimensional Brownian Motion starting at x (B0 =
x), and let a < x < b. In this question we are going to find the probability of the
Brownian Motion hitting b before a using the Optional Stopping Theorem (OST).
Set
(a) Give an easy arguments why Ta , Tb and T are all stopping times with respect
to the natural filtration of the Brownian Motion.
(b) One would like to compute E[BT ] using OST, but (Bt , t ≥ 0) is not a uni-
formly integrable martingale and apply OST would require for T to be a
bounded stopping time. Instead we are using the limiting argument.
Note, that 1T <S MT is FS -measurable. Thus, for the first term we have
One can see that {T ≥ S} ∈ FS∧T . Indeed, for any t ≥ 0 one has
where we have used the Doob’s optional stopping theorem. Combining all these
equalities together, we obtain the claim.
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 21
Exercise 29 Prove the following: (1) A positive right continuous local martingale
(Mt , t ≥ 0) with M0 = 1 is a super-martingale. (2) A positive right continuous
local martingale (Mt , t ≥ 0) is a martingale if E|M0 | < ∞ and EMt = EM0 for
all t > 0.
Solution: (1) Since E|M0 | < ∞, there is a sequence of increasing stopping times
Tn , such that the process MtTn is a martingale. In particular, MtTn is integrable.
For s < t, we use Fatou’s lemma to obtain
E[Mt |Fs ] = E[lim inf Mt∧Tn |Fs ] ≤ lim inf E[Mt∧Tn |Fs ] = lim inf Ms∧Tn = Ms .
n→∞ n→∞ n→∞
(2) Let S ≤ T be two stopping times, bonded by a constant K. Then, from (1), we
have
which implies EMT = EMS . We know that EMT = EMS for any two bounded
stopping times t ≤ T implies that (Mt ) is a martingale, completing the proof.
(1) Let (At ) and (A0t ) be two continuous stochastic processes of finite variation
with initial values 0 and such that (Mt Nt − At ) and (Mt Nt − A0t ) are local
martingales. Prove that (At ) and (A0t ) are indistinguishable.
(2) Prove that hM, N it is symmetric in (Mt ) and (Nt ) and is bilinear.
1
(3) Prove that hM, N it = 4 hM + N, M + N it − hM − N, M − N it .
hM T , N T i = hM, N iT = hM, N T i.
(2) The symmetry equality hM, N i = hN, M i comes from that of the product:
M N = N M . Next, we will prove
The process
M1 N + M2 N − hM1 , N i − hM2 , N i
is a local martingale. Thus, the claim follows from the uniqueness of the bracket
process. Similarly for k ∈ R, kM − khM, i is a local martingale and hkM i =
khM i.
(3) is a consequence of the bi-linearity, proved earlier.
(4) Since (Mt N0 , t ≥ 0) is a local martingale, the bracket process hM, N0 i van-
ishes. By the same reason we have hM0 , N i = hM0 , N0 i = 0. Hence, the claim
follows from the bilinearity.
(5) By the definition M T N T − hM T , N T i and (M N )T − hM, N iT are local
martingales. This implies hM T , N T i = hM, N iT , from the uniqueness of the
bracket process.
Furthermore, both M T N T − hM T , N T i and (M − M T )N T are local mar-
tingales, hence their sum M N T − hM T , N T i is a local martingale as well. This
implies hM, N T i = hM T , N T i, from the uniqueness of the bracket process.
Exercise 31 Let (Mt , t ∈ [0, 1]) and (Nt , t ∈ [0, 1]) be bounded continuous mar-
tingales with M0 = N0 = 0. If (Mt ) and (Nt ) are furthermore independent prove
that their quadratic variation process (hM, N it ) vanishes.
Solution: Let us take a partition 0 = t0 < t1 < . . . < tn = 1. Then we have
n
hX i2
E (Mti − Mti−1 )(Nti − Nti−1 )
i=1
n
X h i h i
= E (Mti − Mti−1 )(Mtj − Mtj−1 ) E (Nti − Nti−1 )(Ntj − Ntj−1 )
i,j=1
X n h i2 h n
i2 X h i h i
= E Mti − Mti−1 E Nti − Nti−1 = E Mt2i − Mt2i−1 E Nt2i − Nt2i−1
i=1 i=1
n
X h i h i Xn h i h i
≤ E Mt2i − Mt2i−1 sup E Nt2j − Nt2j−1 ≤ E Mt2i − Mt2i−1 E sup Nt2j − Nt2j−1 .
i=1 j i=1 j
Because Nt2j is uniformly continuous, supj |Nt2j − Nt2j−1 | → 0 almost surely and
h i
Nt2 is bounded, E supj Nt2j − Nt2j−1 → 0. Since the bracket process is the limit
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 23
Pn
of i=1 (Mti − Mti−1 )(Nti − Nti−1 ) (in probability), this implies hM, N it = 0.
Exercise 32 Prove that (1) for almost surely all ω, the Brownian paths t 7→ Bt (ω)
has infinite total variation on any intervals [a, b]. (2) For almost surely all ω, Bt (ω)
cannot have Hölder continuous path of order α > 21 .
Solution: (1) We know that t is the quadratic variation of (Bt ). We can choose a
(n)
sequence of partitions {tj }j=1,...,Mn such that for almost surely all ω,
n
X 2
Bt(n) (ω) − Bt(n) (ω) → (b − a) , (4)
j j−1
j=1
Mn
X
(Bt(n) (ω) − Bt(n) (ω))2 ≤ max |Bt(n) (ω) − Bt(n) (ω)|TV[a,b] (B(ω)) → 0 .
j j−1 j j j−1
j=1
But for almost surely all ω, this limit is b − a, and hence TV[a,b] (B(ω)) = ∞ for
almost surely all ω.
(2) Suppose that for some α > 12 and some number C, both may depend on ω,
Mn Mn Mn
(n) (n) (n) (n)
X X X
(Bt(n) (ω) − Bt(n) (ω))2 ≤ C 2 |tj − tj−1 |2α ≤ C 2 |∆n |2α−1 |tj − tj−1 |
j j−1
j=1 j=1 j=1
2 2α−1
= C (b − a)|∆n | →0,
what contradicts with (4), unless b − a = 0. Hence for almost surely all ω, the
Brownian path is not Hölder continuous of order α > 1/2 in any interval [a, b].
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 24
Exercise 33 Let (Bt ) be a standard Brownian Motion and let fs , gs ∈ L2loc (B).
Compute the indicated bracket processes, the final expression should not involve
stochastic integration.
Rt Rt
1. h 0 fs dBs , 0 gs dBs i.
R· R·
2. h 0 Bs dBs , 0 Bs3 dBs it .
Solution: (1a) By the definition,
Z · Z · Z t Z t
h fs dBs , gs dBs it = fs gs dhBis = fs gs ds .
0 0 0 0
1
− p1 −
(a) Prove that (Mt ) belongs to C 2 .
(b) If (Ht ) is a bounded process, i.e. sup(t,ω) |Ht (ω)| < ∞, prove that (Mt )
1
belongs to C 2 − .
Rt Rt
Proof: First note that the bracket process of s Hr dBr is s (Hr )2 dr.
(1) Since H is Lp bounded, there exists C s.t. sups∈[0,1] E(|Hs |p ) < C. By
Burkholder-Davis-Gundy inequality,
Z t p2 Z t
p 2 1 p
E|Mt − Ms | ≤ cp E (Hr ) dr ≤ cp (t − s) E 2 (Hr )p dr
s t−s s
p
≤ Ccp (t − s) 2 .
so that (Mt ) has a modification which is Hölder continuous for any γ < 21 − p1 for
any p > 0 and concluding the second assertion, again by Kolmogorov’s continuity
theorem.
Exercise
RT 35 Let T > 0. Let f beR t a left continuous and adapted process such that
E 0 (fs )2 ds < ∞. Prove that ( 0 fs dBs , 0 ≤ t ≤ T ) is a martingale.
Rt
Solution: We know that ( 0 fs dBs , 0 ≤ t ≤ T ) is a local martingale. Furthermore,
Z t Z t Z T
2 2
E( fs dBs , ) ≤ c2 E (fs ) ds ≤ c2 E (fs )2 ds < ∞.
0 0 0
for any N ∈ H 2.
Exercise
Rt 37 Let M ∈ H 2 and K ∈ E. Prove that the elementary integral It :=
0 Ks dMs satisfies
Z t
hI, N it = Ks dhM, N is , ∀t ≥ 0 ,
0
for any N ∈ H 2.
Solution: See lecture notes.
Rt
Exercise 39 Give an expression for 0 s dBs that does not involve stochastic inte-
gration.
Solution: Applying the classical integration by parts formula, we obtain
Z t Z t
s dBs = tBt − Bs ds.
0 0
Rt Rs
Exercise 40 Write 0 (2Bs + 1)d 0 Br d(Br + r) as a function of (Bt ), not in-
volving stochastic integrals.
Solution: Firstly,
Z t Z s Z t
(2Bs + 1)d Br d(Br + r) = (2Bs + 1)Bs d(Bs + s)
0 0 0
Z t Z t Z t
2
= 2Bs dBs + Bs dBs + (2Bs2 + Bs )ds.
0 0 0
7 Itô’s Formula
Exercise 42 If (Nt ) is a continuous local martingale with N0 = 0, show that
Nt − 21 hN,N it Nt − 21 hN,N it
(e ) is a local martingale, and E e ≤ 1.
1
Solution: Let us denote Xt = eNt − 2 hN,N it . The process (Nt − 12 hN, N it ) is a
semi-martingale, and we can apply the Itô formula to the function ex :
Z t
Xt = 1 + Xs dNs .
0
Since the stochastic integral is a local martingale, we conclude that (Xt ) is a local
martingale.
Moreover, let Tn be the sequence of increasing stopping times such that XtTn
is a uniformly integrable martingale. Then, applying the Fatou lemma, we get
2. Compute hexp (Mt − 21 hM it ), exp (Nt − 21 hN it )i, where (Mt ) and (Nt )
are continuous local martingales.
Solution: (1) By Itô formula we have
Z t
1 t 00
Z
0
f (Bt ) = f (0) + f (Bs )dBs + f (Bs )ds ,
0 2 0
Z t
1 t 00
Z
0
g(Bt ) = g(0) + g (Bs )dBs + g (Bs )ds .
0 2 0
Hence, we obtain
Z · Z ·
hf (B), g(B)it = h f 0 (Bs )dBs , g 0 (Bs )dBs it
0 0
Z t
= f 0 (Bs )g 0 (Bs )ds ,
0
Hence, we conclude
Z · Z · Z t
hX, Y it = h Xs dMs , Ys dNs it = Xs Ys dhM, N is .
0 0 0
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 31
Solution: (1) Let us take a function f (x) = |x|, for x ∈ Rn . Its derivatives are
given by
P R t Xsi
To finish the proof, we have to show that B̃t := ni=1 0 |X s|
dBsi is a Brownian
motion. Firstly it is a continuous martingale, starting at 0. It has the quadratic
variation
n Z t
X (Bsi )2
hB̃, B̃it = 2
ds = t.
0 |Bs | i=1
Solution: Let us denote g(x) = h−1 h(x0 ) + x . Then its derivatives are given
by
g 0 (x) = f (g(x)) , g 00 (x) = f 0 (g(x))f (g(x)) .
The claim now follows by applying the Itô formula to g(Bt ).
Exercise 48 Let (Bt ) be a Brownian motion and τ its hitting time of the set [2, ∞).
Is the stochastic process ( Bt1−2 , t < τ ) a solution to a stochastic differential equa-
tion? (The time τ is the natural life time of the process Bt1−2 .)
Solution: Note that τ = inf{t : Bt ∈ [2, ∞)}. Let τn be a sequence of stopping
time increasing to τ . Let f (x) = (x − 2)−1 . Where f is differentiable, f 0 (x) =
−(x − 2)−2 = −f 2 and f 00 (x) = 2(x − 2)−3 = 2f 3 . Applying the Itô formula to
f and to the stopped process Xt := (Bt − 2)−1 we obtain
Z t Z t
τn τn 1 τn 2
Xt ≡ f (Bt ) = − − (Xs ) dBs + (Xsτn )3 ds ,
2 0 0
Consequently,
Z · Z · Z t
hX 2 , Xit = h dXs , 2 Xs Xs it = 2 Xs dhX, Xis .
0 0 0
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 33
Rt
and f : Rd → R, prove that f (xt ) − f (x0 ) − 0 Lf (xs )ds is a local mar-
tingale and give the semi-martingale decomposition for f (xt ).
∂f ∂2f
(x) = 2xi , (x) = 2δi,j .
∂xi ∂xi ∂xj
d Z t d Z
X ∂f i 1 X t ∂2f
f (xt ) =f (x0 ) + (xs )dxs + (xs )dhxi , xj is
0 ∂xi 2 0 ∂xi ∂xj
i=1 i,j=1
d Z t m d Z t
X ∂f X X ∂f
=f (x0 ) + (xs ) σki (xs ) dBsk + (xs )bi (xs )ds
0 ∂xi 0 ∂xi
i=1 k=1 i=1
d Z m
1 X t ∂2f
σki (xs )σkj (xs )ds
X
+ (xs )
2 0 ∂xi ∂xj
i,j=1 k=1
m Z tX
d
X ∂f
=f (x0 ) + (xs )σki (xs ) dBsk
0 ∂xi
k=1 i=1
d m d
Z t !
2
1 σki (xs )σkj (xs )
∂ f ∂f
X X X
+ (xs ) + (xs )bi (xs ) ds
0 2 ∂xi ∂xj ∂xi
i,j=1 k=1 i=1
Rt Pm R t Pd
∂f
Thus f (xt ) − f (x0 ) − 0 Lf (xs )ds = k=1 0 i=1 ∂x i
(xs )σki (xs ) dBsk is a
local martingale and the semi-martingale decomposition is as given in the identity
earlier.
Exercise 51 Let T be a bounded stopping time and (Bt ) a one dimensional Brow-
nian motion. Prove that (BT +s − BT , s ≥ 0) is a Brownian motion.
Solution: Let us denote Ws = BT +s − BT . It is obvious, that W0 = 0 and Wt has
continuous sample paths. Moreover, by the OST we have, for any s < t,
Let Gt = FT +t . Then (Wt ) is a (Gt )-martingale. Next, we take s < t and derive
where we have used the fact that (BT +t )2 − (T + t) is a martingale. This im-
plies that Wt2 − t is a (Gt )-martingale, and hence hW, W it = t. Using the Lévy
characterization theorem, we conclude that Wt is a (Gt )-Brownian motion.
Exercise 52 Let {Bt , Wt1 , Wt2 } be independent one dimensional Brownian mo-
tions.
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 35
1. Let (xs ) be an adapted continuous stochastic process. Prove that (Wt ) de-
fined below is a Brownian motion,
Z t Z t
Wt = cos(xs )dWs1 + sin(xs )dWs2 .
0 0
3. Prove that the process (Xt , Yt ) is a Brownian motion if they satisfy the fol-
lowing relations,
Z t Z t
1
Xt = 1Xs >Ys dWs + 1Xs ≤Ys dWs2
0 0
Z t Z t
Yt = 1Xs ≤Ys dWs1 + 1Xs >Ys dWs2 .
0 0
Solution: (1) The process (Wt ) is a continuous martingale with the quadratic vari-
ation Z t
(cos(xs ))2 + (sin(xs ))2 ds = t .
< W >t =
0
Hence, by Lévy characterization theorem, we conclude that (Wt ) is a Brownian
motion.
(2) Can be shown in the same way.
(3) The processes (Xt ) and (Yt ) are martingales, and their quadratic variations are
Z t
12Xs >Ys + 12Xs ≤Ys ds = t ,
< X >t =
0
Z t
12Xs ≤Ys + 12Xs >Ys ds = t .
< Y >t =
0
where we have used the fact that 1Xs >Ys 1Xs ≤Ys = 0. The claim now follows from
the Lévy characterization theorem.
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 36
(3) Define
Z A−1 (r)
dBs
Wr = , 0 ≤ r < ∞.
0 1−s
Let Gr = FA−1 (r) . Prove that (Wr ) is an (Gr )-martingale.
(2) We have Z · Z t
dBs d < B >s t
= 2
= .
0 1−s t 0 (1 − s) 1−t
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 37
(3) It is obvious that (Wr ) is integrable and adapted to (Gr ). Moreover, for any
t > s, we have
"Z # Z
t/(t+1) s/(s+1)
dBr dBr
E[Wt |Gs ] = E Fs/(s+1) = = Ws ,
0 1−r 0 1−r
because s/(s + 1) < t/(t + 1). Thus the Itô integral (Wt ) is a martingale with
respect to (Ft ). This shows that (Wr ) is a martingale with respect to (Gr ).
(4) By (2), we can calculate
A−1 (t)
hW it = =t.
1 − A−1 (t)
1
(5) Since the function 1−s has finite total variation on [0, c], for every c < 1,
Rt 1
the integral 0 Bs d 1−s , for t ≤ c, can be defined in the Stieltjes sense. Using
integration by parts, we see that the process Wt is defined in the Stieltjes sense as
well. We can use the classical analysis to derive the following equalities:
Z t Z t t Z t
xs
ds = WA(s) ds = sWA(s) − s dWA(s)
0 1−s 0 0 0
Z t Z t Z t Z t Z t
dBs s dBs dBs
=t − dBs = t − + dBs
0 1−s 0 1−s 0 1−s 0 1−s 0
= −xt + Bt ,
which is the claimed equality, for any t ≤ c. Passing c → 1, we conclude that the
equality holds for any t ∈ [0, 1].
(6) Since, xs is defined as an integral of a non-random process with respect to the
Brownian motion, it has the normal distribution, whose parameters are easily seen
to be 0 and s(1 − s) (the latter follows from (1)). Thus, we apply Fubini lemma to
derive
Z 1 Z 1 Z 1p
|xs | E|xs | s(1 − s)
E ds = ds = C ds < ∞ ,
0 1 − s 0 1 − s 0 1−s
R 1 |xs |
for some constant C > 0. This implies that 0 1−s ds < ∞ almost surely. Defining
R t xs
the process Yt = 0 1−s ds and taking any partition 0 = t0 < t1 < . . . < tn = 1,
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 38
we derive
n−1 n−1
X Z ti+1 n−1 Z t
X xs X i+1 |xs |
|Yti+1 − Yti | = ds ≤ ds
1−s 1−s
i=0 i=0 ti i=0 ti
Z 1
|xs |
= ds < ∞ ,
0 1 −s
from what the claim follows.
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 39
and Z t∧τn Z τ
lim lim E g(xs )ds = E g(xs )ds.
n→∞ t→∞ 0 0
This completes the proof.
[Hint: Show that if y0 > 0 then yt is positive and hence the SDE can be considered
to be defined on the upper half plane. Compute its infinitesimal generator L.]
2
Solution: Just observe that yt = y0 eBt −(1/2)t so yt > 0 if y0 > 0. Compute the
generator by Itô’s formula.
Exercise 56 Discuss the uniqueness and existence problem for the SDE
dxt = sin(xt )dBt1 + cos(xt )dBt2 .
Solution: The functions sin(x) and cos(x) are C 1 with bounded derivative, and
hence Lipschitz continuous. For each initial value there is a unique global strong
solution.
Exercise 57 Let (xt ) and (yt ) be solutions to the following respective SDEs (in
Stratnovitch form),
Z t Z t
xt = x0 − ys ◦ dBs , yt = y0 + xs ◦ dBs .
0 0
Show that x2t + yt2 is independent of t.
Solution: Let us rewrite the SDEs in the Itô form:
Z t Z t
1 1
xt = x0 − ys dBs − < y, B >t , yt = y0 + xs dBs + < x, B >t .
0 2 0 2
We can calculate the bracket processes in these expressions:
Z · Z t Z t
< y, B >t =< xs dBs , B >t = xs ds , < x, B >t = ys ds .
0 0 0
Moreover, the quadratic variations of (xt ) and (yt ) are
Z · Z t Z t
2
< y >t =< xs dBs >t = xs ds , < x >t = ys2 ds .
0 0 0
Applying now the Itô formula to the function f (x, y) = + x2 y2,
we obtain
Z t Z t
f (xt , yt ) =f (x0 , y0 ) + 2 xs dxs + 2 ys dys + < x >t + < y >t
0 0
Z t Z t Z t
=f (x0 , y0 ) − 2 xs ys dBs − xs d < y, B >s +2 xs ys dBs
0 0 0
Z t
+ ys d < x, B >s + < x >t + < y >t
0
Z t Z t
=f (x0 , y0 ) − x2s ds − ys2 ds+ < x >t + < y >t = f (x0 , y0 ) ,
0 0
Stochastic Analysis (2014). Lecturer: Xue-Mei Li, Support Class: Andris Gerasimovics 41
which is independent of t.
R∞
Exercise 58 1. Let (ht ) be a deterministic real valued process with 0 (hs )2 ds <
∞. Let (Bt ) be an Ft -Brownian motion on (Ω, F, Ft , P ). Prove that
Z t
1 t
Z
2
exp hs dBs − (hs ) ds
0 2 0
is a martingale.
2. Let (ht ) be a bounded continuous and adapted real valued stochastic process.
Prove that Z t
1 t
Z
2
exp hs dBs − (hs ) ds
0 2 0
is a martingale.
Thus Z t Z t
2
EhM, M it ≤ 2 (hs ) ds + c (hs )2 EhM, M is ds.
0 0
By a version of Grownall’s inequality,
Z t Z t
2 2
EhM, M it ≤ 2 (hs ) ds exp c (hr ) dr .
0 0
and B̃t = Bt − < B, N >t . First observe that the exponential martingale of
(Nt ) is a martingale, c.f. part (1) and Q is a probability measure, equivalent to
P . By Girsanov theorem, (B̃t ) is an Ft local -martingale with respect to Q. Since
< B̃ >t =< B >t = t, it follows from the Lévy characterisation theorem that (B̃t )
is a Brownian motion with respect to Q.