Brownian motion and Itô calculus

8
Brownian motion is a continuous analogue of simple random walks (as described in the previous
part), which is very important in many practical applications. This importance has its origin in
the universal properties of Brownian motion, which appear as the continuous scaling limit of many
simple processes. Moreover, it is also intimately related to martingales and bounded-variation
processes in continuous time. Brownian motion is a very rich structure that inherits properties from
various fields of mathematics [à compléter].
This chapter presents in a first section the main properties of Brownian as well as various
constructions. The second section presents Itô calculus for Brownian motion: this construction is
only a particular case of stochastic calculus for semi-martingales; we choose however to present it
here for two reasons: many applications do not require the general framework and moreover this
general framework is abstract enough, so that a short introduction in a simple case may help for
the next chapter.
Before going any further in this chapter, the reader is encouraged to read again the reminder of
the properties of normal random variables presented in the first part.

8.1 Definition, construction and properties

8.1.1 Definition and structure
Definition 8.1.1 (Brownian motion). Let (Ω, F, P) be a probability space. A Brownian motion
(Bt )t∈R+ is a process with values in R such that:

1. B0 = 0,

2. the maps t 7→ Bt (ω) are almost surely continuous,

3. the increments Bt+s − Bt are independent from σ(Bu , u ≤ t) and are centered normal r.v.
with variance s, for all t, s > 0.

Its existence is postponed to the next subsection and we assume now that it exists. The Brownian
motion exhibits many interesting features mentioned in the previous chapter, that relates it to other
probabilistic tools.

Property 8.1.1. Let (Bt )t∈R+ be a Brownian motion. We considered its associated filtration.
2
• the processes (Bt )t∈R+ , (Bt2 − t)t∈R+ and (euBt −tu /2
)t∈R+ are martingales.

117
118 Chapter 8. Brownian motion and Itô calculus

• the process (Bt )t∈R+ is a Markov process with initial law δ0 and transition semi-group (Qt )t∈R+
given by:
(x − y)2
 
1
Qt (x, dy) = √ exp − dy (8.1)
2πt 2t
Proof. Integrability of Bt = Bt − B0 is provided by the integrability of normal random variables. A
short computation gives:

E [Bt |Fs ] = E [Bt − Bs |Fs ] + E [Bs |Fs ] = E [Bt − Bs ] + Bs = Bs (8.2)

for all s > t, from the independence of the increments and their centered laws. The same
decomposition of Bt as (Bt − Bs ) + Bs also gives the two other martingale properties. The Markov
transition semi-group follows from the characterization:

E [f (Bt+u )g(Bt )] = E [f ((Bt − 0) + (Bt+u − Bt ))g(Bt − 0)]

 2
y2

1 x
ZZ
= f (x + y)g(y) √ exp − −
2π tu 2t 2u
2
y2
 
1 (z − y)
ZZ
= f (z)g(y) √ exp − −
2π tu 2t 2u
Z  
2
−(z−Bt ) /(2u) 1
=E f (z)e √ dz g(Bt )
2πu
for all t, u > 0 through Fubini’s theorem with arbitrary bounded measurable functions f and g.
We have seen that the strong Markov property is not trivial in general. However, it holds for
Brownian motion:
Theorem 8.1.1. Brownian motion satisfies the weak and strong Markov properties. Let T be
a stopping time and (Bt )t∈R+ be a Brownian motion; conditionally on {T < ∞}, the process
(BT +t − BT )t∈R+ is a Brownian motion independent of FT .
Proof. Either we deduce it from general results about Markov processes with càdlàg trajectories
and values in a Polish space, or we may check it by hand. We know from the general theory of
Markov process with càdlàg trajectories and we need only to prove:

E [X1T <∞ Fn (Bt1 +T − BT , . . . , Btn +T − BT )] = E [X1T <∞ ] E [Fn (Bt1 , . . . , Btn )] (8.3)

for arbitrary FT -measurable integrable r.v. X and arbitrary bounded measurable functions Fn :
Rn → R.
The proof is very similar to the discrete case, except that the event {T < ∞} cannot be
written as a countable union of events {T = tk }. In order to avoid this difficulty, we introduce
approximations of T defined by:
⌊2n T ⌋ + 1
Tn = (8.4)
2n
such that Tn is a decreasing sequence of stopping times that converges to T and takes only rational
values.
We first prove the strong Markov property for a rational-valued stopping time S.
X
E [X1S<∞ Fn (Bt1 +S − BS , . . . , Btn +S − BS )] = E [X1S=s Fn (Bt1 +s − Bs , . . . , Btn +s − Bs )]
s∈Q+
X
= E [X1S=s ] E [Fn (Bt1 , . . . , Btn )]
s∈Q+

= E [X1S<∞ ] E [Fn (Bt1 , . . . , Btn )]

as in discrete time. The property is proved for all the approximations Tn and thus, by convergence
of Tn and right-continuity of Brownian motion, we obtain the strong Markov property for any
stopping time T .
8.1 Definition, construction and properties 119

8.1.2 Marginal laws, scaling properties, some stopping times.

Using the independent increments properties, many marginal laws of Brownian motions can be
computed. In particular, we have:
Proposition 8.1.1. Let (Bt )t∈R+ be a Brownian motion. The bounded-dimensional marginal laws
are given by:
P (Bt1 ∈ A1 , Bt2 ∈ A2 , . . . Btn ∈ An )
n n
! !
1 (xk − xk−1 )2 (8.5)
Z Y X
= p exp − dx1 dx2 . . . dxn
A1 ×A2 ×...×An 2π(tk − tk−1 ) 2(tk − tk−1 )
k=1 k=1

for any sequence 0 = t0 < t1 < t2 < . . . < tn with the convention x0 = 0.
Important remark: using this expression of marginal laws and integrating it in order to
compute expectation values is at the time a sure way of evaluation but often the worst one! In
many cases, using the independent increments property together with expectation values is much
more efficient.
Proposition 8.1.2. Let (Bt )t∈R+ be a Brownian motion. As a Gaussian process, it is fully
characterized by its mean and its covariance. It holds:
E [Bt ] = 0 (8.6)
E [Bt Bs ] = min(s, t) (8.7)
Moreover several transformations maps a Brownian motion to another Brownian motion.
Proposition 8.1.3. Let (Bt )t∈R+ be a Brownian motion.
1. time translation invariance: for all u > 0, the centered shifted process (Bt+u − Bu )t∈R+ is a
Brownian motion.
2. invariance under scaling: for all α > 0, the renormalized process (αBα−2 t )t∈R+ is a Brownian
motion.
3. invariance under reflexion: the process (−Bt )t∈R+ is a Brownian motion.
4. invariance under time inversion: the process (tB1/t )t∈R+ (restricted on the set of probability 1
on which tB1/t → 0 as t → 0) is a Brownian motion (in its own filtration).
Proof. The scheme of proof is the same for the four invariances: we check the three point of the
definition.
As for simple random walks described previously, various hitting times of the Brownian motion
are easy to study. The trajectories of the Brownian motion are continuous and thus, for any closed
set B of R, the hitting time is a stopping time.
Proposition 8.1.4. Let (Bt )t∈R+ be a Brownian motion. Let Ta be the stopping time defined by
Ta = inf {t ∈ R+ ; Bt = a} (8.8)
and the process (St )t∈R+ defined by
St = sup Bs . (8.9)
0≤s≤t

For any a ≥ 0, it holds

P (Ta ≤ t) = P (St ≥ a) = 2P (Bt ≥ a) . (8.10)
1
The r.v. St has the same law as |Bt | and the r.v. Ta has the same law as the r.v. a /B12 and 2

admits the density given by  2

a a
√ exp − . (8.11)
2πt 3 2s

1 Be careful, it may not be true as equality in law for the whole process!
120 Chapter 8. Brownian motion and Itô calculus

Proof. The equality {Ta ≤ t} = {St ≥ a} is trivial since the Brownian motion has continuous
trajectories. This continuity also implies {Bt ≥ a} ⊂ {Ta ≤ t} for a ≥ 0. Then, we prove the last
equality by using the strong Markov and independent increments properties:

P (Bt ≥ a) = P (Bt ≥ a, Ta ≤ t) = P ((Bt − BTa ) + BTa ≥ a, Ta ≤ t)

 
= P (Bt − BTa ≥ 0, Ta ≤ t) = E 1Ta ≤t 1Bt −BTa ≥0
  1
= E 1Ta ≤t 1Bt−Ta ≥0 = P (Ta ≥ t)
2
since, for any s > 0, P (Bs ≥ 0) = 1/2. The densities are deduced directly from the equality.

8.1.3 Regularity and trajectories

This section is devoted to the fine study of the trajectories of a Brownian motion. We know from
the definition that any trajectory t 7→ Bt (ω) is continuous and we look for a stronger regularity. To
this purpose, we focus on local properties of trajectories and we deduce from the time invariance
that a study at t = 0 is enough. We start with the following property.
Theorem 8.1.2. Let (Bt )t∈R+ be a Brownian motion. It holds:

Bt
lim sup √ = +∞, a.s.
t→0 + t
Bt
lim inf √ = −∞, a.s.
t→0+ t
√
Proof. The random variable Y = lim supt→0+ Bt / t is a typical example of r.v. which is F0+ -
measurable where Ft is the filtration associated to the Brownian motion. A 0 − 1 law exists for
such a process with independent increments.
Lemma 8.1.1. Let (Bt )t∈R+ be a Brownian motion and (Ft )t∈R+ its filtration. Then, for any
B ∈ F0+ , P (B) = 0 or 1.
Proof. We first prove that (Bt ) is also a Brownian motion in the filtration (Ft+ )t∈R+ . We consider a
sequence t < t′ < t1 < t2 < . . . < tn : the independent increment property implies that (Bt′ +s −Bt′ )s
are independent of Ft′ for any t′ > t and t′ < t1 and thus are independent of Ft+ . Moreover, the
continuity of the trajectories implies that Bt = limt′ ↓ Bt′ and we obtain the result.
We deduce from third property that Bt = Bt − B0 is independent from F0+ for all t > 0 and
thus, for all t > 0, Ft is independent from F0+ , and F0+ = ∩t>0 Ft is independent from itself, hence
the lemma.
The r.v. Y is
√ thus
almost surely constant (eventually infinite). If Y < a for some a finite, then
we have P Bt / t > a → 0 as t → ∞ (if not, Borel Cantelli lemma implies that Y is above a an
infinite number of times). However, this probability is easy to evaluate:
  Z ∞ Z ∞
Bt 1 −x2 /(2t) 1 2
P √ >a = √ √ e dx = √ e−x /2 dx > 0 (8.12)
t a t 2πt a 2π
We thus conclude that Y = +∞. The reflexion invariance gives immediately the second equality.
This theorem gives a good hint of the irregularity of√the Brownian: √ in a time as short as we wish,
any trajectory goes a infinite number of times above t and below − t. The choice of the square
root is also the optimal choice since it is the “last” scaling in (8.12) that gives a non-zero limit.
Corollary 8.1.1. We deduce the following properties:
1. A Brownian trajectory t 7→ Bt (ω) is almost surely nowhere derivable.
√ √
2. lim supt→∞ Bt / t = +∞ and lim inf t→∞ Bt / t = −∞
3. every point is visited a infinite number of times.
8.2 Itô calculus and stochastic integration 121

Proof. The derivability at 0 is in contradiction with the previous theorem. Time translation
invariance extends this property to any t ∈ R+ . The second point is a consequence of the invariance
under time inversion t 7→ 1/t after rescaling. The third point is deduced from the second through
the continuity hypothesis of the trajectories.

We have seen in the previous chapter a new notion of regularity which is the bounded-variation
property, between continuity and derivability.

Property 8.1.2. Let (Bt )t∈R+ be a Brownian motion. It has a non-zero quadratic variation (equal
to t over the interval [0, t]) and thus does not have a bounded total variation.

Proof. We use the characterization with a limit over nested subdivisions. We fix a time t and a
subdivision 0 = t0 < t1 < . . . < tn = t and we perform the following L2 estimation:
 !2 
Xn n
X X n
X
E (Btk − Btk−1 )2 − t  = 3(tk − tk−1 )2 + (tk − tk−1 )(tl − tl−1 ) − 2 (tk − tk−1 )t + t2
k=1 k=1 k6=l k=1
n
X
= 2(tk − tk−1 )2
k=1

(N ) (N )
If we consider sequences of subdivisions (tk )1≤k≤nN whose mesh size goes to 0 (i.e. supk∈{1,...,nN |tk −
(N )
tk−1 | → 0 as N → ∞, then the previous expectation also goes to zero. We thus have:
nN
X L2 ,N →∞
(Bt(N ) − Bt(N ) )2 −−−−−−→ t (8.13)
k k−1
k=1

8.1.4 Construction and existence

8.1.5 Numerical simulation

8.2 Itô calculus and stochastic integration

We have already seen that Brownian motion trajectories are not bounded-variation functions;
however, it has a bounded quadratic variation and it may help to build quantities such as
Z t
Hs dBs (8.14)
0

where Ht is a process with some adequate hypothesis, even if it does not enter directly the
construction made in the previous chapter for bounded-variation functions. Understanding how
this construction
Rt works leads us to Itô calculus, which is one of the main tool to handle quantities
such as 0 h(Bs )dBs .
This section has to be considered more as computational tools rather than general variation
theory: general stochastic calculus for more general martingales is developed in the next chapter.

8.2.1 Stochastic integral

As for Riemann integral, we first build stochastic integrals for staircase functions.

Definition 8.2.1. Let (Bt )t∈R+ be a Brownian motion and let (Ht )t∈R+ be an adapted process with
staircase trajectories, written as:
n
X
Hs = Htk 1[tk ,tk+1 [ (s) (8.15)
k=1
122 Chapter 8. Brownian motion and Itô calculus

where (tk ) is a subdivision of R+ . We define the stochastic integral as:

Z t n
X
Hs dBs = Htk (Btk+1 ∧t − Btk ∧t ). (8.16)
0 k=1

This definition coincides with the usual formula for bounded-variation functions. This definition
is linear in (Hs ) and we can estimate easily the first moments.
Property 8.2.1. Let (Bt )t∈R+ be a Brownian motion and let (Ht )t∈R+ be an adapted process with
staircase trajectories and such that, for all s ∈ R+ , the r.v. Hs is square-integrable (i.e. is in L2 ).
Then, it holds:
Z t 
E Hs dBs = 0 (8.17)
"Z 0 2 #
t Z  t
E Hs dBs =E Hs2 ds (8.18)
0 0

Proof. We use the martingale property of Brownian motion:

Z t  X n n
  X    
E Hs dBs = E Htk (Btk+1 − Btk ) = E E Htk (Btk+1 ∧t − Btk ∧t )Ftk
0 k=1 k=1
Xn
  
= E Htk E (Btk+1 ∧t − Btk ∧t ) = 0
k=1

The second computation uses the same type of decomposition:

"Z
t 2 # X n
n X
 
E Hs dBs = E Htk Htl (Btk+1 ∧t − Btk ∧t )(Btl+1 ∧t − Btl ∧t )
0 k=1 l=1

Using the independent increments property and introducing the adequate filtrations give:
"Z
t 2 # Xn X n Z t
E Hs2 ds
 
E Hs dBs = E [Htk Htl δkl (tk+1 ∧ t − tk ∧ t)] =
0 k=1 l=1 0

and we prove the property.

In order to generalize this stochastic integral, we may use the density of processes with staircase
trajectories to more general processes. For example, if (Ht )t∈R+ is an adapted bounded process with
continuous trajectories, then the sequence of approximations
n
(n)
X
Ht = HlT /n 1[kT /n,(k+1)T /n[ (t) (8.19)
l=0

converges to (Ht ) on [0, T ] for the following convergence:

Z t h i
n→∞
∀t ∈ [0, T ], E (Hs(n) − Hs )2 ds −−−−→ 0 (8.20)
0

This can be proved in a straightforward way using dominated convergence theorems. The hypothesis
of boundedness can be weakened to the hypothesis
"Z #
T
E Hs2 ds < ∞ (8.21)
0

however the corresponding proof is a bit longer and is postponed to the next chapter for a general
framework: we use it here without further proof.
8.2 Itô calculus and stochastic integration 123

Theorem 8.2.1. Let (Bt )t∈R be a Brownian motion and (Ht )t∈R+ be a process adapted to the
filtration associated to (Bt ), with continuous trajectories and such that
"Z #
T
E Hs2 ds <∞ (8.22)
0

Rt
for some T . Then the stochastic integral 0 Hs dBs can be defined in a unique way for t ≤ T as the
R t (n) (n)
limit of 0 Hs dBs for any approximation (Ht ) of (Ht ) in the sense of (8.20).
The following equality are then valid:
Z t 
E Hs dBs = 0, (8.23)
0
"Z
t 2 # Z t 
E Hs dBs =E Hs2 ds . (8.24)
0 0

R t (n)
Proof. We admit here that such approximations exist. Then, we show that the sequence 0 Hs dBs
is a Cauchy sequence in the L2 space and thus, by completeness of L2 spaces, it converges to a
limit (and this limit is a.s. unique). The next equalities are obtained from the similar ones proved
for staircase processes.

8.2.2 Variations and Itô calculus

8.2.2.1 Necessity
Rt
The previous theorems show the existence of r.v. defined as stochastic integrals 0 Hs dBs . This
integral is linear in (Hs ); however many standard properties of integrals such as integration by
parts or differentiation are not standard anymore in their standard versions. For example, the
following intuitive equalities break down:
Z t
Bt2 6= 2Bt dBt (8.25)
0
Z t
Bt4 6= 4Bt3 dBt (8.26)
0
Z t
euBt 6 1+u
= euBt dBt (8.27)
0

2
If there were true, their expectations would give t = 0, 3t2 = 0 and eu t/2 = 1 for any t > 0.
However, we present here without any proof (see below) the correct versions:
t
1 t
Z Z
Bt2 = 2Bt dBt + 2dt (8.28)
0 2 0
Z t Z t
1
Bt4 = 4Bt3 dBt + 12Bt2 dt (8.29)
0 2 0
Z t
1 t 2 uBt
Z
uBt uBt
e =1+u e dBt + u e dt (8.30)
0 2 0

For example, even if this not a proof, the reader may check easily from the properties of stochastic
integrals that the first moments coincide. Some quick further investigations suggest that the
additional terms involve the second derivative of the function f giving the l.h.s. f (Bt ).
Itô calculus gives the rules that govern these additional terms and explains their apparition (and
the irrelevance of higher derivatives): this is strongly related the regularity properties of Brownian,
and in particular its quadratic variation.
124 Chapter 8. Brownian motion and Itô calculus

8.2.2.2 Itô’s formula

In order to have some L2 control of the integrals of the type f (Bt )dBt , we require the following
R

quadratic variation theorem, which is a generalization of (8.13).

Theorem 8.2.2. Let (Bt )t∈R+ be a Brownian motion and let f : R → R be a bounded measurable
(n) (n) (n) (n)
function. Let (tk ) be a sequence of subdivisions (tk )k=0,...,Nn (i.e. such that t0 = 0 < t1 <
(n) (n) (n) (n)
. . . < tNn −1 < tNn = t) whose mesh size supk |tk+1 − tk | goes to zero as n → ∞. It holds,
NX
n −1
 L2
f (Bt(n) ) (Btk+1 − Btk )2 − (tnk+1 − tnk ) −−−−→ 0

(8.31)
k n→∞
k=0

This theorem is useful since it says that every computation of the expectation value of the
square (order 2) of a stochastic integral can be reduced (in L2 to a deterministic term of order 1 in
time. This reduction from an apparent second order term to a first order one explains, as we will
see, the presence of a second derivative in Itô’s formulae.
Proof. We note In the l.h.s. of (8.31), expand the square, and use the independent increments
property:

  NX n −1 h i h 2 i
E In2 = E f (Btnk )2 E (Btk+1 − Btk )2 − (tnk+1 − tnk )
k=0
X h i 
E f (Btnk )f (Btnl ) (Btk+1 − Btk )2 − (tnk+1 − tnk ) E (Btl+1 − Btl )2 − (tnl+1 − tnl )
 
+2
0≤k<lNn

Standard computations of normal r.v. give directly:

E (Btl+1 − Btl )2 − (tnl+1 − tnl ) = 0

 
h 2 i
E (Btk+1 − Btk )2 − (tnk+1 − tnk ) = 2(tnk+1 − tk )2

and thus
NX
n −1 h i  
 2 2 (n) (n) 2 (n) (n)
E In =2 E f (Bt(n) ) (tk+1 − tk ) ≤ 2||f ||∞ sup |tk+1 − tk | t
k k
k=0

and thus goes to 0: the L2 convergence is proved.

We can now state Itô’s formula, which is one of the most important result of this section.
Theorem 8.2.3 (Itô’s formula I). Let (Bt )t∈R be a Brownian motion. Let f : R → R a function
that is twice derivable and whose second derivative is continuous (i.e. f is C 2 ). Then,
Z t
1 t ′′
Z
′
f (Bt ) = f (B0 ) + f (Bs )dBs + f (Bs )ds (8.32)
0 2 0
Proof. We first assume that f is bounded and measurable. We expand the increment f (Bt ) − f (B0 )
as a sum of infinitesimal increments and use, for each increment, Taylor’s formula, which gives the
(n) (n) (n)
existence of quantities xk ∈ [tk , tk+1 ] such that:
NX
n −1  
1 (n)
f (Bt ) − f (B0 ) = f ′ (Bt(n) )(Bt(n) − Bt(n) ) + f ′′ (xk )(Bt(n) − Bt(n) )2
k k+1 k 2 k+1 k
k=0
NX
n −1 h i
(n) (n) (n)
= f ′ (Bt(n) )(Bt(n) − Bt(n) ) + f ′′ (xk )(tk+1 − tk )2
k k+1 k
k=0
Nn −1
1 X (n) 
f ′′ (xk ) (Btk+1 − Btk )2 − (tnk+1 − tnk )

+
2
k=0
8.2 Itô calculus and stochastic integration 125

The r.h.s. does not depend on the subdivision that is used and thus, along a sequence of subdivisions
whose mesh size goes to zero, the limit exists and the previous theorem cancels the second term:
Itô’s formula is proved.
We then go to general C 2 functions with density theorems that are made precise in the next
chapter.
The control of the variations of functions f (Bt ) of the Brownian motion at first order requires
the knowledge of the first two derivatives of f : this requirement, as the proof shows it, uses the
finite quadratic variation of Brownian motion. Another way of understanding this formula is the
interpretation of the stochastic integration formula: we made the choice of Htk (Btk+1 − Btk ), i.e.
we evaluate Htk at the beginning of the infinitesimal integration interval. However, during this
interval, especially if Htk is a Brownian motion, it may have relevant variations and furthermore
these variations are not independent from (Btk+1 − Btk ); controlling these variations is exactly what
we did and it requires the second derivative of f when Hs = f (Bs ).
From the beginning, we could have taken other conventions in the approximation scheme used
for the stochastic integral: it leads for example to Stratonovitch’ calculus, which is an alternative
to Itô’s calculus,
R that lead to the same results but with different computations. In particular, the
property E Hs dBs = 0 breaks down for other conventions.
The same proof gives the generalization to space-time C 2 functions:
Theorem 8.2.4 (Itô’s formula II). Let (Bt )t∈R be a Brownian motion. Let f : R+ × R → R,
(t, x) 7→ f (t, x) a function that is twice derivable and whose second derivatives are continuous (i.e.
f is C 2 ). Then,
Z t Z t
∂f ∂f 1 t ∂2f
Z
f (Bt ) = f (B0 ) + (s, Bs )ds + (s, Bs )dBs + (s, Bs )ds (8.33)
0 ∂t 0 ∂x 2 0 ∂x2

8.2.2.3 Applications
We present now several consequences of Itô’s formula on the evaluation of various moments related
to Brownian motion. We first consider f (x) = xn for integers n ≥ 2 and apply Itô’s formula:
Z t
n(n − 1) t n−2
Z
Btn = n Btn−1 dBt + Bt dt (8.34)
0 2 0

Expectation value maks the first term of the r.h.s. disappear. Writing an (t) = E [Btn ] gives:

n(n − 1) t
Z
an (t) = an−2 (t)dt (8.35)
2 0

Solving this recursion is a second and complicated way of computing the moments of centered
reduced normal random variables:
(
−x2 /2
ne 0 if n is odd
Z
x √ dx = n!
(8.36)
R 2π 2n/2
if n is even

We consider now f (x) = eux . Itô’s formula gives

Z t
u2 t uBt
Z
uBt uBs
e −1=u e dBs + e dt (8.37)
0 2 0
and expectation values are given by:
t
u2
Z
E euBt − 1 = E euBt dt
   
(8.38)
2 0
2
Solving this differential equation gives the well-known result E euBt = etu /2 .
 

All these computations can be unified in the following result which is a one of the key results of
the bridge between PDE and stochastic calculus.
126 Chapter 8. Brownian motion and Itô calculus

Theorem 8.2.5. Let u : R+ × R → R be a C 2 solution of the partial differential equation

1
∂t u − ∂x2 u = 0 (8.39)
2
such that, for all δ > 0 and for all t ≥ 0, there exist Cδ,t > 0 such that u(s, x) ≤ C exp δx2 for
s ∈ [0, t]. Then, the quantity f (t, Bt ) is a martingale.

Proof. The inequality on u is only present to ensure the integrability of f (t, Bt ). Itô’s calculus and
additivity of the integral gives the conditional expectation requirement for martingales. A detailed
proof is given in the next chapter.

This theorem, together with Feynman-Kac formula, explains how Brownian motion is the
probabilistic counterpart of diffusion PDE and how its properties are related to the heat kernel.

8.2.2.4 Feynman-Kac formula

The last important result we present in this section is Feynman-Kac formula, which describes
solutions of diffusion PDE as well-chose expectation values over Brownian motion.

Theorem 8.2.6 (Feynman-Kac formula (a particular case)). Let g : R+ × R → R, φ : R → R and

V : R → R+ be measurable functions such that there exists a function u that is continuous and has
continuous derivatives and that is solution of the PDE:
1 2
∂t u(t, x) = ∂ u(t, x) − V (x)u(t, x) + g(t, x) (8.40)
2 x
with initial conditions u(0, x) = φ(x). Then, u(t, x) can be represented as an expectation value
 Rt 
− V (Bs )ds
u(t, x) =Ex φ(Bt )e 0
Z t Rs  (8.41)
− V (Bu )du
+ Ex g(t − s, Bs )e 0 ds
0

where (Bt )t∈R+ is a Brownian motion started2 at x.

Before the proof, several remarks must be done.

First, we do not give precise hypotheses about the functions g, c and φ and invite the reader to
refer to any reference about differential equations; in most cases, we have explicit functions and
shortcuts can be used to obtain the existence of u easily.
The relation between stochastic evolution of the Brownian motion and the PDE evolution is
non-trivial: u(t, x) is the value of u(t, x) at x after the evolution time t, whereas x appears in the
r.h.s. as the initial point of the Brownian motion. Conversely, φ appears the initial condition of the
PDE, whereas φ is functional of Bt (and not B0 ). There is an underlying time reversal, which is
indeed used in the proof below.

Proof. The main difficulty of this proof is to recognize functions of Brownian motion such that Itô’s
formula can be applied. We fix t > 0 and introduce the time-reversed solution of (8.40):

u∗ (s, x) = u(t − s, x)

The process defined as Rs

− V (Bu )du
Vs = e 0

does not require any Itô’s formula since it does not contain any stochastic integral but only an
integral over Brownian motion: we check that Vs has derivable trajectories and is a bounded

2 i.e. it can be written as x + Wt where Wt is a Brownian motion starting at 0, as constructed previously.

8.2 Itô calculus and stochastic integration 127

variation process: the formalism Rof stochastic integral is useless and standard variational calculus
t
apply (in particular Vs − V0 = 0 Vs′ ds and Vs′ = V (Bs )Vs ). The interested reader may check
3
easily or read in the next chapter (or prove the very simple case V = 0) that Itô’s formula can be
generalized as:
Z t Z t Z t 
′
u∗ (s, Bs )Vs − u∗ (0, x)V0 = ∂x u∗ (s, Bs )Vs dBs + ∂t u∗ (s, Bs )Vs ds + u∗ (s, Bs )Vs ds
0 0 0
t
1
Z
+ ∂x2 u∗ (s, Bs )Vs ds
2 0

Reordering the terms of r.h.s. gives:

Z t Z t  
1
u∗ (s, Bs )Vs − u∗ (0, x)V0 = ∂x u∗ (s, Bs )Vs dBs + Vs ∂t u∗ (s, Bs ) + ∂x2 u∗ (s, Bs ) + u∗ (s, Bs )V (Bs ) ds
0 0 2
Using the time-reversal and (8.40) gives
Z t Z t
u∗ (s, Bs )Vs − u∗ (0, x)V0 = ∂x u∗ (s, Bs )Vs dBs − Vs g(t − s, Bs )ds
0 0

The expectation of the first term of the r.h.s. is zero by construction of the stochastic Itô integral.
We have u∗ (0, x)V0 = u(t, x) and thus the final formula:
Z t 
u(t, x) = Ex [u(0, Bs )Vs ] + Ex Vs g(t − s, Bs )ds
0

An easy example. We look for a solution u of the PDE

(
∂t u(t, x) = 21 ∂x2 u(t, x) + x2 t5
u(0, x) = 0

that is continuous and with continuous derivatives. We first apply Feynman-Kac formula and
represent u as:
Z t  Z t Z t
(t − s)5 Bs2 ds = (t − s)5 E0 (x + Ws )2 ds = (t − s)5 E0 (x + Ws )2 ds
   
u(t, x) = Ex
0 0 0
t
x2 t6 + t7
Z
= (t − s)5 (x2 + s)ds =
0 6
The previous example is very simple and gives a result which could be even guessed! In general,
the philosophy behind the use of Feynman-Kac formula is the reduction of a PDE problem to
Brownian questions, with the belief that enough things are known about Brownian motion to solve
this question. In particular, one could recover many results about PDE from probability theorems.
This Feynman-Kac formula is a particular case of a class of general results relating integro-
differential linear equations and probabilistic diffusions: the reader is more invited to understand
how this formula was proved, rather than the single result.
The case of PDE with Dirichlet boundary conditions can also be integrated to such a formalism
and leads to stopped processes and the reader is encouraged to refer to specialized books about
the subject. The next chapter, which is about general stochastic calculus, does not contain an
exploration of such applications.
The case of some non-linear PDE (there are positivity constraints on the non-linearity) can be
integrated in this formalism but, in this case, it requires Markov processes indexed by trees and not
by R+ anymore.

3 Exercise: adapt the proof of Itô’s formula to a product f (s, Bs )Hs where Hs has continuous derivable trajectories.
128 Chapter 8. Brownian motion and Itô calculus

Non-trivial application to a PDE upper bound problem. We consider the PDE

1 2
∂t u = ∂ u − V (x)u (8.42)
2 x
with V ≥ 0 such that u is continuous with continuous derivative and such that (0, x) = 1. We wish
a bound on u(t, x) at large time. To this purpose, we write u as
 Rt 
− V (B )ds
u(t, x) = E exp 0 x s (8.43)

where fx (·) = f (x + ·). We introduce variables d(x, t) ≥ 0 and we want to know whether the
Brownian motion went out of the domain [−d(x, t), d(x, t)] before time t. We thus have
 Rt 
− Vx (Bs )ds
u(t, x) =E exp 0 1sup0≤s≤t |Bt |>d(x,t)
 Rt  (8.44)
− Vx (Bs )ds
+ E exp 0 1sup0≤s≤t |Bt |≤d(x,t)

We now control both terms. On the first event, we have no control on the trajectory: we use
only V (x) ≥ inf R V = V1 but wish that this event is rare: it is indeed the case from the maximal
inequality   √
3 C t
P sup |Bs | ≥ d(x, t) ≤ sup E [|Bs |] = (8.45)
0≤s≤t d(x, t) 0≤s≤t d(x, t)
On the second event, we need only to minimize V over [x − d(x, t), x + d(x, t)] (and call it
mV (x, d(x, t))) but this event may not be rare and we bound its probability by 1. We then obtain:
√
C t
u(t, x) ≤ + exp (−tmV (x, d(x, s))) (8.46)
d(x, t)

for any d(x, t). Depending on the hypothesis on V (behaviour at ∞), one may choose some optimal
choice of d(x, t) and obtain a bound about the variation of u.