3 Itô Calculus Flashcards
Itô calculus
the main tool for studying stochastic differential equations
e.g. consider differential equation
x’(t)=0.5x(t)
x(0)=1
solution is….graph in next example
x(t)= Cexp(0.5t)
In general for diff eq
X’(t)=f(X(t)) sol
can be approximated by defining
X(tₙ₊₁) = X(tₙ) + (tₙ₊₁ − tₙ)(f(X(tₙ)),
as X’(t) is a limiting equation
X’(t) ~ (X(t+h)-X(t)) /h
giving the limit of a recursion
Consider a graph (time t against X(t)) looks like RHS of x^2 (X(1)=35)
x(t)= Cexp(0.5t)
the graph of a trajectory of the dynamical system
X(tₙ₊₁) = X(tₙ) + (tₙ₊₁ − tₙ)f(X(tₙ)),
X(0) = x
with f(x) = 6x, starting from x = 0.1, and time-step given by tₙ₊₁ − tₙ = 1/100 (interpolated
between the discrete time points).
This type of dynamics is a good approximation of the evolution of many quantities found in nature. However, in practice, the majority of systems in physics, finance,
and biology, among others, are noisy
noisy?
Let us look for example a stock price, and more specifically,
lets say the stock price of Apple the last 10 years. While there is obviously an exponential trend,
the price does not follow a “smooth” trajectory as in prev e.g. but rather fluctuates around it
differential equation is limiting!
we observe what happens at discrete times?
X(tₙ₊₁) = X(tₙ) + (tₙ₊₁ − tₙ)f(X(tₙ)),
Adding this noisyness
what we do is we take the original trajectory:
X(tₙ₊₁) = X(tₙ) + (tₙ₊₁ − tₙ)f(X(tₙ)),
and we add in a random variable ξₙ :
X(tₙ₊₁) = X(tₙ) + (tₙ₊₁ − tₙ)f(X(tₙ))+σ(X(tₙ))ξₙ,
RV ξₙ are gaussian :
ξₙ~N(0, tₙ-tₙ₊₁)
X(0) = x,
independent
original trajectory:
X(tₙ₊₁) = X(tₙ) + (tₙ₊₁ − tₙ)f(X(tₙ)),
and we add in a random variable ξₙ :
X(tₙ₊₁) = X(tₙ) + (tₙ₊₁ − tₙ)f(X(tₙ))+σ(X(tₙ))ξₙ,
RV ξₙ are gaussian :
ξₙ~N(0, tₙ-tₙ₊₁)
X(0) = x,
independent
(motivation)
Example in notes:
ξₙ~N(0,1/n)
and
σ(x)=x
where ξₙ can be modelled by Brownian increments W(tₙ)-W(tₙ₊₁)) (these have the property we need)
replacing gives:
[X(tₙ₊₁) - X(tₙ)]/[(tₙ₊₁ − tₙ)] = f(X(tₙ))+σ(X(tₙ))ξₙ/[tₙ₊₁ − tₙ]
[X(tₙ₊₁) - X(tₙ)]/[(tₙ₊₁ − tₙ)] = f(X(tₙ))+σ(X(tₙ))[W(tₙ)-W(tₙ₊₁)][tₙ₊₁ − tₙ]
[X(tₙ₊₁) - X(tₙ)]/[(tₙ₊₁ − tₙ)] = f(X(tₙ))+σ(X(tₙ))[W(tₙ)-W(tₙ₊₁)][tₙ₊₁ − tₙ]
as
[tₙ₊₁ − tₙ]→ 0?
(motivation)
ie when σ ≡ 0 the above system leads to
dX(t)/dt = f(X(t)) X(0)=x
σ not equal to 0:
dX(t)/dt =
f(X(t)) + σ(X(t)) dW(t)/dt ,
X(0) = x.
We have shown W not differentiable, the limit .derivative doesnt exist.
we need more tools to deal with stochastic diff equations!
Simple case:
σ(x) ≡ 1
dX(t)/dt =
f(X(t)) + σ(X(t)) dW(t)/dt ,
X(0) = x.
dX(t)/dt =
f(X(t)) + dW(t)/dt ,
X(0) = x.
integrating
X(t)= x+ ∫₀ᵗ f(X(s)) ds + W(t).
(eq makes sense)
e.g trivial is BM
if f(t)=1 then integral = W(t)
σ(x) ≡ not a constant
dX(t)/dt =
f(X(t)) + σ(X(t)) dW(t)/dt ,
X(0) = x.
integrating
X(t)=
x+ ∫₀ᵗ f(X(s)) ds
+∫₀ᵗ σ(X(t)) dW(s).
BUT
∫₀ᵗ σ(X(s)) dW(s)?
well W is not in C^1 with probability one
we will consider about convergence in different ways and our aim is to look at these integrals
∫₀ᵗ σ(X(s)) dW(s)?
RECALL
∫₀¹ g(s) dv(s).
for continuous funct g
if v ∈ C^1
, then the above integral is defined and is given by
∫₀¹ g(s) dv(s)
=
∫₀¹ g(s) dv(s)/ds . ds
(you’d take a partition and look at partial sums corresponding to it, see notes for further,
diagram Riemann integral, limit doesnt always exist- depends how regular g and v are
consider if alpha-holder continuous)
we will construct the integral
∫₀ᵗ Y(s) dW(s)
as
integral of Y against BM:
With
Y(t) as a stochastic process
We fix a probability space
(Ω, F, P) with a
**complete
and
right continuous filtration **
F := (F_t)_{t∈[0,T]}.
We assume that on Ω we are given an F-Wiener process W.
NOTE:
right continuous filtration **
F := (F_t)_{t∈[0,T]}.
usually
F is a wiener process
Filtration is
Fₜ= σ(W(s), s≤t )
= σ(W⁻¹(A), A in B(R) s≤t)
inverse images of Borel sets
Rememebr if we say a RV is measurable wrt. F_t it means that it is a function of your wiener process/path of your wiener process but only for time before t
Firstly consider Y…
Y(t)= 1_{[0,1]} (t)
Considering
∫₀ᵗ Y(s) dW(s)
Consider Y s.t its a function of time
Y(t)= 1_{[0,1]} (t)
graph
for t>0
1 for t in [0,1]
0 after
this is a deterministic function.
∫₀ᵗ Y(s) dW(s) =
(∫₀ᵗ 1. dW(s) )
(∫₀¹ 1. dW(s) + 0??)
=W(t)
(using increments and function g,
∫₀¹ 1. dg(s) = ∫₀¹ g’(s). ds = g(1)-g(0))
Now consider
∫₀^T of 1_[ₐ,ᵦ] (t) dW(t)
we expect this to be
∫ₐᵇ dW(t)/dt /dt
=W(b)-W(a)
but we will show this?
“=” not =
Definition 3.2.1.
in the class Hₜˢᵗᵉᵖ
We say that a stochastic process Y : Ω × [0, T] → R is in the class
Hₜˢᵗᵉᵖ
if there
exists n ∈ N and 0 ≤ t₁ < … < tₙ < tₙ₊₁ ≤ T
(and RV Y_1,…Y_n)
such that
1) Y(t)= Σᵢ₌₁ⁿ Yₜ_ᵢ 1_[tᵢ,tᵢ₊₁] (t)
2)Yₜ_ᵢ is Fₜ_ᵢ- measurable for all i = 1, …, n,
- ||Yₜ_ᵢ||_L∞ < ∞
for all i = 1, …, n.
first two are most important
elements of class Hₜˢᵗᵉᵖ are called
simple processes
we will construct stochastic processes from these as we know how to integrate the indicator functs/ characteristic funct
not unique can be written in multiple ways
we will consider such simple processes whose “step length” becomes closer to 0 to approximate the stochastic process, like partial Riemann sums to see if they converge
this is one realisation for an omega, if we change this the realisation changes
simple processes look like
look like
step functs
0 everywhere else 1 for interval t_i- t_i+1, multiplying these gives values Y_(t_i), step functions that are right continuous
the heights are random remember
but each function is t_i measurable, we cant take information from the future only the past
if we integrate against W we would expect this as the sum of Y_(t_i)(W(t_i+1)-W(t_i))
- ||Yₜ_ᵢ||_L∞ < ∞
for all i = 1, …, n. meaning
X is in L∞(Ω)
if there exists M in reals
s.t
P(|x|≤M)=1
If X was binomial: can only take values 0 or 1, so yes we can find M>1
If X Gaussian:
it isnt because the density shows X is everywhere by defn
For Y ∈ Hₜˢᵗᵉᵖ
integrals wrt W
DEFN
stochastic integral I(Y )
For Y ∈ H_T ˢᵗᵉᵖ
we set
I(Y)=
∫₀T Y(s) dW(s)
:=Σᵢ₌₁ⁿ Yₜ_ᵢ (W(tᵢ₊₁)-W(tᵢ))
(This is the sum of RVS, thus is a RV itself, we can discuss the expectation and variance)
stochastic integral I(Y )
REMARK
Notice that the stochastic integral I(Y ) is well defined (does not depend on the
partition t_1, …, t_n) and is a linear operator on Hₜˢᵗᵉᵖ. Also notice that for Y ∈ Hₜˢᵗᵉᵖ we have that I(Y ) ∈ L_2(Ω).
Indeed, it suffices to check that this is the case for Y (t) = Y_t_11_[t1,t1+1)
From Proposition 1.5.7 we have
E[|Yₜ_₁²|(W(t₂) − W(t₁))²]
= E[E[|Yₜ_₁|² (W(t₂) − W(t))²|Fₜ_₁]]
= E[|Yₜ_₁|² E[(W(t₂) − W(t))²|Fₜ_₁]]
=(t₂ − t₁)E[|Yₜ_₁|² ]
≤ (t₂ − t₁)||Yₜ_₁||²_L∞ < ∞,
which shows that I(Y ) ∈ L₂(Ω)
Exercise 3.2.3. Show that the stochastic integral I : Hₜˢᵗᵉᵖ → L₂(Ω) is a ** linear operator**,
that is, if Y₁, Y₂ ∈ Hₜ and a₁, a₂ ∈ R, then we have that
I(a₁Y₁ + a₂Y₂) = a₁I(Y₁) + a₂I(Y₂).
∫₀T a₁Y₁(s) + a₂Y₂(s) dW(s)
OPERATOR IS LINEAR
Suppose Y₁, Y₂ ∈ Hₜ and a₁, a₂ ∈ R, then we have that
I(a₁Y₁ + a₂Y₂) =
a₁∫₀T Y₁(s) dW(s)
+ a₂∫₀T Y₂(s) dW(s)
:=a₁Σᵢ₌₁ⁿ Y₁,ₜ_ᵢ (W(tᵢ₊₁)-W(tᵢ)) + a₂Σᵢ₌₁ⁿ Y₂, ₜ_ᵢ (W(tᵢ₊₁)-W(tᵢ))
=:a₁I(Y₁) + a₂I(Y₂).
he used T=1
Suppose we have function f: [0,1] → R
Define
∫₀1 f(s).ds
The Riemann integral will be the area under the curve by partition rectangles and calculating the area under. As the width of the partition interval tends to 0 we approach a limit if it exists
We could also define a function f_n defined as
f_n(s)=
Σᵢ₌₁ⁿ f( tᵢⁿ )1_[tᵢⁿ, tᵢ+₁ ⁿ] (t)
These are n constant functions on n intervals
we know how to find the integral for such functions
∫₀1f_n(s)
= Σᵢ₌₁ⁿ f( tᵢⁿ ) (tᵢ+₁ ⁿ-tᵢⁿ)
If this converges to some limit, this could be the integral if it is the limit for all partitions
STEP 1
find the sequence f_n that converges to f
STEP 2
f_n to be simple
STEP 3
ensure limit doesn’t depend on what partition you chose
Applying the prev steps to consider the integral of a stochastic process
Given a stochastic process (Y(t))_t in [0,1]
STEP 1
(Y_n(t))_ t in [0,1] in Hₜˢᵗᵉᵖ
find a sequence s.t simple and we can find the integrals
∫₀1 Yₙ(t) dW(s)
and that
STEP 2
(Yₙ(t))t in [0,1] converge to stochastic process (Y(t)) t in [0,1]
(Yₙ(t))_t in [0,1]→ (Y(t))_t in [0,1]
STEP 3
I want to show there exists a RV Z s.t
∫₀1 Yₙ(t) dW(s) → Z
Z will be the stochastic integral of Y