We respect your right to privacy. You can choose not to allow some types of cookies. Your cookie preferences will apply across our website.
What's the expectation abd variable of the stochstic process giveb by
[math] dX_t = (X_t \, (1-X_t))^r \, dW_t [/math]
?
I'm mostly interested in r equals 1/2 or 1.
I could add that it came up in modeling a distribution that can't take values outside of [0,1]. For r=1/2 it's a variation of the stochastic term in the "CIR model" and respecta the bounds I want
bump
senpaitachi
no takers at all?
I rewards with chocolate
>>8174227
explain briefly the theory behind this (only studied stuff like AR MA and ARMA kind of thing)
>>8174243
The ARMA model looks like a recurrence scheme where each new term is the previpus one, weighted by a phi and an independent/standalone fluctuation.
A stochastic differential equation scheme has x_i being x_{i-1} plus some function (set zero in the case I ask about) of that point, plus some function weigh multiplied by a fluctuation (dW)
I have 'a' equal zero and 'b' the quadratic function of X in the OP.
I wonder how the path of X will typically go. The function is not among the basic ones discussed in the books I now looked at.
Sorry for the crappy screenshots
>>8174400
ok tell me if this is bullshit for r=1:
separate variables
dXt/(Xt(1-Xt)) = dWt
when integarated, you get something like ln(Xt/(Xt-1)) = integral(dWt). Is this where you say it follows a normal distribution with std equal to sqrt(t) so N(0,t)
does that make any sense to you?
>>8174449
Such manipulations don't get you to an exact solution as (when thinking of a discretization) dW_t is not just a small positive quantity like in calculus, but instead a random value (normally distrubuted here, pic related).
I might use your rewriting in an algorithm to solve for X_t, but the stochastic Euler method sure is a more straight forward way to go and I did that (see third pic with the graph).
It's very unlikely that this equation has an exact silution, since I think to know that less complicated equations don't have one.
However, I'm only interested in the variance (or a similar meausre for the typical drift) of X_t anyway!
Pic related is the full model
>>8174483
I didn't pretend it was an exact solution -_-
my attempt was to show that the random variable Yt = ln(Xt/(1-Xt)) follows a normal distribution of mean 0 and variance t (because if I remember correctly, a random walk over time t has 0 mean and standard deviation sqrt(t))
from there you could probably derive the probability distribution Xt follows from that of Yt
>>8174507
The expectation <W> of
W=int dW
is zero, as dW is symmetric.
dX is random by definition of the equation. I don't know what to take from the expression when classically integrating dX/(X·(1-X)).
That a process X determines a process f(X), and how, that's the statement of the fundamental theorem for the Ito integral
https://en.wikipedia.org/wiki/It%C3%B4%27s_lemma#Mathematical_formulation_of_It.C3.B4.27s_lemma
The evolution of the probability distribution associated with X evolves according to the Fokker–Planck equation
https://en.wikipedia.org/wiki/Fokker%E2%80%93Planck_equation#One_dimension
>>8171965
https://en.wikipedia.org/wiki/Itô_diffusion
In particular, E(X_t)=EE(X_t|F_0)=EX_0 by the martingale property. So the expectation is just the expectation of the initial condition.
>>8174719
What's F and what about the expected drift?
>>8171965
Ito's lemma:
if Y = f(X)
then dY = df/dX * dX + 1/2 * d^2f / dX^2 * (dX)^2 + df/dt * dt
Apply the formula for f(X) = X - X^2
Note that dW * dW = dt
and df/dt is 0 in our case because the original formula (dX) has no dt coefficient.
>>8174957
I don't see what I'm supposed to use
d(X·(1-X))
for, once I've computed it. It it of use to compute the variance of X after a fixed time that passed?
>>8174976
In general the goal is to make a substitution and get a new equation of the form
dY = a * Y dt + b * Y dZ
or equivalently,
dY/Y = a * dt + b * dZ
(with a and b constants).
Because this implies
Y(t) = Y(0) * e^(a*t + b*Z(t)). (that Y follows Geometric Brownian motion and then you could literally just read off the expected value and variance (expected value is a*t, variance b^2 * t).
Of course that may not be possible in your problem. That would really depend on where it came from.
>>8175000
>That would really depend on where it came from.
Why? The differential equation is stated (>>>>8174493).
I have no preference on the specifics of dB. The X·(1-X) term is so that the value of X are confined to [0,1], as it's a distribution (of particular car gadgets over the totality of cars)
If you care, I develop the model here
https://axiomsofchoice.org/kfv_._note
(Paragraph 4 and 5 on linear approaches on the non-stochastic care are irrelevant)
>>8174493
The answer will depend on the distribution of X_0.
Let's fix any function f and consider the expression u(t,x)=E(f(X_t)|X_0=x).
Then this function u satisfies a deterministic PDE called Kolmogorov's backward equation which you can read off from the coefficients of your SDE. It's written out here http://math.stackexchange.com/questions/598811/kolmogorov-backward-equation-for-itô-diffusion or in, e.g. Oksendal.
Now let f(x)=x^2 and solve the KBE using your favorite numerical approximation scheme. This gives you E(X_t^2|X_0=x) and then of course
E(X_t^2)=\int E(X_t^2|X_0=x)p(x)dx where p(x) is the density of X_0. etc.