I have to prove that the following vector has normal distribution:

$$ \left( W_t, \int_0^t W_s \, ds\right) $$

where $W_t$ is a Brownian motion. I tried just by a forward computation on the characteristic function using the Ito derivative. In order to make an easier computation let $$ \Gamma_t=exp\left( i \xi W_t + i \eta \int_0^t W_s \, ds \right) $$ and $$ Y_t = \int_0^t W_s \, ds $$ We have: $$ d \Gamma_t = i\xi \Gamma_t dW_t+i \eta\Gamma_t dY_t + \frac{1}{2}\left( -|\xi|^2\Gamma_t \, dt\, - i\eta \Gamma_t \, d\langle Y_t\rangle -2\xi \eta \Gamma_t d \langle Y_t, W_t\rangle\right) $$

Now, since we have that $$ dY_t = W_t dt $$ it follows that $$ d \langle Y_t \rangle = 0 $$


(this is a standard formal computation) $$ d \langle Y_t ,W_t\rangle= dY_t dW_t = W_t dt dW_t = 0 $$

Passing to the average: \begin{align} \mathbb{E} \left[ \Gamma_t \right] &=1 +i \eta \int_0^t \mathbb{E} \left[\Gamma_s dY_s \right]-\int_0^t\frac{|\xi|^2}{2} \mathbb{E} \left[ \Gamma_s \right] \, ds \\ &= 1 + i \eta \int_0^t \mathbb{E} \left[ \Gamma_s W_s \right] ds -\int_0^t\frac{|\xi|^2}{2} \mathbb{E} \left[ \Gamma_s \right] \, ds \end{align}

At this point I'm trying to get an ODE for $\mathbb{E}[\Gamma_t]$ but I'm a little bit stuck

  • $\begingroup$ $d \langle Y_t W_t\rangle = d \langle Y_t ,W_t\rangle$ ? $\endgroup$ – Behrouz Maleki Dec 14 '16 at 18:11
  • $\begingroup$ $\int_{0}^{t}W_sds\sim\mathcal{N}\left(0,\frac 13 t^3\right)$ $\endgroup$ – Behrouz Maleki Dec 14 '16 at 18:13
  • 1
    $\begingroup$ $\mathbf E[\Gamma_t dY_t]= E[\Gamma_t W_t]dt\neq 0$. But this is a nice try. $\endgroup$ – zhoraster Dec 14 '16 at 18:24
  • $\begingroup$ @zhoraster You are right, that is a huge mistake... $\endgroup$ – clarkmaio Dec 14 '16 at 18:34

At the end I think I've found the solution. Of course the problem is the non linearity of the term: $$ \mathbb{E}[\Gamma_t W_t] $$ In this case one can use this trick: $$ \partial_\xi \Gamma_t = i \Gamma_t W_t. $$ This comes in particoular from the exponential structure of the charateristic function so that it seems like to be a quite usefull trick for the future.

However let me continue the computation: \begin{align} \mathbb{E}[\Gamma_t] & = 1+i \eta\int_0^t \frac{1}{i}\partial_\xi\mathbb{E}[\Gamma_s] \, ds -\frac{|\xi|^2}{2}\int_0^t\mathbb{E}[\Gamma_s] \, ds \\ & = 1+\int^t_0 \eta \partial_\xi\mathbb{E}[\Gamma_t]- \frac{|\xi|^2}{2}\mathbb{E}[\Gamma_t] \, ds \end{align}

This is a good formulation for us since it means that the left side of the equation is a solution for the following PDE: $$ \partial_t V(t, \xi , \eta)=\eta \partial_\xi V(t,\xi,\eta)-\frac{|\xi|^2}{2}V(t,\xi,\eta) $$ with initial condition $V(0,\xi,\eta)=1$.

Now we are at "home" since we can use all our PDE-weapon to win. For example by choosing the solution of the form: $$ V=\exp(P(t,\xi,\eta)). $$

where $$ P(t,\xi,\eta)=a t^A \xi^2+b t^B \eta^2+c t^C \eta \xi. $$ Now the idea is just make a computation and choose suitable parameters. By appling the differentials operators we get the relation: $$ a A t^{A -1} \xi^2 + b B t^{B-1} \eta^2 + c C t^{C-1} \eta \xi = 2 a A t^A \xi \eta + c t^C \eta^2-\frac{\xi^2}{2} $$

This equation, as equivalence between two polynomes in the $\xi$ and $\eta$ variable, leads us to the following system for the coefficients:

$$ \left\{ \begin{array}{rcl} A a t^{A-1} & = & -\frac{1}{2} \\ Bbt^{B-1} & = & c t^C \\ 2 a A t^A & = & cCt^{C-1} \end{array} \right. $$

Ok...now this gives us (finally)

$$ \left\{ \begin{array}{rcl} A & = & 1 \\ a & = & -\frac{1}{2} \\ B & = & 3 \\ b & = & -\frac{1}{6} \\ C & = & 2\\ c & = & -\frac{1}{2}. \end{array} \right. $$

  • $\begingroup$ I would like to stress an huge problem I found while doing these computations: the homogeneous solution is not unique since there is also the variable $\eta$ and since we haven't initial condition on $u$. For example we could choose as solution a function $\psi(\eta)$ or $\exp(\xi / t \eta)$ $\endgroup$ – clarkmaio Dec 20 '16 at 23:31

Recall the following statement:

Let $X: \Omega \to \mathbb{R}^n$ be a Gaussian random vector and $A \in \mathbb{R}^{k \times n}$ a (deterministic) matrix. Then $Y := A \cdot X$ is Gaussian.

Since $(W_t)_{t \geq 0}$ is a Brownian motion, we know that $X:=(W_{t_0},W_{t_1},\ldots,W_{t_n})$ is Gaussian for any $t_0,\ldots,t_n$. If we choose $t_j := \frac{j}{n} t$, $n \in \mathbb{N}$, and define

$$A := \begin{pmatrix} n^{-1} & \ldots & n^{-1} & 0 \\ 0 & \ldots & 0 & 1 \end{pmatrix} \in \mathbb{R}^{2 \times (n+1)}$$

then, by the above statement, the vector

$$A \cdot X = \left( \frac{1}{n} \sum_{j=0}^{n-1} W_{t_j}, W_{t_n} \right) = \left( \sum_{j=0}^{n-1} W_{t_j} (t_{j+1}-t_j), W_t \right)$$

is Gaussian. Consequently, we find that

$$\left( \int_0^t W_s \, ds, W_t \right) = \lim_{n \to \infty} \left( \sum_{j=0}^{n-1} W_{t_j} (t_{j+1}-t_j), W_t \right)$$

is Gaussian as a pointwise limit of Gaussian random variables (see e.g. this question).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.