For $(B_t)_{t\ge0}$ a standard Brownian motion (Wiener process) define the stochastic process $X_t:=\int_0^tB_s^2\ dB_s$. I am currently trying to assess if $(X_t)_{t\ge0}$ is a Gaussian process and if $(X_t)_{t\ge0}$ has independent increments.

I have already shown that $X_t$ is a martingale with $\mathbb E[X_t]=0$ and var$(X_t)=\mathbb E[X_t^2]=t^3$ (for var I used Ito’s isometry). Now my guess is that $(X_t)_{t\ge0}$ is not a Gaussian process and therefore I am trying to show that the moments of $X_t$ do not match the ones of a Gaussian distribution. So, for example, if I could show that $\mathbb E[X_t^4]\ne 3\cdot \text{var}(X_t)^2$, then this would prove that $X_t$ is not Gasussian, but I don't really know how. As for the increments, I don't really have a starting point either.

  • $\begingroup$ You mean $3\sigma^4$ instead of $3\text{var}(X_t)^4,$ right? $\endgroup$ – Ehsan M. Kermani Dec 6 '13 at 23:41
  • $\begingroup$ Oh, yes. Of course, since $\sigma\ne$var$(X_t)$. So $3$var$(X_t)^2$. $\endgroup$ – TinaL Dec 6 '13 at 23:52
  • 1
    $\begingroup$ So please edit it in your question. $\endgroup$ – Ehsan M. Kermani Dec 6 '13 at 23:53
  • 1
    $\begingroup$ please consider acceping this excellent answer by pressing the tick button. $\endgroup$ – Lost1 Jan 17 '14 at 18:31

The process $X$ is not gaussian and its increments are not independent.

Note first that $X$ is a Brownian martingale, hence a Brownian motion with a change of time, thus, it is distributed like $(\beta_{\langle X\rangle_t})$, where $\beta$ is a Brownian motion independent of $X$. For example, $X_1$ has the distribution of $\beta_{\langle X\rangle_1}=\sqrt{\alpha}\cdot\gamma$ where $\gamma$ is standard normal independent of $(X_t)$ and $\alpha=\langle X\rangle_1$. Thus, $E[X_1]=0$, $E[X_1^2]=E[\alpha]\cdot E[\gamma^2]=E[\alpha]$ and $E[X_1^4]=E[\alpha^2]\cdot E[\gamma^4]=3E[\alpha^2]$.

Since $E[Z^4]=3E[Z^2]^2$ for every centered normal random variable $Z$, these remarks show that if $X_1$ is normal then $E[\alpha^2]=E[\alpha]^2$, that is, $\alpha$ is almost surely constant. But $\alpha=\int\limits_0^1B_t^4\,\mathrm dt$ hence this is not so and $X_1$ is not normal.

To study the independence of the increments of $X$, fix some $s\geqslant0$ and consider the sigma-algebras $\mathcal F^X_s=\sigma(X_u;u\leqslant s)$ and $\mathcal F^B_s=\sigma(B_u;u\leqslant s)$, and the Brownian motion $C$ defined by $C_u=B_{s+u}-B_s$ for every $u\geqslant0$. Then $C$ is independent of $\mathcal F^B_s$. Furthermore, for every $t\geqslant0$, $$ X_{t+s}=X_s+\int_0^t(B_s+C_u)^2\mathrm dC_u=X_s+B_s^2C_t+2B_s\int_0^tC_s\mathrm dC_s+\int_0^tC_s^2\mathrm dC_s. $$ Rewrite this as $$ X_{t+s}-X_s=B_s^2C_t+B_sD_t+G_t, $$ where $D_t$ and $G_t$ are functionals of $C$ hence independent of $\mathcal F^B_s$. Thus, $$ E[(X_{t+s}-X_s)^2\mid\mathcal F^B_s]=B_s^4E[C_t^2]+B_s^2E[D_t^2]+E[G_t^2]+2B_s^3E[C_tD_t]+2B_s^2E[C_tG_t]+2B_sE[D_tG_t]. $$ One can check that $E[C_tD_t]=E[D_tG_t]=0$, $E[C_t^2]=t$, $E[D_t^2]=2t^2$, $E[G_t^2]=t^3$ and $E[C_tG_t]=\frac12t^2$ hence $$ E[(X_{t+s}-X_s)^2\mid\mathcal F^B_s]=tB_s^4+3t^2B_s^2+t^3. $$ Note that $\mathrm d\langle X\rangle_s=B_s^4\mathrm ds$ and that $\langle X\rangle$ is $\mathcal F^X$-adapted hence $B_s^4$ and every function of $B_s^4$, for example $B_s^2$, are measurable with respect to $\mathcal F^X_s$. This yields $$ E[(X_{t+s}-X_s)^2\mid\mathcal F^X_s]=tB_s^4+3t^2B_s^2+t^3. $$ The RHS is not almost surely constant hence $(X_{t+s}-X_s)^2$ is not independent of $\mathcal F^X_s$, in particular the increments of $X$ are not independent.

Edit: One may feel that the computation of the conditional expectation of $(X_{t+s}-X_s)^2$ above is rather cumbersome (it is) and try to replace it by the (definitely simpler) computation of the conditional expectation of $X_{t+s}-X_s$. Unfortunately, $$ E[X_{t+s}-X_s\mid\mathcal F^X_s]=0, $$ hence this computation is not sufficient to decide whether the conditional distribution of $X_{t+s}-X_s$ conditionally on $\mathcal F^X_s$ is constant or not (which is the reformulation of the independence of a random variable and a sigma-algebra this solution relies on). Another way of looking at the situation is that, fortunately, already the conditional second moments are not constant.

  • $\begingroup$ Sorry for asking but what does the notation $\langle X\rangle_1$ mean? $\endgroup$ – TinaL Dec 7 '13 at 13:37
  • $\begingroup$ The process $\langle X\rangle$ is the quadratic variation of the process $X$. $\endgroup$ – Did Dec 7 '13 at 14:59
  • $\begingroup$ So does part one show that any Brownian martingale (i.e. any Stochastic integral w.r.t. a Brownian motion) is not a Gaussian process? Because I don't see that we used any particular property of the process $X$ except for the fact that it is a Brownian martingale. $\endgroup$ – TinaL Dec 7 '13 at 16:12
  • $\begingroup$ Look better (for example, B itself is a Brownian martingale, and gaussian). $\endgroup$ – Did Dec 7 '13 at 16:14
  • $\begingroup$ Would you mind elaborating on that? I don't see what you mean. $\endgroup$ – TinaL Dec 7 '13 at 16:18

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.