Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Could someone help me to figure out solutions of following problems?:

Let $X = (X_t)_{t \geq 0}$ be a Gaussian, zero-mean stochastic process starting from $0$, i.e. $X_0 = 0$. Moreover, assume that the process has stationary increments, meaning that for all $t_1 \geq s_1, t_2 \geq s_2, ... , t_n \geq s_n$, the distribution of the vector $(X_{t_1} - X_{s_1} , ... ,X_{t_n} - X_{s_n})$ only depends on the time points through the differences $t_1 - s_1, ... , t_n - s_n$.

$(a)$

Show that for all $s, t \geq 0$

$$EX_sX_t =1/2 (v(s) + v(t) - v(|t - s|))$$

where the function $v$ is given by $v(t) = EX^2_t$

$(b)$

In addition to stationarity of the increments we now assume that $X$ is $H$-self similar for some parameter $H > 0$. Recall that this means that for every $a > 0$, the process $(X_{at})_t$ has the same finite dimensional distributions as $(a^HX_t)_t$.

Show that the variance function $v(t) = EX^2_t$ must be of the form $v(t) = Ct^{2H}$ for some constant $C>0$.

$(c)$

In view of the $(a),(b)$ we now assume that $X$ is a zero-mean Gaussian process with covariance function $EX_sX_t = 1/2(s^{2H} + t^{2H} - |t - s|^{2H})$ for some $H > 0$.

Show that we must have $H \leq 1$. (Hint: you may use that by Cauchy-Schwarz, the (semi-)metric $d(s, t) = \sqrt{ E(X_s - X_t)^2}$ on $[0, 1)$ satisfies the triangle inequality).

$(d)$

Show that for $H = 1$, we have $X_t = tZ$ a.s., for a standard normal random variable $Z$ not depending on $t$.

$(e)$

Show that for every value of the parameter $H \in (0, 1]$, the process $X$ has a continuous modification.


For now I have some ideas about $(a)$ and $(d)$ but have no clue about $(b), (c), (e)$

For $(a)$ my reasoning would be as follows:

$$EX_sX_t = E[(X_s-X_t)+X_t][(X_t - X_s)+X_s]$$ $$=-E(X_s-X_t)^2+E(X_s-X_t)X_s+EX_t(X_t-X_s)+EX_tX_s$$ $$=-E(X_s-X_t)^2+EX_s^2-EX_tX_s+EX_t^2-EX_tX_s+EX_tX_s$$

so we get

$$EX_sX_t=1/2 (v(s) + v(t) - E(X_s-X_t)^2)$$

but how to show that $E(X_s-X_t)^2 = v(|t - s|)$?

I would say that if $s \leq t$ then $E(X_s-X_t)^2 = E(X_{t-s}-X_0)^2= E(X_{t-s})^2=v(t - s)$ and by symmetry in the case of $s \geq t$ we get $E(X_s-X_t)^2 = v(|t - s|)$. Is this correct?

For $(d)$ I think in this way:

If $H=1$ then $EX_sX_t = 1/2(s^{2H} + t^{2H} - |t - s|^{2H})$ simplyfies to $1/2(s^2 + t^2 - |t - s|^2)$ and this is equal to $st$

on the other hand

$EtZsZ=stEZ^2=st$ in the case of $Z$ being standard Gaussian. Because we are talking about Gaussian proccess it is described by mean $(=0)$ and covariance (st) we are done. Is this the right way?

As I wrote above I have no reasonable ideas about $(b), (c), (e)$.

share|improve this question
    
In order to show $(e)$ you might want to look at Kolmogorov continuity theorem. –  Stefan Hansen Sep 27 '12 at 8:47

1 Answer 1

For $(a)$ it can be convenient to look at the right-hand side. Suppose $t\geq s\geq 0$, then $v(|t-s|)=v(t-s)=E[X_{t-s}^2]=E[(X_t-X_s)^2]$. Therefore $$ \frac{1}{2}\left(v(t)+v(s)-v(|t-s|)\right)=\frac{1}{2}\left(E[X_t^2]+E[X_s^2]-E[(X_t-X_s)^2]\right)\\ =\frac{1}{2}\left(2E[X_tX_s]\right)=E[X_tX_s]. $$ If on the other hand $s>t\geq 0$, then $v(|t-s|)=v(s-t)=E[(X_s-X_t)^2]$ and here the same calculations apply.

Concerning $(b)$ we have that $a^H X_t\sim X_{at}$ for every $t\geq 0$ and $a>0$. In particular $$ a^{2H}E[X_t^2]=E[(a^H X_t)^2]=E[X_{at}^2], $$ and hence $v(at)=a^{2H} v(t)$ holds for every $t\geq 0$ and $a>0$. I am fairly certain, that the only function satisfying this property is the function $t\mapsto Ct^{2H}$ for some constant $C>0$. The positivity of $C$ is due to the fact that $t\mapsto v(t)$ is a non-negative function.

For $(c)$ we may use that the pseudo-metric $d(s,t)=\sqrt{E[(X_s-X_t)^2]}$ satisfies the triangle inequality. Note that $d(s,t)=\sqrt{v(|s-t|)}$ and that $v(t)=t^{2H}$ in this setup. Now the triangle inequality says that $$ d(s,t)\leq d(s,u)+d(u,t) $$ for every $s,t,u\geq 0$. By means of $v$ this is $$ \sqrt{|s-t|^{2H}}\leq \sqrt{|s-u|^{2H}}+\sqrt{|u-t|^{2H}} $$ which implies $$ |s-t|^H\leq |s-u|^H+|u-t|^H $$ for every $s,t,u\geq 0$. For this inequality to hold we must have $H\leq 1$.

For $(d)$ you are right in using that a Gaussian process is uniquely determined by its mean- and covariance function.

share|improve this answer
    
Please, by which argument you affirm that H is less than or equal 1 in the last step. –  Zbigniew May 10 '13 at 6:49
    
@Jonas Try $(s,u,t)=(0,1,2)$. –  Did May 14 '13 at 17:08
    
@Did God protect you for this clarification. –  Zbigniew May 15 '13 at 18:44
4  
@Jonas Who? $ $ –  Did May 15 '13 at 18:45
2  
@Did Existence and uniqueness of God is assumed. –  Zbigniew May 16 '13 at 19:35

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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