0
$\begingroup$

Consider the integral of a stationary random process $X(t)$ with mean $\mu$, variance $\sigma^2$ and stationary correlation function $\rho(t_1 - t_2)$:

$$I = \int_0^L X(t)\,\mathrm{dt}$$

In my previous post the moments of $I$ was found to be:

$$E[I] = \mu_I = L\mu$$

$$\text{Var}[I] = \sigma_I^2 =\sigma^2 \int_0^L \int_0^L \rho(t_1-t_2)\,\mathrm{dt_1\,dt_2} = 2\sigma^2\int_{0}^{L} \rho(\tau)(L-|\tau|)\,\mathrm d\tau$$

Where $\tau = t_1 - t_2$. I would now like the calculate the skewness of $I$:

$$\text{Skew}[I] = E\left[\left(\frac{I-\mu_I}{\sigma_I}\right)^3\right] = \frac{E[I^3] - 3\mu_I\sigma_I^2 - \mu_I^3}{\sigma_I^3}$$

I know $\mu_I$ and $\sigma_I^2$ so the problem is to evaluate $E[I^3]$:

$$E[I^3] = E\left[\int_0^L X(t)\,\mathrm{dt}\int_0^L X(u)\,\mathrm{du} \int_0^L X(v)\,\mathrm{dv}\right]$$

$$ = E\left[\int_0^L \int_0^L \int_0^L X(t) X(u)X(v)\,\mathrm{dt}\,\mathrm{du}\,\mathrm{dv}\right]$$

$$ = \int_0^L \int_0^L \int_0^L E\left[X(t) X(u)X(v)\right]\,\mathrm{dt}\,\mathrm{du}\,\mathrm{dv}$$

In the previous post the expected value was then written in terms of the covariance function however I don't know how to proceed here.

Can you show me how to evaluate the 3rd moment (if possible)?

$\endgroup$
  • $\begingroup$ Is $X(t)$ a Gaussian process? If so $I$ is also Gaussian with zero skew. $\endgroup$ – Jarle Tufto Aug 8 '16 at 16:52
  • $\begingroup$ I am interested in a lognormal process $X(t)$. My understanding is that it is not possible to obtain the probability density function of the sum of lognormal random variables. So I am trying to use the moments (mean, variance etc) instead. It is my suspicion that as the autocorrelation of $X(t)$ decreases the skewness of the distribution of $I$ decreases. Heuristically the distribution of $I$ looks more normal and less lognormal as the autocorrelation of $X(t)$ decreases. If the $\rho(s_1,s_2)$ is the Dirac delta function then $X(t)$ is a white-noise process and $I$ will be a constant. $\endgroup$ – egg Aug 9 '16 at 8:50
2
$\begingroup$

The above expectation inside your triple integral can be found via the following trick.

When $X(t)$, as you say, is a log-normal process, then $(Y(r),Y(s), Y(t)) = (\ln X(r), \ln X(s), \ln X(t))$ is multinormal with moment generating function $$ M(u,v,w) = E(e^{u Y(r) + v Y(s) + w Y(t)})=e^{\mu(u+v+w) + \frac12(u,v,w)^T\Sigma (u,v,w)} $$ where $\Sigma$ is the covariance matrix of $Y(r),Y(s),Y(t)$. Presumably you know the elements of $\Sigma$ if you know the autocovariance function of $Y(t)=\ln X(t)$? From this you can work out the above expectation since this equals $$ E(X(r)X(s)X(t)) = E(e^{Y(r)+Y(s)+Y(t)}) = M(1,1,1), $$ by definition of the moment generating function.

$\endgroup$
  • $\begingroup$ $$E[X(t)X(u)X(v)] = M(1,1,1) = \exp\left(\mu(1+1+1) + \frac{1}{2}(1,1,1)^T\boldsymbol\Sigma(1,1,1)\right)$$ $$= \exp\left(3\mu + \frac{\sigma^2}{2}(1,1,1)^T\boldsymbol\rho(1,1,1)\right)$$ $$\boldsymbol\rho = \begin{bmatrix} 1 & \rho(t-u) & \rho(t-v)\\ \rho(t-u) & 1 & \rho(u-v) \\ \rho(t-v) & \rho(u-v) & 1 \end{bmatrix} $$ $$= \exp\left(3\mu + \frac{\sigma^2}{2}\left(3+2(\rho(t-u) + \rho(t-v)+\rho(v-u)\right)\right)$$ I think I need to triple integrate this now. $\endgroup$ – egg Aug 9 '16 at 11:26
-1
$\begingroup$

You could appeal to Fubini's theorem to rewrite the third moment as: $$E[X^3] = E_X\left[\int_0^L\int_0^L\int_0^L E[X(t)X(u)]X(v) dt dudv \right],$$ and apply what you have already calculated for $E[X(t)X(u)]$. You may end up with a not-so-nice expression.

$\endgroup$
  • $\begingroup$ There are too many expectation operators and misplaced square brackets. Also, you cannot separate $E[X(t)X(u)X(v)]$ into $E_X\left[E[X(t)X(u)]X(v)\right]$ $\endgroup$ – Dilip Sarwate Aug 30 '16 at 15:05

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.