I understand the discete case i.e. the sum of $N$ identically distributed random variables $X_i$ with variance $\sigma^2$. The correlation between these random variables is given by the correlation matrix $\mathbf{\rho}(X_i,X_j)$.

The variance of the linear combination of random variables $X_i$ is given by:

$$\operatorname{Var}\left( \sum_{i=1}^N X_i\right) = N\sigma^2+2\sigma^2\sum_{1\le i<j\le N}\mathbf{\rho}(X_i,X_j)$$

Source: Wikipedia - Variance - Sum of correlated variables

I would like to consider the continuous case of a stochastic process which will be denoted as $X(t)$. The process is stationary with constant variance $\sigma^2$ and correlation function $\rho(X(t),X(h)$. Similar to above I would like to calculate the variance of the linear combination of the random variables $X(t)$. I think that the linear combination over some domain $t \in [0,L]$ can be expressed as

$$I = \int_0^L X(t) dt$$

I would like to know the variance of $I$:


I speculate that if the process $Z(X_i)$ is completely correlated i.e. $\rho(X(t),X(h)) = 1$ then the variance of $I$ is minimised maximised and is given by: $$\operatorname{Var}(I)=L^2\sigma^2$$

If the variables are uncorrelated i.e. $\rho(X(t),X(h))=0$ then I suspect that the variance of $I$ is maximised minimised. It may be infinite zero?

I find the continuous case (i.e. infinite random variables over some domain [0,L]) difficult to understand. Could anyone provide me an expression for the variance of $I$?


1 Answer 1


Interchanging the order of integration and expectation you get $$E(I)=E\int_0^L X(t) dt = \int_0^L EX(t) dt = \int_0^L \mu dt = L\mu $$ and similarly, the second moment of $I$ becomes \begin{align} E(I^2)&=E\left(\int_0^LX(t)dt\int_0^LX(u)du\right) \\ &= E \int_0^L \int_0^L X(t)X(u)du dt \\ &= \int_0^L \int_0^L E[X(t)X(u)]du dt \\ &= \int_0^L \int_0^L [\operatorname{Cov}(X(t),X(u))+EX(t)EX(u)] du dt \\ &= \sigma^2 \int_0^L \int_0^L \rho(t-u)dudt + L^2 \mu^2. \end{align} If the correlation function $\rho(h)$ is for example exponential the double intergral can be easily solved. The variance of $I$ can be found from the general formula $\operatorname{Var}I =E(I^2)-(EI)^2$.

  • $\begingroup$ Excellent. This makes sense to me. I will try to verify this with simulation. $\endgroup$
    – egg
    Aug 1, 2016 at 13:43
  • 3
    $\begingroup$ This assumes that the process is wide-sense stationary or weakly stationary. Also, for this special case, the double integral simplifies to a single integral $$\int_{-L}^{L} \rho(s)(L-|s|)\,\mathrm ds$$ which is easier to work with. $\endgroup$ Aug 2, 2016 at 17:52
  • 1
    $\begingroup$ Good comment. Integration using Wolfram Mathematica gives the same result using both expressions for $$\rho(s) = \exp\left(-\pi\left(\frac{s}{\theta}\right)^2\right)$$ Is it correct that your expression works for all isotropic correlation functions i.e. $s = |t - u|$? Could you explain how you got to your expression? My understanding of double integrals is poor. $\endgroup$
    – egg
    Aug 8, 2016 at 14:33
  • $\begingroup$ Even if unrelated to the question itself, I find a little confusing the statement $\text{Var}[I]=\text{E}[I^2]-\text{E}[I]^2$ as a consequence of the exponential correlation. Isn't it true for any random variable? $\endgroup$ Jan 10, 2020 at 11:49
  • $\begingroup$ @RiccardoBuscicchio Yes $\endgroup$ Jan 10, 2020 at 16:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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