0
$\begingroup$

Consider two random variables $Z$ and $W$. Given the variances of $Z$ and $W$, how can we compute the variance of their convolution $Z \circledast W $?

As an example, please consider the case of noise ($Z$) in an image being modeled by a Gaussian distribution of zero mean and a given variance $\sigma_Z^2$. Suppose we perform a convolution of $Z$ with a Gaussian kernel denoted by $W$. How can we find the variance of $Z \circledast W $? I am not able to find a good reference for this. Thanks for any help.

Improve this question
$\endgroup$
5
  • 2
    $\begingroup$ Can you explicitly show us the mathematical definition of $Z \circledast W $? To the best of my knowledge, in probability, the operands of a convolution operator are two distributions, i.e., the convolution of $F$ and $G$ is defined as $(F * G)(y) = \int G(y - x)dF(x)$, which corresponds to the distribution of the sum of two independent r.v.s $X$ (whose distribution is $F$) and $Y$ (whose distribution is $G$). How do you define the convolution of two r.v.s directly at the r.v.-level? $\endgroup$
    – Zhanxiong
    Commented Mar 18 at 13:18
  • $\begingroup$ Are you convolving the variables or their densities? The two things are very different! The convolution of the densities is the density of the sum of the variables: see stats.stackexchange.com/questions/331973 Your example suggests the latter, in which case the variances add provided the "noise" is independent of the image. See stats.stackexchange.com/questions/38721 $\endgroup$
    – whuber ♦
    Commented Mar 18 at 15:03
  • $\begingroup$ @Zhanxiong Thanks for your comment. My apologies for the confusion. I was referring to the convolution of random variables. I think $W$ may be treated as a constant kernel without any statistics associated with it. I am wondering how the convolution of $W$ with $Z$ would affect the variance of $Z$. $\endgroup$
    – user409495
    Commented Mar 19 at 18:38
  • $\begingroup$ @whuber Thanks for the reply, and sorry for the confusion. I was referring to convolution of the variables. I think in this particular case, $W$ may be treated as a constant kernel. I am not sure how would the convolution with $W$ affect the variance of $Z$. $\endgroup$
    – user409495
    Commented Mar 19 at 18:42
  • 1
    $\begingroup$ @user409495 Can you define $Z \circledast W$ in mathematics symbols, as opposed to plain English? Or can you provide any reference to such definition? $\endgroup$
    – Zhanxiong
    Commented Mar 19 at 18:45

1 Answer 1

Reset to default
-1
$\begingroup$

Like the previous answer posted, there's talk of convolution of the distributions of the 2 random variables. If these are independent, the convolution gives the distribution of the sum of 2 random variables. The characteristic function is the product of the 2 given ones. You can get the variance from this product; it is the coefficient of lambda^2. More precisely it is the second moment of the sum Z=X+Y. Or you can compute it directly as E[(X+Y)^2]. I hope this will help.

Improve this answer
$\endgroup$
1
  • 1
    $\begingroup$ Re "get the variance from this product:" that's unnecessarily complicated. It is basic (and easily proven) that the variances simply add. $\endgroup$
    – whuber ♦
    Commented Mar 19 at 13:25

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.