Question about Convolution terminology (analog to mean, variance, skewness, kurtosis,...)

57 Views Asked by At

In probability theory we have mean, variance, skewness, kurtosis and etc. Can we use the same terms for a convolution (kernel) that sums to 1?

For example, can we say a $1$-D convolution kernel $[3/2,-2,3/2]$ has mean $0$ and $\sigma= 3$?

I'm wondering what basic results generalize from probability distributions to convolutions, or if there are any references that people would recommend (background: math and phy undergrad and grad.). Thanks!

1

There are 1 best solutions below

3
On BEST ANSWER

If you work with non-negative kernels, then after the normalization, you can treat it as a distribution of a delta-signal after one convolution step. So it's natural to talk about statistics of such distribution. Important: it's not the values statistics, it's position statistics assuming the values are probabilities. So if you have a kernel $[0.25, 0.5, 0.25]$ and assume the positions are $[-1,0,1]$, then the mean and sigma are $$m=-1\times0.25 +0\times 0.5+1\times0.25 = 0\\ \sigma=(-1)^2\times0.25+0^2\times0.5+1^2\times0.25=0.5$$

However, if you want to treat the numbers in the kernel as a set and calculate statistics over those, it makes little sense.