I understand why assumption of non-gaussianity is needed in ICA-model. I just can't find any source for why some signals e.g sound signals can be assumed to be non-gaussian. Doesn't everything in nature follow the gaussian distribution?

  • $\begingroup$ What about things in nature which cannot be negative or have an infinite range on values? E.g. Count data, say the number of times you observe an event in an hour, cannot be negative (and is usually considered to be poisson distributed). $\endgroup$
    – Dale C
    Commented Mar 6, 2020 at 2:52
  • $\begingroup$ First, not every variable is found in "nature" (at least, if I understand what you mean). Many variables are not close to normally distributed (e.g. count data as @DaleC mentioned but also income, wealth and other financial data; temporal variables; percntage valuse; other variables with various limits. Second, even in nature, variables aren't perfectly normal and may be far from normal. For instance, the weight of adults in the USA is somewhat fat-tailed. $\endgroup$
    – Peter Flom
    Commented Mar 8, 2020 at 12:13
  • $\begingroup$ This is about multivariate signals, and even if each marginal is normal, it might not be jointly normal! See stats.stackexchange.com/questions/30159/… $\endgroup$ Commented Mar 11, 2023 at 14:57


Your Answer

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