6
$\begingroup$

Say you are given one-dimensional data $X$, with mean $\mu$ and central moments $a_n$ which you know. Can you construct a function $f(x)$ which transforms the data such that $f(X)$ has the central moments $b_n$, also given?

For instance, say we want our second moment (variance) to be $b_2$, then $f(x)=(x-\mu) \sqrt{\frac{b_2}{a_2}}$. Can we extend this function to higher orders as well, such that we can also determine skew, kurtosis or even higher central moments?

If possible, can we have this $f(x)$ monotonically increasing?

I am looking for this in the context of ANN, where such a function might be interesting to work as some kind of copula to make the data more gaussian-like (or maybe uniform?).

Improve this question
$\endgroup$

1 Answer 1

Reset to default
2
$\begingroup$

The function $f(x)$ should be monotonically increasing, such that its inverse $g(y)=f^{-1}(y)$ is defined. Then you can use the theorem about random variable transformations, which says that, when $\varphi(x)$ is the probability density of $X$, then $Y=f(X)$ has the probability density $$\frac{d}{dy}\Phi(g(y))=\varphi(g(y))\cdot\frac{dg(y)}{dx}$$ When you choose a parametrized ansatz for the function $g(y)$, your moment condition transforms into a system of equations for the parameters. Depending on the parametric ansatz and the density function $\varphi$, this system of equations will require a numeric solution, but this is merely a technical point. I would thus conclude that the answer to your question is yes.

Improve this answer
$\endgroup$
4
  • 1
    $\begingroup$ There seems to be a serious technical point to address: how do you know that both random variables have probability densities at all?? $\endgroup$
    – whuber ♦
    Commented Apr 23, 2019 at 14:54
  • $\begingroup$ @whuber Yes, you are right: I made this assumption. Formally, this could be circumvented by setting $$\varphi(x) = \sum_{a\inX(\Omega)} P(X=a)\cdot(\delta(x-a)$$ $\endgroup$
    – cdalitz
    Commented Apr 23, 2019 at 15:02
  • $\begingroup$ I don't see how this works. Suppose, for example, that one set of moments gives a discrete distribution and the other set of moments gives a continuous distribution. There cannot possibly exist an invertible function $g$ that accomplishes what you claim in this case. $\endgroup$
    – whuber ♦
    Commented Apr 23, 2019 at 15:05
  • $\begingroup$ @whuber Sorry, I pressed the "enter" key for adding a newline, but this closed my comment. Hiere is the complete comment: Yes, you are right: I made this assumption, and it might be that my suggestion only works for continuous random variables. Formally, this could be generalized by setting $$\varphi(x) = \sum_{a\in X(\Omega)} P(X=a)\cdot\delta(x-a)$$ for a discrede random variable, where $\delta$ is Dirac's delta distribution. This would transform the integral for computing the central moments into a sum and still yield a set of equations for the parameters. $\endgroup$
    – cdalitz
    Commented Apr 23, 2019 at 15:10

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.