Theory tells me that random variable $Z$ distributes with known mean $\mu$, over support $(1,\infty)$. Unfortunately, I cannot observe $Z$. However, I do observe $X$, which sample mean is $\bar u$, over support $(0,1)$. Importantly, theory tells me that when $X=0$, $Z=1$, and that when $X=1$, $Z=\infty$.

I am looking for a function $f(X)$ which does two things, simultaneously:

  1. Its range is the support of $Z$. For example:

$$f(X) = \frac{1}{1-X}$$

where it holds that $f(0)=1$ and $f(1)=\infty$.

  1. Transforms the mean of $X$ into that of $Z$. For example:

$$ g(X) = \frac{\mu}{\bar u}X $$

where it holds that

$$ E(g(X) = \frac{\mu}{\bar u}E(X) = \mu = E(Z) $$

I am struggling to find a transformation that does both things simultaneously. Clearly $f(X)$ is not a solution because $E[f(X)]\neq \mu$ (except by chance). Similarly, $g(X)$ has range $(0,\dfrac{\mu}{\bar u})$.

Such transformation (which must be non-linear) will change the distribution of $X$. Because I do not know the distribution of $Z$, I am for the moment not concerned about this.

Any ideas? The combination of functions 1 and 2 cannot do the trick, because to keep the range it must be that, for an arbitrary function

$$ h(X) = \frac{c}{c-dX}$$

in order to keep the range fixed, $d=c$, thereby reducing $h(X)$ to $f(X)$, which we know is not a solution.

I am looking at logarithms, exponentials, and so on, but it seems impossible to get both things at the same time. Maybe I am hitting a wall that has been systematised in a theorem?

  • 1
    $\begingroup$ Wikipedia has a detailed page on log and logit transformations of the Beta distribution. While these map [0, 1] to [0, infinity] There's no reason you can't shift the resulting distribution, e.g. the 1+log(U) with U a uniform RV generates an shifted exponential RV with mean 2. $\endgroup$ – AdamO Mar 22 '18 at 16:57
  • $\begingroup$ @AdamO Thanks, but log(x) map into [0,infinity] with the wrong limits. This is, 0 goes into infinity, and 1 into 0. This is contrary to my theoretical restriction. The logit maps into all the real line, so not helpful either. $\endgroup$ – luchonacho Mar 22 '18 at 17:37
  • $\begingroup$ Then do log(1-X) $\endgroup$ – AdamO Mar 22 '18 at 17:40
  • $\begingroup$ @AdamO 1-log(1-X) works in terms of having the same limits than Z, but its mean is fixed. I need to transform it to that of Z. Without a further parameter in the transformation, this is not possible. But any parameter you add will break the limits. Back to nothing. $\endgroup$ – luchonacho Mar 22 '18 at 17:59
  • $\begingroup$ @AdamO Error! Adding a C in front of log gives the constant I need. Then I can adjust C to match any mean. Amazing! Thanks! Do you want to post the answer? $\endgroup$ – luchonacho Mar 22 '18 at 18:15

Probability theory may lead to a quicker solution than pure mathematics here. Begin with simpler requirements by taking a known random variable and transformation. For example, a uniform random variable $U$ has a [0-1] support, and $-log(1-U)$ is exponentially distributed as $E$ with a [0, infinity] support. Note here that 0 maps to 0 and 1 maps to infinity. The mean of an exponential random variable is 1. As exponentials are part of a scale family, $cE$ has a mean of $c$. A shifted exponential is generated by $1+cE$. Inverting this distribution arrives at $1-\exp(-cE)$ which takes a $\text{Beta}(c, 1)$ distribution. Inverse mapping theorems will allow you to map any RV distributed on [0,1] to the quantiles of the Beta RV with (c,1) parametrization, then after applying the transform, generate a targeted outcome with [1, $\infty$] support and known mean.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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