Probably a dumb question but I was struggling to reconcile the fact that a uniform distribution over the reals is impossible with the fact that we can define a uniform over a finite subset of the reals and then put this subset in bijection with the entire set of reals.

For example if we have P(x) = U(0,1), can we not simply define Q(x) as, say, P(tan-1(x)), which will just "stretch out" the distribution P. It seems like for every point in [0, $ \mathbb R^+ $), it will be mapped onto a point in [0,1] and will therefore by definition have density of 1, so we end up with a uniform over the reals with density 1 everywhere.

I understand that this is impossible and I also see how intuitively it doesn't look like a real because intervals of equal length in [0,1] don't get mapped onto equal intervals in [0, $ \mathbb R^+ $). But I'm not seeing the formal reasoning behind this since surely all finite intervals have measure 0 anyway so they are equal

My gut feel is that the answer is something like "There's no guarantee that the property of beng a valid distribution of preserved under bijection etc. So go ahead and define your Q(x), it wouldn't automatically make it a distribution". Is this on the right track?



1 Answer 1


What you describe does not define a uniform distribution, since the measure of an interval $[x,x+1]$ far from 0 will be very small, as its preimage in $[0,1]$ is $[tan^{-1} x, tan^{-1} (x+1)]$. For a uniform distribution you would need that the derivative of your function be constant.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .