3

Usual non-constructive mathematics leads to some paradoxes (e.g. the Banach-Tarski paradox), which are directly related to the axiom of choice. In non-constructive mathematics, the axiom of choice (as well as proofs by contradiction) are not accepted.

For statistics, the axiom of choice implies that for any $c$ there must be a model such that the BIC (or also the AIC) is below $c$ (see this answer, we can arbitrarily increase the likelihood without increasing $N$, the number of parameters), which kind of defeats the purpose of the BIC as a model selection criterion. This method would not work in constructive math, because the space-filling curves or bijections don't exists in constructive mathematics.

Is there a general treatment of statistics in constructive mathematics? AFAIK a lot of parts of statistics are non-constructive (e.g. Gaussian distribution, Beta distribution, etc). It would be interesting to see which parts of statistics are different when the constructive approach is taken.

EDIT:

It seems that my understanding of constructive mathematics is still to limited for this. So I may have been wrong about the existence of the distributions or the space-filling curves in constructive mathematics. I will edit this question, once I get a better understanding. Feel free to edit, if you can shed more light on this topic.

LiKao
  • 2,329
  • 1
  • 17
  • 25
  • Can you please explain what is non-constructive with the Gaussian distribution? Why do I need the axiom of choice to simulate random variates from a Gaussian? – kjetil b halvorsen Sep 12 '19 at 14:21
  • What makes you think that space-filling curves are not possible in constructive mathematics? From what I recall the theorem that for any infinite A, there exists bijection between A and AxA in fact relies on nonconstructive assumptions, but note that negation of that statement just means that there exist sets for which it doesn't hold, so doesn't exclude space-filling curves – Jakub Bartczuk Sep 12 '19 at 15:03
  • @JakubBartczuk According to https://www.reddit.com/r/askscience/comments/6opvae/simplest_proof_that_r2_r/ the existence of a bijection between $A$ and $A\times A$ is equivalent to the axiom of choice (see answer by midtek). However, I am still trying to understand constructive mathematics, so I can't really see how this happens. – LiKao Sep 13 '19 at 08:22
  • @JakubBartczuk Ok, I should have read the reddit further. It seems only the general case $|A|=|A\times A|$ is equivalent to the AOC, but $|\mathbb{R}|=|\mathbb{R}\times \mathbb{R}|$ can actually be shown without it... Interesting. – LiKao Sep 13 '19 at 08:24
  • @kjetilbhalvorsen Ok, I see that I should invest more time into grasping constructive mathematics. I thought that, because the integral over the gaussian distribution has no closed form, it would not be possible to show that it actually is a distribution. But the more I think about it, there may be a way around this. I will close this question and edit once I get a better understanding of constructive mathematics. – LiKao Sep 13 '19 at 08:26
  • @LiKao “Closed form” isn’t a precise term. For instance, you may think that the integral of cosine has a closed form because it’s sine, but sine is just a name that we have to a useful infinite polynomial. So why not follow the same kind of logic and say that the integral of the standard normal is some function? – Dave Sep 13 '19 at 08:54
  • There is a book entirely dedicated to this ["constructive measure theory"](https://bookstore.ams.org/memo-1-116/) by Errett Bishop and Henry Cheng. And there is more if you look around (for instance on [mathoverflow](https://mathoverflow.net/questions/141764/reference-request-constructive-measure-theory)). Not being constructionist will lead to [non-physical conclusions](https://stats.stackexchange.com/questions/315502) that are nice for the mathematicians, but they make physically no sense (and personally I believe/feel that statistics is/should be about real, tangible, objects). – Sextus Empiricus Sep 16 '19 at 14:06

0 Answers0