4
$\begingroup$

Let $W\sim\mathcal{N}(0,\sigma^2)$ be a "reference" Gaussian random variable.

Suppose I have a set of distributions, $\mathcal{W}$, where $W_a\in\mathcal{W}$ if it satisfies the following criteria:

  1. Symmetry: $E[W_a^{2k-1}]=0$ for all $k=1,2,\ldots$
  2. Moments close to reference: $E[W_a^{2k}]\in\left[E[W^{2k}]/\rho^k,\rho^kE[W^{2k}]\right]$ for all $k=1,2,\ldots$ and some $\rho>1$.

I am wondering if this has implications for the total variation distance between the distributions in $\mathcal{W}$, that is do the above constraints yield an upper bound on $\frac{1}{2}\|W-W_a\|_1$?

This question arose as I was reading this paper (an early pre-print is available here for free). In particular, the authors seem to claim (in the proof of Theorem 1) that the distributions in $\mathcal{W}$ cannot be separated via a hypothesis test even given an infinite number of samples. However, their mathematics are rather sloppy, and, thus, I am a little skeptical of their claim.

$\endgroup$

1 Answer 1

5
$\begingroup$

No, you need some other condition if you want to restrict the total variation distance. The total variation distance is maximal between a discrete random variable and a continuous one, and you can take a discrete approximation of your Gaussian by rounding the values to the nearest tenth, say, without affecting the moments much.

$\endgroup$
2
  • $\begingroup$ Great point -- this completely makes sense. Now I wonder what are these other conditions on the distributions in $\mathcal{W}$ that would result in the total variation distance bound. Obviously, if $\mathcal{W}$ contains a set of slightly disturbed $W$'s, the bound should trivially hold. I wonder if the set is larger... $\endgroup$
    – Bullmoose
    Jan 16, 2015 at 21:16
  • $\begingroup$ @Bullmoose: One possibility would be to impose conditions such as that the density is unimodal. Another is to look at a different distance than the total variation distance. $\endgroup$ Jan 16, 2015 at 21:51

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.