Convolution of a pdf $f$ with a Gaussian $g$: distance between $g\ast f$ and $g$?

431 Views Asked by At

I have been looking for references on the following matter: let $f$ be the pdf of any real-value random variable ($f$ is not necessarily continuous wrt Lebesgue measure), and $g=g_{\mu,\sigma}$ be a Gaussian pdf. Writing $d_{\rm TV}$ for the total variation distance and $\ast$ for the convolution, is there anything that can be said between $d_{\rm TV}(f,g)$ and $d_{\rm TV}(g\ast f,g)$?

In particular, is there any result of the form $$ d_{\rm TV}(g\ast f,g) \leq \alpha\cdot d_{\rm TV}(f,g)^\beta $$ for some universal constants $\alpha,\beta >0$?

(if not, what would be a sufficient assumption on $f$? E.g., $f = a\mu + (1-a)\nu$ with $a > 0$ and $\mu$ a measure continuous wrt Lebesgue — would that be enough?)

If this is a straightforward exercise or a known and simple result (or, on the contrary, easily seen to be false), can someone point me towards either a proof of it, or a way/hint to prove it, or anything of that sort?

Edit: in case this could be easier (or hold while the above does not(?)), I am also interested in the same sort of result, but with the Earthmover's Distance (Wasserstein metric) instead of the more stringent total variation.

1

There are 1 best solutions below

3
On

For the Wasserstein distance, perhaps Lemma 7.1.10 in the book ``Gradient Flows in Metric Spaces and the Space of Probability Measures'' by Ambrosio, Gigli, and Savaré is what you are looking for.