# multivariate Gaussian approximation in total variation distance

I'm wondering if there's any general technique that gives the total variation distance between a distribution on $\mathbb{R}^n$ and $N(0, I_n)$.

My understanding is that Stein's method gives only Wasserstein distance in higher dimension because the characterization of multivariate Gaussian is a second-order differential equation (while it is a first-order differential equation in one-dimensional case) so more regularity is required on test functions and thus it yields a weaker distance. And I understand that it is possible to improve Wasserstein distance to total variation distance if the distribution is log-concave.

What is the usual way to handle the total variation distance to multivariate Gaussian? I'm primarily interested in approximating $N(0,I_n)$ but the approximating distribution is not necessarily log-concave. Perhaps there's some easy way for this special case? Or is there any impossibility result?

-
Unfortunately I cannot help you, since I just found out about the Stein's method. However, I was wondering if you could point me in the direction of the material behind the second paragraph of your question. I have a similar problem: I am trying to upper-bound the total variation distance between $N(0,I_n)$ and another distribution. My other distribution happens to be a mixture of multivariate Gaussians with unit variance, but non-zero vectors of means. Thus, I am wondering about the applicability of the last sentence of the second paragraph to mixtures of log-concave distributions. –  Bullmoose May 31 '13 at 8:30

Stein's method doesn't give total variation approximation in one dimension, either, without some kind of additional assumptions. This has nothing to do with Stein's method; for an impossibility result, any discrete distribution has maximal (1 or 2 depending on your normalization convention) total variation distance to any continuous (e.g. Gaussian) distribution. But of course you can approximate any distribution by a discrete distribution, in Wasserstein distance for example.

-

I am not sure the following helps.

If $f$ is a prob.distribution on $\mathbb{R}^n$ having the same covariance matrix as $N(0, I_n)$, then \begin{eqnarray*}D(f\|N(0, I_n))&=& H(N(0, I_n))-H(f)\\\ &=& \frac{1}{2}\log((2\pi e)^n)-H(f)\end{eqnarray*} where $D(\cdot \| \cdot)$ denotes the KL divergence $H$ denotes the Shannon entropy.

Now, by Pinsker's inequality, \begin{eqnarray*}\|f-N(0, I_n)\|_1 &\le& \sqrt{2 D(f\|N(0, I_n))}\\\ &=& \sqrt{\log((2\pi e)^n)-2 H(f)}\end{eqnarray*}

So the conclusion is that if the entropy $f$ is closer to that of Gaussian, $f$ will be closer to $N(0, I_n)$ in total variation. But it works only for those $f$ which also has covariance matrix $I_n$.

-