1: \begin{abstract}

2: 	Consider an empirical measure $\PP_n$ induced by $n$ iid samples from a $d$-dimensional

3: 	$K$-subgaussian distribution $\PP$ and let $\gamma = \mathcal{N}(0,\sigma^2 I_d)$ be the

4: 	isotropic Gaussian measure.

5: 	We study the speed of convergence of the smoothed Wasserstein distance $W_2(\PP_n *

6: 	\gamma, \PP*\gamma) = n^{-\alpha + o(1)}$ with $*$ being the convolution of measures. For $K<\sigma$ and in any dimension $d\ge 1$ we show

7: 	that $\alpha = {1\over2}$. For $K>\sigma$ in dimension $d=1$ we

8: 	show that the rate is slower and is given by $\alpha = {(\sigma^2 + K^2)^2\over 4 (\sigma^4 + K^4)} <

9: 	1/2$. 

10: 	This resolves several open problems in~\cite{goldfeld2020convergence}, and in particular

11: 	precisely identifies the amount of smoothing $\sigma$ needed to obtain a parametric rate. 

12: 	In addition, we also establish that $D_{KL}(\PP_n * \gamma \|\PP*\gamma)$ has rate

13: 	$O(1/n)$ for $K<\sigma$ but only slows

14: 	down to $O({(\log n)^{d+1}\over n})$ for $K>\sigma$. The surprising difference of the

15: 	behavior of $W_2^2$ and KL implies the failure of $T_{2}$-transportation inequality when

16: 	$\sigma < K$. Consequently, it follows that for $K>\sigma$ the log-Sobolev inequality (LSI) for the Gaussian mixture $\mathbb{P} *

17: 	\mathcal{N}(0, \sigma^{2})$ cannot hold. This closes an open problem in~\cite{wang2016functional}, who

18: 	established the LSI under the condition $K<\sigma$ and asked if their bound can be

19: 	improved.

20: \end{abstract}

21: