Inequality for Subgaussian Norms and Lipschitz Functions

85 Views Asked by At

If I have a $L$-Lipschitz function $\psi \colon \mathbb{R} \longrightarrow \mathbb{R}$ with $\psi(0) = 0$, and a random variable $X \sim \mathcal{N}(0, \sigma^2)$, is it true that

$$ \|\psi(X)\|^2_{\phi_2} \leq L^2 \sigma^2 \|Y\|^2_{\phi_2},$$

where $Y \sim \mathcal{N}(0,1)$ and $\| \cdot \|_{\phi_2}$ denotes the subgaussian norm?

I found a similar result which states $$ | \mathbb{E}[\psi(X)^2] - \mathbb{E}[\psi(Z)^2] | \leq L^2 |\sigma^2 - \rho^2| $$ for another random variable $Z \sim \mathcal{N}(0, \rho^2)$, but I don't see how to use this or if it is even helpful. I don't think I can set $Z$ to zero can I? And even if I could, I would be missing the term with the norm of the standard normal in the desired result... any help would be appreciated.

1

There are 1 best solutions below

2
On BEST ANSWER

This seems to follow directly from the definition of the Orlicz norm. By the Lipshitz condition $|\psi(X)| \le L |X|$. Hence, since $\phi_2$ is increasing,

$$ \| \psi(X) \|_{\phi_2 } = \inf \{ u> 0 : E \left[\phi_2\left( \frac{|\psi(X)|}{u} \right)\right] \le 1\} \le \inf \{ u> 0 : E \left[\phi_2\left( \frac{L| X|}{u} \right)\right] \le 1\} = L\|X \|_{\phi_2 }. $$ Replacing $X=\sigma Y$, where $Y$ is standard normal, we see $L\|X \|_{\phi_2 }=L\sigma \| Y \|_{\phi_2 }.$