Let $X$ be a centered (that is, $\mathbb{E}X=0$) sub-gaussian random variable. The sub-gaussian norm of $X$ is defined as $\|X\|_{\psi_2}=\inf\{t>0: \mathbb{E}[\exp(X^2/t^2 )]\leq 2\}$. In other words, $\mathbb{E}[\exp(X^2/\|X\|^2_{\psi_2} )]\leq 2$.

Sub-gaussian random variable $X$ can also be characterized by $\mathbb{E}[e^{\lambda X}]\leq e^{C\lambda^2\|X\|^2_{\psi_2}}$ for all $\lambda\in\mathbb{R}$. I am interested in finding the constant $C$.

How do I do it from first principle?

For example, we have the inequality $e^x\leq x+ e^{x^2}$ for all $x\in\mathbb{R}$. Then, taking expectation on both sides, $\mathbb{E}[e^{\lambda X}]\leq \mathbb{E}[e^{\lambda^2 X^2}]=\mathbb{E}\left[\exp\left(\lambda^2\|X\|^2_{\psi_2}\cdot\frac{X^2}{\|X\|^2_{\psi_2}}\right)\right]$.

How do I go from here?


1 Answer 1


Since $$ \int e^{\alpha x^2/2}\,e^{-x^2/2}\,\frac{\text dx}{\sqrt{2\pi}} =\int e^{-(1-\alpha)x^2/2}\,\frac{\text dx}{\sqrt{2\pi}} =\int e^{-(1-\alpha)x^2/2}\,\frac{\frac{1}{\sqrt{1-\alpha}}\,\text dx}{\sqrt{\frac{2\pi}{(1-\alpha)}}}=\{1-\alpha\}^{-1/2} $$ one can deduce that $$\|X\|_{\psi_2}^2=8/3$$ for a Gaussian variable. Computing it for a sub-Gaussian variable will depend on its density.


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.