1
$\begingroup$

Define $\Psi_{X}(\lambda) = \log E e^{\lambda X}$, and suppose that $EX=0$. We say that the random variable $X$ is sub-Gaussian with variance factor $v$ if: $$\Psi_{X}(\lambda) \leq v\lambda^{2}/2$$ for all $\lambda \in \mathbb{R}$. In other words, the moment generating function of $X$ is "dominated" by the moment generating function of a mean zero Gaussian random variable with variance $v$.

As far as I know, this indicates that sub-gaussian random variables with variance factor $v$ have tails that are "thinner" than the tails of a Gaussian random variable with variance $v$.

Sub-Gaussian random variables are particularly important in the study of concentration inequalities.

Now I have found cases (for example, equation (1.1) in the supplementary material of the paper Spokoiny (2012)) where concentration inequalities are stated for the case when: $$\Psi_{X}(\lambda) \leq v\lambda^{2}/2$$ for all $\lambda$ satisfying $|\lambda|\leq c$ for some $c$.

I am wondering if anyone can offer an interpretation of this? In particular, what can we say about the tails of a random variable that satisfy the sub-Gaussian inequality only for values of $\lambda$ satisfying $|\lambda|\leq c$?

$\endgroup$
2
$\begingroup$

It means that the random variable is sub-exponential, i.e., has heavier tails than the sub-Gaussian distribution.

A random variable is $K$-sub-exponential if $ P(|X| \geq t) \leq 2 \exp(-K/t) $. One of its equivalent definition is that $\Psi_X(t) \leq C^2 \lambda^2$ for all $|\lambda| \leq \frac{1}{C}$. See Proposition 2.7.1 in HDP[1]. You can consult Section 2.7 titled 'Sub-Exponential Distributions' in the HDP[1] to know more.

[1]: High-Dimensional Probability, Roman Vershynin. https://www.math.uci.edu/~rvershyn/papers/HDP-book/HDP-book.html

$\endgroup$
  • $\begingroup$ Thanks for the answer and the reference, this is quite helpful. $\endgroup$ – möbius Feb 20 at 21:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.