1
$\begingroup$

I am reading through a set of notes about concentration of Gaussian measure, and on page 56, they make the following claim that I am failing to see the proof of:

Now we estimate the second summand at the right hand side. Using Young's inequality with $y = e^{\lambda \tilde Z} \geq 0$ and $x = \Sigma^2 - (e-1)E\Sigma^2$, we get \begin{align*}E(\Sigma^2e^{\lambda\tilde Z} &= (e-1)(E\Sigma^2)Ee^{\lambda \tilde Z}+E\left[(\Sigma^2 - (e-1)E\Sigma^2)e^{\lambda \tilde Z}\right]\\ & \leq (e-1)(E\Sigma^2)Ee^{\lambda \tilde Z} +\lambda E(\tilde Ze^{\lambda \tilde Z})-Ee^{\lambda \tilde Z} + E e^{\Sigma^2 - (e-1)E\Sigma^2}\end{align*}

This inequality is baffling to me. I'm assuming they mean Young's product inequality, but stated in its usual form, Young's product inequality does not allow us to put "$x$" in an exponent. The best I could think of was using the fact that $\require{enclose} \enclose{horizontalstrike}{\frac{x^p}{p} = \frac{e^{p\log x}}{p} = \frac{x e^p}{p}}$ so that if we chose $p = x$, we could get from Young's inequality: $$\require{enclose} \enclose{horizontalstrike}{x\cdot y \leq e^{x} +y^{\frac{x}{x-1}}\cdot \frac{x-1}{x}}$$ for $\require{enclose} \enclose{horizontalstrike}{x,y \geq 0}$, but even then, it seems to me that the second term becomes quite intractable. Where are the authors getting this inequality from?

Note: A lot of variables in the above quote are defined in the preceding parts of the proof, so I may have left out important context for this problem that can be found in the link.

$\endgroup$
1
$\begingroup$

They use a generalized version of Young's inequality (see here, the last example under "Generalization using Fenchel-Legendre transform"). In particular, given two real numbers $x > 0$, $y$, one gets:

$$xy \leq x \ln x - x +e^y.$$

This version of Young's inequality was already used in the document you linked, at the bottom of page 43.

Note by the way that there is a mistake in your computation ($e^{p \ln x} \neq x e^p$).

$\endgroup$

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.