0
$\begingroup$

In the notes Fundamentals of Stein’s method in page 7 one reads

enter image description here

here $\phi(x) = \int_{-\infty}^x \frac{1}{\sqrt{2\pi}}e^{-t^2/2}\, dt$

I would like to see first that $[1 - \phi(w)] \leq \frac{1}{2}e^{-w^2/2}$ for $w >0$ So since $\phi(0) = 1/2$ we have $1 - \phi(0) =\frac{1}{2}e^{-0^2/2} $, now taking derivatives

$$ \frac{d}{dw}[1 - \phi(w)] = - \frac{1}{\sqrt{2\pi}}e^{-w^2/2} $$ and

$$\frac{d}{dw}\frac{1}{2}e^{-w^2/2} = \frac{-w}{2}e^{-w^2/2}$$

So $$\frac{d}{dw}\big[\frac{1}{2}e^{-w^2/2}\big] -\frac{d}{dw}[1 - \phi(w)] = \bigg(\frac{1}{\sqrt{2\pi}} - \frac{w}{2}\bigg) e^{-w^2/2} $$

which is negative for $w > \frac{2}{\sqrt{2\pi}}$. In this case, it may be that $$\frac{1}{2}e^{-w^2/2} <[1 - \phi(w)] $$

How do I see that this is not the case?

Now, consider the next inequality implied by the display in the figure:

$$ [1 - \phi(w)] \leq \frac{1}{w\sqrt{2\pi}}e^{-w^2/2}$$

At $0$ the right hand side is $\infty$. checking the derivatives now we see

$$\frac{d}{dw}\big[\frac{1}{w\sqrt{2\pi}}e^{-w^2/2}\big] =-\frac{1}{\sqrt{2\pi}}e^{-w^2/2} - \frac{1}{w^2\sqrt{2\pi}}e^{-w^2/2} $$ and so

$$ \frac{d}{dw}\big[\frac{1}{w\sqrt{2\pi}}e^{-w^2/2}\big] -\frac{d}{dw}[1 - \phi(w)] = - \frac{1}{w^2\sqrt{2\pi}}e^{-w^2/2} $$

which is negative so It is also not clear that

$$ [1 - \phi(w)] \leq \frac{1}{w\sqrt{2\pi}}e^{-w^2/2}$$

Maybe the author meant

$$ [1 - \phi(w)] \leq \max\big\{\frac{1}{2},\frac{1}{w\sqrt{2\pi}}\big\}e^{-w^2/2}$$

Still it is not clear that the inequality will hold, since for $w \geq \frac{2}{\sqrt{2\pi}}$ the inequality we would like to hold is $$ [1 - \phi(w)] \leq \frac{1}{2}e^{-w^2/2} $$

which is from the derivative sign we see that $\frac{1}{2}e^{-w^2/2} - [1 - \phi(w)] $ is decreasing and we still need to check if this is always positive.

Any ideas?

Edit For the first bound and a simple argument using no derivatives see Normal distribution tail probability inequality

$\endgroup$
2
$\begingroup$

Note that for $x>0$:

$$1-\Phi(x)=\frac{1}{\sqrt{2\pi}}\int_x^{\infty}e^{-t^2/2}\,dt\leq \frac{1}{\sqrt{2\pi}}\int_x^{\infty}\frac{t}{x}e^{-t^2/2}\,dt=\frac{e^{-x^2/2}}{x\sqrt{2\pi}}$$

and that $1-\Phi(x)<\frac{1}{2}$ for every $x>0$, so the author meant what he wrote.

If one insists on using the derivative, then one may note that the inequality to prove is equivalent to this one: $$\Phi(x)+\frac{\Phi^{'}(x)}{x}\geq 1$$ Differentiating the function $g(x)=\Phi(x)+\frac{\Phi^{'}(x)}{x}$ one obtains: $$g'(x)=\frac{e^{-x^2/2}}{\sqrt{2\pi}}\left(1-\frac{x^2+1}{x^2}\right)$$ since $g'(x)<0$ for all $x>0$ and $\lim_{x\to\infty}g(x)=1$, one deduces, (by considering derivatives!) that $g(x)\geq 1$ for all $x>0$.

Evidently, the proof obtained by considering derivatives is somewhat longer and much less to the point than the concise inequality above.

$\endgroup$
  • $\begingroup$ Great! Yet the argument does't use derivatives $\endgroup$ – Conrado Costa Aug 16 '17 at 13:22

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.