2
$\begingroup$

I have this inequality $$\frac{1}{a}\exp\bigl\{-\frac{4}{h^2}\bigr\} \geq \frac{1}{f}$$ where $$ a \leq \Bigl(\pi^{d/2}\Gamma(\frac{1}{2}d+1)^{-1} + 1\Bigr) \left(\frac{h^{d+1}}{2} \Gamma \left(\frac{d+1}{2} \right) + h^d \left(\frac{d}{2}\right)^\frac{d}{2} \exp\Bigl\{-\frac{d}{2}\Bigr\} \right) $$ I tried to lowerbound $h$ in the inequality using the properties of the Gamma function and Stirling's approximation but I still see it complicated to be used in order to bound $h$.

Here is my work

Again \begin{align} a & \leq \Bigl(\pi^{d/2}\Gamma(\frac{1}{2}d+1)^{-1} + 1\Bigr) \left(\frac{h^{d+1}}{2} \Gamma \left(\frac{d+1}{2} \right) + h^d \left(\frac{d}{2}\right)^\frac{d}{2} \exp\Bigl\{-\frac{d}{2}\Bigr\} \right) \nonumber \\ \text{let $\frac{d}{2}=n$, then} \nonumber \\ a & \leq \Bigl(\pi^{n}\Gamma(n+1)^{-1} + 1\Bigr)\left( \frac{h^{2n+1}}{2} \Gamma \left( n+\frac{1}{2} \right) + h^{2n} n^n \exp\Bigl\{-n\Bigr\} \right) \nonumber \end{align} If $$\Gamma(n+1) = n\Gamma(n)=n(n-1)!=n! $$ and $$\Gamma(n+\frac{1}{2})=\frac{(2n)!}{4^n n!}\sqrt(\pi)=\frac{(2n-1)!!}{2^n}\sqrt(\pi)=\binom {n-\frac{1}{2}}{n}n! \sqrt(\pi) $$ then

\begin{align} a & \leq \Bigl(\frac{\pi^{n}}{n!} + 1\Bigr)\left( \frac{h^{2n+1}}{2} \frac{(2n)!}{4^n n!}\sqrt(\pi) + h^{2n} n^n \exp\Bigl\{-n\Bigr\} \right)\nonumber \\ & \leq h^{2n+2} \Bigl( \Bigl(\frac{\pi^{n}}{n!} + 1\Bigr) \left( \frac{(2n)!}{2^{2n }n!}\sqrt(\pi) + n^n \exp\Bigl\{-n\Bigr\} \right) \Bigr) \nonumber \end{align} Using stirling's approximation \begin{align} a & \leq h^{2n+2} \Bigl( \Bigl( \frac{1}{\sqrt{2}} \frac{\pi^{n-(1/2)}e^n}{n^{n+(1/2)}} + 1\Bigr)\left(\frac{1}{\sqrt{2}} \frac{e^n}{n^{n+(1/2)}} + \left(\frac{n}{e}\right)^n\right) \Bigr) \nonumber \\ & \leq h^{2n+2}\Bigl( \Bigl( \pi^{n-\frac{1}{2}} \left( \frac{e}{n}\right)^n +1 \Bigr) \Bigl(\left( \frac{e}{n}\right)^n + \left( \frac{n}{e}\right)^n \Bigr) \Bigr) \nonumber \\ & = h^{2n+2}\Bigl( \pi^{n-\frac{1}{2}} \Bigl( \left( \frac{e}{n}\right)^{2n} +1 \Bigr) + \left( \frac{e}{n}\right)^n + \left( \frac{n}{e}\right)^n \Bigr) \nonumber \end{align} What I can do more to simplify $a$ in order to use it to bound $h$ or is there a better way to bound $h$.

$\endgroup$
3
$\begingroup$

I understand from the reference to Stirling that you are looking for a large-$d$ approximation of

$$a_{\rm max}= \Bigl(\pi^{d/2}\Gamma(\frac{1}{2}d+1)^{-1} + 1\Bigr) \left(\frac{h^{d+1}}{2} \Gamma \left(\frac{d+1}{2} \right) + h^d \left(\frac{d}{2}\right)^\frac{d}{2} \exp\left(-\frac{d}{2}\right)\right).$$ With some algebra I arrived at

$$a_{\rm max}\approx\left(h^{d+1}\sqrt{\frac{\pi}{2} } +h^d\right) \exp\left[\frac{d}{2} \left(\ln \left(\frac{d}{2}\right)-1\right)\right]\equiv a_{\rm approx},\;\;d\gg 1.$$

This is already quite accurate for moderately small $d$, here is a plot for $d=10$. Shown are $a_{\rm max}/f$ (blue), $a_{\rm approx}/f$ (green), and $\exp(-4/h^2)$ (orange) as a function of $h$ for $f=10^5$.

The condition $\exp(-4/h^2)\geq a_{\rm max}/f$ is reached in an interval $(h_{\rm min},h_{\rm max})$ around $h_0$ given by

$$4/h_0^2=\ln f-\frac{d}{2} \left(\ln \left(\frac{d}{2}\right)-1\right)$$

This interval only exists if $f$ is large enough, you need $$f\gtrsim\exp\left[\frac{d}{2} \left(\ln \left(\frac{d}{2}\right)-1\right)\right]\equiv f_{\rm min}.$$


Irrespective of these large-$d$ approximations, the inequality in the OP is always violated for large enough $h$, because $a_{\rm max}$ grows as $h^{d+1}$ for large $h$, while $\exp(-4/h^2)$ tends to unity. So $h_{\rm max}<\infty$. Moreover, $h_{\rm min}>0$ because $a_{\rm max}$ vanishes as $h^d$ for small $h$, while $\exp(-4/h^2)$ vanishes more rapidly.

$\endgroup$
  • $\begingroup$ Thanks for your help and cooperation. Concerning $d$ it is not necessarily too large but I can say greater than $10$. Maybe I wan't clear enough but I am trying to show that (using a lowerbound) $h$ is greater than a positive number! but according to your work it's still negative bound $\endgroup$ – Noah16 Dec 3 '18 at 14:27
  • $\begingroup$ why negative bound? $h_0$ is a real positive number if $f$ is large enough and the interval in $h$ in which your inequality holds is a narrow interval centered on $h_0$; so you can conclude that $h$ is near $h_0>0$ provided $f>f_{\rm min}$. If $f<f_{\rm min}$ you cannot satisfy your inequality. $\endgroup$ – Carlo Beenakker Dec 3 '18 at 14:45
  • $\begingroup$ ah ya I understand you. could you please add some of the calculations to show how you got $a_{approx}$ above...THNAKS $\endgroup$ – Noah16 Dec 3 '18 at 15:10
  • $\begingroup$ Thanks for your cooperation and sorry for this late response. It seems that there is something wrong in the calculations. I think you just expanded $$ \left(\frac{h^{d+1}}{2} \Gamma \left(\frac{d+1}{2} \right) + h^d \left(\frac{d}{2}\right)^\frac{d}{2} \exp\left(-\frac{d}{2}\right)\right)$$ and neglect this amount $$ \Bigl(\pi^{d/2}\Gamma(\frac{1}{2}d+1)^{-1} + 1\Bigr) $$ which is much more than one $\endgroup$ – Noah16 Dec 19 '18 at 17:51
  • $\begingroup$ unfortunately, if you considered the term you neglect then the bound is negative and we come back to the same problem ! $\endgroup$ – Noah16 Dec 19 '18 at 18:20

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.