2
$\begingroup$

In a formula in my self-study of a summation method based on the matrix of Eulerian numbers (which I thus call "Eulerian summation") I am considering terms like $$ \Big({n \over e}\Big)^n \cdot {1 \over n!} \cdot {1 \over n+1} \tag 1 $$ For the estimate of the growth there is the Stirling-approximation-formula for the factorial: $$ m! \approx \sqrt{2 \pi m} \Big({m \over e}\Big)^m \tag 2$$

so that I arrive at the estimate

$$ \Big({n \over e}\Big)^n {1 \over n!} \approx {n! \over \sqrt{2 \pi n} }\cdot {1 \over n!} = {1 \over \sqrt{2 \pi n} } \tag 3 $$
(where I omit here and in the following the $ {1 \over n+1}$-term)

But by inspection of some results with increasing n I found, that a much better approximation than(2) is $$ \Big({n \over e}\Big)^n \approx \Gamma(1+n-1/2) \cdot {1 \over \sqrt{2 \pi }} \tag 4$$ From this, and from the coincidence that $\sqrt{\pi} = \Gamma(1/2)=\Gamma(1-1/2)$ one might then write $$ \Big({n \over e}\Big)^n {1 \over n!} \approx {\Gamma(1+n-1/2) \over \Gamma(1-1/2) \cdot \Gamma(1+n) }\cdot {1 \over \sqrt{2}}\\ = \binom{n-1/2}{n}\cdot {1 \over \sqrt{2} } \tag 5 $$
which is then a simple (generalized) binomial-expression.
By comparision , the series $$ s = \sum_{n=0}^\infty \Big({n \over e}\Big)^n {1 \over n!} \cdot {1 \over n+1} \tag 6$$ should then be a convergent expression and nicely approximated by. $$ t = {1\over \sqrt 2}\sum_{n=0}^\infty \binom{n-1/2}{n} \cdot {1 \over n+1} \sim s\tag 7$$

I thought, after that improvement the formula (4) might be even more improvable, so I looked first at Peter Luschny's "factorial" site where I find a related expression at the paragraph "A simple expression(recommended)" with the "LuschnyCF4" where however $m-1/2$ is replaced by $n$ and $n+1/2$ by $N$ and the approximation reads then as $$ n! \approx \sqrt{2 \pi } \Big({P \over e}\Big)^N \\ \text{where } N=n+1/2 \tag 8$$ Here $P$ is a complicated scaling of $N$ so I cannot simply adapt that formula for an improved approximation. So my question is:

Q: How can I improve my (heuristical) approximation in the rhs in (4) (while keeping it focused at positive integers $n$ in the term $(n/e)^n$)?

(P.s.: I did not have a good idea for tagging, perhaps someone could improve that)

$\endgroup$
  • $\begingroup$ I'm a little confused on what you are asking here. Are you looking for a proof of $(4)$? What do you mean by "while keeping it focused at positive integers $n$"? $\endgroup$ – EuYu Jul 4 '14 at 12:17
  • $\begingroup$ Isn't simpler to just plugin the next terms of the Stirling approximation? $\endgroup$ – leonbloy Jul 4 '14 at 12:23
  • $\begingroup$ @EuYu: hmm, in the Stirling-formula the $n$ is also under the square-root; in (4) it has moved mystically into the gamma-expression, so this is a variant of the Stirling-formula. Maybe I've not well expressed what I wanted to say. My problem is that I have the lhs in (4) and at integer n. Then I found that heuristically I can express that by Gamma( f(n)) where f(n) is some function of n (in this case f(n)=n+1/2) . After this ansatz, I can't see how to get an even better approximation by a variation of the expression on the rhs. Hope I could make it clearer? $\endgroup$ – Gottfried Helms Jul 4 '14 at 12:32
  • $\begingroup$ @Leonbloy: please see my comment to EuYu $\endgroup$ – Gottfried Helms Jul 4 '14 at 12:34
  • 1
    $\begingroup$ @GottfriedHelms I'm not sure if this helps, but Stirling's approximation doesn't just hold for integer $n$. It's in fact the first term of the asymptotic approximation for $\Gamma(n)$. If you substitute $n+\frac{1}{2}$ into Stirling's formula, you should be able to reduce it (asymptotically) to $(4)$. The fact that the $n$ has moved out of the square root is effectively because $\Gamma(n+\frac{1}{2})$ has introduced another $\sqrt{n}$ to the approximation. $\endgroup$ – EuYu Jul 4 '14 at 12:43
3
$\begingroup$

The asymptotic expansion for $\Gamma(z)$ is given by $$\Gamma(z) \sim \sqrt{\frac{2\pi}{z}}\left(\frac{z}{e}\right)^z\left(1 + \frac{1}{12z} + \cdots\right)$$ Therefore we have $$\begin{align}\Gamma\left(n+\frac{1}{2}\right) &\sim \sqrt{\frac{2\pi}{\left(n+\frac{1}{2}\right)}}\left(\frac{n+\frac{1}{2}}{e}\right)^{n+\frac{1}{2}} \\&= \sqrt{2\pi}\frac{\left(n+\frac{1}{2}\right)^n}{e^{n+\frac{1}{2}}} \\&= \sqrt{2\pi}\left(\frac{n}{e}\right)^n\frac{\left(1+\frac{1}{2n}\right)^n}{\sqrt{e}}\end{align}$$

Now since we have $$\lim_{n\rightarrow \infty}\left(1 + \frac{x}{n}\right)^n = e^x$$ It follows that $$\left(1+\frac{1}{2n}\right)^n \sim \sqrt{e}$$ And therefore we recover your approximation $(4)$ with the above substitution: $$\Gamma\left(n+\frac{1}{2}\right) \sim \sqrt{2\pi}\left(\frac{n}{e}\right)^n\frac{\left(1+\frac{1}{2n}\right)^n}{\sqrt{e}} \sim \sqrt{2\pi}\left(\frac{n}{e}\right)^n$$ Of course the asymptotics can be kept more accurate without making the latter approximation, but that comes at the expensive of a more complex expression. Also, note that you can obtain increasingly accurate asymptotics by taking more and more terms from the asymptotic series.

$\endgroup$

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.