2
$\begingroup$

I was looking at the graph of $$\prod_{n=1}^\infty\frac{\left(\Gamma(n+1)\right)^2}{\Gamma\left(n+x+1\right)\Gamma\left(n-x+1\right)}=\prod_{n=1}^\infty\prod_{k=1}^\infty\left(1-\frac{x^2}{\left(n+k\right)^2}\right)$$

I noticed that it looks much like a normal curve of height $1$.

So is this equal to the form $e^{\frac{-x^2}{v}}$ for some $v$?

Or essentially the same question: does $$\sum_{n=1}^\infty\sum_{k=1}^\infty\ln\left(1-\frac{x}{\left(n+k\right)^2}\right)=-\frac{x}{v}$$ for some $v$?

Edit Related Question on the partial product.

$\endgroup$
  • $\begingroup$ Is the upper limit of $k$ correct? $\endgroup$ – M. Vinay Apr 11 at 2:02
  • 1
    $\begingroup$ Yes. I'm using the known result, $\frac{\left(\Gamma(n+1)\right)^2}{\Gamma\left(n+xi+1\right)\Gamma\left(n-xi+1\right)}=\prod_{k=1}^\infty\left(1+\frac{x^2}{\left(n+k\right)^2}\right)$ $\endgroup$ – tyobrien Apr 11 at 2:12
  • 1
    $\begingroup$ Major update ! Have a look at my last edit. Cheers :-) $\endgroup$ – Claude Leibovici Apr 12 at 3:28
  • $\begingroup$ It is possible to justify $(1)$ and $(2)$ from $(3)$. If you are concerned, I could elaborate. $\endgroup$ – Claude Leibovici Apr 12 at 8:06
  • $\begingroup$ Please, yes. I would be most interest. @ClaudeLeibovici $\endgroup$ – tyobrien Apr 12 at 11:21
4
$\begingroup$

A problem is that the sum diverges.

$\begin{array}\\ s(x) &=-\sum_{n=1}^\infty\sum_{k=1}^\infty\ln\left(1-\frac{x}{\left(n+k\right)^2}\right)\\ &\ge\sum_{n=1}^\infty\sum_{k=1}^\infty \dfrac{x}{(n+k)^{2}} \quad\text{since }-\ln(1-z) \ge z \text{ if } z \ge 0\\ &=x\sum_{n=1}^\infty\sum_{k=n+1}^\infty \dfrac{1}{k^{2}}\\ &\ge x\sum_{n=1}^\infty\sum_{k=n+1}^\infty \dfrac{1}{k(k+1)}\\ &= x\sum_{n=1}^\infty\sum_{k=n+1}^\infty (\dfrac1{k}-\dfrac1{k+1})\\ &= x\sum_{n=1}^\infty\dfrac1{n+1}\\ \end{array} $

and this diverges.

$\endgroup$
5
$\begingroup$

The observation is interesting per se.

In fact, if we consider $$f_p(x)=\prod_{n=1}^p\frac{\left(\Gamma(n+1)\right)^2}{\Gamma\left(n+x+1\right)\,\Gamma\left(n-x+1\right)}$$ we can find (using a CAS) that $$f_p(x)=\frac{ G(p+2)^2 \, G(2-x)\, G(2+x) }{G(p+2-x)\, G(p+2+x)}\left(-\frac{\sin(\pi x)}{\pi x(x^2-1)}\frac{G(3-x)\, G(3+x) }{ G(2-x)\, G(2+x)}\right)^p$$ where appear the Barnes G-function.

For the simplest $p=1$, $$f_1(x)=-\frac{\sin (\pi x)}{\pi x \left(x^2-1\right)}$$ does not show such a shape but, increasing $p$ more and more, we can effectively notice what you observed.

The problem, as @marty cohen answered, is that there is no limit to this function.

What is numerically interesting is the computation of $x$ such that $f_p(x)=\frac 12$. Here are some results ( the table has been updated after the edit). $$\left( \begin{array}{ccccc}\ p & \text{exact} & \text{using } (1) &\text{using } (2) &\text{using } (3)\\ 10 & 0.543223 & 0.486469 & 0.567259 & 0.545324 \\ 20 & 0.481383 & 0.438933 & 0.497577 & 0.482529 \\ 30 & 0.452924 & 0.416538 & 0.466221 & 0.453767 \\ 40 & 0.435333 & 0.402499 & 0.447040 & 0.436024 \\ 50 & 0.422935 & 0.392505 & 0.433608 & 0.423532 \\ 60 & 0.413516 & 0.384854 & 0.423448 & 0.414049 \\ 70 & 0.406002 & 0.378714 & 0.415372 & 0.406488 \\ 80 & 0.399801 & 0.373622 & 0.408723 & 0.400251 \\ 90 & 0.394552 & 0.369293 & 0.403108 & 0.394973 \\ 100 & 0.390022 & 0.365543 & 0.398270 & 0.390419 \\ 200 & 0.363578 & 0.343397 & 0.370187 & 0.363856 \\ 300 & 0.350356 & 0.332155 & 0.356240 & 0.350588 \\ 400 & 0.341793 & 0.324812 & 0.347238 & 0.341998 \\ 500 & 0.335563 & 0.319439 & 0.340703 & 0.335749 \\ 600 & 0.330715 & 0.315240 & 0.335627 & 0.330888 \\ 700 & 0.326775 & 0.311816 & 0.331507 & 0.326938 \\ 800 & 0.323472 & 0.308938 & 0.328057 & 0.323627 \\ 900 & 0.320640 & 0.306464 & 0.325101 & 0.320788 \\ 1000 & 0.318169 & 0.304301 & 0.322524 & 0.318311 \end{array} \right)$$

Edit

Taking logarithms of the product and using Stirling approximation, a very crude approximation would be $$\color{blue}{f_p(x) \sim \exp\left({-x^2 H_p}\right)}\tag 1$$ A better one would be $$\color{blue}{f_p(x) \sim \exp\left({-x^2 \left(H_p-\frac{1}{2}H_p^{(2)}\right)}\right)}\tag 2$$ Just computing the second derivative at $x=0$ would give as a much much better approximation (this is the exact second order Taylor expansion of $f_p(x)$ buil at $x=0$) $$\color{red}{f_p(x) \sim \exp\left(-x^2 \left(\psi ^{(0)}(p+2)+(p+1)\, \psi ^{(1)}(p+2)+\gamma-\frac{\pi ^2}{6} \right)\right)}\tag 3$$

Considering the asymptotics of the constant $k_{i,p}$ which appears in equation $(i)$, we can see that they are very closely related to each other $$k_{1,p}=\log(p)+\gamma +\frac{1}{2 p}+O\left(\frac{1}{p^2}\right)$$

$$k_{2,p}= \log(p)+\left(\gamma-\frac{\pi ^2}{12}\right) +\frac{1}{p}+O\left(\frac{1}{p^2}\right) $$

$$k_{3,p}=\log(p)+\left(\gamma-\frac{\pi ^2}{6} +1\right)+\frac{1}{p}+O\left(\frac{1}{p^2}\right) $$

Moreover, computing, we can notice that $k_{3,p}\approx k_{1,p}-\frac 23$.

The advantage of these appromations is that, if we need to solve for $x$, $f_p(x)=a$, we have a very good estimate to start Newton method. For illustration, considering $p=100$ and $a=\frac 14$, we would get the following iterates $$\left( \begin{array}{cc} n & x_n \\ 0 & 0.55213548373373596669 \\ 1 & 0.55099521268758128733 \\ 2 & 0.55099729839656297058 \\ 3 & 0.55099729840353755347 \end{array} \right)$$ which is the solution for twenty significant figures.

$\endgroup$
  • $\begingroup$ Thanks a lot! This is what I was interested to see. $\endgroup$ – tyobrien Apr 11 at 13:15
  • $\begingroup$ @tyobrien. You are welcome ! I had fun with the problem. Cheers :-) $\endgroup$ – Claude Leibovici Apr 11 at 13:17
3
$\begingroup$

I prefer to add another answer for an extension of the work.

Considering $$f_p(x)=\prod_{n=1}^p\frac{\left(\Gamma(n+1)\right)^2}{\Gamma\left(n+x+1\right)\,\Gamma\left(n-x+1\right)}$$

$$\log\left(f_p(x) \right)=\sum_{n=1}^p \log\left(\frac{\left(\Gamma(n+1)\right)^2}{\Gamma\left(n+x+1\right)\,\Gamma\left(n-x+1\right)} \right)$$

Using Taylor expansion $$\log\left(\frac{\left(\Gamma(n+1)\right)^2}{\Gamma\left(n+x+1\right)\,\Gamma\left(n-x+1\right)}\right)=-2\sum_{k=1}^\infty \frac{\psi ^{(2 k-1)}(n+1)}{(2 k)!}x^{2k}$$ and what remain is to compute the sums over $n$.

Writing $$\log\left(f_p(x) \right)=c_1 x^2+c_2 x^4+c_3 x^6 + c_4 x^8+\cdots$$ that is to say $$f_p(x)=\exp(c_1 x^2+c_2 x^4+c_3 x^6 + c_4 x^8+\cdots)$$ we should have $$c_1=-\psi ^{(0)}(p+2)-(p+1) \psi ^{(1)}(p+2)-\gamma+\frac{\pi^2}6$$ $$c_2=\frac{-45 \psi ^{(2)}(p+2)-15 (p+1) \psi ^{(3)}(p+2)-90 \zeta (3)+\pi ^4}{180} $$ $$c_3=\frac{-315 \psi ^{(4)}(p+2)-63 (p+1) \psi ^{(5)}(p+2)+8 \left(\pi ^6-945 \zeta (5)\right)}{22680}$$ $$c_4=\frac{-105 \psi ^{(6)}(p+2)-15 (p+1) \psi ^{(7)}(p+2)+8 \left(\pi ^8-9450 \zeta (7)\right)}{302400}$$ For any value of $p$, all coefficients are negative and they are smaller and smaller; this justifies the approximation given by $(3)$ in the previous answer.

For infinitely large values of $p$, we have $$\frac {c_2}{c_1} \sim \frac{90 \zeta (3)-\pi^4}{30 \left(6 \log (p)-\pi ^2+6 \gamma +6\right)}$$ $$\frac {c_3}{c_2} \sim \frac{4 \left(\pi ^6-945 \zeta (5)\right)}{63 \left(\pi ^4-90 \zeta (3)\right)}\approx 0.109046$$ $$\frac {c_4}{c_3} \sim \frac{3 \left(\pi ^8-9450 \zeta (7)\right)}{40 \left(\pi ^6-945 \zeta (5)\right)}\approx 0.163594$$

Warning

Better material in my answer to this question.

$\endgroup$

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.