2

I need to understand very good how the properties of this formula

$\frac{4}{\pi} = \frac{5}{4} + \sum_{N \geq 1} \left[ 2^{-12N + 1} \times(42N + 5)\times {\binom {2N-1} {N}}^3 \right] $

Taken from the paper "Radian Reduction for Trigonometric Function" (Hanek Payne Algorithm)

Some remarkable properties are stated, specifically these four ones

  1. The $k^{th}$ term of the formula is exactly representable in $6k$ bits;
  2. The first $n$ terms of the sum can be represented exactly in $12n$ bits;
  3. The most significant bit of the $k^{th}$ term has weight at most $2^{1-6k}$ and hence each successive term increases the number of valid bits in the sum by at least $6$;
  4. If $12k < m + 1 \leq 12(k+1)$, then the $m^{th}$ bit of $\frac{4}{\pi}$ may be computed using only terms beyond the $k^{th}$.

My questions are: 1. How to prove the formula? 2. How to prove the properties stated above?

PS. I guess with terms the paper means the generic term $a_N$ of the sum...

user8469759
  • 5,285
  • I missed the $\binom{2n-1}{n}^{\color{red}{3}}$ before. – Jack D'Aurizio Jun 23 '15 at 14:50
  • This formula is listed here :https://sites.google.com/site/tpiezas/0013 It is related to the fact that $e^{\pi \sqrt{7}} \approx 2^{12}-24$ Tito Piezas III frequents this site so he should see this soon. – Noam Shalev - nospoon Jun 23 '15 at 14:56
  • Dear @Lukkio. I see that, although you have already asked 7 question in this site and received answers in 6 of them, you have not mark a best answer in any of them. You can do it so by clicking on the checkmark next to the answer that you think is the one that helped you the most. Please read here for more detail. – Leo Sera Jul 22 '15 at 23:02

2 Answers2

3

The formula in question can be rewritten as $$\frac {1}{\pi}=\sum_{j=0}^{\infty} \binom{2j}{j}^3\frac{42j+5}{2^{12j+4}}\tag{1}$$ It was first given by Ramanujan in his famous paper Modular equations and approximations to $\pi$ in 1914.

Luckily Ramanujan gave a proof of the general theory of such series and left the calculations for reader (except for the simplest of the series).

I have presented the details of Ramanujan's theory alongwith proofs in my blog posts (see part 1, part 2, part 3). Here I will present a brief picture and the details of calculation related to the series in question.

Let $k, l$ (called elliptic moduli) be numbers in interval $(0,1)$ and let $K, L, K', L'$ denote the complete elliptic integrals of first kind $$K(k) =\int_{0}^{\pi}\frac{dx}{\sqrt {1-k^2\sin^2x}},\\ K=K(k), L=K(l),k'=\sqrt {1-k^2},l'=\sqrt{1-l^2},K'=K(k'),L'=K(l')\tag{2}$$ It can proved that the as $K'/K$ is a strictly decreasing function of $k$ and maps $(0,1)$ to $(0,\infty) $. Thus if $n$ is a positive integer then for a given $k\in(0,1)$ there is a unique $l\in(0,1)$ such that $L'/L=nK'/K$. Jacobi also proved that in this case $l$ is an algebraic function of $k$ ie there is a polynomial $P(x, y) $ with integer coefficients such that $P(k, l) =0$ (called modular equation of degree $n$) and further the ratio $K/L$ can also be expressed as an algebraic function of $l, k$.

Let $q=\exp(-\pi K'/K) $ denote the nome corresponding to elliptic modulus $k$ and then $q^n=\exp(-\pi L'/L) $ corresponds to $l$. Ramanujan considered the function $$P(q) =1-24\sum_{j=1}^{\infty} \frac{jq^{2j}}{1-q^{2j}}\tag{2}$$ which is related to Dedekind eta function $$\eta(q) =q^{1/24}\prod_{j=1}^{\infty} (1-q^j)\tag{3}$$ via $$P(q) = 12q\frac{d}{dq}\{\log\eta(q^2)\}\tag{4}$$ and proved that $$nP(q^n) - P(q) =\frac{4KL}{\pi^2}F_n(k,l)\tag{5}$$ where $F_n$ is an algebraic function depending on $n$. Ramanujan gave explicit formulas for $F_n$ for many values of $n$.

Another key identity related to $P(q) $ is $$nP(e^{-\pi\sqrt{n}}) +P(e^{-\pi/\sqrt{n}})=\frac{6\sqrt{n}}{\pi}\tag{6}$$ Using $(5),(6)$ Ramanujan established that $$P(e^{-\pi\sqrt{n}}) - \frac{3}{\pi\sqrt{n}}=\left(\frac{2K}{\pi}\right)^2A_n\tag{7}$$ where $A_n$ is an algebraic number and elliptic integral $K$ corresponds to nome $q=e^{-\pi\sqrt{n}}$.

Next Ramanujan considered the well known formula $$\left(\frac{2K}{\pi}\right) ^2=1+\sum_{j=1}^{\infty} \left(\frac{1\cdot 3\cdot 5\dots(2j-1)}{2\cdot 4\cdot 6\dots(2j)}\right) ^3(2kk')^{2j}=\sum_{j=0}^{\infty} \binom{2j}{j}^32^{-6j}(2kk')^{2j}\tag{8}$$ It is also well known that $$\eta(q^2)=2^{-1/3}\sqrt{\frac {2K}{\pi}}(kk')^{1/6}\tag{9}$$ Using $(8),(9)$ we get $$\eta^4(q^2)=\left(\frac{kk'}{4}\right)^{2/3}\sum_{j=0}^{\infty} \binom{2j}{j}^32^{-6j}(2kk')^{2j}$$ Logarithmic differentiation of the above equation with respect to $k$ gives $$P(q) =(1-2k^2)\sum_{j=0}^{\infty} (3j+1)\binom{2j}{j}^32^{-6j}(2kk')^{2j}\tag{10}$$ Using $(7),(8),(10)$ one can get a series for $1/\pi$.


Let us now do the calculations for $n=7$ and obtain the series in question. We start with the modular equation of degree $7$ given by $$(kl) ^{1/4}+(k'l')^{1/4}=1\tag{11}$$ If $k, l$ correspond to nomes $e^{-\pi\sqrt{7}},e^{-\pi/\sqrt {7}}$ then we have $l=k', k'=l$ and then from $(11)$ we get $kk'=1/16$ or $$k^2(1-k^2)=\frac{1}{256}$$ or $$k^2=\frac{1-\sqrt {1-(1/64)} }{2}=\frac{8-3\sqrt{7}}{16}$$ (we need the smaller root here as the larger root is the value of $k'^2$). Thus from equation $(10)$ we get $$P(e^{-\pi\sqrt {7}})=\frac{3\sqrt{7}}{8}\sum_{j=0}^{\infty}(3j+1)\binom{2j}{j}^32^{-12j}\tag{12}$$ Next Ramanujan derived that $$7P(q^7)-P(q)=\frac{12KL}{\pi}(1+kl+k'l')$$ Putting $q=e^{-\pi/\sqrt{7}}$ we get $$7P(e^{-\pi\sqrt{7}})-P(e^{-\pi/\sqrt {7}})=\frac{27\sqrt {7}}{8}\left(\frac{2K}{\pi}\right)^2\tag{13}$$ where $K$ corresponds to nome $e^{-\pi\sqrt{7}}$. Adding the above to equation $(6)$ (with $n=7$) we get $$P(e^{-\pi\sqrt{7}})=\frac{3}{\pi\sqrt{7}}+\frac{27\sqrt {7}}{112}\left(\frac{2K}{\pi}\right)^2$$ Comparing this equation with $(12)$ and using $(8)$ (with $2kk'=1/8$) we can see that $$\frac{3}{\pi\sqrt{7}}=\frac{3\sqrt {7}}{8}\sum_{j=0}^{\infty} \left(3j+1-\frac{9}{14}\right)\binom{2j}{j}^32^{-12j}$$ or $$\frac{1}{\pi}=\sum_{j=0}^{\infty} \binom{2j}{j}^3\frac{42j+5}{2^{12j+4}}$$

  • I've discovered this post today, (+1). I think this is the first case for that we need modular forms to derive the result without using Chowla-Selberg formula. I think that Ramanujan had a modular equation of degree 29 which used to derive his fastest series with full rigour. – User Oct 29 '23 at 07:19
  • @User: Ramanujan's work in the theory of theta/elliptic functions was fully rigorous (in the sense that he assumed that series/products converged reasonably well so that they could be differentiated or integrated). In particular the theory of series for $1/\pi$ was developed from scratch in India before he met Hardy. He gave reasonable detail in his paper as well but left the details of calculations related to modular equations. – Paramanand Singh Oct 29 '23 at 10:38
  • @User: the same is true of a lot of modern papers on such series which almost exclusively give no details of any calculation. Some authors are more forthcoming and give details of the software/code used for the calculation. In particular Borwein brothers and Chudnovsky brothers never gave any detail of their calculations. – Paramanand Singh Oct 29 '23 at 10:40
  • 1
    @User: the symbolic nature of results (for example equation related to $7P(q^7)-P(q)$) indicates to a reasonable degree that such results can't be deduced by any guesses or some empirical approach. I believe he also had formulas for $37P(q^{37})-P(q)$ and $29P(q^{29})-P(q)$ which were used for two of his fastest series, but he didn't mention them in the paper (perhaps due to their unweildy form). No one so far has discovered these formulas even after so many years. – Paramanand Singh Oct 29 '23 at 10:45
  • @User: lastly I want to point out that the series under discussion in this post does not need any modular forms or Chowla Selberg stuff for a proof. – Paramanand Singh Oct 29 '23 at 10:50
  • What I wanted to say is that the easiest series for the theory of s=1/2 can be proved with Legendre's known results using $k_{1}$,$k_{2}$,$k_{3}$ and $k_{4}$ (This last follows from $k_{1}$ using Landen's transformation). Also for the theory of s=1/6 for which there is a link with the theory of s=1/2. This implies at least 8 Ramanujan's series for $1/\pi$ some gave by Ramanujan others no. – User Oct 29 '23 at 11:15
  • I see Borwein's brothers and Chudnovsky's brothers approaches a bit opaque. And no fomula known for $163P(q^{163})-P(q)$ so seems that we are far from proving these formulas by hand. – User Oct 29 '23 at 11:20
1

We can prove the formula through the theory of elliptic integrals and modular forms.

That boils down to proving that:

$$ \sum_{n\geq 1}\binom{2n-1}{n}^3 x^n = -\frac{1}{8}+\frac{1}{2\pi^2}\left(K\left(\frac{1-\sqrt{1-64x}}{2}\right)\right)^2 $$ where $K$ is the complete elliptic integral of the first kind.

The remaining parts follow from the fact that $\binom{2n-1}{n}$ behaves like $\frac{4^n}{2\sqrt{\pi}} $ by Stirling's approximation.

Jack D'Aurizio
  • 353,855
  • How can i post an image here? I could post the image of the formula taken from the paper if i knew how to do it.. In your last formula, the lower number of the binomial coefficent is "n" instead of 1. – user8469759 Jun 23 '15 at 14:22
  • I suppose that the $1$ in the binomial is $n$. Cheers – Claude Leibovici Jun 23 '15 at 14:27
  • Fixed. @Lukkio: instead of posting images, use $\LaTeX$. A short guide is here. – Jack D'Aurizio Jun 23 '15 at 14:33
  • I've used latex... the formula is on my original post. You said the formula is wrong but it is taken from the paper i said. If you try to implement it you will see it is correct, i just wanted to show the formula taken from the paper. – user8469759 Jun 23 '15 at 14:45
  • What about the properties i listed instead? – user8469759 Jun 23 '15 at 15:04
  • @Lukkio: as written, they all depends on the size of the central binomial coefficient. – Jack D'Aurizio Jun 23 '15 at 15:11
  • I don't have a big understading of the properties of elliptic integral and "general Gosper formula", can you suggest me a good reference? (i just found that specific formula in the paper i mentioned). – user8469759 Jun 24 '15 at 09:04