30

Given the recurrence relation $$ a_{n+1}=\frac{1-\sqrt{1-a_n}}{1+\sqrt{1+a_n}} $$ which is easy to find $$ a_n\to0, \quad b_n=\frac{a_{n+1}}{a_n}\to\frac1{4} $$ hence $a_n\sim4^{-n}$, or with some regular conditions we may prove $$ \lim_{n\to\infty} 4^n a_n = C(a_1) $$ where $a_1$ regards to the first item to generate this sequence. Yet, with a refined insight, some suggest that if $a_1=1$, we will have $$ \lim_{n\to\infty} 4^n a_n = \frac1{8\pi}\Gamma\,^4\!\left(\frac1{4}\right) $$ where $\Gamma$ stands for the Gamma function.

This result is out of my reach. I have no approach to it. Thanks in advance for whoever gives help.

Clement C.
  • 68,437
Nanayajitzuki
  • 2,840
  • 9
  • 18
  • What is $C(a_{1})$ ? – acreativename May 28 '22 at 07:39
  • 2
    @acreativename here, $C(a_1)$ means some constant which can be determined by given the first item of this seqence, hence a unknown function of $a_1$, yet I don't know what exactly it is. – Nanayajitzuki May 28 '22 at 07:43
  • But you have already found the limit you are asking for as $a_n\to0$. So you already answered your own question? – emacs drives me nuts May 28 '22 at 08:15
  • @emacsdrivesmenuts No, the question is not the limit of $a_n$ (that's 0), it's the limit of $4^na_n$ (when $a_1=1$). Or, equivalently, finding the right constant $C>0$ such that $a_n \sim C\cdot 4^{-n}$ (asymptotic rate). (Numerical evidence corroborates the OP's conjecture, incidentally.) – Clement C. May 28 '22 at 08:17
  • (Incidentally, the use of $\sim$ in the question is not quite correct -- $\sim$ has a specific meaning, while in the OP it is used more informally) – Clement C. May 28 '22 at 08:18
  • @Clement C: Ok, then I was just confused by the title. – emacs drives me nuts May 28 '22 at 09:39
  • @emacsdrivesmenuts I agree, the title is (was)... suboptimal. – Clement C. May 28 '22 at 10:02
  • 1
    Can you provide more context? Where does the sequence come from? Where is it claimed that the limit has that value? – Martin R May 28 '22 at 11:26
  • 2
    "with a refined insight, some suggest...": you are being very mysterious! – TonyK May 28 '22 at 11:43
  • 1
    Could this be related to https://math.stackexchange.com/q/2489586/42969 ? – Martin R May 28 '22 at 11:51
  • 1
    @MartinR Not sure where this came from for the OP, but computing the first 10000 in Mathematica and looking up the value on WolframAlpha to get a match, this is what I get too. – Clement C. May 28 '22 at 11:53
  • @Martin R It's my bad I didn't write the question well enough. This question is found in an assignment pool from a local college course on asymptotic analysis. However questions in the pool was collected years ago by former students and professors, some with only an answer. Problem like this one remains unfinished for it's maybe too hard to do as an test or homework. – Nanayajitzuki May 28 '22 at 12:54
  • @MartinR Without a doubt, it's related. Typically , problems like this are solved by finding invariants, which are functions of $x,y$ such that $f(a_n)$ is independent of $n$. The trick for finding the limit involving $e^{\pi}$ is noting that a particular elliptic integral is the invariant. Here, it remains to be seen what the appropriate invariant is (and we can't rule out something like a gamma integral, given the conjecture). – Sarvesh Ravichandran Iyer May 28 '22 at 17:13
  • It seems that $\frac1{8\pi}\Gamma,^4!\left(\frac1{4}\right) = 2 K\left(\frac 1{\sqrt 2}\right)^2$ where $K(k)$ is the elliptic integral which is also used in https://math.stackexchange.com/a/2494934/42969. – Martin R May 28 '22 at 18:30

1 Answers1

32

Someone on social media (where I linked this post), Prof. Antoine Chambert-Loir, gave a solution. The outline below is from him; errors, if any, are mine.

The idea is to set $$a_n = \operatorname{sl}^2(u_n) \tag{1} $$ where $\operatorname{sl}$ is the lemniscate sine function.

Then the recurrence relation becomes $$ \operatorname{sl}^2(u_{n+1}) = \frac{1-\sqrt{1-\operatorname{sl}^2(u_n)}}{1+\sqrt{1+\operatorname{sl}^2(u_n)}} = \operatorname{sl}^2(\tfrac{1}{2}u_{n}) \tag{2} $$ using the bisection formulas and identities relating $\operatorname{sl}$ and $\operatorname{cl}$. This shows that $u_{n+1} = \frac{1}{2}u_n$, and therefore $$ u_n = \frac{2u_1}{2^{n}} \tag{3} $$ But $a_1=1$ implies $\operatorname{sl}(u_1) =1$, which then gives $$ u_1 = \int_0^1 \frac{dx}{\sqrt{1-x^4}} = \frac{\sqrt{\pi}\Gamma(5/4)}{\Gamma(3/4)} \tag{4} $$ which finally leads to $$ a_n = \operatorname{sl}^2(u_n) \operatorname*{\sim}_{n\to\infty} u_n^2 = \frac{4u_1^2}{4^n} = \frac{4\pi\Gamma(5/4)^2}{\Gamma(3/4)^2} \cdot 4^{-n} \tag{5} $$ where $\varpi^2=\frac{4\pi\Gamma(5/4)^2}{\Gamma(3/4)^2}\approx 6.875185818$ ($\varpi$ being the lemniscate constant).


The above, as per the OP's question, corresponds to the initial condition $a_1=1$. Using the same argument, one easily gets that, for $a_1=\alpha\in[0,1]$, the limiting constant for $4^n a_n$ is given by $$ C(\alpha) = 4\operatorname{arcsl}(\sqrt{\alpha})^2 = 4\left(\int_0^{\sqrt{\alpha}} \frac{dx}{\sqrt{1-x^4}}\right)^2 \tag{6} $$ where $\operatorname{arcsl}$ is the lemniscate arcsine function. The behavior of $C(\alpha)$ as $\alpha$ ranges from $0$ to $1$ is depicted below.

Behavior of the limiting constant as the initial condition ranges from 0 to 1.

Kroki
  • 13,619
Clement C.
  • 68,437
  • 2
    Impressive! Thank you so much. Lots of things to learn for me, really helpful. – Nanayajitzuki May 29 '22 at 01:49
  • Well, it's not by me -- Antoine Chambert-Loir is to credit for it (I just tried to learn a bit from, and understand what he wrote, and wrote it here). Lots to learn for me too... @Nanayajitzuki – Clement C. May 29 '22 at 01:51
  • 7
    This is what one calls a gem! (even if inspired from work of Antoine Chambert-Loir). The key idea is to figure that the iteration matches with the half angle formula for some elliptic function. – Paramanand Singh May 29 '22 at 02:06
  • 3
    Incidentally, based on this, for arbitrary initial condition $a_1 = \alpha \in[0,1]$, the resulting constant will be given by $$C(\alpha) = 4\operatorname{arcsl}(\sqrt{\alpha})^2 = 4\left(\int_0^{\sqrt{\alpha}} \frac{dx}{\sqrt{1-x^4}}\right)^2$$ – Clement C. May 29 '22 at 02:48
  • The more I think about it, the more I feel the necessity to write a good self-answered question explaining recurrences like this one, and the connection with integrals of rational functions (yes, I think this goes beyond elliptic integrals just that little more!) – Sarvesh Ravichandran Iyer May 29 '22 at 04:36
  • 1
    If $\text{sl} (\phi/2)=a,\text{ sl} (\phi) =b$ then the relation $b=2a\sqrt{1-a^4}/(1+a^4)$ is well known. That the above can be inverted to express $a$ in terms of $b$ is amazing. David Cox mentions in his Galois Theory that one needs to use an intermediate variable $c$ with $c^2=2ia^2/(1-a^4)$ and then $-2ic^2/(1-c^4)=b^2$ and the value $a$ can be obtained by solving two quadratic equations. That the solution turns out to be $a^2=\frac{1-\sqrt{1-b^2}}{1+\sqrt{1+b^2}}$ is really cool. – Paramanand Singh May 29 '22 at 04:58
  • 2
    @ClementC.: May I suggest that you add the general solution $C(\alpha)$ to your answer as well? Comments are ephemeral and may get deleted at any time. – Martin R May 29 '22 at 09:11
  • @MartinR Of course, good point. I've added it. – Clement C. May 29 '22 at 09:28
  • 1
    Just a wild guess. Without invoking Jacobi/lemniscate elliptic functions, the complete elliptic integral of the first kind $$ K(x) = \frac{\pi}{2}\sum_{n\geq 0}\left(\frac{1}{4^n}\binom{2n}{n}\right)^2 x^n=\int_{0}^{\pi/2}\frac{d\theta}{\sqrt{1-x\cos^2\theta}} $$ is related to Gauss' AGM mean through $$ K(x) = \frac{\pi/2}{\text{AGM}(\sqrt{1-x},1)}=\frac{\pi/2}{\text{AGM}(1-\sqrt{x},1+\sqrt{x})}=\frac{\pi/2}{(1+\sqrt{x})\text{AGM}(\frac{1-\sqrt{x}}{1+\sqrt{x}},1)} $$ – Jack D'Aurizio May 30 '22 at 22:51
  • leading to

    $$ K(x) = \frac{1}{1+\sqrt{x}}\cdot K\left(\frac{4\sqrt{x}}{(1+\sqrt{x})^2}\right) $$ aka Landen's transformation. Is it possible to directly relate the values of $4^n a_n$ to the values of $K\left(\frac{1+x}{2}\right)K\left(\frac{1-x}{2}\right)$ or something similar?

    – Jack D'Aurizio May 30 '22 at 22:52
  • @SarveshRavichandranIyer: I am (and perhaps other users too are) waiting for your post on this topic. – Paramanand Singh Jun 03 '22 at 05:52
  • 1
    @ParamanandSingh Will take some time, sorry about that. I'm really busy with other things. I do have a list of stuff prepared, though. – Sarvesh Ravichandran Iyer Jun 03 '22 at 06:32