9

For every nonnegative $(a_0,b_0,c_0)$, consider

$$a_{n+1}=\sqrt{a_n \frac{b_n+c_n}{2}},\quad b_{n+1}=\sqrt{b_n \frac{c_n+a_n}{2}},\quad c_{n+1}=\sqrt{c_n \frac{a_n+b_n}{2}}$$

$$M(a_0,b_0,c_0)=\lim_{n \to \infty} a_n=\lim_{n \to \infty} b_n=\lim_{n \to \infty} c_n$$

This simple three term iterated mean gives several interesting closed forms for particular cases (confirmed to $14$ digits so far):

$$M(1,1,2)=\frac{3^{3/4}}{\sqrt{\pi}}=1.28607413715749$$

$$M(1,1,\sqrt{2})=\frac{2}{\sqrt{\pi}}=1.12837916709551$$

$$M(1,1,\sqrt{3})=\frac{2^{3/4}}{\sqrt{\arccos(-1/3)}}=1.21670090936316$$

$$M(1,1,\sqrt{3}/2)=\frac{1}{\sqrt{\ln 3}}=0.95406458200000$$

These particular closed forms (if true) imply that the general closed form for $M(a_0,b_0,c_0)$ should also exist and involves the known elementary (or at least special) functions.

I have no idea how to find it so far and would appreciate the help.

It seems that for $M(1,1,x)$ the closed form should involve $\frac{1}{\sqrt{\arccos(f(x))}}$. Edit. Except in one case it has logarithm.


Since this mean can be considered yet another generalization of the Arithmetic Geometric Mean for three numbers, I link related questions: 1, 2.


To anyone interested, there is a companion sequence, which corresponds to the square roots of the $a_n,b_n,c_n$ of the titular sequence:

$$a_{n+1}=\frac{\sqrt{a_nb_n}+\sqrt{a_nc_n}}{2},\quad b_{n+1}=\frac{\sqrt{b_na_n}+\sqrt{b_nc_n}}{2},\quad c_{n+1}=\frac{\sqrt{c_na_n}+\sqrt{c_nb_n}}{2}$$

The limit of this sequence $M'(a_0,b_0,c_0)$ corresponds to the previous one:

$$M'(a_0,b_0,c_0)=(M(\sqrt{a_0},\sqrt{b_0},\sqrt{c_0}))^2$$

For example:

$$M'(1,1,2)=(M(1,1,\sqrt{2}))^2=\frac{4}{\pi}$$

Yuriy S
  • 32,728
  • 2
    But the question does ask about general initial conditions, not only $(1,1,x)$, right? // Sorry but why do you insist on giving the definition late in the question although it is necessary to the reader who wants to understand anything at all in the question? – Did Aug 25 '16 at 08:28
  • you should define $L(x) = M(1,1,x)$. Then it's pretty easy to check that $L(x) = \sqrt{\frac{1+x}2} L(\sqrt{\frac{2x}{1+x}})$ and $L(1) = 1$ – mercio Aug 25 '16 at 08:30
  • 1
    "The definition is already in the title" This is a quite misguided reason, posts should stand on their own, with no reference to their title. Let me suggest to revert to the version starting with the definition. – Did Aug 25 '16 at 08:32
  • With Schwab-Borchardt there is a distinction between $x>y$ and $x<y$. Your last case in oposit to the others: $\sqrt{3}/2<1$ (parameter 3 < parameter 1 = parameter 2). Therefore the logarithm. – user90369 Aug 25 '16 at 11:13

2 Answers2

6

Analogous to the Schwab-Borchardt mean:

\begin{align*} B(a,b) &= \frac{\sqrt{b^{2}-a^{2}}}{\cos^{-1} \frac{a}{b}} \\ &= B\left( \frac{a+b}{2}, \sqrt{\frac{(a+b)b}{2}} \right) \end{align*}

which can also be obtained by the iteration: $$ \left \{ \begin{array}{rcl} a_{n+1} &=& \frac{a_{n}+b_{n}}{2} \\ b_{n+1} &=& \sqrt{a_{n+1} b_{n}} \end{array} \right.$$

Now your mean is

\begin{align*} M(x,x,y) &= D(x,y) \\ &= D\left( \sqrt{\frac{x(x+y)}{2}} , \sqrt{xy} \right) \\ &= \frac{\sqrt[4]{x^{2}(y^{2}-x^{2})}}{\sqrt{\cos^{-1} \frac{x}{y}}} \end{align*}

Ng Chung Tak
  • 19,693
1

let $L(x) = M(1,1,x)^2$. Using properties of $M$, we have that $L$ is the unique function on $\Bbb R_{>0}$ continuous at $1$ satisfying $L(1) = 1$ and the "recurrence relation" :

$L(x) = M(\frac{\sqrt{1+x}}2,\frac{\sqrt{1+x}}2,\sqrt x)^2 = \frac{1+x}4M(1,1,\sqrt\frac{2x}{1+x})^2 = \frac{1+x}4 L(\sqrt\frac{2x}{1+x})$

$L$ is continuous at $1$ because $M(1,1,x)$ is between $1$ and $x$.
Such a function is unique because if there was another one $f$, then $f$ doesn't vanish, and $L/f$ is continuous at $1$ and satisfies $(L/f) (x) = (L/f) (\sqrt\frac{2x}{1+x})$. Since iterating $x \mapsto \sqrt\frac{2x}{1+x}$ on any positive real gives a sequence that converges to $1$, we must have $(L/f) (x) = \lim_{y \to 1} (L/f) (y) = 1/1 = 1$.


Let $f$ be the function defined by $f(x) = \frac{\sqrt{x^2-1}}{\arccos \frac 1x}$ if $x > 1$ and $f(1) = 1$ (it should also actually make sense for $0<x<1$ if you take the right pair of complex square roots and arccos)

Some trigonometric identities and some algebra shows that $f$ satisfies the recurrence relation of $L$ .

so we only need to show that $f$ is continuous at $1$ to prove that $L=f$.

Since the map $h \mapsto \cos(\sqrt h)$ is differentiable on the right at $0$ (with derivative $-1/2$), its inverse map $y \mapsto (\arccos y)^2$ is differentiable on the left at $1$ with derivative $-2$.

Now, for $x>1$, $f(x)^2 = \frac {(x-1)(x+1)}{(\arccos \frac 1x)^2} = - \frac {\frac 1x-1}{(\arccos \frac 1x)^2}x(x+1)$ and by definition of the derivative, the big fraction converges to $-1/2$ when $x \to 1$, so the whole thing converges to $-(-1/2)\cdot 1 \cdot 2 = 1$. Therefore $f^2$ is continuous at $1$, and then so is $f$.

This proves that $L(x) = f(x)$ for $x \ge 1$.

mercio
  • 51,119
  • Thank you. But how did you find the function initially. Was it a guess? And what about the general case of $M(a,b,c)$? – Yuriy S Aug 25 '16 at 10:20
  • looking at $\cos(2\sqrt{n-1}/L(\sqrt n))$ for $n=2,3,4,5,6$ etc pretty much gives it away – mercio Aug 25 '16 at 10:23
  • the general case looks pretty intractable, especially if you can't give any example where you recognize a conjectured closed form. – mercio Aug 25 '16 at 10:25
  • That's what I was afraid of. I'm searching for some cases, but no luck so far – Yuriy S Aug 25 '16 at 10:27