Given any natural number $m\gt2$, let $z$,$k$ be complex numbers, where $i=\sqrt{-1}$ and consider the general continued fraction

$$\Theta(k,z,m)=\cfrac{(m+1)}{km+\cfrac{z(0m-1)(2m+1)} {3km+\cfrac{z(m-1)(3m+1)}{5km +\cfrac{z(2m-1)(4m+1)}{7km+\cfrac{z(3m-1)(5m+1)}{9km+\ddots}}}}}$$

The variable $\Theta$ has the property of satisfying the polynomial equation of $m$th degree


$k\neq\pm i$ , $k\neq 0$ ,$z\neq-k^2$,$z\neq 0$

Both continued fractions in this post are just the special cases $z =\pm 1$.

It also obeys a very fundamental transformation property


P.S.:After applying property $(1)$ and $(2)$,we arrive at the elegant solution


Q: How do we prove rigorously that the continued fraction does satisfy the polynomial equation?

  • 1
    $\begingroup$ Roots of polynomials of degree $5$ or higher do not necessarily have roots expressable in terms of radicals. $\endgroup$ – Adam Hughes May 13 '16 at 16:32
  • $\begingroup$ @AdamHughes: I've fixed Nicco's equation $(1)$. This particular equation does have a solution in radicals for integer $m>1$. Kindly see this post. $\endgroup$ – Tito Piezas III May 13 '16 at 18:16
  • $\begingroup$ @Nicco: I've revised your formula $(1)$. This special form is solvable in radicals for all integer $m>1$. (I've cited a link in my comment to Adam.) If you observe the two special cases $z=\pm1$, notice that the $LHS$ is just an expression in the unknown variable multiplied by a constant. $\endgroup$ – Tito Piezas III May 13 '16 at 18:18
  • $\begingroup$ @ Tito Piezas III : thanks, but the formula $$\frac{(i+\Theta\sqrt{z})^m}{(i-\Theta\sqrt{z})^m}=\frac{(ik+\sqrt{z})^{m+1}}{(ik-\sqrt{z})^{m+1}}$$ is solvable in radicals(I have tested it). I think it should be incorporated back into the question since it's simpler and also includes k whereas C in this other formula is not known $\endgroup$ – Nicco May 13 '16 at 19:16
  • 1
    $\begingroup$ @mathreadler in algebra finite iterations are requisite, if you allow limits then you're in the realm of completions which is more analysis. And when one says a polynomial is "solvable by radicals" they mean all its roots are in a finite radical extension, which means no infinite towers. $\endgroup$ – Adam Hughes May 15 '16 at 18:20

For simplicity, we will start with the case $z = 1$.
We will assume $k > 0$ and let $\cot\theta = k$ and $\mu_{\pm} = k \pm \sqrt{k^2+1}$.

The CF at hand has the form $$ \def\CF{\mathop{\LARGE\mathrm K}} \CF_{\ell=1}^{\infty} \frac{\alpha_\ell\gamma_{\ell-1}}{\beta_\ell} \quad\text{ where }\quad \gamma_0 = 1\quad\text { and }\quad \begin{cases} \alpha_\ell &= \ell m + 1\\ \beta_\ell &= k m (2\ell - 1)\\ \gamma_\ell &= \ell m - (m+1) \end{cases} \quad\text{ for }\quad\ell > 0 $$ Consider the associated linear recurrence relation:

$$\alpha_\ell f_{\ell-1} - \beta_\ell f_\ell = \gamma_\ell f_{\ell + 1}\quad\text{ for }\ell > 0\tag{*1}$$ Let $(u_\ell)_{\ell=0}^\infty$ and $(v_\ell)_{\ell=0}^\infty$ be the two solutions satisfying $(u_0,u_1) = (1,0)$ and $(v_0,v_1) = (0,1)$.
In terms of $u$ and $v$, it is easy to verify the convergents of CF satisfy $$L_n \stackrel{def}{=} \CF_{\ell=1}^n \frac{\alpha_\ell\gamma_{\ell-1}}{\beta_{\ell}} = -\frac{u_{n+1}}{v_{n+1}} \quad\implies\quad L_\infty \stackrel{def}{=} \CF_{\ell=1}^\infty \frac{\alpha_\ell\gamma_{\ell-1}}{\beta_{\ell}} = -\lim_{n\to\infty} \frac{u_n}{v_n}$$ Notice $\alpha_\ell,\gamma_\ell \ne 0$ for all $\ell$ and following limits exist $$ \alpha \stackrel{def}{=} \lim\limits_{\ell\to\infty}\frac{\alpha_\ell}{\ell m} = 1, \quad \beta \stackrel{def}{=} \lim\limits_{\ell\to\infty}\frac{\beta_\ell}{\ell m} = 2k \quad\text{ and }\quad \gamma \stackrel{def}{=} \lim\limits_{\ell\to\infty}\frac{\gamma_\ell}{\ell m} = 1 $$ Furthermore, the associated polynomial $$\alpha - \beta\lambda - \gamma\lambda^2 = 0 \quad\iff\quad \lambda^2 +2k\lambda - 1 = 0$$ have roots $\lambda_{\pm} = -k \pm \sqrt{k^2+1} = \frac{1}{\mu_{\pm}}$ whose modulus are distinct. By Poincare-Perron theorem${}^{\color{blue}{[1]}}$, $(*1)$ has a pair of fundamental solutions $(p^{+}_\ell)_{\ell=0}^\infty$ and $(p^{-}_\ell)_{\ell=0}^\infty$ satisfying $\lim\limits_{n\to\infty} \frac{p^{\pm}_{\ell+1}}{p^{\pm}_{\ell}} = \lambda_{\pm}$.

Expanding $u$, $v$ in terms of $p^{\pm}$ and notice $|\lambda_{-}| > 1 > |\lambda_{+}|$, we have

$$\begin{cases} u_\ell &= A p^{+}_\ell + B p^{-}_\ell\\ v_\ell &= C p^{+}_\ell + D p^{-}_\ell \end{cases} \quad\implies\quad L_\infty = -\lim_{n\to\infty} \frac{A p^{+}_\ell + B p^{-}_\ell}{C p^{+}_\ell + D p^{-}_\ell} = -\frac{B}{D}$$ Construct a solution $(w_\ell)_{\ell=0}^\infty$ of $(*1)$ by setting $w_\ell = u_\ell + L_\infty v_\ell$, we have

$$w_\ell = ( A p^{+}_\ell + B p^{-}_\ell ) - \frac{B}{D}( C p^{+}_\ell + D p^{-}_\ell ) = \frac{AD-BC}{D} p^{+}_\ell \quad\implies\quad \lim_{\ell\to\infty} \frac{w_{\ell+1}}{w_\ell} = \lambda_{+} $$ The corresponding OGF $\displaystyle\;\quad w(t) \stackrel{def}{=} \sum_{\ell=0}^\infty w_\ell t^\ell\quad$ will be analytic for $t < |\mu_{+}|$.
Let $a = \frac{1}{m}$, one can rescale $(*1)$ to the form: $$(\ell + a) w_{\ell-1} - k(2\ell-1) w_{\ell} + (a + 1 - \ell) w_{\ell+1} = 0\quad\text{ for }\quad \ell > 0\tag{*2}$$

Let us first treat $a$ as a complex parameter with $-\frac32 < \Re a < -1$.

Multiply the $\ell^{th}$ entry of $(*2)$ by $t^\ell$ and start to sum from $\ell = 1$. If we let $g(t) = w(t) - 1$, we find

$$\begin{align} & \left[t\frac{d}{dt} + a\right]\big\{t(1+g(t))\big\} -k\left[2t\frac{d}{dt} - 1\right]g(t) + \left[a+1 - t\frac{d}{dt}\right]\left\{\frac{g(t)}{t} - L_{\infty}\right\} = 0\\ \iff & (t^2 - 2kt - 1)g'(t) + \left((a+1)t + k + \frac{a+2}{t}\right) g(t) + (a+1)(t - L_\infty) = 0\\ \iff & g'(t) + \left[ -\frac{a+2}{t} + \frac{a+\frac32}{t - \mu_{+}} + \frac{a + \frac32}{t - \mu_{-}}\right] g(t) = (a+1)\frac{t - L_\infty}{1+2kt - t^2}\\ \iff & \frac{d}{dt}\left[\frac{(1+2kt-t^2)^{a+\frac32}}{t^{a+2}} g(t)\right] = (a+1)\frac{(1+2kt-t^2)^{a+\frac12}}{t^{a+2}}(t-L_\infty)\\ \implies & w(t) = 1 + (a+1)\frac{t^{a+2}}{(1+2kt-t^2)^{a+\frac32}}\left[\int_0^t \frac{(1+2ks-s^2)^{a+\frac12}}{s^{a+2}}(s-L_\infty) ds\right] \end{align} $$ For the range of choice of $a$, the coefficient in front of the square bracket is singular at $t = \mu_{-}$. Since $|\mu_{-}| < |\mu_{+1}|$, $w(t)$ is regular there. This means the integral in the square bracket need to vanish at $t = \mu_{-}$. From this, we can deduce

$$L_{\infty} = \frac{\Lambda(a,-a)}{\Lambda(a,-a-1)} \quad\text{ where }\quad \Lambda(a,b) \stackrel{def}{=} \int_0^{\mu_{-}} s^{b-1}(1+2ks-s^2)^{a+\frac12} ds $$ Let $c = a + b + \frac32$ and $\eta = \frac{\mu_{-}}{\mu_{+}}$, the integral $\Lambda(a,b)$ can be evaluated as

$$\begin{align}\Lambda(a,b) & = \int_0^{\mu_{-}} s^{b-1}\left(1 - \frac{s}{\mu_{+}}\right)^{a+\frac12}\left(1 - \frac{s}{\mu_{-}}\right)^{a+\frac12} ds\\ &= \mu_{-}^b \int_0^1 s^{b-1}(1-s)^{c-b-1}\left(1-\eta s\right)^{a+\frac12} ds\\ &=^{\color{blue}{[2]}} \mu_{-}^b \frac{\Gamma(b)\Gamma(c-b)}{\Gamma(c)}{}_2F_1\left(-a-\frac12,b;c; \eta\right)\\ &=^{\color{blue}{[3]}} \mu_{-}^b(1 - \eta)^{-b} \frac{\Gamma(b)\Gamma(c-b)}{\Gamma(c)} {}_2F_1\left(c+a+\frac12,b;c; \frac{\eta}{\eta - 1}\right)\\ \end{align} $$ Substitute this back into expression of $L_\infty$ and simplify, we get

$$L_\infty = -\frac{2(a+1)\mu_{-}}{1 - \eta} \frac{ {}_2F_1\left(a+2,-a;\frac32; \frac{\eta}{\eta-1}\right) }{ {}_2F_1\left(a+1,-(a+1);\frac12; \frac{\eta}{\eta-1}\right) } $$ Notice $$ \frac{\mu_{-}}{1-\eta} = -\frac{1}{2\sqrt{k^2+1}} = -\frac{\sin\theta}{2} \quad\text{ and }\quad \frac{\eta}{\eta - 1} = \frac{\sqrt{k^2+1}-k}{2\sqrt{k^2+1}} = \frac{1-\cos\theta}{2} $$ We get $$L_\infty = \frac{(a+1)\sin\theta {}_2F_1\left(a+2,-a;\frac32; \frac{1-\cos\theta}{2}\right) }{ {}_2F_1\left(a+1,-a-1,\frac12; \frac{1-\cos\theta}{2}\right) } $$ Compare this with following representation of generalized Chebyshev's polynomial in terms of hypergeometric functions.

$$\begin{cases} T_a(\cos x) &= {}_2F_1\left(a,-a; \frac12; \frac12(1-\cos x)\right) = \cos(a x)\\ U_a(\cos x) &= (a+1){}_2F_1\left(a+2,-a; \frac32; \frac12(1 - \cos x)\right) = \frac{\sin((a+1)x)}{\sin x} \end{cases} $$ We find $$L_\infty = \tan((a+1)\theta) = \tan\left((a+1)\tan^{-1}\frac{1}{k}\right)$$

Up to what I see, there is nothing to stop us from analytic continuing this result back to $a \in (0,1)$. This implies $$\Theta(k,1,m) = \tan\left(\frac{m+1}{m}\tan^{-1}\frac1k\right)$$

Back to the case $z \ne 1$. The CF can be obtained by scaling $\alpha_\ell, \gamma_\ell$ by $\sqrt{z}$ for all $\ell > 0$ and $\gamma_0$ by $\frac{1}{\sqrt{z}}$. Since $(*1)$ is invariant when one rescale $\alpha_\ell, \beta_\ell, \gamma_\ell$ by same factor, scaling $\alpha_\ell, \gamma_\ell$ by $\sqrt{z}$ is equivalent to replace the $k$ in $\beta_\ell$ by $\frac{k}{\sqrt{z}}$. As long as $z \in \mathbb{C} \setminus (-\infty, 0 ]$, we have $\Re \sqrt{z} > 0$. One can verify the corresponding $\lambda_{\pm}$ have distinct modulus and above arguments remain valid.

As a result, the CF at hand has following closed form expression:

For $k > 0$, $m > 1$ and $z \in \mathbb{C} \setminus (-\infty,0]$, $$\Theta(k,z,m) = \frac{1}{\sqrt{z}}\Theta\left(\frac{k}{\sqrt{z}},1,m\right) = \frac{1}{\sqrt{z}}\tan\left[\frac{m+1}{m}\tan^{-1}\left(\frac{\sqrt{z}}{k}\right)\right]$$

The rest is obvious.


  • $\color{blue}{[1]}$ - For more details of Poincare-Perron theorem and a proof of it, please consult Chapter 8 of Saber Elaydi's book An Introduction to Difference Equations.

  • $\color{blue}{[2]}$ - For $|t| < 1$ and $\Re c > \Re b > 0$, we have following integral formula by Euler $$\int_0^1 x^{b-1}(1-x)^{c-b-1}(1-tx)^{-a} dx = \frac{\Gamma(b)\Gamma(c-b)}{\Gamma(c)} {}_2F_1(a,b;c;t)$$

  • $\color{blue}{[3]}$ - We are using following Pfaff transformation here.

$${}_2F_1(a,b;c;t) = (1-t)^{-b} {}_2F_1\left(c-a,b;c;\frac{t}{t-1}\right)$$

  • $\begingroup$ Nice proof. I see you started by proving the case $z=1$ and then used the functional equation $$\Theta(k,z,m)=\frac{1}{\sqrt{z}}\Theta(\frac{k}{\sqrt{z}},1,m)$$ for the general case $\endgroup$ – Nicco May 16 '16 at 10:07
  • $\begingroup$ @Nicco, yup. In any event, you can also deduce the functional equation through suitable rescaling of $t$ and $k$ in above derivation. $\endgroup$ – achille hui May 16 '16 at 10:13
  • $\begingroup$ @ ahille hui :Indeed that's possible. Good Job achille $\endgroup$ – Nicco May 16 '16 at 10:25
  • $\begingroup$ @ achille hui : perhaps you may be interested in this problem $\endgroup$ – Nicco May 16 '16 at 16:30
  • $\begingroup$ Assume analytic continuation continue to work, the RHS on the other problem is simply $\frac{4(a+b)}{b-a}\Theta\left(1,1,-\frac{a+b}{2a}\right) = \frac{4(a+b)}{b-a}\tan\left(\frac{\pi}{4}\frac{b-a}{b+a}\right)$. The $-\frac{a+b}{2a}$ there is taking the role of $m$ here. $\endgroup$ – achille hui May 17 '16 at 3:07

The case with $z \neq 1$ can be obtained from the case $z=1$, because you can just factor them out of the infinite fraction while changing $k$ accordingly.

Consider $k,m$ fixed, and let $(f_n)$ be the sequence of homographies $f_n(x) = a_n + b_n / x$, with $a_n = (2n-1)km$ and $b_n = ((n-1)m-1)((n+1)m+1)$ so that
$\Theta = \lim_{n \to \infty} (m+1)/ f_0 \circ f_1 \circ f_2 \circ f_3 \dots f_n(0)$" (well I guess this is where you put the $0$ but it shouldn't really matter)

(I think you should be able to define this more purely, if you call $x_n$ the attractive fixpoint of $f_0 \circ \dots \circ f_n$, then $\Theta = \lim (m+1)/x_n$, but leaving a zero is fine).

Now consider a nonzero complex number $z$ and let $g(x) = x/z$. After conjugation by $g$, you get $g^{-1} \circ f_n \circ g (x) = za_n + (z^2b_n)/x$, and so $\lim_{n \to \infty} g^{-1} \circ f_0 \circ f_1 \circ \dots f_n (0)$ is given by the infinite fraction where we replace $a_n$ and $b_n$ with $za_n$ and $z^2 b_n$. This amounts to replacing $k$ with $zk$ and adding a $z^2$ factor in front of every numerator.

Thus, $g^{-1}((m+1)/\Theta(k,1)) = (m+1)/\Theta(kz,z^2)$, so $z\Theta(kz,z^2) = \Theta(k,1)$

If we know the conjecture for $\Theta(k,1)$, we have $\frac{(i+\Theta(k,1))^m}{(i-\Theta(k,1))^m}=\frac{(ik+1)^{m+1}}{(ik-1)^{m+1}}$

And $\frac{(i+z\Theta(kz,z^2))^m}{(i-z\Theta(kz,z^2))^m}=\frac{(ik+1)^{m+1}}{(ik-1)^{m+1}} = \frac{(ikz+z)^{m+1}}{(ikz-z)^{m+1}}$. So letting $w =z^2$ and $k' = kz$ we obtain $\frac{(i+\sqrt w \Theta(k',w))^m}{(i-\sqrt w\Theta(k',w))^m} = \frac{(ik'+\sqrt w)^{m+1}}{(ik'- \sqrt w)^{m+1}}$ (the choice of square root doesn't matter as long as we pick the same on both side)

  • $\begingroup$ yes i only show that the case $z \neq 1$ reduces to the case $z=1$ after rescaling things by $z^2$. so it would make this a duplicate of an earlier question $\endgroup$ – mercio May 13 '16 at 23:49
  • $\begingroup$ @ mercio :that's because of the tranformation property $z\Theta(kz,z^2)=\Theta(k,1)$ which is very fundamental. $\endgroup$ – Nicco May 14 '16 at 1:23

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.