10
$\begingroup$

I'm looking for the following second derivative

$$ \kappa_2 := \left . \frac{d^2}{d\lambda^2} \ln \left({_2F_1}\!\left(\tfrac{1}{2},\,- \lambda;\,1;\,\alpha\right)\right) \right \vert_{\lambda = 0} , $$

where $\alpha$ is a real parameter in $[0,1]$. As you may have guessed I'm trying to compute the variance of a certain probability distribution.

The first moment is simple and is given by

$$ m_1 = 2 \ln \frac{1 + \sqrt{1-\alpha}}{2}. $$

The second moment can be shown to be equal to the following

$$ m_2 = \frac{2}{\pi} \int_{0}^{1} \frac{\left [ \ln(1-\alpha y^2) \right ]^2}{\sqrt{1-y^2}} dy $$

Mathematica can evaluate the above integral in terms of various function and polylogarithms (in fact dilogarithms). However the resulting expression is not even manifestly real. The above integral expression for $m_2$ is so far the best I found to deal with but there are many others. In essence I am trying to find a "nice" expression for that integral ($m_2$).

I'd be happy if I'm given an expression which is manifestly real. I suspect that the combination $\kappa_2 = m_2 - (m_1)^2 $ (which is equal to the first equation) might look nicer and that some identity involving dilogarithms should be used.

Added

An alternative representation for $m_2$ (obtained using the series of the Hypergeometric) is the following

$$ m_{2}=2\sum_{n=2}^{\infty}\left(\begin{array}{c} -1/2\\ n \end{array}\right)\frac{H_{n-1}}{n}\left(-\alpha\right)^{n} \, , $$

where $H_n$ are the Harmonic numbers, i.e.

$$ H_n \, = \, \sum_{p=1}^n \frac{1}{p} $$

So far this is the best I could get:

$$ \kappa_2 \,=\, -4 \log ^2\left(\sqrt{1-\alpha }+1\right)+4 \log \left(4-4 \sqrt{1-\alpha }\right) \log \left(\sqrt{1-\alpha }+1\right)+4 i \pi \log \left(\frac{2}{\sqrt{1-\alpha }+1}\right)+4 \log (2) \log \left(\frac{1}{\alpha }\right)+4 \text{Li}_2\left(\frac{2 \left(\sqrt{1-\alpha }+1\right)}{\alpha }\right)-4 \text{Li}_2\left(\frac{-\alpha +2 \sqrt{1-\alpha }+2}{\alpha }\right) $$

I would like to avoid the explicitly imaginary (third) term which is compensated by the dilogarithm to produce a real result. Does anybody know an identity for dilogarithm that can be used here?

$\endgroup$
12
  • 1
    $\begingroup$ Can you include Mathematica's output? We might be able to make its reality manifest. $\endgroup$ Sep 6, 2014 at 3:13
  • $\begingroup$ Can you post Mathematica's input ? $\endgroup$ Sep 6, 2014 at 3:45
  • $\begingroup$ I am unable to get the analytical expression of $m_2$. Did you make a change of variable ? $\endgroup$ Sep 6, 2014 at 4:07
  • $\begingroup$ @ClaudeLeibovici: Here is the output. $\endgroup$
    – Lucian
    Sep 6, 2014 at 6:19
  • 1
    $\begingroup$ I know this is an old question, but the series the OP obtained can be written in the form $$2 \sum_{n=2}^\infty \frac{(2n)! ~H_{n-1}}{n!^2 ~n} \frac{\alpha^n}{4^n}$$ which is perfectly real and converges very fast (and has rational terms for rational $\alpha$). $\endgroup$
    – Yuriy S
    Apr 23, 2018 at 13:10

2 Answers 2

2
$\begingroup$

I will start from scratch. For any $R\geq 1$ we have $$ (R-e^{i\theta})(R-e^{i\theta}) = (R^2+1)-2R\cos\theta \tag{1}$$ $$ 1-\frac{2R}{R^2+1}\cos\theta = \frac{R^2}{R^2+1}\left(1-\frac{e^{i\theta}}{R}\right)\left(1-\frac{e^{-i\theta}}{R}\right)\tag{2}$$ hence by setting $\frac{2R}{R^2+1}=\frac{1}{A}$ we have $A\geq 1$ and $$ \log\left(1-\frac{1}{A}\cos\theta\right) = \log\left(\frac{A+\sqrt{A^2-1}}{2A}\right)+\sum_{n\geq 1}\frac{\cos(n\theta)}{n R^n}.\tag{3} $$ By integrating both sides of $(3)$ on $(0,\pi/2)$ we have $$ \int_{0}^{1}\frac{\log\left(1-\frac{x}{A}\right)}{\sqrt{1-x^2}}\,dx = \frac{\pi}{2}\log\left(\frac{A+\sqrt{A^2-1}}{2A}\right)-2\operatorname{Im}\operatorname{Li}_2\left(iA-i\sqrt{A^2-1}\right)\tag{4} $$ and similarly $$ \int_{0}^{1}\frac{\log\left(1+\frac{x}{A}\right)}{\sqrt{1-x^2}}\,dx = \frac{\pi}{2}\log\left(\frac{A-\sqrt{A^2-1}}{2A}\right)+2\operatorname{Im}\operatorname{Li}_2\left(iA+i\sqrt{A^2-1}\right)\tag{5} $$ such that by summing $(4)$ and $(5)$ $$ \int_{0}^{1}\frac{\log\left(1-\frac{x^2}{A^2}\right)}{\sqrt{1-x^2}}\,dx =\\= -\frac{\pi}{2}\log\left(2A\right)+2\operatorname{Im}\operatorname{Li}_2\left(iA+i\sqrt{A^2-1}\right)-2\operatorname{Im}\operatorname{Li}_2\left(iA-i\sqrt{A^2-1}\right).\tag{6} $$ Similarly we may integrate $\frac{\log\left(1\pm\frac{x}{A}\right)}{\sqrt{1-x^2}}$ on sub-intervals of $(0,1)$, always in terms of $\log$ and $\operatorname{Im}\operatorname{Li}_2$. Actually it turns out you do not need $\text{Im}\text{Li}_2$ to compute $\kappa_2$, just $\text{Li}_2$. From $H_{n}=\int_{0}^{1}\frac{1-x^n}{1-x}\,dx$ it follows that the computation of $ \sum_{n\geq 2}\left[\frac{1}{4^n}\binom{2n}{n}\right]\frac{\alpha^n H_{n-1}}{n}$ boils down to the computation of

$$ \int_{0}^{1}\left[\frac{1}{x}\log\left(\frac{1+\sqrt{1-\alpha x}}{2}\right)-\log\left(\frac{1+\sqrt{1-\alpha}}{2}\right)\right]\frac{dx}{1-x},$$ which by integration by parts and partial fraction decomposition reduces to the determination of the integrals over $(0,1)$ of $$ \frac{\log(x)}{\sqrt{1-\alpha x}},\frac{\log(1-x)}{\sqrt{1-\alpha x}},\frac{\log(x)}{1+\sqrt{1-\alpha x}},\frac{\log(1-x)}{1+\sqrt{1-\alpha x}}.$$ The former two are equivalent to each other and elementary, the latter two depend on $$ \int_{0}^{1}\frac{\log(x)}{1+\sqrt{1-\alpha x}}\,dx = \frac{1}{2}\,\phantom{}_4 F_3\left(\tfrac{1}{2},1,1,1;2,2,2;\alpha\right) $$ which reduces (up to elementary terms) to $$ \int_{\sqrt{1-\alpha}}^{1}\frac{x\log(1-x^2)}{1+x}\,dx $$ then to $\operatorname{Li}_2\left(\frac{1-\sqrt{1-\alpha}}{2}\right)$. Some details needs to be filled but you have all the necessary ingredients.

$\endgroup$
2
  • $\begingroup$ Thank you, this seems a lot of work. I appreciate it, but I don't see how does that answer the question. $\endgroup$
    – lcv
    Mar 25, 2019 at 8:19
  • $\begingroup$ @lcv: it finds $\kappa_2$. $\endgroup$ Mar 25, 2019 at 14:39
1
$\begingroup$

The definition of $\operatorname{Li}_2$ used in the last formula is $$\operatorname{Li}_2(z) = -\operatorname{Li}_2 \left( \frac 1 z \right) - \frac {3 \ln^2(-z) + \pi^2} 6, \quad |z| > 1.$$ Therefore $$\kappa_2 = 4 \operatorname{Li}_2 \left( \frac \alpha {\omega - \alpha} \right) - 4 \operatorname{Li}_2 \left( \frac \alpha \omega \right) + 2\ln^2 \left( \frac \omega \alpha - 1 \right) - 2\ln^2 \left( \frac \omega \alpha \right) + \\ 4 \ln \left( \frac \omega 2 \right) \ln \left( \frac {4 (4 - \omega - \alpha)} \alpha \right) - \ln(16) \ln(\alpha), \\ \omega = 2 + 2 \sqrt {1 - \alpha}.$$

$\endgroup$
2
  • $\begingroup$ Thank you. I haven't checked your formula yet, but how did you remove the explicitly imaginary term? $\endgroup$
    – lcv
    Mar 25, 2019 at 8:21
  • 1
    $\begingroup$ The polylogarithm terms are real after the substitution, and $\ln^2 z = (\ln(-z) + i \pi)^2$ for negative $z$. The three imaginary terms cancel out. $\endgroup$
    – Maxim
    Mar 25, 2019 at 16:57

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.