12

Let $E(n)=\text{expectation of }\dfrac{\prod_{k=1}^n\tan x_k}{\sum_{k=1}^n\tan x_k}$ where $x_k$ are independent uniformly random real numbers in $\left(0,\frac{\pi}{2}\right)$.

Is the following conjecture true:

$E(n)=\left(\frac{\pi}{2}\right)^{2n-6}$ for $n>2$.

Context

Earlier I found that $E(2)=\left(\frac{2}{\pi}\right)^2\int_0^{\pi/2}\int_0^{\pi/2}\frac{(\tan x_1)(\tan x_2)}{\tan x_1+\tan x_2}dx_1dx_2=\frac{2}{\pi}$.

Naturally, I wondered if $E(n)$ has closed forms for other $n$ values.

Basis of my conjecture

$E(3)=\left(\frac{2}{\pi}\right)^3\int_0^{\pi/2}\int_0^{\pi/2}\int_0^{\pi/2}\frac{(\tan x_1) (\tan x_2)(\tan x_3)}{\tan x_1+\tan x_2+\tan x_3}dx_1dx_2dx_3=0.9999996\dots$, according to Desmos. I guess it's actually $1$.

For $n>3$, Desmos and Wolfram Cloud don't work, so I used Excel with about a million trials. I got:

  • $E(4)\approx2.480$, whereas $\left(\frac{\pi}{2}\right)^{2(4)-6}\approx2.467$
  • $E(5)\approx6.047$, whereas $\left(\frac{\pi}{2}\right)^{2(5)-6}\approx6.088$
  • $E(6)\approx16.794$, whereas $\left(\frac{\pi}{2}\right)^{2(6)-6}\approx15.022$
  • $E(7)\approx37.495$, whereas $\left(\frac{\pi}{2}\right)^{2(7)-6}\approx37.065$

As $n$ increases, the values of $\dfrac{\prod_{k=1}^n\tan x_k}{\sum_{k=1}^n\tan x_k}$ fluctuate more and more, so this numerical approach becomes less reliable.

If my conjecture is true, I can't imagine what a proof would look like.

Dan
  • 35,053

3 Answers3

9

The conjecture is true for $n=3$.

Consider the integral $$ I_3=\int_{[0,\pi/2]^3}\frac{\tan x_1\tan x_2\tan x_3}{\tan x_1+\tan x_2+\tan x_3}dx_1dx_2dx_3=\int_{[0,\infty]^3}\frac{xyz}{x+y+z}\frac{dxdydz}{(1+x^2)(1+y^2)(1+z^2)} $$ With change of variables $x=\tan x_1,y=\tan x_2,z=\tan x_3$. Now switch to spherical coordinates. $$ I_3=\int_0^\infty r^2dr\int_0^{\pi/2}\sin(\theta) d\theta\int_0^{\pi/2}d\varphi \frac{r^2\sin ^2(\theta ) \cos (\theta ) \sin (\varphi ) \cos (\varphi )}{\sin (\theta ) (\sin (\varphi )+\cos (\varphi ))+\cos (\theta )}\frac{1}{\left(r^2 \cos ^2(\theta )+1\right) \left(r^2 \sin ^2(\theta ) \sin ^2(\varphi )+1\right) \left(r^2 \sin ^2(\theta ) \cos ^2(\varphi )+1\right)} $$ The $r$ part is just rational integral, which is easily done from partial fractions $$ \frac2\pi\int_0^\infty\frac{r^4}{\left(r^2 \cos ^2(\theta )+1\right) \left(r^2 \sin ^2(\theta ) \sin ^2(\varphi )+1\right) \left(r^2 \sin ^2(\theta ) \cos ^2(\varphi )+1\right)}dr\\ =\frac{\sec (\theta )}{\left(\cos ^2(\theta )-\sin ^2(\theta ) \cos ^2(\varphi )\right) \left(\cos ^2(\theta )-\sin ^2(\theta ) \sin ^2(\varphi )\right)}+\frac{\csc ^5(\theta ) \sec (\varphi ) \sec (2 \varphi )}{\cos ^2(\varphi )-\cot ^2(\theta )}-\frac{\csc ^5(\theta ) \csc (\varphi ) \sec (2 \varphi )}{\sin ^2(\varphi )-\cot ^2(\theta )}\\ =\frac{\sin (\theta ) (\sin (\varphi )+\cos (\varphi ))+\cos (\theta )}{((\sin (\varphi )+\cos (\varphi )) (\sin (\theta ) \cos (\varphi )+\cos (\theta )) (\sin (\theta ) \sin (\varphi )+\cos (\theta ))) \left(\sin ^3(\theta ) \cos (\theta ) \sin (\varphi ) \cos (\varphi )\right)} $$ Plug back into $I_3$, lots of terms cancel out, and $$ I_3=\frac\pi2\int_0^{\pi/2}\frac{d\varphi}{\sin (\varphi )+\cos (\varphi )}\int_0^{\pi/2}\frac{d\theta}{ (\sin (\theta ) \cos (\varphi )+\cos (\theta )) (\sin (\theta ) \sin (\varphi )+\cos (\theta ))}\\ =\frac\pi2\int_0^{\pi/2}\frac{\log(\cot(\varphi))}{\cos(2\varphi)}d\varphi=\int_0^\infty \frac{\log t}{t^2-1}dt=\frac{\pi^3}8\qquad t=\cot(\varphi) $$ The conjecture for $n=3$ is thus proved.

Unfortunately, such massive cancellation fails to occur when $n=4$, so at least this approach is not easy.

Appendix

For numerical evaluation, $$ I_n=\int_{[0,\infty]^n}\frac{1}{x_1+\cdots+x_n}\frac{x_1dx_1}{1+x_1^2}\cdots\frac{x_ndx_n}{1+x_n^2}=\int_{[0,\infty]^n}\int_0^\infty e^{-t(x_1+\cdots+x_n)}dt\frac{x_1dx_1}{1+x_1^2}\cdots\frac{x_ndx_n}{1+x_n^2}\\ =\int_0^\infty\left(\int_0^\infty\frac{xe^{-xt}}{1+x^2}dx\right)^ndt=\int_0^\infty\left( \Big(\frac{\pi}{2} -\text{Si}(t)\Big) \sin (t)-\text{Ci}(t) \cos (t)\right)^ndt $$ From this, it holds in high precision $$ \left(\frac2\pi\right)^6I_4\approx 0.95596478957 $$ This result differs quite largely from the conjecture, so the latter is probably false. Indeed, finding a closed form is still tempting. I suspect polylogarithms could possibly appear in the result.

Interesting byproducts appear along my attempt, such as $$ \int_0^{\infty } \frac{\log (x) \left(\left(x^2+2\right) \log \left(x^2+1\right)-2 x \tan ^{-1}(x)\right)}{\left(x^2+1\right) \left(x^2+4\right)} \, dx =\frac{\pi ^3}{9}\\ \int_0^{\infty } \frac{\left(x^2+2\right) \log \left(x^2+1\right)-2 x \tan ^{-1}(x)}{x \left(x^2+1\right) \left(x^2+4\right)} \, dx =\frac{\pi ^2}{36} $$ It is invited to prove them. (The latter is easy)

Po1ynomial
  • 1,866
5

$\newcommand{\Ei}{\operatorname{Ei}} \newcommand{\Ci}{\operatorname{Ci}} \newcommand{\si}{\operatorname{si}} \newcommand{\Li}{\operatorname{Li}_4} $ We have $E(4),E(5)$ in closed forms. $$\begin{aligned} &E(2)=\frac{2}{\pi} ,\\ &E(3)=1,\\ &E(4)=\frac{28}{\pi^3}\zeta(3)+\frac{4}{\pi},\\ &E(5)=-\frac{1280}{3\pi^4}\Li\left(\frac12\right)+\frac{235}{27}-\frac{160}{9\pi^4}\ln(2)^4+\frac{160}{9\pi^2}\ln(2)^2, \end{aligned} $$ where $\operatorname{Li}_n(z)$ denotes polylogarithms.


Proof. The moment integral, as has shown by another answer, is equivalent to evaluate a sort of integrals with powered trigonometric integral functions. $$ \begin{aligned} E(n) & = \left ( \frac{2}{\pi} \right )^n \int_{[0,\pi/2]^n}\frac{\tan x_1...\tan x_n}{ \tan x_1+...+\tan x_n}\text{d}x_1...\text{d}x_n,\\ &=\left ( \frac{2}{\pi} \right )^n \int_{\left [ 0,\infty \right ]^n } \frac{1}{(1+x_1^2)...(1+x_n^2)} \frac{x_1...x_n}{x_1+...+x_n} \text{d}x_1...\text{d}x_n,\\ &=\left ( -1 \right )^n \left ( \frac{2}{\pi} \right )^n \int_{0}^{\infty} A(x)^n\text{d}x( A(x) \text{ to be defined as some trigonometric integrals}). \end{aligned} $$The last line is by replacing $\frac1a=\int_0^\infty e^{-at}\text{d}t$, switching the order of integration and using $$ A(\alpha)=-\int_{0}^{\infty} \frac{xe^{-\alpha x}}{1+x^2}\text{d}x ,\quad \alpha\in\mathbb{R}^{+}.$$

We use complex tricks to convert these trigonometric integrals into exponential integrals, thence to the substantial reduction in computation complexity. Denote two exponential analogues(because they behave like exponent functions), $$ E_1(z)=e^{z}\operatorname{Ei}_1(-z),E_2(z)=e^{-z}\left(\operatorname{Ei}_2(z)-\pi i\right), $$ where $\Ei_{1,2}(.)$ is the exponent integral function with branch cuts chosen. Use the definition, $$ \Ei_1(-z)= \gamma+\log(z)+\int_{0}^{1} \frac{e^{-zt}-1}{t} \text{d}t,\quad \Ei_2(z)= \gamma+\log(z)+\int_{0}^{1} \frac{e^{zt}-1}{t} \text{d}t, $$ and assume $\log(z)=\ln(|z|)+i\arg(z),\arg(z)\in[0,2\pi)$ with branch cut $[0,\infty).$ Note that for a postive real $z$, one have $\Ei_1(-z)=\Ei(-z),\Ei_2(z)=\Ei(z)$.


Next we show that, defining $$A(x)=\Ci(x)\cos(x)+\si(x)\sin(x),\quad B(x)=\si(x)\cos(x)-\Ci(x)\sin(x), $$ the following relations are valid for $x\in\mathbb{R}^{+}$. $$ E_1(ix)=A(x)-i B(x),E_2(ix)=A(x)+iB(x). $$ We start as follows. $$ \begin{aligned} \Ei_1(-ix) &=\gamma+\log(ix)+\int_{0}^{1} \frac{e^{-ixt}-1}{t} \text{d}t,\\ &=\left(\gamma+\log(x)+\int_{0}^{1} \frac{\cos(xt)-1}{t} \text{d}t\right) +i\left ( \frac\pi2-\int_{0}^{1} \frac{\sin(xt)}{t} \text{d}t\right ),\\ &=\Ci(x)-i \si(x).\\ \Rightarrow E_1(-ix)&=e^{ix}E_1(-ix),\\ &=\left ( \cos(x)+i \sin(x) \right ) \left(\Ci(x)-i \si(x)\right),\\ &=A(x)-iB(x). \end{aligned} $$The other can be done simultaneously. And it's worth noticing that for a positive real $\alpha$ $$ A(\alpha)=-\int_{0}^{\infty} \frac{xe^{-\alpha x}}{1+x^2}\text{d}x ,B(\alpha)=-\int_{0}^{\infty} \frac{e^{-\alpha x}}{1+x^2}\text{d}x. $$ A short proof is given by integrating respectively $$ f(z)=\frac{ze^{-\alpha z}\left(\Ei_2(\alpha z)-\pi i\right)}{1+z^2},\frac{e^{-\alpha z}\left(\Ei_2(\alpha z)-\pi i\right)}{1+z^2} $$ along the real-axis. Compare the imaginary part, giving $$ \begin{aligned} & \int_{0}^{\infty} \frac{xe^{-\alpha x}}{1+x^2}\text{d}x =\left ( -\frac{1}{\pi} \right ) \Im\left ( 2\pi i\operatorname{Res} (f(z),i)\right ) =-\Re E_2(i\alpha)=-A(\alpha),\\ &\int_{0}^{\infty} \frac{e^{-\alpha x}}{1+x^2}\text{d}x =\left ( -\frac{1}{\pi} \right ) \Im\left ( 2\pi i\operatorname{Res} (f(z),i)\right ) =-\Im E_2(i\alpha)=-B(\alpha). \end{aligned} $$


Then we apply some simple contour integration. Define the contour enclosed in the first quadrant with $R$ sufficiently large: $$C:[0,R)\cup\left \{ Re^{i\theta} \mid \theta\in\left[0,\frac{\pi}{2}\right]\right \} \cup\left \{ ix\mid x\in[0,R) \right \},$$ and integrate the function $$ f(z)=E_1(z)^mE_2(z)^n,m,n\in\mathbb{N},m+n\ge2, $$ giving $$ \int_{0}^{\infty}e^{(m-n)x}\Ei(-x)^m\left ( \Ei(x)-\pi i \right )^n \text{d} x =i\int_{0}^{\infty}\left ( A(x)-i B(x) \right )^m \left ( A(x)+iB(x) \right )^n\text{d}x. $$ The LHS are, by definition, some common logarithmical integrals.(If interested, click here.)

  • To calculate $E(4)$, substituting $(m,n)=(4,0),(3,1),(2,2)$ to obtain $$ \begin{aligned} &\int_{0}^{\infty}e^{4x}\Ei(-x)^4\text{d} x =i\int_{0}^{\infty}\left ( A(x)-i B(x) \right )^4\text{d}x,\\ &\int_{0}^{\infty}e^{2x}\Ei(-x)^3\left ( \Ei(x)-\pi i \right ) \text{d} x =i\int_{0}^{\infty}\left ( A(x)-i B(x) \right )^2 \left ( A(x)^2+B(x)^2\right )\text{d}x,\\ &\int_{0}^{\infty}\Ei(-x)^2\left ( \Ei(x)-\pi i \right )^2 \text{d} x =i\int_0^\infty\left ( A(x)^2+B(x)^2 \right )^2\text{d}x. \end{aligned} $$ Discard the real part of every equality, solve the linear system and make use of these known evaluations, $$ \int_{0}^{\infty}e^{2x}\Ei(-x)^3\text{d}x = -\frac72 \zeta(3),\qquad\int_{0}^{\infty}\Ei(-x)^2\Ei(x)\text{d}x =- \frac{\pi^2}{3}. $$ We obtain $$ \color{blue}{\int_{0}^{\infty} A(x)^4\text{d}x =\frac{7\pi}{4}\zeta(3)+\frac{\pi^3}4},\qquad \int_{0}^{\infty} A(x)^2B(x)^2\text{d}x =\frac{\pi^3}{12},\qquad \int_{0}^{\infty} B(x)^4\text{d}x =-\frac{7\pi}{4}\zeta(3)+\frac{\pi^3}4. $$ The first integral is exactly $(\frac\pi2)^4E(4)$, which completes the proof.

  • To calculate $E(5)$, substituting $(m,n)=(5,0),(4,1),(3,2)$ to obtain $$ \begin{aligned} &\int_{0}^{\infty}e^{5x}\Ei(-x)^5\text{d} x =i\int_{0}^{\infty}\left ( A(x)-i B(x) \right )^5\text{d}x,\\ &\int_{0}^{\infty}e^{3x}\Ei(-x)^4\left ( \Ei(x)-\pi i \right ) \text{d} x =i\int_{0}^{\infty}\left ( A(x)-i B(x) \right )^3 \left ( A(x)^2+B(x)^2\right )\text{d}x,\\ &\int_{0}^{\infty}e^{x}\Ei(-x)^3\left ( \Ei(x)-\pi i \right )^2 \text{d} x =i\int_{0}^{\infty}\left ( A(x)-i B(x) \right ) \left ( A(x)^2+B(x)^2 \right )^2\text{d}x. \end{aligned} $$ Discard the real part of every equality, solve the linear system and make use of these known evaluations, $$ \begin{aligned} &\int_{0}^{\infty}e^{3x}\Ei(-x)^4\text{d}x = \frac{26\pi^4}{135} +\frac{4\pi^2}{9}\ln(2)^2 -\frac{4}{9}\ln(2)^4-\frac{32}{3}\Li\left ( \frac12 \right ) ,\\ &\int_{0}^{\infty}e^{x}\Ei(-x)^3\Ei(x)\text{d}x = \frac{61\pi^4}{360} +\frac{\pi^2}{3}\ln(2)^2 -\frac{1}{3}\ln(2)^4-8\Li\left ( \frac12 \right ). \end{aligned} $$ We obtain $$ \begin{aligned} &\color{purple}{\int_{0}^{\infty} A(x)^5\text{d}x =\frac{40\pi}{3}\,\text{Li}_4\left(\frac{1}{2}\right)-\frac{235\pi^5}{864}+\frac{5\pi}{9}\ln(2)^4-\frac{5\pi^3}{9}\ln(2)^2},\\ &\int_{0}^{\infty} A(x)^3B(x)^2\text{d}x =\frac{4\pi}{3}\,\text{Li}_4\left(\frac{1}{2}\right)-\frac{131\pi^5}{4320}+\frac{\pi}{18}\ln(2)^4-\frac{\pi^3}{18}\ln(2)^2,\\ &\int_{0}^{\infty} A(x)B(x)^4\text{d}x =-\frac{\pi^5}{160} . \end{aligned} $$ The first integral is exactly $-(\frac\pi2)^5E(5)$, which completes the proof.
    Remark. It would be exhaustive to evaluate each integral separately. Moreover, the result of the last integral shows great simplicity. In fact, using integral representations we have $$ \int_{0}^{\infty} A(x)B(x)^4\text{d}x =-\int_{[0,\infty]^5} \frac{1}{\left ( 1+x_1^2\right ) \left ( 1+x_2^2\right )\left ( 1+x_3^2\right ) \left ( 1+x_4^2\right )\left ( 1+x_5^2\right ) } \frac{x_1}{x_1+x_2+x_3+x_4+x_5}\text{d}x_1...\text{d}x_5=-\frac{1}{5} \left ( \frac{\pi}{2} \right )^5 $$ by symmetry.

0

If $L(s)=\int_0^{\infty}\frac{u}{1+u^2}e^{-su}du$ then $E(n)=\int_0^{\infty}f(s)^nds$ by applying $1/A=\int_0^{\infty}e^{-sA}ds$ to $A=\tan x_1+\cdots+x_n.$

(Oh, too late!)