10
$\begingroup$

$$\frac{\left(\beta_{2}+\beta_{1}\right)^{\alpha_{1}+\alpha_{2}-1}B\left(\alpha_{1},\alpha_{2}\right)}{\beta_{1}^{\alpha_{1}}\beta_{2}^{\alpha_{2}}}=\frac{1}{\alpha_{1}\beta_{2}}{_{2}F_{1}}\left({1-\alpha_{2},1\atop 1+\alpha_{1}};-\tfrac{\beta_{1}}{\beta_{2}}\right)+\frac{1}{\alpha_{2}\beta_{1}}{_{2}F_{1}}\left({1-\alpha_{1},1\atop 1+\alpha_{2}};-\tfrac{\beta_{2}}{\beta_{1}}\right)$$
Where $B(\alpha_{1},\alpha_{2})=\frac{\Gamma\left(\alpha_{1}\right)\cdot\Gamma\left(\alpha_{2}\right)}{\Gamma\left(\alpha_{1}+\alpha_{2}\right)}$ is the Beta function
To me it's interesting because:

  1. It has 4 independent parameters/variables.

  2. It relates $\frac{\beta_{1}}{\beta_{2}} and \frac{\beta_{2}}{\beta_{1}}$.

  3. The 1 in the upper term can be manipulated by Hypergeometric relations or Generalized Hypergeometric relations:

• DLMF 15.5: https://dlmf.nist.gov/15.5

• DLMF 16.3(i): https://dlmf.nist.gov/16.3.i
If this type of identity is known I would be interested in any information. Is it possible that the original probability problem (see below) could be a source of new Hypergeometric identities?

Background:
While kibitzing on a paper by Aaron Hendrickson I noticed an equivalent form arising from the calculation of conditional probabilities for the gamma-difference distribution. Specifically, for $Y\sim\mathcal{GD}(\alpha_{1},\alpha_{2},\beta_{1},\beta_{2})$ which is the distribution of the difference of independent gamma variables we have the following results. $$\operatorname{pr}(Y\leq0)=C_{Y}\frac{\Gamma(\alpha_{1}+\alpha_{2})}{\beta_{2}\alpha_{1}}{_{2}F_{1}}\left({1-\alpha_{2},1\atop 1+\alpha_{1}};-\tfrac{\beta_{1}}{\beta_{2}}\right),$$
$$\operatorname{pr}(Y\geq0)=C_{Y}\frac{\Gamma(\alpha_{1}+\alpha_{2})}{\beta_{1}\alpha_{2}}{_{2}F_{1}}\left({1-\alpha_{1},1\atop 1+\alpha_{2}};-\tfrac{\beta_{2}}{\beta_{1}}\right),$$
where
$$C_{Y}=\beta_{1}^{\alpha_{1}}\beta_{2}^{\alpha_{2}}(\beta_{1}+\beta_{2})^{1-\alpha_{1}-\alpha_{2}},$$
for $\alpha_{1},\alpha_{2},\beta_{1},\beta_{2}>0$. If true then it follows that $\operatorname{pr}(Y\leq0)+\operatorname{pr}(Y\geq0)=1$ (with an overlap of measure zero) and the equation is true.
As far as I can tell the statement is true numerically but I have no idea how to prove it in terms of Hypergeometric functions.

$\endgroup$
6
$\begingroup$

This expression can be obtained from the connection formula \begin{equation} w_{1}(z)=\frac{\Gamma\left(c\right)\Gamma\left(b-a\right)}{\Gamma\left(b\right)\Gamma\left(c-a\right)}w_{5}(z)+\frac{\Gamma\left(c\right)\Gamma\left(a-b% \right)}{\Gamma\left(a\right)\Gamma\left(c-b\right)}w_{6}(z) \end{equation} where \begin{align} w_{1}(z)&={_{2}F_{1}}\left({a,b\atop c};z\right)\\ w_{5}(z)&=e^{a\pi\mathrm{i}}z^{-a}{_{2}F_{1}}\left({a,a-c+1\atop a-b+1};\frac{1}{z}\right)\\ w_{6}(z)&=e^{b\pi\mathrm{i}}z^{-b}{_{2}F_{1}}\left({b,b-c+1\atop b-a+1};\frac{1}{z}\right) \end{align} Choosing $a=1,b=1-\alpha_2,c=1+\alpha_1$ and $z=-\beta_1/\beta_2$, we have \begin{align} w_{1}(z)&={_{2}F_{1}}\left({1,1-\alpha_2\atop 1+\alpha_1};-\frac{\beta_1}{\beta_2}\right)\\ w_{5}(z)&=\frac{\beta_2}{\beta_1}{_{2}F_{1}}\left({1,1-\alpha_1\atop 1+\alpha_2};-\frac{\beta_2}{\beta_1}\right)\\ w_{6}(z)&=e^{(1-\alpha_2)\pi\mathrm{i}}\left( -\frac{\beta_1}{\beta_2} \right)^{-(1-\alpha_2)}{_{2}F_{1}}\left({1-\alpha_2,1-\alpha_1-\alpha_2\atop 1-\alpha_2};-\frac{\beta_2}{\beta_1}\right) \end{align} The last expression can be simplified as (see here) \begin{equation} w_{6}(z)=\left( \frac{\beta_1}{\beta_2} \right)^{-(1-\alpha_2)}(1+\frac{\beta_2}{\beta_1})^{\alpha_1+\alpha_2-1} \end{equation} The connection formula reads thus \begin{align} {_{2}F_{1}}\left({1,1-\alpha_2\atop 1+\alpha_1};-\frac{\beta_1}{\beta_2}\right)=& \frac{\Gamma(1+\alpha_1)\Gamma(-\alpha_2)}{\Gamma(1-\alpha_2)\Gamma(\alpha_1)}\frac{\beta_2}{\beta_1}{_{2}F_{1}}\left({1,1-\alpha_1\atop 1+\alpha_2};-\frac{\beta_2}{\beta_1}\right)\\ &+\frac{\Gamma(1+\alpha_1)\Gamma(\alpha_2)}{\Gamma(1)\Gamma(\alpha_1+\alpha_2)}\left( \frac{\beta_1}{\beta_2} \right)^{-(1-\alpha_2)}(1+\frac{\beta_2}{\beta_1})^{\alpha_1+\alpha_2-1} \end{align} which after some simplifications leads to \begin{align} {_{2}F_{1}}\left({1,1-\alpha_2\atop 1+\alpha_1};-\frac{\beta_1}{\beta_2}\right)=& -\frac{\alpha_1}{\alpha_2}\frac{\beta_2}{\beta_1}{_{2}F_{1}}\left({1,1-\alpha_1\atop 1+\alpha_2};-\frac{\beta_2}{\beta_1}\right)\\ &+\alpha_1 B(\alpha_1,\alpha_2)\left( \beta_1+\beta_2 \right)^{\alpha_1+\alpha_2-1}\beta_1^{-\alpha_1}\beta_2^{1-\alpha_2} \end{align} or to \begin{equation} \frac{1}{\alpha_1\beta_2}{_{2}F_{1}}\left({1,1-\alpha_2\atop 1+\alpha_1};-\frac{\beta_1}{\beta_2}\right)+ \frac{1}{\alpha_2\beta_1}{_{2}F_{1}}\left({1,1-\alpha_1\atop 1+\alpha_2};-\frac{\beta_2}{\beta_1}\right) = B(\alpha_1,\alpha_2)\left( \beta_1+\beta_2 \right)^{\alpha_1+\alpha_2-1}\beta_1^{-\alpha_1}\beta_2^{-\alpha_2} \end{equation} as expected.

$\endgroup$
  • 1
    $\begingroup$ Thanks! Neat, clean, and instructive :) Is there some way to give you extra credit? I've basically ignored Kummer's solution set (20); which I now see was obviously foolish. $\endgroup$ – rrogers Jun 4 '18 at 18:42
  • 2
    $\begingroup$ You're welcome! That was a nice exercise for me... $\endgroup$ – Paul Enta Jun 4 '18 at 18:48

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.