I have already read Gamma and Beta function proof.

Here is my work. By definition, the Beta function is $$B(a, b) = \int_{0}^{1}x^{a-1}(1-x)^{b-1}\text{ d}x$$ Let $y = \dfrac{x}{1-x}$, the "odds ratio" of $x$. Then after some algebra, it can be shown that $x = \dfrac{y}{1+y}$, $\text{d}x = \dfrac{1}{(1+y)^2}\text{ d}y$, and $1-x=\dfrac{1}{1+y}$, yielding $$B(a, b) = \int_{0}^{\infty}\dfrac{y^{a-1}}{(1+y)^{a+b}}\text{ d}y\text{.}$$ Now we have, for the Gamma function, $$\Gamma(a) = \int_{0}^{\infty}y^{a-1}e^{-y}\text{ d}y\text{.}$$ For $k > 0$, let $ku = y$, so that $k\text{ d}u = \text{d}y$, hence after some work, we obtain $$\Gamma(a) = k^a\int_{0}^{\infty}u^{a-1}e^{-ku}\text{ d}u$$ from which we obtain $$\dfrac{1}{k^a}=\dfrac{\displaystyle \int_{0}^{\infty}u^{a-1}e^{-ku}\text{ d}u}{\Gamma(a)}$$ Now, the linked question above says - if I'm interpreting it correctly - to notice $$\dfrac{1}{(1+y)^{a+b}}=\dfrac{\displaystyle \int_{0}^{\infty}u^{a+b-1}e^{-(1+y)u}\text{ d}u}{\Gamma(a+b)}$$ and then substitute this into $B(a, b)$: $$\begin{align}B(a, b) &= \dfrac{1}{\Gamma(a+b)}\int_{0}^{\infty}y^{a-1}\left[\int_{0}^{\infty}u^{a+b-1}e^{-(1+y)u}\text{ d}u \right]\text{ d}y \\ &= \dfrac{1}{\Gamma(a+b)}\int_{0}^{\infty}y^{a+b-1}\left[\dfrac{\Gamma(a+b)}{(1+y)^{a+b}} \right]\text{ d}y \end{align}$$ ... and I'm not sure how this helps me.


You have a double integral at some point, so your approach is using multivariate calculus.
But I can outline an approach which really avoids it. For any $a,b>0$, it is pretty clear that $B(a,b)=B(b,a)$ and that both $B(\cdot,b), B(a,\cdot)$ are continuous, non-negative functions on $\mathbb{R}^+$. We have $B(a,1)=\frac{1}{a}$ and by integration by parts we also have

$$ \frac{B(a+1,b)}{B(a,b)}= \frac{a}{a+b}$$ so for a fixed value of $b\in\mathbb{R}^+$ the functions $$ f(a) = B(a,b)\Gamma(a+b),\qquad g(a)=\Gamma(a)\Gamma(b)$$ fulfill $f(1)=g(1)=\Gamma(b)$ and $\frac{f(a+1)}{f(a)}=\frac{g(a+1)}{g(a)}=a$ for any $a\in\mathbb{R}^+$.

By the Bohr-Mollerup theorem $\Gamma$ is log-convex on $\mathbb{R}^+$.
By the Cauchy-Schwarz inequality $B(a,b)$ is midpoint-log-convex with respect to the $a$ variable:

$$ \int_{0}^{1}x^{a_1}(1-x)^b\frac{dx}{x(1-x)}\int_{0}^{1}x^{a_2}(1-x)^b\frac{dx}{x(1-x)} \geq \left(\int_{0}^{1}x^{\frac{a_1+a_2}{2}}(1-x)^b\frac{dx}{x(1-x)}\right)^2$$ and since $B(a,b)$ is non-negative and continuous, it is log-convex.
By the Bohr-Mollerup theorem again, $f(a)$ and $g(a)$ are the same function on $\mathbb{R}^+$, hence $$ B(a,b)=\frac{\Gamma(a)\Gamma(b)}{\Gamma(a+b)} $$ as wanted.


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.