0
$\begingroup$

Can you explain, how does $\text{B}(\alpha, \beta)$ transfrom to $\text{B}(s+\alpha, f+\beta)$ in the following equation?

$$ \begin{align*} p(\left. q=x \right| s,f) &= {{{s+f \choose s} x^{s+\alpha-1}(1-x)^{f+\beta-1} / \operatorname{B}(\alpha,\beta)} \over \int_{y=0}^1 \left({s+f \choose s} y^{s+\alpha-1}(1-y)^{f+\beta-1} / \operatorname{B}(\alpha,\beta)\right) dy} \\ ~\\ ~\\ ~\\ ~\\ & = {x^{s+\alpha-1}(1-x)^{f+\beta-1} \over \operatorname{B}(s+\alpha,f+\beta)} \end{align*} $$ (This equation is taken from Wikipedia's article on conjugate priors.)

However, I started an investigation from parameter estimation for text analysis by G. Heinrich (in that paper, equations 23 and 24).

I see, that in many papers, this is an "of course" statement, but I would be grateful for a derivation of the formula, or hints on how to do it on my own.

$\endgroup$
  • 2
    $\begingroup$ Full reference for Heinrich paper please: are you expecting people to Google it themselves? $\endgroup$ – Nick Cox Aug 4 '13 at 14:38
  • 1
    $\begingroup$ No math whatsoever is needed (which is why this is an "of course" result): notice that only the numerator depends on $x$ and that the $x$-dependence is accounted for entirely by the terms $x^{s+\alpha-1}(1-x)^{f+\beta-1}$ (whose values are defined and non-negative only for $0\le x\le 1$). Every bit of the rest is a constant and therefore has to equal the value needed to normalize the total probability to unity. Because a Beta distribution with parameters $s+\alpha, f+\beta$ is defined to have this form (up to the normalization constant), $q$ must be that Beta distribution. $\endgroup$ – whuber Aug 4 '13 at 15:32
  • $\begingroup$ @diabel: Is my answer below acceptable, or are you looking for something else? $\endgroup$ – Tyler Streeter Aug 8 '13 at 12:26
2
$\begingroup$

In the equation:

$$ \begin{align*} p(\left. q=x \right| s,f) &= \frac{ {s+f \choose s} x^{s+\alpha-1}(1-x)^{f+\beta-1} / \operatorname{B}(\alpha,\beta) }{ \int_{y=0}^1 \left({s+f \choose s} y^{s+\alpha-1}(1-y)^{f+\beta-1} / \operatorname{B}(\alpha,\beta)\right) dy } \\ &= \frac{ {s+f \choose s} x^{s+\alpha-1}(1-x)^{f+\beta-1} / \operatorname{B}(\alpha,\beta) }{ {s+f \choose s} \left( \int_{y=0}^1 y^{s+\alpha-1}(1-y)^{f+\beta-1} \, dy \right) / \operatorname{B}(\alpha \, \beta) } \\ &= \frac{ x^{s+\alpha-1}(1-x)^{f+\beta-1} }{ \int_{y=0}^1 y^{s+\alpha-1}(1-y)^{f+\beta-1} \, dy } \end{align*} $$

the denominator matches the form of the beta function:

$$ \operatorname{B}(x,y) = \int_0^1 t^{x-1}(1-t)^{y-1} \, dt $$

so the denominator becomes:

$$ \int_{y=0}^1 y^{s+\alpha-1}(1-y)^{f+\beta-1} \, dy = \operatorname{B}(s+\alpha,f+\beta) $$

and the original equation becomes:

$$ p(\left. q=x \right| s,f) = \frac{ x^{s+\alpha-1}(1-x)^{f+\beta-1} }{ \operatorname{B}(s+\alpha,f+\beta) } $$

$\endgroup$

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.