Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

According to Wikipedia, the Beta distribution is related to the gamma distribution by the following relation:

$$\lim_{n\to\infty}n B(k, n) = \Gamma(k, 1)$$

Can you point me to a derivation of this fact? Can it be generalized? For example, is there a similar relation that results in something other than a constant 1 for the Gamma second parameter? What if we have

$$\lim_{n\to\infty,m\to\infty,n=mb}n B(k, m) $$

That is, the two variables go to infinity while maintaining a constant ratio b.

The reason I'm asking is because I'm trying to figure out how to simplify a hieraerchical bayesian model involving the beta distribution.

(This is my first post; sorry for the math notation, the MathJaX syntax was too daunting, but I'll try to learn)

share|improve this question
    
It's not MathJaX, it's LaTex, and there's a very very helpful post here: meta.math.stackexchange.com/questions/5020/… :) –  Daniel Sep 3 '12 at 22:32
    
Your "fact" is wrong. For any integer $k \ge 2$, $$\lim_{n \to \infty} n B(n,k) = \lim_{n \to \infty} \dfrac{\Gamma(k)}{(n+1)(n+2)\ldots(n+k-1)} = 0$$ –  Robert Israel Sep 3 '12 at 23:04
    
Perhaps you meant $$ \lim_{n \to \infty} n^k B(n,k) = \Gamma(k)$$ –  Robert Israel Sep 3 '12 at 23:06
    
For fixed $b > 0$, $$B(n,bn) \sim \sqrt{2\pi/n} b^{bn-1/2}(1+b)^{1/2-n-bn} \ \text{as}\ n \to \infty $$ –  Robert Israel Sep 3 '12 at 23:12
1  
Robert, I think you are talking about the Beta and Gamma functions, whereas my question concerns the Beta and Gamma distributions. –  Sten Linnarsson Sep 6 '12 at 7:49

3 Answers 3

up vote 3 down vote accepted

Let $X_n$ denote a random variable with beta distribution $\mathrm B(k,n)$ and $Y_n=nX_n$. Then, for every $s\geqslant0$, $\mathrm E(Y_n^s)=n^s\mathrm E(X_n^s)$ and one knows $\mathrm E(X_n^s)$, hence $$ \mathrm E(Y_n^s)=n^s\frac{\mathrm B(k+s,n)}{\mathrm B(k,n)}=n^s\frac{\Gamma(k+s)\Gamma(k+n)}{\Gamma(k+s+n)\Gamma(k)}\longrightarrow\frac{\Gamma(k+s)}{\Gamma(k)}. $$ This is $\mathrm E(Z^s)$ for any random variable $Z$ with gamma distribution $\Gamma(k)$ hence $Y_n\to Z$ in distribution.

Let $\widetilde X_n$ denote a random variable with beta distribution $\mathrm B(k,n/b)$, and $\widetilde Y_n=n\widetilde X_n$. Then, $\widetilde X_n$ is distributed like $X_{n/b}$ hence $\widetilde Y_n$ is distributed like $bY_{n/b}$ and $\widetilde Y_n\to bZ$ in distribution.

share|improve this answer
    
Thanks! I'm actually shocked and amazed at the speed and quality of answers on the stackexchange sites. It's like having an extra brain with just slightly higher latency! –  Sten Linnarsson Sep 6 '12 at 8:07
    
I would upvote your answer if I could, but I get "please login or register to vote", even though I'm already logged-in. –  Sten Linnarsson Sep 6 '12 at 8:10
    
Thanks. Since I am not an expert on these matters of login/register/vote, I am flagging this answer so that your comment is read by some moderator... –  Did Sep 6 '12 at 9:51
    
@Sten & did, according to math.stackexchange.com/priviledges it requires 15 rep before a user can vote Q/A up and down. Also, while Sten is logged in, his account is not registered. Registering has many benefits (user identity more consistently kept track of etc.) and I strongly encourage users to do it. –  Willie Wong Sep 6 '12 at 11:53
    
@Sten See? These MSE mods are amazing, aren't they? :-) –  Did Sep 6 '12 at 12:13

This concerns the relationship between the Gamma and Beta distributions as opposed to the Gamma and Beta functions. Let $X \sim \mbox{Gamma}(\alpha, 1)$ and $Y \sim \mbox{Gamma}(\beta, 1)$ where the paramaterization is such that $\alpha$ is the shape parameter. Then
$$ \frac{X}{X + Y} \sim \mbox{Beta}(\alpha, \beta). $$

To prove this, write the joint pdf $f_{X, Y} (x, y) = \frac{1}{\Gamma(\alpha) \Gamma(\beta)} x^{\alpha - 1} y^{\beta - 1} e^{-(x + y)}$ (on $\mathbb R^2_+$) and make the transformation $U = \frac{X}{X + Y}$ and $V = X + Y$. The Jacobian of the transformation $X = VU, Y = V(1 - U)$ is equal to $V$ so the joint distribution of $U$ and $V$ has pdf $$ \frac{v}{\Gamma(\alpha)\Gamma(\beta)} (vu)^{\alpha - 1} (v (1 - u))^{\beta - 1} e^{-v} = \frac{1}{\Gamma(\alpha)\Gamma(\beta)}v^{\alpha + \beta - 1} e^{-v} u^{\alpha - 1} (1 - u)^{\beta - 1} $$ (on $\mathbb R_+ \times [0, 1]$) and hence $U$ and $V$ are independent (because the pdf factors over $u$ and $v$) with $V \sim \mbox{Gamma}(\alpha + \beta, 1)$ and $U \sim \mbox{Beta}(\alpha, \beta)$ which is apparent from the terms $v^{\alpha + \beta - 1} e^{-v}$ and $u^{\alpha - 1}(1 - u)^{\beta - 1}$ respectively.

share|improve this answer

There is another way to view the relationship between the gamma distribution and the beta distribution through the Dirichlet distribution. This post (http://math.stackexchange.com/q/190695) talks about exactly how they are related without the Dirichlet distribution, but here is a slightly broader view:

Let $Z_1, Z_2, \ldots, Z_n$ be independent random variables such that $$ Z_i \sim \text{Gamma}(\alpha_i,1), \quad i=1,\ldots, n$$ where $\alpha_i \geq 0$ are the shape parameters. If $Y_j$ is defined as follows, $$Y_j = Z_j/\sum_{i=1}^n Z_i, \quad j=1, \ldots, n$$ then $(Y_1, \ldots, Y_n)$ is Dirichlet distributed with parameter $(\alpha_1, \ldots, \alpha_n)$. When $n=2$, the Dirichlet distribution reduces to the Beta distribution, denoted by $\text{Beta}(\alpha_1, \alpha_2)$.

Here is a link to Ferguson's paper that mentions the above relationship: http://www.cis.upenn.edu/~taskar/courses/cis700-sp08/papers/ferguson.pdf

share|improve this answer

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.