Distribution of exponential family

AI Thread Summary
The discussion centers on expressing a two-variable probability function in canonical form to determine if it belongs to the exponential family. The probability function is given as p(y1,y2)=Γ(y1+y2+γ)/((y1+y2)!*Γ(γ)), with γ known and greater than zero. Participants suggest that the function can be analyzed by considering the random variable Z=Y1+Y2, which simplifies the expression. They inquire about the existence of functions h(z) and g(γ) to facilitate the transformation into canonical form. The goal is to derive the expected value and variance matrices from this model.
the_dane
Messages
29
Reaction score
0
Let's say my probability function is given by: p(y1,y2)=Γ(y1+y2+γ)/((y1+y2)!*Γ(γ)), when γ>0 is known. I suppose it is from an exponential family but I can't write in canonical form because I'm only familiar with exponential family with one variable so I'm confused now when there's to variable. Can someone help me out here.
 
Physics news on Phys.org
For the distribution of a 2-vector random variable like that to be in the exponential family it must be able to be written in the factorised form

$$p(y_1,y_2)=h(y_1,y_2)g(\gamma)\exp\left(\sum_{k=1}^s\eta_k(\gamma)T_k(y_1,y_2)\right)$$

where ##s## is a non-negative integer and ##h,g, \eta_1,...,\eta_s, T_1,...,T_s## are all known functions.
 
the_dane said:
Let's say my probability function is given by: p(y1,y2)=Γ(y1+y2+γ)/((y1+y2)!*Γ(γ)), when γ>0 is known. I suppose it is from an exponential family but I can't write in canonical form because I'm only familiar with exponential family with one variable so I'm confused now when there's to variable. Can someone help me out here.
Consider the random variable ##Z=Y_1+Y_2##.

Therefore, you have:

##
\begin{eqnarray*}
\frac{\Gamma(y_1+y_2+\gamma)}{(y_1+y_2)! \ \Gamma(\gamma)} = \frac{\Gamma(z+\gamma)}{z! \ \Gamma(\gamma)} = \frac{1}{z!} \cdot \frac{1}{\Gamma(\gamma)} \cdot \Gamma(z+\gamma) \\
\end{eqnarray*}##

Do you have a function ##h(z)## and a function ##g(\gamma)## now?

Also, remember that ##a=\exp(\log(a))##! :wink:
 
Thanks for the answers. I have edited my model a lot and I'm now looking at this model. probability function for Y=(Y1,Y2) is given by p. Can anyone bring this to canonical form so I can find expected value and variance matrixes.
https://dl.dropboxusercontent.com/u/17974596/Sk%C3%A6rmbillede%202016-02-02%20kl.%2007.35.26.png
 
Last edited by a moderator:
... or just help me find the variance and expected value :)
 
Back
Top