6

From what I understand, the exponential family is defined as

$$f(y;\theta,\phi) = \exp\left(\frac{y\theta - b(\theta)}{a(\phi)}+c(y,\phi)\right) $$

I've read (but not seen shown anywhere), that the t distribution is not a member of the exponential family. But I don't understand why.

For instance, suppose I set $\theta = 0$, $b(\theta)=0$, and set $$c(y,\phi) = \ln\left(\frac{\Gamma\left(\frac{\phi+1}{2}\right)}{\sqrt{\phi\pi}\Gamma(\frac{\phi}{2})} \left(1+\frac{x^2}{\phi} \right)^{-\frac{\phi+1}{2}}\right)$$,

wouldn't then the t distribution then appear because

you would have

$$\exp\left(0+\ln\left(\frac{\Gamma\left(\frac{\phi+1}{2}\right)}{\sqrt{\phi\pi}\Gamma(\frac{\phi}{2})} \left(1+\frac{x^2}{\phi} \right)^{-\frac{\phi+1}{2}}\right)\right)=\frac{\Gamma\left(\frac{\phi+1}{2}\right)}{\sqrt{\phi\pi}\Gamma(\frac{\phi}{2})} \left(1+\frac{x^2}{\phi} \right)^{-\frac{\phi+1}{2}}$$

which is the t-distribution.

Why doesn't this work?

EDIT:

Ideally, I'd like to know why also my logic above is wrong (which I'm certain it is). So if you can fit that into your answer, that would be great.

Stan Shunpike
  • 3,623
  • 2
  • 27
  • 36
  • 3
    Because the random variable and the parameter cannot separate as a product. – Xi'an Jan 12 '20 at 20:19
  • 2
    Your characterization of an exponential family is meaningless because *every* density function $f$ can be written in such a form: simply set $\theta=0=b(\theta)$ and $c(y;\phi)=\log f(y;\phi).$ – whuber Jan 12 '20 at 20:38
  • 1
    Your first equation is the definition of an *exponential dispersion model*, not the exponential family, and note that $a()$ is required to be positive. – Gordon Smyth Jan 13 '20 at 04:34
  • 1
    If $\phi$ is constant then it becomes a linear exponential family, but if $\theta$ is constant (as in your calculation) then it isn't anything at all. – Gordon Smyth Jan 13 '20 at 04:40

1 Answers1

6

No, the t distribution is not an exponential family. Exponential family distributions do have existing moment generating functions, and the t distribution do not. See also Why doesn't the exponential family include all distributions?

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467