# Exponential family canonical form

I cannot figure this myself - suppose I have some distribution $$f(y; \theta)$$ from the exponential family. By Wikipedia (https://en.wikipedia.org/wiki/Exponential_family) let us assume this form:

$$f(y; \theta) = h(y)exp\big[\eta(\theta)T(y) - A(\theta)\big]$$

The wiki defines the canonical form as

If $$\eta(\theta) = \theta$$, then the exponential family is said to be in canonical form

Then, it adds the part I do not get:

By defining a transformed parameter $$\eta(\theta) = \theta$$, it is always possible to convert an exponential family to canonical form

I do not get what they do with parameters $$\theta$$. For example, let us have the Normal with unknown mean $$\mu$$ AND unknown variance $$\sigma^2$$. There is immediately example in wiki, where they define:

$$\eta = \big(\frac{\mu}{\sigma^2}, -\frac{1}{2\sigma^2}\big)$$

and then conclude: then our form is canonical. Why is that? They initially define the normal as $$f(y; \mu, \sigma)$$! According to the previous definition, it is clear that $$\eta(\theta) \neq (\mu, \sigma^2)$$! Why do they call it a canonical form then?

Actually this can be shown almost for all distributions when written in canonical form, their parameter vector $$\eta$$ is clearly not equal to the parameters in the definition.

• It's a typo. They define $\theta=(\cdot,\cdot)^{\top}$ (look at the next lines where they define $A(\eta)$). – d.k.o. Jan 31 at 18:44
• Then does it mean that the real density should be defined as $f(y; \frac{\mu}{\sigma^2}, -\frac{1}{2\sigma^2})$? They kinda no where mention that parameters can be swapped like that (apart from second quote "By defining a transformed parameter..." - which i did not get - is it what it implies?). – John Jan 31 at 18:56