Let $$ f(y | \eta) = h(y) \exp\left( \eta^\top T(y) + A(\eta) \right)$$ be the exponential family with base density/pmf $h$, sufficient statistic $T$, log partition function $A$ and natural parameter $\eta$. If it helps, we may restrict it to the case where $\eta$ is scalar. Obviously $h$ must be nonnegative. One can show that $A$ must also be convex.

The Wikipedia page states:

The function $A(\eta)$ ... is automatically determined once the other functions have been chosen, since it must assume a form that causes the distribution to be normalized (sum or integrate to one over the entire domain).

Bold s on functions by me for emphasis.

My Question:

Given some fixed $T$, am I free to choose any convex $A$ such that there exists an $h$ such that the resulting function is a valid pdf/pmf? More precisely, given some $T$ consider the set $$\mathcal{A}_T = \{ A:\mathbb{R} \to \mathbb{R} \mid \exists \, h:\mathbb{R} \to \mathbb{R} \mid A(\eta) = \log\int h(y)\exp \left(\eta T(y) \right)\, dy\}.$$ Is there a useful alternative characterisation of this set?


I do not understand the "free" in the question. Or the question at all..

As rightly pointed by the Wikipedia page, $A(\cdot)$ is completely determined by the choice of the triplet

dominating measure $\text dy$ x base density $h(\cdot)$ x statistic $T(\cdot)$

Up to a constant, $A(\cdot)$ is a Laplace transform and hence satisfies all properties of Laplace transforms, including identifying the associated distribution (of $T$).

Choosing first $A(\cdot)$ out of the blue and then looking for an associated triplet does not seem like a fruitful endeavour.

  • $\begingroup$ I appreciate your response, even though my question is evidently not clear. I understand that $A(\cdot)$ is must satisfy $A(\eta) = \log \int h(y) \exp\left( \eta^\top T(y) \right) \, d y$ and can see how it can be used to identify the cumulants of $T$. The reason I would like to choose $A$ first is that I find myself in the situation where I am forced to use a specific $A$ and am wondering whether the associated subset of exponential families is useful. Is there anything I can add to my question to make it better? $\endgroup$
    – Student
    Jul 15 at 8:11
  • $\begingroup$ A useful item of information would be to explain why you choose $A(\cdot)$ first. Otherwise, I do not for one see how one can answer the question. $\endgroup$
    – Xi'an
    Jul 15 at 15:56

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.