While the correct definition of exponential family is

$$ f_X(x\mid\theta) = h(x) \exp \left (\eta(\theta) \cdot T(x) -A(\theta)\right ), $$

it seems that in many materials I read, they don't pay much attention to $h(x)$. Sometimes, authors just drop this term.

Can somebody tell me more about the meaning of this $h(x)$ term? The most detailed description about this term I've found is "simply reflects the underlying measure w.r.t. which $p(x\mid\theta)$ is a density." (See http://stat-www.berkeley.edu/pub/users/mjwain/Fall2012_Stat241a/reader_ch8.pdf). Also, some people call it "base measure".

I kind of understand that $h(x)$ may not be important because it's trivial in most cases (1 or a constant like $1/\sqrt{2\pi}$), but for some distributions, e.g. Poisson, this term is nontrivial ($1/x!$), and it moderates the exponential term so greatly.

My understanding of this term is 1) it assigns a base probability to each element in the space of $x$, and the exponential term modifies this base probability. 2) it's here so that we don't get unbounded log-partition function $A(\theta)$, in cases like Poisson.

Can anybody tell me more about this $h(x)$, and any comment on my understanding of it is welcomed.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.