Suppose we have a random variable $X$ having p.d.f of the form $$f(x|\theta)=\exp\{c(\theta)'T(x)-B(\theta)\}h(x),$$ then we say $X$ is from exponential family. Further, we say that an exponential family is minimal if the functions $c(\theta)$ and the statistics $T(X)$, are linearly independent respectively.

I am trying to determine the minimal exponential family form of the following case.

Here, $X$ has p.d.f $$f(x|\alpha,\beta)=\exp\left\{\sum_{i=1}^nx_{i}(\alpha+\beta z_i)-\sum_{i=1}^n\log(1+\exp(\alpha+\beta z_i))\right\},$$

which is equal to

$$f(x|\alpha,\beta)=\exp\left\{\sum_{i=1}^nx_{i}\alpha + \sum_{i=1}^n\beta z_ix_i-\sum_{i=1}^n\log(1+\exp(\alpha+\beta z_i))\right\}$$

For the first equation the natural sufficient statistic $T$ is , $T=(X_1, \dots,X_n)$, while the second is $T=(\sum_{i=1}^nX_i,\sum_{i=1}^nX_iz_i)$.

I wonder which one is the minimal exponential family form?


1 Answer 1


How can you have a sufficient statistic with more dimensions than the number of parameters? If that were true, then the expectation parametrization would have more dimensions than the natural parametrization.

I think if $n \le 1$, it's the first form (and the parameter is just $\alpha + z_1 \beta$), otherwise it's the second form.

  • 1
    $\begingroup$ Look into curved exponential families $\endgroup$ Feb 23, 2021 at 7:54
  • $\begingroup$ @kjetilbhalvorsen I don't think that the sufficient statistic of a curved exponential family can have more dimensions than the number of parameters. Can you provide an example of such a sufficient statistic? It's certainly not true for, for example, the normal distribution with known variance (one sufficient statistic, one parameter). $\endgroup$
    – Neil G
    Feb 23, 2021 at 7:59

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.