This question arises from the study of the k-parameter exponential family of distributions.

When the canonical form is specified i.e. the family $\mathcal{P}_\theta$ is reparametrized by $\eta_i = \eta_i(\theta), i=\{1,\dotsc,k\}$ with $\theta \in \Theta \subset \mathbb{R}^k$, there is an implicit assumption that the dimension of $\Theta$ and the dimension of the image under the map $(\theta_1,\dotsc,\theta_k)\rightarrow (\eta_1(\theta),\dotsc,\eta_k(\theta)) $ are equal to k. Can we conclude that the mapping is one-to-one, which is crucial for identifiability of $\theta$?

I did not see this in the notes I find so far. Can someone offer his/her wisdom on the matter?


For arbitrary continuous functions, no such implication holds. The map $e^{i\theta}$ is continuous and surjective from the real line onto a circle. Both have dimension $1$, but the map rather dramatically fails to be one to one, as the inverse image of every point is infinite.

  • $\begingroup$ Thanks Matt, I totally agree with you. I suppose identifiability is too elementary a topic that the statisticians hardly brush on it if ever. $\endgroup$ – David KWH Sep 6 '16 at 18:16
  • $\begingroup$ @DKWH No problem. $\endgroup$ – Matt Samuel Sep 6 '16 at 18:19

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.