2
$\begingroup$

In Wikipedia it is stated that:

A vector exponential family is said to be curved if the dimension of $$ {\boldsymbol \theta} = \left (\theta_1, \theta_2, \ldots, \theta_d \right )^T$$ is less than the dimension of the vector $$ {\boldsymbol \eta}(\boldsymbol \theta) = \left (\eta_1(\boldsymbol \theta), \eta_2(\boldsymbol \theta), \ldots, \eta_s(\boldsymbol \theta) \right )^T.$$

with

$$ f_X(x|\boldsymbol \theta) = h(x) g(\boldsymbol \theta) \exp\Big(\boldsymbol\eta({\boldsymbol \theta}) \cdot \mathbf{T}(x)\Big) $$

but what happens when the dimension of $\boldsymbol \eta$ is smaller than the one of $\boldsymbol \theta$ ?

$\endgroup$
2
$\begingroup$

The dimension of the parameter vector $\theta$ can't be larger than that of the natural parameters $\eta$ because then you have necessarily redundant $\theta$. The idea of a curved exponential family is that you're slicing out part of the natural parameter space. If you have extra parameters then you are still taking the whole natural parameter space, and it's just a regular exponential family.

$\endgroup$
  • 1
    $\begingroup$ Thanks. From "The Bayesian choice", C. Robert p 138: $x \sim N(\theta, \theta^2)$ induces an exponential family of dimension 2. Is not it contradictory with our answer ? Thanks in advance. $\endgroup$ – peuhp Jun 12 '14 at 12:42
  • 1
    $\begingroup$ @peuhp There seem to be two senses of the word "dimension" in operation here: one is the dimension of a manifold (coordinatized by the parameters $\theta$, of dimension $d$) and the other is a dimension of a space in which it is embedded (with coordinates $\eta$, of dimension $s$). The latter was likely meant in the C. Robert quotation--but whether that is the correct interpretation depends on exactly how Robert defines "dimension." $\endgroup$ – whuber Jun 12 '14 at 14:27
  • $\begingroup$ @whuber. Thanks. From the book (and roughly speaking), the dimension of the exponential family is the one of $\nu(\theta)$, so I am a bit lost. It seems to me still contradictory: $\theta$ is a scalar ($d=1$) but $\nu(\theta)$ is cleary a 2d vector ($s=2$). So $d>s$. Thanks for helping me to clarify my mind. $\endgroup$ – peuhp Jun 12 '14 at 15:13
  • 1
    $\begingroup$ @peuhp: Isn't your example exactly the one defined to be a curved exponential family? The dimension of $\theta$ is 1; the dimension of $\eta$ is 2. It's a parabolic slice through the cartesian plane. $\endgroup$ – Neil G Jun 12 '14 at 15:24
  • $\begingroup$ @Neil G. You right. I realise I must be tired because my question makes no sense. I ask by the negative exactly what I state preliminarly to my question. Thanks. $\endgroup$ – peuhp Jun 12 '14 at 15:28

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.