It is known that (given a $\sigma$-finite Borel reference measure $\nu$ on $\mathbb{R}$) the parameter space of an exponential family is convex in Euclidean space. However, my question is, for an the associated exponential manifold $\sqrt{EM(c)}$ a convex subset of $L^2_{\nu}$? Where:

  • $$ EM(c)\triangleq \left\{ p(\cdot,\theta):\, \theta \in \Theta,\, p(x,\theta)=exp(\theta^Tc(x)-\psi(\theta)) \right\}, $$ for some fixed continuously-differentiable scalar functions $c=\{c_1,\dots,c_n\}$ on $\mathbb{R}^d$,
  • We assume that $EM(c)\subseteq L^2_{\nu}$.

Assuming that $\sqrt{EM(c)}$ is defined as $\{\sqrt p\colon p\in EM(c)\}$, the answer is "of course, no". Indeed, write $t$ and $T$ for $\theta$ and $\Theta$, respectively, and $p_t$ for $p(\cdot,t)$. Then the convexity of $\sqrt{EM(c)}$ would imply that, for any $s$ and $t$ in $T$, the function $\frac14\,p_s+\frac14\,p_t+\frac12\,\sqrt{p_s}\sqrt{p_t}=(\frac12 \sqrt{p_s}+\frac12 \sqrt{p_t})^2$ is in $EM(c)$ and hence a pdf, which (in view of the Cauchy--Schwarz inequality) will be the case only when $p_s=p_t$ $\nu$-almost everywhere for all $s$ and $t$ in $T$, i.e., when the exponential family has (essentially) only one member.


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.