We have that $$ \Delta_{\lambda} KL(\lambda) = \mathbb{E}_{\theta\sim q_{\lambda}(\theta),z\sim g_{N}(z|\theta)}(\Delta_{\lambda}[\log \: q_{\lambda}(\theta)](\log \: q_{\lambda}(\theta)-\hat{h}(\theta,z))) $$

We also have that $$q_{\lambda}(\theta)=\exp (T(\theta)'\lambda - Z(\lambda))$$

where $T(\theta)$ is vector of sufficient statistics and $\lambda$ is vector of natural parameters, so that $q_{\lambda}(\theta)$ has an exponential family form. Then $\Delta_{\lambda}Z(\lambda)=\mathbb{E}_{q_{\lambda}}[T(\theta)]$. Then, $\Delta_{\lambda} KL(\lambda)$ becomes

$$ \begin{eqnarray} \Delta_{\lambda} KL(\lambda) &=& \mathbb{E_{\theta}} [(T(\theta)-\mathbb{E}_{\theta}[T(\theta)])(T(\theta)'\lambda-\mathbb{E}_{z}\hat{h}(\theta,z))]\\ &=& \mathbb{E_{\theta}} [(T(\theta)-\mathbb{E}_{\theta}[T(\theta)])T(\theta)']\lambda-\mathbb{E}_{\theta}[(T(\theta)-\mathbb{E}_{\theta})\mathbb{E}_{z}\hat{h}(\theta,z)]\\ &=& \Cov _{q_{\lambda}}(T(\theta),T(\theta))\lambda - \mathbb{E}_{\theta}[(T(\theta)-\mathbb{E}_{\theta})\mathbb{E}_{z}\hat{h}(\theta,z)] \end{eqnarray} $$

I have three questions:

  1. Why is $\Delta_{\lambda}Z(\lambda)=\mathbb{E}_{q_{\lambda}}[T(\theta)]$ true?
  2. In $\Delta_{\lambda} KL(\lambda) = \mathbb{E_{\theta}} [(T(\theta)-\mathbb{E}_{\theta}[T(\theta)])(T(\theta)'\lambda-\mathbb{E}_{z}\hat{h}(\theta,z))]$, how did we get rid of $Z(\lambda)$?
  3. Why is $\mathbb{E_{\theta}} [(T(\theta)-\mathbb{E}_{\theta}[T(\theta)])T(\theta)']=\Cov _{q_{\lambda}}(T(\theta),T(\theta))$? Shouldn't it be $\mathbb{E_{\theta}} [(T(\theta)-\mathbb{E}_{\theta}[T(\theta)])(T(\theta)-\mathbb{E}_{\theta}[T(\theta)])']=\Cov _{q_{\lambda}}(T(\theta),T(\theta))$

Note: taken from the top of p. 09 on http://xxx.tau.ac.il/pdf/1503.08621v1.pdf

  • $\begingroup$ Please add the [self-study] tag & read its wiki. Then tell us what you understand thus far, what you've tried & where you're stuck. We'll provide hints to help you get unstuck. $\endgroup$ – gung - Reinstate Monica Jul 11 '15 at 15:58
  • $\begingroup$ Hi, this is not a homework/textbook problem. It's individual steps from a paper that I'm confused by, and I detail which steps I'm confused by in 1, 2, 3, so I don't think this qualifies. $\endgroup$ – mlstudent Jul 11 '15 at 16:00

All it involves is a little bit of calculation:

  1. Take the equation $q_{\lambda}(\theta) = \exp(T(\theta)'\lambda -Z(\lambda)$) and differentate both sides with respect to $\lambda$. The derivative of the exponential function is itself, and by the chain rule we obtain $$ \Delta_\lambda q_{\lambda}(\theta) = T(\theta)q_{\lambda}(\theta) - \Delta_\lambda Z(\lambda)q_{\lambda}(\theta)$$

Solving for $\Delta_\lambda Z(\lambda)$ we get $$ \Delta_\lambda Z(\lambda) = T(\theta) - \Delta_\lambda (\log q_{\lambda}(\theta))$$.

The paper notes on page 5 that $E_\theta(\Delta_\lambda (\log q_{\lambda}(\theta))) = 0$, hence taking expected values in the above equation we arrive at $$ \Delta_\lambda Z(\lambda) = E_\theta(\Delta_\lambda Z(\lambda)) = E_\theta T(\theta) - E_\theta(\Delta_\lambda (\log q_{\lambda}(\theta))) = E_\theta T(\theta) .$$

  1. $Z(\lambda)$ does not depend on $\theta$, so it's a constant when taking expected values over $\theta$ an $z$: $$ E_{\theta,z}( (T(\theta) - \Delta_\lambda Z(\lambda))Z(\lambda)) = Z(\lambda) E_{\theta,z}( (T(\theta) - \Delta_\lambda Z(\lambda)) = Z(\lambda) E_{\theta,z}( (T(\theta) - E_\theta T(\theta)) = 0 $$

  2. Yes, since $E_\theta ( (T(\theta) - E_\theta T(\theta)) E_\theta (T(\theta)')) = 0$, the two expressions are the same.

  • $\begingroup$ Hi, I was away when the bounty expired. I will look over this and try to award the bounty if I still can and I understand it. $\endgroup$ – mlstudent Jul 31 '15 at 19:24
  • $\begingroup$ No worries, let me know if it's unclear or mistaken. $\endgroup$ – sandris Aug 1 '15 at 12:46
  • $\begingroup$ Accepting it is appreciated, though, if you find it useful ;-) $\endgroup$ – sandris Aug 4 '15 at 6:49

I think that I have the answer for part (1): check out section 4 on this pdf. It's a result of being an exponential family.


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.