We have a random variable $X$ that belongs to the exponential family with p.d.f.

$$ P_X(x|\boldsymbol \theta) = h(x) \exp\left(\eta({\boldsymbol \theta}) . T(x) - A({\boldsymbol \theta}) \right) $$

where ${\boldsymbol \theta} = \left(\theta_1, \theta_2, \cdots, \theta_s \right )^T$ is the parameter vector and $\mathbf{T}(x)= \left(T_1(x), T_2(x), \cdots,T_s(x) \right)^T$ is the joint sufficient statistic and $A({\boldsymbol \theta}) = \log \int_x h(x)\exp( \eta(\boldsymbol \theta).T(x))dx$

(The notation is following the Wikipedia page on the exponential family of distributions)

Let the data be given labels such that the joint distribution is now associated with $(x, y) \in \mathcal{X}\times\mathcal{Y}$.

EDIT The sufficient statistics for this joint distribution is given by $\mathbf{T}(x, y)$

I am unable to derive the following expression for the exponential form of the conditional distribution of labels given data (ignoring the reference measure $h(x)$)

$$ P(y | x; \theta) = \exp\left(\eta({\boldsymbol \theta}) . T(x,y) - A({\boldsymbol \theta | x}) \right) $$

with $A({\boldsymbol \theta|x}) = \log \int_{\mathcal{Y}} \exp( \eta(\boldsymbol\theta).T(x,y))dy$

This expression is used in a paper on missing variables that I am trying to implement. I have tried writing out the conditional in terms of the joint probability but did not get any clean decomposition of terms.

Is there any standard proof or text that derives the expression for conditional probability of an exponential family distribution? Any hints or references would be great. Thanks.

  • 1
    $h(x)$ is not the natural parameter but the reference measure. – Xi'an Feb 5 '15 at 18:08
  • @Xi'an You're right, I am sorry - missed out on adding the joint sufficient statistics. Does it make sense now? Thanks. – AruniRC Feb 6 '15 at 0:41
  • Defining the "joint sufficient statistics" does not tell about the joint distribution. Is it still an exponential family with natural sufficient statistics $T(x,y)$? – Xi'an Feb 6 '15 at 7:39
  • Yes, the joint distribution is still an exponential family with the natural sufficient statistics $T(x,y)$ – AruniRC Feb 6 '15 at 23:02
up vote 3 down vote accepted

I have answered my own question. It turned out to be a rather obvious application of Bayes Rule only after making a somewhat arbitrary assumption. My question was not very clear, mostly due to my own tenuous understanding at that time.

However, this result is used quite a lot in machine learning literature involving integrating out missing variables. I am including the proof in case others find it helpful when seeing the result.

$$ P(x, y|\boldsymbol \theta) = h(x) \exp\left(\eta({\boldsymbol \theta}) . T(x, y) - A({\boldsymbol \theta}) \right) $$

By Bayes Rule,

$$ P(y|x, \theta) = \frac{ P(x|y, \theta)}{ \int_{y^{'}} P(x|{y^{'}}, \theta) P(y^{'}|\theta)d{y^{'}}} = \frac{ P(x, y| \theta)}{ \int_{y^{'}} P(x,{y^{'}}| \theta) d{y^{'}}} = \frac{h(x) \exp (\eta (\theta) . T(x,y) - A(\theta))}{ \int_{y^{'}} h(x) \exp (\eta (\theta) . T(x,y^{'}) - A(\theta))dy{'}} $$

Assumed the $h(x)$ base reference measure to be a function only of $x$ so that we can cancel it from numerator and denominator in the last step above, getting

$$ \frac{\exp ( \eta(\theta).T(x,y))}{\int_{y^{'}} \exp ( \eta(\theta).T(x,y^{'}))dy^{'}} = \exp ( \eta(\theta).T(x,y) - \log(\int_{y^{'}} \exp ( \eta(\theta).T(x,y^{'}))dy^{'}) ) = \exp ( \eta(\theta).T(x,y) - A(\theta|x) ) $$

Your Answer


By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.