Say, $X$ is dependent on $\alpha$. Rigorously speaking,

  • if $X$ and $\alpha$ are both random variables, we could write $p(X\mid\alpha)$;

  • however, if $X$ is a random variable and $\alpha$ is a parameter, we have to write $p(X; \alpha)$.

I notice several times that the machine learning community seems to ignore the differences and abuse the terms.

For example, in the famous LDA model, where $\alpha$ is the Dirichlet parameter instead of a random variable.

enter image description here

Shouldn't it be $p(\theta;\alpha)$? I see a lot of people, including the LDA paper's original authors, write it as $p(\theta\mid\alpha)$.

  • 8
    $\begingroup$ Mathematically speaking, you can always condition on a constant, since this is a limiting case of random variable. From a Bayesian point of view, all unknowns are treated as random variables, so it makes sense to use the conditioning notation all over. $\endgroup$
    – Xi'an
    Commented Jan 1, 2015 at 9:38
  • 1
    $\begingroup$ @Xi'an I understand your point on "conditioning on a constant". But imagine I draw $X$ from a categorical distribution of parameter $\theta$, i.e., $X\sim Cat(\theta)$. Can I write the distribution as $p(X\mid\theta)$? That looks weird to me, since one can always set a fixed $\theta$. $p(X;\theta)$ looks more comfortable to me. $\endgroup$ Commented Jan 1, 2015 at 9:43
  • 6
    $\begingroup$ I do not see the problem in writing $p(X\mid\theta)$ in this special case. Once again, using conditional notations paves the way to introduce prior distributions on every unknown parameter. $\endgroup$
    – Xi'an
    Commented Jan 1, 2015 at 9:47

1 Answer 1


I think this is more about Bayesian/non-Bayesian statistics than machine learning vs.. statistics.

In Bayesian statistics parameter are modelled as random variables, too. If you have a joint distribution for $X,\alpha$, $p(X \mid \alpha)$ is a conditional distribution, no matter what the physical interpretation of $X$ and $\alpha$. If one considers only fixed $\alpha$s or otherwise does not put a probability distribution over $\alpha$, the computations with $p(X; \alpha)$ are exactly the same as with $p(X \mid \alpha)$ with $p(\alpha)$. Furthermore, one can at any point decide to extend the model with fixed values of $\alpha$ to one where there is a prior distribution over $\alpha$. To me at least, it seems strange that the notation for the distribution-given-$\alpha$ should change at this point, wherefore some Bayesians prefer to use the conditioning notation even if one has not (yet?) bothered to define all parameters as random variables.

Argument about whether one can write $p(X ; \alpha)$ as $p(X \mid \alpha)$ has also arisen in comments of Andrew Gelman's blog post Misunderstanding the $p$-value. For example, Larry Wasserman had the opinion that $\mid$ is not allowed when there is no conditioning-from-joint while Andrew Gelman had the opposite opinion.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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