Why are we using log(pdf) everywhere?

[maxnoe]

Wouldn't it be more intuitive to use the pdf and not logp when defining the classes? kogp could be automatically created by the base class (Distribution)

[ibab]

Sometimes there are constant terms in the log likelihood that can be left out. These could be detected, but I think it's easier to leave them out explicitly.
Also, this is how pymc and Stan do it, so using logp as well means we can compare our PDF implementations more easily.

[maxnoe]

Don't we get problems when we multiply pdfs if we leave out constants in log(pdf)?

[ibab]

Multiplication of PDFs becomes addition under log(pdf).

[maxnoe]

of course