Add base class for distributions with non-conjugate priors and an example use #67
Conversation
| // Transition the current latent values. | ||
| virtual void transition_theta(std::mt19937* prng) = 0; | ||
|
|
||
| void incorporate(const T& x) { |
There was a problem hiding this comment.
In general we'll want to override incorporate/unincorporate to maintain sufficient statistics and avoid the loop at each call to logp_score, right? Is that not possible for Skellam, or is this more efficient?
For distributions with mostly unique samples, keeping a running tally of logp_score at each call to incorporate/unincorporate could be better as it would avoid looping over all the data in logp_score. Ultimately I think we might want to leave incorporate/unincorporate/logp_score undefined in the base class and have each distribution explicitly do what's best, but I think this is fine for now.
There was a problem hiding this comment.
I've been told in places like https://www.johndcook.com/CompendiumOfConjugatePriors.pdf (bottom of page 2) that having fixed dimensional sufficient statistics implies the existence of a conjugate prior family. That said, I've never actually read the precise statement or proof of that claim.
I'm fine with a running tally, though. I'll upgrade the code to do that now.
No description provided.