Stochastic determinant evaluation functionality

[rkarur]

I have a use case in my sampling applications to lattice field theory where I would like to define a logdensity function with a determinant term, but rather than fully evaluating the determinant, Id like to evaluate it in a stochastic fashion to reduce computational cost; another popular use case of this scenario could be to evaluate any likelihood function with a $1/\sqrt{\det A}$ term, where the $\det A$ term depends on parameters that are being sampled. This means that I would have to modify the logdensity_fn's in blackjax to carry an extra argument rng_key to be able to create random vectors at each step of a logdensity_fn's evaluation. I wonder what the best line of action is to implement this into blackjax, whether it is redefining all logdensity_fn's to (by default) take two arguments rather than one, or something else.

[reubenharry]

So (as discussed elsewhere, but to port the discussion here), it would first be good to understand what the downside of a partially applied log_density fuction is, i.e. make_logdensity_fn(key=key) which gets passed to the kernel at each step