\(\log q_\phi(x \mid \theta) + \log p(\theta)\) for an nle() fit and
\(\log r_\phi(\theta, x) + \log p(\theta)\) for an nre() fit, returning
-Inf outside the prior support. This is the potential the MCMC samplers
target, and the two fits differ only in what surrogate_ops() hands back.
Arguments
- fit
An
nsbi_nleornsbi_nrefit.- x_obs
The observation to condition on. Rows are independent observations of the same parameter.
- max_batch
Largest number of
(theta, x)pairs evaluated at once. No caller overrides it today, but it is a real tuning knob – the batch size the MCMC evaluations are chunked into – that a future caller would plausibly want to change, so it stays a parameter.