Extracting gradient of log posterior and unconstrained parameters

[simonrmaskell]

We're keen to enable Stan's users to capitalise on control variates. To make that possible (in R), one route would be to add functionality to posterior and to CmdStanR (which I now understand to use posterior to generate the summaries that control variates could sensibly augment). See here: stan-dev/posterior#273.

To make that possible, we'll need to get the gradient of the log posterior with respect to the unconstrained parameters and the unconstrained parameters themselves out of CmdStanR. We're happy to expend effort to make this happen, but would welcome perspectives on how this might best be achieved.

[jgabry]

In the latest version of cmdstanr @andrjohns added methods grad_log_prob, unconstrain_variables, and unconstrain_draws. Does that give you what you need?

• https://mc-stan.org/cmdstanr/reference/fit-method-init_model_methods.html
• https://mc-stan.org/cmdstanr/reference/fit-method-grad_log_prob.html

[simonrmaskell]

Awesome. Yes, I think so. Thank you @jgabry and @andrjohns !!