Using MH() with a custom proposal and acceptance rate of 1 to emulate traditional Gibbs?

[aws]

Hello,

Thank you in advance for reading. I am a researcher considering whether to use Turing for a project. As background, I will be implementing and extending a sparse infinite Bayesian factor model as described in Sparse Bayesian infinite factor models - PMC.

While this paper specifies the full conditional posterior distribution for each parameter and thus allows Gibbs sampling, I plan to introduce some additional latent variables that will require Metropolis-Hastings sampling within Gibbs. Thus, Turing’s compositional sampler is very appealing.

My question is, can I use the MH() functionality in Turing to specify the conditional posterior as a proposal and force an acceptance rate of 1 for the parameters whose conditional posteriors I know? I will use either MH() or NUTS() for the remaining parameters. As far as I can tell it is possible to specify the proposal for MH() but I did not see an obvious way to force Turing to accept the proposal to emulate traditional Gibbs sampling.

Thank you for your time.

Best,
Aaron

[bertschi]

@aws

My question is, can I use the MH() functionality in Turing to specify the conditional posterior as a proposal and force an acceptance rate of 1 for the parameters whose conditional posteriors I know? I will use either MH() or NUTS() for the remaining parameters. As far as I can tell it is possible to specify the proposal for MH() but I did not see an obvious way to force Turing to accept the proposal to emulate traditional Gibbs sampling.

From what I recall, Gibbs can be considered as a special case of MH which happens to accept with probability 1. Thus, no need to force it, i.e., if your proposal is correctly specified it should just always be accepted.