Applying an MCMC Step to Sample from Distribution with KL Divergence Term

Questions v5
11

No. The var argument to Potential is an arbitrary PyTensor graph that will be added directly to the model logp. For example this is a convoluted way of adding a normal RV with potential

import pymc as pm

with pm.Model() as m:
  # for a likelihood, value is a constant
  value = pm.Flat("value")  # no logp
  pm.Potential(
    "normal_logp", 
    (
      -0.5 * pm.math.pow((value - mu) / sigma, 2) 
      - pm.math.log(pm.math.sqrt(2.0 * np.pi)) 
      - pm.math.log(sigma),
    ),
)