Thanks for the post! Quick question about your last equation: if each h is a distribution over a coarser partition of W (rather than W), then how are we drawing w from h for the inner geometric expectation?
You draw a part, which is a subset of W, and thus has a probability according to Q.
Thanks for the post! Quick question about your last equation: if each h is a distribution over a coarser partition of W (rather than W), then how are we drawing w from h for the inner geometric expectation?
You draw a part, which is a subset of W, and thus has a probability according to Q.