Alexey Radul

The empirical frequencies bound remains correct for the case of the symmetric Dirichlet prior, but since that prior is so constrained, there is hope of being able to find something...

In fact, if there are at least two distinct results, the probability can't be more than 1/k^2. Why? By exchangeability, we can take the second trial to have produced a...

Now that #455 is done, the uncollapsed case looks like it will use the same bounds methods as the collapsed case, exposing them through the `weightBound` method of the `PosteriorAAALKernel`.

@vkmvkmvkmvkm Is this a priority? Presumably not for NIPS, since no one has asked for it; but in general? Opinions on the design?

Project management decision: What are we doing with this? What are the trigger events that cause us to reconsider this?

Some data on how helpful this is likely to be. I profiled the command ``` venture puma -e "assume x = normal(0, 1); observe normal(x, 1) = 2; infer mh(default,...

From conversation with @riastradh-probcomp: - Shared pointers may be more expensive than they look to the profiler, because they imply memory barriers and inter-core communication (even if the "application" is...

For the particular case of `PSP`, Lite has a better thought-out class hierarchy, with a handful of base classes with appropriate names that supply appropriate default implementations. Is it possible...

Not quite. - incorporate and unincorporate are always on, and do nothing by default. Else, they are expected to adjust the auxiliary state (to wit, every possible call to `incorporate`...

Looks reasonable.