I am happy leaving it to the user to enforce constraints they require, rather than having explicitly constrained distributions, however:
We need a Log/Exp bijection
We should ensure that Distribution.log_prob calculations involving arctanh and log return -jnp.inf where appropriate (for samples out of support), rather than jnp.nan.
We should have an example of learning a distribution with bounded support
I am happy leaving it to the user to enforce constraints they require, rather than having explicitly constrained distributions, however:
Log/ExpbijectionDistribution.log_probcalculations involving arctanh and log return-jnp.infwhere appropriate (for samples out of support), rather thanjnp.nan.