yes, the dimention is 2 and I would prefer to do numerical intergration. But I am not really sure how to do that, because I am not sure how to write pyro.sample statement for observables in this case. And as I said I feel the way to do it is to have a new type of distribution where the log_prob is calculated by numerical integration (or any customed method). Could you give me some advices on that? Thanks.