As a side note, I’m not sure if you are aware that your post has been plagiarized at this blog: http://www.yindawei.com/2014/12/05/metropolis-hastings-mcmc-when-the-proposal-and-target-have-differing-support/

I’m not sure if you care to confront the other blogger, but just want to put a word out there.

]]>This looks really nice. There is quite a lot of interest in probabilistic programming in the clojure community too, I’ve yet to try this out on Scala. I was hoping to pick your brains about the limitations of AD. It is my understanding that many AD libraries have difficulty when your likelihood involves a lot of linear algebra operatoins. For instance imagine a Gaussian process likelihood, where you would like to do inference on the variance and length scale of the GP. When the linear algebra operations concern matrices which are themselves functions of the hyper parameters, is it still possible for AD to compute the derivative? I don’t think many AD libraries can do this, but I also haven’t exhaustively tried them out. Do you have a feel for this?

]]>