I am building a model in numpyro, and I would like a multivariate normal distribution that is truncated such that no dimension is ever < a. Is there a straightforward way to do this?
you might find this helpful
This is perfect, thanks!
This is a slight change of direction, but what are the advantages and disadvantages of truncating a distribution vs subjecting it to something like a softmax/softplus?
Obviously it changes the shape of the distribution. The resultant function stays continuous/differentiable. It’s also less coding. Anything you can say a priori about the performance?
i think it’s impossible to say in general. it’s all about what modeling assumptions you want to make. if you’re happy with the distribution implicitly defined by some deterministic transformation you introduce, go for it