Comment by MiroF

I haven't read this paper yet, but to answer your question:

because bucket choice is a discrete decision - and discrete decisions are hard to pass gradients through