I loved this post, thanks for writing it. I get the argument why one shouldn't expect local minima in very high dimensions. But then, what's wrong with the informal argument that there has to be a minimum because, well, the expected loss cannot be negative?