Thank you Professor for your reply. It makes sense now. I think trying to consider the space of hypothesis "actually explored" is not that useful, as you said, the space of possible hypothesis
is independent of
and
and the resultant bound is a much more general and useful characterization of the learning model.