How do we try to approximate o() (the error of the entire distribution)?

a) Train on a small sample of M.
b) Train on the entire dataset M.
c) Use a validation set separate from M.
d) Ignore the error term.