We show that a fully polynomial time approximation scheme given
for an optimization problem can always be simply modified to a
polynomial time algorithm solving the problem optimally if the
computation model is the deterministic Turing Machine or the
logarithmic cost RAM and if the range of the error bound is the
rational numbers or a subset (0,b).
Moreover, we prove that a P-reduction is not necessarily an
A-reduction for some suitable error bound transformation but we
give a sufficient criteria.