If there are known rate or magnitude bounds on the system inputs, then the maximum error magnitude can be obtained from l_1 norms of the sensitivity function (see, for example, the books of Dahleh & Diaz-Bobillo, Boyd & Barratt or Whidborne & Liu).
If there are both rate and magnitude bounds, then less conservative methods exist - there are some pointers in Boyd & Barratt or (for the continuous time case) see the works of Arunsawatwong.