A new method of sensitivity analysis for mixed integer/linear programming (
MILP) is derived from the idea of inference duality. The inference dual of
an optimization problem asks how the optimal value can be deduced from the
constraints. In MILP, a deduction based on the resolution method of theorem
proving can be obtained from the branch-and-cut tree that solves the prima
l problem. One can then investigate which perturbations of the problem leav
e this proof intact. On this basis it is shown that, in a minimization prob
lem, any perturbation that satisfies a certain system of linear inequalitie
s will reduce the optimal value no more than a prespecified amount. One can
also give an upper bound on the increase in the optimal value that results
from a given perturbation. The method is illustrated on two realistic prob
lems.