Bounds for Row-Aggregation in Linear Programming

Most applied linear programs reflect a certain degree of aggregation-either explicit or implicit-of some larger, more detailed problem. This paper develops methods for assessing the loss in accuracy resulting from aggregation. We showed previously that, when columns only are aggregated, a feasible solution to the larger problem can be recovered. This may not be the case under row-aggregation. Several reasonable measures of "accuracy loss" for this case are defined, and the bounds on these quantities derived. These results enable the modeler to compare and evaluate alternative approximate models of the same problem.