We describe a general framework for extracting knowledge-based energy function from a set of native protein structures. In this scheme, the energy function is optimal when there is least chance that a random structure has a lower energy than the corresponding native structure. We first show that subject to certain approximations, most current database-derived energy functions fall within this framework, including mean-field potentials, Z-score optimization, and constraint satisfaction methods. We then propose a simple method for energy function parametrization derived from our analysis. We go on to compare our method to other methods using a simple lattice model in the context of three different energy function scenarios. We show that our method, which is based on the most stringent criteria, performs best in all cases. The power and limitations of each method for deriving knowledge-based energy function is examined.