We present a new approach to inferring a probability distribution which is incompletely specified by a number of linear constraints. We argue that the currently most popular approach of entropy maximization depends on a ``constraints as knowledge'' interpretation of the constraints, and that a different ``constraints as data'' perspective leads to a completely different type of inference procedures by statistical methods. With statistical methods some of the counterintuitive results of entropy maximization can be avoided, and inconsistent sets of constraints can be handled just like consistent ones. A particular statistical inference method is developed and shown to have a nice robustness property.
Get the Full paper