Coherent lower previsions are general models of uncertainty in probabilitydistributions. They are often approximated by some less general models, such ascoherent lower probabilities or in terms of some finite set of constraints. Theamount of error induced by the approximations has been often neglected in theliterature, despite the fact that it can be quite substantial. The aim of thispaper is to provide a general method for estimating the exact degree of errorfor given approximations of coherent lower previsions on finite probabilityspaces. The information on the maximal error is especially useful in caseswhere the approximations require a lot of effort to be calculated. Our methodis based on convex analysis on the corresponding credal sets, which can berepresented as convex polyhedra. It provides the exact maximal possible amountof error for a given finite approximation of a coherent lower prevision. Analgorithm based on quadratic programming is also provided.
展开▼