Conditional probability in the light of qualitative belief change (Q535329)

The author reminds the reader on Borel's example to illustrate the need for allowing propositions of measure 0 to be used conditionally (what is the probability that a point lies in the Western hemisphere given that it lies on the equator? 1/2 is of course the expected answer). On this basis, there are various ways of defining the function that, given a \textit{consistent} proposition \(a\) with \(p(a)=0\), maps a proposition \(x\) to \(p(x|a)\), the probability of \(x\) given \(a\): set it to \(\mathbf 1\) (the constant function equal to 1), or define it as a proper Kolmogorov function for all such \(a\)s (Hosiasson-Lindenbaum system), or for no such \(a\) (unit system), or for some such \(a\)s (Popper). Carnap's approach is to avoid the issue, imposing that there be no such \(a\). A parallel is drawn with notions from belief revision: \(\mathbf 1\) is the counterpart to the expansion function that, given an initial belief set \(K\) and a consistent belief \(a\) that is inconsistent with \(K\), maps \((K,a)\) to the set of logical consequences of \(K\cup\{a\}\), that is, the whole language, whereas a Kolmogorov function is the counterpart to a revision function that, in the same situation, eliminates some beliefs from \(K\) in a way that guarantees that the resulting set is consistent with \(a\). The Hosiasson-Lindenbaum system can then be seen as \textit{radically revisionary}, whereas Popper's system is only \textit{variably} so. The relationships between these approaches and the van Frassen system, an axiomatic framework which is equivalent to adding \(\mathbf 1\) to Popper's system, are described and discussed. The author claims that no approach is right or wrong, but that, depending on the application, an approach might be more appropriate than the other, and the respective benefits of those approaches are discussed. Then the author reflects on the relationships between belief revision and conditional probability. He mentions the framework of \textit{screened revision}, in which a pre-processing step decides whether \(K\) should be revised in the face of \(a\) in case \(K\) is inconsistent with \(a\): this suggests to use \(p\) rather than \(\mathbf 1\) in the cases where \(a\) is ``too unbelievable to take seriously as a condition.'' He also discusses hyper-revisionary probability function frameworks in which the function that maps \(x\) to \(p(x|a)\) is replaced by a more plausible one in case experience reveals \(a\) but \(p(a)\) is very small. Transposed to the field of belief revision, this suggests that \(K\ast a\) should not always be defined as the set of logical consequences of \(K\cup\{a\}\) in case \(a\) is consistent with \(K\), and that \(K\) could be contracted when \(a\) is too improbable in the face of \(K\). Finally, \textit{proto-probability functions}, defined by purely order-theoretic conditions weaker than the axioms of the usual systems for conditional probability, are discussed and shown to follow from the axioms that define the Hosiasson-Lindenbaum, Popper, Carnap, and the unit systems. As proto-probability functions support Hawthorne's system of qualitative uncertain inference, this could provide still new perspectives on conditional probability. Comparisons with other plausibility measures are also sketched. All proofs and more technical considerations are grouped together in a long appendix.