The quantitative/qualitative watershed for rules of uncertain inference (Q2454645)

The qualitative approach to the formal analysis of uncertain reasoning is exemplified by the family \({\mathbf P}\) of consequence relations \(\vsim\) determined by the class of stoppered preferential models, as formulated by \textit{S. Kraus}, \textit{D. Lehmann} and \textit{M. Magidor} [Artif. Intell. 44, No. 1/2, 167--207 (1990; Zbl 0782.03012)], whereby \(a\vsim x\) iff every state \(s\) that is minimal among those satisfying \(a\) satisfies \(x\). The quantitative approach is exemplified by the family \({\mathbf O}\) of consequence relations determined by probability functions, \(p\), and thresholds, \(t\), whereby \(a\vsim x\) iff either \(p(a)= 0\) or \(p_a(x)\geq t\). This paper investigates significant similarities and differences between these two approaches. The paper is gratifyingly rich. A brief review like this cannot attempt to capture its wealth of results, or the clarity of its presentation; it can only recommend that the interested reader seek the paper itself. Part I contrasts the qualitative approach with the quantitative when the threshold, \(t\), could be any number within \([0, 1]\); Part II then looks at some threshold-sensitive rules. In Part I, Section 2 presents the families \({\mathbf P}\) and \({\mathbf O}\). The former satisfies familiar closure conditions on the relation \(\vsim \): reflexivity, right weakening, left classical equivalence, `very cautious monotony', if \(a\vsim x\wedge y\) then \(a\wedge x\vsim y\), the rule OR, if \(a\vsim x\) and \(b\vsim x\) then \(a\vee b\vsim x\), and the rule AND, if \(a\vsim x\) and \(a\vsim y\) then \(a\vsim x\wedge y\). The latter satisfies the first four of these, but only weakened versions of OR and AND, viz. WOR, if \(a\wedge b\vsim x\) and \(a\wedge \neg b\vsim x\) then \(a\vsim x\), which is equivalent to XOR, if \(a\vsim x\) and \(b\vsim x\) and \(\vdash\neg(a\wedge b)\) then \(a\vee b\vsim x\), and WAND, if \(a\vsim x\) and \(a\wedge\neg y\vsim y\) then \(a\vsim x\wedge y\). (Both WOR and WAND are derivable in \({\mathbf O}\).) Section 3 explores the role of `failsafe' conditions like \(a\wedge\neg y\vsim y\) in WAND. If a relation \(a\vapprox y\) is defined equivalent to \(a\wedge\neg y\vsim y\), then if \(\vsim \) satisfies all the rules of \({\mathbf O}\) then \(\vapprox\) satisfies all the rules of \({\mathbf P}\). Section 4 examines the power of the rules AND, OR and the rules of cautious monotony and cumulative transitivity in the framework of \({\mathbf O}\). AND is revealed as the `watershed' of the title that demarcates the two approaches, in the sense that if \({\mathbf O}\) is extended to include it then the others are derivable, but AND is not derivable from the others with \({\mathbf O}\). Section 5 raises the questions of representation and completeness for probabilistic consequence in terms of \({\mathbf O}\). This relation is not representable by any family of probabilistically sound finite-premise Horn rules, nor is there any set of finite-premise Horn rules that is complete with respect to all sound, possibly countably-premised, Horn rules. Section 6 turns to extensions of \({\mathbf P}\) and \({\mathbf O}\) to families of relations that satisfy closure conditions given by non-Horn rules, such as the (qualitative) family \({\mathbf R}\) determined by the class of all ranked stoppered preferential models that validates the rule of rational monotony (RM), if \(a\vsim x\) then \(a\vsim \neg b\) or \(a\wedge b\vsim x\), or the (quantitative) family \({\mathbf Q}\) that extends \({\mathbf O}\) with the rule of negation rationality (NR), if \(a\vsim x\) then if \(a\wedge b\vsim x\) or if \(a\wedge\neg b\vsim x\). A new Horn rule PREF(1), if \(a\vsim x\) and if \(a\vsim\neg x\) then if \(a\vsim\, \perp\), is introduced; this with RM provides a bridge between \({\mathbf Q}\) and the strictly stronger \({\mathbf R}\), in that \({\mathbf O}\) plus PREF(1) and RM is equivalent to \({\mathbf R}\). Part II of the paper extends the preceding results to investigate families of threshold-sensitive probabilistically sound rules. Each \(t\) in \([0, 1]\) determines a family of consequence relations \({\mathbf O}(t)\). In Part I, \(t\) could be any member of \([0, 1]\); here it is stipulated to be sufficiently high or to lie within a range of specified values, and the question is, what rules are probabilistically sound with respect to what thresholds? Section 7 considers Horn rules, and isolates a family of `preface' rules (so-called from their similarity to the preface paradox), for each \(n\geq 1\), PREF\((n)\) is: if \(a\vsim x_1\cdots a\vsim x_n\) and if \(a\vsim \neg(x_1\cdots x_n)\) then \(a\vsim\, \perp\). Such rules are sound on the qualitative approach. Probabilistically, they are sound when for all probability functions \(p\) and all thresholds \(t> n/(n+1)\). Section 8 generalizes those rules. Section 9 introduces some threshold-sensitive non-Horn rules, especially some that correspond to the lottery paradox. Section 10 concludes with some open questions, notably whether a representation theorem can be given for \({\mathbf O}\) plus NR and another `Archimedean rule', whether \({\mathbf O}\) is complete with respect to finite-premise Horn rules, and similar questions of completeness for threshold-sensitive probabilistic consequence relations.