Zero-one law and definability of linear order

§1. Introduction. A logic ℒ has a limit law, if the asymptotic probability of every query definable in ℒ converges. It has a 0–1-law if the probability converges to 0 or 1. The 0–1-law for first-order logic on relational vocabularies was independently found by Glebski et al. [6] and Fagin [5]. Later it has been shown for many other logics, for instance for fragments of second order logic [12], for finite variable logic [13] and for FO extended with the rigidity quantifier [3]. Lynch [14] has shown a limit law for first-order logic on vocabularies with unary functions.We say that two formulas or two logics are almost everywhere equivalent, if they are equivalent on a class of structures whose asymptotic probability measure is one [7]. A 0–1-law is usually proved by showing that every quantifier of the logic has almost everywhere quantifier elimination, i.e., every formula with just one quantifier in front of it is almost everywhere equivalent to a quantifier-free formula. Besides proving 0–1-law, this implies that the logic is (weakly) almost everywhere equivalent to first-order logic.The aim of this paper is to study, whether a logic with a 0–1-law can have greater expressive power than FO in the almost everywhere sense and to what extent. In particular, we are interested on the definability of linear order. Because a 0–1-law determines the almost everywhere expressive power of the sentences of the logic completely, but does not say anything about formulas explicitly, we have to assume some regularity on logics. We will therefore mostly consider extensions of first-order logic with generalized quantifiers.

Finite variable logic, stability and finite models

We will study complete Ln-theories and their models, where Ln is the set of first order formulas in which at most n distinct variables occur. Here, by a complete Ln-theory we mean a theory such that for every Ln-sentence, it or its negation is implied by the theory. Hence, a complete Ln-theory need not necessarily be complete in the usual sense. Our approach is to transfer concepts and methods from stability theory, such as the order property and counting types, to the context of Ln-theories. So, in one sense, we will develop some rudimentary stability theory for a particular class of (possibly) incomplete theories. To make the ‘stability theoretic’ arguments work, we need to assume that models of the complete Ln-theory T which we consider can be amalgamated in certain ways. If this condition is satisfied and T has infinite models then there will exist models of T which are sufficiently saturated with respect to Ln. This allows us to use some counting types arguments from stability theory. If, moreover, we impose some finiteness conditions on the number of Ln-types and the length of Ln-definable orders then a sufficiently saturated model of T will be ω-categorical and ω-stable. Using the theory of ω-categorical and ω-stable structures we derive that T has arbitrarily large finite models.A different approach to combining stability theory with finite model theory is made by Hyttinen in [9] and [10].