A Sequent Calculus for Urn Logic
Abstract
Approximately speaking, an urn model for firstorder logic is a model where the domain of quantification changes depending on the values of variables which have been bound by quantifiers previously. In this paper we introduce a modelchanging semantics for urnmodels, and then give a sequent calculus for urn logic by introducing formulas which can be read as saying that “after the individuals \(a_{1}, \ldots , a_{n}\) have been drawn, \(A\) is the case”.
Keywords
Urn logic Sequent calculus Nonstandard quantification Proof theory1 Introduction
Urn logic is a generalisation of firstorder logic introduced in Rantala (1975). The idea behind urn logic is to think of the domain of quantification in a way reminiscent of the eponymous urns of balls used in probability theory. The metaphor here is to think of quantifying over objects in a domain as being analogous to drawing balls from an urn. Typically in firstorder logic what we are doing is always quantifying over the same objects—our domain stays fixed every time we ‘draw’ from the domain. But, just as we can alter the contents of the urn based upon what we’ve drawn from it (for example, by simply not replacing it, or by putting more objects back into the urn of a kind based on what we’ve drawn out), we can also think about models for the quantifiers where the domain of quantification changes in the course of evaluating a formula.
Hintikka (1975) argues that urn models can be used to give a satisfactory model theoretic account of impossible worlds (and thus deal with issues of logical omniscience). Hintikka’s idea is, when giving our firstorder epistemic logic, to have two kinds of worlds—invariant, and almostinvariant worlds—the invariant ones being those which are used for determining logical truths, and all worlds (invariant and almostinvariant alike) being used in the interpretation of the epistemic operators. These almostinvariant worlds are envisaged as being urn models where the domain from which we can ‘draw’ is constant up until we have made \(n\) draws, the \(n\) being determined by the agent’s cognitive capacities. We will dwell no further on this potential application, outlined by Hintikka, of the notion of an urn model. Instead our focus will be largely technical (although we will mention another potential application at the end of the next section).
Relatively little work has been done on urn logic—notable exceptions being Cresswell (1982) and Olin (1978). In Cresswell (1982) a ‘classical analogue’ of Rantala’s urn logic is investigated, using a modification of the standard Tarskian semantics for firstorder logic as opposed to the gametheoretic semantics used by Rantala. Using these semantics Cresswell gives a Henkinstyle completeness proof for a range of urn logics with various constraints placed upon how we can draw individuals from the domain. Olin (1978), also using a classical version of urn Logic, has investigated the modeltheory of urn models, with particular attention to issues concerning satisfiability and categoricity.
Our plan here is as follows. In Sect. 2 we will describe the (standard/Tarskian) semantics for urn logic which we will be using, and comparing it to the semantic treatments given by Rantala, Cresswell and Olin. We will then proceed in Sect. 3 to give a sequent calculus for urn logic obtained by ‘decorating’ formulas so as to give us a way to keep track of changes in the domain. We then prove this sequent calculus to be sound and complete w.r.t. the semantics described in Sect. 2 before going on to show how to extend our sequent calculus in Sect. 5 so that it is sound and complete w.r.t. the class of urn models which Rantala is concerned with.
2 Semantics for Urn Logic
The semantic discussion of urn logic in Rantala (1975) is a modification of the gametheoretic semantics for firstorder logic from Hintikka (1979). Rather than working within the framework of gametheoretic semantics, though, we will instead work within a more familiar looking semantic framework, somewhat reminiscent of that used in Cresswell (1982) and Olin (1978). Rantala’s reasons (Rantala 1975, p. 351) for using the gametheoretic semantics is so that we can have a ‘modelchanging’ semantics in the sense discussed in Sect. 4 of Humberstone (2008), the semantics for urn logic not requiring any of the independentchoice style phenomena which gametheoretic semantics is particularly amenable to.
In the semantics for urn logic which we will be using, we will keep our model fixed, but evaluate formulas for truth relative, not just to a variable assignment as with standard Tarskian treatments of firstorder logic, but also relative to a draw.
Definition 1
Given an urn model \({\mathcal {M}} = \langle U, D, V\rangle \) a sequence \(\langle s_{1}, \ldots , s_{n}\rangle \) of objects from \(U\) is a draw from \({\mathcal {M}}\) iff (1) \(s_{1}\in D(\emptyset )\) and (2) \(s_{i+1} \in D(\langle s_{1},\ldots , s_{i}\rangle )\) for all \(1\le i \le n1\).
One of the effects of the condition \((Draw)\) is that whenever the domain determined by a sequence \(\underline{s}\) is nonempty, then \(\underline{s}\) is automatically a draw.
Definition 2
An assignment \(\nu \) is an \(x\) variant of \(\mu \) iff \(\nu (y) = \mu (y)\) for all variables \(y\) except (possibly) \(x\).

\(U = \{a, b\}\).

\(D(\emptyset ) = \{a, b\}\), \(D(a) = \{b\}\), \(D(b) = \{a\}\), and \(D(\underline{s}) = U\) for all sequences of length 2 or more.

\(V(F) = \{a\}\).
The above example also shows how our semantics differs from that of Cresswell and Olin. According to our semantics the truth of a quantified formula like \(\exists x\varphi \) in a model is evaluated relative to a sequence of ‘draws’ from the domain which, when we are considering sentences, will be determined by the values which have been assigned to variables which do not occur in the formula under consideration. By contrast, (again considering sentences) Cresswell and Olin evaluate formulas relative to the effect which such assigning has had upon the formula being evaluated, having the relevant domain of quantification be determined by the values assigned to the free variables in the formula in question. In the above case, for example, Cresswell and Olin would evaluate \(\forall xFx\) relative to \(D(\emptyset )\) (and consequently judge the above formula as valid).
2.1 Globally and Locally Determined Domains
This difference between the approach we are taking here, and that taken by Cresswell and Olin corresponds quite neatly to the distinction between the strongly and weakly exclusive interpretation of the quantifiers discussed in (Hintikka 1956, p. 230) and given a prooftheoretic treatment in Wehmeier (2004, 2009). The following terminology is more suggestive, though: call any semantics for the quantifiers where the domain quantified over is a nonconstant function of the formula under investigation one with locally determined domains. Similarly, we will call any semantics for the quantifiers in which the domain quantified over is a nonconstant function of the broader sententialcontext in which the quantified expression occurs one with globally determined domains.^{1} So the stronglyexclusive interpretation of the quantifiers (discussed in Wehmeier (2009)) is one where we have globally determined domains, and the weaklyexclusive interpretation (discussed in Wehmeier (2004)) one where we have locally determined domains. In the case of the semantics for urn logic discussed above, the semantics used by Cresswell and Olin is one with local domain determination, and that used by Rantala and myself one with global domain determination.

\((\star )\) In all of John’s classes he fails three students.

\((\star ^{\prime })\) Whatever office you go to, the supervisor is always unavailable.
In each of the cases we have discussed, the domain of the second quantified expression varies with the values introduced by the initial quantifier expression. Therefore, given what we have been assuming about the relation between semantic binding and syntactic binding outlined above [i.e. the Binding Assumption], it follows that there are bindable variables in the logical form of sentences containing quantifier expressions whose values are quantifier domains. (Stanley 2000, p. 421)
What this argument tells us, then, is that due to the fact that in order for sentences like \((\star )\) and \((\star ^{\prime })\) to be given the straightforward truth conditions mentioned above the outer quantified expression must semantically bind the inner quantified expression. According to Stanley’s Binding Assumption this can only be so if the outer quantified expression binds variables which are otherwise free in the inner quantified expression, and thus the true logical form of sentences like \((\star )\) and \((\star ^{\prime })\) must contain such bound variables, unlike the surface form given above.
Quantifiers with globally determined domains, like the urn quantifiers we will be investigating here, give us the technical machinery to undermine part of the plausibility of the Binding Assumption. If our quantifiers are to work just like standard Tarskian quantifiers then the only way for there to be semantic binding of quantifier expressions is, as the Binding Assumption tells us, by having freevariables in the semantically bound expression being syntactically bound by the expression which is doing the binding. If the only treatment of quantifiers we are countenancing is the Tarskian one then the Binding Assumption seems warranted, but in the presence of alternative ways of registering semantic binding without having syntactic binding this warrant is at least somewhat weakened.
To see that we can do without the Binding Assumption, note that if we were to use quantifiers with globally determined domains then we could get the correct interpretation for \((\star )\) using its surface logical form. For example, we will end up with the semantically bound reading of \((\star )\) whenever, for all \(a \in D(\emptyset )\cap \mathsf{John\hbox {}Class}\) we have \(D(a)\cap \mathsf{Student} = \mathsf{InClass}(a)\), where \(F\) is the extension of the predicate \(F\). Here the predicate \(\mathsf{InClass}\), which is implicit in the standard interpretation of \((\star )\), is used implicitly the semantically restrict our domain of quantification. We might think of the process of domain determination on this picture as being a topdown pragmatic process, based on an assessment of speaker intent and features of the conversational context.
This kind of semantic (as opposed to syntactic) binding also has the additional advantage of allowing us to accommodate deviant interpretations of \((\star )\) in a relatively straightforward manner. For example, suppose John finds teaching so stressful that for each class he teaches he finds that the only way he can unwind is by failing three students, and that he is thoroughly unconcerned which classes the students are in. In such a situation we can imagine an (understandably) confused interlocutor engaging with someone who has just uttered \((\star )\) querying whether John fails three students from each class he teaches and being told “No, John just gets very stressed and can only get relief by failing students. He’s not overly concerned which class they’re in!”. Inappropriate behaviour for a teacher, perhaps, but a perfectly legitimate reading of \((\star )\) all the same.
Our aim is not to necessarily endorse such a view here, but mostly to illustrate how being clearer about the various options we have for interpreting the quantifiers provides us with greater flexibility in giving the semantics of natural language without having to postulate additional bindable variables or hidden constituents.
3 A Sequent Calculus for Urn Logic
3.1 The System UC
Let us define the following sequent calculus, which we will call UC (mnemonic for ‘Urn Calculus’). Formally speaking our sequents are pairs of multisets of \(L^{U}\)formulas \(\langle \varGamma , \varDelta \rangle \), which we will write as \(\varGamma {>\!\!\!}\varDelta \). Throughout we will use \(\underline{s}\) and \(\underline{u}\) as variables over sequences of terms, \(A\) and \(B\) as \(L\)formulas. We will also occasionally find it useful to use \(\alpha \) and \(\beta \) as schematic letters for \(L^{U}\)formulas. Throughout \(t\) and \(u\) be schematic for terms, and \(a\) be schematic for freevariables (variously, in both cases, with and without subscripts). Where \(\underline{s}\) is the sequence \(\langle s_{1}, \ldots , s_{n}\rangle \) we will write \(\underline{s}\cdot t\) to denote the sequence \(\langle s_{1}, \ldots , s_{n}, t\rangle \).
3.2 Soundness
Let us abbreviate the claim that \(\underline{s}\) is a draw from \({\mathcal {M}}\) by writing \(draw^{{\mathcal {M}}}(\underline{s})\). Furthermore we will occasionally write \(\mu (\underline{s})\) where \(\mu \) is a variable assignment and \(\underline{s}\) is a sequence of terms from \(L\), understanding that as being the result of applying the variable assignment to each member of the sequence in turn (e.g. \(\mu (\langle a, b\rangle ) = \langle \mu (a), \mu (b)\rangle \)).
Definition 3
Definition 4
Say that a sequent \(\varGamma {>\!\!\!}\varDelta \) fails in an UrnModel \({\mathcal {M}}\) on an assignment \(\mu \) whenever for all \(\alpha \in \varGamma \) we have \({\mathcal {M}}\Vdash ^{+} \alpha [\mu ]\) and for all \(\beta \in \varDelta \) we have \({\mathcal {M}}\Vdash ^{} \beta [\mu ]\). Otherwise we will say that it holds in \({\mathcal {M}}\) on \(\mu \) and write \({\mathcal {M}}\models \varGamma {>\!\!\!}\varDelta [\mu ]\).
We will often find it convenient to write \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ]\) (or \({\mathcal {M}}\Vdash ^{}\varGamma [\mu ]\)) for some set of formulas \(\varGamma \), which should be understood as meaning that \({\mathcal {M}}\Vdash ^{+}\alpha [\mu ]\) (resp. \({\mathcal {M}}\Vdash ^{}\alpha [\mu ]\)) for all formulas \(\alpha \in \varGamma \), in which case the above definition tells us that a sequent fails in a model \({\mathcal {M}}\) on an assignment \(\mu \) whenever we have \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ]\) and \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ]\).
We will occasionally find the following lemma useful to appeal to in what follows.
Lemma 1
Proposition 1
(Soundness) If \(\varGamma {>\!\!\!}\varDelta \) is derivable in UC then, for all models \({\mathcal {M}}\) and assignments \(\mu \), we have \({\mathcal {M}}\models \varGamma {>\!\!\!}\varDelta [\mu ]\).
Proof
What we want to show is that if a sequent \(\varGamma {>\!\!\!}\varDelta \) is derivable in UC then, for all models \({\mathcal {M}}\) and all assignments \(mu\) we have either it is not the case that \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ]\) or it is not the case that \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ]\)—i.e. that a sequent is derivable if it holds in all models. Put more explicitly, the consequent of this conditional means that, for all models \({\mathcal {M}}\) and all assignments \(\mu \) that either we have \({\mathcal {M}}\not \Vdash ^{+}\alpha \) for some \(\alpha \in \varGamma \) or \({\mathcal {M}}\not \Vdash ^{}\beta \) for some \(\beta \in \varDelta \).
To prove this we will show that the axioms hold in all models, and show that each of the rules preserve the property of holding in a model (in particular by showing that if the conclusion of a rule fails in a model on an assignment, then (one of its) premise sequents also fails in a model on an assignment. Thus, given that the axioms hold in all models, and the rules preserve the property of holding in a model it follows by a simple induction on the length of derivations that if a sequent is derivable then it holds in all models, as desired.
Axiom Case: Suppose that \(\varGamma {>\!\!\!}\varDelta \) is an axiom. Then we have some formula \(Ft_1\ldots t_n\) and some term sequences \(\underline{s}\) and \(\underline{s^{\prime }}\) s.t. \(\underline{s}:Ft_1\ldots t_n\in \varGamma \) and \(\underline{s^{\prime }}:Ft_1\ldots t_n\in \varDelta \). If we have either \(\underline{s}\) or \(\underline{s^{\prime }}\) fail to be a draw then we’re done, as this will mean that either \({\mathcal {M}}\not \Vdash ^{+}\underline{s}:Ft_1\ldots t_n[\mu ]\) or \({\mathcal {M}}\not \Vdash ^{}\underline{s^{\prime }}:Ft_1\ldots t_n[\mu ]\). So suppose that we have both \(draw^{{\mathcal {M}}}(\underline{s})\) and \(draw^{{\mathcal {M}}}(\underline{s^{\prime }})\). It follows now that either (1)\({\mathcal {M}}, \underline{s}\models Ft_1\ldots t_n[\mu ]\) or (2)\({\mathcal {M}}, \underline{s}\not \models Ft_1\ldots t_n[\mu ]\). If (1) then we have \({\mathcal {M}}, \underline{s^{\prime }}\models Ft_1\ldots t_n[\mu ]\) by Lemma 1, and so \({\mathcal {M}}\not \Vdash ^{}\underline{s^{\prime }}:Ft_1\ldots t_n[\mu ]\). The case for (2) follows similarly.
We now cover some representative cases of the rules.
\(\lnot \) cases: For \((\lnot L)\) suppose that we have \(\varGamma , \underline{s}:\lnot A{>\!\!\!}\varDelta \) failing in some model \({\mathcal {M}}\) on an assignment \(\mu \). That is we have \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ]\), \({\mathcal {M}}\Vdash ^{+} \underline{s}:\lnot A[\mu ]\) and \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ]\). As we have \({\mathcal {M}}\Vdash ^{+} \underline{s}:\lnot A[\mu ]\) it follows that \(draw^{{\mathcal {M}}}(\mu (\underline{s}))\) and \({\mathcal {M}},\mu (\underline{s})\models \lnot A\), and hence that \({\mathcal {M}}, \mu (\underline{s})\not \models A\), from which it follows that \({\mathcal {M}}\Vdash ^{} \underline{s}:A[\mu ]\), resulting in \(\varGamma {>\!\!\!}\underline{s}:A, \varDelta \) failing in \({\mathcal {M}}\) on \(\mu \) as desired. For \((\lnot R)\) suppose that we have \(\varGamma {>\!\!\!}\varDelta , \underline{s}:\lnot A\) failing in some model \({\mathcal {M}}\) on an assignment \(\mu \). That is, we have \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ]\), \({\mathcal {M}}\Vdash ^{} \underline{s}:\lnot A[\mu ]\) and \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ]\). From \({\mathcal {M}}\Vdash ^{} \underline{s}:\lnot A[\mu ]\) it follows that \(draw^{{\mathcal {M}}}(\mu (\underline{s}))\) and \({\mathcal {M}},\mu (\underline{s})\not \models \lnot A[\mu ]\), and hence that \({\mathcal {M}},\mu (\underline{s})\models A[\mu ]\) from which it follows that \({\mathcal {M}}\Vdash ^{+}\underline{s} A[\mu ]\), resulting in \(\varGamma , \underline{s}:A{>\!\!\!}\varDelta \) failing in \({\mathcal {M}}\) on \(\mu \) as desired.
\(\forall \) cases: For \((\forall \hbox {L})\) suppose that we have \(\varGamma , \underline{s}:\forall x A(x){>\!\!\!}\varDelta \) failing in a model \({\mathcal {M}}\) on an assignment \(\mu \). That is to say, we have \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ]\), \({\mathcal {M}}\Vdash ^{+} \underline{s}:\forall x A(x)[\mu ]\) and \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ]\). As we have \({\mathcal {M}}\Vdash ^{+} \underline{s}:\forall x A(x)[\mu ]\) it follows that \(draw^{{\mathcal {M}}}(\mu (\underline{s}))\) and \({\mathcal {M}},\mu (\underline{s})\models \forall x A(x)[\mu ]\). So for all \(e \in D(\underline{s})\) we have \({\mathcal {M}}, \mu (\underline{s})\cdot e\models A(x)[\mu ^{\prime }]\) where \(\mu ^{\prime }\) is an \(x\)variant of \(\mu \) for which \(\mu (x) = e\). Now by the side condition for the rule we know that for some formula \(B\) and (possibly empty) term sequence \(\underline{s^{\prime }}\) we have \(\underline{s}\cdot t\cdot \underline{s^{\prime }}: B\in \varGamma \cup \varDelta \) from which it follows that \(draw^{{\mathcal {M}}}(\mu (\underline{s}\cdot t\cdot \underline{s^{\prime }}))\), and thus that \(draw^{{\mathcal {M}}}(\mu (\underline{s}\cdot t))\) by the (Draw) condition. Thus, letting \(\mu ^{\prime }(x)=\mu (t)\) it follows that \({\mathcal {M}}, \mu (\underline{s}\cdot x)\models A(x)[\mu ^{\prime }]\). As \(\mu ^{\prime }\) is an \(x\)variant of \(\mu \) it follows that \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ^{\prime }]\) and \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ^{\prime }]\). Further, as \(\mu ^{\prime }(x)=\mu ^{\prime }(t)\) we have \({\mathcal {M}}\Vdash ^{+}\underline{s}\cdot t:A(t)[\mu ^{\prime }]\), from which we have \(\varGamma , \underline{s}\cdot t:A(t){>\!\!\!}\varDelta \) failing in \({\mathcal {M}}\) on \(\mu ^{\prime }\) as desired.
For \((\forall R)\) suppose that we have \(\varGamma {>\!\!\!}\varDelta , \underline{s}:\forall x A(x)\) failing in some model \({\mathcal {M}}\) on an assignment \(\mu \). That is to say, we have \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ]\), \({\mathcal {M}}\Vdash ^{} \underline{s}:\forall x A(x)[\mu ]\) and \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ]\). It follows, then, that \(draw^{{\mathcal {M}}}(\mu (\underline{s}))\) and that there is an object \(e \in D(\mu (\underline{s}))\) s.t. \({\mathcal {M}},\mu (\underline{s})\cdot e\not \models A(x)[\mu ^{\prime }]\) where \(\mu ^{\prime }\) is an \(x\)variant of \(\mu \) for which \(\mu (x) = e\). Given that, by the side condition on \((\forall R)\), the term \(a\) does not occur in \(\varGamma \), \(\varDelta \) or \(\underline{s}:\forall x A(x)\) we can simply let our variable assignment assign \(e\) to \(a\) in which case we have \({\mathcal {M}}, \mu ^{\prime }(\underline{s})\cdot a \not \models A(a)[\mu ^{\prime }]\)–which given \(draw^{{\mathcal {M}}}(\mu (\underline{s}\cdot a))\) gives us \({\mathcal {M}}\Vdash ^{}\underline{s}\cdot a: A(a)[\mu ^{\prime }]\). Further, as \(\mu ^{\prime }\) is an \(x\)variant of \(\mu \) it follows that \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ^{\prime }]\) and \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ^{\prime }]\) from which it follows that \(\varGamma {>\!\!\!}\varDelta ,\underline{s}\cdot a:A(a)\) fails in \({\mathcal {M}}\) on \(\mu ^{\prime }\) as desired.
\((Draw)\) cases: For the case of \((Draw\hbox {}L)\) suppose that \(\varGamma , \underline{s}\cdot t:\varphi {>\!\!\!}\varDelta \) fails in some model \({\mathcal {M}}\) on an assignment \(\mu \), that is to say that we have \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ]\), \({\mathcal {M}}\Vdash ^{+} \underline{s}\cdot t: \varphi [\mu ]\) and \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ]\). It follows directly that \(draw^{{\mathcal {M}}}(\underline{s}\cdot t)\) and hence that \(draw^{{\mathcal {M}}}(\underline{s})\) by the (Draw) condition, and that \({\mathcal {M}}, \underline{s}\cdot t\models \varphi [\mu ]\). Thus applying Lemma 1 it follows that \({\mathcal {M}}, \underline{s}\models \varphi [\mu ]\), and so \({\mathcal {M}}\Vdash ^{+}\underline{s}:\varphi [\mu ]\) as desired. The case for \((Draw\hbox {}R)\) follows similarly.
\((dCut)\) case: For the case of the rule of \((dCut)\) suppose that the sequent \(\varGamma {>\!\!\!}\varDelta \) fails in the model \({\mathcal {M}}\) on an assignment \(\mu \)—i.e. that \({\mathcal {M}}\Vdash ^{+}\varGamma [\mu ]\), \({\mathcal {M}}\Vdash ^{}\varDelta [\mu ]\), and further that for some formula \(B\) and term sequent \(\underline{s^{\prime }}\) we have \(\underline{s}\cdot \underline{s^{\prime }}: B\in \varGamma \cup \varDelta \). From the above it follows by the (Draw) condition that \(draw^{{\mathcal {M}}}(\mu (\underline{s}))\) and so either \({\mathcal {M}}, \mu (\underline{s})\models A[\mu ]\) or \({\mathcal {M}}, \mu (\underline{s})\not \models A[\mu ]\)—i.e. that either \({\mathcal {M}}\Vdash ^{+}\underline{s}:A[\mu ]\) or \({\mathcal {M}}\Vdash ^{}\underline{s}:A[\mu ]\), meaning that one of the sequents \(\varGamma {>\!\!\!}\varDelta , \underline{s}:A\) or \(\varGamma , \underline{s}:A{>\!\!\!}\varDelta \) must fail in \({\mathcal {M}}\) on the variable assignment \(\mu \), as desired. \(\square \)
4 Completeness
What we are going to do here is to construct a reduction tree. We will use reduction trees a few times in what is to follow so we will define the procedure in general terms here so that we can use it in the completeness proofs below. Our presentation below is largely a generalisation of that given in Takeuti (1987).
4.1 Reduction Rules
What this means is that in order for us to apply this rule to a formula \(\underline{s}:\forall x A(x)\) appearing on the LHS of the sequent turnstile we require that there be some formula of the appropriate form appearing somewhere in the sequent. If the side conditions to a reduction rule are not met, then we do not extend the branch.
We will now explain how we can use reduction rules to construct a reduction tree. This is a tree, each node of which is a sequent. If the tip of a branch in the reduction tree is an axiom then we will say that that branch is closed, otherwise it is open.
Definition 5

Stage 0 At stage 0 we begin with \(\varGamma {>\!\!\!}\varDelta \) as the root node of the tree.

Stage \(n+1\) At stage \(n+1\) we apply all of the reduction procedures in \({\mathbb {R}}\) in the order given as many times as possible to every formula introduced at an earlier stage which occurs in the tip of an open branch.
Theorem 1
(Completeness) Let \({\mathcal {S}} = \varGamma {>\!\!\!}\varDelta \) be a sequent. Then either there is a \((dCut)\)free proof of \({\mathcal {S}}\), or there is an urn model in which \({\mathcal {S}}\) fails.
Proof
 \(U^{{\mathcal {S}}}\) is the set of all terms from \(L^{U}\).Table 1
The \({\mathbb {UC}}\)reduction rules
\((\lnot L\hbox {}\mathbf{Red})\):
\((\lnot R\hbox {}\mathbf{Red})\):
\((\wedge L\hbox {}\mathbf{Red})\):
\((\wedge R\hbox {}\mathbf{Red})\):
\((\forall L\hbox {}\mathbf{Red})\):
where for some \(B\) and some \(\underline{s^{\prime }}\) we have \(\underline{s}\cdot t\cdot \underline{s^{\prime }}:B\in \varGamma \cup \varDelta \)
\((\forall R\hbox {}\mathbf{Red})\):
where \(a\) is the first free variable which does not occur in \(\varGamma \cup \varDelta \cup \{\underline{s}:\forall x A(x)\}\).
\((Draw\ L\hbox {}\mathbf{Red})\):
where \(\varphi \) is quantifier free.
\((Draw\ R\hbox {}\mathbf{Red})\):
where \(\varphi \) is quantifier free.

\(D^{{\mathcal {S}}}(\emptyset ) = \{t  \text { for some}\, L\text {formula}\,A\text { and term sequence}\,\underline{s}, t\cdot \underline{s}{:}A\in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\} \)

\(D^{{\mathcal {S}}}(\underline{s}) = \{t  \text { for some}\, L\text {formula}\,A\text { and term sequence}\,\underline{s^{\prime }}, \underline{s}\cdot t\cdot \underline{s^{\prime }}{:}A\in {\varvec{\Gamma }} \cup {\varvec{\Delta }}\}\)

\(V^{{\mathcal {S}}}(F) = \{\langle t_1,\ldots , t_n\rangle  \text {for some sequence of terms}\,s_1, \ldots , s_n,\langle s_1,\ldots , s_n\rangle {:} Ft_1\ldots t_n\in {\varvec{\Gamma }}\}\)
\({\mathcal {M}}^{{\mathcal {S}}}\) is an urn model: It is easy to verify that \({\mathcal {M}}^{{\mathcal {S}}}\) satisfies the condition \((Draw)\) and thus is an urn model. Suppose that \(D^{{\mathcal {S}}}(\underline{s}\cdot s_{n}) \not = \emptyset \). Then there is some \(b\) and some (possibly empty) term sequence \(\underline{s^{\prime }}\) s.t. for some formula \(A\) we have \(\underline{s}\cdot s_{n}\cdot b\cdot \underline{s^{\prime }}:A\in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\). But this also means that there is some term sequence \(\underline{u}\) s.t. \(\underline{s}\cdot s_{n}\cdot \underline{u}:A\in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\), and so \(s_{n}\in D^{{\mathcal {S}}}(\underline{s})\), as desired.
We proceed by induction on the complexity of formulas, the basis case following from the definition of \(V^{{\mathcal {S}}}\), the cases for the boolean connectives are standard, the only case of interest being that of the quantifiers. We begin by noting that, by the construction of \(D^{S}\) that for every \(L\)formula \(B\) and term sequence \(\underline{s}\), if \(\underline{s}:B\in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\) then \(draw^{{\mathcal {M}}^{{\mathcal {S}}}}(i(\underline{s}))\).
This gives us the following corollary.
Corollary 1
The rule of \((dCut)\) is eliminable from UC.
5 Rantala’s Conditions
5.1 The Condition \((Ran2)\)
These three rules force every draw to always be extendable. To see this first note that all three rules are sound in urn models which satisfy \((Ran2)\).
Proposition 2
UCR2 is sound for the class of all urn models which satisfy \((Ran2)\).
Proof
For \((\forall L)\hbox {}{Ran2}\) suppose that we have \({\mathcal {M}}\Vdash ^{+}\underline{s}: \forall x A(x)[\mu ]\)— i.e. that \(draw^{{\mathcal {M}}}(\mu (\underline{s}))\) and \({\mathcal {M}}, \mu (\underline{s})\models \forall x A(x)[\mu ]\). From \(draw^{{\mathcal {M}}}(\mu (\underline{s}))\) we get that there must be some object \(e\in D(\mu (\underline{s}))\) by \((Ran2)\). So we can simply let \(\mu (a) = e\) and thus we have \({\mathcal {M}}, \mu (\underline{s}\cdot a)\models A(a)[\mu ]\) as desired.
For the \((Ran2L)\) case suppose that we have \({\mathcal {M}}\Vdash ^{+}\underline{s}:\varphi [\mu ]\). Then we have that \(\mu (\underline{s})\) is a draw from \({\mathcal {M}}\) and thus, by \((Ran2)\) that there is some object \(\mu (\underline{s})\cdot b\) is also a draw. Given that we also have \({\mathcal {M}}, \mu (\underline{s})\models \varphi [\mu ]\) it follows by Lemma 1 that \({\mathcal {M}}, \mu (\underline{s}\cdot a)\models \varphi [\mu ]\) where \(\mu (a) = e\), yielding \({\mathcal {M}}\Vdash ^{+}\underline{s}\cdot a: \varphi [\mu ]\) as desired. The case of \((Ran2R)\) follows similarly. \(\square \)
Lemma 2
Proof
We prove something stronger, namely that for every formula \(A\) there is some such atomic formula (or formulas) \(\varphi _{A}\) s.t. for some sequence of terms \(\underline{s^{\prime }}\) we have \(\underline{s}\underline{s^{\prime }}:\varphi _{A}\in \varGamma _{j}\cup \varDelta _{j}\).
To prove this we show that for every formula \(\underline{s}:A\in \varGamma _{k}\cup \varDelta _{k}\), there is a formula \(B\) of lower complexity such that, for some sequence of terms \(\underline{s^{\prime }}\), \(\underline{s}\underline{s^{\prime }}:B\in \varGamma _{k}\cup \varDelta _{k}\).To see this note that the reduction steps for \({\mathbb {UC}}\) guarantee that at each stage we can produce a formula of the required kind whenever the main connective of \(A\) is not a quantifier. In the case where the main connective of \(A\) is \(\forall \) either we can produce the appropriate formula using \((\forall R\hbox {}\mathbf{Red})\), or using \((\forall L\hbox {}\mathbf{Red})\) (if the conditions are met), or (and here is where we require \({\mathbb {UCR}}2\)) we can produce it using \((\forall L\hbox {}Ran2\hbox {}\mathbf{Red})\). \(\square \)
Proposition 3
UCR2 is complete for the class of all urn models which satisfy \((Ran2)\).
Proof
The proof is just like that of Theorem 1, except that we now create a \({\mathbb {UCR}}2\) reduction tree for our sequent \({\mathcal {S}}\) (Table 2). To verify that this results in a countermodel to \({\mathcal {S}}\) which satisfies condition \((Ran2)\) suppose that \(s_{n+1} \in D^{{\mathcal {S}}}(\langle s_1, \ldots , s_n\rangle )\). Then by the construction of \({\mathcal {M}}^{{\mathcal {S}}}\) there must be some formula of the form \(\langle s_1, \ldots , s_{n+1}\rangle \cdot \underline{s^{\prime }}: B\in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\). If any such formula has a nonempty sequence \(\underline{s^{\prime }}\) then we are done, so suppose that all such formulas are of the form \(\langle s_1, \ldots , s_{n+1}\rangle : B\). Such a formula must have been in the tree at some stage \(k\), and so we know by Lemma 2 that there is some later stage at which we have some sequence of terms \(\underline{s^{\prime \prime }}\) and some quantifierfree formula \(\varphi \) s.t. \(\langle s_1, \ldots , s_{n+1}\rangle \cdot \underline{s^{\prime \prime }}: \varphi \in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\). Again, if \(\underline{s^{\prime \prime }}\) is nonempty then we are done, so suppose it is empty. Then by (depending on whether it is in \({\varvec{\Gamma }}\) or \({\varvec{\Delta }}\)) either the \((Ran2\hbox {}L\hbox {}\mathbf{Red})\) or \((Ran2\hbox {}R\hbox {}\mathbf{Red})\) reductions we know that, for some term \(a\) that \(\langle s_1, \ldots , s_{n+1}\rangle \cdot a:\varphi \in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\), and so by construction of \({\mathcal {M}}^{{\mathcal {S}}}\) we have that for some term \(a\), \(a\in D^{{\mathcal {S}}}(\langle s_1, \ldots , s_{n+1}\rangle )\), as desired. \(\square \)
5.2 \((Ran1)\) and AlmostRantala Models
The \({\mathbb {UCR}}2\)reduction rules
\((\forall L\hbox {}Ran2\hbox {}\mathbf{Red})\):  
where \(a\) is the first free variable which does not occur in \(\varGamma \cup \varDelta \cup \{\underline{s}:\forall x A(x)\}\).  
\((Ran2L\hbox {}\mathbf{Red})\):  
where \(a\) is the first free variable which does not occur in \(\varGamma \cup \varDelta \cup \{\underline{s}:\varphi \}\), and \(\varphi \) is quantifier free.  
\((Ran2R\hbox {}\mathbf{Red})\):  
where \(a\) is the first free variable which does not occur in \(\varGamma \cup \varDelta \cup \{\underline{s}:\varphi \}\), and \(\varphi \) is quantifier free. 
Proposition 4
UCRant is sound for the class of all urn models which satisfy both \((Ran1')\) and \((Ran2)\).
Proof
What needs to be shown is that \((Ran1'\hbox {}L)\) and \((Ran1'\hbox {}R)\) are sound in models which satisfy \((Ran1')\). We treat just the cases of the left rule here (the right rule following similarly).
Suppose that a conclusion of an application of \((Ran1'\hbox {}L)\) were to fail in a model \({\mathcal {M}}\) on an assignment \(\mu \). Then in particular we would have that \({\mathcal {M}}\Vdash ^{+}\underline{s}\cdot t\cdot \underline{u}:\varphi [\mu ]\), and so \(draw^{{\mathcal {M}}}(\mu (\underline{s}\cdot t\cdot \underline{u}))\). and \({\mathcal {M}}, \mu (\underline{s}\cdot t\cdot \underline{u})\models \varphi [\mu ]\). By \((Draw)\) it follows that \(\mu (t)\in D(\underline{s})\) and so by \((Ran1')\) that \(\mu (t)\in D(\emptyset )\), and thus that \(\mu (t)\) is a draw from \({\mathcal {M}}\). Then by Lemma 1 it follows that \({\mathcal {M}}, \mu (t)\models \varphi [\mu ]\) and thus \({\mathcal {M}}\Vdash ^{+}t:\varphi [\mu ]\) as desired. \(\square \)
Proposition 5
UCRant is complete for the class of all urn models which satisfy \((Ran1^{\prime })\) and \((Ran2)\).
Proof
The \({\mathbb {UCRANT}}\)reduction rules
\({(Ran1^{\prime }\hbox {}L)\hbox {}\mathbf{Red}}\):  
where \(\varphi \) is quantifier free.  
\({(Ran1^{\prime }\hbox {}R)}\hbox {}\mathbf{Red}\):  
where \(\varphi \) is quantifier free. 
Suppose, then, that \(t\in D^{{\mathcal {S}}}(\underline{s})\) for some sequence \(\underline{s}\). Then by the construction of \({\mathcal {M}}^{{\mathcal {S}}}\) it follows that for some formula \(B\) and sequence of terms \(\underline{s^{\prime }}\) we have \(\underline{s}\cdot t\cdot \underline{s^{\prime }}:B\in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\). Then by Lemma 2 there is some sequence of terms \(\underline{u}\) and variable free formula \(\varphi \) s.t. \(\underline{s}\cdot t\cdot \underline{s^{\prime }} \cdot \underline{u}:\varphi \in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\). So by either \({(Ran1^{\prime }\hbox {}L)\hbox {}\mathbf{Red}}\) or \({(Ran1^{\prime }\hbox {}R)\hbox {}\mathbf{Red}}\) it follows that \(t:\varphi \in {\varvec{\Gamma }}\cup {\varvec{\Delta }}\), and so by the construction of \(D^{{\mathcal {S}}}\) that \(t\in D^{{\mathcal {S}}}(\emptyset )\) as desired. \(\square \)
6 Conclusion
One of the things which we have endeavoured to do here in our proof theoretic investigation of urn logic is to provide the first steps towards a more general framework for thinking about the quantifiers more generally. In particular the alterations to the language we made in Sect. 3 were made specifically in order to induce extra structure into our formulas (and thus into our sequents) which we could adjust without altering the operational rules for our quantifiers in order to validate further inferences. Our approach here has been directed specifically at the goal of giving a proof theoretic characterisation of urn logic, rather than a general investigation of this approach to the proof theory of the quantifiers, the more general study having to wait until another occasion.
Footnotes
 1.
Rather than limiting ourselves to sententialcontext determining the domain (so that the broader formula in which a quantified expression is embedded determines the relevant domain) we could instead let the domain simply be determined by some context of assessment \(c\), which is only in part determined by the sententialcontext. On this approach we can give a pleasant account of (semantically) restricted quantification. For example suppose \(c\) was a context in which the speaker has just opened the fridge—a context which makes the fridge’s contents especially salient—and consider the utterance of a sentence such as “There’s no beer” relative to this context. Then we could think of this as being of the form \(\lnot \exists x \mathsf{Beer}(x)\) with the quantifier ranging over the domain determined by \(c\)—namely the contents of the fridge. We will consider another similar example below.
 2.
It is important to recall at this point that the languages under consideration here do not contain individual constants. Up until now this has simply been a matter of convenience of exposition, here it matters.
Notes
Acknowledgments
I would like to thank Dave Ripley, Greg Restall and Lloyd Humberstone for their helpful comments and suggestions. I’m also grateful to an anonymous referee fro this journal who provided a number of invaluable comments, and noticed a rather serious (but thankfully repairable) problem in an earlier version of this paper.
References
 Cresswell, M. (1982). Urn models: A classical exposition. Studia Logica, 41, 109–130.CrossRefGoogle Scholar
 Hintikka, J. (1956). Identity, variables, and impredicative definitions. Journal of Symbolic Logic, 21, 225–245.CrossRefGoogle Scholar
 Hintikka, J. (1975). Impossible possible worlds vindicated. Journal of Philosophical Logic, 4, 475–484.CrossRefGoogle Scholar
 Hintikka, J. (1979). Quantifiers in logic and quantifiers in natural language. In E. Saarinen (Ed.), Gametheoretical semantics (pp. 27–47). Dordrecht: D. Reidel Publishing Co.Google Scholar
 Humberstone, L. (2008). Can every modifier be treated as a sentence modifier? Philosophical Perspectives, 22, 241–275.CrossRefGoogle Scholar
 Olin, P. (1978). Urn models and categoricity. Journal of Philosophical Logic, 7, 331–345.CrossRefGoogle Scholar
 Rantala, V. (1975). Urn models: A new kind of nonstandard model for firstorder logic. Journal of Philosophical Logic, 4, 445–474.CrossRefGoogle Scholar
 Stanley, J. (2000). Context and logical form. Linguistics and Philosophy, 23, 391–434.CrossRefGoogle Scholar
 Takeuti, G. (1987). Proof theory. Amsterdam: NorthHolland.Google Scholar
 Wehmeier, K. (2004). Wittgensteinian predicate logic. Notre Dame Journal of Formal Logic, 45(1), 1–11.CrossRefGoogle Scholar
 Wehmeier, K. (2009). On ramsey’s ‘silly delusion’ regarding tractatus 5.53. In G. Primiero & S. Rahman (Eds.), Acts of knowledgehistory, philosophy and logic (pp. 353–368). London: College Publications.Google Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.