Introduction
The Craig interpolation property states that whenever an implication φ → ψ \varphi\to\psi φ → ψ is valid, there exists a formula θ \theta θ , using only the propositional variables common to φ \varphi φ and ψ \psi ψ , such that both φ → θ \varphi\to\theta φ → θ and θ → ψ \theta\to\psi θ → ψ are valid [5 ] . Uniform interpolation strengthens this requirement: given a formula φ \varphi φ and a retained signature σ \sigma σ , one seeks a single formula over σ \sigma σ that captures all consequences of φ \varphi φ expressible in σ \sigma σ and variables not used in φ \varphi φ [37 ] . Both properties have been extensively studied in modal logic; see [46 ] for a recent introduction and surveys of interpolation and its applications.
The traditional study of interpolation is qualitative: a logic either enjoys an interpolation property or it does not. While fundamental, this distinction provides little information about the prevalence of interpolation failure. A logic may fail interpolation because of a small collection of exceptional configurations, or failures may be widespread among its formulas. This raises a natural quantitative question:
How frequent are formulas, or pairs of formulas, for which interpolants exist?
In this paper, we initiate a systematic investigation of this question for quasi-normal propositional modal logics. We introduce densities of uniform and Craig interpolants by counting formulas modulo logical equivalence and study their asymptotic behaviour as the number of retained or forgotten propositional variables grows. We also investigate an alternative counting method that identifies formulas after existential quantification of the forgotten variables. Our results show that the choice of counting method matters. After existential quantification the densities can exhibit substantially different behaviour, revealing distinctions between modal logics that are invisible under ordinary counting.
To obtain a finite counting space, we fix two disjoint finite signatures σ \sigma σ and ρ \rho ρ of retained and forgotten propositional variables, respectively, and a modal depth bound d d d . At bounded depth, formulas over a finite signature form only finitely many equivalence classes. For locally tabular logics, the depth bound can be dropped.
Formulas of depth at most d d d over σ ∪ ρ \sigma\cup\rho σ ∪ ρ can be identified modulo logical equivalence with subsets of the finite set of complete ( σ ∪ ρ , d ) (\sigma\cup\rho,d) ( σ ∪ ρ , d ) -types realised at distinguished worlds of Kripke models. We call these types ( σ ∪ ρ , d ) (\sigma\cup\rho,d) ( σ ∪ ρ , d ) -atoms. We sample formulas uniformly modulo logical equivalence by choosing a subset of these atoms uniformly at random. Writing n = ∣ σ ∣ n=|\sigma| n = ∣ σ ∣ and m = ∣ ρ ∣ m=|\rho| m = ∣ ρ ∣ , we define the uniform interpolation density
UnifInt n , m d \operatorname{UnifInt}^{d}_{n,m} UnifInt n , m d
as the proportion of these formulas that admit a uniform interpolant over σ \sigma σ .
For Craig interpolation, we work with pairs of formulas φ , ψ \varphi,\psi φ , ψ over signatures σ ∪ ρ \sigma\cup\rho σ ∪ ρ and σ ∪ ρ ′ \sigma\cup\rho' σ ∪ ρ ′ , respectively, such that ∣ ρ ∣ = ∣ ρ ′ ∣ |\rho|=|\rho'| ∣ ρ ∣ = ∣ ρ ′ ∣ . We consider pairs for which φ ∧ ψ \varphi\land\psi φ ∧ ψ is inconsistent and ask whether their inconsistency can be witnessed by a formula over the common signature σ \sigma σ . This is equivalent to the usual formulation of Craig interpolation for valid implications. We define D C r a i g n , m d \mathrm{DCraig}^{d}_{n,m} DCraig n , m d as the proportion of such inconsistent pairs admitting a Craig interpolant.
A first observation of the paper is that these densities can be decomposed into densities over σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cells, where a σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell is a set of ( σ ∪ ρ , d ) (\sigma\cup\rho,d) ( σ ∪ ρ , d ) -atoms which cannot be distinguished, up to depth d d d , using the language over σ \sigma σ alone. We prove that interpolation failure is local to cells and the global densities can be computed as products of local densities:
U n i f I n t n , m d = ∏ C U n i f I n t C , D C r a i g n , m d = ∏ C D C r a i g C , \mathrm{UnifInt}^{d}_{n,m}=\prod_C\mathrm{UnifInt}_C,\qquad\mathrm{DCraig}^{d}_{n,m}=\prod_C\mathrm{DCraig}_C, UnifInt n , m d = C ∏ UnifInt C , DCraig n , m d = C ∏ DCraig C ,
where C C C ranges over the σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cells. This basic cell reduction principle underlies all of our asymptotic results.
We begin with the regime in which ρ \rho ρ is fixed and σ \sigma σ grows. Suppose that interpolation fails for σ 0 , ρ \sigma_0,\rho σ 0 , ρ , and modal depth d d d . For any σ ⊇ σ 0 \sigma\supseteq\sigma_0 σ ⊇ σ 0 of size n n n , the original ‘obstruction’ can be copied into 2 n 2^n 2 n many distinct cells of the same size as the original one. Each copy contributes a fixed factor strictly below 1 to the product decomposition, giving
lim n → ∞ U n i f I n t n , m d = 0 if uniform interpolation fails, \lim_{n\to\infty}\mathrm{UnifInt}^{d}_{n,m}=0\quad\text{if uniform interpolation fails,} n → ∞ lim UnifInt n , m d = 0 if uniform interpolation fails,
and
lim n → ∞ D C r a i g n , m d = 0 if Craig interpolation fails. \lim_{n\to\infty}\mathrm{DCraig}^{d}_{n,m}=0\quad\text{if Craig interpolation fails.} n → ∞ lim DCraig n , m d = 0 if Craig interpolation fails.
The opposite regime, in which σ \sigma σ is fixed and ρ \rho ρ grows, presents a different picture. As we show below, uniform interpolation admits a general convergence theorem in this regime, whereas the asymptotic behaviour of Craig interpolation remains open for arbitrary quasi-normal modal logics. For uniform interpolation, inside every cell there is a large family of atoms (2 m 2^m 2 m many) which cannot occur in any formula where uniform interpolation fails. Consequently the failure probability inside a cell is at most 2 − 2 m 2^{-2^m} 2 − 2 m . Since the number of cells depends only on σ \sigma σ , it follows that
lim m → ∞ U n i f I n t n , m d = 1. \lim_{m\to\infty}\mathrm{UnifInt}^{d}_{n,m}=1. m → ∞ lim UnifInt n , m d = 1.
For Craig interpolation, we focus on locally tabular quasi-normal modal logics, where no modal depth bound is needed. Within each cell C C C , inconsistent pairs correspond to ordered bicliques in an incompatibility graph H C H_C H C on its atoms. Unlike in the uniform case, the large family of universally compatible atoms does not suffice to control the density, since large bicliques may still occur. To address this, we introduce the canonical profile of H C H_C H C as its coarsest quotient preserving and reflecting compatibility. Craig interpolation holds precisely when every cell has a singleton canonical profile. Relaxing this condition, we say that a logic L L L has the uniformly bounded intra-cell incompatibility property (uniform BICIP ) if the sizes of canonical profiles are bounded by a constant depending only on L L L . We say that L L L has BICIP if the bound may also depend on n n n , but not on m m m . Both properties imply that, for every fixed n n n ,
lim m → ∞ D C r a i g n , m = 1. \lim_{m\to\infty}\mathrm{DCraig}_{n,m}=1. m → ∞ lim DCraig n , m = 1.
We show that every tabular quasi-normal modal logic has uniform BICIP.
A natural alternative way of counting is to abstract from distinctions between formulas that disappear once the forgotten or non-shared variables are existentially quantified away. Thus, instead of counting the number of φ \varphi φ itself, we count the number of ∃ p ⃗ . φ \exists\vec{p}.\varphi ∃ p . φ (up to logical equivalence). For Craig interpolation, the two components are treated separately, using ∃ p ⃗ . φ \exists\vec{p}.\varphi ∃ p . φ and ∃ p ′ ⃗ . ψ \exists\vec{p'}.\psi ∃ p ′ . ψ . We denote the corresponding densities by
U n i f I n t n , m Π , d and D C r a i g n , m Π , d . \mathrm{UnifInt}^{\Pi,d}_{n,m}\qquad\text{and}\qquad\mathrm{DCraig}^{\Pi,d}_{n,m}. UnifInt n , m Π , d and DCraig n , m Π , d .
We call these the uniform and Craig interpolation densities after existential quantification . Unlike ordinary modal equivalence, equivalence after existential quantification may depend on the chosen class of frames since two classes determining the same modal logic may induce different equivalence relations on existentially quantified formulas. Thus, counting up to equivalence after existential quantification depends not only on the logic but also on the chosen frame class. The natural choice is the class of all frames validating the logic, although we shall also consider well-motivated alternatives.
For this new counting measure, the cell decomposition remains valid, but the asymptotic behaviour can change substantially. When m m m is fixed and n n n grows, the previous asymptotic results continue to hold. The more interesting case is that of fixed n n n and growing m m m , where existential quantification can identify formulas that are distinct under ordinary modal equivalence.
Table 1 summarises the results for this regime. With the exception of the difference logic Diff (see below), our results on densities after existential quantification are shown for the class of all frames validating the logic. For tabular logics, there are only finitely many equivalence classes of existentially quantified formulas for any fixed n n n . Hence, no substantive asymptotic analysis is required: the density sequences U n i f I n t n , m Π \mathrm{UnifInt}^{\Pi}_{n,m} UnifInt n , m Π and D C r a i g n , m Π \mathrm{DCraig}^{\Pi}_{n,m} DCraig n , m Π eventually stabilise as m → ∞ m \to\infty m → ∞ . This contrasts with their ordinary counterparts, which converge to 1. The difference arises because large families of distinct atoms that determine convergence under ordinary counting collapse to a single equivalence class after existential quantification.
Table 1. Asymptotic behaviour of interpolant densities for fixed n = ∣ σ ∣ n = |\sigma| n = ∣ σ ∣ as m = ∣ ρ ∣ → ∞ m = |\rho| \to\infty m = ∣ ρ ∣ → ∞ . The superscript Π \Pi Π denotes counting modulo logical equivalence after existential quantification over the forgotten or non-shared variables. A question mark indicates that convergence to 1 is open. A dash indicates that no universal results holds. The results for Diff marked with * hold for the class of difference frames.
In case studies, we investigate the locally tabular logics G r z [ Δ ] \mathrm{Grz}[\Delta] Grz [ Δ ] , for Δ ≥ 1 \Delta\ge1 Δ ≥ 1 , and the difference logic Diff. The former are determined by finite partial orders of depth at most Δ \Delta Δ . Craig interpolation fails for G r z [ Δ ] \mathrm{Grz}[\Delta] Grz [ Δ ] if and only if Δ ≥ 3 \Delta\ge3 Δ ≥ 3 [13 Chapter 8] . Counterexamples are, intuitively, rare as they require intricate incompatible arrangements of paths. By contrast, Diff, the logic of difference frames ( W , ≠ ) (W,\ne) ( W , = ) [49 , 40 ] , admits straightforward counterexamples to Craig interpolation enforcing models of finite sizes [28 ] .
This distinction is not directly visible under ordinary counting, where all the corresponding interpolation densities converge to 1. For G r z [ Δ ] \mathrm{Grz}[\Delta] Grz [ Δ ] , convergence of the Craig interpolation density follows from BICIP (and even uniform BICIP when Δ ≤ 3 \Delta\le3 Δ ≤ 3 ). For Diff, it follows from a more involved analysis of quotients of incompatibility graphs.
The distinction becomes apparent after existential quantification. For G r z [ Δ ] \mathrm{Grz}[\Delta] Grz [ Δ ] , the quantified uniform interpolation density still converges to 1, while the quantified Craig interpolation density remains bounded away from 0. By contrast, for Diff, both quantified densities converge to 0 over the class of difference frames. We note that not every frame validating Diff is a difference frame, but leave the asymptotic analysis over the full class of Diff-validating frames for future work.
Related work
Quantitative and asymptotic questions about logical properties have been studied along two broadly independent lines; the density of interpolant existence considered here sits at their intersection but does not seem to have been addressed before. Also related is work on the complexity of interpolant existence for logics without Craig or uniform interpolation and work on repairing Craig interpolation minimally.
Density of truth. Closest in method is the density of truth programme, which fixes a signature and asks for the fraction of length-n n n formulas that are tautologies as n → ∞ n \to\infty n → ∞ [36 , 50 , 25 , 27 , 9 , 10 ] . Of particular relevance, Kostrzycka and Zaionc compute the density of truth for the { → , □ } \{\to,\Box\} { → , □ } fragments of S5 and S4 [26 ] . Related fragment computations include Dummett’s logic and monadic intermediate logics [23 , 24 ] . In contrast to our approach, counting is not done modulo logical equivalence in this programme. In fact, since all tautologies are logically equivalent, this assumption would trivialise the density of truth as one would count only one tautology.
Zero-one laws. The older strand concerns randomness over structures rather than formulas. Glebskii et al. and Fagin established that every first-order sentence holds in almost all or almost no finite structures [16 , 8 ] . In the modal setting, Halpern and Kapron proved zero-one laws for model validity in K, T, S4, S5 [19 ] ; see also [30 , 17 , 20 ] . More recently Verbrugge established zero-one laws for provability logic and Grzegorczyk’s logic [47 ] . Our results are analogous in spirit to zero-one laws, although our probability spaces consist of formulas modulo logical equivalence rather than finite structures.
Deciding Interpolant Existence. The property whose density we measure has been studied qualitatively, as a decision problem, when Craig or uniform interpolation fails. For decidability of Craig interpolant existence, there are results for modal logics with nominals [2 , 21 ] , decidable fragments of FO [22 ] , difference logic and weak K4 [28 ] , and logics extending K4.3 [29 ] . Decidability of uniform interpolant existence has mainly been investigated in description logic [15 , 31 ] .
Repairing Craig Interpolation. Another way to deal with the failure of interpolation is to restore it by extending the expressive power of the language [1 ] . This raises the natural question of whether there is a minimal extension, in a suitable sense, that has interpolation, and if so, what it is. This question has been studied for, among others, modal logic with nominals [44 ] , linear temporal logic [14 ] , and fragments of first-order logic [45 ] .
Preliminaries
We introduce basic notation for modal logic [3 , 18 ] . Fix a countably infinite set P r o p Prop P r o p of propositional variables . A signature τ \tau τ is a finite subset of P r o p Prop P r o p . The language ML of modal formulas is built from propositional variables using ⊤ \top ⊤ , ¬ \neg ¬ , ∧ \land ∧ and the modal operator □ \Box □ ; we use standard connectives ⊥ \bot ⊥ , ∨ \lor ∨ and → \to → as abbreviations and set ◊ φ : = ¬ □ ¬ φ \Diamond\varphi:= \neg\Box\neg\varphi ◊ φ := ¬ □ ¬ φ . We also use □ ≤ d \Box^{\le d} □ ≤ d defined by taking ⋀ i = 0 d □ i φ \bigwedge_{i=0}^{d}\Box^{i}\varphi ⋀ i = 0 d □ i φ , where □ i \Box^{i} □ i stands for i i i operators □ \Box □ . We write sig ( φ ) \operatorname{sig}(\varphi) sig ( φ ) for the signature of φ \varphi φ , the finite set of propositional variables occurring in φ \varphi φ , and md ( φ ) \operatorname{md}(\varphi) md ( φ ) for the modal depth of φ \varphi φ , the maximal nesting of □ \Box □ . For a signature τ ⊆ P r o p \tau\subseteq Prop τ ⊆ P r o p and a depth d ≥ 0 d \ge0 d ≥ 0 , we put
M L τ d : = { φ ∈ M L ∣ sig ( φ ) ⊆ τ , md ( φ ) ≤ d } and M L τ : = ⋃ d ≥ 0 M L τ d . \mathrm{ML}_{\tau}^{d}:=\{\varphi\in\mathrm{ML}\mid\operatorname{sig}(\varphi)\subseteq\tau,\ \operatorname{md}(\varphi)\le d\}\quad\text{and}\quad\mathrm{ML}_{\tau}:=\bigcup_{d\ge0}\mathrm{ML}_{\tau}^{d}. ML τ d := { φ ∈ ML ∣ sig ( φ ) ⊆ τ , md ( φ ) ≤ d } and ML τ := d ≥ 0 ⋃ ML τ d .
Modal formulas are interpreted in Kripke frames, or simply frames. A frame is a triple F = ( W , R , A ) \mathfrak{F}=(W,R,A) F = ( W , R , A ) , where W W W is a nonempty set of worlds , R ⊆ W × W R\subseteq W\times W R ⊆ W × W is a binary accessibility relation on W W W , and A ⊆ 2 W A\subseteq2^{W} A ⊆ 2 W is a set of admissible valuations on W W W , closed under the Boolean operations ∪ \cup ∪ , ∩ \cap ∩ and ∼ \mathord{\sim} ∼ and defined by taking F ( U ) = { w ∈ W ∣ w ′ ∈ U , for all w ′ ∈ W with w R w ′ } F(U)=\{w\in W\mid w' \in U,\ \text{for all }w'\in W\text{ with }wRw'\} F ( U ) = { w ∈ W ∣ w ′ ∈ U , for all w ′ ∈ W with w R w ′ } . If every set of worlds is admissible (that is, A = 2 W A=2^{W} A = 2 W ), then we omit A A A and write ( W , R ) (W,R) ( W , R ) . A model based on F \mathfrak{F} F is a pair M = ( F , V ) \mathfrak{M}=(\mathfrak{F},V) M = ( F , V ) with a valuation V : P r o p → A V:Prop\to A V : P r o p → A . A pointed model is a pair of the form ( M , w ) (\mathfrak{M},w) ( M , w ) with w ∈ W w\in W w ∈ W . The satisfaction relation between pointed models ( M , w ) (\mathfrak{M},w) ( M , w ) and formulas is defined as follows:
M , w ⊩ ¬ φ \mathfrak{M},w \Vdash\neg\varphi M , w ⊩ ¬ φ iff M , w ⊮ φ \mathfrak{M},w \nVdash\varphi M , w ⊮ φ ;
M , w ⊨ φ ∧ ψ \mathcal{M}, w \models\varphi\land\psi M , w ⊨ φ ∧ ψ iff M , w ⊨ φ \mathcal{M}, w \models\varphi M , w ⊨ φ and M , w ⊨ ψ \mathcal{M}, w \models\psi M , w ⊨ ψ ;
M , w ⊨ □ φ \mathcal{M}, w \models\Box\varphi M , w ⊨ □ φ iff w ∈ F ( { w ′ ∈ W ∣ M , w ′ ⊨ φ } ) w \in F(\{w' \in W \mid\mathcal{M}, w' \models\varphi\}) w ∈ F ({ w ′ ∈ W ∣ M , w ′ ⊨ φ }) .
A formula φ \varphi φ is valid in a pointed frame ( F , w ) (\mathfrak{F}, w) ( F , w ) , written F , w ⊨ φ \mathfrak{F}, w \models\varphi F , w ⊨ φ , if M , w ⊨ φ \mathcal{M}, w \models\varphi M , w ⊨ φ for every model M \mathcal{M} M based on F \mathfrak{F} F . For a class K \mathcal{K} K of pointed frames,
Log ( K ) : = { φ ∈ M L ∣ F , w ⊨ φ for all ( F , w ) ∈ K } \operatorname{Log}(\mathcal{K}) := \{\varphi\in\mathrm{ML} \mid\mathfrak{F}, w \models\varphi\text{ for all }(\mathfrak{F}, w) \in\mathcal{K}\} Log ( K ) := { φ ∈ ML ∣ F , w ⊨ φ for all ( F , w ) ∈ K }
is the logic determined by K \mathcal{K} K ; it is a quasi-normal modal logic. Unless stated otherwise, we assume that pointed frames ( F , w ) (\mathfrak{F}, w) ( F , w ) and models ( M , w ) (\mathcal{M}, w) ( M , w ) are rooted with root w w w in the sense that all v ∈ W v \in W v ∈ W are reachable from w w w along R R R -paths. Every quasi-normal modal logic is determined by rooted frames [3 ] .
In this article, many results are shown for classes of logics such as the class of all quasi-normal modal logics, the class of all tabular such logics, and the class of all such logics with (uniform) BICIP. However, we also consider a few concrete logics including
Grzegorczyk’s logic, Grz, determined by the class of finite partial orders (reflexive, transitive, and antisymmetric frames);
for Δ ≥ 1 \Delta\ge1 Δ ≥ 1 , the logics G r z [ Δ ] \mathrm{Grz}[\Delta] Grz [ Δ ] determined by finite partial orders of depth at most Δ \Delta Δ . Here, depth counts worlds in a strict chain; in particular, a singleton has depth 1, see [39 ] and [3 Section 8.6] .
The logic Diff determined by all difference frames ( W , ≠ ) (W, \ne) ( W , = ) which was first considered as the ‘logic of elsewhere’ [49 , 40 ] ; see also [7 ] .
Throughout we fix a quasi-normal logic L L L and the class K \mathcal{K} K of all pointed rooted frames in which L L L is valid. Let P M o d L \mathrm{PMod}_L PMod L denote the class of pointed models over the full set of variables P r o p \mathrm{Prop} Prop based on frames in K \mathcal{K} K . Semantic consequence is taken over this class: for φ , ψ ∈ M L \varphi, \psi\in\mathrm{ML} φ , ψ ∈ ML , we write φ ⊨ ψ \varphi\models\psi φ ⊨ ψ if every ( M , w ) ∈ P M o d L (\mathcal{M}, w) \in\mathrm{PMod}_L ( M , w ) ∈ PMod L with M , w ⊨ φ \mathcal{M}, w \models\varphi M , w ⊨ φ also satisfies M , w ⊨ ψ \mathcal{M}, w \models\psi M , w ⊨ ψ , and φ ≡ ψ \varphi\equiv\psi φ ≡ ψ if φ ⊨ ψ \varphi\models\psi φ ⊨ ψ and ψ ⊨ φ \psi\models\varphi ψ ⊨ φ . So, all entailments below are over L L L , but we omit the subscript to avoid clutter.
For pointed models ( M , w ) , ( M ′ , w ′ ) ∈ P M o d L (\mathcal{M}, w),(\mathcal{M}', w') \in\mathrm{PMod}_L ( M , w ) , ( M ′ , w ′ ) ∈ PMod L , a signature τ \tau τ and a depth d d d , put
( M , w ) ≡ τ d ( M ′ , w ′ ) iff M , w ⊨ φ ⇔ M ′ , w ′ ⊨ φ , for all φ ∈ M L τ d . (\mathcal{M}, w) \equiv_{\tau}^{d} (\mathcal{M}', w') \quad\text{iff}\quad\mathcal{M}, w \models\varphi\Leftrightarrow\mathcal{M}', w' \models\varphi,\text{ for all }\varphi\in\mathrm{ML}_{\tau}^{d}. ( M , w ) ≡ τ d ( M ′ , w ′ ) iff M , w ⊨ φ ⇔ M ′ , w ′ ⊨ φ , for all φ ∈ ML τ d .
Each ≡ τ d \equiv_{\tau}^{d} ≡ τ d is an equivalence relation on P M o d L \mathrm{PMod}_L PMod L . For every signature τ \tau τ and every d ≥ 0 d \ge0 d ≥ 0 , the set M L τ d \mathrm{ML}_{\tau}^{d} ML τ d is finite up to L L L -equivalence, and so ≡ τ d \equiv_{\tau}^{d} ≡ τ d has finitely many equivalence classes on P M o d L \mathrm{PMod}_L PMod L . This is the standard normal-form bound: over a finite set of variables and bounded depth, there are, up to provable equivalence in the minimal quasi-normal logic, finitely many formulas, and L L L -equivalence can only identify more.
Definition 1 (atom). The ≡ τ d \equiv_{\tau}^{d} ≡ τ d -class of ( M , w ) (\mathcal{M}, w) ( M , w ) is denoted [ M , w ] τ d [\mathcal{M}, w]_{\tau}^{d} [ M , w ] τ d and called a ( τ , d ) (\tau,d) ( τ , d ) -atom. The (finite) set of all ( τ , d ) (\tau,d) ( τ , d ) -atoms is denoted A t τ d \mathrm{At}_{\tau}^{d} At τ d ; in other words, each ≡ τ d \equiv_{\tau}^{d} ≡ τ d partitions P M o d L \mathrm{PMod}_L PMod L , and so A t τ d = P M o d L / ≡ τ d \mathrm{At}_{\tau}^{d} = \mathrm{PMod}_L/{\equiv_{\tau}^{d}} At τ d = PMod L / ≡ τ d .
Atoms are denoted by M , N , … \mathcal{M},\mathcal{N},\ldots M , N , … ; being equivalence classes, they are literally sets of pointed models, and we use set-theoretic notation for them accordingly: for instance, we write M ∩ N = ∅ \mathcal{M} \cap\mathcal{N} = \varnothing M ∩ N = ∅ to say that no pointed model lies in both M \mathcal{M} M and N \mathcal{N} N . Also, for φ ∈ M L τ d \varphi\in\mathrm{ML}_{\tau}^{d} φ ∈ ML τ d and a ( τ , d ) (\tau,d) ( τ , d ) -atom M \mathcal{M} M , we write
M ⊨ φ iff M , w ⊨ φ , for some (equivalently, all) ( M , w ) ∈ M \mathcal{M} \models\varphi\quad\text{iff}\quad\mathcal{M}, w \models\varphi,\text{ for some (equivalently, all) }(\mathcal{M}, w) \in\mathcal{M} M ⊨ φ iff M , w ⊨ φ , for some (equivalently, all) ( M , w ) ∈ M
because all members of M \mathcal{M} M agree on M L τ d \mathrm{ML}_{\tau}^{d} ML τ d -formulas.
Definition 2 (characteristic formulas). The characteristic formula of an atom M ∈ A t τ d \mathcal{M} \in\mathrm{At}_{\tau}^{d} M ∈ At τ d is
Φ ( M ) : = ⋀ { φ ∈ M L τ d ∣ M ⊨ φ } , \Phi(\mathcal{M}) := \bigwedge\{\varphi\in\mathrm{ML}_{\tau}^{d} \mid\mathcal{M} \models\varphi\}, Φ ( M ) := ⋀ { φ ∈ ML τ d ∣ M ⊨ φ } , which is a finite (up to equivalence) conjunction. Note that we omit sub- and superscripts for Φ \Phi Φ as they should be clear from M \mathcal{M} M .
The characteristic formula of an atom is true exactly on its pointed models:
Proposition 1. For any M ∈ A t τ d \mathcal{M} \in\mathrm{At}_{\tau}^{d} M ∈ At τ d , we have M , w ⊨ Φ ( M ) \mathcal{M}, w \models\Phi(\mathcal{M}) M , w ⊨ Φ ( M ) iff ( M , w ) ∈ M (\mathcal{M}, w) \in\mathcal{M} ( M , w ) ∈ M , for every ( M , w ) ∈ P M o d L (\mathcal{M}, w) \in\mathrm{PMod}_L ( M , w ) ∈ PMod L .
We extend the notion of characteristic formulas to sets D ⊆ A t τ d D \subseteq\mathrm{At}_{\tau}^{d} D ⊆ At τ d of atoms:
Φ ( D ) : = ⋁ M ∈ D Φ ( M ) . \Phi(D) := \bigvee_{\mathcal{M}\in D} \Phi(\mathcal{M}). Φ ( D ) := M ∈ D ⋁ Φ ( M ) .
The map D ↦ Φ ( D ) D \mapsto\Phi(D) D ↦ Φ ( D ) is a bijection between subsets of A t τ d \mathrm{At}_{\tau}^{d} At τ d and M L τ d \mathrm{ML}_{\tau}^{d} ML τ d modulo L L L -equivalence: every φ ∈ M L τ d \varphi\in\mathrm{ML}_{\tau}^{d} φ ∈ ML τ d is equivalent to Φ ( D φ ) \Phi(D_{\varphi}) Φ ( D φ ) for the unique D φ : = { M ∈ A t τ d ∣ M ⊨ φ } D_{\varphi} := \{\mathcal{M} \in\mathrm{At}_{\tau}^{d} \mid\mathcal{M} \models\varphi\} D φ := { M ∈ At τ d ∣ M ⊨ φ } . We may therefore identify a “formula of depth ≤ d \leq d ≤ d over τ \tau τ ” with a “subset of A t τ d \mathrm{At}_{\tau}^{d} At τ d ”.
Locally Tabular Logics
Call a logic L L L locally tabular if, over every signature, there are only finitely many formulas up to L L L -equivalence. For a locally tabular logic L L L , for each signature τ \tau τ , the descending chain of ≡ τ d \equiv_{\tau}^{d} ≡ τ d -equivalences on pointed models
≡ τ 0 ⊇ ≡ τ 1 ⊇ ⋯ \equiv_{\tau}^{0} \supseteq\equiv_{\tau}^{1} \supseteq\cdots ≡ τ 0 ⊇ ≡ τ 1 ⊇ ⋯
stabilises: there is a depth d ∗ d^{*} d ∗ such that ≡ τ d ∗ \equiv_{\tau}^{d^{*}} ≡ τ d ∗ coincides with ≡ τ d \equiv_{\tau}^{d} ≡ τ d , for all d ≥ d ∗ d \geq d^{*} d ≥ d ∗ . It is an open question whether for every locally tabular L L L there exists a d ∗ d^{*} d ∗ such that the sequence ≡ τ d \equiv_{\tau}^{d} ≡ τ d stabilises at d ∗ d^{*} d ∗ for all signatures τ \tau τ [42 Problem 2] , see also [43 , 41 ] . This is the case for all modal logics L L L of transitive frames and such an L L L is locally tabular if and only if its frames have bounded depth [3 Theorem 12.21] . In Diff, every modal formula is logically equivalent to a formula of modal depth at most two.
When dealing with locally tabular logics, we define τ \tau τ -atoms over the set of all formulas in M L τ \mathrm{ML}_{\tau} ML τ without the depth restriction, hence ( M , w ) ≡ τ ( M ′ , w ′ ) (\mathcal{M},w) \equiv_{\tau} (\mathcal{M}',w') ( M , w ) ≡ τ ( M ′ , w ′ ) if ( M , w ) ≡ τ d ( M ′ , w ′ ) (\mathcal{M},w) \equiv_{\tau}^{d} (\mathcal{M}',w') ( M , w ) ≡ τ d ( M ′ , w ′ ) for all d ≥ 0 d \geq0 d ≥ 0 , and a τ \tau τ -atom M \mathcal{M} M is defined as the ≡ τ \equiv_{\tau} ≡ τ -class of some ( M , w ) (\mathcal{M},w) ( M , w ) and then denoted [ M , w ] τ [\mathcal{M},w]_{\tau} [ M , w ] τ . Accordingly, the characteristic formula Φ ( M ) \Phi(\mathcal{M}) Φ ( M ) is also defined over all formulas in M L τ \mathrm{ML}_{\tau} ML τ . Results obtained in the general case of bounded modal depth transfer to the locally tabular case by reading ≡ τ \equiv_{\tau} ≡ τ as ≡ τ d \equiv_{\tau}^{d} ≡ τ d , where d d d is such that ≡ τ d ′ = ≡ τ d \equiv_{\tau}^{d'} = \equiv_{\tau}^{d} ≡ τ d ′ = ≡ τ d for all d ′ ≥ d d' \geq d d ′ ≥ d ; such a d d d exists by local tabularity.
A logic L L L is tabular if L = Log ( K ) L = \operatorname{Log}(\mathcal{K}) L = Log ( K ) for a finite class of finite pointed frames K \mathcal{K} K . Every tabular logic is locally tabular, but the converse does not hold as witnessed, for example, by Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] for Δ ≥ 2 \Delta\geq2 Δ ≥ 2 and Diff.
When describing the structure of τ \tau τ -atoms, we refer to maximal consistent set of τ \tau τ -literals as τ \tau τ -colours; in other words, a τ \tau τ -colour is a set containing either p p p or ¬ p \neg p ¬ p , for each propositional variable p ∈ τ p \in\tau p ∈ τ . Also, given a set of τ \tau τ -colours c 1 , … , c k c_{1}, \ldots, c_{k} c 1 , … , c k , we denote
∇ ( c 1 , … , c k ) : = [ □ ⋁ i ⋀ c i ] ∧ [ ⋀ i ◊ ⋀ c i ] . \nabla(c_{1},\ldots,c_{k}) := \left[\Box\bigvee_{i}\bigwedge c_{i}\right]\land\left[\bigwedge_{i}\Diamond\bigwedge c_{i}\right]. ∇ ( c 1 , … , c k ) := [ □ i ⋁ ⋀ c i ] ∧ [ i ⋀ ◊ ⋀ c i ] .
We illustrate these definitions on the following example of a tabular logic.
Example 1 (3-fork). Let K \mathcal{K} K consist of three pointed frames ( ( W k , R k ) , w 0 ) ((W_{k},R_{k}),w_{0}) (( W k , R k ) , w 0 ) , k = 1 , 2 , 3 k = 1,2,3 k = 1 , 2 , 3 , where
W i = { w 0 , w 1 , … , w k } and w 0 R w i , for 1 ≤ i ≤ k ; W_{i} = \{w_{0},w_{1},\ldots,w_{k}\}\quad\text{and}\quad w_{0}R w_{i},\ \text{for }1 \leq i \leq k; W i = { w 0 , w 1 , … , w k } and w 0 R w i , for 1 ≤ i ≤ k ; the frame with three successors is depicted in Figure 1 a. It should be clear that there are three groups of τ \tau τ -atoms, with each successor having the same τ \tau τ -colour, one of two τ \tau τ -colours and one of three τ \tau τ -colours; see Figure 1 b, where we illustrate representative pointed models and the characteristic formulas of their atoms. One can easily verify that the number of τ \tau τ -atoms is
2 ∣ τ ∣ [ ( 2 ∣ τ ∣ 1 ) + ( 2 ∣ τ ∣ 2 ) + ( 2 ∣ τ ∣ 3 ) ] = 1 6 2 4 ∣ τ ∣ + 5 6 2 2 ∣ τ ∣ . 2^{|\tau|}\left[\binom{2^{|\tau|}}{1}+\binom{2^{|\tau|}}{2}+\binom{2^{|\tau|}}{3}\right]=\frac{1}{6}2^{4|\tau|}+\frac{5}{6}2^{2|\tau|}. 2 ∣ τ ∣ [ ( 1 2 ∣ τ ∣ ) + ( 2 2 ∣ τ ∣ ) + ( 3 2 ∣ τ ∣ ) ] = 6 1 2 4∣ τ ∣ + 6 5 2 2∣ τ ∣ . Obviously, atoms can contain pointed models with identical successors, but it will be useful for us to identify so-called standard pointed models for atoms without such duplicates (as depicted in Figure 1 b).
For locally tabular logic, atoms can always be represented by a single model. We use the well-known notion of p-morphism (or, equivalently, surjective bounded morphism) [18 ] .
Definition 3 (τ \tau τ -p-morphism). Let M = ( W , R , V ) \mathcal{M} = (W,R,V) M = ( W , R , V ) , M ′ = ( W ′ , R ′ , V ′ ) \mathcal{M}' = (W',R',V') M ′ = ( W ′ , R ′ , V ′ ) be models. A map f : W → W ′ f: W \to W' f : W → W ′ is a τ \tau τ -p-morphism from M \mathcal{M} M to M ′ \mathcal{M}' M ′ if, for all w , v ∈ W w,v \in W w , v ∈ W , u ′ ∈ W ′ u' \in W' u ′ ∈ W ′ , the following holds:
Figure 1. a) The 3-fork pointed frame and b) its three groups of τ \tau τ -atoms, for distinct τ \tau τ -colours c 1 , c 2 , c 3 c_1,c_2,c_3 c 1 , c 2 , c 3 , with their characteristic formulas.
(p1) M , w ⊨ p \mathcal{M}, w \models p M , w ⊨ p iff M ′ , f ( w ) ⊨ p \mathcal{M}', f(w) \models p M ′ , f ( w ) ⊨ p , for all p ∈ τ p \in\tau p ∈ τ ;
(p2) w R v wRv w R v implies f ( w ) R ′ f ( v ) f(w)R'f(v) f ( w ) R ′ f ( v ) ;
(p3) f ( w ) R ′ u ′ f(w)R'u' f ( w ) R ′ u ′ implies f ( u ) = u ′ f(u)=u' f ( u ) = u ′ for some u ∈ W u \in W u ∈ W with w R u wRu w R u .
A p-morphism of pointed models additionally maps the distinguished world to the distinguished world.
If f f f is a τ \tau τ -p-morphism of pointed models from ( M , w ) (\mathcal{M}, w) ( M , w ) to ( M ′ , w ′ ) (\mathcal{M}', w') ( M ′ , w ′ ) and ( M ′ , w ′ ) (\mathcal{M}', w') ( M ′ , w ′ ) is rooted, then (p3) implies that f f f is surjective. We use the following invariance property of p-morphisms [18 Corollary 16] .
Lemma 2. If f f f is a τ \tau τ -p-morphism from M \mathcal{M} M to M ′ \mathcal{M}' M ′ , then w w w and f ( w ) f(w) f ( w ) satisfy the same τ \tau τ -formulas, for every w ∈ W w \in W w ∈ W .
Definition 4 (τ \tau τ -standard pointed model). Let M \mathcal{M} M be a τ \tau τ -atom for a locally tabular logic L L L . The τ \tau τ -standard pointed model ( S M , w M ) (\mathcal{S}_{\mathcal{M}}, w_{\mathcal{M}}) ( S M , w M ) of M \mathcal{M} M is defined by taking the quotient of any pointed model ( M , w ) ∈ M (\mathcal{M}, w) \in\mathcal{M} ( M , w ) ∈ M by ≡ τ \equiv_\tau ≡ τ and discarding all worlds not reachable from w M w_{\mathcal{M}} w M .
On finite models, τ \tau τ -bisimulations characterise logical equivalence. Hence, the construction above is also known as the bisimulation quotient of a model [18 Definition 44] . We summarise its main properties.
Lemma 3. Let M \mathcal{M} M be a τ \tau τ -atom for a locally tabular logic L L L . Then the τ \tau τ -standard pointed model ( S M , w M ) (\mathcal{S}_{\mathcal{M}}, w_{\mathcal{M}}) ( S M , w M ) of M \mathcal{M} M is well-defined (it does not depend on the choice of ( M , w ) ∈ M (\mathcal{M}, w) \in\mathcal{M} ( M , w ) ∈ M ) and has the following properties:
( S M , w M ) ∈ M (\mathcal{S}_{\mathcal{M}}, w_{\mathcal{M}}) \in\mathcal{M} ( S M , w M ) ∈ M ;
( S M , w M ) (\mathcal{S}_{\mathcal{M}}, w_{\mathcal{M}}) ( S M , w M ) is extensional, that is, distinct worlds are not ≡ τ \equiv_\tau ≡ τ -equivalent;
a rooted model ( N , u ) (\mathcal{N}, u) ( N , u ) based on a frame for L L L is in M \mathcal{M} M iff there is a τ \tau τ -p-morphism from ( N , u ) (\mathcal{N}, u) ( N , u ) onto ( S M , w M ) (\mathcal{S}_{\mathcal{M}}, w_{\mathcal{M}}) ( S M , w M ) .
Uniform Interpolants
We now fix two disjoint signatures σ \sigma σ (retained) and ρ \rho ρ (forgotten) and a modal depth d ≥ 0 d \ge0 d ≥ 0 . Let φ ∈ M L σ ∪ ρ d \varphi\in\mathrm{ML}_{\sigma\cup\rho}^{d} φ ∈ ML σ ∪ ρ d . A formula θ ∈ M L σ \theta\in\mathrm{ML}_{\sigma} θ ∈ ML σ is a uniform σ \sigma σ -interpolant of φ \varphi φ if
φ ⊨ θ \varphi\models\theta φ ⊨ θ ;
for every χ ∈ M L \chi\in\mathrm{ML} χ ∈ ML with sig ( χ ) ∩ ρ = ∅ \operatorname{sig}(\chi) \cap\rho= \varnothing sig ( χ ) ∩ ρ = ∅ , if φ ⊨ χ \varphi\models\chi φ ⊨ χ , then θ ⊨ χ \theta\models\chi θ ⊨ χ .
A logic L L L has the uniform interpolation property if every formula has a uniform σ \sigma σ -interpolant for every choice of retained signature σ \sigma σ . For φ ∈ M L \varphi\in\mathrm{ML} φ ∈ ML , define the strongest σ \sigma σ -consequence of φ \varphi φ of depth ≤ d \le d ≤ d as follows:
Cons σ d ( φ ) : = ⋀ { ψ ∈ M L σ d ∣ φ ⊨ ψ } ; \operatorname{Cons}_{\sigma}^{d}(\varphi) := \bigwedge\{\psi\in\mathrm{ML}_{\sigma}^{d} \mid\varphi\models\psi\}; Cons σ d ( φ ) := ⋀ { ψ ∈ ML σ d ∣ φ ⊨ ψ } ;
this is a finite conjunction up to equivalence. The sets grow with depth: Cons σ e ( φ ) ⊨ Cons σ d ( φ ) \operatorname{Cons}_{\sigma}^{e}(\varphi) \models\operatorname{Cons}_{\sigma}^{d}(\varphi) Cons σ e ( φ ) ⊨ Cons σ d ( φ ) for e ≥ d e \ge d e ≥ d . Then a uniform σ \sigma σ -interpolant of φ \varphi φ , if it exists, is equivalent to Cons σ e ( φ ) \operatorname{Cons}_{\sigma}^{e}(\varphi) Cons σ e ( φ ) for some e e e : taking e e e at least its modal depth, the interpolant θ \theta θ is a σ \sigma σ -consequence of φ \varphi φ of depth ≤ e \leq e ≤ e , hence Conse σ e ( φ ) ⊨ θ \operatorname{Conse}^{e}_{\sigma}(\varphi) \models\theta Conse σ e ( φ ) ⊨ θ ; and θ \theta θ entails every ρ \rho ρ -free consequence, in particular, every conjunct of Conse σ e ( φ ) \operatorname{Conse}^{e}_{\sigma}(\varphi) Conse σ e ( φ ) , so θ ⊨ Conse σ e ( φ ) \theta\models\operatorname{Conse}^{e}_{\sigma}(\varphi) θ ⊨ Conse σ e ( φ ) . Therefore, non-existence means precisely that, for every e e e , the candidate Conse σ e ( φ ) \operatorname{Conse}^{e}_{\sigma}(\varphi) Conse σ e ( φ ) fails to entail some ρ \rho ρ -free consequence χ e \chi_e χ e of φ \varphi φ ; recording χ e \chi_e χ e over σ ∪ ρ ′ \sigma\cup\rho' σ ∪ ρ ′ with ρ ′ : = sig ( χ e ) ∖ σ \rho' := \operatorname{sig}(\chi_e) \setminus\sigma ρ ′ := sig ( χ e ) ∖ σ disjoint from ρ \rho ρ gives us the following working characterisation:
Lemma 4 (uniform interpolation criterion). A formula φ ∈ M L σ ∪ ρ d \varphi\in\mathrm{ML}^{d}_{\sigma\cup\rho} φ ∈ ML σ ∪ ρ d has no uniform σ \sigma σ -interpolant if and only if, for every e ≥ d e \geq d e ≥ d , there is a ρ \rho ρ -free formula χ e \chi_e χ e such that
φ ⊭ χ e and Conse σ e ( φ ) ⊭ χ e . \varphi\not\models\chi_e \qquad\text{and} \qquad\operatorname{Conse}^{e}_{\sigma}(\varphi) \not\models\chi_e. φ ⊨ χ e and Conse σ e ( φ ) ⊨ χ e .
Example 2. Consider the following two formulas for the 3-fork logic from Example 1:
φ = p ∧ q ∧ ∇ ( p ∧ q , p ∧ ¬ q , ¬ p ∧ q ) , ψ = p ∧ q ′ ∧ ∇ ( p ∧ q ′ , ¬ p ∧ q ′ , ¬ p ∧ ¬ q ′ ) . \begin{aligned}
\varphi&= p \land q \land\nabla(p \land q, p \land\neg q, \neg p \land q),\\
\psi&= p \land q' \land\nabla(p \land q', \neg p \land q', \neg p \land\neg q').
\end{aligned} φ ψ = p ∧ q ∧ ∇ ( p ∧ q , p ∧ ¬ q , ¬ p ∧ q ) , = p ∧ q ′ ∧ ∇ ( p ∧ q ′ , ¬ p ∧ q ′ , ¬ p ∧ ¬ q ′ ) .
The first requires two successors satisfying p p p and one satisfying ¬ p \neg p ¬ p , whereas the second requires one successor satisfying p p p and two satisfying ¬ p \neg p ¬ p . Each formula describes a single atom over its respective signature.
Let σ = { p } \sigma= \{p\} σ = { p } . Then φ \varphi φ has no uniform σ \sigma σ -interpolant: its strongest σ \sigma σ -consequence over σ \sigma σ is θ = p ∧ ∇ ( p , ¬ p ) \theta= p \land\nabla(p,\neg p) θ = p ∧ ∇ ( p , ¬ p ) . However, θ \theta θ is not a uniform σ \sigma σ -interpolant of φ \varphi φ because φ ⊨ ¬ ψ \varphi\models\neg\psi φ ⊨ ¬ ψ but θ ⊭ ¬ ψ \theta\not\models\neg\psi θ ⊨ ¬ ψ . Thus, no σ \sigma σ -formula captures all consequences of φ \varphi φ over signatures disjoint from { q } \{q\} { q } .
Cell Factorisation
We note that standard criteria for uniform interpolants in the literature are bisimulation-based [48 , 6 , 11 , 32 ] . For our purposes, it is more convenient to work with logical indistinguishability directly and use atoms. We first lift the relation ≡ σ d \equiv_{\sigma}^{d} ≡ σ d from pointed models to atoms (≡ σ d \equiv_{\sigma}^{d} ≡ σ d is used for both pointed models and atoms).
Definition 5 (lifted equivalence). Let M 1 ∈ A t σ ∪ ρ 1 d \mathcal{M}_1 \in\mathrm{At}^{d}_{\sigma\cup\rho_1} M 1 ∈ At σ ∪ ρ 1 d and M 2 ∈ A t σ ∪ ρ 2 d \mathcal{M}_2 \in\mathrm{At}^{d}_{\sigma\cup\rho_2} M 2 ∈ At σ ∪ ρ 2 d . We write M 1 ≡ σ d M 2 \mathcal{M}_1 \equiv_{\sigma}^{d} \mathcal{M}_2 M 1 ≡ σ d M 2 if ( M 1 , w 1 ) ≡ σ d ( M 2 , w 2 ) (\mathcal{M}_1,w_1) \equiv_{\sigma}^{d} (\mathcal{M}_2,w_2) ( M 1 , w 1 ) ≡ σ d ( M 2 , w 2 ) , for some ( M i , w i ) ∈ M i (\mathcal{M}_i,w_i) \in\mathcal{M}_i ( M i , w i ) ∈ M i , for i = 1 , 2 i=1,2 i = 1 , 2 .
Although the lifted relation ≡ σ d \equiv_{\sigma}^{d} ≡ σ d is defined by requiring some representatives to agree, it is in fact independent of the choice of representatives.
Proposition 5. Let M 1 ∈ A t σ ∪ ρ 1 d \mathcal{M}_1 \in\mathrm{At}^{d}_{\sigma\cup\rho_1} M 1 ∈ At σ ∪ ρ 1 d and M 2 ∈ A t σ ∪ ρ 2 d \mathcal{M}_2 \in\mathrm{At}^{d}_{\sigma\cup\rho_2} M 2 ∈ At σ ∪ ρ 2 d . If M 1 ≡ σ d M 2 \mathcal{M}_1 \equiv_{\sigma}^{d} \mathcal{M}_2 M 1 ≡ σ d M 2 , then ( M 1 , w 1 ) ≡ σ d ( M 2 , w 2 ) (\mathcal{M}_1,w_1) \equiv_{\sigma}^{d} (\mathcal{M}_2,w_2) ( M 1 , w 1 ) ≡ σ d ( M 2 , w 2 ) , for every ( M i , w i ) ∈ M i (\mathcal{M}_i,w_i) \in\mathcal{M}_i ( M i , w i ) ∈ M i , i = 1 , 2 i=1,2 i = 1 , 2 .
Indeed, since each ( σ ∪ ρ i , d ) (\sigma\cup\rho_i,d) ( σ ∪ ρ i , d ) -atom refines the ≡ σ d \equiv_{\sigma}^{d} ≡ σ d -equivalence on pointed models, and so all members of M i \mathcal{M}_i M i are pairwise ≡ σ d \equiv_{\sigma}^{d} ≡ σ d -equivalent, M 1 ≡ σ d M 2 \mathcal{M}_1 \equiv_{\sigma}^{d} \mathcal{M}_2 M 1 ≡ σ d M 2 implies that all pointed models in M 1 \mathcal{M}_1 M 1 and M 2 \mathcal{M}_2 M 2 are ≡ σ d \equiv_{\sigma}^{d} ≡ σ d -equivalent. It also follows that ≡ σ d \equiv_{\sigma}^{d} ≡ σ d is an equivalence relation on A t σ ∪ ρ d \mathrm{At}^{d}_{\sigma\cup\rho} At σ ∪ ρ d . We now introduce the corresponding partitioning of the set of ( σ ∪ ρ , d ) (\sigma\cup\rho,d) ( σ ∪ ρ , d ) -atoms.
Definition 6 (cell). The ≡ σ d \equiv_{\sigma}^{d} ≡ σ d -class of a ( σ ∪ ρ , d ) (\sigma\cup\rho,d) ( σ ∪ ρ , d ) -atom M \mathcal{M} M is denoted [ M ] σ d [\mathcal{M}]_{\sigma}^{d} [ M ] σ d and called a σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell. The set of all σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cells is denoted C e l l σ , ρ d \mathrm{Cell}^{d}_{\sigma,\rho} Cell σ , ρ d ; in other words, C e l l σ , ρ d = A t σ ∪ ρ d / ≡ σ d \mathrm{Cell}^{d}_{\sigma,\rho} = \mathrm{At}^{d}_{\sigma\cup\rho}/\equiv_{\sigma}^{d} Cell σ , ρ d = At σ ∪ ρ d / ≡ σ d .
Note that we use comma to separate the signatures in C e l l σ , ρ d \mathrm{Cell}^{d}_{\sigma,\rho} Cell σ , ρ d . Since the pointed models of the ( σ ∪ ρ , d ) (\sigma\cup\rho,d) ( σ ∪ ρ , d ) -atoms in a given σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell C ∈ C e l l σ , ρ d C \in\mathrm{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d are all ≡ σ d \equiv_{\sigma}^{d} ≡ σ d -equivalent, they form a ( σ , d ) (\sigma,d) ( σ , d ) -atom themselves:
Proposition 6. For any C ∈ C e l l σ , ρ d C \in\mathrm{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d we have ⋃ C = [ M , w ] σ d \bigcup C = [\mathcal{M},w]_{\sigma}^{d} ⋃ C = [ M , w ] σ d , for any ( M , w ) ∈ M (\mathcal{M},w) \in\mathcal{M} ( M , w ) ∈ M with M ∈ C \mathcal{M} \in C M ∈ C .
It follows, in particular, that the cell characteristic formula Γ ( C ) = Φ ( ⋃ C ) \Gamma(C) = \Phi(\bigcup C) Γ ( C ) = Φ ( ⋃ C ) is an M L σ d \mathrm{ML}^{d}_{\sigma} ML σ d -formula. Observe that the σ \sigma σ -formula Γ ( C ) \Gamma(C) Γ ( C ) is L \mathcal{L} L -equivalent to the ( σ ∪ ρ ) \left(\sigma\cup\rho\right) ( σ ∪ ρ ) -formula ⋁ M ∈ C Φ ( M ) \bigvee_{\mathcal{M}\in C} \Phi(\mathcal{M}) ⋁ M ∈ C Φ ( M ) . Since the formulas Γ ( C ) \Gamma(C) Γ ( C ) , C ∈ C e l l σ , ρ d C \in\mathrm{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d , are pairwise inconsistent, and their disjunction is valid, we have
φ ≡ ⋁ C ∈ C e l l σ , ρ d ( φ ∧ Γ ( C ) ) , for any φ ∈ M L σ ∪ ρ d . \varphi\equiv\bigvee_{C\in\mathrm{Cell}^{d}_{\sigma,\rho}} \left(\varphi\land\Gamma(C)\right), \qquad\text{for any } \varphi\in\mathrm{ML}^{d}_{\sigma\cup\rho}. φ ≡ C ∈ Cell σ , ρ d ⋁ ( φ ∧ Γ ( C ) ) , for any φ ∈ ML σ ∪ ρ d .
In particular, Φ ( D ) ∧ Γ ( C ) = Φ ( D ∩ C ) \Phi(D) \land\Gamma(C) = \Phi(D \cap C) Φ ( D ) ∧ Γ ( C ) = Φ ( D ∩ C ) . We are now in a position to state our main cell factorisation result:
Lemma 7 (cell factorisation). A formula φ ∈ M L σ ∪ ρ d \varphi\in\mathrm{ML}^{d}_{\sigma\cup\rho} φ ∈ ML σ ∪ ρ d has a uniform σ \sigma σ -interpolant if and only if φ ∧ Γ ( C ) \varphi\land\Gamma(C) φ ∧ Γ ( C ) has a uniform σ \sigma σ -interpolant, for each σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell C ∈ C e l l σ , ρ d C \in\mathrm{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d .
Proof. Suppose first that θ \theta θ is a uniform σ \sigma σ -interpolant of φ \varphi φ . We show that θ ∧ Γ ( C ) \theta\land\Gamma(C) θ ∧ Γ ( C ) is a uniform σ \sigma σ -interpolant of φ ∧ Γ ( C ) \varphi\land\Gamma(C) φ ∧ Γ ( C ) , for each C C C . Trivially, φ ∧ Γ ( C ) ⊨ θ ∧ Γ ( C ) \varphi\land\Gamma(C) \models\theta\land\Gamma(C) φ ∧ Γ ( C ) ⊨ θ ∧ Γ ( C ) . Take any ρ \rho ρ -free formula ψ \psi ψ with φ ∧ Γ ( C ) ⊨ ψ \varphi\land\Gamma(C) \models\psi φ ∧ Γ ( C ) ⊨ ψ . Then φ ⊨ Γ ( C ) → ψ \varphi\models\Gamma(C) \to\psi φ ⊨ Γ ( C ) → ψ . As Γ ( C ) → ψ \Gamma(C) \to\psi Γ ( C ) → ψ is ρ \rho ρ -free, uniform interpolation gives θ ⊨ Γ ( C ) → ψ \theta\models\Gamma(C) \to\psi θ ⊨ Γ ( C ) → ψ , whence θ ∧ Γ ( C ) ⊨ ψ \theta\land\Gamma(C) \models\psi θ ∧ Γ ( C ) ⊨ ψ .
Conversely, suppose that φ ∧ Γ ( C ) \varphi\land\Gamma(C) φ ∧ Γ ( C ) has a uniform σ \sigma σ -interpolant θ C \theta_C θ C , for each C C C . We show that ⋁ C θ C \bigvee_C \theta_C ⋁ C θ C is a uniform σ \sigma σ -interpolant of φ \varphi φ . Since φ ≡ ⋁ C ( φ ∧ Γ ( C ) ) \varphi\equiv\bigvee_C(\varphi\land\Gamma(C)) φ ≡ ⋁ C ( φ ∧ Γ ( C )) , we have φ ⊨ ⋁ C θ C \varphi\models\bigvee_C \theta_C φ ⊨ ⋁ C θ C . If φ ⊨ ψ \varphi\models\psi φ ⊨ ψ for a ρ \rho ρ -free formula ψ \psi ψ , then φ ∧ Γ ( C ) ⊨ ψ \varphi\land\Gamma(C) \models\psi φ ∧ Γ ( C ) ⊨ ψ for every C C C . Thus, θ C ⊨ ψ \theta_C \models\psi θ C ⊨ ψ for every C C C , and consequently ⋁ C θ C ⊨ ψ \bigvee_C \theta_C \models\psi ⋁ C θ C ⊨ ψ . □
Example 3. Let us revisit the logic of 3-forks from Example 1. For a signature τ \tau τ , let Λ τ n \Lambda^n_\tau Λ τ n denote the set of all n n n -element subsets of the set of τ \tau τ -colours (we slightly abuse notation and assume that Λ τ 1 \Lambda^1_\tau Λ τ 1 contains τ \tau τ -colours themselves rather than their singleton sets). Thus
∣ Λ τ n ∣ = ( 2 ∣ τ ∣ n ) . \lvert\Lambda^n_\tau\rvert= \binom{2^{\lvert\tau\rvert}}{n}. ∣ Λ τ n ∣ = ( n 2 ∣ τ ∣ ) .
Given a σ \sigma σ -colour s s s and a ρ \rho ρ -colour r r r , we denote by s r sr sr the ( σ ∪ ρ ) (\sigma\cup\rho) ( σ ∪ ρ ) -colour that is obtained by taking the union of s s s and r r r (which corresponds to the conjunction of their literals if used in a formula); in the examples, we often write, say, p q pq pq for a colour comprising p p p and q q q . With this notation at hand, we can describe the three groups of σ ρ \sigma^\rho σ ρ -cells for the logic of 3-forks; see Figure 2 .
Figure 2. σ^ρ-cells for 3-forks: the σ- and ρ-colours are indicated by the shape and the shading of the node, respectively.
C s 0 , s 1 ρ C^\rho_{s_0,s_1} C s 0 , s 1 ρ : for s 0 , s 1 ∈ Λ σ 1 s_0,s_1 \in\Lambda^1_\sigma s 0 , s 1 ∈ Λ σ 1 , containing the following atoms with characteristic formulas:
M s 0 r 0 , s 1 r 1 s 0 r 0 ∧ ∇ ( s 1 r 1 ) , for r 0 , r 1 ∈ Λ ρ 1 ; M s 0 r 0 , s 1 { r 1 , r 2 } s 0 r 0 ∧ ∇ ( s 1 r 1 , s 1 r 2 ) , for r 0 ∈ Λ ρ 1 and { r 1 , r 2 } ∈ Λ ρ 2 ; M s 0 r 0 , s 1 { r 1 , r 2 , r 3 } s 0 r 0 ∧ ∇ ( s 1 r 1 , s 1 r 2 , s 1 r 3 ) , for r 0 ∈ Λ ρ 1 and { r 1 , r 2 , r 3 } ∈ Λ ρ 3 . \begin{aligned}
\mathcal{M}_{s_0r_0,s_1r_1} &\quad s_0r_0 \land\nabla(s_1r_1), &&\text{for } r_0,r_1 \in\Lambda^1_\rho;\\
\mathcal{M}_{s_0r_0,s_1\{r_1,r_2\}} &\quad s_0r_0 \land\nabla(s_1r_1,s_1r_2), &&\text{for } r_0 \in\Lambda^1_\rho\text{ and } \{r_1,r_2\} \in\Lambda^2_\rho;\\
\mathcal{M}_{s_0r_0,s_1\{r_1,r_2,r_3\}} &\quad s_0r_0 \land\nabla(s_1r_1,s_1r_2,s_1r_3), &&\text{for } r_0 \in\Lambda^1_\rho\text{ and } \{r_1,r_2,r_3\} \in\Lambda^3_\rho.
\end{aligned} M s 0 r 0 , s 1 r 1 M s 0 r 0 , s 1 { r 1 , r 2 } M s 0 r 0 , s 1 { r 1 , r 2 , r 3 } s 0 r 0 ∧ ∇ ( s 1 r 1 ) , s 0 r 0 ∧ ∇ ( s 1 r 1 , s 1 r 2 ) , s 0 r 0 ∧ ∇ ( s 1 r 1 , s 1 r 2 , s 1 r 3 ) , for r 0 , r 1 ∈ Λ ρ 1 ; for r 0 ∈ Λ ρ 1 and { r 1 , r 2 } ∈ Λ ρ 2 ; for r 0 ∈ Λ ρ 1 and { r 1 , r 2 , r 3 } ∈ Λ ρ 3 .
C s 0 , { s 1 , s 2 } ρ C^\rho_{s_0,\{s_1,s_2\}} C s 0 , { s 1 , s 2 } ρ : for s 0 ∈ Λ σ 1 s_0 \in\Lambda^1_\sigma s 0 ∈ Λ σ 1 and { s 1 , s 2 } ∈ Λ σ 2 \{s_1,s_2\} \in\Lambda^2_\sigma { s 1 , s 2 } ∈ Λ σ 2 , containing the following atoms with characteristic formulas:
M s 0 r 0 , { s 1 r 1 , s 2 r 2 } s 0 r 0 ∧ ∇ ( s 1 r 1 , s 2 r 2 ) , for r 0 , r 1 , r 2 ∈ Λ ρ 1 ; M s 0 r 0 , s 1 { r 1 , r 2 } , s 2 r ′ s 0 r 0 ∧ ∇ ( s 1 r 1 , s 1 r 2 , s 2 r ′ ) , for r 0 , r ′ ∈ Λ ρ 1 and { r 1 , r 2 } ∈ Λ ρ 2 ; M s 0 r 0 , s 1 r ′ , s 2 { r 1 , r 2 } s 0 r 0 ∧ ∇ ( s 1 r ′ , s 2 r 1 , s 2 r 2 ) , for r 0 , r ′ ∈ Λ ρ 1 and { r 1 , r 2 } ∈ Λ ρ 2 . \begin{aligned}
\mathcal{M}_{s_0r_0,\{s_1r_1,s_2r_2\}} &\quad s_0r_0 \land\nabla(s_1r_1,s_2r_2), &&\text{for } r_0,r_1,r_2 \in\Lambda^1_\rho;\\
\mathcal{M}_{s_0r_0,s_1\{r_1,r_2\},s_2r'} &\quad s_0r_0 \land\nabla(s_1r_1,s_1r_2,s_2r'), &&\text{for } r_0,r' \in\Lambda^1_\rho\text{ and } \{r_1,r_2\} \in\Lambda^2_\rho;\\
\mathcal{M}_{s_0r_0,s_1r',s_2\{r_1,r_2\}} &\quad s_0r_0 \land\nabla(s_1r',s_2r_1,s_2r_2), &&\text{for } r_0,r' \in\Lambda^1_\rho\text{ and } \{r_1,r_2\} \in\Lambda^2_\rho.
\end{aligned} M s 0 r 0 , { s 1 r 1 , s 2 r 2 } M s 0 r 0 , s 1 { r 1 , r 2 } , s 2 r ′ M s 0 r 0 , s 1 r ′ , s 2 { r 1 , r 2 } s 0 r 0 ∧ ∇ ( s 1 r 1 , s 2 r 2 ) , s 0 r 0 ∧ ∇ ( s 1 r 1 , s 1 r 2 , s 2 r ′ ) , s 0 r 0 ∧ ∇ ( s 1 r ′ , s 2 r 1 , s 2 r 2 ) , for r 0 , r 1 , r 2 ∈ Λ ρ 1 ; for r 0 , r ′ ∈ Λ ρ 1 and { r 1 , r 2 } ∈ Λ ρ 2 ; for r 0 , r ′ ∈ Λ ρ 1 and { r 1 , r 2 } ∈ Λ ρ 2 .
C s 0 , { s 1 , s 2 , s 3 } ρ C^\rho_{s_0,\{s_1,s_2,s_3\}} C s 0 , { s 1 , s 2 , s 3 } ρ : for s 0 ∈ Λ σ 1 s_0 \in\Lambda^1_\sigma s 0 ∈ Λ σ 1 and { s 1 , s 2 , s 3 } ∈ Λ σ 3 \{s_1,s_2,s_3\} \in\Lambda^3_\sigma { s 1 , s 2 , s 3 } ∈ Λ σ 3 , containing the following atoms with characteristic formulas:
M s 0 r 0 , { s 1 r 1 , s 2 r 2 , s 3 r 3 } s 0 r 0 ∧ ∇ ( s 1 r 1 , s 2 r 2 , s 3 r 3 ) , for r 0 , r 1 , r 2 , r 3 ∈ Λ ρ 1 . \mathcal{M}_{s_0r_0,\{s_1r_1,s_2r_2,s_3r_3\}} \quad s_0r_0 \land\nabla(s_1r_1,s_2r_2,s_3r_3), \qquad\text{for } r_0,r_1,r_2,r_3 \in\Lambda^1_\rho. M s 0 r 0 , { s 1 r 1 , s 2 r 2 , s 3 r 3 } s 0 r 0 ∧ ∇ ( s 1 r 1 , s 2 r 2 , s 3 r 3 ) , for r 0 , r 1 , r 2 , r 3 ∈ Λ ρ 1 .
Observe that only the cells C s 0 , { s 1 , s 2 } ρ C^\rho_{s_0,\{s_1,s_2\}} C s 0 , { s 1 , s 2 } ρ can contain counterexamples for uniform σ \sigma σ -interpolation. For example, with σ = { p } \sigma= \{p\} σ = { p } and ρ = { q } \rho= \{q\} ρ = { q } , consider the atom M φ = M p q , p { q , − q } , − p q \mathcal{M}_\varphi= \mathcal{M}_{pq,p\{q,-q\},-pq} M φ = M pq , p { q , − q } , − pq corresponding to φ \varphi φ in Example 2. It belongs to the cell C p , { p , − p } ρ C^\rho_{p,\{p,-p\}} C p , { p , − p } ρ , since its root satisfies p p p and its successors realise exactly the two σ \sigma σ -colours, { p } \{p\} { p } and { − p } \{-p\} { − p } . The atom M ψ \mathcal{M}_\psi M ψ defined by ψ \psi ψ is obtained from M p q , p q , − p { q , − q } \mathcal{M}_{pq,pq,-p\{q,-q\}} M pq , pq , − p { q , − q } by renaming q q q to q ′ q' q ′ . It belongs to the same cell. Thus, the two atoms agree over σ \sigma σ , but their refinements over the forgotten variables differ.
Pinning Signatures and Flipping Variables
In the constructions below, we require two operations: one globally assigns either ⊤ \top ⊤ or ⊥ \bot ⊥ to each variable in a given signature, while the other flips the truth value of each variable in a given signature. Let τ \tau τ be a fixed signature.
Definition 7 (pinning a signature). A pinning of τ \tau τ is a function π : τ → { ⊤ , ⊥ } \pi:\tau\to\{\top,\bot\} π : τ → { ⊤ , ⊥ } . For a model M = ( F , V ) \mathcal{M}=(\mathcal{F},V) M = ( F , V ) , let M π = ( F , V π ) \mathcal{M}^{\pi}=(\mathcal{F},V^{\pi}) M π = ( F , V π ) be the model that agrees with M \mathcal{M} M on symbols outside τ \tau τ and sets each s ∈ τ s\in\tau s ∈ τ to a constant value according to π \pi π :
V π ( s ) : = { W , if π ( s ) = ⊤ , ∅ , if π ( s ) = ⊥ , for each s ∈ τ , V^{\pi}(s):=\begin{cases}
W, & \text{if }\pi(s)=\top,\\
\varnothing, & \text{if }\pi(s)=\bot,
\end{cases}
\qquad\text{for each }s\in\tau, V π ( s ) := { W , ∅ , if π ( s ) = ⊤ , if π ( s ) = ⊥ , for each s ∈ τ , which is admissible since ∅ , W ∈ A \varnothing,W\in A ∅ , W ∈ A . Similarly, given a formula φ \varphi φ , we write φ − π \varphi^{-\pi} φ − π for the simultaneous substitution of π ( s ) \pi(s) π ( s ) for each s ∈ τ s\in\tau s ∈ τ .
Observe that
sig ( φ − π ) ⊆ sig ( φ ) ∖ τ and md ( φ − π ) ≤ md ( φ ) , for any formula φ . \operatorname{sig}(\varphi^{-\pi})\subseteq\operatorname{sig}(\varphi)\setminus\tau
\qquad\text{and}\qquad
\operatorname{md}(\varphi^{-\pi})\leq\operatorname{md}(\varphi),
\qquad\text{for any formula }\varphi. sig ( φ − π ) ⊆ sig ( φ ) ∖ τ and md ( φ − π ) ≤ md ( φ ) , for any formula φ .
The relevant preservation properties of pinning are as follows:
Lemma 8. Let π \pi π be a pinning of τ \tau τ .
( M , w ) ⊨ φ − π (\mathcal{M},w)\models\varphi^{-\pi} ( M , w ) ⊨ φ − π if and only if ( M π , w ) ⊨ φ (\mathcal{M}^{\pi},w)\models\varphi ( M π , w ) ⊨ φ , for every φ \varphi φ and every ( M , w ) ∈ PMod L (\mathcal{M},w)\in\operatorname{PMod}_{L} ( M , w ) ∈ PMod L ;
( M , w ) ≡ μ d ( N , u ) (\mathcal{M},w)\equiv_{\mu}^{d}(\mathcal{N},u) ( M , w ) ≡ μ d ( N , u ) if and only if ( M π , w ) ≡ μ ∪ τ d ( N π , u ) (\mathcal{M}^{\pi},w)\equiv_{\mu\cup\tau}^{d}(\mathcal{N}^{\pi},u) ( M π , w ) ≡ μ ∪ τ d ( N π , u ) , for every ( M , w ) , ( N , u ) ∈ PMod L (\mathcal{M},w),(\mathcal{N},u)\in\operatorname{PMod}_{L} ( M , w ) , ( N , u ) ∈ PMod L and every μ \mu μ disjoint from τ \tau τ .
Proof. (1) By induction on the structure of $\varphi.
(2) Suppose ( M , w ) ≡ μ d ( N , u ) (\mathcal{M},w)\equiv_{\mu}^{d}(\mathcal{N},u) ( M , w ) ≡ μ d ( N , u ) . Take any φ ∈ ML μ d \varphi\in\operatorname{ML}_{\mu}^{d} φ ∈ ML μ d . Since φ − π ∈ ML μ d \varphi^{-\pi}\in\operatorname{ML}_{\mu}^{d} φ − π ∈ ML μ d , we have M , w ⊨ φ − π \mathcal{M},w\models\varphi^{-\pi} M , w ⊨ φ − π iff N , u ⊨ φ − π \mathcal{N},u\models\varphi^{-\pi} N , u ⊨ φ − π ; two applications of Item (1) turn this into M π , w ⊨ φ \mathcal{M}^{\pi},w\models\varphi M π , w ⊨ φ iff N π , u ⊨ φ \mathcal{N}^{\pi},u\models\varphi N π , u ⊨ φ .
Conversely, let ψ ∈ ML μ d ⊆ ML μ ∪ τ d \psi\in\operatorname{ML}_{\mu}^{d}\subseteq\operatorname{ML}_{\mu\cup\tau}^{d} ψ ∈ ML μ d ⊆ ML μ ∪ τ d . As ψ \psi ψ has no variable of τ \tau τ , ψ − π = ψ \psi^{-\pi}=\psi ψ − π = ψ , so Item (1) gives M π , w ⊨ ψ \mathcal{M}^{\pi},w\models\psi M π , w ⊨ ψ iff M , w ⊨ ψ \mathcal{M},w\models\psi M , w ⊨ ψ and likewise for N \mathcal{N} N . Hence M , w ⊨ ψ \mathcal{M},w\models\psi M , w ⊨ ψ iff M π , w ⊨ ψ \mathcal{M}^{\pi},w\models\psi M π , w ⊨ ψ iff N π , u ⊨ ψ \mathcal{N}^{\pi},u\models\psi N π , u ⊨ ψ iff N , u ⊨ ψ \mathcal{N},u\models\psi N , u ⊨ ψ .
:::
We now extend the pinning operation to atoms: for M ∈ At μ d \mathcal{M}\in\operatorname{At}_{\mu}^{d} M ∈ At μ d and a pinning π \pi π of τ \tau τ (disjoint from μ \mu μ ), set
M π = [ ( M π , w ) ] μ ∪ τ d ∈ At μ ∪ τ d , for some ( M , w ) ∈ M . \mathcal{M}^{\pi}=[(\mathcal{M}^{\pi},w)]_{\mu\cup\tau}^{d}\in\operatorname{At}_{\mu\cup\tau}^{d},
\qquad\text{for some }(\mathcal{M},w)\in\mathcal{M}. M π = [( M π , w ) ] μ ∪ τ d ∈ At μ ∪ τ d , for some ( M , w ) ∈ M .
By Lemma 8(2), M π \mathcal{M}^{\pi} M π is well-defined and does not depend on the choice of ( M , w ) ∈ M (\mathcal{M},w)\in\mathcal{M} ( M , w ) ∈ M . Pinning can then be naturally lifted to sets of atoms and cells: for D ⊆ At μ d D\subseteq\operatorname{At}_{\mu}^{d} D ⊆ At μ d and a pinning π \pi π of τ \tau τ (disjoint from μ \mu μ ), set
D π = { M π ∣ M ∈ D } ⊆ At μ ∪ τ d . D^{\pi}=\{\mathcal{M}^{\pi}\mid\mathcal{M}\in D\}\subseteq\operatorname{At}_{\mu\cup\tau}^{d}. D π = { M π ∣ M ∈ D } ⊆ At μ ∪ τ d .
Lemma 9. Let μ \mu μ be a signature disjoint from τ \tau τ and π \pi π a pinning of τ \tau τ .
Φ ( D π ) ≡ Φ ( D ) ∧ ⋀ p ∈ τ □ ≤ d ( p ↔ π ( p ) ) \Phi(D^{\pi})\equiv\Phi(D)\land\bigwedge_{p\in\tau}\Box^{\leq d}(p\leftrightarrow\pi(p)) Φ ( D π ) ≡ Φ ( D ) ∧ ⋀ p ∈ τ □ ≤ d ( p ↔ π ( p )) , for every D ⊆ At μ d D\subseteq\operatorname{At}_{\mu}^{d} D ⊆ At μ d .
The map ⋅ π : M ↦ M π \cdot^{\pi}:\mathcal{M}\mapsto\mathcal{M}^{\pi} ⋅ π : M ↦ M π is a bijection from D D D onto D π D^{\pi} D π , for every D ⊆ At μ d D\subseteq\operatorname{At}_{\mu}^{d} D ⊆ At μ d .
C π ∈ Cell σ ∗ ∪ τ , ρ d C^{\pi}\in\operatorname{Cell}_{\sigma^{*}\cup\tau,\rho}^{d} C π ∈ Cell σ ∗ ∪ τ , ρ d , for each cell C ∈ Cell σ ∗ , ρ d C\in\operatorname{Cell}_{\sigma^{*},\rho}^{d} C ∈ Cell σ ∗ , ρ d , for any pairwise disjoint signatures σ ∗ \sigma^{*} σ ∗ , τ \tau τ and ρ \rho ρ .
Proof. (1) Follows from Lemma 8(1).
(2) Surjectivity is by definition, while injectivity is by observing that the restriction of M π \mathcal{M}^{\pi} M π to μ \mu μ is M \mathcal{M} M .
(3) By definition, C π C^{\pi} C π is a set of ( σ ∗ ∪ τ ∪ ρ , d ) (\sigma^{*}\cup\tau\cup\rho,d) ( σ ∗ ∪ τ ∪ ρ , d ) -atoms with, by (1), Φ ( C π ) ≡ Φ ( C ) ∧ ⋀ p ∈ τ □ ≤ d ( p ↔ π ( p ) ) \Phi(C^{\pi})\equiv\Phi(C)\land\bigwedge_{p\in\tau}\Box^{\leq d}(p\leftrightarrow\pi(p)) Φ ( C π ) ≡ Φ ( C ) ∧ ⋀ p ∈ τ □ ≤ d ( p ↔ π ( p )) . On the other hand, consider ( σ ∗ , d ) (\sigma^{*},d) ( σ ∗ , d ) -atom M = ⋃ C \mathcal{M}=\bigcup C M = ⋃ C . By definition, M π \mathcal{M}^{\pi} M π is a ( σ ∗ ∪ τ , d ) (\sigma^{*}\cup\tau,d) ( σ ∗ ∪ τ , d ) -atom with, by (1), Φ ( M π ) ≡ Φ ( M ) ∧ ⋀ p ∈ τ □ ≤ d ( p ↔ π ( p ) ) \Phi(\mathcal{M}^{\pi})\equiv\Phi(\mathcal{M})\land\bigwedge_{p\in\tau}\Box^{\leq d}(p\leftrightarrow\pi(p)) Φ ( M π ) ≡ Φ ( M ) ∧ ⋀ p ∈ τ □ ≤ d ( p ↔ π ( p )) . It follows that Φ ( C π ) ≡ Φ ( M π ) \Phi(C^{\pi})\equiv\Phi(\mathcal{M}^{\pi}) Φ ( C π ) ≡ Φ ( M π ) , and so ⋃ C π = M π \bigcup C^{\pi}=\mathcal{M}^{\pi} ⋃ C π = M π . Thus, C π C^{\pi} C π is a ( σ ∗ ∪ τ ) ρ (\sigma^{*}\cup\tau)^{\rho} ( σ ∗ ∪ τ ) ρ -cell.
:::
We next introduce the second operation.
Definition 8 (flipping). Let v ⊆ τ v \subseteq\tau v ⊆ τ . For any model M = ( F , V ) \mathcal{M}=(\mathfrak{F},V) M = ( F , V ) , let M v = ( F , V v ) \mathcal{M}^{v}=(\mathfrak{F},V^{v}) M v = ( F , V v ) flip the variables selected by v v v :
V v ( p ) : = { W ∖ V ( p ) , if p ∈ v , V ( p ) , otherwise . V^{v}(p):=
\begin{cases}
W\setminus V(p), & \text{if }p\in v,\\
V(p), & \text{otherwise}.
\end{cases} V v ( p ) := { W ∖ V ( p ) , V ( p ) , if p ∈ v , otherwise . which is admissible since A A A is closed under complement. Similarly, for a formula φ \varphi φ , denote by φ v \varphi^{v} φ v the result of replacing each p ∈ v p\in v p ∈ v by ¬ p \neg p ¬ p .
Observe that ⋅ v \cdot^{v} ⋅ v is involutive and preserves both the signature and the modal depth of formulas. Moreover, we have the following:
Lemma 10. ( M v , w ) ⊨ φ (\mathcal{M}^{v},w)\models\varphi ( M v , w ) ⊨ φ if and only if ( M , w ) ⊨ φ v (\mathcal{M},w)\models\varphi^{v} ( M , w ) ⊨ φ v , for every formula φ \varphi φ and every ( M , w ) ∈ P M o d L (\mathcal{M},w)\in\mathrm{PMod}_{L} ( M , w ) ∈ PMod L .
We now extend the flipping operation to atoms: for M ∈ A t μ ∪ τ d \mathcal{M}\in\mathrm{At}^{d}_{\mu\cup\tau} M ∈ At μ ∪ τ d with μ \mu μ disjoint from τ \tau τ and v ⊆ τ v\subseteq\tau v ⊆ τ , let
M v = { ( M v , w ) ∣ ( M , w ) ∈ M } ∈ A t μ ∪ τ d . \mathcal{M}^{v}=\{(\mathcal{M}^{v},w)\mid(\mathcal{M},w)\in\mathcal{M}\}\in\mathrm{At}^{d}_{\mu\cup\tau}. M v = {( M v , w ) ∣ ( M , w ) ∈ M } ∈ At μ ∪ τ d .
Next, for any D ⊆ A t μ ∪ τ d D\subseteq\mathrm{At}^{d}_{\mu\cup\tau} D ⊆ At μ ∪ τ d , set D v = { M v ∣ M ∈ D } ⊆ A t μ ∪ τ d D^{v}=\{\mathcal{M}^{v}\mid\mathcal{M}\in D\}\subseteq\mathrm{At}^{d}_{\mu\cup\tau} D v = { M v ∣ M ∈ D } ⊆ At μ ∪ τ d .
Lemma 11. Let μ \mu μ be a signature disjoint from τ \tau τ and v ⊆ τ v\subseteq\tau v ⊆ τ .
Φ ( D v ) = ( Φ ( D ) ) v \Phi(D^{v})=\bigl(\Phi(D)\bigr)^{v} Φ ( D v ) = ( Φ ( D ) ) v , for every D ⊆ A t μ ∪ τ d D\subseteq\mathrm{At}^{d}_{\mu\cup\tau} D ⊆ At μ ∪ τ d .
The map ⋅ v : M ↦ M v \cdot^{v}:\mathcal{M}\mapsto\mathcal{M}^{v} ⋅ v : M ↦ M v is a bijection from D D D onto D v D^{v} D v , for every D ⊆ A t μ ∪ τ d D\subseteq\mathrm{At}^{d}_{\mu\cup\tau} D ⊆ At μ ∪ τ d .
C v ∈ C e l l τ , μ d C^{v}\in\mathrm{Cell}^{d}_{\tau,\mu} C v ∈ Cell τ , μ d , for all cells C ∈ C e l l τ , μ d C\in\mathrm{Cell}^{d}_{\tau,\mu} C ∈ Cell τ , μ d . Moreover, C v 1 ≠ C v 2 C^{v_{1}}\ne C^{v_{2}} C v 1 = C v 2 , for distinct v 1 , v 2 ⊆ τ v_{1},v_{2}\subseteq\tau v 1 , v 2 ⊆ τ .
Proof. (1) and (2) are immediate from the definitions.
(3) Two different flips already differ at depth 0 0 0 : ( M v 1 , w ) ̸ ≡ τ 0 ( M v 2 , w ) (\mathcal{M}^{v_{1}},w)\not\equiv^{0}_{\tau}(\mathcal{M}^{v_{2}},w) ( M v 1 , w ) ≡ τ 0 ( M v 2 , w ) , for distinct v 1 , v 2 ⊆ τ v_{1},v_{2}\subseteq\tau v 1 , v 2 ⊆ τ , since some p ∈ τ p\in\tau p ∈ τ receives opposite truth values. Consequently, the 2 ∣ τ ∣ 2^{|\tau|} 2 ∣ τ ∣ flips of a fixed atom lie in 2 ∣ τ ∣ 2^{|\tau|} 2 ∣ τ ∣ distinct τ μ \tau^{\mu} τ μ -cells, all of the same size as the original by the same relabelling bijection.
Density of Uniform Interpolants
In this section, we introduce the density of uniform interpolants and show that it factorises into a product of intra-cell densities. We then investigate its asymptotic behaviour as either the number of retained variables or the number of forgotten variables tends to infinity.
Fix a logic L L L , a modal depth d ≥ 0 d\geq0 d ≥ 0 and disjoint retained and forgotten signatures, σ \sigma σ and ρ \rho ρ . Identifying, as in Definition 2, a formula of depth ≤ d \leq d ≤ d over σ ∪ ρ \sigma\cup\rho σ ∪ ρ with the subset of A t σ ∪ ρ d \mathrm{At}^{d}_{\sigma\cup\rho} At σ ∪ ρ d it defines, set
U σ , ρ d : = { D ⊆ A t σ ∪ ρ d ∣ Φ ( D ) has a uniform σ -interpolant } . \mathcal{U}^{d}_{\sigma,\rho}:=\{D\subseteq\mathrm{At}^{d}_{\sigma\cup\rho}\mid\Phi(D)\text{ has a uniform }\sigma\text{-interpolant}\}. U σ , ρ d := { D ⊆ At σ ∪ ρ d ∣ Φ ( D ) has a uniform σ -interpolant } .
The density of uniform interpolation is the proportion of L L L -equivalence classes of formulas of depth at most d d d over σ ∪ ρ \sigma\cup\rho σ ∪ ρ that admit a uniform σ \sigma σ -interpolant. Equivalently, we equip the powerset of A t σ ∪ ρ d \mathrm{At}^{d}_{\sigma\cup\rho} At σ ∪ ρ d with the uniform distribution and put
U n i f I n t σ , ρ d : = ∣ U σ , ρ d ∣ 2 ∣ A t σ ∪ ρ d ∣ . \mathrm{UnifInt}^{d}_{\sigma,\rho}:=\frac{|\mathcal{U}^{d}_{\sigma,\rho}|}{2^{|\mathrm{At}^{d}_{\sigma\cup\rho}|}}. UnifInt σ , ρ d := 2 ∣ At σ ∪ ρ d ∣ ∣ U σ , ρ d ∣ .
For a cell C ∈ C e l l σ , ρ d C\in\mathrm{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d , we define the corresponding intra-cell density by restricting to subsets of C C C :
U C : = { D ⊆ C ∣ Φ ( D ) has a uniform σ -interpolant } and U n i f I n t C : = ∣ U C ∣ 2 ∣ C ∣ . \mathcal{U}_{C}:=\{D\subseteq C\mid\Phi(D)\text{ has a uniform }\sigma\text{-interpolant}\}
\qquad\text{and}\qquad
\mathrm{UnifInt}_{C}:=\frac{|\mathcal{U}_{C}|}{2^{|C|}}. U C := { D ⊆ C ∣ Φ ( D ) has a uniform σ -interpolant } and UnifInt C := 2 ∣ C ∣ ∣ U C ∣ .
Theorem 12 (cell reduction principle).
U n i f I n t σ , ρ d = ∏ C ∈ C e l l σ , ρ d U n i f I n t C . \mathrm{UnifInt}^{d}_{\sigma,\rho}=\prod_{C\in\mathrm{Cell}^{d}_{\sigma,\rho}}\mathrm{UnifInt}_{C}. UnifInt σ , ρ d = C ∈ Cell σ , ρ d ∏ UnifInt C .
Proof. Since the cells partition A σ ∪ ρ t , d A_{\sigma\cup\rho}^{t,d} A σ ∪ ρ t , d , the map D ↦ ( D ∩ C ) C ∈ C e l l σ , ρ d D \mapsto(D \cap C)_{C\in\mathrm{Cell}_{\sigma,\rho}^{d}} D ↦ ( D ∩ C ) C ∈ Cell σ , ρ d is a bijection from 2 A σ ∪ ρ t , d 2^{A_{\sigma\cup\rho}^{t,d}} 2 A σ ∪ ρ t , d onto ∏ C ∈ C e l l σ , ρ d 2 C \prod_{C\in\mathrm{Cell}_{\sigma,\rho}^{d}}2^C ∏ C ∈ Cell σ , ρ d 2 C . By Lemma 7, this bijection restricts to a bijection U σ , ρ d → ∏ C ∈ C e l l σ , ρ d U C U_{\sigma,\rho}^{d} \to\prod_{C\in\mathrm{Cell}_{\sigma,\rho}^{d}} U_C U σ , ρ d → ∏ C ∈ Cell σ , ρ d U C . Consequently, ∣ U σ , ρ d ∣ = ∏ C ∣ U C ∣ |U_{\sigma,\rho}^{d}|=\prod_C|U_C| ∣ U σ , ρ d ∣ = ∏ C ∣ U C ∣ . Since 2 ∣ A σ ∪ ρ t , d ∣ = ∏ C 2 ∣ C ∣ 2^{|A_{\sigma\cup\rho}^{t,d}|}=\prod_C2^{|C|} 2 ∣ A σ ∪ ρ t , d ∣ = ∏ C 2 ∣ C ∣ , division gives the claimed identity.
As 0 ≤ UnifInt C ≤ 1 0 \leq\operatorname{UnifInt}_C \leq1 0 ≤ UnifInt C ≤ 1 , it follows that a single cell C C C with UnifInt C < 1 \operatorname{UnifInt}_C < 1 UnifInt C < 1 ensures UnifInt σ , ρ d < 1 \operatorname{UnifInt}_{\sigma,\rho}^{d}<1 UnifInt σ , ρ d < 1 . We also write UnifInt ‾ σ , ρ d : = 1 − UnifInt σ , ρ d \overline{\operatorname{UnifInt}}_{\sigma,\rho}^{d}:=1-\operatorname{UnifInt}_{\sigma,\rho}^{d} UnifInt σ , ρ d := 1 − UnifInt σ , ρ d and UnifInt ‾ C : = 1 − UnifInt C \overline{\operatorname{UnifInt}}_C:=1-\operatorname{UnifInt}_C UnifInt C := 1 − UnifInt C for the densities of interpolation failure.
Example 4. We return to Example 3, where we presented the σ ⊤ ρ \sigma^{\top\rho} σ ⊤ ρ -cells for the logic of 3-forks. As we have observed, cells C C C of the form C s 0 , s 1 ρ C_{s_0,s_1}^{\rho} C s 0 , s 1 ρ and C s 0 , { s 1 , s 2 , s 3 } ρ C_{s_0,\{s_1,s_2,s_3\}}^{\rho} C s 0 , { s 1 , s 2 , s 3 } ρ contain no counterexamples to uniform σ \sigma σ -interpolation, and so UnifInt C = 1 \operatorname{UnifInt}_C=1 UnifInt C = 1 . On the other hand, each of the cells C C C of the form C s 0 , { s 1 , s 2 } ρ C_{s_0,\{s_1,s_2\}}^{\rho} C s 0 , { s 1 , s 2 } ρ contains counterexamples. It can be seen that every Φ ( D ) \Phi(D) Φ ( D ) , for a non-empty D ⊆ C D\subseteq C D ⊆ C , does not have uniform σ \sigma σ -interpolation iff D D D consists entirely of atoms of the form M s 0 r 0 , s 1 { r 1 , r 2 } , s 2 r ′ \mathcal{M}_{s_0r_0,s_1\{r_1,r_2\},s_2r'} M s 0 r 0 , s 1 { r 1 , r 2 } , s 2 r ′ or entirely of atoms of the form M s 0 r 0 , s 1 r ′ , s 2 { r 1 , r 2 } \mathcal{M}_{s_0r_0,s_1r',s_2\{r_1,r_2\}} M s 0 r 0 , s 1 r ′ , s 2 { r 1 , r 2 } . Let M = 2 ∣ ρ ∣ M=2^{|\rho|} M = 2 ∣ ρ ∣ . Then C C C contains M 3 M^3 M 3 atoms of the form M s 0 r 0 , { s 1 r 1 , s 2 r 2 } \mathcal{M}_{s_0r_0,\{s_1r_1,s_2r_2\}} M s 0 r 0 , { s 1 r 1 , s 2 r 2 } , 1 2 ( M − 1 ) M 3 \frac{1}{2}(M-1)M^3 2 1 ( M − 1 ) M 3 atoms of the form M s 0 r 0 , s 1 { r 1 , r 2 } , s 2 r ′ \mathcal{M}_{s_0r_0,s_1\{r_1,r_2\},s_2r'} M s 0 r 0 , s 1 { r 1 , r 2 } , s 2 r ′ and the same number of the form M s 0 r 0 , s 1 r ′ , s 2 { r 1 , r 2 } \mathcal{M}_{s_0r_0,s_1r',s_2\{r_1,r_2\}} M s 0 r 0 , s 1 r ′ , s 2 { r 1 , r 2 } ; in total, C C C contains M 4 M^4 M 4 atoms. Since the D D D without uniform σ \sigma σ -interpolation coincide with the non-empty subsets of the two latter types of atoms, we have
UnifInt ‾ C = 2 ⋅ 2 1 2 ( M − 1 ) M 3 − 1 2 M 4 . \overline{\operatorname{UnifInt}}_C=2\cdot\frac{2^{\frac{1}{2}(M-1)M^3}-1}{2^{M^4}}. UnifInt C = 2 ⋅ 2 M 4 2 2 1 ( M − 1 ) M 3 − 1 .
Let now N = 2 ∣ σ ∣ N=2^{|\sigma|} N = 2 ∣ σ ∣ . Since there are 1 2 ( N − 1 ) N 2 \frac{1}{2}(N-1)N^2 2 1 ( N − 1 ) N 2 cells of the form C s 0 , { s 1 , s 2 } ρ C_{s_0,\{s_1,s_2\}}^{\rho} C s 0 , { s 1 , s 2 } ρ , we obtain
UnifInt σ , ρ = ( 1 − 2 1 2 ( M − 1 ) M 3 − 1 2 M 4 − 1 ) 1 2 ( N − 1 ) N 2 . \operatorname{UnifInt}_{\sigma,\rho}=\left(1-\frac{2^{\frac{1}{2}(M-1)M^3}-1}{2^{M^4}-1}\right)^{\frac{1}{2}(N-1)N^2}. UnifInt σ , ρ = ( 1 − 2 M 4 − 1 2 2 1 ( M − 1 ) M 3 − 1 ) 2 1 ( N − 1 ) N 2 .
(We drop the modal depth d d d in UnifInt σ , ρ \operatorname{UnifInt}_{\sigma,\rho} UnifInt σ , ρ since every formula in the logic of 3-forks is equivalent to a formula of modal depth at most 1.)
We next consider the asymptotic behaviour of UnifInt σ n , ρ m d \operatorname{UnifInt}_{\sigma_n,\rho_m}^{d} UnifInt σ n , ρ m d on increasing signatures, which are the initial segments of infinite lists:
σ n : = { p 1 , … , p n } and ρ m : = { q 1 , … , q m } . \sigma_n:=\{p_1,\ldots,p_n\}\qquad\text{and}\qquad\rho_m:=\{q_1,\ldots,q_m\}. σ n := { p 1 , … , p n } and ρ m := { q 1 , … , q m } .
The modal depth d ≥ 0 d\geq0 d ≥ 0 , on the other hand, is fixed. We abbreviate UnifInt σ n , ρ m d \operatorname{UnifInt}_{\sigma_n,\rho_m}^{d} UnifInt σ n , ρ m d and UnifInt ‾ σ n , ρ m d \overline{\operatorname{UnifInt}}_{\sigma_n,\rho_m}^{d} UnifInt σ n , ρ m d by UnifInt n , m d \operatorname{UnifInt}_{n,m}^{d} UnifInt n , m d and UnifInt ‾ n , m d \overline{\operatorname{UnifInt}}_{n,m}^{d} UnifInt n , m d , respectively.
Asymptotics for n n n fixed and m → ∞ m\to\infty m → ∞
We begin by proving that, within any single σ ⊤ ρ \sigma^{\top\rho} σ ⊤ ρ -cell, the density of failure of uniform interpolation is at most 2 − 2 ∣ ρ ∣ 2^{-2^{|\rho|}} 2 − 2 ∣ ρ ∣ , and deduce that UnifInt n , m d → 1 \operatorname{UnifInt}_{n,m}^{d}\to1 UnifInt n , m d → 1 as m → ∞ m\to\infty m → ∞ , for each fixed n ≥ 0 n\geq0 n ≥ 0 .
Consider two disjoint signatures σ \sigma σ and ρ \rho ρ . Recall that a ρ \rho ρ -colour r r r is a set containing either q q q or ¬ q \neg q ¬ q , for each q ∈ ρ q\in\rho q ∈ ρ . Each ρ \rho ρ -colour r r r clearly defines a pinning of ρ \rho ρ , which we also denote by r r r . The next lemma identifies atoms that cannot occur in a subset D D D whose characteristic formula lacks a uniform interpolant.
Lemma 13. For every cell C ∈ C e l l σ , ρ d C\in\mathrm{Cell}_{\sigma,\rho}^{d} C ∈ Cell σ , ρ d , consider
M C : = { M r ∣ r a ρ -colour } ⊆ A σ ∪ ρ t , d , where M = ⋃ C . M_C:=\{\mathcal{M}^{r}\mid r\text{ a }\rho\text{-colour}\}\subseteq A_{\sigma\cup\rho}^{t,d},\qquad\text{where }\mathcal{M}=\bigcup C. M C := { M r ∣ r a ρ -colour } ⊆ A σ ∪ ρ t , d , where M = ⋃ C . Then M C ⊆ C ∖ D M_C\subseteq C\setminus D M C ⊆ C ∖ D , for every D ⊆ C D\subseteq C D ⊆ C such that Φ ( D ) \Phi(D) Φ ( D ) has no uniform σ \sigma σ -interpolant, and ∣ M C ∣ = 2 ∣ ρ ∣ |M_C|=2^{|\rho|} ∣ M C ∣ = 2 ∣ ρ ∣ .
Proof. First, each M r \mathcal{M}^{r} M r shares the σ ⊤ ρ \sigma^{\top\rho} σ ⊤ ρ -cell C C C , since pinning ρ \rho ρ leaves σ \sigma σ -formulas untouched; thus, M C ⊆ C M_C\subseteq C M C ⊆ C .
Next, suppose Φ ( D ) \Phi(D) Φ ( D ) , for D ⊆ C D\subseteq C D ⊆ C , has no uniform σ \sigma σ -interpolant. We show M C ∩ D = ∅ M_C\cap D=\varnothing M C ∩ D = ∅ . Suppose, towards a contradiction, that M r ∈ D \mathcal{M}^{r}\in D M r ∈ D for some ρ \rho ρ -colour r r r . Then Φ ( M r ) ⊨ Φ ( D ) \Phi(\mathcal{M}^{r})\models\Phi(D) Φ ( M r ) ⊨ Φ ( D ) , and so ( Φ ( M r ) ) − r ⊨ Φ ( D ) − r (\Phi(\mathcal{M}^{r}))^{-r}\models\Phi(D)^{-r} ( Φ ( M r ) ) − r ⊨ Φ ( D ) − r . By Lemma 9(1), ( Φ ( M r ) ) − r ≡ Φ ( M ) = Γ ( C ) (\Phi(\mathcal{M}^{r}))^{-r}\equiv\Phi(\mathcal{M})=\Gamma(C) ( Φ ( M r ) ) − r ≡ Φ ( M ) = Γ ( C ) , and so Γ ( C ) ⊨ Φ ( D ) − r \Gamma(C)\models\Phi(D)^{-r} Γ ( C ) ⊨ Φ ( D ) − r . On the other hand, as D ⊆ C D\subseteq C D ⊆ C , we have Φ ( D ) ⊨ Γ ( C ) \Phi(D)\models\Gamma(C) Φ ( D ) ⊨ Γ ( C ) . It follows that Φ ( D ) − r ⊨ Γ ( C ) \Phi(D)^{-r}\models\Gamma(C) Φ ( D ) − r ⊨ Γ ( C ) because Γ ( C ) \Gamma(C) Γ ( C ) is ρ \rho ρ -free and so ( Γ ( C ) ) − r = Γ ( C ) (\Gamma(C))^{-r}=\Gamma(C) ( Γ ( C ) ) − r = Γ ( C ) . Thus,
Φ ( D ) − r ≡ Γ ( C ) . (1) \Phi(D)^{-r}\equiv\Gamma(C). \tag*{(1)} Φ ( D ) − r ≡ Γ ( C ) . (1)
Let χ \chi χ be an arbitrary ρ \rho ρ -free formula with Φ ( D ) ⊨ χ \Phi(D) \models\chi Φ ( D ) ⊨ χ . Then Φ ( D ) ⊨ χ − r \Phi(D) \models\chi^{-r} Φ ( D ) ⊨ χ − r . As χ − r = χ \chi^{-r} = \chi χ − r = χ , we get Γ ( C ) ⊨ χ \Gamma(C) \models\chi Γ ( C ) ⊨ χ . So, formula Γ ( C ) \Gamma(C) Γ ( C ) is a uniform σ \sigma σ -interpolant of Φ ( D ) \Phi(D) Φ ( D ) , a contradiction. So, M C ∩ D = ∅ M_C \cap D = \varnothing M C ∩ D = ∅ .
Finally, atoms M r M^{r} M r and M r ′ M^{r'} M r ′ for distinct ρ \rho ρ -colours r ≠ r ′ r \ne r' r = r ′ , differ as some atomic formula q ∈ ρ ⊆ M L ∪ d q \in\rho\subseteq\mathrm{ML}_{\cup}^{d} q ∈ ρ ⊆ ML ∪ d has opposite values. This gives 2 ∣ ρ ∣ 2^{|\rho|} 2 ∣ ρ ∣ distinct members of M C M_C M C .
Observe that the proof actually shows that Γ ( C ) \Gamma(C) Γ ( C ) is a uniform σ \sigma σ -interpolant because every model of Γ ( C ) \Gamma(C) Γ ( C ) can be turned into a model of Φ ( D ) \Phi(D) Φ ( D ) by interpreting ρ \rho ρ according to the pinning function r r r .
Theorem 14. For every cell C ∈ C e l l σ , ρ d C \in\mathrm{Cell}_{\sigma,\rho}^{d} C ∈ Cell σ , ρ d ,
U n i f I n t ‾ C ≤ 1 2 2 ∣ ρ ∣ . \overline{\mathrm{UnifInt}}_{C} \le\frac{1}{2^{2^{|\rho|}}}. UnifInt C ≤ 2 2 ∣ ρ ∣ 1 .
Proof. Consider, for a cell C C C , the set M C M_C M C of Lemma 13. We have ∣ M C ∣ = 2 ∣ ρ ∣ |M_C| = 2^{|\rho|} ∣ M C ∣ = 2 ∣ ρ ∣ . Let D ⊆ C D \subseteq C D ⊆ C be any subset whose characteristic formula Φ ( D ) \Phi(D) Φ ( D ) lacks a uniform σ \sigma σ -interpolant. By Lemma 13, D ∩ M C = ∅ D \cap M_C = \varnothing D ∩ M C = ∅ . Thus, D ⊆ C ∖ M C D \subseteq C \setminus M_C D ⊆ C ∖ M C , and so there are at most 2 ∣ C ∣ − ∣ M C ∣ = 2 ∣ C ∣ − 2 ∣ ρ ∣ 2^{|C|-|M_C|} = 2^{|C|-2^{|\rho|}} 2 ∣ C ∣ − ∣ M C ∣ = 2 ∣ C ∣ − 2 ∣ ρ ∣ many such D D D . Dividing by 2 ∣ C ∣ 2^{|C|} 2 ∣ C ∣ , we obtain
U n i f I n t ‾ C = ∣ { D ⊆ C ∣ Φ ( D ) has no uniform σ -interpolant } ∣ 2 ∣ C ∣ ≤ 2 ∣ C ∣ − 2 ∣ ρ ∣ 2 ∣ C ∣ = 2 − 2 ∣ ρ ∣ . \overline{\mathrm{UnifInt}}_{C} = \frac{\left|\left\{D \subseteq C \mid\Phi(D)\text{ has no uniform }\sigma\text{-interpolant}\right\}\right|}{2^{|C|}} \le\frac{2^{|C|-2^{|\rho|}}}{2^{|C|}} = 2^{-2^{|\rho|}}. UnifInt C = 2 ∣ C ∣ ∣ { D ⊆ C ∣ Φ ( D ) has no uniform σ -interpolant } ∣ ≤ 2 ∣ C ∣ 2 ∣ C ∣ − 2 ∣ ρ ∣ = 2 − 2 ∣ ρ ∣ .
We can now show the main result of this section for increasing forgotten signatures ρ 0 , ρ 1 , … \rho_0,\rho_1,\ldots ρ 0 , ρ 1 , … :
Theorem 15. For any L L L and any fixed n n n , we have U n i f I n t ‾ n , m d ≤ ∣ A t σ n d ∣ ⋅ 2 − 2 m \overline{\mathrm{UnifInt}}_{n,m}^{d} \le|\mathrm{At}_{\sigma_n}^{d}| \cdot2^{-2^{m}} UnifInt n , m d ≤ ∣ At σ n d ∣ ⋅ 2 − 2 m , and so lim m → ∞ U n i f I n t ‾ n , m d = 1 \lim_{m\to\infty}\overline{\mathrm{UnifInt}}_{n,m}^{d} = 1 lim m → ∞ UnifInt n , m d = 1 .
Proof. By Proposition 6, there are exactly ∣ A t σ n d ∣ |\mathrm{At}_{\sigma_n}^{d}| ∣ At σ n d ∣ -many σ n ρ m \sigma_n\rho_m σ n ρ m -cells of A t σ n ∪ ρ m d \mathrm{At}_{\sigma_n\cup\rho_m}^{d} At σ n ∪ ρ m d . By the cell reduction principle (Theorem 12) and the upper bound on the number of counterexamples in a cell (Theorem 14), we obtain
U n i f I n t ‾ n , m d = ∏ C ∈ C e l l σ n , ρ m d U n i f I n t ‾ C ≥ ( 1 − 2 − 2 m ) ∣ A t σ n d ∣ ≥ 1 − ∣ A t σ n d ∣ ⋅ 2 − 2 m , \overline{\mathrm{UnifInt}}_{n,m}^{d} = \prod_{C\in\mathrm{Cell}_{\sigma_n,\rho_m}^{d}} \overline{\mathrm{UnifInt}}_{C} \ge(1-2^{-2^{m}})^{|\mathrm{At}_{\sigma_n}^{d}|} \ge1-|\mathrm{At}_{\sigma_n}^{d}|\cdot2^{-2^{m}}, UnifInt n , m d = C ∈ Cell σ n , ρ m d ∏ UnifInt C ≥ ( 1 − 2 − 2 m ) ∣ At σ n d ∣ ≥ 1 − ∣ At σ n d ∣ ⋅ 2 − 2 m , using ( 1 − x ) k ≥ 1 − k x (1-x)^k \ge1-kx ( 1 − x ) k ≥ 1 − k x for x ∈ [ 0 , 1 ] x \in[0,1] x ∈ [ 0 , 1 ] . Since ∣ A t σ n d ∣ |\mathrm{At}_{\sigma_n}^{d}| ∣ At σ n d ∣ does not depend on m m m , the right-hand side tends to 1.
Asymptotics for m m m fixed and n → ∞ n \to\infty n → ∞
Assume that L L L lacks uniform interpolation for some formula of depth at most d d d . Then there are n ∗ ≥ 0 n^* \ge0 n ∗ ≥ 0 , m ∗ ≥ 1 m^* \ge1 m ∗ ≥ 1 and D m ∗ ⊆ A t σ n ∗ ∪ ρ m ∗ d D_{m^*} \subseteq\mathrm{At}_{\sigma_{n^*}\cup\rho_{m^*}}^{d} D m ∗ ⊆ At σ n ∗ ∪ ρ m ∗ d such that Φ ( D m ∗ ) \Phi(D_{m^*}) Φ ( D m ∗ ) has no uniform σ n ∗ \sigma_{n^*} σ n ∗ -interpolant. D m ∗ D_{m^*} D m ∗ serves as a seed counterexample to uniform interpolation, from which further counterexamples in any A t σ n ∪ ρ m d \mathrm{At}_{\sigma_n\cup\rho_m}^{d} At σ n ∪ ρ m d with n ≥ n ∗ n \ge n^* n ≥ n ∗ and m ≥ m ∗ m \ge m^* m ≥ m ∗ are constructed. We then show U n i f I n t ‾ n , m d → 0 \overline{\mathrm{UnifInt}}_{n,m}^{d} \to0 UnifInt n , m d → 0 as n → ∞ n \to\infty n → ∞ , for any fixed m ≥ m ∗ m \ge m^* m ≥ m ∗ .
Consider three pairwise disjoint signatures σ ∗ \sigma^* σ ∗ , τ \tau τ and ρ \rho ρ , with σ = σ ∗ ∪ τ \sigma= \sigma^* \cup\tau σ = σ ∗ ∪ τ . Let ↑ τ \uparrow\tau ↑ τ be a pinning for τ \tau τ defined by taking ↑ τ ( p ) = ⊤ \uparrow\tau(p) = \top ↑ τ ( p ) = ⊤ , for all p ∈ τ p \in\tau p ∈ τ . We call ↑ τ \uparrow\tau ↑ τ a lifting. Then denote by D ↑ τ D^{\uparrow\tau} D ↑ τ and φ ↑ τ \varphi^{\uparrow\tau} φ ↑ τ the results of pinning τ \tau τ in D D D and φ \varphi φ , respectively, with ↑ τ \uparrow\tau ↑ τ . We show that the lifting operation preserves non-existence of uniform interpolation.
Lemma 16. For any D ⊆ A t σ ∗ ∪ ρ d D \subseteq\mathrm{At}_{\sigma^*\cup\rho}^{d} D ⊆ At σ ∗ ∪ ρ d , if Φ ( D ) \Phi(D) Φ ( D ) does not have a uniform σ ∗ \sigma^* σ ∗ -interpolant, then Φ ( D ↑ τ ) \Phi(D^{\uparrow\tau}) Φ ( D ↑ τ ) does not have a uniform σ \sigma σ -interpolant.
Proof. Suppose Φ ( D ↑ τ ) \Phi(D^{\uparrow\tau}) Φ ( D ↑ τ ) has a uniform σ \sigma σ -interpolant θ \theta θ . We have Φ ( D ↑ τ ) ⊨ θ \Phi(D^{\uparrow\tau}) \models\theta Φ ( D ↑ τ ) ⊨ θ , whence Φ ( D ) ⊨ θ ↓ τ \Phi(D) \models\theta^{\downarrow\tau} Φ ( D ) ⊨ θ ↓ τ because, by Lemma 9(1), Φ ( D ↑ τ ) ↓ τ ≡ Φ ( D ) \Phi(D^{\uparrow\tau})^{\downarrow\tau} \equiv\Phi(D) Φ ( D ↑ τ ) ↓ τ ≡ Φ ( D ) . Let ψ \psi ψ be any ρ \rho ρ -free consequence of Φ ( D ) \Phi(D) Φ ( D ) . Since τ \tau τ is disjoint from σ ∗ ∪ ρ \sigma^*\cup\rho σ ∗ ∪ ρ , variables of ψ \psi ψ outside σ ∗ \sigma^* σ ∗ may be renamed so that ψ \psi ψ contains no variable of τ \tau τ . Then if Φ ( D ↑ τ ) ⊨ Φ ( D ) ⊨ ψ \Phi(D^{\uparrow\tau}) \models\Phi(D) \models\psi Φ ( D ↑ τ ) ⊨ Φ ( D ) ⊨ ψ , and uniform σ \sigma σ -interpolation yields θ ⊨ ψ \theta\models\psi θ ⊨ ψ , whence θ ↓ τ ⊨ ψ \theta^{\downarrow\tau} \models\psi θ ↓ τ ⊨ ψ . The renaming can be reversed, as θ ↓ τ ⊨ ψ \theta^{\downarrow\tau} \models\psi θ ↓ τ ⊨ ψ uses only σ ∗ \sigma^* σ ∗ . Thus, θ ↓ τ \theta^{\downarrow\tau} θ ↓ τ is a uniform σ ∗ \sigma^* σ ∗ -interpolant of Φ ( D ) \Phi(D) Φ ( D ) , a contradiction.
On the other hand, since the flipping operation is involutive, it preserves both the existence and non-existence of uniform interpolants.
Lemma 17. For any ν ⊆ σ \nu\subseteq\sigma ν ⊆ σ and D ⊆ A t σ ∪ ρ d D \subseteq\mathrm{At}_{\sigma\cup\rho}^{d} D ⊆ At σ ∪ ρ d , formula Φ ( D ) \Phi(D) Φ ( D ) has a uniform σ \sigma σ -interpolant if and only if Φ ( D ν ) \Phi(D^\nu) Φ ( D ν ) does.
Proof. By Lemma 11(1), if θ \theta θ is a uniform σ \sigma σ -interpolant of Φ ( D ) \Phi(D) Φ ( D ) , then θ v \theta^v θ v is a uniform σ \sigma σ -interpolant of Φ ( D v ) \Phi(D^v) Φ ( D v ) : apply the same involutive substitution to each entailment, including consequences in arbitrary additional variables.
We now apply the lifting and flipping operations to show the main convergence result in this section.
Theorem 18. Let d ≥ 0 d \ge0 d ≥ 0 . If Φ ( D ) \Phi(D) Φ ( D ) , for D ⊆ At σ n ∗ ∪ ρ m ∗ d D \subseteq\operatorname{At}_{\sigma_{n^*}\cup\rho_{m^*}}^d D ⊆ At σ n ∗ ∪ ρ m ∗ d , has no uniform σ n ∗ \sigma_{n^*} σ n ∗ -interpolant, then, for every fixed m ≥ m ∗ m \ge m^* m ≥ m ∗ , there is a σ n ∗ ↑ ρ m \sigma_{n^*}^{\uparrow\rho_m} σ n ∗ ↑ ρ m -cell C m C_m C m such that
UnifInt n , m d ≤ ( 1 − 2 − ∣ C m ∣ ) 2 n , for all n ≥ n ∗ . \operatorname{UnifInt}_{n,m}^d \le\left(1-2^{-\lvert C_m\rvert}\right)^{2^n}, \qquad\text{for all } n \ge n^*. UnifInt n , m d ≤ ( 1 − 2 − ∣ C m ∣ ) 2 n , for all n ≥ n ∗ . Consequently, lim n → ∞ UnifInt n , m d = 0 \lim_{n\to\infty}\operatorname{UnifInt}_{n,m}^d=0 lim n → ∞ UnifInt n , m d = 0 .
Proof. Assume Φ ( D ) \Phi(D) Φ ( D ) with D ⊆ At σ n ∗ ∪ ρ m ∗ d D \subseteq\operatorname{At}_{\sigma_{n^*}\cup\rho_{m^*}}^d D ⊆ At σ n ∗ ∪ ρ m ∗ d , n ∗ ≥ 0 n^* \ge0 n ∗ ≥ 0 and m ∗ ≥ 1 m^* \ge1 m ∗ ≥ 1 , has no uniform σ n ∗ \sigma_{n^*} σ n ∗ -interpolant. Then, by Lemma 7, there is a cell C m ∗ ∈ Cell σ n ∗ , ρ m ∗ d C_{m^*}\in\operatorname{Cell}_{\sigma_{n^*},\rho_{m^*}}^d C m ∗ ∈ Cell σ n ∗ , ρ m ∗ d such that Φ ( C m ∗ ∩ D ) \Phi(C_{m^*}\cap D) Φ ( C m ∗ ∩ D ) has no uniform σ n ∗ \sigma_{n^*} σ n ∗ -interpolant. Consider now the trail ( C m ) m ≥ m ∗ (C_m)_{m\ge m^*} ( C m ) m ≥ m ∗ of C m ∗ C_{m^*} C m ∗ , the sequence defined by taking
C m = { M ∈ At σ n ∗ ∪ ρ m d ∣ M ⊆ ⋃ C m ∗ } ∈ Cell σ n ∗ , ρ m d , for m ≥ m ∗ . C_m=\{\mathcal{M}\in\operatorname{At}_{\sigma_{n^*}\cup\rho_m}^d\mid\mathcal{M}\subseteq\bigcup C_{m^*}\}\in\operatorname{Cell}_{\sigma_{n^*},\rho_m}^d,\qquad\text{for }m\ge m^*. C m = { M ∈ At σ n ∗ ∪ ρ m d ∣ M ⊆ ⋃ C m ∗ } ∈ Cell σ n ∗ , ρ m d , for m ≥ m ∗ .
Equivalently, M ∈ At σ n ∗ ∪ ρ m d \mathcal{M}\in\operatorname{At}_{\sigma_{n^*}\cup\rho_m}^d M ∈ At σ n ∗ ∪ ρ m d belongs to C m C_m C m if some (equivalently, all) ( M , w ) ∈ M (\mathcal{M},w)\in\mathcal{M} ( M , w ) ∈ M satisfy the cell characteristic formula Γ ( C m ∗ ) \Gamma(C_{m^*}) Γ ( C m ∗ ) . By Proposition 6, ⋃ C m = ⋃ C m ∗ \bigcup C_m=\bigcup C_{m^*} ⋃ C m = ⋃ C m ∗ , for each m ≥ m ∗ m\ge m^* m ≥ m ∗ , and so C m C_m C m trivially inherits from C m ∗ C_{m^*} C m ∗ a set D m ⊆ C m D_m\subseteq C_m D m ⊆ C m such that Φ ( D m ) \Phi(D_m) Φ ( D m ) has no uniform σ n ∗ \sigma_{n^*} σ n ∗ -interpolant. Fix m ≥ m ∗ m\ge m^* m ≥ m ∗ and consider any n ≥ n ∗ n\ge n^* n ≥ n ∗ . For every v ⊆ σ n v\subseteq\sigma_n v ⊆ σ n , let C m n , v : = ( C m ↑ τ n ) v C_m^{n,v}:=(C_m^{\uparrow\tau_n})^v C m n , v := ( C m ↑ τ n ) v , where τ n : = σ n ∖ σ n ∗ \tau_n:=\sigma_n\setminus\sigma_{n^*} τ n := σ n ∖ σ n ∗ . By Lemmas 9 and 11, C m n , v C_m^{n,v} C m n , v is a σ n ↑ ρ m \sigma_n^{\uparrow\rho_m} σ n ↑ ρ m -cell, which, by Lemmas 16 and 17, contains a counterexample to uniform σ n \sigma_n σ n -interpolation. Thus,
UnifInt ‾ C m n , v ≥ 1 2 ∣ C m n , v ∣ , for all n ≥ n ∗ and v ⊆ σ n . \overline{\operatorname{UnifInt}}_{C_m^{n,v}}\ge\frac{1}{2\lvert C_m^{n,v}\rvert},\qquad\text{for all }n\ge n^*\text{ and }v\subseteq\sigma_n. UnifInt C m n , v ≥ 2 ∣ C m n , v ∣ 1 , for all n ≥ n ∗ and v ⊆ σ n .
By Lemmas 9 and 11, the size of the cells is invariant under flipping and lifting: ∣ C m n , v ∣ = ∣ C m ∣ \lvert C_m^{n,v}\rvert=\lvert C_m\rvert ∣ C m n , v ∣ = ∣ C m ∣ , for all n ≥ n ∗ n\ge n^* n ≥ n ∗ and v ⊆ σ n v\subseteq\sigma_n v ⊆ σ n . By the cell reduction principle (Theorem 12), as 0 ≤ UnifInt C ≤ 1 0\le\operatorname{UnifInt}_C\le1 0 ≤ UnifInt C ≤ 1 , we obtain
UnifInt n , m d = ∏ C ∈ Cell σ n , ρ m d UnifInt C ≤ ∏ v ⊆ σ n UnifInt C m n , v ≤ ( 1 − 1 2 ∣ C m ∣ ) 2 n . \operatorname{UnifInt}_{n,m}^d
=\prod_{C\in\operatorname{Cell}_{\sigma_n,\rho_m}^d}\operatorname{UnifInt}_C
\le\prod_{v\subseteq\sigma_n}\operatorname{UnifInt}_{C_m^{n,v}}
\le\left(1-\frac{1}{2\lvert C_m\rvert}\right)^{2^n}. UnifInt n , m d = C ∈ Cell σ n , ρ m d ∏ UnifInt C ≤ v ⊆ σ n ∏ UnifInt C m n , v ≤ ( 1 − 2 ∣ C m ∣ 1 ) 2 n .
Since ∣ C m ∣ \lvert C_m\rvert ∣ C m ∣ does not depend on n n n , the claim follows.
We have determined the asymptotic behaviour of UnifInt n , m d \operatorname{UnifInt}_{n,m}^d UnifInt n , m d if either n n n or m m m is fixed. It would also be of interest to investigate what happens if both grow.
Example 5. In the 3-fork example, we obtain the following sharp transition when both signature sizes grow. Put
n c ( m ) : = 2 4 m + 2 3 m 6 . n_c(m):=\frac{2^{4m}+2^{3m}}{6}. n c ( m ) := 6 2 4 m + 2 3 m .
As n , m → ∞ n,m\to\infty n , m → ∞ , we have
UnifInt n , m = e − ( 1 + o ( 1 ) ) 2 3 ( n − n c ( m ) ) . \operatorname{UnifInt}_{n,m}=e^{-(1+o(1))2^{3(n-n_c(m))}}. UnifInt n , m = e − ( 1 + o ( 1 )) 2 3 ( n − n c ( m )) .
Consequently, for any sequence n = n ( m ) → ∞ n=n(m)\to\infty n = n ( m ) → ∞ as m → ∞ m\to\infty m → ∞ ,
lim m → ∞ UnifInt n ( m ) , m = { 1 , if n ( m ) − n c ( m ) → − ∞ , e − 2 3 t , if n ( m ) − n c ( m ) → t , 0 , if n ( m ) − n c ( m ) → + ∞ . \lim_{m\to\infty}\operatorname{UnifInt}_{n(m),m}
=
\begin{cases}
1, & \text{if }n(m)-n_c(m)\to-\infty,\\
e^{-2^{3t}}, & \text{if }n(m)-n_c(m)\to t,\\
0, & \text{if }n(m)-n_c(m)\to+\infty.
\end{cases} m → ∞ lim UnifInt n ( m ) , m = ⎩ ⎨ ⎧ 1 , e − 2 3 t , 0 , if n ( m ) − n c ( m ) → − ∞ , if n ( m ) − n c ( m ) → t , if n ( m ) − n c ( m ) → + ∞.
The result shows that the transition from almost all formulas having uniform interpolants to almost none is extremely sharp. It occurs when the number of retained variables is about n c ( m ) n_c(m) n c ( m ) , where m m m is the number of forgotten variables. Near this threshold, adding just a couple of retained variable can dramatically reduce the density: for example, for m = 3 m=3 m = 3 , we have n c ( m ) = 768 n_c(m)=768 n c ( m ) = 768 , and UnifInt n , m \operatorname{UnifInt}_{n,m} UnifInt n , m goes down from ≈ 0.88 \approx0.88 ≈ 0.88 for n = 767 n=767 n = 767 to ≈ 0.0003 \approx0.0003 ≈ 0.0003 for n = 769 n=769 n = 769 .
Craig Interpolants
A formula θ \theta θ is a Craig interpolant for formulas φ \varphi φ and ψ \psi ψ if
sig ( θ ) ⊆ sig ( φ ) ∩ sig ( ψ ) , φ ⊨ θ , ψ ⊨ ¬ θ . \operatorname{sig}(\theta) \subseteq\operatorname{sig}(\varphi) \cap\operatorname{sig}(\psi), \qquad\varphi\models\theta, \qquad\psi\models\neg\theta. sig ( θ ) ⊆ sig ( φ ) ∩ sig ( ψ ) , φ ⊨ θ , ψ ⊨ ¬ θ .
A logic L L L has the Craig interpolation property (CIP) if every pair φ , ψ \varphi,\psi φ , ψ with φ ∧ ψ ⊨ ⊥ \varphi\land\psi\models\bot φ ∧ ψ ⊨ ⊥ admits a Craig interpolant. For quasi-normal modal logics, this is equivalent, by the deduction theorem for ⊨ \models ⊨ , to the usual formulation of CIP in terms of implications, also called local interpolation ; see [12 ] .
As in the uniform-interpolation setting, we fix a modal depth d ≥ 0 d \ge0 d ≥ 0 and disjoint signatures σ \sigma σ and ρ \rho ρ . Let ρ ′ \rho' ρ ′ be a fresh disjoint copy of ρ \rho ρ . We consider pairs of formulas φ , ψ \varphi,\psi φ , ψ of depth at most d d d with
sig ( φ ) = σ ∪ ρ and sig ( ψ ) = σ ∪ ρ ′ . \operatorname{sig}(\varphi) = \sigma\cup\rho\qquad\text{and} \qquad\operatorname{sig}(\psi) = \sigma\cup\rho'. sig ( φ ) = σ ∪ ρ and sig ( ψ ) = σ ∪ ρ ′ .
Thus, σ \sigma σ is precisely their shared signature, while ρ \rho ρ and ρ ′ \rho' ρ ′ contain the variables occurring only on the left and right, respectively. As formulas are counted up to L L L -equivalence, requiring every variable in the prescribed signatures to occur does not affect the count.
The fixed bijection from ρ \rho ρ to ρ ′ \rho' ρ ′ induces a bijection A t σ ∪ ρ d → A t σ ∪ ρ ′ d \mathrm{At}^{d}_{\sigma\cup\rho} \to\mathrm{At}^{d}_{\sigma\cup\rho'} At σ ∪ ρ d → At σ ∪ ρ ′ d on atoms: the result of renaming ρ \rho ρ to ρ ′ \rho' ρ ′ in M \mathcal{M} M is denoted by M ′ \mathcal{M}' M ′ , and D ′ = { M ′ ∣ M ∈ D } D' = \{\mathcal{M}' \mid\mathcal{M} \in D\} D ′ = { M ′ ∣ M ∈ D } , for D ⊆ A t σ ∪ ρ d D \subseteq\mathrm{At}^{d}_{\sigma\cup\rho} D ⊆ At σ ∪ ρ d .
Definition 9. Let D 1 , D 2 ⊆ A t σ ∪ ρ d D_1,D_2 \subseteq\mathrm{At}^{d}_{\sigma\cup\rho} D 1 , D 2 ⊆ At σ ∪ ρ d .
The pair ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent if Φ ( D 1 ) ∧ Φ ( D 2 ′ ) ⊨ ⊥ \Phi(D_1) \land\Phi(D'_2) \models\bot Φ ( D 1 ) ∧ Φ ( D 2 ′ ) ⊨ ⊥ . The set of all inconsistent pairs in A t σ ∪ ρ d \mathrm{At}^{d}_{\sigma\cup\rho} At σ ∪ ρ d is denoted by I n C o n σ , ρ d \mathrm{InCon}^{d}_{\sigma,\rho} InCon σ , ρ d .
A Craig σ \sigma σ -interpolant for ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is a formula θ ∈ M L σ \theta\in\mathrm{ML}_{\sigma} θ ∈ ML σ such that
Φ ( D 1 ) ⊨ θ , θ ⊨ ¬ Φ ( D 2 ′ ) . \Phi(D_1) \models\theta, \qquad\theta\models\neg\Phi(D'_2). Φ ( D 1 ) ⊨ θ , θ ⊨ ¬Φ ( D 2 ′ ) . The set of all pairs in A t σ ∪ ρ d \mathrm{At}^{d}_{\sigma\cup\rho} At σ ∪ ρ d admitting a Craig σ \sigma σ -interpolant is denoted by C r a i g σ , ρ d \mathrm{Craig}^{d}_{\sigma,\rho} Craig σ , ρ d .
Clearly, C r a i g σ , ρ d ⊆ I n C o n σ , ρ d \mathrm{Craig}^{d}_{\sigma,\rho} \subseteq\mathrm{InCon}^{d}_{\sigma,\rho} Craig σ , ρ d ⊆ InCon σ , ρ d . In particular, pairs with an empty component belong to both sets. We next introduce a relation on atoms that describes inconsistency of pairs. Two atoms M 1 , M 2 ∈ A t σ ∪ ρ d \mathcal{M}_1,\mathcal{M}_2 \in\mathrm{At}^{d}_{\sigma\cup\rho} M 1 , M 2 ∈ At σ ∪ ρ d are incompatible , written M 1 ⌢ M 2 \mathcal{M}_1 \mathrel{\smallfrown} \mathcal{M}_2 M 1 ⌢ M 2 , if
M 1 ∩ M 2 ′ = ∅ . \mathcal{M}_1 \cap\mathcal{M}'_2 = \varnothing. M 1 ∩ M 2 ′ = ∅ .
Otherwise, they are compatible .
Proposition 19. Let D 1 , D 2 ⊆ A t σ ∪ ρ d D_1,D_2 \subseteq\mathrm{At}^{d}_{\sigma\cup\rho} D 1 , D 2 ⊆ At σ ∪ ρ d . Then ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent if and only if
M 1 ⌢ M 2 , for all M 1 ∈ D 1 and M 2 ∈ D 2 . \mathcal{M}_1 \mathrel{\smallfrown} \mathcal{M}_2, \qquad\text{for all } \mathcal{M}_1 \in D_1 \text{ and } \mathcal{M}_2 \in D_2. M 1 ⌢ M 2 , for all M 1 ∈ D 1 and M 2 ∈ D 2 .
Proof. The conjunction Φ ( D 1 ) ∧ Φ ( D 2 ′ ) \Phi(D_1) \land\Phi(D'_2) Φ ( D 1 ) ∧ Φ ( D 2 ′ ) is satisfiable if and only if M 1 ∩ M 2 ′ ≠ ∅ \mathcal{M}_1 \cap\mathcal{M}'_2 \ne\varnothing M 1 ∩ M 2 ′ = ∅ for some M 1 ∈ D 1 \mathcal{M}_1 \in D_1 M 1 ∈ D 1 and M 2 ∈ D 2 \mathcal{M}_2 \in D_2 M 2 ∈ D 2 .
Example 6. The 3-fork formulas φ \varphi φ and ψ \psi ψ from Examples 2 and 3 illustrate incompatibility. Their conjunction is inconsistent: φ \varphi φ requires at least two successors satisfying p p p , while ψ \psi ψ requires at least two satisfying ¬ p \neg p ¬ p , and the frame has only three successors. On the other hand, the two formulas have the same { p } \{p\} { p } -consequences. Consequently, although φ ⊨ ¬ ψ \varphi\models\neg\psi φ ⊨ ¬ ψ , there is no Craig interpolant over { p } \{p\} { p } . Equivalently, if M φ \mathcal{M}_{\varphi} M φ and M ψ \mathcal{M}_{\psi} M ψ denote the corresponding atoms after identifying q ′ q' q ′ with q q q , then M φ ⌢ M ψ \mathcal{M}_{\varphi} \mathrel{\smallfrown} \mathcal{M}_{\psi} M φ ⌢ M ψ , even though both atoms belong to the same σ ⌈ p ⌉ \sigma^{\lceil p\rceil} σ ⌈ p ⌉ -cell.
For every cell C ∈ C e l l σ , ρ d C \in\mathrm{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d , let I n C o n C \mathrm{InCon}_{C} InCon C and C r a i g C \mathrm{Craig}_{C} Craig C denote the sets of inconsistent pairs and pairs admitting a Craig σ \sigma σ -interpolant, respectively, among D 1 , D 2 ⊆ C D_1,D_2 \subseteq C D 1 , D 2 ⊆ C .
Lemma 20. For all D 1 , D 2 ⊆ A t σ ∪ ρ d D_1,D_2 \subseteq\mathrm{At}^{d}_{\sigma\cup\rho} D 1 , D 2 ⊆ At σ ∪ ρ d ,
( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent if and only if ( D 1 ∩ C , D 2 ∩ C ) ∈ I n C o n C (D_1 \cap C,D_2 \cap C) \in\mathrm{InCon}_{C} ( D 1 ∩ C , D 2 ∩ C ) ∈ InCon C , for every C ∈ C e l l σ , ρ d C \in\mathrm{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d ;
( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) admits a Craig σ \sigma σ -interpolant if and only if ( D 1 ∩ C , D 2 ∩ C ) ∈ C r a i g C (D_1 \cap C,D_2 \cap C) \in\mathrm{Craig}_{C} ( D 1 ∩ C , D 2 ∩ C ) ∈ Craig C , for every C ∈ C e l l σ , ρ d C \in\mathrm{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d .
Proof. (1) By Proposition 19, the pair ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent if and only if every M 1 ∈ D 1 \mathcal{M}_1 \in D_1 M 1 ∈ D 1 is incompatible with every M 2 ∈ D 2 \mathcal{M}_2 \in D_2 M 2 ∈ D 2 . If M 1 \mathcal{M}_1 M 1 and M 2 \mathcal{M}_2 M 2 are compatible, then M 1 ∩ M 2 ′ ≠ ∅ \mathcal{M}_1 \cap\mathcal{M}'_2 \ne\varnothing M 1 ∩ M 2 ′ = ∅ implies that M 1 \mathcal{M}_1 M 1 and M 2 \mathcal{M}_2 M 2 belong to the same σ τ ρ \sigma^{\tau\rho} σ τ ρ -cell. Thus, ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent if and only if ( D 1 ∩ C , D 2 ∩ C ) (D_1 \cap C,D_2 \cap C) ( D 1 ∩ C , D 2 ∩ C ) is inconsistent for every σ τ ρ \sigma^{\tau\rho} σ τ ρ -cell C C C .
(2) Recall from Proposition 6 and the definition of cell characteristic formulas preceding Lemma 7 that Γ ( C ) \Gamma(C) Γ ( C ) is a σ \sigma σ -formula and, for every cell C C C , Φ ( D 1 ) ∧ Γ ( C ) ≡ Φ ( D 1 ∩ C ) \Phi(D_1) \land\Gamma(C) \equiv\Phi(D_1 \cap C) Φ ( D 1 ) ∧ Γ ( C ) ≡ Φ ( D 1 ∩ C ) and Φ ( D 2 ′ ) ∧ Γ ( C ) ≡ Φ ( ( D 2 ∩ C ) ′ ) \Phi(D'_2) \land\Gamma(C) \equiv\Phi((D_2 \cap C)') Φ ( D 2 ′ ) ∧ Γ ( C ) ≡ Φ (( D 2 ∩ C ) ′ ) . The second equivalence follows from the first by renaming ρ \rho ρ to ρ ′ \rho' ρ ′ , since Γ ( C ) \Gamma(C) Γ ( C ) contains only variables from σ \sigma σ . Moreover, the disjunction of all Γ ( C ) \Gamma(C) Γ ( C ) is valid. Now, if θ ∈ M L σ \theta\in\mathrm{ML}_{\sigma} θ ∈ ML σ is a Craig σ \sigma σ -interpolant for ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) , then θ ∧ Γ ( C ) \theta\land\Gamma(C) θ ∧ Γ ( C ) is a Craig σ \sigma σ -interpolant for ( D 1 ∩ C , D 2 ∩ C ) (D_1 \cap C,D_2 \cap C) ( D 1 ∩ C , D 2 ∩ C ) . Conversely, if, for every cell C C C , there is a Craig σ \sigma σ -interpolant θ C \theta_C θ C for ( D 1 ∩ C , D 2 ∩ C ) (D_1 \cap C,D_2 \cap C) ( D 1 ∩ C , D 2 ∩ C ) , then ⋁ C ∈ C e l l σ , ρ d ( Γ ( C ) ∧ θ C ) \bigvee_{C \in\mathrm{Cell}^{d}_{\sigma,\rho}}(\Gamma(C) \land\theta_C) ⋁ C ∈ Cell σ , ρ d ( Γ ( C ) ∧ θ C ) is a Craig σ \sigma σ -interpolant for ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) . □
Density of Craig Interpolants
We now introduce the density of Craig interpolants and establish a cell reduction principle analogous to Theorem 12.
Theorem 21 (cell reduction principle for Craig interpolation).
∣ I n C o n σ , ρ d ∣ = ∏ C ∈ C e l l σ , ρ d ∣ I n C o n C ∣ , ∣ C r a i g σ , ρ d ∣ = ∏ C ∈ C e l l σ , ρ d ∣ C r a i g C ∣ . |\mathrm{InCon}^{d}_{\sigma,\rho}| = \prod_{C \in\mathrm{Cell}^{d}_{\sigma,\rho}} |\mathrm{InCon}_C|,\qquad|\mathrm{Craig}^{d}_{\sigma,\rho}| = \prod_{C \in\mathrm{Cell}^{d}_{\sigma,\rho}} |\mathrm{Craig}_C|. ∣ InCon σ , ρ d ∣ = C ∈ Cell σ , ρ d ∏ ∣ InCon C ∣ , ∣ Craig σ , ρ d ∣ = C ∈ Cell σ , ρ d ∏ ∣ Craig C ∣.
Proof. Since the cells partition A t σ ∪ ρ d \mathrm{At}^{d}_{\sigma\cup\rho} At σ ∪ ρ d , the map ( D 1 , D 2 ) ↦ ( ( D 1 ∩ C , D 2 ∩ C ) ) C ∈ C e l l σ , ρ d (D_1,D_2) \mapsto((D_1 \cap C,D_2 \cap C))_{C \in\mathrm{Cell}^{d}_{\sigma,\rho}} ( D 1 , D 2 ) ↦ (( D 1 ∩ C , D 2 ∩ C ) ) C ∈ Cell σ , ρ d is a bijection from 2 A t σ ∪ ρ d × 2 A t σ ∪ ρ d 2^{\mathrm{At}^{d}_{\sigma\cup\rho}} \times2^{\mathrm{At}^{d}_{\sigma\cup\rho}} 2 At σ ∪ ρ d × 2 At σ ∪ ρ d onto ∏ C ∈ C e l l σ , ρ d ( 2 C × 2 C ) \prod_{C \in\mathrm{Cell}^{d}_{\sigma,\rho}}(2^C \times2^C) ∏ C ∈ Cell σ , ρ d ( 2 C × 2 C ) . By Lemma 20, this bijection restricts to bijections
I n C o n σ , ρ d ⟶ ∏ C ∈ C e l l σ , ρ d I n C o n C and C r a i g σ , ρ d ⟶ ∏ C ∈ C e l l σ , ρ d C r a i g C . \mathrm{InCon}^{d}_{\sigma,\rho} \longrightarrow\prod_{C \in\mathrm{Cell}^{d}_{\sigma,\rho}} \mathrm{InCon}_C \qquad\text{and}\qquad\mathrm{Craig}^{d}_{\sigma,\rho} \longrightarrow\prod_{C \in\mathrm{Cell}^{d}_{\sigma,\rho}} \mathrm{Craig}_C. InCon σ , ρ d ⟶ C ∈ Cell σ , ρ d ∏ InCon C and Craig σ , ρ d ⟶ C ∈ Cell σ , ρ d ∏ Craig C .
The result follows by taking cardinalities. □
We define the density of Craig interpolation as the proportion of inconsistent pairs that admit a Craig interpolant:
D C r a i g σ , ρ d : = ∣ C r a i g σ , ρ d ∣ ∣ I n C o n σ , ρ d ∣ , D C r a i g C : = ∣ C r a i g C ∣ ∣ I n C o n C ∣ . \mathrm{DCraig}^{d}_{\sigma,\rho} := \frac{|\mathrm{Craig}^{d}_{\sigma,\rho}|}{|\mathrm{InCon}^{d}_{\sigma,\rho}|},\qquad\mathrm{DCraig}_C := \frac{|\mathrm{Craig}_C|}{|\mathrm{InCon}_C|}. DCraig σ , ρ d := ∣ InCon σ , ρ d ∣ ∣ Craig σ , ρ d ∣ , DCraig C := ∣ InCon C ∣ ∣ Craig C ∣ .
The denominators are non-zero, since pairs with an empty component are inconsistent. By Theorem 21,
D C r a i g σ , ρ d = ∏ C ∈ C e l l σ , ρ d D C r a i g C . \mathrm{DCraig}^{d}_{\sigma,\rho} = \prod_{C \in\mathrm{Cell}^{d}_{\sigma,\rho}} \mathrm{DCraig}_C. DCraig σ , ρ d = C ∈ Cell σ , ρ d ∏ DCraig C .
In particular, D C r a i g σ , ρ d = 1 \mathrm{DCraig}^{d}_{\sigma,\rho} = 1 DCraig σ , ρ d = 1 if and only if D C r a i g C = 1 \mathrm{DCraig}_C = 1 DCraig C = 1 for every cell C ∈ C e l l σ , ρ d C \in\mathrm{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d . Thus, at each fixed depth, every inconsistent pair admits a Craig interpolant if and only if this holds within every cell. We also write D C r a i g ‾ σ , ρ d : = 1 − D C r a i g σ , ρ d \overline{\mathrm{DCraig}}^{d}_{\sigma,\rho} := 1 - \mathrm{DCraig}^{d}_{\sigma,\rho} DCraig σ , ρ d := 1 − DCraig σ , ρ d and D C r a i g ‾ C : = 1 − D C r a i g C \overline{\mathrm{DCraig}}_C := 1 - \mathrm{DCraig}_C DCraig C := 1 − DCraig C for the densities of Craig interpolation failure .
We next consider the asymptotic behaviour of D C r a i g σ n , ρ m d \mathrm{DCraig}^{d}_{\sigma_n,\rho_m} DCraig σ n , ρ m d on increasing signatures (but fixed modal depth). We abbreviate D C r a i g σ n , ρ m d \mathrm{DCraig}^{d}_{\sigma_n,\rho_m} DCraig σ n , ρ m d by D C r a i g n , m d \mathrm{DCraig}^{d}_{n,m} DCraig n , m d .
Asymptotics for m m m fixed and n → ∞ n \to\infty n → ∞
Suppose that, at the fixed depth d d d , there is an inconsistent pair without a Craig interpolant. Thus, for some n ∗ ≥ 0 n^* \ge0 n ∗ ≥ 0 and m ∗ ≥ 1 m^* \ge1 m ∗ ≥ 1 , there are D 1 , D 2 ⊆ A t σ n ∗ ∪ ρ m ∗ d D_1,D_2 \subseteq\mathrm{At}^{d}_{\sigma_{n^*} \cup\rho_{m^*}} D 1 , D 2 ⊆ At σ n ∗ ∪ ρ m ∗ d such that ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent and admits no Craig σ n ∗ \sigma_{n^*} σ n ∗ -interpolant. We show that, for every fixed m ≥ m ∗ m \ge m^* m ≥ m ∗ , the density D C r a i g n , m d \mathrm{DCraig}^{d}_{n,m} DCraig n , m d tends to 0 0 0 as n → ∞ n \to\infty n → ∞ . We use the lifting and flipping operations introduced in Section 4.2 (see also Section 3.2). The following lemma shows that lifting preserves both inconsistency and Craig interpolation.
Lemma 22 (Craig lifting). Let σ ∗ \sigma^* σ ∗ , τ \tau τ and ρ \rho ρ be pairwise disjoint signatures. For all D 1 , D 2 ⊆ A t σ ∗ ∪ ρ d D_1,D_2 \subseteq\mathrm{At}^{d}_{\sigma^* \cup\rho} D 1 , D 2 ⊆ At σ ∗ ∪ ρ d
( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent if and only if ( D 1 ↑ τ , D 2 ↑ τ ) (D_1^{\uparrow\tau},D_2^{\uparrow\tau}) ( D 1 ↑ τ , D 2 ↑ τ ) is inconsistent;
(ii) ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) admits a Craig σ ∗ \sigma^* σ ∗ -interpolant if and only if ( D 1 ↑ τ , D 2 ↑ τ ) (D_1^{\uparrow\tau},D_2^{\uparrow\tau}) ( D 1 ↑ τ , D 2 ↑ τ ) admits a Craig ( σ ∗ ∪ τ ) (\sigma^*\cup\tau) ( σ ∗ ∪ τ ) -interpolant.
Proof. Since ρ ′ \rho' ρ ′ is disjoint from τ \tau τ , lifting commutes with the renaming of ρ \rho ρ to ρ ′ \rho' ρ ′ . Thus, by Lemma 9(1), Φ ( D 1 ↑ τ ) ≡ Φ ( D 1 ) ∧ ⋀ p ∈ τ □ ≤ d p \Phi(D_1^{\uparrow\tau}) \equiv\Phi(D_1)\land\bigwedge_{p\in\tau}\Box_{\leq d}p Φ ( D 1 ↑ τ ) ≡ Φ ( D 1 ) ∧ ⋀ p ∈ τ □ ≤ d p and Φ ( ( D 2 ↑ τ ) ′ ) ≡ Φ ( D 2 ′ ) ∧ ⋀ p ∈ τ □ ≤ d p \Phi((D_2^{\uparrow\tau})')\equiv\Phi(D_2')\land\bigwedge_{p\in\tau}\Box_{\leq d}p Φ (( D 2 ↑ τ ) ′ ) ≡ Φ ( D 2 ′ ) ∧ ⋀ p ∈ τ □ ≤ d p . In particular, Φ ( D 1 ↑ τ ) ↓ τ ≡ Φ ( D 1 ) \Phi(D_1^{\uparrow\tau})^{\downarrow\tau}\equiv\Phi(D_1) Φ ( D 1 ↑ τ ) ↓ τ ≡ Φ ( D 1 ) and Φ ( ( D 2 ↑ τ ) ′ ) ↓ τ ≡ Φ ( D 2 ′ ) \Phi((D_2^{\uparrow\tau})')^{\downarrow\tau}\equiv\Phi(D_2') Φ (( D 2 ↑ τ ) ′ ) ↓ τ ≡ Φ ( D 2 ′ ) . Recall that, by Lemma 8(1), pinning preserves entailment.
(i) If ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent, then Φ ( D 1 ↑ τ ) ⊨ Φ ( D 1 ) \Phi(D_1^{\uparrow\tau})\models\Phi(D_1) Φ ( D 1 ↑ τ ) ⊨ Φ ( D 1 ) and Φ ( ( D 2 ↑ τ ) ′ ) ⊨ Φ ( D 2 ′ ) \Phi((D_2^{\uparrow\tau})')\models\Phi(D_2') Φ (( D 2 ↑ τ ) ′ ) ⊨ Φ ( D 2 ′ ) imply that ( D 1 ↑ τ , ( D 2 ↑ τ ) ′ ) (D_1^{\uparrow\tau},(D_2^{\uparrow\tau})') ( D 1 ↑ τ , ( D 2 ↑ τ ) ′ ) is inconsistent. Conversely, suppose that ( D 1 ↑ τ , D 2 ↑ τ ) (D_1^{\uparrow\tau},D_2^{\uparrow\tau}) ( D 1 ↑ τ , D 2 ↑ τ ) is inconsistent. Applying the pinning substitution ⋅ ↓ τ \cdot^{\downarrow\tau} ⋅ ↓ τ and using the equivalences above gives Φ ( D 1 ) ∧ Φ ( D 2 ′ ) ⊨ ⊥ \Phi(D_1)\land\Phi(D_2')\models\bot Φ ( D 1 ) ∧ Φ ( D 2 ′ ) ⊨ ⊥ . Hence, ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent.
(ii) Suppose first that θ ∈ M L σ ∗ \theta\in\mathrm{ML}_{\sigma^*} θ ∈ ML σ ∗ is a Craig σ ∗ \sigma^* σ ∗ -interpolant for ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) . Then Φ ( D 1 ) ⊨ θ \Phi(D_1)\models\theta Φ ( D 1 ) ⊨ θ and θ ⊨ ¬ Φ ( D 2 ′ ) \theta\models\neg\Phi(D_2') θ ⊨ ¬Φ ( D 2 ′ ) . The equivalences established above imply Φ ( D 1 ↑ τ ) ⊨ θ \Phi(D_1^{\uparrow\tau})\models\theta Φ ( D 1 ↑ τ ) ⊨ θ and θ ⊨ ¬ Φ ( ( D 2 ↑ τ ) ′ ) \theta\models\neg\Phi((D_2^{\uparrow\tau})') θ ⊨ ¬Φ (( D 2 ↑ τ ) ′ ) . Since θ ∈ M L σ ∗ ⊆ M L σ ∗ ∪ τ \theta\in\mathrm{ML}_{\sigma^*}\subseteq\mathrm{ML}_{\sigma^*\cup\tau} θ ∈ ML σ ∗ ⊆ ML σ ∗ ∪ τ , it is also a Craig ( σ ∗ ∪ τ ) (\sigma^*\cup\tau) ( σ ∗ ∪ τ ) -interpolant for ( D 1 ↑ τ , D 2 ↑ τ ) (D_1^{\uparrow\tau},D_2^{\uparrow\tau}) ( D 1 ↑ τ , D 2 ↑ τ ) .
Conversely, suppose that θ ∈ M L σ ∗ ∪ τ \theta\in\mathrm{ML}_{\sigma^*\cup\tau} θ ∈ ML σ ∗ ∪ τ is a Craig ( σ ∗ ∪ τ ) (\sigma^*\cup\tau) ( σ ∗ ∪ τ ) -interpolant for ( D 1 ↑ τ , D 2 ↑ τ ) (D_1^{\uparrow\tau},D_2^{\uparrow\tau}) ( D 1 ↑ τ , D 2 ↑ τ ) . Applying ⋅ ↓ τ \cdot^{\downarrow\tau} ⋅ ↓ τ to both entailments and using Lemma 8(1) together with the equivalences above, we obtain Φ ( D 1 ) ⊨ θ ↓ τ \Phi(D_1)\models\theta^{\downarrow\tau} Φ ( D 1 ) ⊨ θ ↓ τ and θ ↓ τ ⊨ ¬ Φ ( D 2 ′ ) \theta^{\downarrow\tau}\models\neg\Phi(D_2') θ ↓ τ ⊨ ¬Φ ( D 2 ′ ) . Since θ ∈ M L σ ∗ ∪ τ \theta\in\mathrm{ML}_{\sigma^*\cup\tau} θ ∈ ML σ ∗ ∪ τ , we have θ ↓ τ ∈ M L σ ∗ \theta^{\downarrow\tau}\in\mathrm{ML}_{\sigma^*} θ ↓ τ ∈ ML σ ∗ . Thus θ ↓ τ \theta^{\downarrow\tau} θ ↓ τ is a Craig σ ∗ \sigma^* σ ∗ -interpolant for ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) . □
By Lemma 11, flipping also preserves inconsistency and Craig interpolation:
Lemma 23. Let σ \sigma σ and ρ \rho ρ be disjoint signatures. For all v ⊆ σ v\subseteq\sigma v ⊆ σ and all D 1 , D 2 ⊆ A t σ ∪ ρ d D_1,D_2\subseteq\mathrm{At}_{\sigma\cup\rho}^{d} D 1 , D 2 ⊆ At σ ∪ ρ d ,
(i) ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is inconsistent if and only if ( D 1 v , D 2 v ) (D_1^v,D_2^v) ( D 1 v , D 2 v ) is inconsistent;
(ii) ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) admits a Craig σ \sigma σ -interpolant if and only if ( D 1 v , D 2 v ) (D_1^v,D_2^v) ( D 1 v , D 2 v ) does.
We can now prove the Craig analogue of Theorem 18.
Theorem 24. Let d ≥ 0 d\geq0 d ≥ 0 . If there is an inconsistent pair D 1 , D 2 ⊆ A t σ n ∗ ∪ ρ m ∗ d D_1,D_2\subseteq\mathrm{At}_{\sigma_{n^*}\cup\rho_{m^*}}^{d} D 1 , D 2 ⊆ At σ n ∗ ∪ ρ m ∗ d without a Craig σ n ∗ \sigma_{n^*} σ n ∗ -interpolant, then, for every fixed m ≥ m ∗ m\geq m^* m ≥ m ∗ , there is a σ n ∗ ↑ ρ m ∗ \sigma_{n^*}^{\uparrow\rho_{m^*}} σ n ∗ ↑ ρ m ∗ -cell C m C_m C m such that
D C r a i g n , m d ≤ ( 1 − 2 − 2 ∣ C m ∣ ) 2 n , for all n ≥ n ∗ . \mathrm{DCraig}_{n,m}^{d}\leq\left(1-2^{-2|C_m|}\right)^{2^n},\qquad\text{for all }n\geq n^*. DCraig n , m d ≤ ( 1 − 2 − 2∣ C m ∣ ) 2 n , for all n ≥ n ∗ . Consequently, lim n → ∞ D C r a i g n , m d = 0 \lim_{n\to\infty}\mathrm{DCraig}_{n,m}^{d}=0 lim n → ∞ DCraig n , m d = 0 .
Proof. Assume that D 1 , D 2 ⊆ A t σ n ∗ ∪ ρ m ∗ d D_1,D_2\subseteq\mathrm{At}_{\sigma_{n^*}\cup\rho_{m^*}}^{d} D 1 , D 2 ⊆ At σ n ∗ ∪ ρ m ∗ d is an inconsistent pair without a Craig σ n ∗ \sigma_{n^*} σ n ∗ -interpolant. By Lemma 20 , there is a σ n ∗ ↑ ρ m ∗ \sigma_{n^*}^{\uparrow\rho_{m^*}} σ n ∗ ↑ ρ m ∗ -cell C m ∗ C_{m^*} C m ∗ such that ( D 1 ∩ C m ∗ , D 2 ∩ C m ∗ ) (D_1\cap C_{m^*},D_2\cap C_{m^*}) ( D 1 ∩ C m ∗ , D 2 ∩ C m ∗ ) is inconsistent and admits no Craig σ n ∗ \sigma_{n^*} σ n ∗ -interpolant. Consider the trail ( C m ) m ≥ m ∗ (C_m)_{m\geq m^*} ( C m ) m ≥ m ∗ of C m ∗ C_{m^*} C m ∗ defined in the proof of Theorem 18 and recall that ⋃ C m = ⋃ C m ∗ \bigcup C_m=\bigcup C_{m^*} ⋃ C m = ⋃ C m ∗ , for each m ≥ m ∗ m\geq m^* m ≥ m ∗ . As in the proof of Theorem 18 , for every m ≥ m ∗ m\geq m^* m ≥ m ∗ , there are sets D 1 , m , D 2 , m ⊆ C m D_{1,m},D_{2,m}\subseteq C_m D 1 , m , D 2 , m ⊆ C m such that ( D 1 , m , D 2 , m ) (D_{1,m},D_{2,m}) ( D 1 , m , D 2 , m ) is inconsistent and admits no Craig σ n ∗ \sigma_{n^*} σ n ∗ -interpolant.
Fix m ≥ m ∗ m\geq m^* m ≥ m ∗ and, every n ≥ n ∗ n\geq n^* n ≥ n ∗ , consider τ n : = σ n ∖ σ n ∗ \tau_n:=\sigma_n\setminus\sigma_{n^*} τ n := σ n ∖ σ n ∗ . For each v ⊆ σ n v\subseteq\sigma_n v ⊆ σ n , let as before C m n , v : = ( C m ↑ τ n ) v C_m^{n,v}:=(C_m^{\uparrow\tau_n})^v C m n , v := ( C m ↑ τ n ) v . By Lemmas 22 and 23 , each cell C m n , v C_m^{n,v} C m n , v contains an inconsistent pair without a Craig σ n \sigma_n σ n -interpolant. Consequently,
∣ C r a i g C m n , v ∣ ≤ ∣ I n C o n C m n , v ∣ − 1. |\mathrm{Craig}_{C_m^{n,v}}|\leq|\mathrm{InCon}_{C_m^{n,v}}|-1. ∣ Craig C m n , v ∣ ≤ ∣ InCon C m n , v ∣ − 1. By Lemmas 9 and 11, ∣ C m n , v ∣ = ∣ C m ∣ |C_m^{n,v}|=|C_m| ∣ C m n , v ∣ = ∣ C m ∣ . And, since 1 ≤ ∣ I n C o n C m n , v ∣ ≤ 2 2 ∣ C m ∣ 1\leq|\mathrm{InCon}_{C_m^{n,v}}|\leq2^{2|C_m|} 1 ≤ ∣ InCon C m n , v ∣ ≤ 2 2∣ C m ∣ , we obtain
D C r a i g ‾ C m n , v ≥ 2 − 2 ∣ C m ∣ . \overline{\mathrm{DCraig}}_{C_m^{n,v}}\geq2^{-2|C_m|}. DCraig C m n , v ≥ 2 − 2∣ C m ∣ . By the cell reduction principle for Craig interpolation (Theorem 21 ),
D C r a i g n , m d = ∏ C ∈ C e l l σ n , ρ m d D C r a i g C ≤ ∏ v ⊆ σ n D C r a i g C m n , v ≤ ( 1 − 2 − 2 ∣ C m ∣ ) 2 n . \mathrm{DCraig}_{n,m}^{d}
=
\prod_{C\in\mathrm{Cell}_{\sigma_n,\rho_m}^{d}}\mathrm{DCraig}_{C}
\leq
\prod_{v\subseteq\sigma_n}\mathrm{DCraig}_{C_m^{n,v}}
\leq
\left(1-2^{-2|C_m|}\right)^{2^n}. DCraig n , m d = C ∈ Cell σ n , ρ m d ∏ DCraig C ≤ v ⊆ σ n ∏ DCraig C m n , v ≤ ( 1 − 2 − 2∣ C m ∣ ) 2 n . Since m m m is fixed, ∣ C m ∣ |C_m| ∣ C m ∣ does not depend on n n n , and the final expression tends to 0 0 0 as n → ∞ n\to\infty n → ∞ . □
Asymptotics for n n n fixed and m → ∞ m \to\infty m → ∞
We now consider the case in which the retained signature is fixed and the number of non-shared variables tends to infinity. Here we obtain results for locally tabular logics. As in Section 4.1, we aim to bound the density of interpolation failure within individual cells. For locally tabular logics, we drop the depth restriction and work with all formulas over a finite signature. We show that, if the incompatibility graphs of cells admit profiles whose sizes are bounded by a function of the retained signature size alone, then D C r a i g n , m → 1 \mathrm{DCraig}_{n,m} \to1 DCraig n , m → 1 for fixed n n n . We then apply this result to tabular quasi-normal modal logics.
Fix a locally tabular quasi-normal modal logic L L L and disjoint signatures σ , ρ \sigma, \rho σ , ρ . We write A t σ ∪ ρ \mathrm{At}_{\sigma\cup\rho} At σ ∪ ρ for the finite set of σ ∪ ρ \sigma\cup\rho σ ∪ ρ -atoms defined without a modal depth restriction. The characteristic formulas Φ ( M ) \Phi(M) Φ ( M ) of atoms M M M , as well as the cell characteristic formulas Γ ( C ) = Φ ( ⋃ C ) \Gamma(C) = \Phi(\bigcup C) Γ ( C ) = Φ ( ⋃ C ) , are likewise defined without a modal-depth restriction. The following observation will be useful throughout.
Lemma 25. Let C ∈ C e l l σ , ρ C \in\mathrm{Cell}_{\sigma,\rho} C ∈ Cell σ , ρ and let D , D 1 , D 2 ⊆ C D, D_1, D_2 \subseteq C D , D 1 , D 2 ⊆ C be non-empty. Then
Φ ( D ) \Phi(D) Φ ( D ) has a uniform σ \sigma σ -interpolant if and only if Γ ( C ) \Gamma(C) Γ ( C ) is a uniform σ \sigma σ -interpolant of Φ ( D ) \Phi(D) Φ ( D ) ;
the pair ( D 1 , D 2 ) (D_1, D_2) ( D 1 , D 2 ) does not admit a Craig σ \sigma σ -interpolant.
Proof. (1) By local tabularity, Proposition 6 applies without a depth bound. Thus, Γ ( C ) \Gamma(C) Γ ( C ) is the characteristic formula of a σ \sigma σ -atom. Since D D D is non-empty and all its atoms belong to the same cell, Φ ( D ) \Phi(D) Φ ( D ) and Γ ( C ) \Gamma(C) Γ ( C ) have exactly the same σ \sigma σ -consequences. In particular,
Φ ( D ) ⊨ θ iff Γ ( C ) ⊨ θ , for every θ ∈ M L σ . \Phi(D) \models\theta\quad\text{iff} \quad\Gamma(C) \models\theta,\qquad\text{for every }\theta\in\mathrm{ML}_{\sigma}. Φ ( D ) ⊨ θ iff Γ ( C ) ⊨ θ , for every θ ∈ ML σ .
If θ \theta θ is a uniform σ \sigma σ -interpolant of Φ ( D ) \Phi(D) Φ ( D ) , then θ ⊨ Γ ( C ) \theta\models\Gamma(C) θ ⊨ Γ ( C ) , since Γ ( C ) \Gamma(C) Γ ( C ) is a ρ \rho ρ -free consequence of Φ ( D ) \Phi(D) Φ ( D ) . The above equivalence gives Γ ( C ) ⊨ θ \Gamma(C) \models\theta Γ ( C ) ⊨ θ , so θ ≡ Γ ( C ) \theta\equiv\Gamma(C) θ ≡ Γ ( C ) . The converse is immediate.
(2) Observe that any Craig σ \sigma σ -interpolant θ \theta θ would satisfy Γ ( C ) ⊨ θ \Gamma(C) \models\theta Γ ( C ) ⊨ θ by the same equivalence. Since Φ ( D 2 ′ ) ⊨ Γ ( C ) \Phi(D_2') \models\Gamma(C) Φ ( D 2 ′ ) ⊨ Γ ( C ) and D 2 ≠ ∅ D_2 \ne\varnothing D 2 = ∅ , this contradicts θ ⊨ ¬ Φ ( D 2 ′ ) \theta\models\neg\Phi(D_2') θ ⊨ ¬Φ ( D 2 ′ ) .
We now count the pairs admitting a Craig interpolant within a single cell.
Theorem 26. Let L L L be a locally tabular quasi-normal logic. Then ∣ C r a i g C ∣ = 2 ∣ C ∣ + 1 − 1 |\mathrm{Craig}_C| = 2^{|C|+1}-1 ∣ Craig C ∣ = 2 ∣ C ∣ + 1 − 1 for every σ ρ \sigma^\rho σ ρ -cell C C C .
In particular,
∣ C r a i g σ , ρ ∣ = ∏ C ∈ C e l l σ , ρ ( 2 ∣ C ∣ + 1 − 1 ) . |\mathrm{Craig}_{\sigma,\rho}| = \prod_{C \in\mathrm{Cell}_{\sigma,\rho}} \left(2^{|C|+1}-1\right). ∣ Craig σ , ρ ∣ = C ∈ Cell σ , ρ ∏ ( 2 ∣ C ∣ + 1 − 1 ) .
Proof. The product formula follows from Theorem 21 . We compute ∣ C r a i g C ∣ |\mathrm{Craig}_C| ∣ Craig C ∣ for a fixed cell C C C . By Lemma 25 , no pair ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) with non-empty D 1 , D 2 ⊆ C D_1,D_2 \subseteq C D 1 , D 2 ⊆ C admits a Craig interpolant. On the other hand, if D 1 = ∅ D_1 = \varnothing D 1 = ∅ or D 2 = ∅ D_2 = \varnothing D 2 = ∅ , then ⊥ \bot ⊥ or ⊤ \top ⊤ is a Craig interpolant, respectively. Therefore, the pairs in C C C with Craig interpolants are exactly those with an empty component, giving ∣ C r a i g C ∣ = 2 ∣ C ∣ + 2 ∣ C ∣ − 1 = 2 ∣ C ∣ + 1 − 1 |\mathrm{Craig}_C| = 2^{|C|} + 2^{|C|} - 1 = 2^{|C|+1}-1 ∣ Craig C ∣ = 2 ∣ C ∣ + 2 ∣ C ∣ − 1 = 2 ∣ C ∣ + 1 − 1 .
To count the inconsistent pairs within a single cell, we consider the graph on C C C induced by the incompatibility relation between atoms in C C C .
Definition 10 (incompatibility graph). The incompatibility graph of a cell C C C is H C = ( C , ⌣ ) H_C = (C,\mathrel{\smile}) H C = ( C , ⌣ ) , where ⌣ \smile ⌣ is the incompatibility relation between atoms in C C C .
Example 7. In Example 6, the 3-fork atoms M φ \mathcal{M}_{\varphi} M φ and M ψ \mathcal{M}_{\psi} M ψ form an edge of the incompatibility graph of C { p , p , ¬ p } ρ C^\rho_{\{p,p,\neg p\}} C { p , p , ¬ p } ρ , for ∣ ρ ∣ ≥ 1 |\rho| \ge1 ∣ ρ ∣ ≥ 1 . Thus the singleton pair ( { M φ } , { M ψ } ) (\{\mathcal{M}_{\varphi}\}, \{\mathcal{M}_{\psi}\}) ({ M φ } , { M ψ }) is a non-empty ordered biclique in C { p , p , ¬ p } ρ C^\rho_{\{p,p,\neg p\}} C { p , p , ¬ p } ρ . It represents an inconsistent pair without a Craig interpolant.
Every atom is compatible with itself, so H C H_C H C has no loops. By Proposition 19 , an inconsistent pair ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) with D 1 , D 2 ⊆ C D_1,D_2 \subseteq C D 1 , D 2 ⊆ C is precisely an ordered biclique in H C H_C H C . By Theorem 26 , such a pair fails Craig interpolation exactly when both sides are non-empty. We count them as follows:
N o C r a i g C : = ∣ I n C o n C ∣ − ∣ C r a i g C ∣ = ∣ { ( D 1 , D 2 ) ∣ D 1 , D 2 non-empty ordered biclique in H C } ∣ . (2) \mathrm{NoCraig}_C := |\mathrm{InCon}_C| - |\mathrm{Craig}_C| = \left|\{(D_1,D_2) \mid D_1,D_2\text{ non-empty ordered biclique in }H_C\}\right|. \tag*{(2)} NoCraig C := ∣ InCon C ∣ − ∣ Craig C ∣ = ∣ {( D 1 , D 2 ) ∣ D 1 , D 2 non-empty ordered biclique in H C } ∣ . (2)
Then ∣ I n C o n C ∣ = ( 2 ∣ C ∣ + 1 − 1 ) + N o C r a i g C |\mathrm{InCon}_{C}| = (2^{|C|+1} - 1) + \mathrm{NoCraig}_{C} ∣ InCon C ∣ = ( 2 ∣ C ∣ + 1 − 1 ) + NoCraig C , and
D C r a i g C = 2 ∣ C ∣ + 1 − 1 2 ∣ C ∣ + 1 − 1 + N o C r a i g C , D C r a i g C ‾ = N o C r a i g C ∣ I n C o n C ∣ ≤ N o C r a i g C 2 ∣ C ∣ . (3) \mathrm{DCraig}_{C} = \frac{2^{|C|+1}-1}{2^{|C|+1}-1+\mathrm{NoCraig}_{C}}, \qquad\overline{\mathrm{DCraig}_{C}} = \frac{\mathrm{NoCraig}_{C}}{|\mathrm{InCon}_{C}|} \leq\frac{\mathrm{NoCraig}_{C}}{2^{|C|}}. \tag*{(3)} DCraig C = 2 ∣ C ∣ + 1 − 1 + NoCraig C 2 ∣ C ∣ + 1 − 1 , DCraig C = ∣ InCon C ∣ NoCraig C ≤ 2 ∣ C ∣ NoCraig C . (3)
To bound N o C r a i g C \mathrm{NoCraig}_{C} NoCraig C , we replace H C H_{C} H C by a smaller graph that preserves its incompatibility relation. A graph homomorphism is called strong if it both preserves and reflects edges.
Definition 11. Let C C C be a cell and let H P = ( P , ≺ P ) H_{P} = (P, \mathord{\prec_{P}}) H P = ( P , ≺ P ) be a graph. A profile map for C C C is a surjective strong graph homomorphism π : H C → H P \pi: H_{C} \to H_{P} π : H C → H P . We call H P H_{P} H P a profile graph of C C C , or simply a profile . For p ∈ P p \in P p ∈ P and Q ⊆ P Q \subseteq P Q ⊆ P , put
w ( p ) : = ∣ π − 1 ( p ) ∣ , w ( Q ) : = ∑ p ∈ Q w ( p ) . w(p) := |\pi^{-1}(p)|, \qquad w(Q) := \sum_{p \in Q} w(p). w ( p ) := ∣ π − 1 ( p ) ∣ , w ( Q ) := p ∈ Q ∑ w ( p ) .
As usual, in a graph H G = ( G , ≺ G ) H_{G} = (G, \mathord{\prec_{G}}) H G = ( G , ≺ G ) , the neighbourhood of g ∈ G g \in G g ∈ G is N G ( g ) : = { g ′ ∈ G ∣ g ≺ G g ′ } N_{G}(g) := \{g' \in G \mid g \mathord{\prec_{G}} g'\} N G ( g ) := { g ′ ∈ G ∣ g ≺ G g ′ } , and the common neighbourhood of a non-empty G ′ ⊆ G G' \subseteq G G ′ ⊆ G is N G ( G ′ ) : = ⋂ g ′ ∈ G ′ N G ( g ′ ) N_{G}(G') := \bigcap_{g' \in G'} N_{G}(g') N G ( G ′ ) := ⋂ g ′ ∈ G ′ N G ( g ′ ) .
The following lemma provides an upper bound for N o C r a i g C \mathrm{NoCraig}_{C} NoCraig C in terms of ordered bicliques in H P H_{P} H P and the weight w w w .
Lemma 27. Let C C C be a cell and H P = ( P , ≺ P ) H_{P} = (P, \mathord{\prec_{P}}) H P = ( P , ≺ P ) a profile graph of C C C . Then
N o C r a i g C ≤ ∑ ∅ ≠ Q ⊆ P N P ( Q ) ≠ ∅ 2 w ( Q ) ⋅ 2 w ( N P ( Q ) ) . \mathrm{NoCraig}_{C} \leq\sum_{\substack{\varnothing\neq Q \subseteq P \\ N_{P}(Q) \neq\varnothing}} 2^{w(Q)} \cdot2^{w(N_{P}(Q))}. NoCraig C ≤ ∅ = Q ⊆ P N P ( Q ) = ∅ ∑ 2 w ( Q ) ⋅ 2 w ( N P ( Q )) .
Proof. Let π : H C → H P \pi: H_{C} \to H_{P} π : H C → H P be a profile map for C C C . Observe that ( D 1 , D 2 ) (D_{1}, D_{2}) ( D 1 , D 2 ) is an ordered biclique in H C H_{C} H C with D 1 , D 2 D_{1}, D_{2} D 1 , D 2 non-empty iff there exists a non-empty Q ⊆ P Q \subseteq P Q ⊆ P such that D 1 ⊆ π − 1 ( Q ) D_{1} \subseteq\pi^{-1}(Q) D 1 ⊆ π − 1 ( Q ) , N P ( Q ) ≠ ∅ N_{P}(Q) \neq\varnothing N P ( Q ) = ∅ , and D 2 ⊆ π − 1 ( N P ( Q ) ) D_{2} \subseteq\pi^{-1}(N_{P}(Q)) D 2 ⊆ π − 1 ( N P ( Q )) . The inequality follows from the definition of the weights w w w .
Among all profiles of a cell, there is a coarsest one, obtained by identifying atoms with the same incompatibility neighbourhood.
Definition 12. Let, for any cell C C C and M 1 , M 2 ∈ C \mathcal{M}_{1},\mathcal{M}_{2} \in C M 1 , M 2 ∈ C ,
M 1 ∼ C M 2 ⟺ N C ( M 1 ) = N C ( M 2 ) . \mathcal{M}_{1} \sim_{C} \mathcal{M}_{2} \Longleftrightarrow N_{C}(\mathcal{M}_{1}) = N_{C}(\mathcal{M}_{2}). M 1 ∼ C M 2 ⟺ N C ( M 1 ) = N C ( M 2 ) . Call the quotient graph H C / ∼ C H_{C}/\sim_{C} H C / ∼ C defined by ∼ C \sim_{C} ∼ C with domain C / ∼ C C/\sim_{C} C / ∼ C the canonical profile graph of C C C and π C : C → C / ∼ C \pi_{C} : C \to C/\sim_{C} π C : C → C / ∼ C the canonical profile map . Its number of classes is inc ( C ) : = ∣ C / ∼ C ∣ \operatorname{inc}(C) := |C/\sim_{C}| inc ( C ) := ∣ C / ∼ C ∣ .
The canonical map π C \pi_{C} π C is a profile map. Moreover, every profile map π : H C → H P \pi: H_{C} \to H_{P} π : H C → H P refines the canonical partition, so
inc ( C ) ≤ ∣ P ∣ . \operatorname{inc}(C) \leq|P|. inc ( C ) ≤ ∣ P ∣.
The logic L L L has CIP if and only if H C H_{C} H C has no edges for every cell C C C . Equivalently, inc ( C ) = 1 \operatorname{inc}(C) = 1 inc ( C ) = 1 for every cell.
Example 8. The 3-fork gives a simple example of a canonical profile. For C = C { p } , { { p } , { ¬ p } } ρ C = C^{\rho}_{\{p\},\{\{p\},\{\neg p\}\}} C = C { p } , {{ p } , { ¬ p }} ρ with ρ = { q } \rho= \{q\} ρ = { q } , the atoms fall into three canonical classes: those with two distinct successor colours, those with three distinct successor colours of which two refine p p p , and those with three distinct successor colours of which two refine ¬ p \neg p ¬ p . The only edge of the canonical profile joins the last two classes. In particular, inc ( C ) = 3 \operatorname{inc}(C) = 3 inc ( C ) = 3 , although C C C contains more than three atoms.
We now introduce a hierarchy of bounds on the size of canonical profiles.
Definition 13. A locally tabular quasi-normal modal logic L L L has the uniformly bounded intra-cell incompatibility property (uniform BICIP ) if there is a constant F L F_{L} F L such that inc ( C ) ≤ F L \operatorname{inc}(C) \leq F_{L} inc ( C ) ≤ F L for all σ ↑ \sigma^{\uparrow} σ ↑ -cells C C C . L L L has BICIP if there is a function f L f_{L} f L depending on n = ∣ σ ∣ n = |\sigma| n = ∣ σ ∣ only such that inc ( C ) ≤ f L ( n ) \operatorname{inc}(C) \leq f_{L}(n) inc ( C ) ≤ f L ( n ) for every σ ↑ \sigma^{\uparrow} σ ↑ -cell C C C , independently from ρ \rho ρ .
We next show that the pinned atoms of Lemma 13 are isolated in the incompatibility graph. For a σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell C C C , put M : = ⋃ C ∈ At σ \mathcal{M}:=\bigcup C\in\operatorname{At}_{\sigma} M := ⋃ C ∈ At σ ; see Proposition 6. As in Section 4, for each ρ \rho ρ -colour r r r , define M r : = [ M r , w ] σ ∪ ρ \mathcal{M}^{r}:=[\mathcal{M}^{r},w]_{\sigma\cup\rho} M r := [ M r , w ] σ ∪ ρ for ( M , w ) ∈ M (\mathcal{M},w)\in\mathcal{M} ( M , w ) ∈ M , and let M C : = { M r ∣ r a ρ -colour } M_C:=\{\mathcal{M}^{r}\mid r\text{ a }\rho\text{-colour}\} M C := { M r ∣ r a ρ -colour } . By Lemma 8(2), applied at every modal depth, this definition does not depend on the chosen representative ( M , w ) (\mathcal{M},w) ( M , w ) . Pinning leaves all σ \sigma σ -formulas unchanged, so M C ⊆ C M_C\subseteq C M C ⊆ C . Distinct ρ \rho ρ -colours differ at the distinguished world, and hence ∣ M C ∣ = 2 ∣ ρ ∣ |M_C|=2^{|\rho|} ∣ M C ∣ = 2 ∣ ρ ∣ .
Lemma 28. Every atom in M C M_C M C is compatible with every atom in C C C .
Proof. Fix a ρ \rho ρ -colour r r r and N ∈ C N\in C N ∈ C . Choose ( M , w ) ∈ M (\mathcal{M},w)\in\mathcal{M} ( M , w ) ∈ M and ( N , u ) ∈ N (\mathcal{N},u)\in N ( N , u ) ∈ N . Since ( M , w ) ≡ σ ( N , u ) (\mathcal{M},w)\equiv_{\sigma}(\mathcal{N},u) ( M , w ) ≡ σ ( N , u ) , Lemma 8(2), applied at every depth, gives [ N r , u ] σ ∪ ρ = [ M r , w ] σ ∪ ρ = M r [\mathcal{N}^{r},u]_{\sigma\cup\rho}=[\mathcal{M}^{r},w]_{\sigma\cup\rho}=\mathcal{M}^{r} [ N r , u ] σ ∪ ρ = [ M r , w ] σ ∪ ρ = M r . Let P \mathcal{P} P be based on the same frame as N \mathcal{N} N , interpret the variables in ρ \rho ρ constantly according to r r r , and put V P ( q ′ ) = V N ( q ) V_{\mathcal{P}}(q')=V_{\mathcal{N}}(q) V P ( q ′ ) = V N ( q ) for each corresponding q ∈ ρ q\in\rho q ∈ ρ and q ′ ∈ ρ ′ q'\in\rho' q ′ ∈ ρ ′ . All other valuations agree with N \mathcal{N} N . Then [ P , u ] σ ∪ ρ = M r [\mathcal{P},u]_{\sigma\cup\rho}=\mathcal{M}^{r} [ P , u ] σ ∪ ρ = M r and [ P , u ] σ ∪ ρ ′ = B ′ [\mathcal{P},u]_{\sigma\cup\rho'}=B' [ P , u ] σ ∪ ρ ′ = B ′ . Thus M r ∩ N ′ ≠ ∅ \mathcal{M}^{r}\cap N'\ne\varnothing M r ∩ N ′ = ∅ , which proves compatibility.
Consequently, no atom of M C M_C M C occurs in a non-empty ordered biclique of H C H_C H C .
Theorem 29. Let L L L be a locally tabular quasi-normal modal logic with BICIP, and let f L f_L f L be a corresponding bound. Then, for every σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell C C C ,
DCraig ‾ C ≤ 2 f L ( ∣ σ ∣ ) − 2 ∣ ρ ∣ . \overline{\operatorname{DCraig}}_C \le2^{f_L(|\sigma|)-2^{|\rho|}}. DCraig C ≤ 2 f L ( ∣ σ ∣ ) − 2 ∣ ρ ∣ . Consequently, for every fixed n ≥ 0 n\ge0 n ≥ 0 ,
lim m → ∞ DCraig n , m = 1. \lim_{m\to\infty}\operatorname{DCraig}_{n,m}=1. m → ∞ lim DCraig n , m = 1. Proof. Fix a σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell C C C and let π : H C → H P \pi:H_C\to H_P π : H C → H P be its canonical profile map. By assumption, ∣ P ∣ ≤ f L ( ∣ σ ∣ ) |P|\le f_L(|\sigma|) ∣ P ∣ ≤ f L ( ∣ σ ∣ ) .
By Lemma 28, every non-empty ordered biclique ( D 1 , D 2 ) (D_1,D_2) ( D 1 , D 2 ) is disjoint with M C M_C M C . Since its two sides are disjoint,
∣ D 1 ∣ + ∣ D 2 ∣ ≤ ∣ C ∣ − ∣ M C ∣ ≤ ∣ C ∣ − 2 ∣ ρ ∣ . |D_1|+|D_2|\le|C|-|M_C|\le|C|-2^{|\rho|}. ∣ D 1 ∣ + ∣ D 2 ∣ ≤ ∣ C ∣ − ∣ M C ∣ ≤ ∣ C ∣ − 2 ∣ ρ ∣ . We compute an upper bound for NoCraig C \operatorname{NoCraig}_C NoCraig C using Lemma 27:
NoCraig C ≤ ∑ ∅ ≠ Q ⊆ P N P ( Q ) ≠ ∅ 2 w ( Q ) + w ( N P ( Q ) ) ≤ 2 ∣ P ∣ ⋅ 2 ∣ C ∣ − 2 ∣ ρ ∣ ≤ 2 ∣ C ∣ − ( 2 ∣ ρ ∣ − f L ( ∣ σ ∣ ) ) . \operatorname{NoCraig}_C
\le
\sum_{\substack{\varnothing\ne Q\subseteq P\\N_P(Q)\ne\varnothing}}
2^{w(Q)+w(N_P(Q))}
\le2^{|P|}\cdot2^{|C|-2^{|\rho|}}
\le2^{|C|-(2^{|\rho|}-f_L(|\sigma|))}. NoCraig C ≤ ∅ = Q ⊆ P N P ( Q ) = ∅ ∑ 2 w ( Q ) + w ( N P ( Q )) ≤ 2 ∣ P ∣ ⋅ 2 ∣ C ∣ − 2 ∣ ρ ∣ ≤ 2 ∣ C ∣ − ( 2 ∣ ρ ∣ − f L ( ∣ σ ∣ )) . where the penultimate inequality uses that there are at most 2 ∣ P ∣ 2^{|P|} 2 ∣ P ∣ subsets Q Q Q of P P P and each summand is at most 2 ∣ C ∣ − 2 ∣ ρ ∣ 2^{|C|-2^{|\rho|}} 2 ∣ C ∣ − 2 ∣ ρ ∣ . Dividing by ∣ InCon C ∣ ≥ 2 ∣ C ∣ |\operatorname{InCon}_C|\ge2^{|C|} ∣ InCon C ∣ ≥ 2 ∣ C ∣ via (2),
DCraig ‾ C = NoCraig C ∣ InCon C ∣ ≤ NoCraig C 2 ∣ C ∣ ≤ 2 f L ( ∣ σ ∣ ) − 2 ∣ ρ ∣ , \overline{\operatorname{DCraig}}_C
=
\frac{\operatorname{NoCraig}_C}{|\operatorname{InCon}_C|}
\le
\frac{\operatorname{NoCraig}_C}{2^{|C|}}
\le
2^{f_L(|\sigma|)-2^{|\rho|}}, DCraig C = ∣ InCon C ∣ NoCraig C ≤ 2 ∣ C ∣ NoCraig C ≤ 2 f L ( ∣ σ ∣ ) − 2 ∣ ρ ∣ , as required. By the cell reduction principle for Craig interpolation (Theorem 21), DCraig n , m = ∏ C DCraig C \operatorname{DCraig}_{n,m}=\prod_C\operatorname{DCraig}_C DCraig n , m = ∏ C DCraig C . Since 1 − ∏ C x C ≤ ∑ C ( 1 − x C ) 1-\prod_C x_C\le\sum_C(1-x_C) 1 − ∏ C x C ≤ ∑ C ( 1 − x C ) for x C ∈ [ 0 , 1 ] x_C\in[0,1] x C ∈ [ 0 , 1 ] , we obtain
DCraig ‾ n , m ≤ ∑ C ∈ Cell σ n , ρ m DCraig ‾ C ≤ ∣ At σ n ∣ 2 f L ( n ) − 2 m . \overline{\operatorname{DCraig}}_{n,m}
\le
\sum_{C\in\operatorname{Cell}_{\sigma^n,\rho^m}}
\overline{\operatorname{DCraig}}_C
\le
|\operatorname{At}_{\sigma^n}|\,2^{f_L(n)-2^m}. DCraig n , m ≤ C ∈ Cell σ n , ρ m ∑ DCraig C ≤ ∣ At σ n ∣ 2 f L ( n ) − 2 m . For fixed n n n , the right-hand side tends to 0 0 0 as m → ∞ m\to\infty m → ∞ .
The canonical profile can be hard to compute or work with. We now introduce another profile with nodes that are directly obtained from atoms. Recall that for a locally tabular quasi-normal modal logic we can identify atoms M ∈ At σ ∪ ρ M\in\operatorname{At}_{\sigma\cup\rho} M ∈ At σ ∪ ρ with a pointed extensional σ ∪ ρ \sigma\cup\rho σ ∪ ρ -model S M \mathcal{S}_M S M (Lemma 3).
Definition 14. The σ \sigma σ -shadow of an atom M ∈ At σ ∪ ρ M\in\operatorname{At}_{\sigma\cup\rho} M ∈ At σ ∪ ρ is defined as sh ( M ) : = ( S M ∣ σ , w M ) \operatorname{sh}(M):=(\mathcal{S}_M|_{\sigma},w_M) sh ( M ) := ( S M ∣ σ , w M ) , the pointed σ \sigma σ -model with the same underlying pointed frame as S M \mathcal{S}_M S M and with the valuation restricted to σ \sigma σ .
Note that sh ( M ) \operatorname{sh}(M) sh ( M ) need not be extensional: forgetting ρ \rho ρ may make distinct worlds indistinguishable in M L σ \mathrm{ML}_{\sigma} ML σ . We retain the underlying frame of the standard ( σ ∪ ρ ) (\sigma\cup\rho) ( σ ∪ ρ ) -model and do not take a further quotient.
Definition 15. An L L L -cover of two pointed σ \sigma σ -models ( M 1 , w 1 ) (\mathcal{M}_1,w_1) ( M 1 , w 1 ) and ( M 2 , w 2 ) (\mathcal{M}_2,w_2) ( M 2 , w 2 ) is a pointed σ \sigma σ -model ( M , w ) (\mathcal{M},w) ( M , w ) based on a frame for L L L , together with σ \sigma σ -p p p -morphisms
f i : ( M , w ) → ( M i , w i ) , i = 1 , 2. f_i:(\mathcal{M},w)\to(\mathcal{M}_i,w_i),\qquad i=1,2. f i : ( M , w ) → ( M i , w i ) , i = 1 , 2.
Theorem 30. Let L L L be a locally tabular quasi-normal modal logic. Let C C C be a σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell. Then M 1 , M 2 ∈ C \mathcal{M}_1,\mathcal{M}_2 \in C M 1 , M 2 ∈ C are compatible in H C H_C H C if and only if ( sh ( M 1 ) , sh ( M 2 ) ) (\operatorname{sh}(\mathcal{M}_1),\operatorname{sh}(\mathcal{M}_2)) ( sh ( M 1 ) , sh ( M 2 )) has an L L L -cover. The cover may be chosen finite.
Proof. ( ⇒ ) (\Rightarrow) ( ⇒ ) Suppose that M 1 \mathcal{M}_1 M 1 and M 2 \mathcal{M}_2 M 2 are compatible. Then M 1 ∩ M 2 ′ ≠ ∅ \mathcal{M}_1 \cap\mathcal{M}_2' \ne\varnothing M 1 ∩ M 2 ′ = ∅ . Choose a pointed model in this intersection, and let ( M , w ) (\mathfrak{M},w) ( M , w ) be the finite standard model of its ( σ ∪ ρ ∪ ρ ′ ) (\sigma\cup\rho\cup\rho') ( σ ∪ ρ ∪ ρ ′ ) -atom. By Lemma 3, there are pointed p-morphisms f 1 : ( M , w ) → ( S M 1 , w M 1 ) f_1:(\mathfrak{M},w)\to(\mathcal{S}_{\mathcal{M}_1},w_{\mathcal{M}_1}) f 1 : ( M , w ) → ( S M 1 , w M 1 ) over σ ∪ ρ \sigma\cup\rho σ ∪ ρ , and f 2 : ( M , w ) → ( S M 2 ′ , w M 2 ′ ) f_2:(\mathfrak{M},w)\to(\mathcal{S}_{\mathcal{M}_2'},w_{\mathcal{M}_2'}) f 2 : ( M , w ) → ( S M 2 ′ , w M 2 ′ ) over σ ∪ ρ ′ \sigma\cup\rho' σ ∪ ρ ′ . Restricting the valuations to σ \sigma σ gives σ \sigma σ -p-morphisms f i : ( M ↾ σ , w ) → sh ( M i ) f_i:(\mathfrak{M}\mathord{\upharpoonright}\sigma,w)\to\operatorname{sh}(\mathcal{M}_i) f i : ( M ↾ σ , w ) → sh ( M i ) , for i = 1 , 2 i=1,2 i = 1 , 2 , where the second target is identified with sh ( M 2 ) \operatorname{sh}(\mathcal{M}_2) sh ( M 2 ) via the renaming of ρ \rho ρ to ρ ′ \rho' ρ ′ . Thus ( M ↾ σ , w ) (\mathfrak{M}\mathord{\upharpoonright}\sigma,w) ( M ↾ σ , w ) is a finite L L L -cover of the two shadows.
( ⇐ ) (\Leftarrow) ( ⇐ ) Suppose that the shadows have an L L L -cover ( M , w ) (\mathfrak{M},w) ( M , w ) , where M = ( W , R , V ) \mathfrak{M}=(W,R,V) M = ( W , R , V ) , witnessed by σ \sigma σ -p-morphisms f i : ( M , w ) → sh ( M i ) f_i:(\mathfrak{M},w)\to\operatorname{sh}(\mathcal{M}_i) f i : ( M , w ) → sh ( M i ) , i = 1 , 2 i=1,2 i = 1 , 2 . Let V i V_i V i be the valuation of S M i \mathcal{S}_{\mathcal{M}_i} S M i . Define M + = ( W , R , V + ) \mathfrak{M}^{+}=(W,R,V^{+}) M + = ( W , R , V + ) by
V + ( p ) = { f 1 − 1 ( V 1 ( p ) ) , p ∈ ρ , f 2 − 1 ( V 2 ( q ) ) , p = q ′ ∈ ρ ′ , q ∈ ρ , V ( p ) , p ∉ ρ ∪ ρ ′ . V^{+}(p)=
\begin{cases}
f_1^{-1}(V_1(p)), & p\in\rho,\\
f_2^{-1}(V_2(q)), & p=q'\in\rho',\ q\in\rho,\\
V(p), & p\notin\rho\cup\rho'.
\end{cases} V + ( p ) = ⎩ ⎨ ⎧ f 1 − 1 ( V 1 ( p )) , f 2 − 1 ( V 2 ( q )) , V ( p ) , p ∈ ρ , p = q ′ ∈ ρ ′ , q ∈ ρ , p ∈ / ρ ∪ ρ ′ . Since ρ \rho ρ and ρ ′ \rho' ρ ′ are disjoint and neither meets σ \sigma σ , this defines a valuation extending V V V on σ \sigma σ . By construction, f 1 f_1 f 1 preserves the valuations of all variables in σ ∪ ρ \sigma\cup\rho σ ∪ ρ , and f 2 f_2 f 2 preserves those in σ ∪ ρ ′ \sigma\cup\rho' σ ∪ ρ ′ after renaming ρ \rho ρ to ρ ′ \rho' ρ ′ in its target. Their underlying maps already satisfy the forth and back conditions. Hence they induce pointed p-morphisms f 1 : ( M + ↾ σ ∪ ρ , w ) → ( S M 1 , w M 1 ) f_1:(\mathfrak{M}^{+}\mathord{\upharpoonright}_{\sigma\cup\rho},w)\to(\mathcal{S}_{\mathcal{M}_1},w_{\mathcal{M}_1}) f 1 : ( M + ↾ σ ∪ ρ , w ) → ( S M 1 , w M 1 ) and f 2 : ( M + ↾ σ ∪ ρ ′ , w ) → ( S M 2 ′ , w M 2 ′ ) f_2:(\mathfrak{M}^{+}\mathord{\upharpoonright}_{\sigma\cup\rho'},w)\to(\mathcal{S}_{\mathcal{M}_2'},w_{\mathcal{M}_2'}) f 2 : ( M + ↾ σ ∪ ρ ′ , w ) → ( S M 2 ′ , w M 2 ′ ) . By Lemma 3, it follows that ( M + , w ) ∈ M 1 ∩ M 2 ′ (\mathfrak{M}^{+},w)\in\mathcal{M}_1\cap\mathcal{M}_2' ( M + , w ) ∈ M 1 ∩ M 2 ′ . Thus, M 1 \mathcal{M}_1 M 1 and M 2 \mathcal{M}_2 M 2 are compatible.
Finally, the cover in the first implication is finite, and the converse construction does not require finiteness. This proves the last assertion.
Example 9. Consider again the 3-fork atoms defined by φ \varphi φ and ψ \psi ψ in Example 2. Their { p } \{p\} { p } -shadows have three successors, with two satisfying p p p in the first shadow and two satisfying ¬ p \neg p ¬ p in the second. A common p-morphic predecessor based on a 3-fork would need at least two successors of each colour, which is impossible. Thus, the shadows have no cover, as follows from Theorem 30 .
The notion of an L L L -cover is closely related to Robinson’s joint consistency property [4 Theorem 2.2.23] . Atoms in the same σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell have the same complete σ \sigma σ -theory, and their compatibility means that, after renaming the forgotten variables on one side, the two theories have a common realisation. Theorem 30 characterises this joint realisability in terms of a common p-morphic predecessor of their σ \sigma σ -shadows. Thus covers provide a frame-theoretic version of a local Robinson consistency condition.
Define the shadow incompatibility graph of C C C by
sh ( H C ) : = ( sh ( C ) , ̸ ∼ sh ) . \operatorname{sh}(H_C):=(\operatorname{sh}(C),\mathrel{\not\sim^{\operatorname{sh}}}). sh ( H C ) := ( sh ( C ) , ∼ sh ) .
where sh ( C ) = { sh ( M ) ∣ M ∈ C } \operatorname{sh}(C)=\{\operatorname{sh}(\mathcal{M})\mid\mathcal{M}\in C\} sh ( C ) = { sh ( M ) ∣ M ∈ C } and two shadows are related by ̸ ∼ sh \not\sim^{\operatorname{sh}} ∼ sh if they have no L L L -cover. The shadow profile map is
π sh : M ⟼ sh ( M ) . \pi_{\operatorname{sh}}:\mathcal{M}\longmapsto\operatorname{sh}(\mathcal{M}). π sh : M ⟼ sh ( M ) .
Proposition 31. Let C C C be a σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell. Then π sh \pi_{\operatorname{sh}} π sh is a profile mapping from H C H_C H C onto sh ( H C ) \operatorname{sh}(H_C) sh ( H C ) .
As the shadow incompatibility graph lies between the cell and the canonical profile, we have a strong graph homomorphism κ : sh ( H C ) → H C / ∼ C \kappa:\operatorname{sh}(H_C)\to H_C/{\sim_C} κ : sh ( H C ) → H C / ∼ C with π C = κ ∘ π sh \pi_C=\kappa\circ\pi_{\operatorname{sh}} π C = κ ∘ π sh .
We apply the shadow profile to show that all tabular modal logics have the uniform BICIP. If L L L is tabular, then L = Log ( K ) L=\operatorname{Log}(\mathcal{K}) L = Log ( K ) for the (up to isomorphism) finite class K \mathcal{K} K of finite rooted Kripke frames validating L L L . Assume
K = { ( F 1 , r 1 ) , … , ( F k , r k ) } , F i = ( W i , R i ) , r i ∈ W i . \mathcal{K}=\{(\mathfrak{F}_1,r_1),\ldots,(\mathfrak{F}_k,r_k)\},\qquad\mathfrak{F}_i=(W_i,R_i),\ r_i\in W_i. K = {( F 1 , r 1 ) , … , ( F k , r k )} , F i = ( W i , R i ) , r i ∈ W i .
Theorem 32. Tabular quasi-normal modal logics have the uniform BICIP: for every σ ↑ ρ \sigma^{\uparrow\rho} σ ↑ ρ -cell C C C ,
inc ( C ) ≤ F L : = ∑ i = 1 k N L N i N i = ∣ W i ∣ , N L = max i N i . \operatorname{inc}(C)\le F_L:=\sum_{i=1}^{k}N_L^{N_i}\qquad N_i=|W_i|,\quad N_L=\max_i N_i. inc ( C ) ≤ F L := i = 1 ∑ k N L N i N i = ∣ W i ∣ , N L = i max N i .
Proof. Assume C C C is a σ τ P \sigma\tau_P σ τ P -cell. Because the canonical profile is the coarsest profile, it is sufficient to show that ∣ sh ( C ) ∣ ≤ F L |\operatorname{sh}(C)| \le F_L ∣ sh ( C ) ∣ ≤ F L . Recall that the shadow sh ( M ) \operatorname{sh}(\mathcal{M}) sh ( M ) of M ∈ C \mathcal{M} \in C M ∈ C is obtained by taking the restriction ( S M ↾ σ , w M ) (\mathfrak{S}_{\mathcal{M}}\mathord{\upharpoonright}\sigma,w_{\mathcal{M}}) ( S M ↾ σ , w M ) to σ \sigma σ of the standard model ( S M , w M ) (\mathfrak{S}_{\mathcal{M}},w_{\mathcal{M}}) ( S M , w M ) . Consider the σ \sigma σ -atom A = ⋃ C A = \bigcup C A = ⋃ C and note that its σ \sigma σ -shadow coincides with its standard model. By Lemma 3, we have a σ \sigma σ -p-morphism from sh ( M ) \operatorname{sh}(\mathcal{M}) sh ( M ) onto sh ( A ) \operatorname{sh}(A) sh ( A ) . Observe that once the underlying mapping of a p-morphism is fixed, the valuation of sh ( M ) \operatorname{sh}(\mathcal{M}) sh ( M ) is uniquely determined by the valuation of sh ( A ) \operatorname{sh}(A) sh ( A ) . Because any sh ( M ) \operatorname{sh}(\mathcal{M}) sh ( M ) is based on some ( F i , r i ) ∈ K (\mathfrak{F}_i,r_i) \in\mathcal{K} ( F i , r i ) ∈ K and sh ( A ) \operatorname{sh}(A) sh ( A ) is based on a frame in K \mathcal{K} K , the number of different valuations sh ( M ) \operatorname{sh}(\mathcal{M}) sh ( M ) could have is bounded by the number N L N i N^{N_i}_{L} N L N i of mappings from the underpinning frame ( F i , r i ) (\mathfrak{F}_i,r_i) ( F i , r i ) of sh ( M ) \operatorname{sh}(\mathcal{M}) sh ( M ) onto the frame underpinning sh ( A ) \operatorname{sh}(A) sh ( A ) . The inequality follows.
Theorems 29 and 32 imply the following result for tabular logics.
Theorem 33. For every tabular quasi-normal modal logic L L L and every fixed n ≥ 0 n \ge0 n ≥ 0 ,
lim m → ∞ D C r a i g n m = 1. \lim_{m\to\infty}\mathrm{DCraign}_{m}=1. m → ∞ lim DCraign m = 1.
Density after Existential Quantification
Interpolants are independent of the names of forgotten or non-shared variables and are invariant under natural transformations, such as uniformly replacing any such variable by its negation. In this section, we abstract away from variable names and some of these transformations by identifying input formulas whose existential quantifications over all forgotten or non-shared variables are logically equivalent over a given class of frames.
Throughout this section, we fix a class K \mathcal{K} K of pointed frames and assume that L = Log ( K ) L=\operatorname{Log}(\mathcal{K}) L = Log ( K ) . Let φ ∈ M L τ \varphi\in\mathrm{ML}_{\tau} φ ∈ ML τ and ρ ⊆ τ \rho\subseteq\tau ρ ⊆ τ . The formula ∃ ρ ⃗ . φ \exists\vec{\rho}.\varphi ∃ ρ . φ is interpreted in a model M = ( F , V ) \mathcal{M}=(\mathfrak{F},V) M = ( F , V ) by setting M , w ⊨ ∃ ρ ⃗ . φ \mathcal{M},w \models\exists\vec{\rho}.\varphi M , w ⊨ ∃ ρ . φ if and only if there exists a valuation V ′ V' V ′ coinciding with V V V on all variables outside ρ \rho ρ such that M ′ , w ⊨ φ \mathcal{M}',w \models\varphi M ′ , w ⊨ φ , where M ′ = ( F , V ′ ) \mathcal{M}'=(\mathfrak{F},V') M ′ = ( F , V ′ ) .
We write ∃ ρ ⃗ . φ ≡ ∃ ρ ⃗ . ψ \exists\vec{\rho}.\varphi\equiv\exists\vec{\rho}.\psi ∃ ρ . φ ≡ ∃ ρ . ψ if the two formulas are true at exactly the same pointed models based on members of K \mathcal{K} K , and denote their equivalence class by [ ∃ ρ ⃗ . φ ] [\exists\vec{\rho}.\varphi] [ ∃ ρ . φ ] . Unlike ordinary modal equivalence, equivalence after existential quantification may depend on the chosen frame class since different classes of frames determining the same logic can induce different equivalence relations on existentially quantified formulas (see Example 12 below). Thus, the counting measure considered in this section depends not only on L L L but also on K \mathcal{K} K .
Fix disjoint signatures σ \sigma σ and ρ \rho ρ . The uniform σ \sigma σ -interpolants of a formula φ ∈ M L σ ∪ ρ \varphi\in\mathrm{ML}_{\sigma\cup\rho} φ ∈ ML σ ∪ ρ depend only on the equivalence class [ ∃ ρ ⃗ . φ ] [\exists\vec{\rho}.\varphi] [ ∃ ρ . φ ] . More precisely, if χ \chi χ is a uniform σ \sigma σ -interpolant of φ \varphi φ and ∃ ρ ⃗ . ψ ≡ ∃ ρ ⃗ . φ \exists\vec{\rho}.\psi\equiv\exists\vec{\rho}.\varphi ∃ ρ . ψ ≡ ∃ ρ . φ , then χ \chi χ is also a uniform σ \sigma σ -interpolant of ψ \psi ψ . Similarly, the consistency and Craig interpolants of pairs φ ∈ M L σ ∪ ρ \varphi\in\mathrm{ML}_{\sigma\cup\rho} φ ∈ ML σ ∪ ρ and ψ ∈ M L σ ∪ ρ ′ \psi\in\mathrm{ML}_{\sigma\cup\rho'} ψ ∈ ML σ ∪ ρ ′ depend only on the equivalence classes [ ∃ ρ ⃗ . φ ] [\exists\vec{\rho}.\varphi] [ ∃ ρ . φ ] and [ ∃ ρ ′ ⃗ . ψ ] [\exists\vec{\rho'}.\psi] [ ∃ ρ ′ . ψ ] , respectively. This allows us to define an alternative notion of interpolant density, obtained by counting existentially quantified formulas of the form ∃ ρ ⃗ . φ \exists\vec{\rho}.\varphi ∃ ρ . φ up to logical equivalence. Fix a modal depth d d d . Let
Q σ , ρ d = { [ ∃ ρ ⃗ . φ ] ∣ φ ∈ M L σ ∪ ρ d } , U σ , ρ Π , d = { [ ∃ ρ ⃗ . φ ] ∈ Q σ , ρ d ∣ φ has a uniform σ -interpolant } . Q^{d}_{\sigma,\rho}=\{[\exists\vec{\rho}.\varphi]\mid\varphi\in\mathrm{ML}^{d}_{\sigma\cup\rho}\},\qquad
U^{\Pi,d}_{\sigma,\rho}=\{[\exists\vec{\rho}.\varphi]\in Q^{d}_{\sigma,\rho}\mid\varphi\text{ has a uniform }\sigma\text{-interpolant}\}. Q σ , ρ d = {[ ∃ ρ . φ ] ∣ φ ∈ ML σ ∪ ρ d } , U σ , ρ Π , d = {[ ∃ ρ . φ ] ∈ Q σ , ρ d ∣ φ has a uniform σ -interpolant } .
and define the density of uniform interpolation after existential quantification as
UnifInt σ , ρ Π , d : = ∣ U σ , ρ Π , d ∣ ∣ Q σ , ρ d ∣ . \operatorname{UnifInt}^{\Pi,d}_{\sigma,\rho}:=\frac{|U^{\Pi,d}_{\sigma,\rho}|}{|Q^{d}_{\sigma,\rho}|}. UnifInt σ , ρ Π , d := ∣ Q σ , ρ d ∣ ∣ U σ , ρ Π , d ∣ .
Existentially quantified formulas are neither closed under negation nor conjunction. Normal forms are therefore less transparent than in ML, and counting becomes correspondingly more involved. In particular, we can no longer identify equivalence classes of existentially quantified formulas with all subsets of a set of atoms, as in the unquantified case. Nevertheless, we use a representation based on the atoms introduced earlier.
Every φ ∈ M L σ ∪ ρ d \varphi\in\mathrm{ML}^{d}_{\sigma\cup\rho} φ ∈ ML σ ∪ ρ d is equivalent to a unique Φ ( D ) \Phi(D) Φ ( D ) with D ⊆ A t σ ∪ ρ d D\subseteq\mathrm{At}^{d}_{\sigma\cup\rho} D ⊆ At σ ∪ ρ d . By distribution of existential quantification over disjunction, ∃ ρ ⃗ . φ \exists\vec{\rho}.\varphi ∃ ρ . φ is equivalent to a formula of the form
⋁ M ∈ D ∃ ρ ⃗ . Φ ( M ) ≡ ∃ ρ ⃗ . Φ ( D ) \bigvee_{M\in D}\exists\vec{\rho}.\Phi(M)\ \equiv\ \exists\vec{\rho}.\Phi(D) M ∈ D ⋁ ∃ ρ .Φ ( M ) ≡ ∃ ρ .Φ ( D )
with D ⊆ A t σ ∪ ρ d D\subseteq\mathrm{At}^{d}_{\sigma\cup\rho} D ⊆ At σ ∪ ρ d . We thus work with the equivalence classes [ ∃ ρ ⃗ . Φ ( D ) ] [\exists\vec{\rho}.\Phi(D)] [ ∃ ρ .Φ ( D )] for D ⊆ A t σ ∪ ρ d D\subseteq\mathrm{At}^{d}_{\sigma\cup\rho} D ⊆ At σ ∪ ρ d . These classes form a join-semilattice with bottom element and join defined by
⊥ : = [ ∃ ρ ⃗ . Φ ( ∅ ) ] , \bot\ :=\ [\exists\vec{\rho}.\Phi(\varnothing)], ⊥ := [ ∃ ρ .Φ ( ∅ )] ,
[ ∃ ρ ⃗ . Φ ( D ) ] ∨ [ ∃ ρ ⃗ . Φ ( E ) ] : = [ ∃ ρ ⃗ . Φ ( D ∪ E ) ] . [\exists\vec{\rho}.\Phi(D)]\vee[\exists\vec{\rho}.\Phi(E)]\ :=\ [\exists\vec{\rho}.\Phi(D\cup E)]. [ ∃ ρ .Φ ( D )] ∨ [ ∃ ρ .Φ ( E )] := [ ∃ ρ .Φ ( D ∪ E )] .
Example 10 (Pinned atoms and existential quantification). Recall the set M C M_C M C of pinned atoms from Lemma 13. For every ρ \rho ρ -colour r r r , Lemma 9 gives
Φ ( M r ) ≡ Γ ( C ) ∧ ⋀ q ∈ ρ ≤ d ( q ↔ r ( q ) ) , M = ⋃ C . \Phi(\mathcal{M}^{r}) \equiv\Gamma(C) \land\bigwedge_{q \in\rho}^{\leq d}(q \leftrightarrow r(q)), \qquad\mathcal{M} = \bigcup C. Φ ( M r ) ≡ Γ ( C ) ∧ q ∈ ρ ⋀ ≤ d ( q ↔ r ( q )) , M = ⋃ C . Since Γ ( C ) \Gamma(C) Γ ( C ) is ρ \rho ρ -free and the variables in ρ \rho ρ can be interpreted as true or false according to r r r on every frame, we obtain ∃ ρ ⃗ . Φ ( M r ) ≡ Γ ( C ) \exists\vec{\rho}.\Phi(\mathcal{M}^{r}) \equiv\Gamma(C) ∃ ρ .Φ ( M r ) ≡ Γ ( C ) . In fact, by distribution of existential quantification over disjunction, for every non-empty D ⊆ M C D \subseteq M_C D ⊆ M C , ∃ ρ ⃗ . Φ ( D ) ≡ Γ ( C ) \exists\vec{\rho}.\Phi(D) \equiv\Gamma(C) ∃ ρ .Φ ( D ) ≡ Γ ( C ) . Thus, under ordinary counting, the 2 ∣ ρ ∣ 2^{|\rho|} 2 ∣ ρ ∣ distinct atoms in M C M_C M C give rise to 2 2 ∣ ρ ∣ − 1 2^{2^{|\rho|}}-1 2 2 ∣ ρ ∣ − 1 pairwise non-equivalent formulas Φ ( D ) \Phi(D) Φ ( D ) with ∅ ≠ D ⊆ M C \varnothing\ne D \subseteq M_C ∅ = D ⊆ M C . After existential quantification, however, all these formulas belong to a single equivalence class. This illustrates why the fundamental cell inequality (Theorem 14) does not extend to counting after existential quantification. Its proof exploits the independent choices among the atoms in M C M_C M C , which collapse to a single equivalence class after quantification.
Unlike in the unquantified case, we can now have ∃ ρ ⃗ . Φ ( M 1 ) ⊨ ∃ ρ ⃗ . Φ ( M 2 ) \exists\vec{\rho}.\Phi(\mathcal{M}_1) \models\exists\vec{\rho}.\Phi(\mathcal{M}_2) ∃ ρ .Φ ( M 1 ) ⊨ ∃ ρ .Φ ( M 2 ) , even for distinct M 1 , M 2 \mathcal{M}_1,\mathcal{M}_2 M 1 , M 2 , and such an entailment might only hold in one direction. Observe, however, that for any two different cells C 1 , C 2 ∈ Cell σ , ρ d C_1,C_2 \in\operatorname{Cell}^{d}_{\sigma,\rho} C 1 , C 2 ∈ Cell σ , ρ d and non-empty D i ⊆ C i D_i \subseteq C_i D i ⊆ C i , i = 1 , 2 i=1,2 i = 1 , 2 , we still have
∃ ρ ⃗ . Φ ( D 1 ) ⊨ ¬ ∃ ρ ⃗ . Φ ( D 2 ) . \exists\vec{\rho}.\Phi(D_1) \models\neg\exists\vec{\rho}.\Phi(D_2). ∃ ρ .Φ ( D 1 ) ⊨ ¬∃ ρ .Φ ( D 2 ) .
This allows us to extend the cell factorisation to the counting of equivalence classes after existential quantification. For any cell C ∈ Cell σ , ρ d C \in\operatorname{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d , let
Q C = { [ ∃ ρ ⃗ . Φ ( D ) ] ∣ D ⊆ C } , U C Π : = { [ ∃ ρ ⃗ . φ ] ∈ Q C ∣ φ has a uniform σ -interpolant } . Q_C = \{[\exists\vec{\rho}.\Phi(D)] \mid D \subseteq C\}, \qquad U_C^\Pi:= \{[\exists\vec{\rho}.\varphi] \in Q_C \mid\varphi\text{ has a uniform }\sigma\text{-interpolant}\}. Q C = {[ ∃ ρ .Φ ( D )] ∣ D ⊆ C } , U C Π := {[ ∃ ρ . φ ] ∈ Q C ∣ φ has a uniform σ -interpolant } .
The density of uniform interpolation in C C C after existential quantification is defined by
UnifInt C Π : = ∣ U C Π ∣ ∣ Q C ∣ . \operatorname{UnifInt}_{C}^{\Pi} := \frac{|U_C^\Pi|}{|Q_C|}. UnifInt C Π := ∣ Q C ∣ ∣ U C Π ∣ .
Theorem 34 (projective cell reduction principle). UnifInt σ , ρ Π , d = ∏ C ∈ Cell σ , ρ d UnifInt C Π \operatorname{UnifInt}_{\sigma,\rho}^{\Pi,d} = \prod_{C \in\operatorname{Cell}^{d}_{\sigma,\rho}} \operatorname{UnifInt}_{C}^{\Pi} UnifInt σ , ρ Π , d = ∏ C ∈ Cell σ , ρ d UnifInt C Π .
Proof. The cells partition At σ ∪ ρ d \operatorname{At}^{d}_{\sigma\cup\rho} At σ ∪ ρ d and so consider the map
[ ∃ ρ ⃗ . Φ ( D ) ] ⟼ ( [ ∃ ρ ⃗ . Φ ( D ∩ C ) ] ) C ∈ Cell σ , ρ d . [\exists\vec{\rho}.\Phi(D)] \longmapsto\left([\exists\vec{\rho}.\Phi(D \cap C)]\right)_{C \in\operatorname{Cell}^{d}_{\sigma,\rho}}. [ ∃ ρ .Φ ( D )] ⟼ ( [ ∃ ρ .Φ ( D ∩ C )] ) C ∈ Cell σ , ρ d .
This map is well-defined, since
∃ ρ ⃗ . Φ ( D ∩ C ) ≡ Γ ( C ) ∧ ∃ ρ ⃗ . Φ ( D ) . \exists\vec{\rho}.\Phi(D \cap C) \equiv\Gamma(C) \land\exists\vec{\rho}.\Phi(D). ∃ ρ .Φ ( D ∩ C ) ≡ Γ ( C ) ∧ ∃ ρ .Φ ( D ) .
It is also bijective: existentially quantified formulas supported on distinct cells are mutually inconsistent, and every element of Q σ , ρ d Q^{d}_{\sigma,\rho} Q σ , ρ d is the join of its cell components. By Lemma 7, under this bijection, [ ∃ ρ ⃗ . Φ ( D ) ] ∈ U σ , ρ Π , d [\exists\vec{\rho}.\Phi(D)] \in U_{\sigma,\rho}^{\Pi,d} [ ∃ ρ .Φ ( D )] ∈ U σ , ρ Π , d iff [ ∃ ρ ⃗ . Φ ( C ∩ D ) ] ∈ U C Π [\exists\vec{\rho}.\Phi(C \cap D)] \in U_C^\Pi [ ∃ ρ .Φ ( C ∩ D )] ∈ U C Π for every cell C ∈ Cell σ , ρ d C \in\operatorname{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d . Hence
∣ U σ , ρ Π , d ∣ = ∏ C ∈ Cell σ , ρ d ∣ U C Π ∣ , ∣ Q σ , ρ d ∣ = ∏ C ∈ Cell σ , ρ d ∣ Q C ∣ , |U_{\sigma,\rho}^{\Pi,d}| = \prod_{C \in\operatorname{Cell}^{d}_{\sigma,\rho}} |U_C^\Pi|,\qquad|Q_{\sigma,\rho}^{d}| = \prod_{C \in\operatorname{Cell}^{d}_{\sigma,\rho}} |Q_C|, ∣ U σ , ρ Π , d ∣ = C ∈ Cell σ , ρ d ∏ ∣ U C Π ∣ , ∣ Q σ , ρ d ∣ = C ∈ Cell σ , ρ d ∏ ∣ Q C ∣ ,
and the equation follows. □
The definition of Craig interpolation density after existential quantification is analogous to the uniform case, so we provide only a sketch. Fix again disjoint signatures σ \sigma σ and ρ \rho ρ . Let InCon σ , ρ Π , d \operatorname{InCon}_{\sigma,\rho}^{\Pi,d} InCon σ , ρ Π , d be the set of ordered pairs of the form
( [ ∃ ρ ⃗ . Φ ( D 1 ) ] , [ ∃ ρ ⃗ . Φ ( D 2 ) ] ) ([\exists\vec{\rho}.\Phi(D_1)], [\exists\vec{\rho}.\Phi(D_2)]) ([ ∃ ρ .Φ ( D 1 )] , [ ∃ ρ .Φ ( D 2 )])
with D 1 , D 2 ⊆ At σ ∪ ρ d D_1,D_2 \subseteq\operatorname{At}^{d}_{\sigma\cup\rho} D 1 , D 2 ⊆ At σ ∪ ρ d , such that Φ ( D 1 ) \Phi(D_1) Φ ( D 1 ) and Φ ( D 2 ′ ) \Phi(D'_2) Φ ( D 2 ′ ) are inconsistent. As before, here D 2 ′ D'_2 D 2 ′ denotes the image of D 2 D_2 D 2 under the fixed renaming of ρ \rho ρ to ρ ′ \rho' ρ ′ . Let Craig σ , ρ Π , d \operatorname{Craig}_{\sigma,\rho}^{\Pi,d} Craig σ , ρ Π , d be the set of these pairs admitting a Craig σ \sigma σ -interpolant. The corresponding cellwise sets InCon C Π \operatorname{InCon}_{C}^{\Pi} InCon C Π and Craig C Π \operatorname{Craig}_{C}^{\Pi} Craig C Π are defined analogously for each cell C ∈ Cell σ , ρ d C \in\operatorname{Cell}^{d}_{\sigma,\rho} C ∈ Cell σ , ρ d . As in the case of uniform interpolants, these definitions do not depend on the chosen representatives D 1 , D 2 D_1,D_2 D 1 , D 2 . One can show:
Theorem 35 (projective cell reduction principle for Craig interpolation).
∣ I n C o n σ , ρ Π , d ∣ = ∏ C ∈ C e l l σ , ρ d ∣ I n C o n C Π ∣ , ∣ C r a i g σ , ρ Π , d ∣ = ∏ C ∈ C e l l σ , ρ d ∣ C r a i g C Π ∣ . |\mathrm{InCon}_{\sigma,\rho}^{\Pi,d}|=\prod_{C\in\mathrm{Cell}_{\sigma,\rho}^{d}}|\mathrm{InCon}_{C}^{\Pi}|,\qquad|\mathrm{Craig}_{\sigma,\rho}^{\Pi,d}|=\prod_{C\in\mathrm{Cell}_{\sigma,\rho}^{d}}|\mathrm{Craig}_{C}^{\Pi}|. ∣ InCon σ , ρ Π , d ∣ = C ∈ Cell σ , ρ d ∏ ∣ InCon C Π ∣ , ∣ Craig σ , ρ Π , d ∣ = C ∈ Cell σ , ρ d ∏ ∣ Craig C Π ∣.
The density of Craig interpolation after existential quantification is defined by
D C r a i g σ , ρ Π , d : = ∣ C r a i g σ , ρ Π , d ∣ ∣ I n C o n σ , ρ Π , d ∣ , D C r a i g C Π : = ∣ C r a i g C Π ∣ ∣ I n C o n C Π ∣ . \mathrm{DCraig}_{\sigma,\rho}^{\Pi,d}:=\frac{|\mathrm{Craig}_{\sigma,\rho}^{\Pi,d}|}{|\mathrm{InCon}_{\sigma,\rho}^{\Pi,d}|},\qquad\mathrm{DCraig}_{C}^{\Pi}:=\frac{|\mathrm{Craig}_{C}^{\Pi}|}{|\mathrm{InCon}_{C}^{\Pi}|}. DCraig σ , ρ Π , d := ∣ InCon σ , ρ Π , d ∣ ∣ Craig σ , ρ Π , d ∣ , DCraig C Π := ∣ InCon C Π ∣ ∣ Craig C Π ∣ .
By Theorem 35, we obtain
D C r a i g σ , ρ Π , d = ∏ C ∈ C e l l σ , ρ d D C r a i g C Π . \mathrm{DCraig}_{\sigma,\rho}^{\Pi,d}=\prod_{C\in\mathrm{Cell}_{\sigma,\rho}^{d}}\mathrm{DCraig}_{C}^{\Pi}. DCraig σ , ρ Π , d = C ∈ Cell σ , ρ d ∏ DCraig C Π .
Asymptotics for m m m fixed and n → ∞ n \to\infty n → ∞
We first fix ρ \rho ρ and let σ \sigma σ grow. The analysis is essentially the same as in the ordinary, unquantified case, so we keep the exposition brief. Indeed, the modified counting method concerns the forgotten signature ρ \rho ρ , which remains fixed in this regime.
As before, let σ n = { p 1 , … , p n } \sigma_n=\{p_1,\ldots,p_n\} σ n = { p 1 , … , p n } and ρ m = { q 1 , … , q m } \rho_m=\{q_1,\ldots,q_m\} ρ m = { q 1 , … , q m } be disjoint signatures, and use numerical parameters n , m n,m n , m in place of the signatures. For example, U n i f I n t n , m Π , d \mathrm{UnifInt}_{n,m}^{\Pi,d} UnifInt n , m Π , d denotes the density of formulas in M L σ n ∪ ρ m d \mathrm{ML}_{\sigma_n\cup\rho_m}^{d} ML σ n ∪ ρ m d , counted up to equivalence after existential quantification over ρ m \rho_m ρ m , that admit uniform σ n \sigma_n σ n -interpolants.
The lifting and flipping constructions from the proofs of Theorems 18 and 24 yield the same asymptotic behaviour in the quantified case. As before, we write ( C m ) m ≥ m ∗ (C_m)_{m\ge m^*} ( C m ) m ≥ m ∗ for the trail of a cell C m ∗ C_{m^*} C m ∗ containing a counterexample to interpolation.
Theorem 36. Let L = Log ( K ) L=\operatorname{Log}(\mathcal{K}) L = Log ( K ) and fix d ≥ 0 d\ge0 d ≥ 0 .
If Φ ( D ) \Phi(D) Φ ( D ) , for D ⊆ C m ∗ ∈ C e l l σ n ∗ , ρ m ∗ d D\subseteq C_{m^*}\in\mathrm{Cell}_{\sigma_{n^*},\rho_{m^*}}^{d} D ⊆ C m ∗ ∈ Cell σ n ∗ , ρ m ∗ d , has no uniform σ n ∗ \sigma_{n^*} σ n ∗ -interpolant, then, for every fixed m ≥ m ∗ m\ge m^* m ≥ m ∗ and all n ≥ n ∗ n\ge n^* n ≥ n ∗ ,
U n i f I n t n , m Π , d ≤ ( 1 − 2 − ∣ C m ∣ ) 2 n → n → ∞ 0. \mathrm{UnifInt}_{n,m}^{\Pi,d}\leq\left(1-2^{-|C_m|}\right)^{2^n}\xrightarrow[n\to\infty]{}0. UnifInt n , m Π , d ≤ ( 1 − 2 − ∣ C m ∣ ) 2 n n → ∞ 0. If there is an inconsistent pair, for D 1 , D 2 ⊆ C m ∗ ∈ C e l l σ n ∗ , ρ m ∗ d D_1,D_2\subseteq C_{m^*}\in\mathrm{Cell}_{\sigma_{n^*},\rho_{m^*}}^{d} D 1 , D 2 ⊆ C m ∗ ∈ Cell σ n ∗ , ρ m ∗ d , without a Craig σ n ∗ \sigma_{n^*} σ n ∗ -interpolant, then, for every fixed m ≥ m ∗ m\ge m^* m ≥ m ∗ and all n ≥ n ∗ n\ge n^* n ≥ n ∗ ,
D C r a i g n , m Π , d ≤ ( 1 − 2 − 2 ∣ C m ∣ ) 2 n → n → ∞ 0. \mathrm{DCraig}_{n,m}^{\Pi,d}\leq\left(1-2^{-2|C_m|}\right)^{2^n}\xrightarrow[n\to\infty]{}0. DCraig n , m Π , d ≤ ( 1 − 2 − 2∣ C m ∣ ) 2 n n → ∞ 0.
Proof. The proof follows the same argument as in Theorems 18 and 24. It suffices to observe that, for any σ τ \sigma\tau σ τ -cell C C C , the following hold:
(i) if some Φ ( D ) \Phi(D) Φ ( D ) with D ⊆ C D\subseteq C D ⊆ C has no uniform σ \sigma σ -interpolant, then U n i f I n t C Π ≤ 1 − 1 ∣ Q C ∣ ≤ 1 − 2 − ∣ C ∣ \mathrm{UnifInt}_{C}^{\Pi}\leq1-\frac{1}{|Q_C|}\leq1-2^{-|C|} UnifInt C Π ≤ 1 − ∣ Q C ∣ 1 ≤ 1 − 2 − ∣ C ∣ ;
(ii) if I n C o n C ∖ C r a i g C ≠ ∅ \mathrm{InCon}_{C}\setminus\mathrm{Craig}_{C}\ne\varnothing InCon C ∖ Craig C = ∅ , then D C r a i g C Π ≤ 1 − 1 ∣ Q C ∣ 2 ≤ 1 − 2 − 2 ∣ C ∣ \mathrm{DCraig}_{C}^{\Pi}\leq1-\frac{1}{|Q_C|^2}\leq1-2^{-2|C|} DCraig C Π ≤ 1 − ∣ Q C ∣ 2 1 ≤ 1 − 2 − 2∣ C ∣ .
Indeed, ∣ Q C ∣ ≤ 2 ∣ C ∣ |Q_C|\leq2^{|C|} ∣ Q C ∣ ≤ 2 ∣ C ∣ , and each hypothesis guarantees at least one failure among the corresponding equivalence classes. Applying the same lifting and flipping arguments as in Theorems 18 and 24, together with the projective cell reduction principles, yields the claimed bounds.
Asymptotics for n n n fixed and m → ∞ m \to\infty m → ∞
We now fix σ \sigma σ and let ρ \rho ρ grow. In this regime, the asymptotic behaviour of the density depends on the modal logic and can differ significantly from that of the ordinary density. As in the ordinary-counting case, we focus on locally tabular logics and therefore drop the modal-depth restriction. For every finite signature τ \tau τ , the set M L τ \mathrm{ML}_{\tau} ML τ is then finite up to logical equivalence. The definitions and results from the bounded-depth setting carry over accordingly. For instance, for disjoint signatures σ \sigma σ and ρ \rho ρ , the set Q σ , ρ = { [ ∃ p ˉ . φ ] ∣ φ ∈ M L σ ∪ ρ } Q_{\sigma,\rho}=\{[\exists\bar{p}.\varphi]\mid\varphi\in\mathrm{ML}_{\sigma\cup\rho}\} Q σ , ρ = {[ ∃ p ˉ . φ ] ∣ φ ∈ ML σ ∪ ρ } consists of the equivalence classes of existentially quantified formulas.
Fix disjoint signatures σ \sigma σ and ρ \rho ρ , and let C ∈ C e l l σ , ρ C \in\mathrm{Cell}_{\sigma,\rho} C ∈ Cell σ , ρ be a σ ⊤ ρ \sigma^{\mathord{\top}\rho} σ ⊤ ρ -cell. Let H C = ( C , ⌢ ) H_C=(C,\mathord{\frown}) H C = ( C , ⌢ ) be its incompatibility graph and let
π C : C → P = C / ∼ C \pi_C:C\to P=C/{\sim_C} π C : C → P = C / ∼ C
be the canonical profile map with canonical profile H P = ( P , ⌢ P ) H_P=(P,\mathord{\frown}_P) H P = ( P , ⌢ P ) . Thus ∣ P ∣ = inc ( C ) |P|=\operatorname{inc}(C) ∣ P ∣ = inc ( C ) . For S ⊆ P S\subseteq P S ⊆ P , define
Q C [ S ] : = { [ ∃ ρ ⃗ . Φ ( D ) ] ∣ D ⊆ π C − 1 ( S ) } . \mathcal{Q}_C[S]:=\{[\exists\vec{\rho}.\Phi(D)]\mid D\subseteq\pi_C^{-1}(S)\}. Q C [ S ] := {[ ∃ ρ .Φ ( D )] ∣ D ⊆ π C − 1 ( S )} .
Note that ⊥ = [ ∃ ρ ⃗ . Φ ( ∅ ) ] \bot=[\exists\vec{\rho}.\Phi(\varnothing)] ⊥ = [ ∃ ρ .Φ ( ∅ )] belongs to Q C [ S ] \mathcal{Q}_C[S] Q C [ S ] for every S ⊆ P S\subseteq P S ⊆ P , including S = ∅ S=\varnothing S = ∅ .
As in the non-quantified case (Theorem 26 ), we start with counting the set Craig C Π \operatorname{Craig}^{\Pi}_C Craig C Π of pairs ( x , y ) ∈ Q C × Q C (x,y)\in\mathcal{Q}_C\times\mathcal{Q}_C ( x , y ) ∈ Q C × Q C for which a Craig interpolant exists. As these are exactly the pairs of the form ( x , ⊥ ) (x,\bot) ( x , ⊥ ) or ( ⊥ , y ) (\bot,y) ( ⊥ , y ) with x , y ∈ Q C x,y\in\mathcal{Q}_C x , y ∈ Q C we obtain:
Theorem 37. Let L = Log ( K ) L=\operatorname{Log}(\mathcal{K}) L = Log ( K ) be a locally tabular quasi-normal logic. Then, for every σ ⊤ ρ \sigma^{\mathord{\top}\rho} σ ⊤ ρ -cell C C C ,
∣ Craig C Π ∣ = 2 ∣ Q C ∣ − 1. |\operatorname{Craig}^{\Pi}_C|=2|\mathcal{Q}_C|-1. ∣ Craig C Π ∣ = 2∣ Q C ∣ − 1.
The next observation allows us to bound the number of disjunctions built from formulas in Q C [ S ] \mathcal{Q}_C[S] Q C [ S ] and Q C [ T ] \mathcal{Q}_C[T] Q C [ T ] if ( S , T ) (S,T) ( S , T ) is an ordered biclique in H P H_P H P .
Lemma 38. If ( S , T ) (S,T) ( S , T ) is an ordered biclique of H P H_P H P , then
J S , T : Q C [ S ] × Q C [ T ] ⟶ Q C , J S , T ( x , y ) = x ∨ y , J_{S,T}:\mathcal{Q}_C[S]\times\mathcal{Q}_C[T]\longrightarrow\mathcal{Q}_C,\qquad J_{S,T}(x,y)=x\vee y, J S , T : Q C [ S ] × Q C [ T ] ⟶ Q C , J S , T ( x , y ) = x ∨ y , is injective. Consequently, ∣ Q C [ S ] ∣ ∣ Q C [ T ] ∣ ≤ ∣ Q C ∣ |\mathcal{Q}_C[S]||\mathcal{Q}_C[T]|\leq|\mathcal{Q}_C| ∣ Q C [ S ] ∣∣ Q C [ T ] ∣ ≤ ∣ Q C ∣ .
Proof. Let x , x ′ ∈ Q C [ S ] x,x'\in\mathcal{Q}_C[S] x , x ′ ∈ Q C [ S ] and y , y ′ ∈ Q C [ T ] y,y'\in\mathcal{Q}_C[T] y , y ′ ∈ Q C [ T ] , and suppose that x ∨ y = x ′ ∨ y ′ x\vee y=x'\vee y' x ∨ y = x ′ ∨ y ′ . Since ( S , T ) (S,T) ( S , T ) is an ordered biclique, every element of Q C [ S ] \mathcal{Q}_C[S] Q C [ S ] is inconsistent with every element of Q C [ T ] \mathcal{Q}_C[T] Q C [ T ] . It follows that x = x ′ x=x' x = x ′ and y = y ′ y=y' y = y ′ and so J S , T J_{S,T} J S , T is injective.
We now assume that L L L has BICIP. So assume that inc ( C ) = ∣ P ∣ ≤ f L ( n ) \operatorname{inc}(C)=|P|\leq f_L(n) inc ( C ) = ∣ P ∣ ≤ f L ( n ) for all σ ⊤ ρ \sigma^{\mathord{\top}\rho} σ ⊤ ρ -cells, with ∣ σ ∣ = n |\sigma|=n ∣ σ ∣ = n , independent from m m m . We use that then the number of non-empty ordered pairs of disjoint subsets of P P P is bounded by
κ L ( n ) : = 3 f L ( n ) − 2 f L ( n ) + 1 + 1. (4) \kappa_L(n):=3^{f_L(n)}-2^{f_L(n)+1}+1. \tag*{(4)} κ L ( n ) := 3 f L ( n ) − 2 f L ( n ) + 1 + 1. (4)
Theorem 39. If L = Log ( K ) L=\operatorname{Log}(\mathcal{K}) L = Log ( K ) has BICIP, then lim inf m → ∞ DCraig n , m Π > 0 \liminf_{m\to\infty}\operatorname{DCraig}^{\Pi}_{n,m}>0 lim inf m → ∞ DCraig n , m Π > 0 , for every n ≥ 0 n\geq0 n ≥ 0 .
Proof. Let I C \mathcal{I}_C I C be the set of inconsistent pairs in ( Q C ∖ { ⊥ } ) × ( Q C ∖ { ⊥ } ) (\mathcal{Q}_C\setminus\{\bot\})\times(\mathcal{Q}_C\setminus\{\bot\}) ( Q C ∖ { ⊥ }) × ( Q C ∖ { ⊥ }) . Then
NoCraig C Π = ∣ InCon C Π ∣ − ∣ Craig C Π ∣ = ∣ I C ∣ . \operatorname{NoCraig}^{\Pi}_C=|\operatorname{InCon}^{\Pi}_C|-|\operatorname{Craig}^{\Pi}_C|=|\mathcal{I}_C|. NoCraig C Π = ∣ InCon C Π ∣ − ∣ Craig C Π ∣ = ∣ I C ∣.
We bound NoCraig C Π \operatorname{NoCraig}^{\Pi}_C NoCraig C Π from above. Choose for ( x , y ) ∈ I C (x,y)\in\mathcal{I}_C ( x , y ) ∈ I C any sets D , E ⊆ C D,E\subseteq C D , E ⊆ C with x = ∃ ρ ⃗ . Φ ( D ) x=\exists\vec{\rho}.\Phi(D) x = ∃ ρ .Φ ( D ) and y = ∃ ρ ⃗ . Φ ( E ) y=\exists\vec{\rho}.\Phi(E) y = ∃ ρ .Φ ( E ) . Then the ordered pair of images ( π C [ D ] , π C [ E ] ) (\pi_C[D],\pi_C[E]) ( π C [ D ] , π C [ E ]) is an ordered biclique of non-empty sets in H P H_P H P and ( x , y ) ∈ ( Q C [ π C [ D ] ] , Q C [ π C [ E ] ] ) (x,y)\in(\mathcal{Q}_C[\pi_C[D]],\mathcal{Q}_C[\pi_C[E]]) ( x , y ) ∈ ( Q C [ π C [ D ]] , Q C [ π C [ E ]]) . We have shown
I C ⊆ ⋃ ∅ ≠ S , T ⊆ P S ⌢ P T Q C [ S ] × Q C [ T ] . \mathcal{I}_C\subseteq\bigcup_{\substack{\varnothing\ne S,T\subseteq P\\S\mathrel{\frown_P}T}}\mathcal{Q}_C[S]\times\mathcal{Q}_C[T]. I C ⊆ ∅ = S , T ⊆ P S ⌢ P T ⋃ Q C [ S ] × Q C [ T ] .
There are at most κ L ( n ) \kappa_L(n) κ L ( n ) such ordered bicliques in H P H_P H P and for each pair ( S , T ) (S,T) ( S , T ) in this union, Lemma 38 gives ∣ Q C [ S ] ∣ ∣ Q C [ T ] ∣ ≤ ∣ Q C ∣ |\mathcal{Q}_C[S]||\mathcal{Q}_C[T]|\leq|\mathcal{Q}_C| ∣ Q C [ S ] ∣∣ Q C [ T ] ∣ ≤ ∣ Q C ∣ . Hence we obtain the upper bound
NoCraig C Π ≤ ∑ ∅ ≠ S , T ⊆ P S ⌢ P T ∣ Q C [ S ] ∣ ∣ Q C [ T ] ∣ ≤ κ L ( n ) ∣ Q C ∣ . \operatorname{NoCraig}^{\Pi}_C\leq\sum_{\substack{\varnothing\ne S,T\subseteq P\\S\mathrel{\frown_P}T}}|\mathcal{Q}_C[S]||\mathcal{Q}_C[T]|\leq\kappa_L(n)|\mathcal{Q}_C|. NoCraig C Π ≤ ∅ = S , T ⊆ P S ⌢ P T ∑ ∣ Q C [ S ] ∣∣ Q C [ T ] ∣ ≤ κ L ( n ) ∣ Q C ∣.
Using Theorem 37, we obtain
DCraig C Π = 2 ∣ Q C ∣ − 1 2 ∣ Q C ∣ − 1 + NoCraig C Π ≥ 2 ∣ Q C ∣ − 1 2 ∣ Q C ∣ − 1 + κ L ( n ) ∣ Q C ∣ ≥ 1 1 + κ L ( n ) . \operatorname{DCraig}^{\Pi}_C
=\frac{2|\mathcal{Q}_C|-1}{2|\mathcal{Q}_C|-1+\operatorname{NoCraig}^{\Pi}_C}
\geq\frac{2|\mathcal{Q}_C|-1}{2|\mathcal{Q}_C|-1+\kappa_L(n)|\mathcal{Q}_C|}
\geq\frac{1}{1+\kappa_L(n)}. DCraig C Π = 2∣ Q C ∣ − 1 + NoCraig C Π 2∣ Q C ∣ − 1 ≥ 2∣ Q C ∣ − 1 + κ L ( n ) ∣ Q C ∣ 2∣ Q C ∣ − 1 ≥ 1 + κ L ( n ) 1 .
Using projective cell reduction principle for Craig interpolation (Theorem 35 ) and the fact that there are ∣ At σ n ∣ |\operatorname{At}_{\sigma_n}| ∣ At σ n ∣ factors, independently of m m m , it follows that
DCraig n , m Π ≥ ( 1 1 + κ L ( n ) ) ∣ At σ n ∣ . \operatorname{DCraig}^{\Pi}_{n,m} \ge\left(\frac{1}{1+\kappa_L(n)}\right)^{|\operatorname{At}_{\sigma_n}|}. DCraig n , m Π ≥ ( 1 + κ L ( n ) 1 ) ∣ At σ n ∣ .
Hence we have obtained a lower bound for DCraig n , m Π \operatorname{DCraig}^{\Pi}_{n,m} DCraig n , m Π that does not depend on m m m .
Fix a cell C 0 ∈ Cell σ , ∅ C_0 \in\operatorname{Cell}_{\sigma,\emptyset} C 0 ∈ Cell σ , ∅ and consider its trail ( C m ) m ≥ 0 (C_m)_{m\ge0} ( C m ) m ≥ 0 . For m ≤ ℓ m \le\ell m ≤ ℓ , let ι m , ℓ : C m → C ℓ \iota_{m,\ell}: C_m \to C_\ell ι m , ℓ : C m → C ℓ be the pinning operation on atoms that interpret in every model of every M ∈ C m \mathcal{M} \in C_m M ∈ C m every variable in ρ ℓ ∖ ρ m \rho_\ell\setminus\rho_m ρ ℓ ∖ ρ m as false. Then
M ∼ N ⟺ ι m , ℓ ( M ) ∼ ι m , ℓ ( N ) . (5) \mathcal{M} \mathrel{\sim} \mathcal{N} \quad\Longleftrightarrow\quad\iota_{m,\ell}(\mathcal{M}) \mathrel{\sim} \iota_{m,\ell}(\mathcal{N}). \tag*{(5)} M ∼ N ⟺ ι m , ℓ ( M ) ∼ ι m , ℓ ( N ) . (5)
Indeed, a joint model on the left defines a joint model on the right by interpreting the variables in ρ ℓ ∖ ρ m \rho_\ell\setminus\rho_m ρ ℓ ∖ ρ m as false on both sides. A joint model on the right is also a joint model on the left. Let P ( m ) P(m) P ( m ) be the canonical profile of C m C_m C m and π m \pi_m π m the canonical profile mapping, for all m ≥ 0 m \ge0 m ≥ 0 .
Lemma 40. Assume that L = Log ( K ) L = \operatorname{Log}(\mathcal{K}) L = Log ( K ) has BICIP witnessed by f L ( n ) f_L(n) f L ( n ) . Then there is m 0 > 0 m_0 > 0 m 0 > 0 such that, for every m ≥ m 0 m \ge m_0 m ≥ m 0 and D ⊆ C m D \subseteq C_m D ⊆ C m ,
Φ ( D ) has a uniform σ -interpolant ⟺ D = ∅ or N P ( m ) ( π m [ D ] ) = ∅ . (6) \Phi(D)\text{ has a uniform }\sigma\text{-interpolant} \quad\Longleftrightarrow\quad D = \emptyset\text{ or }N_{P(m)}(\pi_m[D]) = \emptyset. \tag*{(6)} Φ ( D ) has a uniform σ -interpolant ⟺ D = ∅ or N P ( m ) ( π m [ D ]) = ∅. (6)
Proof. For m ≤ ℓ m \le\ell m ≤ ℓ , choose A p ∈ C m A_p \in C_m A p ∈ C m with π m ( A p ) = p \pi_m(A_p) = p π m ( A p ) = p , for each p ∈ P ( m ) p \in P(m) p ∈ P ( m ) . By (4), the map p ↦ π ℓ ( ι m , ℓ ( A p ) ) p \mapsto\pi_\ell(\iota_{m,\ell}(A_p)) p ↦ π ℓ ( ι m , ℓ ( A p )) is injective. Hence ∣ P ( m ) ∣ |P(m)| ∣ P ( m ) ∣ is nondecreasing and, by the hypothesis, bounded by f L ( ∣ σ ∣ ) f_L(|\sigma|) f L ( ∣ σ ∣ ) . Fix m 0 > 0 m_0 > 0 m 0 > 0 from which the ∣ P ( m ) ∣ |P(m)| ∣ P ( m ) ∣ are constant. If m 0 ≤ m ≤ ℓ m_0 \le m \le\ell m 0 ≤ m ≤ ℓ , then for every B ∈ C ℓ B \in C_\ell B ∈ C ℓ there is p ∈ P ( m ) p \in P(m) p ∈ P ( m ) with π ℓ ( B ) = π ℓ ( ι m , ℓ ( A p ) ) \pi_\ell(B) = \pi_\ell(\iota_{m,\ell}(A_p)) π ℓ ( B ) = π ℓ ( ι m , ℓ ( A p )) . Thus, for all M ∈ C m \mathcal{M} \in C_m M ∈ C m ,
ι m , ℓ ( M ) ∼ B ⟺ M ∼ A p . (7) \iota_{m,\ell}(\mathcal{M}) \mathrel{\sim} B \quad\Longleftrightarrow\quad\mathcal{M} \mathrel{\sim} A_p. \tag*{(7)} ι m , ℓ ( M ) ∼ B ⟺ M ∼ A p . (7)
The case D = ∅ D = \emptyset D = ∅ is immediate. Suppose D ≠ ∅ D \ne\emptyset D = ∅ . Then Φ ( D ) \Phi(D) Φ ( D ) has no uniform σ \sigma σ -interpolant if and only if the cell characteristic formula Γ ( C m ) \Gamma(C_m) Γ ( C m ) is not a uniform σ \sigma σ -interpolant for Π ( D ) \Pi(D) Π ( D ) (by Lemma 25 ) if and only if there is a ρ m \rho_m ρ m -free formula χ \chi χ such that Φ ( D ) ⊨ χ \Phi(D) \models\chi Φ ( D ) ⊨ χ and Γ ( C m ) ⊭ χ \Gamma(C_m) \not\models\chi Γ ( C m ) ⊨ χ . Suppose such a χ \chi χ exists. Renaming its variables outside σ \sigma σ if necessary, choose ℓ ≥ m \ell\ge m ℓ ≥ m so that χ \chi χ uses only variables in σ ∪ ( ρ ℓ ∖ ρ m ) \sigma\cup(\rho_\ell\setminus\rho_m) σ ∪ ( ρ ℓ ∖ ρ m ) . There is ( M , w ) ∈ PMod L (\mathcal{M}, w) \in\operatorname{PMod}_L ( M , w ) ∈ PMod L with M , w ⊨ Γ ( C m ) ∧ ¬ χ \mathcal{M}, w \models\Gamma(C_m) \land\neg\chi M , w ⊨ Γ ( C m ) ∧ ¬ χ . Put B = [ M , w ] σ ∪ ρ ℓ B = [\mathcal{M}, w]_{\sigma\cup\rho_\ell} B = [ M , w ] σ ∪ ρ ℓ ; then B ∈ C ℓ B \in C_\ell B ∈ C ℓ and Φ ( B ) ⊨ ¬ χ \Phi(B) \models\neg\chi Φ ( B ) ⊨ ¬ χ . For every M ∈ D \mathcal{M} \in D M ∈ D , the implication Φ ( M ) ⊨ χ \Phi(\mathcal{M}) \models\chi Φ ( M ) ⊨ χ holds, also after renaming the variables of χ \chi χ outside σ \sigma σ . Therefore ι m , ℓ ( M ) ∼ B \iota_{m,\ell}(\mathcal{M}) \mathrel{\sim} B ι m , ℓ ( M ) ∼ B . By (6), some p ∈ P ( m ) p \in P(m) p ∈ P ( m ) satisfies M ∼ A p \mathcal{M} \mathrel{\sim} A_p M ∼ A p for every M ∈ D \mathcal{M} \in D M ∈ D . Hence N P ( m ) ( π m [ D ] ) ≠ ∅ N_{P(m)}(\pi_m[D]) \ne\emptyset N P ( m ) ( π m [ D ]) = ∅ .
Conversely, suppose p ∈ N P ( m ) ( π m [ D ] ) p \in N_{P(m)}(\pi_m[D]) p ∈ N P ( m ) ( π m [ D ]) . Let ρ m ′ \rho'_m ρ m ′ be a fresh copy of ρ m \rho_m ρ m and A p ′ A'_p A p ′ the corresponding renamed atom, using the prime notation from § 5 . Since M ∼ A p \mathcal{M} \mathrel{\sim} A_p M ∼ A p for every M ∈ D \mathcal{M} \in D M ∈ D , Proposition 19 gives Φ ( D ) ⊨ ¬ Φ ( A p ′ ) \Phi(D) \models\neg\Phi(A'_p) Φ ( D ) ⊨ ¬Φ ( A p ′ ) . On the other hand, Φ ( A p ′ ) \Phi(A'_p) Φ ( A p ′ ) is consistent and entails Γ ( C m ) \Gamma(C_m) Γ ( C m ) , because A p ∈ C m A_p \in C_m A p ∈ C m . Thus Γ ( C m ) ⊭ ¬ Φ ( A p ′ ) \Gamma(C_m) \not\models\neg\Phi(A'_p) Γ ( C m ) ⊨ ¬Φ ( A p ′ ) . Taking χ = ¬ Φ ( A p ′ ) \chi= \neg\Phi(A'_p) χ = ¬Φ ( A p ′ ) proves that Φ ( D ) \Phi(D) Φ ( D ) has no uniform σ \sigma σ -interpolant.
Lemma 41. Let H P = ( P , ∼ P ) H_P = (P,\mathrel{\sim_P}) H P = ( P , ∼ P ) be a canonical profile, let ∅ ≠ S ⊆ P \emptyset\ne S \subseteq P ∅ = S ⊆ P and suppose N P ( S ) ≠ ∅ N_P(S) \ne\emptyset N P ( S ) = ∅ . Then there is a nonempty A ⊆ P A \subseteq P A ⊆ P such that ( S , A ) (S,A) ( S , A ) is an ordered biclique of H P H_P H P and N P ( S ∪ A ) = ∅ N_P(S \cup A) = \emptyset N P ( S ∪ A ) = ∅ .
Proof. Put U 0 : = S U_0 := S U 0 := S and assume U i U_i U i has been defined. If N P ( U i ) ≠ ∅ N_P(U_i) \ne\emptyset N P ( U i ) = ∅ , choose a i + 1 ∈ N P ( U i ) a_{i+1} \in N_P(U_i) a i + 1 ∈ N P ( U i ) and put U i + 1 : = U i ∪ { a i + 1 } U_{i+1} := U_i \cup\{a_{i+1}\} U i + 1 := U i ∪ { a i + 1 } . Irreflexivity implies a i + 1 ∉ U i a_{i+1} \notin U_i a i + 1 ∈ / U i , for otherwise a i + 1 ∼ P a i + 1 a_{i+1} \mathrel{\sim_P} a_{i+1} a i + 1 ∼ P a i + 1 . Thus, the cardinality of U i U_i U i increases strictly at every stage. Since P P P is finite, the process terminates after at most ∣ P ∣ |P| ∣ P ∣ steps at a set U k U_k U k with N P ( U k ) = ∅ N_P(U_k) = \emptyset N P ( U k ) = ∅ . Since N P ( S ) ≠ ∅ N_P(S) \ne\emptyset N P ( S ) = ∅ , k ≥ 1 k \ge1 k ≥ 1 .
Put A : = U k ∖ S A := U_k \setminus S A := U k ∖ S . Then A A A is non-empty and N P ( S ∪ A ) = ∅ N_P(S \cup A) = \emptyset N P ( S ∪ A ) = ∅ holds. Every a i a_i a i was chosen incompatible with all members of S ⊆ U i − 1 S \subseteq U_{i-1} S ⊆ U i − 1 , so ( S , A ) (S,A) ( S , A ) is an ordered biclique of H P H_P H P .
Theorem 42. If L = Log ( K ) L = \operatorname{Log}(\mathcal{K}) L = Log ( K ) has BICIP, then lim inf m → ∞ UnifInt n , m Π > 0 \liminf_{m\to\infty}\operatorname{UnifInt}^{\Pi}_{n,m} > 0 lim inf m → ∞ UnifInt n , m Π > 0 for every n ≥ 0 n \ge0 n ≥ 0 .
Proof. Fix n ≥ 0 n \ge0 n ≥ 0 , signature σ \sigma σ with ∣ σ ∣ = n |\sigma| = n ∣ σ ∣ = n and a cell C 0 ∈ Cell σ , ∅ C_0 \in\operatorname{Cell}_{\sigma,\emptyset} C 0 ∈ Cell σ , ∅ . Fix m 0 m_0 m 0 supplied by Lemma 40 . Consider any element C = C m C = C_m C = C m of the trail ( C m ) m ≥ 0 (C_m)_{m\ge0} ( C m ) m ≥ 0 of C 0 C_0 C 0 for some m ≥ m 0 m \ge m_0 m ≥ m 0 . Let
G C : = { x ∈ Q C ∣ x has a uniform σ -interpolant } , B C : = Q C ∖ G C . \mathcal{G}_C := \{x \in Q_C \mid x\text{ has a uniform }\sigma\text{-interpolant}\}, \qquad\mathcal{B}_C := Q_C \setminus\mathcal{G}_C. G C := { x ∈ Q C ∣ x has a uniform σ -interpolant } , B C := Q C ∖ G C .
e n d s m i d = 1 {ends_mid=1} e n d s m i d = 1
Let π C \pi_C π C be the canonical profile mapping from C C C to canonical profile P P P . For every x ∈ B C x \in\mathcal{B}_C x ∈ B C , choose a non-empty representative D x ⊆ C D_x \subseteq C D x ⊆ C and put S x : = π C [ D x ] S_x := \pi_C[D_x] S x := π C [ D x ] . By Lemma 40, N P ( S x ) ≠ ∅ N_P(S_x) \ne\varnothing N P ( S x ) = ∅ . Apply Lemma 41 and choose nonempty A x A_x A x such that ( S x , A x ) (S_x,A_x) ( S x , A x ) is an ordered biclique of H P H_P H P and N P ( S x ∪ A x ) = ∅ N_P(S_x \cup A_x) = \varnothing N P ( S x ∪ A x ) = ∅ . For every a ∈ P a \in P a ∈ P , choose M a ∈ C \mathcal{M}_a \in C M a ∈ C with π C ( M a ) = a \pi_C(\mathcal{M}_a) = a π C ( M a ) = a , and define e a : = [ ∃ ρ ⃗ . Φ σ ∪ ρ m ( { M a } ) ] ∈ Q C [ { a } ] e_a := [\exists\vec{\rho}.\Phi_{\sigma\cup\rho_m}(\{\mathcal{M}_a\})] \in Q_C[\{a\}] e a := [ ∃ ρ . Φ σ ∪ ρ m ({ M a })] ∈ Q C [{ a }] . For non-empty A ⊆ P A \subseteq P A ⊆ P , put z A : = ⋁ a ∈ A e a ∈ Q C [ A ] z_A := \bigvee_{a\in A} e_a \in Q_C[A] z A := ⋁ a ∈ A e a ∈ Q C [ A ] . For each ordered pair ( S , A ) (S,A) ( S , A ) occurring among the chosen pairs ( S x , A x ) (S_x,A_x) ( S x , A x ) , define
B S , A : = { x ∈ B C ∣ ( S x , A x ) = ( S , A ) } \mathcal{B}_{S,A} := \{x \in\mathcal{B}_C \mid(S_x,A_x) = (S,A)\} B S , A := { x ∈ B C ∣ ( S x , A x ) = ( S , A )}
and
F S , A : B S , A ⟶ Q C , F S , A ( x ) : = x ∨ z A . F_{S,A}: \mathcal{B}_{S,A} \longrightarrow Q_C,\qquad F_{S,A}(x) := x \vee z_A. F S , A : B S , A ⟶ Q C , F S , A ( x ) := x ∨ z A .
If D x D_x D x represents x x x , then F S , A ( x ) F_{S,A}(x) F S , A ( x ) is represented by D x ∪ { M a ∣ a ∈ A } D_x \cup\{\mathcal{M}_a \mid a \in A\} D x ∪ { M a ∣ a ∈ A } and π C ( D x ∪ { M a ∣ a ∈ A } ) = S ∪ A \pi_C(D_x \cup\{\mathcal{M}_a \mid a \in A\}) = S \cup A π C ( D x ∪ { M a ∣ a ∈ A }) = S ∪ A . N P ( S ∪ A ) = ∅ N_P(S \cup A) = \varnothing N P ( S ∪ A ) = ∅ , so Lemma 40 gives F S , A ( x ) ∈ G C F_{S,A}(x) \in\mathcal{G}_C F S , A ( x ) ∈ G C . Moreover, ( S , A ) (S,A) ( S , A ) is an ordered biclique of H P H_P H P . Lemma 38, with the second coordinate fixed at z A z_A z A , shows that F S , A F_{S,A} F S , A is injective. Therefore ∣ B S , A ∣ ≤ ∣ G C ∣ |\mathcal{B}_{S,A}| \le|\mathcal{G}_C| ∣ B S , A ∣ ≤ ∣ G C ∣ . Every occurring ( S , A ) (S,A) ( S , A ) is a non-empty ordered biclique and has disjoint sides. There are at most κ L ( n ) \kappa_L(n) κ L ( n ) such pairs. Since the B S , A \mathcal{B}_{S,A} B S , A partition B C \mathcal{B}_C B C , we obtain ∣ B C ∣ ≤ κ L ( n ) ∣ G C ∣ |\mathcal{B}_C| \le\kappa_L(n)|\mathcal{G}_C| ∣ B C ∣ ≤ κ L ( n ) ∣ G C ∣ . It follows that
UnifInt C Π = ∣ G C ∣ ∣ G C ∣ + ∣ B C ∣ ≥ 1 1 + κ L ( n ) . \operatorname{UnifInt}^{\Pi}_{C} = \frac{|\mathcal{G}_C|}{|\mathcal{G}_C|+|\mathcal{B}_C|} \ge\frac{1}{1+\kappa_L(n)}. UnifInt C Π = ∣ G C ∣ + ∣ B C ∣ ∣ G C ∣ ≥ 1 + κ L ( n ) 1 .
As there are only ∣ At σ ∣ |\operatorname{At}_\sigma| ∣ At σ ∣ many cells Cell σ , ∅ \operatorname{Cell}_{\sigma,\varnothing} Cell σ , ∅ , and we can choose as m 0 m_0 m 0 the maximum over all m 0 m_0 m 0 supplied by Lemma 40 for each cell, the projective cell reduction principle (Theorem 34) implies
UnifInt n , m Π ≥ ( 1 1 + κ L ( n ) ) ∣ At σ ∣ > 0. \operatorname{UnifInt}^{\Pi}_{n,m} \ge\left(\frac{1}{1+\kappa_L(n)}\right)^{|\operatorname{At}_\sigma|} > 0. UnifInt n , m Π ≥ ( 1 + κ L ( n ) 1 ) ∣ At σ ∣ > 0.
We close this section by observing that, for every tabular modal logic L L L and fixed signature σ \sigma σ , the set ⋃ ρ Q σ , ρ \bigcup_\rho Q_{\sigma,\rho} ⋃ ρ Q σ , ρ is finite, regardless of the chosen class of pointed frames K \mathcal{K} K with L = Log ( K ) L = \operatorname{Log}(\mathcal{K}) L = Log ( K ) .
Lemma 43. Let L = Log ( K ) L = \operatorname{Log}(\mathcal{K}) L = Log ( K ) be a tabular quasi-normal modal logic. Then, for every n ≥ 0 n \ge0 n ≥ 0 and every σ n ↑ ρ m \sigma^{\uparrow\rho_m}_n σ n ↑ ρ m -cell C C C , ∣ Q C ∣ ≤ 2 F L |Q_C| \le2^{F_L} ∣ Q C ∣ ≤ 2 F L and so ∣ Q σ n , ρ m ∣ ≤ 2 F L ∣ At σ ∣ |Q_{\sigma_n,\rho_m}| \le2^{F_L|\operatorname{At}_\sigma|} ∣ Q σ n , ρ m ∣ ≤ 2 F L ∣ At σ ∣ for all m ≥ 0 m \ge0 m ≥ 0 .
Proof. Fix a σ n ↑ ρ m \sigma^{\uparrow\rho_m}_n σ n ↑ ρ m -cell C C C . Assume first that we define logical equivalence of existentially quantified formulas using the class K L \mathcal{K}_L K L of all rooted frames validating L L L . Hence L L L is given in the same way as in Theorem 32. Then every x ∈ Q C x \in Q_C x ∈ Q C is determined by the set of σ n \sigma_n σ n -shadows sh ( C ) \operatorname{sh}(C) sh ( C ) in the sense that any two different x , y ∈ Q C x,y \in Q_C x , y ∈ Q C have different models in sh ( C ) \operatorname{sh}(C) sh ( C ) . It follows from the proof of Theorem 32 that ∣ Q C ∣ ≤ 2 F L |Q_C| \le2^{F_L} ∣ Q C ∣ ≤ 2 F L with F L F_L F L a constant depending only on L L L . The bound on Q σ n , ρ m Q_{\sigma_n,\rho_m} Q σ n , ρ m follows directly from the cell reduction principle.
Now consider an arbitrary class K \mathcal{K} K with L = Log ( K ) L = \operatorname{Log}(\mathcal{K}) L = Log ( K ) . Since K ⊆ K L \mathcal{K} \subseteq\mathcal{K}_L K ⊆ K L , logical equivalence over K L \mathcal{K}_L K L implies logical equivalence over K \mathcal{K} K . Consequently, the number of equivalence classes in Q C Q_C Q C , defined using equivalence over K \mathcal{K} K , is also bounded by 2 F L 2^{F_L} 2 F L .
Example 11. We analyse UnifInt σ , ρ Π \operatorname{UnifInt}^{\Pi}_{\sigma,\rho} UnifInt σ , ρ Π for the 3-fork. Assume ρ ≠ ∅ \rho\ne\varnothing ρ = ∅ (otherwise, it is trivial). As we have seen in Example 4, only cells of the form C = C s 0 , { s 1 , s 2 } ρ C = C^{\rho}_{s_0,\{s_1,s_2\}} C = C s 0 , { s 1 , s 2 } ρ contain counterexamples to uniform σ \sigma σ -interpolation. After existential quantification, we obtain the same five non-logically equivalent formulas in Q C Q_C Q C :
φ 11 = ∃ r ⃗ 0 r ⃗ 1 r ⃗ 2 . Φ ( M s 0 r 0 , { s 1 r 1 , s 2 r 2 } ) \varphi_{11} = \exists\vec{r}_0\vec{r}_1\vec{r}_2.\ \Phi(\mathcal{M}_{s_0r_0,\{s_1r_1,s_2r_2\}}) φ 11 = ∃ r 0 r 1 r 2 . Φ ( M s 0 r 0 , { s 1 r 1 , s 2 r 2 } ) φ 12 = ∃ r ⃗ 0 r ⃗ 1 r ⃗ 2 ′ . Φ ( M s 0 r 0 , s 1 r 1 ′ , { s 2 r 2 } ) \varphi_{12} = \exists\vec{r}_0\vec{r}_1\vec{r}_2'.\ \Phi(\mathcal{M}_{s_0r_0,s_1r_1',\{s_2r_2\}}) φ 12 = ∃ r 0 r 1 r 2 ′ . Φ ( M s 0 r 0 , s 1 r 1 ′ , { s 2 r 2 } ) φ 21 = ∃ r ⃗ 0 r ⃗ 1 r ⃗ 2 ′ . Φ ( M s 0 r 0 , s 1 { r 1 , r 2 } , s 2 r 2 ′ ) \varphi_{21} = \exists\vec{r}_0\vec{r}_1\vec{r}_2'.\ \Phi(\mathcal{M}_{s_0r_0,s_1\{r_1,r_2\},s_2r_2'}) φ 21 = ∃ r 0 r 1 r 2 ′ . Φ ( M s 0 r 0 , s 1 { r 1 , r 2 } , s 2 r 2 ′ ) as well as ⊥ \bot ⊥ and φ 12 ∨ φ 21 \varphi_{12} \vee\varphi_{21} φ 12 ∨ φ 21 , where r ⃗ \vec{r} r is the list of variables in an ρ \rho ρ -colour r r r . Then UnifInt n , m Π = 0.6 \operatorname{UnifInt}^{\Pi}_{n,m} = 0.6 UnifInt n , m Π = 0.6 since exactly φ 12 \varphi_{12} φ 12 and φ 21 \varphi_{21} φ 21 lack uniform σ \sigma σ -interpolants. Thus, we obtain
UnifInt n , m Π = ( 3 5 ) 1 2 ( 2 n − 1 ) 2 2 n . \operatorname{UnifInt}^{\Pi}_{n,m} = \binom{3}{5}^{\frac{1}{2}(2^n-1)2^{2n}}. UnifInt n , m Π = ( 5 3 ) 2 1 ( 2 n − 1 ) 2 2 n .
Case Study: Bounded-Depth Extensions of Grz
We consider the bounded-depth extensions G r z [ Δ ] \mathrm{Grz}_{[\Delta]} Grz [ Δ ] of G r z \mathrm{Grz} Grz introduced in the preliminaries, where Δ ≥ 1 \Delta\ge1 Δ ≥ 1 . They are all locally tabular and have CIP and uniform interpolation if and only if Δ ≤ 2 \Delta\le2 Δ ≤ 2 [13 Chapter 8] . We prove BICIP for every Δ ≥ 1 \Delta\ge1 Δ ≥ 1 . By Theorems 15 and 29 , it follows that lim m → ∞ U n i f I n t n , m = lim m → ∞ D C r a i g n , m = 1 \lim_{m \to\infty}\mathrm{UnifInt}_{n,m} = \lim_{m \to\infty}\mathrm{DCraig}_{n,m} = 1 lim m → ∞ UnifInt n , m = lim m → ∞ DCraig n , m = 1 for every n ≥ 0 n \ge0 n ≥ 0 and Δ ≥ 1 \Delta\ge1 Δ ≥ 1 . We also show that G r z [ Δ ] \mathrm{Grz}_{[\Delta]} Grz [ Δ ] has uniform BICIP exactly when Δ ≤ 3 \Delta\le3 Δ ≤ 3 , with inc ( C ) ≤ 3 \operatorname{inc}(C) \le3 inc ( C ) ≤ 3 for Δ = 3 \Delta= 3 Δ = 3 and, by CIP, inc ( C ) = 1 \operatorname{inc}(C) = 1 inc ( C ) = 1 for Δ ∈ { 1 , 2 } \Delta\in\{1,2\} Δ ∈ { 1 , 2 } .
After existential quantification, Theorems 42 and 39 give positive lower bounds for U n i f I n t n , m Π \mathrm{UnifInt}^{\Pi}_{n,m} UnifInt n , m Π and D C r a i g n , m Π \mathrm{DCraig}^{\Pi}_{n,m} DCraig n , m Π , respectively, again for every fixed n ≥ 0 n \ge0 n ≥ 0 and Δ ≥ 1 \Delta\ge1 Δ ≥ 1 , for any class K \mathcal{K} K with G r z [ Δ ] = Log ( K ) \mathrm{Grz}_{[\Delta]} = \operatorname{Log}(\mathcal{K}) Grz [ Δ ] = Log ( K ) . We further prove in this section lim m → ∞ U n i f I n t n , m Π = 1 \lim_{m \to\infty}\mathrm{UnifInt}^{\Pi}_{n,m} = 1 lim m → ∞ UnifInt n , m Π = 1 for any class K \mathcal{K} K with G r z [ Δ ] = Log ( K ) \mathrm{Grz}_{[\Delta]} = \operatorname{Log}(\mathcal{K}) Grz [ Δ ] = Log ( K ) containing all finite tree-shaped frames of depth bounded by Δ \Delta Δ . Whether lim m → ∞ D C r a i g n , m Π = 1 \lim_{m \to\infty}\mathrm{DCraig}^{\Pi}_{n,m} = 1 lim m → ∞ DCraig n , m Π = 1 on such K \mathcal{K} K remains open.
Our main tool in this section is the shadow profile of cells introduced in Section 6.2 and the characterisation of compatibility of atoms via covers.
For a finite poset F = ( W , R ) \mathfrak{F} = (W,R) F = ( W , R ) , put R + = R ∖ { ( w , w ) ∣ w ∈ W } R^{+} = R \setminus\{(w,w) \mid w \in W\} R + = R ∖ {( w , w ) ∣ w ∈ W } . Its depth dpt ( F ) \operatorname{dpt}(\mathfrak{F}) dpt ( F ) is the maximal number of worlds in an R + R^{+} R + -chain; thus a singleton has depth 1. Write ↑ u = { x ∈ W ∣ u R x } \mathord{\uparrow}u = \{x \in W \mid uRx\} ↑ u = { x ∈ W ∣ u R x } . A world w w w is a root if W = ↑ w W = \mathord{\uparrow}w W = ↑ w . As before, distinguished worlds are assumed to be roots. A finite rooted pointed τ \tau τ -model on a poset is called a τ \tau τ -poset.
To establish BICIP, we seek a bound inc ( C ) = ∣ C / ∼ C ∣ ≤ f L ( ∣ σ ∣ ) \operatorname{inc}(C) = |C/{\sim_C}| \le f_L(|\sigma|) inc ( C ) = ∣ C / ∼ C ∣ ≤ f L ( ∣ σ ∣ ) for any σ τ ρ \sigma^{\tau^\rho} σ τ ρ -cell C C C , independent of ∣ ρ ∣ |\rho| ∣ ρ ∣ . For G r z [ Δ ] \mathrm{Grz}_{[\Delta]} Grz [ Δ ] , the shadows in sh ( H C ) \operatorname{sh}(H_C) sh ( H C ) are σ \sigma σ -posets, and Theorem 30 gives
p ⌣ s h p ′ ⟺ { p , p ′ } has no G r z [ Δ ] -cover . p \mathrel{\mathop{\smile}\limits^{\mathrm{sh}}} p' \iff\{p,p'\}\text{ has no }\mathrm{Grz}_{[\Delta]}\text{-cover}. p ⌣ sh p ′ ⟺ { p , p ′ } has no Grz [ Δ ] -cover .
Let P = C / ∼ C P = C/{\sim_C} P = C / ∼ C . The canonical profile map π C : C → P \pi_C : C \to P π C : C → P factors through the strong graph homomorphism κ : sh ( H C ) → P \kappa: \operatorname{sh}(H_C) \to P κ : sh ( H C ) → P :
π C = κ ∘ π s h . \pi_C = \kappa\circ\pi_{\mathrm{sh}}. π C = κ ∘ π sh .
We say that a shadow p p p represents q ∈ P q \in P q ∈ P iff κ ( p ) = q \kappa(p) = q κ ( p ) = q .
A σ \sigma σ -poset ( M , w ) (\mathcal{M},w) ( M , w ) is tree-shaped if its order is the reflexive transitive closure of a finite rooted tree with root w w w . Every non-root world then has a unique immediate predecessor, its parent ; its immediate successors are its children . A tree-shaped σ \sigma σ -poset is normalized if no two distinct children u 1 , u 2 u_1,u_2 u 1 , u 2 of a common parent have isomorphic rooted σ \sigma σ -submodels, that is, there is no σ \sigma σ -isomorphism I : ↑ u 1 → ↑ u 2 I : \mathord{\uparrow}u_1 \to\mathord{\uparrow}u_2 I : ↑ u 1 → ↑ u 2 with I ( u 1 ) = u 2 I(u_1) = u_2 I ( u 1 ) = u 2 .
To count normalized σ \sigma σ -posets up to isomorphism, put N = 2 n N = 2^n N = 2 n , and define
B 1 ( n ) = N , B k + 1 ( n ) = N ⋅ 2 B k ( n ) . B_1(n) = N,\qquad B_{k+1}(n) = N \cdot2^{B_k(n)}. B 1 ( n ) = N , B k + 1 ( n ) = N ⋅ 2 B k ( n ) .
Theorem 44. For Δ ≥ 1 \Delta\ge1 Δ ≥ 1 ,
the number of normalized σ \sigma σ -posets of depth bounded by Δ \Delta Δ is, up to isomorphism, bounded by B Δ ( ∣ σ ∣ ) B_\Delta(|\sigma|) B Δ ( ∣ σ ∣ ) ;
for any signature ρ \rho ρ disjoint from σ \sigma σ and every σ τ ρ \sigma^{\tau^\rho} σ τ ρ -cell C C C in G r z [ Δ ] \mathrm{Grz}_{[\Delta]} Grz [ Δ ] : for every p ∈ sh ( C ) p \in\operatorname{sh}(C) p ∈ sh ( C ) , there is a normalized σ \sigma σ -poset q q q of depth ≤ Δ \le\Delta ≤ Δ such that q ≡ σ p q \equiv_\sigma p q ≡ σ p , and, for every r ∈ sh ( C ) r \in\operatorname{sh}(C) r ∈ sh ( C ) , the pair { p , r } \{p,r\} { p , r } has a G r z [ Δ ] \mathrm{Grz}_{[\Delta]} Grz [ Δ ] -cover iff the pair { q , r } \{q,r\} { q , r } has a G r z [ Δ ] \mathrm{Grz}_{[\Delta]} Grz [ Δ ] -cover.
It follows that inc ( C ) \operatorname{inc}(C) inc ( C ) is bounded by B Δ ( ∣ σ ∣ ) B_\Delta(|\sigma|) B Δ ( ∣ σ ∣ ) for any σ τ ρ \sigma^{\tau^\rho} σ τ ρ -cell C C C . To show this, take for every p ∈ sh ( C ) p \in\operatorname{sh}(C) p ∈ sh ( C ) a normalized norm ( p ) \operatorname{norm}(p) norm ( p ) satisfying Point 2. If κ ( p 1 ) ≠ κ ( p 2 ) \kappa(p_1) \ne\kappa(p_2) κ ( p 1 ) = κ ( p 2 ) , then there is r ∈ sh ( C ) r \in\operatorname{sh}(C) r ∈ sh ( C ) such that { r , p 1 } \{r,p_1\} { r , p 1 } has a cover but { r , p 2 } \{r,p_2\} { r , p 2 } does not, or vice versa. Hence { norm ( p 1 ) , r } \{\operatorname{norm}(p_1),r\} { norm ( p 1 ) , r } has a cover but { norm ( p 2 ) , r } \{\operatorname{norm}(p_2),r\} { norm ( p 2 ) , r } does not, or vice versa, and therefore norm ( p 1 ) ≇ norm ( p 2 ) \operatorname{norm}(p_1) \not\cong\operatorname{norm}(p_2) norm ( p 1 ) ≅ norm ( p 2 ) . Thus
inc ( C ) = ∣ κ [ sh ( H C ) ] ∣ ≤ ∣ { [ norm ( p ) ] ≅ : p ∈ sh ( C ) } ∣ ≤ B Δ ( ∣ σ ∣ ) . \operatorname{inc}(C) = |\kappa[\operatorname{sh}(H_C)]| \le\bigl|\{[\operatorname{norm}(p)]_{\cong} : p \in\operatorname{sh}(C)\}\bigr| \le B_\Delta(|\sigma|). inc ( C ) = ∣ κ [ sh ( H C )] ∣ ≤ {[ norm ( p ) ] ≅ : p ∈ sh ( C )} ≤ B Δ ( ∣ σ ∣ ) .
We obtain:
Theorem 45. For every Δ ≥ 1 \Delta\ge1 Δ ≥ 1 , G r z [ Δ ] \mathrm{Grz}_{[\Delta]} Grz [ Δ ] has BICIP: for every σ τ ρ \sigma^{\tau^\rho} σ τ ρ -cell C C C , we have inc ( C ) ≤ B Δ ( ∣ σ ∣ ) \operatorname{inc}(C) \le B_\Delta(|\sigma|) inc ( C ) ≤ B Δ ( ∣ σ ∣ ) .
The proof of Theorem 44 is in two steps. First, we associate with every σ \sigma σ -poset its tree-shaped unravelling [3 Chapter 3] . Let ( M , w ) (\mathcal{M},w) ( M , w ) with M = ( W , R , V ) \mathcal{M} = (W,R,V) M = ( W , R , V ) be a σ \sigma σ -poset. A branch is a chain γ = \gamma= γ = ( x 0 R + x 1 R + … R + x k ) (x_0R^{+}x_1R^{+}\ldots R^{+}x_k) ( x 0 R + x 1 R + … R + x k ) starting at the root w = x 0 w=x_0 w = x 0 . Let end M ( γ ) \operatorname{end}_{\mathcal{M}}(\gamma) end M ( γ ) map γ \gamma γ to its endpoint x k x_k x k . The unravelling U ( M , w ) = ( ( W U , R U , V U ) , ( w ) ) U(\mathcal{M},w)=((W^U,R^U,V^U),(w)) U ( M , w ) = (( W U , R U , V U ) , ( w )) is the σ \sigma σ -model whose worlds W U W^U W U are the branches of ( M , w ) (\mathcal{M},w) ( M , w ) and
γ R U γ ′ if γ is a prefix of γ ′ , γ ∈ V U ( p ) if end M ( γ ) ∈ V ( p ) , \gamma R^U\gamma' \quad\text{if }\gamma\text{ is a prefix of }\gamma',
\qquad
\gamma\in V^U(p)\text{ if }\operatorname{end}_{\mathcal{M}}(\gamma)\in V(p), γ R U γ ′ if γ is a prefix of γ ′ , γ ∈ V U ( p ) if end M ( γ ) ∈ V ( p ) ,
and with distinguished world the length-0 branch ( w ) (w) ( w ) . Then R U R^U R U is a partial order, and since every branch begins with w w w , every world is R U R^U R U -reachable from ( w ) (w) ( w ) , so U ( M , w ) U(\mathcal{M},w) U ( M , w ) is a rooted pointed σ \sigma σ -poset.
Lemma 46 (unravelling). Let ( M , w ) (\mathcal{M},w) ( M , w ) be a σ \sigma σ -poset. Then dpt ( U ( M , w ) ) = dpt ( M ) \operatorname{dpt}(U(\mathcal{M},w))=\operatorname{dpt}(\mathcal{M}) dpt ( U ( M , w )) = dpt ( M ) , end M \operatorname{end}_{\mathcal{M}} end M is a surjective σ \sigma σ -p p p -morphism, and, for every σ \sigma σ -poset ( N , v ) (\mathcal{N},v) ( N , v ) , the pair { ( M , w ) , ( N , v ) } \{(\mathcal{M},w),(\mathcal{N},v)\} {( M , w ) , ( N , v )} has a Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] -cover iff the pair { U ( M , w ) , ( N , v ) } \{U(\mathcal{M},w),(\mathcal{N},v)\} { U ( M , w ) , ( N , v )} has a Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] -cover.
Proof. That end M \operatorname{end}_{\mathcal{M}} end M is a surjective p p p -morphism and dpt ( U ( M , w ) ) = dpt ( M ) \operatorname{dpt}(U(\mathcal{M},w))=\operatorname{dpt}(\mathcal{M}) dpt ( U ( M , w )) = dpt ( M ) is immediate from the definition. If ( K , r ) (\mathcal{K},r) ( K , r ) covers U ( M , w ) , ( N , v ) U(\mathcal{M},w),(\mathcal{N},v) U ( M , w ) , ( N , v ) then composing a witness p p p -morphism with end M \operatorname{end}_{\mathcal{M}} end M show that it covers ( M , w ) , ( N , v ) (\mathcal{M},w),(\mathcal{N},v) ( M , w ) , ( N , v ) . Conversely, let ( K , r ) (\mathcal{K},r) ( K , r ) cover ( M , w ) , ( N , v ) (\mathcal{M},w),(\mathcal{N},v) ( M , w ) , ( N , v ) via f , g f,g f , g . Consider U ( K , r ) U(\mathcal{K},r) U ( K , r ) with end K : U ( K , r ) → ( K , r ) \operatorname{end}_{\mathcal{K}}:U(\mathcal{K},r)\to(\mathcal{K},r) end K : U ( K , r ) → ( K , r ) ; then dpt ( U ( K , r ) ) = dpt ( K ) ≤ Δ \operatorname{dpt}(U(\mathcal{K},r))=\operatorname{dpt}(\mathcal{K})\leq\Delta dpt ( U ( K , r )) = dpt ( K ) ≤ Δ and g ∘ end K : U ( K , r ) → ( N , v ) g\circ\operatorname{end}_{\mathcal{K}}:U(\mathcal{K},r)\to(\mathcal{N},v) g ∘ end K : U ( K , r ) → ( N , v ) . Define q : U ( K , r ) → U ( M , w ) q:U(\mathcal{K},r)\to U(\mathcal{M},w) q : U ( K , r ) → U ( M , w ) by sending a branch γ \gamma γ to the branch obtained from its f f f -image by deleting consecutive repetitions. Since f f f is monotone and M \mathcal{M} M is a poset, an image cannot recur after a different image. Thus, q ( γ ) q(\gamma) q ( γ ) is a branch and
end M ( q ( γ ) ) = f ( end K ( γ ) ) . \operatorname{end}_{\mathcal{M}}(q(\gamma))=f(\operatorname{end}_{\mathcal{K}}(\gamma)). end M ( q ( γ )) = f ( end K ( γ )) .
It preserves σ \sigma σ -colours since f f f does, and is monotone since deleting repetitions preserves the prefix order. For (p3), let γ \gamma γ end at t 0 t_0 t 0 and let δ \delta δ extend q ( γ ) q(\gamma) q ( γ ) by worlds x 1 R + … R + x j x_1R^{+}\ldots R^{+}x_j x 1 R + … R + x j ; applying (p3) for f f f successively yields t 0 R + t 1 R + ⋯ R + t j t_0R^{+}t_1R^{+}\cdots R^{+}t_j t 0 R + t 1 R + ⋯ R + t j in K \mathcal{K} K with f ( t i ) = x i f(t_i)=x_i f ( t i ) = x i , and γ \gamma γ extended by t 1 , … , t j t_1,\ldots,t_j t 1 , … , t j maps to δ \delta δ . So q q q is a p p p -morphism and U ( K , r ) U(\mathcal{K},r) U ( K , r ) covers U ( M , w ) , ( N , v ) U(\mathcal{M},w),(\mathcal{N},v) U ( M , w ) , ( N , v ) .
Let ( M , w ) (\mathcal{M},w) ( M , w ) be tree-shaped. Suppose two distinct children u 1 , u 2 u_1,u_2 u 1 , u 2 of a common parent z z z have isomorphic subtrees, witnessed by an isomorphism I : ↑ u 1 → ↑ u 2 I:\mathord{\uparrow}u_1\to\mathord{\uparrow}u_2 I : ↑ u 1 → ↑ u 2 of σ \sigma σ -models with I ( u 1 ) = u 2 I(u_1)=u_2 I ( u 1 ) = u 2 . The sibling merge of u 2 u_2 u 2 into u 1 u_1 u 1 is the σ \sigma σ -model ( M − , w ) (\mathcal{M}^{-},w) ( M − , w ) with worlds W ∖ ↑ u 2 W\setminus\mathord{\uparrow}u_2 W ∖ ↑ u 2 and R R R and V V V restricted. As ↑ u 2 \mathord{\uparrow}u_2 ↑ u 2 is upward closed and omits the root, ( M − , w ) (\mathcal{M}^{-},w) ( M − , w ) is again a tree-shaped, σ \sigma σ -poset with dpt ( M − ) ≤ dpt ( M ) \operatorname{dpt}(\mathcal{M}^{-})\leq\operatorname{dpt}(\mathcal{M}) dpt ( M − ) ≤ dpt ( M ) .
Lemma 47 (sibling merge). Let ( M − , w ) (\mathcal{M}^{-},w) ( M − , w ) be the sibling merge of u 2 u_2 u 2 into u 1 u_1 u 1 in a tree-shaped σ \sigma σ -poset ( M , w ) (\mathcal{M},w) ( M , w ) . Then, for every σ \sigma σ -poset ( N , v ) (\mathcal{N},v) ( N , v ) , the pair { ( M , w ) , ( N , v ) } \{(\mathcal{M},w),(\mathcal{N},v)\} {( M , w ) , ( N , v )} has a Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] -cover iff the pair { ( M − , w ) , ( N , v ) } \{(\mathcal{M}^{-},w),(\mathcal{N},v)\} {( M − , w ) , ( N , v )} has a Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] -cover.
Proof. Retain z , u 1 , u 2 z,u_1,u_2 z , u 1 , u 2 and the isomorphism I : ↑ u 1 → ↑ u 2 I:\mathord{\uparrow}u_1\to\mathord{\uparrow}u_2 I : ↑ u 1 → ↑ u 2 . We write ≤ \leq ≤ for the orders in the target models.
( ⇒ ) (\Rightarrow) ( ⇒ ) Suppose { ( M , w ) , ( N , v ) } \{(\mathcal{M},w),(\mathcal{N},v)\} {( M , w ) , ( N , v )} has a cover ( K , u ) (\mathcal{K},u) ( K , u ) via f , g f,g f , g . Define e : ( K , u ) → ( M − , w ) e:(\mathcal{K},u)\to(\mathcal{M}^{-},w) e : ( K , u ) → ( M − , w ) by
e ( x ) = { f ( x ) , f ( x ) ∉ ↑ u 2 , I − 1 ( f ( x ) ) , f ( x ) ∈ ↑ u 2 . e(x)=
\begin{cases}
f(x), & f(x)\notin\mathord{\uparrow}u_2,\\
I^{-1}(f(x)), & f(x)\in\mathord{\uparrow}u_2.
\end{cases} e ( x ) = { f ( x ) , I − 1 ( f ( x )) , f ( x ) ∈ / ↑ u 2 , f ( x ) ∈ ↑ u 2 .
Condition (p1) follows from f f f and I I I . For (p2), an edge entering ↑ u 2 \mathord{\uparrow}u_2 ↑ u 2 from outside starts at or below z z z , and therefore also below u 1 u_1 u 1 . For (p3), use I − 1 I^{-1} I − 1 inside ↑ u 2 \mathord{\uparrow}u_2 ↑ u 2 and the original back witnesses elsewhere. Thus e , g e,g e , g witness a cover of { ( M − , w ) , ( N , v ) } \{(\mathcal{M}^{-},w),(\mathcal{N},v)\} {( M − , w ) , ( N , v )} .
( ⇐ ) (\Leftarrow) ( ⇐ ) Suppose { ( M − , w ) , ( N , v ) } \{(\mathcal{M}^{-},w),(\mathcal{N},v)\} {( M − , w ) , ( N , v )} has a finite cover ( K − , u ) (\mathcal{K}^{-},u) ( K − , u ) via f , g f,g f , g , where K − = ( ( W − , R − ) , V − ) \mathcal{K}^{-}=((W^{-},R^{-}),V^{-}) K − = (( W − , R − ) , V − ) . Put X = f − 1 ( ↑ u 1 ) X=f^{-1}(\mathord{\uparrow}u_1) X = f − 1 ( ↑ u 1 ) and let X ′ = { x ′ : x ∈ X } X'=\{x':x\in X\} X ′ = { x ′ : x ∈ X } be disjoint from W − W^{-} W − . Define K = ( ( W , R ) , V ) \mathcal{K}=((W,R),V) K = (( W , R ) , V ) by
W = W − ⊔ X ′ , W=W^{-}\mathbin{\sqcup}X', W = W − ⊔ X ′ ,
R = R − ∪ { ( x ′ , y ′ ) ∣ ( x , y ) ∈ R − ∩ ( X × X ) } R=R^{-}\mathbin{\cup}\{(x',y')\mid(x,y)\in R^{-}\cap(X\times X)\} R = R − ∪ {( x ′ , y ′ ) ∣ ( x , y ) ∈ R − ∩ ( X × X )}
∪ { ( x , y ′ ) ∣ ( x , y ) ∈ R − ∩ ( ( W − ∖ X ) × X ) } , \mathbin{\cup}\{(x,y')\mid(x,y)\in R^{-}\cap((W^{-}\setminus X)\times X)\}, ∪ {( x , y ′ ) ∣ ( x , y ) ∈ R − ∩ (( W − ∖ X ) × X )} ,
V ( p ) = V − ( p ) ∪ { x ′ ∣ x ∈ V − ( p ) ∩ X } ( p ∈ σ ) . V(p)=V^{-}(p)\mathbin{\cup}\{x'\mid x\in V^{-}(p)\cap X\}
\qquad(p\in\sigma). V ( p ) = V − ( p ) ∪ { x ′ ∣ x ∈ V − ( p ) ∩ X } ( p ∈ σ ) .
Since X X X is upward closed, R R R is reflexive, antisymmetric and transitive. There are no edges between X X X and X ′ X' X ′ in either direction. Also u ∉ X u\notin X u ∈ / X , and u u u reaches every world of X ′ X' X ′ .
Let c : K → K − c:\mathcal{K}\to\mathcal{K}^{-} c : K → K − be given by c ( x ) = x c(x)=x c ( x ) = x for x ∈ W − x\in W^{-} x ∈ W − and c ( x ′ ) = x c(x')=x c ( x ′ ) = x for x ∈ X x\in X x ∈ X . Conditions (p1) and (p2) are immediate. For (p3), old worlds retain their old back witnesses; if x ′ ∈ X ′ x'\in X' x ′ ∈ X ′ and x R − y xR^{-}y x R − y , then y ∈ X y\in X y ∈ X and x ′ R ′ y ′ x'R'y' x ′ R ′ y ′ . Thus c c c is a σ \sigma σ -p p p -morphism. Moreover, no chain meets both X X X and X ′ X' X ′ , so c c c is injective on chains. Hence
dpt ( K ) = dpt ( K − ) ≤ Δ . \operatorname{dpt}(\mathcal{K})=\operatorname{dpt}(\mathcal{K}^{-})\leq\Delta. dpt ( K ) = dpt ( K − ) ≤ Δ.
Define
f ′ ( x ) = { f ( x ) , x ∈ W − , I ( f ( y ) ) , x = y ′ ∈ X ′ , g ′ = g ∘ c . f'(x)=
\begin{cases}
f(x), & x\in W^{-},\\
I(f(y)), & x=y'\in X',
\end{cases}
\qquad g'=g\circ c. f ′ ( x ) = { f ( x ) , I ( f ( y )) , x ∈ W − , x = y ′ ∈ X ′ , g ′ = g ∘ c .
Then g ′ g' g ′ is a σ \sigma σ -p-morphism.
For f ′ f' f ′ , condition (p1) follows from f f f and I I I . For (p2), only mixed edges need checking. If x R y ′ xRy' x R y ′ with x ∈ W − ∖ X x\in W^{-}\setminus X x ∈ W − ∖ X and y ∈ X y\in X y ∈ X , then
f ( x ) ≤ f ( y ) , f ( x ) ∉ ↑ u 1 , f ( y ) ∈ ↑ u 1 . f(x)\leq f(y),\qquad f(x)\notin\uparrow u_{1},\qquad f(y)\in\uparrow u_{1}. f ( x ) ≤ f ( y ) , f ( x ) ∈ / ↑ u 1 , f ( y ) ∈↑ u 1 .
Since M \mathcal{M} M is tree-shaped, f ( x ) ≤ z ≤ I ( f ( y ) ) f(x)\leq z\leq I(f(y)) f ( x ) ≤ z ≤ I ( f ( y )) .
For (p3), let f ′ ( x ) ≤ a f'(x)\leq a f ′ ( x ) ≤ a . If x = y ′ ∈ X ′ x=y'\in X' x = y ′ ∈ X ′ , write a = I ( b ) a=I(b) a = I ( b ) with f ( y ) ≤ b f(y)\leq b f ( y ) ≤ b . Choose t t t with y R − t yR^{-}t y R − t and f ( t ) = b f(t)=b f ( t ) = b . Then t ∈ X t\in X t ∈ X , y ′ R ′ t ′ y'R't' y ′ R ′ t ′ and f ′ ( t ′ ) = a f'(t')=a f ′ ( t ′ ) = a . If x ∈ X x\in X x ∈ X , then a ∈ ↑ u 1 a\in\uparrow u_{1} a ∈↑ u 1 , so an old back witness for f f f suffices. If x ∈ W − ∖ X x\in W^{-}\setminus X x ∈ W − ∖ X and a ∈ M − a\in\mathcal{M}^{-} a ∈ M − , again use an old back witness. Finally, if x ∈ W − ∖ X x\in W^{-}\setminus X x ∈ W − ∖ X and a ∈ ↑ u 2 a\in\uparrow u_{2} a ∈↑ u 2 , then f ( x ) ≤ z f(x)\leq z f ( x ) ≤ z . Put b = I − 1 ( a ) b=I^{-1}(a) b = I − 1 ( a ) and choose y y y with x R − y xR^{-}y x R − y and f ( y ) = b f(y)=b f ( y ) = b . Then y ∈ X y\in X y ∈ X , x R y ′ xRy' x R y ′ and f ′ ( y ′ ) = a f'(y')=a f ′ ( y ′ ) = a . Thus f ′ f' f ′ is a σ \sigma σ -p-morphism, and f ′ , g ′ f',g' f ′ , g ′ witness the required Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] -cover.
Proof of Theorem 44. Point 1 follows by induction from the definition of the B k B_k B k : a normalized model is determined up to isomorphism by its root colour and the set of isomorphism types of its immediate successor submodels.
For Point 2, unravel p p p and repeatedly merge isomorphic sibling subtrees. Every merge removes worlds, so the construction terminates in a normalized model norm ( p ) \operatorname{norm}(p) norm ( p ) . Its depth is at most Δ \Delta Δ , and Lemmas 46 and 47 imply that, for every r ∈ sh ( H C ) r\in\operatorname{sh}(H_C) r ∈ sh ( H C ) ,
{ p , r } has a Grz [ Δ ] -cover ⟺ { norm ( p ) , r } has a Grz [ Δ ] -cover . \{p,r\}\text{ has a }\operatorname{Grz}_{[\Delta]}\text{-cover}
\Longleftrightarrow
\{\operatorname{norm}(p),r\}\text{ has a }\operatorname{Grz}_{[\Delta]}\text{-cover}. { p , r } has a Grz [ Δ ] -cover ⟺ { norm ( p ) , r } has a Grz [ Δ ] -cover .
The endpoint and folding maps are σ \sigma σ -p-morphisms, so norm ( p ) ≡ σ p \operatorname{norm}(p)\equiv_{\sigma}p norm ( p ) ≡ σ p .
We have shown the following consequence of Theorems 45 and 29.
Corollary 48 (Craig density for Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] ). For n ≥ 0 n\geq0 n ≥ 0 and Δ ≥ 1 \Delta\geq1 Δ ≥ 1 , Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] satisfies lim m → ∞ DCraig n , m = 1 \lim_{m\to\infty}\operatorname{DCraig}_{n,m}=1 lim m → ∞ DCraig n , m = 1 .
We next show that Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] has uniform BICIP iff Δ ≤ 3 \Delta\leq3 Δ ≤ 3 . For Δ ≤ 2 \Delta\leq2 Δ ≤ 2 , this follows from CIP. For Δ = 3 \Delta=3 Δ = 3 , we show that inc ( C ) ≤ 3 \operatorname{inc}(C)\leq3 inc ( C ) ≤ 3 for every cell C C C , with equality for a suitable cell. The result indicates that, for Δ = 3 \Delta=3 Δ = 3 , exactly two ‘formula types’ are responsible for the failure of CIP. We begin by reproving that Grz [ 2 ] \operatorname{Grz}_{[2]} Grz [ 2 ] enjoys CIP.
Lemma 49. If ( M , w ) ≡ σ ( N , v ) (\mathcal{M},w)\equiv_{\sigma}(\mathcal{N},v) ( M , w ) ≡ σ ( N , v ) and dpt ( M ) , dpt ( N ) ≤ 2 \operatorname{dpt}(\mathcal{M}),\operatorname{dpt}(\mathcal{N})\leq2 dpt ( M ) , dpt ( N ) ≤ 2 , then { ( M , w ) , ( N , v ) } \{(\mathcal{M},w),(\mathcal{N},v)\} {( M , w ) , ( N , v )} has a cover of depth at most 2.
Proof. It follows from ( M , w ) ≡ σ ( N , v ) (\mathcal{M},w)\equiv_{\sigma}(\mathcal{N},v) ( M , w ) ≡ σ ( N , v ) that the σ \sigma σ -colours of w w w and v v v coincide and that the sets of σ \sigma σ -colours of maximal worlds also coincide. Take a new root t t t of σ \sigma σ -colour of w w w and, above t t t , one maximal world ( x , y ) (x,y) ( x , y ) for every pair of maximal worlds with the same σ \sigma σ -colour. Define f ( t ) = w f(t)=w f ( t ) = w , g ( t ) = v g(t)=v g ( t ) = v , f ( x , y ) = x f(x,y)=x f ( x , y ) = x and g ( x , y ) = y g(x,y)=y g ( x , y ) = y . Every proper successor of either target root is maximal and occurs in such a pair. Thus f , g f,g f , g satisfy (p1)–(p3). The construction also applies if either target is a singleton.
We use the following construction. Let p = ( M , w ) p=(\mathcal{M},w) p = ( M , w ) and q = ( N , v ) q=(\mathcal{N},v) q = ( N , v ) have depth at most 3 and w w w and v v v the same σ \sigma σ -colour. Suppose that every proper successor in either model can be matched with a σ \sigma σ -equivalent world in the other model, with both generated models of depth at most 2. Take a depth-2 cover of each matched pair, using Lemma 49, and put a new root below their disjoint union. The component maps extend to σ \sigma σ -p-morphisms onto p , q p,q p , q ; the matched pairs supply (p3) at the new root. Thus { p , q } \{p,q\} { p , q } has a cover of depth at most 3.
Theorem 50. For Δ ≥ 1 \Delta\geq1 Δ ≥ 1 , Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] has uniform BICIP iff Δ ≤ 3 \Delta\leq3 Δ ≤ 3 . For Δ = 3 \Delta=3 Δ = 3 , the bound inc ( C ) ≤ 3 \operatorname{inc}(C)\leq3 inc ( C ) ≤ 3 is sharp.
Proof. For Δ = 1 \Delta= 1 Δ = 1 , sh ( C ) \operatorname{sh}(C) sh ( C ) has a single element, and so inc ( C ) = 1 \operatorname{inc}(C) = 1 inc ( C ) = 1 , for every cell C C C . At Δ = 2 \Delta= 2 Δ = 2 , Lemma 49 and Theorem 30 give the same conclusion. In particular, we reproved that these two logics have CIP.
Let Δ = 3 \Delta= 3 Δ = 3 and C ∈ Cell σ , ρ C \in\operatorname{Cell}_{\sigma,\rho} C ∈ Cell σ , ρ . Take the σ \sigma σ -atom A = ⋃ C A = \bigcup C A = ⋃ C and let ( S A , r ) (\mathcal{S}_A,r) ( S A , r ) be its standard model. By Lemma 3, every σ \sigma σ -shadow ( M , w ) (\mathcal{M},w) ( M , w ) of M ∈ C \mathcal{M} \in C M ∈ C has a σ \sigma σ -p-morphism π : ( M , w ) → ( S A , r ) \pi: (\mathcal{M},w) \to(\mathcal{S}_A,r) π : ( M , w ) → ( S A , r ) . Write c ( x ) c(x) c ( x ) for the σ \sigma σ -colour of a world x x x .
If dpt ( S A ) = 1 \operatorname{dpt}(\mathcal{S}_A) = 1 dpt ( S A ) = 1 , all worlds in all shadows have the same σ \sigma σ -colour. Hence inc ( C ) = 1 \operatorname{inc}(C) = 1 inc ( C ) = 1 . The same conclusion holds if dpt ( S A ) = 3 \operatorname{dpt}(\mathcal{S}_A) = 3 dpt ( S A ) = 3 : no non-root world can be mapped by π \pi π to r r r , since (p3) would then give a four-world chain. Thus equal images under the maps π \pi π to S A \mathcal{S}_A S A provide the required matches.
Suppose now dpt ( S A ) = 2 \operatorname{dpt}(\mathcal{S}_A) = 2 dpt ( S A ) = 2 . Its worlds are r , a 1 , … , a k r,a_1,\ldots,a_k r , a 1 , … , a k , where the a i a_i a i are maximal and c ( a 1 ) , … , c ( a k ) c(a_1),\ldots,c(a_k) c ( a 1 ) , … , c ( a k ) are pairwise distinct. For k = 3 k = 3 k = 3 , this is G 0 \mathcal{G}_0 G 0 in Figure 3 . A σ \sigma σ -shadow has a root stutter if π ( x ) = r \pi(x) = r π ( x ) = r for some non-root x x x , and a lower stutter if π ( x ) = a i \pi(x) = a_i π ( x ) = a i for some x x x with a proper successor. Figure 3 illustrates these in G 1 \mathcal{G}_1 G 1 and G 2 \mathcal{G}_2 G 2 , respectively.
Figure 3. A standard σ \sigma σ -model G 0 \mathfrak{G}_0 G 0 of depth 2 and root- and lower-stutter refinements. The pair { G 1 , G 2 } \{\mathfrak{G}_1,\mathfrak{G}_2\} { G 1 , G 2 } has no G r z [ 3 ] \mathrm{Grz}_{[3]} Grz [ 3 ] -cover.
The only incompatible pairs are those in which one shadow has a root stutter and the other has a lower stutter but no root stutter. Indeed, if both shadows have root stutters, or neither does, every proper world has a σ \sigma σ -equivalent proper match in the other shadow. If one shadow has neither stutter, its depth is at most 2, and its root can also be used as a match. The preceding construction therefore covers all these pairs.
For the remaining case, suppose a cover exists. Lift a non-root world mapping to r r r in the first shadow. Its image in the second shadow must be the root: the two cover maps induce the same image in the extensional model S A \mathcal{S}_A S A . Lifting a lower stutter from this world then gives a four-world chain, a contradiction.
Hence, there are at most three distinct neighbourhoods, corresponding to shadows with neither stutter, with a root stutter, and with a lower stutter but no root stutter. Thus inc ( C ) ≤ 3 \operatorname{inc}(C) \leq3 inc ( C ) ≤ 3 . The bound is attained by the three models in Figure 3 , choosing c ( r ) , c ( a 1 ) , c ( a 2 ) , c ( a 3 ) c(r),c(a_1),c(a_2),c(a_3) c ( r ) , c ( a 1 ) , c ( a 2 ) , c ( a 3 ) pairwise distinct. Assign distinct ρ \rho ρ -colours within each model, using one common signature ρ \rho ρ . The expansions are extensional standard models of atoms in one cell, and their shadows have distinct neighbourhoods.
For Δ = 4 \Delta= 4 Δ = 4 , let F ~ k \widetilde{\mathcal{F}}_k F k have branches r < a i < b i r < a_i < b_i r < a i < b i , 1 ≤ i ≤ k 1 \leq i \leq k 1 ≤ i ≤ k , with pairwise distinct σ \sigma σ -colours; take 2 ∣ σ ∣ ≥ 2 k + 1 2^{|\sigma|} \geq2k+1 2 ∣ σ ∣ ≥ 2 k + 1 . For X , Y ⊆ { 1 , … , k } X,Y \subseteq\{1,\ldots,k\} X , Y ⊆ { 1 , … , k } , let M X \mathcal{M}_X M X introduce an a i a_i a i -stutter iff i ∈ X i \in X i ∈ X , and let N Y \mathcal{N}_Y N Y introduce a b i b_i b i -stutter iff i ∈ Y i \in Y i ∈ Y . Figure 4 shows k = 3 k = 3 k = 3 , X = { 1 , 3 } X = \{1,3\} X = { 1 , 3 } and Y = { 2 , 3 } Y = \{2,3\} Y = { 2 , 3 } . Then
Figure 4. Depth-4 refinements of F ~ 3 \widetilde{\mathfrak{F}}_3 F 3 . The common index 3 forces a five-world chain in any cover of { M { 1 , 3 } , N { 2 , 3 } } \{\mathcal{M}_{\{1,3\}},\mathcal{N}_{\{2,3\}}\} { M { 1 , 3 } , N { 2 , 3 } } .
{ M X , N Y } has a Grz [ 4 ] -cover ⟺ X ∩ Y = ∅ . \{\mathcal{M}_X,\mathcal{N}_Y\}\text{ has a Grz}_{[4]}\text{-cover}\quad\Longleftrightarrow\quad X \cap Y = \varnothing. { M X , N Y } has a Grz [ 4 ] -cover ⟺ X ∩ Y = ∅ .
If X ∩ Y = ∅ X \cap Y = \varnothing X ∩ Y = ∅ , combine the stutters in one model; each branch has at most one stutter, so its depth is at most 4. Collapsing stutters gives the cover maps. If i ∈ X ∩ Y i \in X \cap Y i ∈ X ∩ Y , lifting the two a i a_i a i -worlds and then the two b i b_i b i -worlds forces a five-world chain. The distinct σ \sigma σ -colours prevent these lifts from moving to another branch.
Choose one ρ \rho ρ with 2 ∣ ρ ∣ ≥ 3 k + 1 2^{|\rho|} \geq3k+1 2 ∣ ρ ∣ ≥ 3 k + 1 and assign distinct ρ \rho ρ -colours within each witness model. Their expansions are extensional standard models of atoms in one cell C C C , since collapsing stutters gives σ \sigma σ -p-morphisms onto F ~ k \widetilde{\mathcal{F}}_k F k . If X ≠ X ′ X \neq X' X = X ′ and, say, i ∈ X ∖ X ′ i \in X \setminus X' i ∈ X ∖ X ′ , then N { i } \mathcal{N}_{\{i\}} N { i } is incompatible with M X \mathcal{M}_X M X but compatible with M X ′ \mathcal{M}_{X'} M X ′ . Thus the M X \mathcal{M}_X M X represent 2 k 2^k 2 k distinct profile nodes, and inc ( C ) ≥ 2 k \operatorname{inc}(C) \geq2^k inc ( C ) ≥ 2 k .
For Δ > 4 \Delta> 4 Δ > 4 , prepend a chain of Δ − 4 \Delta- 4 Δ − 4 worlds with distinct new σ \sigma σ -colours to all witness models. The disjoint-stutter covers then have depth at most Δ \Delta Δ , while overlapping stutters force depth at least Δ + 1 \Delta+ 1 Δ + 1 . The same construction therefore gives unbounded profile sizes for every fixed Δ ≥ 4. \Delta\geq4. Δ ≥ 4. □ \square □
Uniform Interpolation Density after Existential Quantification
We show that for any class K \mathcal{K} K of rooted frames determining G r z [ Δ ] \mathrm{Grz}_{[\Delta]} Grz [ Δ ] and containing all finite tree-shaped frames of depth at most Δ \Delta Δ , we have lim m → ∞ UnifInt n , m Π → 1 \lim_{m\to\infty}\operatorname{UnifInt}^{\Pi}_{n,m}\to1 lim m → ∞ UnifInt n , m Π → 1 , for any fixed n n n . 1 To this end, we first show a general sufficient condition for UnifInt n , m Π → 1 \operatorname{UnifInt}^{\Pi}_{n,m}\to1 UnifInt n , m Π → 1 for logics with BCIP. We use the notation from Section 7.2.
Lemma 51. Let L = Log ( K ) L = \operatorname{Log}(\mathcal{K}) L = Log ( K ) be a locally tabular quasi-normal modal logic with BCIP. Assume that for logical equivalence in K \mathcal{K} K for every n ≥ 0 n \geq0 n ≥ 0 there is a function L n ( m ) > 0 L_n(m) > 0 L n ( m ) > 0 with L n ( m ) → ∞ L_n(m) \to\infty L n ( m ) → ∞ as m → ∞ m \to\infty m → ∞ , such that the following holds for all sufficiently large m m m : for every σ n Π ρ \sigma^{\Pi\rho}_n σ n Π ρ -cell C C C and every q ∈ P q \in P q ∈ P with P P P the canonical profile of C C C and N P ( q ) ≠ ∅ N_P(q) \neq\varnothing N P ( q ) = ∅ ,
∣ Q C [ { q } ] ∣ ≥ L n ( m ) . (8) \lvert Q_C[\{q\}]\rvert\geq L_n(m). \tag*{(8)} ∣ Q C [{ q }]∣ ≥ L n ( m ) . (8)
Then lim m → ∞ UnifInt n , m Π = 1 \lim_{m\to\infty}\operatorname{UnifInt}^{\Pi}_{n,m}=1 lim m → ∞ UnifInt n , m Π = 1 for every n ≥ 0 n\ge0 n ≥ 0 .
Proof. Fix n ≥ 0 n\ge0 n ≥ 0 . For each C σ n ∈ Cell σ n , ∅ C_{\sigma_n}\in\operatorname{Cell}_{\sigma_n,\emptyset} C σ n ∈ Cell σ n , ∅ , take the stabilisation index supplied by Lemma 40, and let m 0 ( n ) m_0(n) m 0 ( n ) be their maximum. Fix m ≥ m 0 ( n ) m\ge m_0(n) m ≥ m 0 ( n ) sufficiently large for the hypothesis to hold. Also fix a σ n ↑ ρ m \sigma_n^{\uparrow\rho_m} σ n ↑ ρ m -cell C C C with canonical profile P P P .
Let S ⊆ P S\subseteq P S ⊆ P be non-empty with N P ( S ) ≠ ∅ N_P(S)\ne\emptyset N P ( S ) = ∅ . Choose q S ∈ N P ( S ) q_S\in N_P(S) q S ∈ N P ( S ) . Then ( S , { q S } ) (S,\{q_S\}) ( S , { q S }) is an ordered biclique of H P H_P H P . Lemma 38 gives an injection Q C [ S ] × Q C [ { q S } ] ⟶ Q C Q_C[S]\times Q_C[\{q_S\}]\longrightarrow Q_C Q C [ S ] × Q C [{ q S }] ⟶ Q C . Consequently,
∣ Q C [ S ] ∣ ≤ ∣ Q C ∣ ∣ Q C [ { q S } ] ∣ ≤ ∣ Q C ∣ L n ( m ) . (9) \lvert Q_C[S]\rvert\le\frac{\lvert Q_C\rvert}{\lvert Q_C[\{q_S\}]\rvert}\le\frac{\lvert Q_C\rvert}{L_n(m)}. \tag*{(9)} ∣ Q C [ S ]∣ ≤ ∣ Q C [{ q S }]∣ ∣ Q C ∣ ≤ L n ( m ) ∣ Q C ∣ . (9)
Let B C B_C B C be the set of x ∈ Q C x\in Q_C x ∈ Q C without a uniform σ n \sigma_n σ n -interpolant. By Lemma 40,
B C ⊆ ⋃ ∅ ≠ S ⊆ P N P ( S ) ≠ ∅ Q C [ S ] , B_C\subseteq\bigcup_{\substack{\emptyset\ne S\subseteq P\\N_P(S)\ne\emptyset}}Q_C[S], B C ⊆ ∅ = S ⊆ P N P ( S ) = ∅ ⋃ Q C [ S ] ,
and, with 2 ∣ P ∣ ≤ 2 f L ( n ) 2^{\lvert P\rvert}\le2^{f_L(n)} 2 ∣ P ∣ ≤ 2 f L ( n ) , we obtain from (7) the following:
∣ B C ∣ ≤ 2 f L ( n ) ∣ Q C ∣ L n ( m ) . \lvert B_C\rvert\le2^{f_L(n)}\frac{\lvert Q_C\rvert}{L_n(m)}. ∣ B C ∣ ≤ 2 f L ( n ) L n ( m ) ∣ Q C ∣ .
Dividing by ∣ Q C ∣ \lvert Q_C\rvert ∣ Q C ∣ shows
1 − UnifInt C Π ≤ 2 f L ( n ) L n ( m ) . 1-\operatorname{UnifInt}^{\Pi}_C\le\frac{2^{f_L(n)}}{L_n(m)}. 1 − UnifInt C Π ≤ L n ( m ) 2 f L ( n ) .
By the projective cell reduction principle (Theorem 34),
1 − UnifInt n , m Π = 1 − ∏ C ∈ Cell σ n , ρ m UnifInt C Π ≤ ∑ C ∈ Cell σ n , ρ m ( 1 − UnifInt C Π ) ≤ ∣ At σ n ∣ 2 f L ( n ) L n ( m ) ⟶ 0. 1-\operatorname{UnifInt}^{\Pi}_{n,m}
=
1-\prod_{C\in\operatorname{Cell}_{\sigma_n,\rho_m}}\operatorname{UnifInt}^{\Pi}_C
\le
\sum_{C\in\operatorname{Cell}_{\sigma_n,\rho_m}}\left(1-\operatorname{UnifInt}^{\Pi}_C\right)
\le
\lvert\operatorname{At}_{\sigma_n}\rvert\frac{2^{f_L(n)}}{L_n(m)}
\longrightarrow0. 1 − UnifInt n , m Π = 1 − C ∈ Cell σ n , ρ m ∏ UnifInt C Π ≤ C ∈ Cell σ n , ρ m ∑ ( 1 − UnifInt C Π ) ≤ ∣ At σ n ∣ L n ( m ) 2 f L ( n ) ⟶ 0.
We verify the conditions of Lemma 51 for any class K \mathcal{K} K of rooted frames determining Grz [ Δ ] \operatorname{Grz}[\Delta] Grz [ Δ ] and containing all finite tree-shaped frames of depth at most Δ \Delta Δ . We set f L ( n ) = B Δ ( n ) f_L(n)=B_\Delta(n) f L ( n ) = B Δ ( n ) , fix Δ ≥ 1 \Delta\ge1 Δ ≥ 1 , K \mathcal{K} K , n ≥ 0 n\ge0 n ≥ 0 , and let K n K_n K n bound the numbers of worlds in all normalized σ n \sigma_n σ n -posets of depth at most Δ \Delta Δ . Such a bound exists by Theorem 44.
Fix an arbitrary integer v > 0 v>0 v > 0 . It suffices to provide m v m_v m v such that, for all m ≥ m v m\ge m_v m ≥ m v , every σ n ↑ ρ m \sigma_n^{\uparrow\rho_m} σ n ↑ ρ m -cell C C C , and every q ∈ P q\in P q ∈ P with P P P the canonical profile of C C C and N P ( q ) ≠ ∅ N_P(q)\ne\emptyset N P ( q ) = ∅ it holds that ∣ Q C [ { q } ] ∣ > v \lvert Q_C[\{q\}]\rvert>v ∣ Q C [{ q }]∣ > v . Put
m v = ⌈ log 2 ( ( v + 2 ) K n ) ⌉ . m_v=\left\lceil\log_2((v+2)K_n)\right\rceil. m v = ⌈ log 2 (( v + 2 ) K n ) ⌉ .
Let m ≥ m v m\ge m_v m ≥ m v , let C C C be such a cell, and choose p = ( M ′ , w ) ∈ sh ( H C ) p=(\mathcal{M}',w)\in\operatorname{sh}(H_C) p = ( M ′ , w ) ∈ sh ( H C ) with κ ( p ) = q \kappa(p)=q κ ( p ) = q . By Theorem 44, there is a normalized σ n \sigma_n σ n -poset ( M , w ) (\mathcal{M},w) ( M , w ) with ( M , w ) ≡ σ n p (\mathcal{M},w)\equiv_{\sigma_n}p ( M , w ) ≡ σ n p such that, for every r ∈ sh ( C ) r\in\operatorname{sh}(C) r ∈ sh ( C ) ,
{ ( M , w ) , r } has a cover ⟺ { p , r } has a cover . \{(\mathcal{M},w),r\}\text{ has a cover}\iff\{p,r\}\text{ has a cover}. {( M , w ) , r } has a cover ⟺ { p , r } has a cover .
At this stage, ( M , w ) (\mathcal{M},w) ( M , w ) need not itself belong to sh ( C ) \operatorname{sh}(C) sh ( C ) .
Since N P ( q ) ≠ ∅ N_P(q)\ne\emptyset N P ( q ) = ∅ , M \mathcal{M} M has a non-root world u u u . Indeed, if M \mathcal{M} M were a singleton of colour a a a , every world in every shadow in C C C would have colour a a a . Each shadow would then map constantly onto M \mathcal{M} M and would itself cover its pair with M \mathcal{M} M , contradicting N P ( q ) ≠ ∅ N_P(q)\ne\emptyset N P ( q ) = ∅ .
Put U = ↑ u U=\mathord{\uparrow}u U = ↑ u and let M [ u ; k ] \mathcal{M}[u;k] M [ u ; k ] replace U U U by k k k disjoint isomorphic sibling copies. For 1 ≤ k ≤ v + 1 1\le k\le v+1 1 ≤ k ≤ v + 1 ,
∣ M [ u ; k ] ∣ = ∣ W ∖ U ∣ + k ∣ U ∣ ≤ ( v + 2 ) K n ≤ 2 m . \lvert\mathcal{M}[u;k]\rvert=\lvert W\setminus U\rvert+k\lvert U\rvert\le(v+2)K_n\le2^m. ∣ M [ u ; k ]∣ = ∣ W ∖ U ∣ + k ∣ U ∣ ≤ ( v + 2 ) K n ≤ 2 m .
Assign distinct ρ m \rho_m ρ m -colours to its worlds. The resulting model ( M [ u ; k ] , w ) (\mathcal{M}[u;k],w) ( M [ u ; k ] , w ) is extensional, hence the standard model of an atom A k A_k A k . The folding map gives
( M [ u ; k ] , w ) ≡ σ n ( M , w ) ≡ σ n p , (\mathcal{M}[u;k],w)\equiv_{\sigma_n}(\mathcal{M},w)\equiv_{\sigma_n}p, ( M [ u ; k ] , w ) ≡ σ n ( M , w ) ≡ σ n p ,
so A k ∈ C \mathcal{A}_k \in C A k ∈ C and sh ( A k ) = ( M [ u ; k ] , w ) \operatorname{sh}(\mathcal{A}_k)=(\mathcal{M}[u;k],w) sh ( A k ) = ( M [ u ; k ] , w ) . Repeated sibling merging reduces M [ u ; k ] \mathcal{M}[u;k] M [ u ; k ] to M \mathcal{M} M . Thus, for every r ∈ sh ( C ) r \in\operatorname{sh}(C) r ∈ sh ( C ) ,
{ sh ( A k ) , r } has a cover ⟺ { p , r } has a cover . \{\operatorname{sh}(\mathcal{A}_k),r\}\text{ has a cover}\quad\Longleftrightarrow\quad\{p,r\}\text{ has a cover}. { sh ( A k ) , r } has a cover ⟺ { p , r } has a cover .
Consequently, π C ( A k ) = κ ( sh ( A k ) ) = q \pi_C(\mathcal{A}_k)=\kappa(\operatorname{sh}(\mathcal{A}_k))=q π C ( A k ) = κ ( sh ( A k )) = q . Put f k = [ ∃ ρ ⃗ m . Φ ( { A k } ) ] ∈ Q C [ { q } ] f_k=[\exists\vec{\rho}_m.\Phi(\{\mathcal{A}_k\})]\in Q_C[\{q\}] f k = [ ∃ ρ m .Φ ({ A k })] ∈ Q C [{ q }] . We claim that f k ≠ f ℓ f_k\ne f_\ell f k = f ℓ whenever k < ℓ k<\ell k < ℓ . The equivalence classes f k , f ℓ f_k,f_\ell f k , f ℓ are defined on the class K \mathcal{K} K containing all tree-shaped frames of depth at most Δ \Delta Δ . Hence, the σ n \sigma_n σ n -model ( M [ u ; k ] , w ) (\mathcal{M}[u;k],w) ( M [ u ; k ] , w ) satisfies f k f_k f k and is based on a frame in K \mathcal{K} K . If it satisfied f ℓ f_\ell f ℓ , some ρ m \rho_m ρ m -valuation on the same frame would realize A ℓ \mathcal{A}_\ell A ℓ . By Lemma 3, there would be a surjective p-morphism onto M [ u ; ℓ ] \mathcal{M}[u;\ell] M [ u ; ℓ ] . But
∣ M [ u ; k ] ∣ = ∣ W ∖ U ∣ + k ∣ U ∣ < ∣ W ∖ U ∣ + ℓ ∣ U ∣ = ∣ M [ u ; ℓ ] ∣ . |\mathcal{M}[u;k]|=|W\setminus U|+k|U|<|W\setminus U|+\ell|U|=|\mathcal{M}[u;\ell]|. ∣ M [ u ; k ] ∣ = ∣ W ∖ U ∣ + k ∣ U ∣ < ∣ W ∖ U ∣ + ℓ ∣ U ∣ = ∣ M [ u ; ℓ ] ∣.
a contradiction. Hence ∣ Q C [ { q } ] ∣ > v |Q_C[\{q\}]|>v ∣ Q C [{ q }] ∣ > v .
For sufficiently large m m m , take v = ⌊ 2 m K n ⌋ − 2 v=\left\lfloor\frac{2^m}{K_n}\right\rfloor-2 v = ⌊ K n 2 m ⌋ − 2 . Then m ≥ m v m\ge m_v m ≥ m v , so the hypothesis of Lemma 51 holds with
L n ( m ) = ⌊ 2 m K n ⌋ − 1 ⟶ ∞ . L_n(m)=\left\lfloor\frac{2^m}{K_n}\right\rfloor-1\longrightarrow\infty. L n ( m ) = ⌊ K n 2 m ⌋ − 1 ⟶ ∞.
Corollary 52. For any Δ ≥ 1 \Delta\ge1 Δ ≥ 1 and any class K \mathcal{K} K of rooted frames determining Grz [ Δ ] \operatorname{Grz}_{[\Delta]} Grz [ Δ ] and containing all finite tree-shaped frames of depth at most Δ \Delta Δ : if n ≥ 0 n\ge0 n ≥ 0 is fixed, then lim m → ∞ UnifInt n , m Π = 1 \lim_{m\to\infty}\operatorname{UnifInt}^{\Pi}_{n,m}=1 lim m → ∞ UnifInt n , m Π = 1 .
For DCraig n , m Π \operatorname{DCraig}^{\Pi}_{n,m} DCraig n , m Π , Theorem 39 gives a positive lower bound depending on n n n and Δ \Delta Δ . Whether DCraig n , m Π → 1 \operatorname{DCraig}^{\Pi}_{n,m}\to1 DCraig n , m Π → 1 for fixed n , Δ n,\Delta n , Δ and m → ∞ m\to\infty m → ∞ remains open.
The Difference Logic Diff
We consider the logic Diff of modal formulas valid in difference frames ( W , ≠ ) (W,\ne) ( W , = ) . Thus, w ⊨ ◊ p w\models\Diamond p w ⊨ ◊ p if p p p holds at some world different from w w w . We show that lim m → ∞ DCraig n , m = 1 \lim_{m\to\infty}\operatorname{DCraig}_{n,m}=1 lim m → ∞ DCraig n , m = 1 , whereas on difference frames lim m → ∞ DCraig n , m Π = lim m → ∞ UnifInt n , m Π = 0 \lim_{m\to\infty}\operatorname{DCraig}^{\Pi}_{n,m}=\lim_{m\to\infty}\operatorname{UnifInt}^{\Pi}_{n,m}=0 lim m → ∞ DCraig n , m Π = lim m → ∞ UnifInt n , m Π = 0 .
Not every frame validating Diff is a difference frame. In fact, let K ≠ \mathcal{K}_{\ne} K = be the class of all difference frames and K D i f f \mathcal{K}_{\mathrm{Diff}} K Diff be the class of frames ( W , R ) (W,R) ( W , R ) with R ⊇ { ( w , v ) ∈ W 2 ∣ w ≠ v } R\supseteq\{(w,v)\in W^2\mid w\ne v\} R ⊇ {( w , v ) ∈ W 2 ∣ w = v } .
Lemma 53 (Segerberg [40 ] ). D i f f = Log ( K D i f f ) = Log ( K ≠ ) \mathrm{Diff}=\operatorname{Log}(\mathcal{K}_{\mathrm{Diff}})=\operatorname{Log}(\mathcal{K}_{\ne}) Diff = Log ( K Diff ) = Log ( K = ) , and a connected frame ( W , R ) (W,R) ( W , R ) validates Diff iff ( W , R ) ∈ K D i f f (W,R)\in\mathcal{K}_{\mathrm{Diff}} ( W , R ) ∈ K Diff .
Proof. Follows from the observation that a reflexive world can be obtained as a p-morphic image of at least two distinct irreflexive worlds.
The logic Diff is locally tabular. Hence, by Lemma 3, its τ \tau τ -atoms can be identified with τ \tau τ -standard pointed models. In such models, each τ \tau τ -colour is realised at most once, since worlds of the same colour satisfy the same formulas in M L τ \mathrm{ML}_{\tau} ML τ . Consequently, a colour c c c satisfying c ∧ ◊ c c\wedge\Diamond c c ∧ ◊ c must be represented by a reflexive world, and standard pointed models need not be based on difference frames. For our purposes, it is nevertheless more convenient to work with difference frames and record whether a colour is realised by zero, one, or at least two worlds. We therefore identify a τ \tau τ -atom with a pair ( F , c 0 ) (F,c_0) ( F , c 0 ) , where F F F maps each τ \tau τ -colour c c c to an element of { 0 , 1 , ≥ 2 } \{0,1,\ge2\} { 0 , 1 , ≥ 2 } , and c 0 c_0 c 0 is the τ \tau τ -colour of the distinguished world, with F ( c 0 ) ≠ 0 F(c_0)\ne0 F ( c 0 ) = 0 .
Note that a characteristic formula of ( F , c 0 ) (F,c_0) ( F , c 0 ) is given by
Φ ( F , c 0 ) = c 0 ∧ ⋀ F ( c ) = 0 ¬ ( c ∨ ◊ c ) ∧ ⋀ F ( c ) = 1 ( ( c ∨ ◊ c ) ∧ ¬ ◊ ( c ∧ ◊ c ) ) ∧ ⋀ F ( c ) ≥ 2 ( c ∧ ◊ c ) . \Phi(F,c_0)=c_0\ \wedge\ \bigwedge_{F(c)=0}\neg(c\vee\Diamond c)\ \wedge\ \bigwedge_{F(c)=1}\bigl((c\vee\Diamond c)\wedge\neg\Diamond(c\wedge\Diamond c)\bigr)\ \wedge\ \bigwedge_{F(c)\ge2}(c\wedge\Diamond c). Φ ( F , c 0 ) = c 0 ∧ F ( c ) = 0 ⋀ ¬ ( c ∨ ◊ c ) ∧ F ( c ) = 1 ⋀ ( ( c ∨ ◊ c ) ∧ ¬ ◊ ( c ∧ ◊ c ) ) ∧ F ( c ) ≥ 2 ⋀ ( c ∧ ◊ c ) .
As before, let σ = { p 1 , … , p n } \sigma=\{p_1,\ldots,p_n\} σ = { p 1 , … , p n } and ρ = { q 1 , … , q m } \rho=\{q_1,\ldots,q_m\} ρ = { q 1 , … , q m } be disjoint signatures, and put N = 2 n N=2^n N = 2 n , M = 2 m M=2^m M = 2 m , and K = 2 n + m K=2^{n+m} K = 2 n + m . There are K K K choices for the distinguished colour, two choices for its multiplicity, and three choices for each of the remaining K − 1 K-1 K − 1 colours. Hence
∣ At σ ∪ ρ ∣ = 2 K 3 K − 1 , ∣ At σ ∣ = 2 N 3 N − 1 . (10) |\operatorname{At}_{\sigma\cup\rho}|=2K3^{K-1},\qquad|\operatorname{At}_{\sigma}|=2N3^{N-1}. \tag*{(10)} ∣ At σ ∪ ρ ∣ = 2 K 3 K − 1 , ∣ At σ ∣ = 2 N 3 N − 1 . (10)
Fix a cell C ∈ Cell σ , ρ C \in\operatorname{Cell}_{\sigma,\rho} C ∈ Cell σ , ρ . Its atoms can be regarded as refinements of the σ \sigma σ -atom ( F σ , c σ ) (F_{\sigma},c_{\sigma}) ( F σ , c σ ) underlying C C C . We describe these refinements in detail. Partition the σ \sigma σ -colours into F σ 0 F_{\sigma}^{0} F σ 0 , F σ 1 F_{\sigma}^{1} F σ 1 , and F σ ≥ 2 F_{\sigma}^{\ge2} F σ ≥ 2 , according to whether their multiplicity in F σ F_{\sigma} F σ is zero, exactly one, or at least two. We call the colours in F σ ≥ 2 F_{\sigma}^{\ge2} F σ ≥ 2 heavy and those in F σ 1 F_{\sigma}^{1} F σ 1 light. Each atom in C C C is determined by a family of refinement maps δ s \delta_s δ s , for s ∈ F σ ≥ 2 ∪ F σ 1 s \in F_{\sigma}^{\ge2} \cup F_{\sigma}^{1} s ∈ F σ ≥ 2 ∪ F σ 1 , which are functions of the form
{ ρ -colours } ⟶ { 0 , 1 , ≥ 2 } , \{\rho\text{-colours}\}\ \longrightarrow\ \{0,1,\ge2\}, { ρ -colours } ⟶ { 0 , 1 , ≥ 2 } ,
together with a choice of a ρ \rho ρ -colour r σ r_{\sigma} r σ for the distinguished world, satisfying δ c σ ( r σ ) ≠ 0 \delta_{c_{\sigma}}(r_{\sigma}) \ne0 δ c σ ( r σ ) = 0 .
For a refinement map δ : { ρ -colours } ⟶ { 0 , 1 , ≥ 2 } \delta: \{\rho\text{-colours}\} \longrightarrow\{0,1,\ge2\} δ : { ρ -colours } ⟶ { 0 , 1 , ≥ 2 } , set
λ ( δ ) : = ∣ { r ∣ δ ( r ) = 1 } ∣ + 2 ∣ { r ∣ δ ( r ) = ≥ 2 } ∣ . \lambda(\delta) := \lvert\{r \mid\delta(r)=1\}\rvert+ 2\lvert\{r \mid\delta(r)=\ge2\}\rvert. λ ( δ ) := ∣{ r ∣ δ ( r ) = 1 }∣ + 2 ∣{ r ∣ δ ( r ) =≥ 2 }∣ .
The value λ ( δ s ) \lambda(\delta_s) λ ( δ s ) is the least possible number of worlds whose ( σ ∪ ρ ) (\sigma\cup\rho) ( σ ∪ ρ ) -colours refine s s s according to δ s \delta_s δ s . To define a member of C C C , we require λ ( δ s ) ≥ 2 \lambda(\delta_s) \ge2 λ ( δ s ) ≥ 2 for s ∈ F σ ≥ 2 s \in F_{\sigma}^{\ge2} s ∈ F σ ≥ 2 and λ ( δ s ) = 1 \lambda(\delta_s)=1 λ ( δ s ) = 1 for s ∈ F σ 1 s \in F_{\sigma}^{1} s ∈ F σ 1 : a heavy colour is realised at least twice, while a light colour has exactly one refinement, realised at a unique world.
We call a refinement map δ \delta δ rigid if δ ( r ) ∈ { 0 , 1 } \delta(r) \in\{0,1\} δ ( r ) ∈ { 0 , 1 } for all ρ \rho ρ -colours r r r and flexible otherwise. If δ s \delta_s δ s is rigid, then λ ( δ s ) \lambda(\delta_s) λ ( δ s ) is the exact number of worlds that satisfy ( σ ∪ ρ ) (\sigma\cup\rho) ( σ ∪ ρ ) -colours refining s s s according to δ s \delta_s δ s ; if δ s \delta_s δ s is flexible (and so s ∈ F σ ≥ 2 s \in F_{\sigma}^{\ge2} s ∈ F σ ≥ 2 ), then any k ∈ [ λ ( δ s ) , ∞ ) k \in[\lambda(\delta_s),\infty) k ∈ [ λ ( δ s ) , ∞ ) can be the number of worlds that satisfy ( σ ∪ ρ ) (\sigma\cup\rho) ( σ ∪ ρ ) -colours refining s s s . So, for a refinement map δ \delta δ , define the interval specifying a possible number of worlds satisfying the refinements:
I ( δ ) = { { λ ( δ ) } , δ rigid , [ λ ( δ ) , ∞ ) , δ flexible . I(\delta)=
\begin{cases}
\{\lambda(\delta)\}, & \delta\text{ rigid},\\
[\lambda(\delta),\infty), & \delta\text{ flexible}.
\end{cases} I ( δ ) = { { λ ( δ )} , [ λ ( δ ) , ∞ ) , δ rigid , δ flexible .
We now construct a profile of the incompatibility graph of C C C . It records, for each heavy σ \sigma σ -colour, the possible numbers of worlds realising that colour, while disregarding the individual ρ \rho ρ -refinements and the choice of the distinguished world. Observe that, for a rigid δ s \delta_s δ s , the value of λ ( δ s ) \lambda(\delta_s) λ ( δ s ) is bounded by 2 ≤ λ ( δ s ) ≤ M 2 \le\lambda(\delta_s) \le M 2 ≤ λ ( δ s ) ≤ M , while for a flexible δ s \delta_s δ s , it is bounded by 2 ≤ λ ( δ s ) ≤ 2 M 2 \le\lambda(\delta_s) \le2M 2 ≤ λ ( δ s ) ≤ 2 M . So, consider the set of possible values of I ( δ s ) I(\delta_s) I ( δ s ) :
Π M = Π M R ∪ Π M F , where Π M R = { { λ } ∣ 2 ≤ λ ≤ M } and Π M F = { [ λ , ∞ ) ∣ 2 ≤ λ ≤ 2 M } . \Pi_M=\Pi_M^{\mathrm{R}}\cup\Pi_M^{\mathrm{F}},\quad\text{where}\quad
\Pi_M^{\mathrm{R}}=\{\{\lambda\}\mid2\le\lambda\le M\}
\quad\text{and}\quad
\Pi_M^{\mathrm{F}}=\{[\lambda,\infty)\mid2\le\lambda\le2M\}. Π M = Π M R ∪ Π M F , where Π M R = {{ λ } ∣ 2 ≤ λ ≤ M } and Π M F = {[ λ , ∞ ) ∣ 2 ≤ λ ≤ 2 M } .
Let P = ( Π M ) F σ ≥ 2 P=(\Pi_M)^{F_{\sigma}^{\ge2}} P = ( Π M ) F σ ≥ 2 be the set of mappings p p p from F σ ≥ 2 F_{\sigma}^{\ge2} F σ ≥ 2 to Π M \Pi_M Π M . For s ∈ F σ ≥ 2 s \in F_{\sigma}^{\ge2} s ∈ F σ ≥ 2 , we call each p ( s ) ∈ Π M R p(s)\in\Pi_M^{\mathrm{R}} p ( s ) ∈ Π M R a rigid component of p p p and each p ( s ) ∈ Π M F p(s)\in\Pi_M^{\mathrm{F}} p ( s ) ∈ Π M F a flexible component of p p p . Define incompatibility in P P P by setting p ⌢ P q p\mathrel{\frown^{P}}q p ⌢ P q if p ( s ) ∩ q ( s ) = ∅ p(s)\cap q(s)=\varnothing p ( s ) ∩ q ( s ) = ∅ for some s ∈ F σ ≥ 2 s\in F_{\sigma}^{\ge2} s ∈ F σ ≥ 2 . The graph H P = ( P , ⌢ P ) H_P=(P,\mathrel{\frown^{P}}) H P = ( P , ⌢ P ) is called the box-profile graph of C C C . We define the profile mapping π : C → P \pi:C\to P π : C → P as follows. Using I I I , we can map any M ∈ C \mathcal{M}\in C M ∈ C with refinement mappings ( δ s ) s ∈ F σ ≥ 2 (\delta_s)_{s\in F_{\sigma}^{\ge2}} ( δ s ) s ∈ F σ ≥ 2 to an element of P P P by setting
B ( M ) = { s ↦ I ( δ s ) ∣ s ∈ F σ ≥ 2 } ∈ ( Π M ) F σ ≥ 2 . B(\mathcal{M})=\{s\mapsto I(\delta_s)\mid s\in F_{\sigma}^{\ge2}\}\in(\Pi_M)^{F_{\sigma}^{\ge2}}. B ( M ) = { s ↦ I ( δ s ) ∣ s ∈ F σ ≥ 2 } ∈ ( Π M ) F σ ≥ 2 .
Observe that, for M , N ∈ C \mathcal{M},\mathcal{N}\in C M , N ∈ C with refinement mappings ( δ s M ) s ∈ F σ ≥ 2 (\delta_s^{\mathcal{M}})_{s\in F_{\sigma}^{\ge2}} ( δ s M ) s ∈ F σ ≥ 2 and ( δ s N ) s ∈ F σ ≥ 2 (\delta_s^{\mathcal{N}})_{s\in F_{\sigma}^{\ge2}} ( δ s N ) s ∈ F σ ≥ 2 , we have
M ⌢ N ⟺ I ( δ s M ) ∩ I ( δ s N ) = ∅ for some s ∈ F σ ≥ 2 . \mathcal{M}\mathrel{\frown}\mathcal{N}
\quad\Longleftrightarrow\quad
I(\delta_s^{\mathcal{M}})\cap I(\delta_s^{\mathcal{N}})=\varnothing
\quad\text{for some }s\in F_{\sigma}^{\ge2}. M ⌢ N ⟺ I ( δ s M ) ∩ I ( δ s N ) = ∅ for some s ∈ F σ ≥ 2 .
Now one can easily show the following.
Lemma 54. π = B \pi=B π = B is a profile map for C C C onto ( P , ⌢ P ) (P,\mathrel{\frown^{P}}) ( P , ⌢ P ) , and ∣ P ∣ = ( 3 M − 2 ) h ≤ ( 3 M ) h |P|=(3M-2)^h\le(3M)^h ∣ P ∣ = ( 3 M − 2 ) h ≤ ( 3 M ) h , where h = ∣ F σ ≥ 2 ∣ h=|F_{\sigma}^{\ge2}| h = ∣ F σ ≥ 2 ∣ .
As before, the weights of a profile p ∈ P p\in P p ∈ P and set Q ⊆ P Q\subseteq P Q ⊆ P for mapping π \pi π are defined as w ( p ) = ∣ π − 1 ( p ) ∣ w(p)=|\pi^{-1}(p)| w ( p ) = ∣ π − 1 ( p ) ∣ and w ( Q ) = ∑ p ∈ Q w ( p ) w(Q)=\sum_{p\in Q}w(p) w ( Q ) = ∑ p ∈ Q w ( p ) , respectively. To compute these weights, we first compute weights w 1 w_1 w 1 and w 1 ( ∘ w_1^{\mathstrut\circ} w 1 ( ∘ for the component mappings I I I and then lift these to π \pi π . For J ∈ Π M J\in\Pi_M J ∈ Π M , let
w 1 ( J ) : = ∣ I − 1 ( J ) ∣ . w_1(J):=|I^{-1}(J)|. w 1 ( J ) := ∣ I − 1 ( J ) ∣.
Abbreviate
R λ R = I − 1 ( { λ } ) = { δ is a rigid refinement map ∣ λ ( δ ) = λ } , for { λ } ∈ Π M R , \mathcal{R}_{\lambda}^{\mathrm{R}}=I^{-1}(\{\lambda\})=\{\delta\text{ is a rigid refinement map}\mid\lambda(\delta)=\lambda\},
\qquad\text{for }\{\lambda\}\in\Pi_M^{\mathrm{R}}, R λ R = I − 1 ({ λ }) = { δ is a rigid refinement map ∣ λ ( δ ) = λ } , for { λ } ∈ Π M R ,
R λ F = I − 1 ( [ λ , ∞ ) ) = { δ is a flexible refinement map ∣ λ ( δ ) = λ } , for [ λ , ∞ ) ∈ Π M F . \mathcal{R}_{\lambda}^{\mathrm{F}}=I^{-1}([\lambda,\infty))=\{\delta\text{ is a flexible refinement map}\mid\lambda(\delta)=\lambda\},
\qquad\text{for }[\lambda,\infty)\in\Pi_M^{\mathrm{F}}. R λ F = I − 1 ([ λ , ∞ )) = { δ is a flexible refinement map ∣ λ ( δ ) = λ } , for [ λ , ∞ ) ∈ Π M F .
so that
w 1 ( { λ } ) = ∣ R λ R ∣ , w 1 ( [ λ , ∞ ) ) = ∣ R λ F ∣ . w_1(\{\lambda\}) = \lvert\mathcal{R}^{R}_{\lambda}\rvert,\qquad w_1([\lambda,\infty)) = \lvert\mathcal{R}^{F}_{\lambda}\rvert. w 1 ({ λ }) = ∣ R λ R ∣ , w 1 ([ λ , ∞ )) = ∣ R λ F ∣ .
Counting refinement maps gives
w 1 ( { λ } ) = ( M λ ) , w 1 ( Π M F ) = ∑ λ = 2 2 M w 1 ( [ λ , ∞ ) ) = 3 M − 2 M , w 1 ( [ 2 , ∞ ) ) = M . (11) w_1(\{\lambda\}) = \binom{M}{\lambda},\qquad w_1(\Pi^{F}_{M}) = \sum_{\lambda=2}^{2M} w_1([\lambda,\infty)) = 3^M - 2^M,\qquad w_1([2,\infty)) = M. \tag*{(11)} w 1 ({ λ }) = ( λ M ) , w 1 ( Π M F ) = λ = 2 ∑ 2 M w 1 ([ λ , ∞ )) = 3 M − 2 M , w 1 ([ 2 , ∞ )) = M . (11)
To take into account the distinguished σ \sigma σ -colour c σ c_{\sigma} c σ , we also count marked refinements , i.e., pairs ( δ , r ) (\delta,r) ( δ , r ) such that r r r is a ρ \rho ρ -colour with r ∈ supp ( δ ) r \in\operatorname{supp}(\delta) r ∈ supp ( δ ) , where supp ( δ ) = { r ∣ δ ( r ) ≠ 0 } \operatorname{supp}(\delta) = \{r \mid\delta(r) \ne0\} supp ( δ ) = { r ∣ δ ( r ) = 0 } :
w 1 ∙ ( { λ } ) = ∣ { ( δ , r ) : δ ∈ R λ R , r ∈ supp ( δ ) } ∣ , w 1 ∙ ( [ λ , ∞ ) ) = ∣ { ( δ , r ) : δ ∈ R λ F , r ∈ supp ( δ ) } ∣ . w^{\bullet}_1(\{\lambda\}) = \lvert\{(\delta,r) : \delta\in\mathcal{R}^{R}_{\lambda},\ r \in\operatorname{supp}(\delta)\}\rvert,\qquad
w^{\bullet}_1([\lambda,\infty)) = \lvert\{(\delta,r) : \delta\in\mathcal{R}^{F}_{\lambda},\ r \in\operatorname{supp}(\delta)\}\rvert. w 1 ∙ ({ λ }) = ∣{( δ , r ) : δ ∈ R λ R , r ∈ supp ( δ )}∣ , w 1 ∙ ([ λ , ∞ )) = ∣{( δ , r ) : δ ∈ R λ F , r ∈ supp ( δ )}∣ .
Then
w 1 ∙ ( { λ } ) = ∑ δ ∈ R λ R ∣ supp ( δ ) ∣ = λ ( M λ ) , w 1 ∙ ( [ λ , ∞ ) ) = ∑ δ ∈ R λ F ∣ supp ( δ ) ∣ . w^{\bullet}_1(\{\lambda\}) = \sum_{\delta\in\mathcal{R}^{R}_{\lambda}} \lvert\operatorname{supp}(\delta)\rvert= \lambda\binom{M}{\lambda},\qquad
w^{\bullet}_1([\lambda,\infty)) = \sum_{\delta\in\mathcal{R}^{F}_{\lambda}} \lvert\operatorname{supp}(\delta)\rvert. w 1 ∙ ({ λ }) = δ ∈ R λ R ∑ ∣ supp ( δ )∣ = λ ( λ M ) , w 1 ∙ ([ λ , ∞ )) = δ ∈ R λ F ∑ ∣ supp ( δ )∣ .
As ∣ supp ( δ ) ∣ ≥ 1 \lvert\operatorname{supp}(\delta)\rvert\ge1 ∣ supp ( δ )∣ ≥ 1 in general and ∣ supp ( δ ) ∣ = 1 \lvert\operatorname{supp}(\delta)\rvert= 1 ∣ supp ( δ )∣ = 1 on R λ F \mathcal{R}^{F}_{\lambda} R λ F ,
w 1 ∙ ( { λ } ) ≥ w 1 ( { λ } ) , w 1 ∙ ( [ λ , ∞ ) ) ≥ w 1 ( [ λ , ∞ ) ) , w 1 ∙ ( [ 2 , ∞ ) ) = M . (12) w^{\bullet}_1(\{\lambda\}) \ge w_1(\{\lambda\}),\qquad
w^{\bullet}_1([\lambda,\infty)) \ge w_1([\lambda,\infty)),\qquad
w^{\bullet}_1([2,\infty)) = M. \tag*{(12)} w 1 ∙ ({ λ }) ≥ w 1 ({ λ }) , w 1 ∙ ([ λ , ∞ )) ≥ w 1 ([ λ , ∞ )) , w 1 ∙ ([ 2 , ∞ )) = M . (12)
We lift these weights to π \pi π . Fix a box-profile q ∈ P q \in P q ∈ P and consider the inverse image
π − 1 ( q ) = { M ∈ C ∣ π ( M ) = q } . \pi^{-1}(q) = \{\mathcal{M} \in C \mid\pi(\mathcal{M}) = q\}. π − 1 ( q ) = { M ∈ C ∣ π ( M ) = q } .
The refinement maps for distinct σ \sigma σ -colours can be chosen independently, subject only to the choice of the refinement of the distinguished world. Consequently,
π − 1 ( q ) ≅ { ∏ s ∈ F σ 1 2 ρ × ∏ s ∈ F σ ≥ 2 I − 1 ( q ( s ) ) , c σ ∈ F σ 1 , ∏ s ∈ F σ 1 2 ρ × { ( δ , r ) ∣ δ ∈ I − 1 ( q ( c σ ) ) , r ∈ supp ( δ ) } × ∏ s ∈ F σ ≥ 2 ∖ { c σ } I − 1 ( q ( s ) ) , c σ ∈ F σ ≥ 2 . \pi^{-1}(q) \cong
\begin{cases}
\prod_{s\in F^{1}_{\sigma}} 2^{\rho}\times\prod_{s\in F^{\ge2}_{\sigma}} I^{-1}(q(s)), & c_{\sigma}\in F^{1}_{\sigma},\\[6pt]
\prod_{s\in F^{1}_{\sigma}}2^{\rho}\times\{(\delta,r)\mid\delta\in I^{-1}(q(c_{\sigma})),\ r\in\operatorname{supp}(\delta)\}\times\prod_{s\in F^{\ge2}_{\sigma}\setminus\{c_{\sigma}\}}I^{-1}(q(s)), & c_{\sigma}\in F^{\ge2}_{\sigma}.
\end{cases} π − 1 ( q ) ≅ ⎩ ⎨ ⎧ ∏ s ∈ F σ 1 2 ρ × ∏ s ∈ F σ ≥ 2 I − 1 ( q ( s )) , ∏ s ∈ F σ 1 2 ρ × {( δ , r ) ∣ δ ∈ I − 1 ( q ( c σ )) , r ∈ supp ( δ )} × ∏ s ∈ F σ ≥ 2 ∖ { c σ } I − 1 ( q ( s )) , c σ ∈ F σ 1 , c σ ∈ F σ ≥ 2 .
Indeed, if c σ c_{\sigma} c σ is light, its unique refinement determines the colour of the distinguished world. If c σ c_{\sigma} c σ is heavy, an atom requires, in addition to its refinement maps, a choice of a ρ \rho ρ -colour r r r with δ c σ ( r ) ≠ 0 \delta_{c_{\sigma}}(r) \ne0 δ c σ ( r ) = 0 for the distinguished world. These are exactly the choices recorded by the marked refinements.
Taking cardinalities in, and writing ℓ = ∣ F σ 1 ∣ \ell= \lvert F^{1}_{\sigma}\rvert ℓ = ∣ F σ 1 ∣ , we obtain
w ( q ) = ∣ π − 1 ( q ) ∣ = { M ℓ × ∏ s ∈ F σ ≥ 2 w 1 ( q ( s ) ) , c σ ∈ F σ 1 , M ℓ × w 1 ∙ ( q ( c σ ) ) × ∏ s ∈ F σ ≥ 2 ∖ { c σ } w 1 ( q ( s ) ) , c σ ∈ F σ ≥ 2 . (13) w(q) = \lvert\pi^{-1}(q)\rvert=
\begin{cases}
M^{\ell}\times\prod_{s\in F^{\ge2}_{\sigma}}w_1(q(s)), & c_{\sigma}\in F^{1}_{\sigma},\\[6pt]
M^{\ell}\times w^{\bullet}_1(q(c_{\sigma}))\times\prod_{s\in F^{\ge2}_{\sigma}\setminus\{c_{\sigma}\}}w_1(q(s)), & c_{\sigma}\in F^{\ge2}_{\sigma}.
\end{cases} \tag*{(13)} w ( q ) = ∣ π − 1 ( q )∣ = ⎩ ⎨ ⎧ M ℓ × ∏ s ∈ F σ ≥ 2 w 1 ( q ( s )) , M ℓ × w 1 ∙ ( q ( c σ )) × ∏ s ∈ F σ ≥ 2 ∖ { c σ } w 1 ( q ( s )) , c σ ∈ F σ 1 , c σ ∈ F σ ≥ 2 . (13)
This concludes the computation of individual weights.
As before, for Q ⊆ P Q \subseteq P Q ⊆ P , the common neighbourhood of Q Q Q in P P P is denoted by N P ( Q ) N_P(Q) N P ( Q ) , while its complement by N P ( Q ) ‾ : = P ∖ N P ( Q ) \overline{N_P(Q)} := P \setminus N_P(Q) N P ( Q ) := P ∖ N P ( Q ) . Set g ( Q ) = w ( N P ( Q ) ‾ ) − w ( Q ) g(Q) = w(\overline{N_P(Q)}) - w(Q) g ( Q ) = w ( N P ( Q ) ) − w ( Q ) . Then w ( N P ( Q ) ) = ∣ C ∣ − w ( N P ( Q ) ‾ ) w(N_P(Q)) = \lvert C\rvert- w(\overline{N_P(Q)}) w ( N P ( Q )) = ∣ C ∣ − w ( N P ( Q ) ) and so
w ( Q ) + w ( N P ( Q ) ) = ∣ C ∣ − g ( Q ) . (14) w(Q) + w(N_P(Q)) = \lvert C\rvert- g(Q). \tag*{(14)} w ( Q ) + w ( N P ( Q )) = ∣ C ∣ − g ( Q ) . (14)
We use the function g ( Q ) g(Q) g ( Q ) to bound the number of inconsistent pairs without a Craig interpolant within cells C C C , NoCraig C = ∣ InCon C ∣ − ∣ Craig C ∣ \operatorname{NoCraig}_C = \lvert\operatorname{InCon}_C\rvert- \lvert\operatorname{Craig}_C\rvert NoCraig C = ∣ InCon C ∣ − ∣ Craig C ∣ ; see (2) .
Lemma 55. Let G ⊆ 2 P ∖ { ∅ } \mathcal{G} \subseteq2^P \setminus\{\varnothing\} G ⊆ 2 P ∖ { ∅ } and assume that every non-empty ordered biclique in P P P can be oriented so that its first side belongs to G \mathcal{G} G . Then
NoCraig C ≤ 2 ∣ C ∣ + 1 ∑ Q ∈ G 2 − g ( Q ) . (15) \operatorname{NoCraig}_C \le2^{\lvert C\rvert+1}\sum_{Q\in\mathcal{G}}2^{-g(Q)}. \tag*{(15)} NoCraig C ≤ 2 ∣ C ∣ + 1 Q ∈ G ∑ 2 − g ( Q ) . (15) In particular, if g ( Q ) ≥ Δ g(Q) \ge\Delta g ( Q ) ≥ Δ for every Q ∈ G Q \in\mathcal{G} Q ∈ G , then
1 − DCraig C ≤ 2 ∣ P ∣ + 1 − Δ . (16) 1-\operatorname{DCraig}_C \le2^{\lvert P\rvert+1-\Delta}. \tag*{(16)} 1 − DCraig C ≤ 2 ∣ P ∣ + 1 − Δ . (16)
Proof. The first inequality follows from Lemma 27 and (12) with a factor 2 to take into account the orientation of bicliques. Divide by 2 ∣ C ∣ 2^{|C|} 2 ∣ C ∣ and use (2) for the second inequality.
We now determine a sufficiently large Δ \Delta Δ for an appropriate G \mathcal{G} G . Assume F σ ≥ 2 ≠ ∅ F_{\sigma}^{\ge2} \ne\varnothing F σ ≥ 2 = ∅ , that is, h ≥ 1 h \ge1 h ≥ 1 . For a box-profile p p p , let
Flex ( p ) = { s ∣ p ( s ) is flexible } and Rigid ( p ) = { s ∣ p ( s ) is rigid } . \operatorname{Flex}(p)=\{s \mid p(s)\text{ is flexible}\}
\qquad\text{and}\qquad
\operatorname{Rigid}(p)=\{s \mid p(s)\text{ is rigid}\}. Flex ( p ) = { s ∣ p ( s ) is flexible } and Rigid ( p ) = { s ∣ p ( s ) is rigid } .
Call a box-profile p p p somewhere rigid if ∣ Rigid ( p ) ∣ > 0 |\operatorname{Rigid}(p)|>0 ∣ Rigid ( p ) ∣ > 0 , otherwise call it everywhere flexible . Let S \mathcal{S} S be the set of non-empty Q ⊆ P Q\subseteq P Q ⊆ P such that each p ∈ Q p\in Q p ∈ Q is somewhere rigid.
Lemma 56. Every non-empty ordered biclique in P P P can be oriented so that its first side belongs to S \mathcal{S} S .
Proof. Clearly q 1 ̸ ∼ P q 2 q_1\not\sim_P q_2 q 1 ∼ P q 2 for any two everywhere flexible q 1 , q 2 ∈ P q_1,q_2\in P q 1 , q 2 ∈ P . Hence an ordered biclique in P P P cannot contain an everywhere flexible profile on both sides. Orient it so that the first side contains none.
We are in a position now to prove the main lower bound on g ( Q ) g(Q) g ( Q ) and NoCraig G \operatorname{NoCraig}_{\mathcal{G}} NoCraig G .
Theorem 57. Assume h = ∣ F σ ≥ 2 ∣ h=|F_{\sigma}^{\ge2}| h = ∣ F σ ≥ 2 ∣ , ℓ = ∣ F σ 1 ∣ \ell=|F_{\sigma}^{1}| ℓ = ∣ F σ 1 ∣ , and M = 2 m ≥ 2 M=2^m\ge2 M = 2 m ≥ 2 . For every Q ∈ S Q\in\mathcal{S} Q ∈ S ,
g ( Q ) ≥ 1 4 h − 1 M ℓ + 2 h − 1 . (17) g(Q)\ge\frac{1}{4^{h-1}}M^{\ell+2h-1}. \tag*{(17)} g ( Q ) ≥ 4 h − 1 1 M ℓ + 2 h − 1 . (17)
Proof. Choose q ∈ Q q\in Q q ∈ Q with
k = ∣ Flex ( q ) ∣ k=|\operatorname{Flex}(q)| k = ∣ Flex ( q ) ∣
maximal among Q Q Q . Since Q ∈ S Q\in\mathcal{S} Q ∈ S , we have ∣ Rigid ( q ) ∣ > 0 |\operatorname{Rigid}(q)|>0 ∣ Rigid ( q ) ∣ > 0 . The shell E ( q ) ⊆ P \mathcal{E}(q)\subseteq P E ( q ) ⊆ P consists of the box-profiles p p p satisfying the following:
(i) s ∈ Flex ( p ) s\in\operatorname{Flex}(p) s ∈ Flex ( p ) , for every s ∈ Flex ( q ) s\in\operatorname{Flex}(q) s ∈ Flex ( q ) ;
(ii) q ( s ) ⊆ p ( s ) q(s)\subseteq p(s) q ( s ) ⊆ p ( s ) , for every s ∈ Rigid ( q ) s\in\operatorname{Rigid}(q) s ∈ Rigid ( q ) ;
(iii) s ∈ Flex ( p ) s\in\operatorname{Flex}(p) s ∈ Flex ( p ) , for at least one s ∈ Rigid ( q ) s\in\operatorname{Rigid}(q) s ∈ Rigid ( q ) .
Claim 57.1. E ( q ) ⊆ N ‾ ( Q ) ∖ Q \mathcal{E}(q)\subseteq\overline{N}(Q)\setminus Q E ( q ) ⊆ N ( Q ) ∖ Q . Consequently g ( Q ) ≥ w ( E ( q ) ) g(Q)\ge w(\mathcal{E}(q)) g ( Q ) ≥ w ( E ( q )) .
Proof. First observe that, for every p ∈ E ( q ) p\in\mathcal{E}(q) p ∈ E ( q ) , p ( s ) ∩ q ( s ) ≠ ∅ p(s)\cap q(s)\ne\varnothing p ( s ) ∩ q ( s ) = ∅ for all s ∈ F σ ≥ 2 s\in F_{\sigma}^{\ge2} s ∈ F σ ≥ 2 . Indeed, if q ( s ) q(s) q ( s ) is flexible, then, by (i), p ( s ) p(s) p ( s ) is flexible and p ( s ) ∩ q ( s ) ≠ ∅ p(s)\cap q(s)\ne\varnothing p ( s ) ∩ q ( s ) = ∅ follows. If q ( s ) = { λ } q(s)=\{\lambda\} q ( s ) = { λ } , then, by (ii), either p ( s ) p(s) p ( s ) is the same rigid component or p ( s ) = [ μ , ∞ ) p(s)=[\mu,\infty) p ( s ) = [ μ , ∞ ) with μ ≤ λ \mu\le\lambda μ ≤ λ , in which case λ ∈ [ μ , ∞ ) \lambda\in[\mu,\infty) λ ∈ [ μ , ∞ ) . Hence p ∈ N ‾ ( { q } ) ⊆ N ‾ ( Q ) p\in\overline{N}(\{q\})\subseteq\overline{N}(Q) p ∈ N ({ q }) ⊆ N ( Q ) .
By (i) and (iii), every member p p p of the shell has at least ( k + 1 ) (k+1) ( k + 1 ) -many flexible components, whereas maximality of q q q says that no member of Q Q Q has more than k k k . Thus E ( q ) ∩ Q = ∅ \mathcal{E}(q)\cap Q=\varnothing E ( q ) ∩ Q = ∅ .
We next show a lower bound for the weight of the shell.
Claim 57.2. Let L : = w 1 ( Π M F ) L:=w_1(\Pi_M^F) L := w 1 ( Π M F ) . For any box-profile q q q with ∣ Flex ( q ) ∣ = k |\operatorname{Flex}(q)|=k ∣ Flex ( q ) ∣ = k and ∣ Rigid ( q ) ∣ > 0 |\operatorname{Rigid}(q)|>0 ∣ Rigid ( q ) ∣ > 0 , we have
w ( E ( q ) ) ≥ M ℓ + 1 L k ( M 2 ) h − k − 1 . (18) w(\mathcal{E}(q))\ge M^{\ell+1}L^k\binom{M}{2}^{h-k-1}. \tag*{(18)} w ( E ( q )) ≥ M ℓ + 1 L k ( 2 M ) h − k − 1 . (18)
Proof. We say that a rigid component { λ } ∈ Π M R \{\lambda\}\in\Pi_M^R { λ } ∈ Π M R is compatible with a component J ∈ Π M J\in\Pi_M J ∈ Π M if { λ } ⊆ J \{\lambda\}\subseteq J { λ } ⊆ J . Observe that { λ } ⊆ [ μ , ∞ ) \{\lambda\}\subseteq[\mu,\infty) { λ } ⊆ [ μ , ∞ ) iff μ ≤ λ \mu\le\lambda μ ≤ λ . So, for each { λ } ∈ Π M R \{\lambda\}\in\Pi_M^R { λ } ∈ Π M R , we introduce the ordinary weights of, respectively, rigid, flexible and all components compatible with { λ } \{\lambda\} { λ } :
w { λ } R = w 1 ( { λ } ) , w { λ } F = ∑ J ∈ Π M F { λ } ⊆ J w 1 ( J ) = ∑ μ = 2 λ w 1 ( [ μ , ∞ ) ) , w { λ } = ∑ J ∈ Π M { λ } ⊆ J w 1 ( J ) = w { λ } R + w { λ } F . w_{\{\lambda\}}^R=w_1(\{\lambda\}),\qquad
w_{\{\lambda\}}^F=\sum_{\substack{J\in\Pi_M^F\\\{\lambda\}\subseteq J}}w_1(J)=\sum_{\mu=2}^{\lambda}w_1([\mu,\infty)),
\qquad
w_{\{\lambda\}}=\sum_{\substack{J\in\Pi_M\\\{\lambda\}\subseteq J}}w_1(J)=w_{\{\lambda\}}^R+w_{\{\lambda\}}^F. w { λ } R = w 1 ({ λ }) , w { λ } F = J ∈ Π M F { λ } ⊆ J ∑ w 1 ( J ) = μ = 2 ∑ λ w 1 ([ μ , ∞ )) , w { λ } = J ∈ Π M { λ } ⊆ J ∑ w 1 ( J ) = w { λ } R + w { λ } F .
We can describe the shell as follows:
E ( q ) = ( ∏ s ∈ F σ ≥ 2 P s ) ∖ D , (19) \mathcal{E}(q)=\left(\prod_{s\in F_{\sigma}^{\ge2}}P_s\right)\setminus\mathcal{D}, \tag*{(19)} E ( q ) = s ∈ F σ ≥ 2 ∏ P s ∖ D , (19)
where
P s = Π M F P_s=\Pi^{\mathrm{F}}_M P s = Π M F , for each s ∈ Flex ( q ) s\in\operatorname{Flex}(q) s ∈ Flex ( q ) , see (i);
P s = { J ∈ Π M ∣ q ( s ) ⊆ J } P_s=\{J\in\Pi_M\mid q(s)\subseteq J\} P s = { J ∈ Π M ∣ q ( s ) ⊆ J } , for each s ∈ Rigid ( q ) s\in\operatorname{Rigid}(q) s ∈ Rigid ( q ) , see (ii);
and D = { p ∈ Π s ∈ F σ ≥ 2 P s ∣ p ( s ) = q ( s ) , for all s ∈ Rigid ( q ) } \mathcal{D}=\{p\in\Pi_{s\in F_{\sigma}^{\ge2}}P_s\mid p(s)=q(s),\text{ for all }s\in\operatorname{Rigid}(q)\} D = { p ∈ Π s ∈ F σ ≥ 2 P s ∣ p ( s ) = q ( s ) , for all s ∈ Rigid ( q )} , see (iii).
Assume first c σ ∈ F σ 1 c_\sigma\in F^1_\sigma c σ ∈ F σ 1 , so, by (11), the weight w ( p ) = M ℓ ∏ s ∈ F σ ≥ 2 w 1 ( p ( s ) ) w(p)=M^\ell\prod_{s\in F_{\sigma}^{\ge2}}w_1(p(s)) w ( p ) = M ℓ ∏ s ∈ F σ ≥ 2 w 1 ( p ( s )) is multiplicative over s ∈ F σ ≥ 2 s\in F_{\sigma}^{\ge2} s ∈ F σ ≥ 2 . Hence, for any product family ∏ s ∈ F σ ≥ 2 C s \prod_{s\in F_{\sigma}^{\ge2}}C_s ∏ s ∈ F σ ≥ 2 C s with C s ⊆ Π M C_s\subseteq\Pi_M C s ⊆ Π M ,
w ( ∏ s ∈ F σ ≥ 2 C s ) = M ℓ ∏ s ∈ F σ ≥ 2 w 1 ( C s ) , w 1 ( C s ) : = ∑ x ∈ C s w 1 ( x ) . w\left(\prod_{s\in F_{\sigma}^{\ge2}}C_s\right)=M^\ell\prod_{s\in F_{\sigma}^{\ge2}}w_1(C_s),\qquad w_1(C_s):=\sum_{x\in C_s}w_1(x). w s ∈ F σ ≥ 2 ∏ C s = M ℓ s ∈ F σ ≥ 2 ∏ w 1 ( C s ) , w 1 ( C s ) := x ∈ C s ∑ w 1 ( x ) .
By the definitions of P s P_s P s , L L L and w { λ } w_{\{\lambda\}} w { λ } , we have w 1 ( P s ) = L w_1(P_s)=L w 1 ( P s ) = L , for each s ∈ Flex ( q ) s\in\operatorname{Flex}(q) s ∈ Flex ( q ) , and w 1 ( P s ) = w q ( s ) w_1(P_s)=w^{}_{q(s)} w 1 ( P s ) = w q ( s ) , for each s ∈ Rigid ( q ) s\in\operatorname{Rigid}(q) s ∈ Rigid ( q ) . On D \mathcal{D} D , the weight of each s ∈ Flex ( q ) s\in\operatorname{Flex}(q) s ∈ Flex ( q ) is L L L and of each s ∈ Rigid ( q ) s\in\operatorname{Rigid}(q) s ∈ Rigid ( q ) is w q ( s ) R w^{\mathrm{R}}_{q(s)} w q ( s ) R ; thus, w ( D ) = M ℓ L k ∏ s ∈ Rigid ( q ) w q ( s ) R w(\mathcal{D})=M^\ell L^k\prod_{s\in\operatorname{Rigid}(q)}w^{\mathrm{R}}_{q(s)} w ( D ) = M ℓ L k ∏ s ∈ Rigid ( q ) w q ( s ) R . Subtracting, using D ⊆ ∏ s P s \mathcal{D}\subseteq\prod_sP_s D ⊆ ∏ s P s and (17), we obtain
w ( E ( q ) ) = M ℓ L k [ ∏ s ∈ Rigid ( q ) w q ( s ) − ∏ s ∈ Rigid ( q ) w q ( s ) R ] . (20) w(\mathcal{E}(q))=M^\ell L^k\left[\prod_{s\in\operatorname{Rigid}(q)}w_{q(s)}-\prod_{s\in\operatorname{Rigid}(q)}w^{\mathrm{R}}_{q(s)}\right]. \tag*{(20)} w ( E ( q )) = M ℓ L k s ∈ Rigid ( q ) ∏ w q ( s ) − s ∈ Rigid ( q ) ∏ w q ( s ) R . (20)
So far we assumed c σ ∈ F σ 1 c_\sigma\in F^1_\sigma c σ ∈ F σ 1 , so the weight w 1 w_1 w 1 applies. We now check that c σ ∈ F σ ≥ 2 c_\sigma\in F_{\sigma}^{\ge2} c σ ∈ F σ ≥ 2 can only increase w ( E ( q ) ) w(\mathcal{E}(q)) w ( E ( q )) , so that (18) stays valid as a lower bound. Recall from (11) that, if c σ ∈ F σ ≥ 2 c_\sigma\in F_{\sigma}^{\ge2} c σ ∈ F σ ≥ 2 , then the component p ( c σ ) p(c_\sigma) p ( c σ ) uses w 1 ∙ w_1^\bullet w 1 ∙ in place of w 1 w_1 w 1 , all others are unchanged.
Case c σ ∈ Flex ( q ) c_\sigma\in\operatorname{Flex}(q) c σ ∈ Flex ( q ) . Then its factor changes from w 1 ( P c σ ) = L w_1(P_{c_\sigma})=L w 1 ( P c σ ) = L to w 1 ∙ ( P c σ ) : = w 1 ∙ ( Π M F ) ≥ w 1 ( Π M F ) = L w_1^\bullet(P_{c_\sigma}):=w_1^\bullet(\Pi^{\mathrm{F}}_M)\ge w_1(\Pi^{\mathrm{F}}_M)=L w 1 ∙ ( P c σ ) := w 1 ∙ ( Π M F ) ≥ w 1 ( Π M F ) = L by (10). This same factor multiplies both w ( ∏ s P s ) w(\prod_sP_s) w ( ∏ s P s ) and w ( D ) w(\mathcal{D}) w ( D ) , and so (18) holds as a lower bound.
Case c σ ∈ Rigid ( q ) c_\sigma\in\operatorname{Rigid}(q) c σ ∈ Rigid ( q ) . Consider the products over the remaining rigid components (empty products = 1 =1 = 1 when Rigid ( q ) = { c σ } \operatorname{Rigid}(q)=\{c_\sigma\} Rigid ( q ) = { c σ } ):
w = ∏ s ∈ Rigid ( q ) ∖ { c σ } w q ( s ) , w R = ∏ s ∈ Rigid ( q ) ∖ { c σ } w q ( s ) R ; w=\prod_{s\in\operatorname{Rigid}(q)\setminus\{c_\sigma\}}w_{q(s)},\qquad w^{\mathrm{R}}=\prod_{s\in\operatorname{Rigid}(q)\setminus\{c_\sigma\}}w^{\mathrm{R}}_{q(s)}; w = s ∈ Rigid ( q ) ∖ { c σ } ∏ w q ( s ) , w R = s ∈ Rigid ( q ) ∖ { c σ } ∏ w q ( s ) R ;
then w ≥ w R ≥ 0 w\ge w^{\mathrm{R}}\ge0 w ≥ w R ≥ 0 since w q ( s ) ≥ w q ( s ) R ≥ 0 w_{q(s)}\ge w^{\mathrm{R}}_{q(s)}\ge0 w q ( s ) ≥ w q ( s ) R ≥ 0 , for each s ∈ Rigid ( q ) s\in\operatorname{Rigid}(q) s ∈ Rigid ( q ) . The k k k coordinates in Flex ( q ) \operatorname{Flex}(q) Flex ( q ) contribute L k L^k L k . Partition E ( q ) \mathcal{E}(q) E ( q ) by the value of p ( c σ ) p(c_\sigma) p ( c σ ) :
if p ( c σ ) ∈ Π M R p(c_\sigma)\in\Pi^{\mathrm{R}}_M p ( c σ ) ∈ Π M R , then, by (iii), another rigid component must be flexible, so those ( h − k − 1 ) (h-k-1) ( h − k − 1 ) components contribute w − w R w-w^{\mathrm{R}} w − w R , and p ( c σ ) p(c_\sigma) p ( c σ ) contributes w R ∙ = w 1 ∙ ( q ( c σ ) ) w^{\mathrm{R}\bullet}=w_1^\bullet(q(c_\sigma)) w R ∙ = w 1 ∙ ( q ( c σ )) , giving weight M ℓ L k w R ∙ ( w − w R ) ≥ M ℓ L k w q ( c σ ) R ( w − w R ) M^\ell L^k w^{\mathrm{R}\bullet}(w-w^{\mathrm{R}})\ge M^\ell L^k w^{\mathrm{R}}_{q(c_\sigma)}(w-w^{\mathrm{R}}) M ℓ L k w R ∙ ( w − w R ) ≥ M ℓ L k w q ( c σ ) R ( w − w R ) ;
if q ( c σ ) ⊆ p ( c σ ) ∈ Π M F q(c_\sigma)\subseteq p(c_\sigma)\in\Pi^{\mathrm{F}}_M q ( c σ ) ⊆ p ( c σ ) ∈ Π M F , then, by (iii), the other ( h − k − 1 ) (h-k-1) ( h − k − 1 ) components range without constraints (w w w ), and p ( c σ ) p(c_\sigma) p ( c σ ) contributes w F ∙ = ∑ J ∈ Π M F , q ( c σ ) ⊆ J w 1 ∙ ( J ) w^{\mathrm{F}\bullet}=\sum_{J\in\Pi^{\mathrm{F}}_M,\ q(c_\sigma)\subseteq J}w_1^\bullet(J) w F ∙ = ∑ J ∈ Π M F , q ( c σ ) ⊆ J w 1 ∙ ( J ) , giving weight M ℓ L k w F ∙ w ≥ M ℓ L k w q ( c σ ) F w M^\ell L^k w^{\mathrm{F}\bullet}w\ge M^\ell L^k w^{\mathrm{F}}_{q(c_\sigma)}w M ℓ L k w F ∙ w ≥ M ℓ L k w q ( c σ ) F w .
Hence, we obtain the following lower bound for w ( E ( q ) ) w(\mathcal{E}(q)) w ( E ( q )) , cf. (18):
w ( E ( q ) ) ≥ M ℓ L k [ w q ( c σ ) R ( w − w R ) + w q ( c σ ) F w ] = M ℓ L k [ ( w q ( c σ ) R + w q ( c σ ) F ) w − w q ( c σ ) R w R ] = M ℓ L k [ ∏ s ∈ Rigid ( q ) w q ( s ) − ∏ s ∈ Rigid ( q ) w q ( s ) R ] . \begin{aligned}
w(\mathcal{E}(q))&\ge M^\ell L^k\left[w^{\mathrm{R}}_{q(c_\sigma)}(w-w^{\mathrm{R}})+w^{\mathrm{F}}_{q(c_\sigma)}w\right]\\
&=M^\ell L^k\left[\left(w^{\mathrm{R}}_{q(c_\sigma)}+w^{\mathrm{F}}_{q(c_\sigma)}\right)w-w^{\mathrm{R}}_{q(c_\sigma)}w^{\mathrm{R}}\right]\\
&=M^\ell L^k\left[\prod_{s\in\operatorname{Rigid}(q)}w_{q(s)}-\prod_{s\in\operatorname{Rigid}(q)}w^{\mathrm{R}}_{q(s)}\right].
\end{aligned} w ( E ( q )) ≥ M ℓ L k [ w q ( c σ ) R ( w − w R ) + w q ( c σ ) F w ] = M ℓ L k [ ( w q ( c σ ) R + w q ( c σ ) F ) w − w q ( c σ ) R w R ] = M ℓ L k s ∈ Rigid ( q ) ∏ w q ( s ) − s ∈ Rigid ( q ) ∏ w q ( s ) R .
Observe now that, for any w i w_i w i and w i R w_i^{\mathrm{R}} w i R with w i ≥ w i R ≥ 0 w_i\ge w_i^{\mathrm{R}}\ge0 w i ≥ w i R ≥ 0 , we have the following:
∏ i = 1 r w i − ∏ i = 1 r w i R = ∑ j = 1 r [ ( ∏ i < j w i ) ⋅ ( w j − w j R ) ⋅ ( ∏ i > j w i R ) ] ≥ ( ∏ i < r w i ) ⋅ ( w r − w r R ) ; \prod_{i=1}^{r}w_i-\prod_{i=1}^{r}w_i^{\mathrm{R}}=\sum_{j=1}^{r}\left[\left(\prod_{i<j}w_i\right)\cdot(w_j-w_j^{\mathrm{R}})\cdot\left(\prod_{i>j}w_i^{\mathrm{R}}\right)\right]\ge\left(\prod_{i<r}w_i\right)\cdot(w_r-w_r^{\mathrm{R}}); i = 1 ∏ r w i − i = 1 ∏ r w i R = j = 1 ∑ r [ ( i < j ∏ w i ) ⋅ ( w j − w j R ) ⋅ ( i > j ∏ w i R ) ] ≥ ( i < r ∏ w i ) ⋅ ( w r − w r R ) ;
the inequality holds because the summands are non-negative, and so the last one does not exceed the total. Also, for every { λ } ∈ Π M R \{\lambda\}\in\Pi^{\mathrm{R}}_M { λ } ∈ Π M R , we have
w { λ } R = ( M λ ) , w { λ } F ≥ M , w { λ } ≥ ( M 2 ) . w^{\mathrm{R}}_{\{\lambda\}}=\binom{M}{\lambda},\qquad w^{\mathrm{F}}_{\{\lambda\}}\ge M,\qquad w_{\{\lambda\}}\ge\binom{M}{2}. w { λ } R = ( λ M ) , w { λ } F ≥ M , w { λ } ≥ ( 2 M ) .
Indeed [ 2 , ∞ ) [2,\infty) [ 2 , ∞ ) intersects with every { λ } \{\lambda\} { λ } , so it contributes to w { λ } F w^{\mathrm{F}}_{\{\lambda\}} w { λ } F , whence w { λ } F ≥ w 1 ( [ 2 , ∞ ) ) = M w^{\mathrm{F}}_{\{\lambda\}} \ge w_1([2,\infty)) = M w { λ } F ≥ w 1 ([ 2 , ∞ )) = M by (9).
For λ = 2 \lambda=2 λ = 2 , w { 2 } = w { 2 } F + w { 2 } F ≥ w { 2 } R = ( M 2 ) w_{\{2\}}=w^{\mathrm{F}}_{\{2\}}+w^{\mathrm{F}}_{\{2\}}\ge w^{\mathrm{R}}_{\{2\}}=\binom{M}{2} w { 2 } = w { 2 } F + w { 2 } F ≥ w { 2 } R = ( 2 M ) . For λ ≥ 3 \lambda\ge3 λ ≥ 3 , w { λ } F ≥ w 1 ( [ 3 , ∞ ) ) w^{\mathrm{F}}_{\{\lambda\}}\ge w_1([3,\infty)) w { λ } F ≥ w 1 ([ 3 , ∞ )) . Hence w { λ } ≥ 2 ( M 2 ) ≥ ( M 2 ) w_{\{\lambda\}}\ge2\binom{M}{2}\ge\binom{M}{2} w { λ } ≥ 2 ( 2 M ) ≥ ( 2 M ) .
With these observations at hand, we derive the required lower bound for w ( E ( q ) ) w(\mathcal{E}(q)) w ( E ( q )) :
w ( E ( q ) ) ≥ M ℓ L k [ ∏ s ∈ Rigid ( q ) w q ( s ) − ∏ s ∈ Rigid ( q ) w q ( s ) R ] ≥ M ℓ L k [ ( ∏ s ∈ Rigid ( q ) ∖ { s 0 } w q ( s ) ) ⋅ ( w q ( s 0 ) − w q ( s 0 ) R ) ] ≥ M ℓ L k ⋅ ( M 2 ) h − k − 1 ⋅ M . w(\mathcal{E}(q)) \ge M^{\ell}L^k\left[\prod_{s\in\operatorname{Rigid}(q)}w_{q(s)}-\prod_{s\in\operatorname{Rigid}(q)}w^{\mathrm{R}}_{q(s)}\right]\ge M^{\ell}L^k\left[\left(\prod_{s\in\operatorname{Rigid}(q)\setminus\{s_0\}}w_{q(s)}\right)\cdot\left(w_{q(s_0)}-w^{\mathrm{R}}_{q(s_0)}\right)\right]\ge M^{\ell}L^k\cdot\left(\frac{M}{2}\right)^{h-k-1}\cdot M. w ( E ( q )) ≥ M ℓ L k s ∈ Rigid ( q ) ∏ w q ( s ) − s ∈ Rigid ( q ) ∏ w q ( s ) R ≥ M ℓ L k s ∈ Rigid ( q ) ∖ { s 0 } ∏ w q ( s ) ⋅ ( w q ( s 0 ) − w q ( s 0 ) R ) ≥ M ℓ L k ⋅ ( 2 M ) h − k − 1 ⋅ M .
where s 0 s_0 s 0 is any element of Rigid ( q ) \operatorname{Rigid}(q) Rigid ( q ) and ∣ Rigid ( q ) ∣ = h − k |\operatorname{Rigid}(q)|=h-k ∣ Rigid ( q ) ∣ = h − k . This finishes the proof of Claim 57.2. □ \square □
Claims 57.1 and 57.2 give
g ( Q ) ≥ M ℓ + 1 L k ( M 2 ) h − k − 1 . g(Q)\ge M^{\ell+1}L^k\left(\frac{M}{2}\right)^{h-k-1}. g ( Q ) ≥ M ℓ + 1 L k ( 2 M ) h − k − 1 .
Since M = 2 m M=2^m M = 2 m with m ≥ 1 m\ge1 m ≥ 1 , we have 2 m ≥ 2 m 2^m\ge2m 2 m ≥ 2 m (true for m = 1 , 2 m=1,2 m = 1 , 2 , then inductively from 2 m + 1 ≥ 4 m ≥ 2 ( m + 1 ) 2^{m+1}\ge4m\ge2(m+1) 2 m + 1 ≥ 4 m ≥ 2 ( m + 1 ) ), which is exactly 2 M ≥ M 2 2^M\ge M^2 2 M ≥ M 2 , and so L = 3 M − 2 M ≥ 2 M ≥ M 2 L=3^M-2^M\ge2^M\ge M^2 L = 3 M − 2 M ≥ 2 M ≥ M 2 . Also, since M ( M − 1 ) 2 ≥ M 2 4 \frac{M(M-1)}{2}\ge\frac{M^2}{4} 2 M ( M − 1 ) ≥ 4 M 2 , we have, using k ≥ 0 k\ge0 k ≥ 0 ,
g ( Q ) ≥ M ℓ + 1 M 2 k ( M 2 4 ) h − k − 1 = 1 4 h − k − 1 M ℓ + 2 h − 1 ≥ 1 4 h − 1 M ℓ + 2 h − 1 . g(Q)\ge M^{\ell+1}M^{2k}\left(\frac{M^2}{4}\right)^{h-k-1}=\frac{1}{4^{h-k-1}}M^{\ell+2h-1}\ge\frac{1}{4^{h-1}}M^{\ell+2h-1}. g ( Q ) ≥ M ℓ + 1 M 2 k ( 4 M 2 ) h − k − 1 = 4 h − k − 1 1 M ℓ + 2 h − 1 ≥ 4 h − 1 1 M ℓ + 2 h − 1 .
Theorem 58
Theorem 58. For every fixed n n n , lim m → ∞ D C r a i g n , m = 1 \lim_{m\to\infty}\mathrm{DCraig}_{n,m}=1 lim m → ∞ DCraig n , m = 1 .
Proof. Fix σ = { p 1 , … , p n } \sigma=\{p_1,\ldots,p_n\} σ = { p 1 , … , p n } and let ρ m = { q 1 , … , q m } \rho_m=\{q_1,\ldots,q_m\} ρ m = { q 1 , … , q m } for m ≥ 1 m\ge1 m ≥ 1 . Observe that for any σ ↑ ρ m \sigma^{\uparrow\rho_m} σ ↑ ρ m -cell C C C , h = ∣ F σ 2 ∣ h=|F^2_\sigma| h = ∣ F σ 2 ∣ and ℓ = ∣ F σ 1 ∣ \ell=|F^1_\sigma| ℓ = ∣ F σ 1 ∣ are determined by the restriction F σ F_\sigma F σ and do not depend on m m m . By the cell reduction principle for Craig interpolation (Theorem 21 ), it then suffices to show that, for all σ ↑ ρ m \sigma^{\uparrow\rho_m} σ ↑ ρ m -cells C C C with fixed h h h and ℓ \ell ℓ , we have 1 − D C r a i g C ≤ f h , ℓ ( n , m ) 1-\mathrm{DCraig}_C\le f_{h,\ell}(n,m) 1 − DCraig C ≤ f h , ℓ ( n , m ) for a function f h , ℓ ( n , m ) f_{h,\ell}(n,m) f h , ℓ ( n , m ) converging to 0 0 0 for n n n fixed and m → ∞ m\to\infty m → ∞ .
We make a case distinction according to the value of ( h , ℓ ) (h,\ell) ( h , ℓ ) .
*Case h ≥ 1 h\ge1 h ≥ 1 and ( h , ℓ ) ≠ ( 1 , 0 ) (h,\ell)\ne(1,0) ( h , ℓ ) = ( 1 , 0 ) . By Lemma 56, every ordered biclique in P P P can be oriented so that the first side is in S S S . Now Lemma 55, Theorem 57, and ∣ P ∣ ≤ ( 3 M ) h |P|\le(3M)^h ∣ P ∣ ≤ ( 3 M ) h (Lemma 54) give
1 − D C r a i g C ≤ 2 1 + ( 3 M ) h − 1 4 h − 1 M ℓ + 2 h − 1 . (21) 1-\mathrm{DCraig}_C\le2^{1+(3M)^h-\frac{1}{4^{h-1}}M^{\ell+2h-1}}. \tag*{(21)} 1 − DCraig C ≤ 2 1 + ( 3 M ) h − 4 h − 1 1 M ℓ + 2 h − 1 . (21)
For fixed h , ℓ h,\ell h , ℓ in this range,
1 4 h − 1 M ℓ + 2 h − 1 ( 3 M ) h = 1 4 h − 1 3 h M ℓ + h − 1 ⟶ ∞ , \frac{\frac{1}{4^{h-1}}M^{\ell+2h-1}}{(3M)^h}=\frac{1}{4^{h-1}3^h}M^{\ell+h-1}\longrightarrow\infty, ( 3 M ) h 4 h − 1 1 M ℓ + 2 h − 1 = 4 h − 1 3 h 1 M ℓ + h − 1 ⟶ ∞ ,
because ℓ + h − 1 ≥ 1 \ell+h-1\ge1 ℓ + h − 1 ≥ 1 . Thus (19) tends to 0 0 0 , as required.
*Case ( h , ℓ ) = ( 1 , 0 ) (h,\ell)=(1,0) ( h , ℓ ) = ( 1 , 0 ) . Then there is exactly one s ∈ F σ 2 s\in F^2_\sigma s ∈ F σ 2 and F σ 1 = ∅ F^1_\sigma=\varnothing F σ 1 = ∅ . A rigid box-profile takes the form { λ } \{\lambda\} { λ } for 2 ≤ λ ≤ M 2\le\lambda\le M 2 ≤ λ ≤ M . Hence any Q ∈ S Q\in S Q ∈ S is a non-empty subset of { { 2 } , … , { M } } \{\{2\},\ldots,\{M\}\} {{ 2 } , … , { M }} .
Let m 0 m_0 m 0 be its largest value. Then N ‾ P ( Q ) = Q ∪ { [ μ , ∞ ) ∣ 2 ≤ μ ≤ m 0 } \overline{N}_P(Q)=Q\cup\{[\mu,\infty)\mid2\le\mu\le m_0\} N P ( Q ) = Q ∪ {[ μ , ∞ ) ∣ 2 ≤ μ ≤ m 0 } . Put
Λ ∙ ( m 0 ) = ∑ μ = 2 m 0 w 1 ∙ ( [ μ , ∞ ) ) , Λ ( m 0 ) = ∑ μ = 2 m 0 w 1 ( [ μ , ∞ ) ) . \Lambda^\bullet(m_0)=\sum_{\mu=2}^{m_0}w^\bullet_1([\mu,\infty)),\qquad\Lambda(m_0)=\sum_{\mu=2}^{m_0}w_1([\mu,\infty)). Λ ∙ ( m 0 ) = μ = 2 ∑ m 0 w 1 ∙ ([ μ , ∞ )) , Λ ( m 0 ) = μ = 2 ∑ m 0 w 1 ([ μ , ∞ )) .
Then g ( Q ) = Λ ∙ ( m 0 ) ≥ Λ ( m 0 ) g(Q)=\Lambda^\bullet(m_0)\ge\Lambda(m_0) g ( Q ) = Λ ∙ ( m 0 ) ≥ Λ ( m 0 ) . There are exactly 2 m 0 − 2 2^{m_0-2} 2 m 0 − 2 non-empty sets Q Q Q of rigid box-profiles with largest value m 0 m_0 m 0 . Applying (13) and dividing by 2 ∣ C ∣ 2^{|C|} 2 ∣ C ∣ gives
1 − D C r a i g C ≤ 2 ∑ m 0 = 2 M 2 m 0 − 2 − Λ ∙ ( m 0 ) ≤ 2 ∑ m 0 = 2 M 2 m 0 − 2 − Λ ( m 0 ) . (22) 1-\mathrm{DCraig}_C\le2\sum_{m_0=2}^{M}2^{m_0-2-\Lambda^\bullet(m_0)}\le2\sum_{m_0=2}^{M}2^{m_0-2-\Lambda(m_0)}. \tag*{(22)} 1 − DCraig C ≤ 2 m 0 = 2 ∑ M 2 m 0 − 2 − Λ ∙ ( m 0 ) ≤ 2 m 0 = 2 ∑ M 2 m 0 − 2 − Λ ( m 0 ) . (22)
Now Λ ( 2 ) = M \Lambda(2)=M Λ ( 2 ) = M by (9) and for m 0 ≥ 3 m_0\ge3 m 0 ≥ 3 one can easily show Λ ( m 0 ) ≥ M + M ( M − 1 ) = M 2 \Lambda(m_0)\ge M+M(M-1)=M^2 Λ ( m 0 ) ≥ M + M ( M − 1 ) = M 2 . Consequently
1 − D C r a i g C ≤ 2 ( 2 − M + ( M − 2 ) 2 M − 2 − M 2 ) ⟶ 0 , (23) 1-\mathrm{DCraig}_C\le2\left(2^{-M}+(M-2)2^{M-2-M^2}\right)\longrightarrow0, \tag*{(23)} 1 − DCraig C ≤ 2 ( 2 − M + ( M − 2 ) 2 M − 2 − M 2 ) ⟶ 0 , (23)
as required.
*Case h = 0 h=0 h = 0 . Then F σ 2 = ∅ F^2_\sigma=\varnothing F σ 2 = ∅ and any two atoms in C C C are compatible. It follows that D C r a i g C = 1 \mathrm{DCraig}_C=1 DCraig C = 1 . □ \square □
Density in Diff after Quantification
We now consider the densities UnifInt n , m Π \operatorname{UnifInt}^{\Pi}_{n,m} UnifInt n , m Π and DCraig n , m Π \operatorname{DCraig}^{\Pi}_{n,m} DCraig n , m Π . We show that, for fixed n n n , both converge to 0 0 0 as m → ∞ m \to\infty m → ∞ when equivalence after existential quantification is defined over difference frames. Equivalence of existentially quantified modal formulas over K ≠ \mathcal{K}_{\ne} K = does not imply equivalence over K D i f f \mathcal{K}_{\mathrm{Diff}} K Diff , as the following example illustrates.
Example 12. Consider the formulas φ = ◊ ≥ 2 q ∧ ◊ ◊ = 1 ¬ q \varphi= \Diamond^{\geq2}q \land\Diamond\Diamond^{=1}\neg q φ = ◊ ≥ 2 q ∧ ◊ ◊ = 1 ¬ q and ψ = ◊ = 1 q ∧ ◊ ◊ ≥ 2 ¬ q \psi= \Diamond^{=1}q \land\Diamond\Diamond^{\geq2}\neg q ψ = ◊ = 1 q ∧ ◊ ◊ ≥ 2 ¬ q , where ◊ ≥ 2 χ = χ ∧ ◊ χ \Diamond^{\geq2}\chi= \chi\land\Diamond\chi ◊ ≥ 2 χ = χ ∧ ◊ χ and ◊ = 1 χ = χ ∧ ¬ ◊ χ \Diamond^{=1}\chi= \chi\land\neg\Diamond\chi ◊ = 1 χ = χ ∧ ¬ ◊ χ . Over K ≠ \mathcal{K}_{\ne} K = , the formulas ∃ q . φ \exists q.\varphi ∃ q . φ and ∃ q . ψ \exists q.\psi ∃ q . ψ are equivalent, since both hold precisely on pointed difference frames with at least three worlds. By contrast, let W = { 0 , 1 } W = \{0,1\} W = { 0 , 1 } and R = { ( 0 , 0 ) , ( 0 , 1 ) , ( 1 , 0 ) } R = \{(0,0),(0,1),(1,0)\} R = {( 0 , 0 ) , ( 0 , 1 ) , ( 1 , 0 )} . Then ( W , R ) (W,R) ( W , R ) validates Diff, but ( W , R ) , 0 ⊨ ∃ q . φ (W,R),0 \models\exists q.\varphi ( W , R ) , 0 ⊨ ∃ q . φ and ( W , R ) , 0 ⊭ ∃ q . ψ (W,R),0 \not\models\exists q.\psi ( W , R ) , 0 ⊨ ∃ q . ψ . Thus the two formulas are not equivalent over K D i f f \mathcal{K}_{\mathrm{Diff}} K Diff .
Hence, the choice between K ≠ \mathcal{K}_{\ne} K = and K D i f f \mathcal{K}_{\mathrm{Diff}} K Diff affects the counting of equivalence classes after existential quantification. Since difference frames provide the intended semantics for Diff, we restrict attention to this class throughout the section.
To show converge to 0 0 0 as m → ∞ m \to\infty m → ∞ , by the projective cell reduction principles, it suffices to define a sequence C m C_m C m of σ ↑ ρ m \sigma^{\uparrow\rho_m} σ ↑ ρ m -cells whose densities tend to 0 0 0 .
Fix σ \sigma σ and a σ \sigma σ -colour c 0 c_0 c 0 and let ( F 0 , c 0 ) (F_0,c_0) ( F 0 , c 0 ) be the σ \sigma σ -atom with F 0 ( c 0 ) = ≥ 2 F_0(c_0) = \geq2 F 0 ( c 0 ) =≥ 2 and F 0 ( t ) = 0 F_0(t) = 0 F 0 ( t ) = 0 for all t ≠ c 0 t \ne c_0 t = c 0 . For m ≥ 1 m \geq1 m ≥ 1 , let C m C_m C m be the σ ↑ ρ m \sigma^{\uparrow\rho_m} σ ↑ ρ m -cell with σ \sigma σ -reduct ( F 0 , c 0 ) (F_0,c_0) ( F 0 , c 0 ) . In the notation above, this cell has ( h , ℓ ) = ( 1 , 0 ) (h,\ell) = (1,0) ( h , ℓ ) = ( 1 , 0 ) . For the asymptotic argument we may assume m ≥ 1 m \geq1 m ≥ 1 , so M = 2 m ≥ 2 M = 2^m \geq2 M = 2 m ≥ 2 .
For D ⊆ C m D \subseteq C_m D ⊆ C m , the formula ∃ ρ ⃗ m . Φ ( D ) \exists\vec{\rho}_m.\Phi(D) ∃ ρ m .Φ ( D ) holds in a pointed σ \sigma σ -model if all worlds have colour c 0 c_0 c 0 and its cardinality satisfies the constraint ∃ ρ ⃗ m . Φ ( M ) \exists\vec{\rho}_m.\Phi(M) ∃ ρ m .Φ ( M ) associated with some atom M ∈ D \mathcal{M} \in D M ∈ D . As each constraint ∃ ρ ⃗ m . Φ ( M ) \exists\vec{\rho}_m.\Phi(\mathcal{M}) ∃ ρ m .Φ ( M ) belongs to Π M \Pi_{\mathcal{M}} Π M , in K ≠ \mathcal{K}_{\ne} K = the equivalence classes in Q C m Q_{C_m} Q C m correspond to unions of sets in Π M \Pi_{\mathcal{M}} Π M . Let S M S_{\mathcal{M}} S M be the set of these unions.
Lemma 59. In K ≠ \mathcal{K}_{\ne} K = , ∣ S M ∣ = ∣ Q C m ∣ = ( M + 1 ) 2 M − 1 |S_{\mathcal{M}}| = |Q_{C_m}| = (M+1)2^{M-1} ∣ S M ∣ = ∣ Q C m ∣ = ( M + 1 ) 2 M − 1 .
Proof. A union U ∈ S M U \in S_{\mathcal{M}} U ∈ S M containing no interval of the form [ λ , ∞ ) [\lambda,\infty) [ λ , ∞ ) is an arbitrary subset of { 2 , … , M } \{2,\ldots,M\} { 2 , … , M } , giving 2 M − 1 2^{M-1} 2 M − 1 possibilities. A union U U U containing an interval [ λ , ∞ ) [\lambda,\infty) [ λ , ∞ ) has a least threshold s ∈ { 2 , … , 2 M } s \in\{2,\ldots,2M\} s ∈ { 2 , … , 2 M } such that [ s , ∞ ) ⊆ U [s,\infty) \subseteq U [ s , ∞ ) ⊆ U . Then s − 1 ∉ U s-1 \notin U s − 1 ∈ / U , while points below s − 1 s-1 s − 1 may be chosen freely. The thresholds 2 ≤ s ≤ M + 1 2 \leq s \leq M+1 2 ≤ s ≤ M + 1 therefore give
1 + 1 + ∑ s = 4 M + 1 2 s − 3 = 2 M − 1 1+1+\sum_{s=4}^{M+1}2^{s-3}=2^{M-1} 1 + 1 + s = 4 ∑ M + 1 2 s − 3 = 2 M − 1
unions. Each of the remaining M − 1 M-1 M − 1 thresholds allows an arbitrary subset of { 2 , … , M } \{2,\ldots,M\} { 2 , … , M } and hence gives 2 M − 1 2^{M-1} 2 M − 1 unions. Thus there are M 2 M − 1 M2^{M-1} M 2 M − 1 unions containing an interval of the form [ λ , ∞ ) [\lambda,\infty) [ λ , ∞ ) , and the claim follows. □ \square □
Proposition 60. For the cell C m C_m C m ,
DCraig C m Π ≤ 3 ( M + 1 ) ( 2 3 ) M , UnifInt C m Π = 2 ( M + 1 ) 2 M − 1 . \operatorname{DCraig}^{\Pi}_{C_m} \leq3(M+1)\left(\frac{2}{3}\right)^M,\qquad\operatorname{UnifInt}^{\Pi}_{C_m} = \frac{2}{(M+1)2^{M-1}}. DCraig C m Π ≤ 3 ( M + 1 ) ( 3 2 ) M , UnifInt C m Π = ( M + 1 ) 2 M − 1 2 .
Proof. By Theorem 37 , ∣ Craig C m Π ∣ = 2 ∣ S M ∣ − 1 |\operatorname{Craig}^{\Pi}_{C_m}| = 2|S_{\mathcal{M}}|-1 ∣ Craig C m Π ∣ = 2∣ S M ∣ − 1 . All pairs of disjoint non-empty subsets of the M − 1 M-1 M − 1 -element set { 2 , … , M } \{2,\ldots,M\} { 2 , … , M } correspond to inconsistent pairs of formulas. There are 3 M − 1 − 2 ⋅ 2 M − 1 + 1 3^{M-1}-2\cdot2^{M-1}+1 3 M − 1 − 2 ⋅ 2 M − 1 + 1 such pairs. Consequently
∣ InCon C m Π ∣ ≥ ( 2 ∣ S M ∣ − 1 ) + 3 M − 1 − 2 ⋅ 2 M − 1 + 1 = 3 M − 1 + M 2 M . |\operatorname{InCon}_{C_m}^{\Pi}| \geq(2|S_{\mathcal{M}}|-1)+3^{M-1}-2\cdot2^{M-1}+1 = 3^{M-1}+M2^M. ∣ InCon C m Π ∣ ≥ ( 2∣ S M ∣ − 1 ) + 3 M − 1 − 2 ⋅ 2 M − 1 + 1 = 3 M − 1 + M 2 M .
Therefore, using Lemma 59,
DCraig C m Π ≤ 2 ∣ S M ∣ − 1 3 M − 1 + M 2 M ≤ ( M + 1 ) 2 M 3 M − 1 = 3 ( M + 1 ) ( 2 3 ) M . \operatorname{DCraig}^{\Pi}_{C_m} \leq\frac{2|S_{\mathcal{M}}|-1}{3^{M-1}+M2^M} \leq\frac{(M+1)2^M}{3^{M-1}} = 3(M+1)\left(\frac{2}{3}\right)^M. DCraig C m Π ≤ 3 M − 1 + M 2 M 2∣ S M ∣ − 1 ≤ 3 M − 1 ( M + 1 ) 2 M = 3 ( M + 1 ) ( 3 2 ) M .
For uniform interpolation, the bottom and top elements of Q C m Q_{C_m} Q C m have uniform σ \sigma σ -interpolants. We show that no other element of Q C m Q_{C_m} Q C m does. If ∅ ≠ U ⊊ { 2 , 3 , … } \varnothing\ne U \subsetneq\{2,3,\ldots\} ∅ = U ⊊ { 2 , 3 , … } is represented by a union in S M S_{\mathcal{M}} S M , choose k ∉ U k \notin U k ∈ / U . Choose D ⊆ C m D \subseteq C_m D ⊆ C m representing U U U . Then ∃ ρ ⃗ m . Φ ( D ) \exists\vec{\rho}_m.\Phi(D) ∃ ρ m .Φ ( D ) holds precisely in the pointed models based on difference frames satisfying Φ ( ⋃ C m ) \Phi(\bigcup C_m) Φ ( ⋃ C m ) and whose number of worlds belongs to U U U . Take a formula ψ k \psi_k ψ k using symbols from ρ k \rho_k ρ k with ( σ ∪ ρ m ) ∩ ρ k ′ = ∅ (\sigma\cup\rho_m) \cap\rho_k' = \varnothing ( σ ∪ ρ m ) ∩ ρ k ′ = ∅ such that all pointed models based on difference frames satisfying ψ k \psi_k ψ k have exactly k k k worlds. By definition, Φ ( D ) ⊨ ¬ ψ k \Phi(D) \models\neg\psi_k Φ ( D ) ⊨ ¬ ψ k but Γ ( C m ) ⊭ ¬ ψ k \Gamma(C_m) \not\models\neg\psi_k Γ ( C m ) ⊨ ¬ ψ k for the cell characteristic formula Γ ( C m ) \Gamma(C_m) Γ ( C m ) . It follows that Γ ( C m ) \Gamma(C_m) Γ ( C m ) is not a uniform σ \sigma σ -interpolant of Φ ( D ) \Phi(D) Φ ( D ) , and so by Lemma 25 there is none.
Theorem 61. For every fixed n n n , lim m → ∞ D C r a i g n , m Π = lim m → ∞ U n i f I n t n , m Π = 0 \lim_{m\to\infty}\mathrm{DCraig}^{\Pi}_{n,m}=\lim_{m\to\infty}\mathrm{UnifInt}^{\Pi}_{n,m}=0 lim m → ∞ DCraig n , m Π = lim m → ∞ UnifInt n , m Π = 0 .
Proof. By the two projective cell reduction principles, Theorems 35 and 34, both densities are products of cell factors in [ 0 , 1 ] [0,1] [ 0 , 1 ] . Hence each product is bounded above by D C r a i g C m Π \mathrm{DCraig}^{\Pi}_{C_m} DCraig C m Π and U n i f I n t C m Π \mathrm{UnifInt}^{\Pi}_{C_m} UnifInt C m Π , respectively. Proposition 60 gives the two bounds, and both tend to zero.
Conclusion
We have obtained initial results on the density of formulas with Craig and uniform interpolants for quasi-normal modal logics. Many questions remain open. We begin with those arising directly from our results.
Is there any (quasi-)normal modal logic with lim m → ∞ D C r a i g n , m ≠ 1 \lim_{m\to\infty}\mathrm{DCraig}_{n,m}\ne1 lim m → ∞ DCraig n , m = 1 for some fixed n n n ? If no general results can be obtained, how do standard modal logics lacking CIP, such as K4.3, S4.3, weak K4, and product modal logics, behave?
Do we have lim m → ∞ D C r a i g n , m Π = 1 \lim_{m\to\infty}\mathrm{DCraig}^{\Pi}_{n,m}=1 lim m → ∞ DCraig n , m Π = 1 for the class of a a a frames validating G r z [ Δ ] \mathrm{Grz}[\Delta] Grz [ Δ ] when Δ = 3 \Delta=3 Δ = 3 or, more generally, when Δ ≥ 3 \Delta\ge3 Δ ≥ 3 ? How do other standard modal logics, such as those listed under Point 1, behave with respect to lim m → ∞ D C r a i g n , m Π \lim_{m\to\infty}\mathrm{DCraig}^{\Pi}_{n,m} lim m → ∞ DCraig n , m Π and lim m → ∞ U n i f I n t n , m Π \lim_{m\to\infty}\mathrm{UnifInt}^{\Pi}_{n,m} lim m → ∞ UnifInt n , m Π ?
Inspired by Maksimova’s classification of CIP for normal extensions of S4 [33 , 34 , 35 , 38 ] , can we also classify normal extensions of S4 according to the behaviour of lim m → ∞ D C r a i g n , m Π \lim_{m\to\infty}\mathrm{DCraig}^{\Pi}_{n,m} lim m → ∞ DCraig n , m Π and/or (uniform) BICIP?
In this article, we have always kept one parameter fixed. A natural next question is to determine phase transition windows when the retained and forgotten signatures grow simultaneously. For the logic of the 3-fork, this has been done in Examples 4 and 11, and our quantitative estimates also already provide partial information. For instance, for ordinary uniform interpolation, the density tends to 1 whenever ∣ A t σ n ∣ 2 − 2 m → 0 |\mathrm{At}_{\sigma_n}|2^{-2^m}\to0 ∣ At σ n ∣ 2 − 2 m → 0 , while it tends to 0 whenever 2 n 2 − ∣ C m ∣ → ∞ 2^{n2^{-|C_m|}}\to\infty 2 n 2 − ∣ C m ∣ → ∞ . Thus, the present results already delimit regions on either side of a possible phase transition, but they are generally not sharp enough to identify the window. Determining these regions more precisely and whether sharp thresholds exist for natural (classes of) modal logics remains an open problem.
Throughout, the modal depth bound d d d has been kept fixed. Allowing d d d to grow together with the signatures introduces a further asymptotic parameter. Note that the number and size of atoms and cells can depend strongly on d d d .
The results presented in this paper can easily be extended to quasi-normal modal logics with multiple modal operators of arbitrary arity. More care is needed for classical modal logics without the normality axiom. For these logics, all proofs still go through, except for the proof of Lemma 9 about pinning. Thus, the asymptotic behaviour of uniform and Craig interpolants when the number of retained variables grows remains open. On the other hand, it is easy to see that if the universal modality (a modal operator □ a \Box_a □ a such that w ⊨ □ a p w\models\Box_a p w ⊨ □ a p iff p p p is true in all worlds) is definable in a classical modal logic, then the lifting lemmas can be proved as before. This applies, for instance, to temporal modal logics such as linear temporal logic (LTL).
A critical property on which some of our proofs rely is that any propositional variable can be interpreted in the same way as ⊤ \top ⊤ or ⊥ \bot ⊥ and can be uniformly replaced by its negation without affecting satisfiability. This is not the case, for instance, for modal logics with nominals, and their asymptotic behaviour could therefore be rather different.
It would be interesting to investigate counting measures intermediate between ordinary counting and counting after existential quantification. For example, one could identify formulas up to permutations of variables and uniform replacement of variables by their negations. Such identifications could also be extended to the retained variables.
Moving towards non-modal logics, the techniques developed in this paper can be directly applied to some natural fragments of first-order logic without function symbols and lacking CIP, such as the 2-variable fragment [22 ] . Non-classical logics, such as super-intuitionistic and many-valued logics, might require rather different approaches [12 ] .