Model Theory of Monadic Predicate Logic with the Infinity Quantifier

A preservation result involves some fragment $\mathtt{L}_{\mathfrak{P}}$ of a given yardstick logic $\mathtt{L}$, related to a certain semantic property $\mathfrak{P}$. It is usually formulated as

In this work, our main yardstick logic will be $\mathtt{M}\mathtt{E}^{\infty}$. Table 1 summarises the semantic properties ($\mathfrak{P}$) we are going to consider, the corresponding expressively complete fragment ($\mathtt{L}_{\mathfrak{P}}$) and preservation theorem.

The proof of each preservation theorem is composed of two parts. The first, simpler one concerns the claim that each sentence in the fragment satisfies the concerned property. It is usually proved by induction on the structure of the sentence. The other direction is the expressive completeness statement, stating that within the considered logic, the fragment is expressively complete for the property. Its verification generally requires more effort. In this paper, we will actually verify a stronger expressive completeness statement. Namely, for each semantic property $\mathfrak{P}$ and corresponding fragment $\mathtt{L}_{\mathfrak{P}}$ from Table 1, we are going to provide an effective translation operation $(\cdot)^{\sf p}:\mathtt{M}\mathtt{E}^{\infty}\to\mathtt{L}_{\mathfrak{P}}$ such that

Since the satisfiability problem for $\mathtt{M}\mathtt{E}^{\infty}$ is decidable and the translation $(\cdot)^{\sf p}$ is effectively computable, we obtain, as an immediate corollary of (2), that for each property $\mathfrak{P}$ listed in Table 1

The proof of each instance of (2) will follow an uniform pattern, analogous to the one employed in the aim of obtaining similar results in the context of the modal $\mu$-calculus [15, 8, 11]. The crux of the adopted proof method is that, extending known results on monadic first-order logic, for each sentence $\varphi$ in $\mathtt{M}\mathtt{E}^{\infty}$ it is possible to compute a logically equivalent sentence in basic normal norm. Such normal forms will take the shape of a disjunction $\bigvee\nabla_{\mathtt{M}\mathtt{E}^{\infty}}$, where each disjunct $\nabla_{\mathtt{M}\mathtt{E}^{\infty}}$ characterises a class of models of $\varphi$ satisfying the same set of ${\mathtt{M}\mathtt{E}^{\infty}}$-sentences of equal quantifier rank as $\varphi$. Based on this, it will therefore be enough to define an effective translation $(\cdot)^{\sf p}$ for sentences in normal form, point-wise in each disjunct $\nabla_{\mathtt{M}\mathtt{E}^{\infty}}$, and then verify that it indeed satisfies (2).

As a corollary of the employed proof method, we thus obtain effective normal forms for sentences satisfying the considered property.

In addition to $\mathtt{M}\mathtt{E}^{\infty}$, in this paper we are also going to consider monadic first-order logic with and without equality, denoted respectively by $\mathtt{M}\mathtt{E}$ and $\mathtt{M}$. Table 2 shows a summary of the expressive completeness and normal form results presented in this paper.

Parity automata are finite-state systems playing a crucial role in obtaining decidability and expressiveness results in fixpoint logic (see e.g. [27]). They are specified by a finite set of states $A$, a distinguished, initial state $a\in A$, a function $\Omega$ assigning to each states a priority (a natural number), and a transition function $\Delta$ whose co-domain is usually given by a monadic logic in which the set of (monadic) predicates coincides with $A$. Hence, each monadic logic $\mathtt{L}$ induces its own class of automata $\mathtt{Aut}(\mathtt{L})$.

A landmark result in this area is Janin and Walukiewicz’s theorem stating that the bisimulation-invariant fragment of monadic second order logic coincides with the modal $\mu$-calculus [15], and the proof of this result is an interesting mix of the theory of parity automata and the model theory of monadic predicate logic. First, preservation and normal forms results are used to verify that (on tree models) $\mathtt{Aut}(\mathtt{Pos}(\mathtt{M}\mathtt{E}))$ is the class of automata characterising the expressive power of monadic second order logic [29], whereas $\mathtt{Aut}(\mathtt{Pos}(\mathtt{M}))$ corresponds to the modal $\mu$-calculus [14], where $\mathtt{Pos}(\mathtt{L})$ denote the positive fragment of the monadic logic $\mathtt{L}$. Then, Janin-Walukieiwcz’s expressiveness theorem is a consequence of these automata characterisations and the fact that positive monadic first-order logic without equality provides the quotient-invariant fragment of positive monadic first-order logic with equality (see Theorem 6.15).

In our companion paper [7], we provide a Janin-Walukiewicz type characterisation result for weak monadic second order logic. Our proof, analogously to the case of full monadic second order logic discussed previously, crucially employs preservation and normal form results for $\mathtt{M}\mathtt{E}^{\infty}$ listed in Tables 1 and 2.

Results in this paper first appeared in the first author’s PhD thesis ([5, Chapter 5]); this journal version largely expands material first published as part of the conference papers [10, 6]. In particular, the whole of Section 6 below contains new results.

We start by introducing a normal form for monadic first-order logic without equality.

Clearly the meaning of the formula $\nabla_{\mathtt{M}}(\Sigma)$ is that $\Sigma$ is a complete description of the collection of types that are realised in a monadic model. Notice that $\nabla_{\mathtt{M}}(\Sigma,\Pi)=\nabla_{\mathtt{M}}(\Sigma)=\forall x.\bot$, for $\Sigma=\Pi=\varnothing$.

This observation is easy to prove using Ehrenfeucht-Fraïssé games – proof sketches can be found in [12, Lemma 16.23] and [28, Proposition 4.14] –, and the decidability of the satisfiability problem for $\mathtt{M}$ (Fact 2.6). We omit a full proof because it is very similar to the following more complex cases.

When considering a normal form for $\mathtt{M}\mathtt{E}$, the fact that we can ‘count types’ using equality yields a more involved basic form.

We prove that every sentence of monadic first-order logic with equality is equivalent to a formula in basic form. Although this result seems to be folklore, we provide a detailed proof because some of its ingredients will be used later, when we give a normal form for $\mathtt{M}\mathtt{E}^{\infty}$. We start by defining the following relation between monadic models.

Intuitively, two models are related by $\sim^{=}_{k}$ when their type information coincides ‘modulo $k$’. Later on we prove that this is the same as saying that they cannot be distinguished by a sentence of $\mathtt{M}\mathtt{E}$ with quantifier rank at most $k$. As a special case, observe that any two monadic models are related by $\sim^{=}_{0}$.

For the moment, we record the following properties of these relations.

Proof. We only prove the second statement, and first we consider the case where $k=0$. The equivalence relation $\sim^{=}_{0}$ has the class $\mathfrak{M}$ of all monadic models as its unique equivalence class, so here we may define $\varphi^{=}_{\mathfrak{M}}\mathrel{:=}\top$.

From now on we assume that $k>0$. Let $E\in\mathfrak{M}/{\sim^{=}_{k}}$ and let $\mathbb{D}\in E$ be a representative. Call $S_{1},\dots,S_{n}\subseteq A$ to the types such that $|S_{i}|_{\mathbb{D}}=n_{i}<k$ and $S^{\prime}_{1},\dots,S^{\prime}_{m}\subseteq A$ to those satisfying $|S^{\prime}_{i}|_{\mathbb{D}}\geq k$. Note that the union of all the $S_{i}$ and $S^{\prime}_{i}$ yields all the possible $A$-types, and that if a type $S_{j}$ is not realised at all, we take $n_{j}=0$. Now define

where we understand that any conjunct of the form $\exists x_{1},\dots,x_{l}.\psi$ with $l=0$ is simply omitted (or, to the same effect, defined as $\top$). It is easy to see that $\mathtt{qr}(\varphi^{=}_{E})=k$ and that $\mathbb{D}^{\prime}\models\varphi^{=}_{E}$ iff $\mathbb{D}^{\prime}\in E$. Intuitively, $\varphi^{=}_{E}$ gives a specification of $E$ “type by type”; in particular observe that $\varphi^{=}_{\mathbb{D}_{\varnothing}}\equiv\forall x.\bot$. qed

Next we recall a (standard) notion of Ehrenfeucht-Fraïssé game for $\mathtt{M}\mathtt{E}$ which will be used to establish the connection between ${\sim^{=}_{k}}$ and $\equiv_{k}^{\mathtt{M}\mathtt{E}}$.

Proof. Step (1) to (2) is direct by Proposition 3.6. For (2) to (3) we give a winning strategy for $\exists$ in $\mathrm{EF}_{k}^{=}(\mathbb{D}_{0},\mathbb{D}_{1})$ by showing the following claim.

Proof of Claim Let $\forall$ pick $d_{i}\in D_{i}$ such that the type of $d_{i}$ is $T\subseteq A$. If $d_{i}$ had already been played then $\exists$ picks the same element as before and $f_{n+1}=f_{n}$. If $d_{i}$ is new and $|T|_{\mathbb{D}_{i}}\geq k$ then, as at most $n<k$ elements have been played, there is always some new $d_{-i}\in D_{-i}$ that $\exists$ can choose to match $d_{i}$. If $|T|_{\mathbb{D}_{i}}=m<k$ then we know that $|T|_{\mathbb{D}_{-i}}=m$. Therefore, as $d_{i}$ is new and $f_{n}$ is injective, there must be a $d_{-i}\in D_{-i}$ that $\exists$ can choose. $\blacktriangleleft$

Step (3) to (1) is a standard result [9, Corollary 2.2.9] which we prove anyway because we will need to extend it later. We prove the following loaded statement.

Proof of Claim If $\varphi$ is atomic the claim holds because of $f_{n}:\overline{{\mathbf{s_{0}}}}\mapsto\overline{{\mathbf{s_{1}}}}$ being a partial isomorphism. The Boolean cases are straightforward. Let $\varphi(z_{1},\dots,z_{n})=\exists x.\psi(z_{1},\dots,z_{n},x)$ and suppose $\mathbb{D}_{0}\models\varphi(\overline{{\mathbf{s_{0}}}})$. Hence, there exists $d_{0}\in D_{0}$ such that $\mathbb{D}_{0}\models\psi(\overline{{\mathbf{s_{0}}}},d_{0})$. By hypothesis we know that $\exists$ has a winning strategy for $\mathrm{EF}_{k}^{=}(\mathbb{D}_{0},\mathbb{D}_{1})@(\overline{{\mathbf{s_{0}}}},\overline{{\mathbf{s_{1}}}})$. Therefore, if $\forall$ picks $d_{0}\in D_{0}$ she can respond with some $d_{1}\in D_{1}$ and have a winning strategy for $\mathrm{EF}_{k}^{=}(\mathbb{D}_{0},\mathbb{D}_{1})@(\overline{{\mathbf{s_{0}}}}{\cdot}d_{0},\overline{{\mathbf{s_{1}}}}{\cdot}d_{1})$. By induction hypothesis, because $\mathtt{qr}(\psi)\leq k-(n+1)$, we have that $\mathbb{D}_{0}\models\psi(\overline{{\mathbf{s_{0}}}},d_{0})$ iff $\mathbb{D}_{1}\models\psi(\overline{{\mathbf{s_{1}}}},d_{1})$ and hence $\mathbb{D}_{1}\models\exists x.\psi(\overline{{\mathbf{s_{1}}}},x)$. The opposite direction is proved by a symmetric argument. $\blacktriangleleft$

Proof. Let $\mathtt{qr}(\psi)=k$ and let $\llbracket\psi\rrbracket$ be the class of models satisfying $\psi$. As $\mathfrak{M}/{\equiv_{k}^{\mathtt{M}\mathtt{E}}}$ is the same as $\mathfrak{M}/{\sim_{k}^{=}}$ by Proposition 3.8, it is easy to see that $\psi$ is equivalent to $\bigvee\{\varphi^{=}_{E}\mid E\in\llbracket\psi\rrbracket/{\sim_{k}^{=}}\}$. Now it only remains to see that each $\varphi^{=}_{E}$ is equivalent to the sentence $\nabla_{\mathtt{M}\mathtt{E}}(\overline{{\mathbf{T}}},\Pi)$ for some $\overline{{\mathbf{T}}},\Pi\subseteq\wp(A)$ with $\Pi\subseteq\overline{{\mathbf{T}}}$.

The crucial observation is that we will use $\overline{{\mathbf{T}}}$ and $\Pi$ to give a specification of the types “element by element”. Let $\mathbb{D}$ be a representative of the equivalence class $E$. Call $S_{1},\dots,S_{n}\subseteq A$ to the types such that $|S_{i}|_{\mathbb{D}}=n_{i}<k$ and $S^{\prime}_{1},\dots,S^{\prime}_{m}\subseteq A$ to those satisfying $|S^{\prime}_{j}|_{\mathbb{D}}\geq k$. The size of the sequence $\overline{{\mathbf{T}}}$ is defined to be $(\sum_{i=1}^{n}n_{i})+k\times m$ where $\overline{{\mathbf{T}}}$ contains exactly $n_{i}$ occurrences of type $S_{i}$ and at least $k$ occurrences of each $S^{\prime}_{j}$. On the other hand we set $\Pi\mathrel{:=}\{S^{\prime}_{1},\dots,S^{\prime}_{m}\}$. It is straightforward to check that $\Pi\subseteq\overline{{\mathbf{T}}}$ and $\varphi^{=}_{E}$ is equivalent to $\nabla_{\mathtt{M}\mathtt{E}}(\overline{{\mathbf{T}}},\Pi)$. (Observe however, that the quantifier rank of the latter is only bounded by $k\times 2^{|A|}+1$.) In particular $\varphi^{=}_{\mathbb{D}_{\varnothing}}\equiv\nabla_{\mathtt{M}\mathtt{E}}(\varnothing,\varnothing)=\forall x.\bot$.

The effectiveness of the procedure hence follows from the fact that, given the previous bound on the size of a normal form, it is possible to non-deterministically guess the number of disjuncts, types and associated parameters for each conjunct and repeatedly check whether the formulas $\varphi$ and $\bigvee\nabla_{\mathtt{M}\mathtt{E}}(\overline{{\mathbf{T}}},\Pi)$ are equivalent, this latter problem being decidable by Fact 2.6. qed

The logic $\mathtt{M}\mathtt{E}^{\infty}$ extends $\mathtt{M}\mathtt{E}$ with the capacity to tear apart finite and infinite sets of elements. This is reflected in the normal form for $\mathtt{M}\mathtt{E}^{\infty}$ by adding extra information to the normal form of $\mathtt{M}\mathtt{E}$.

Intuitively, the formula $\nabla_{\!\!\infty}(\Sigma)$ says that (1) for every type $S\in\Sigma$, there are infinitely many elements satisfying $S$ and (2) only finitely many elements do not satisfy any type in $\Sigma$. As a special case, the formula $\nabla_{\!\!\infty}(\varnothing)$ expresses that the model is finite. A short argument reveals that, intuitively, every disjunct of the form $\nabla_{\mathtt{M}\mathtt{E}^{\infty}}(\overline{{\mathbf{T}}},\Pi,\Sigma)$ expresses that any monadic model satisfying it admits a partition of its domain in three parts:

1. (i)

   distinct elements $t_{1},\dots,t_{n}$ with respective types $T_{1},\dots,T_{n}$,

2. (ii)

   finitely many elements whose types belong to $\Pi$, and

3. (iii)

   for each $S\in\Sigma$, infinitely many elements with type $S$.

Observe that basic formulas of $\mathtt{M}\mathtt{E}$ are not basic formulas of $\mathtt{M}\mathtt{E}^{\infty}$.

In the same way as before, we define an equivalence relation $\sim^{\infty}_{k}$ on monadic models which refines $\sim^{=}_{k}$ by adding information about the (in-)finiteness of the types.

Proof. We only prove the last point, for $k>0$. Let $E\in\mathfrak{M}/{\sim^{\infty}_{k}}$ and let $\mathbb{D}\in E$ be a representative of the class. Let $E^{\prime}\in\mathfrak{M}/{\sim^{=}_{k}}$ be the equivalence class of $\mathbb{D}$ with respect to $\sim^{=}_{k}$. Let $S_{1},\dots,S_{n}\subseteq A$ be all the types such that $|S_{i}|_{\mathbb{D}}\geq\omega$, and define

It is not difficult to see that $\mathtt{qr}(\varphi^{\infty}_{E})=k$ and that $\mathbb{D}^{\prime}\models\varphi^{\infty}_{E}$ iff $\mathbb{D}^{\prime}\in E$. In particular $\varphi^{\infty}_{\mathbb{D}_{\varnothing}}\equiv\nabla_{\mathtt{M}\mathtt{E}^{\infty}}(\varnothing,\varnothing,\varnothing)=\forall x.\bot\land\forall^{\infty}y.\bot$. qed

Now we give a version of the Ehrenfeucht-Fras̈sé game for $\mathtt{M}\mathtt{E}^{\infty}$. This game, which extends $\mathrm{EF}^{=}_{k}$ with moves for $\exists^{\infty}$, is the adaptation of the Ehrenfeucht-Fraïssé game for monotone generalised quantifiers found in [16] to the case of full monadic first-order logic.

Proof. Step (1) to (2) is direct by Proposition 3.12. For (2) to (3) we show the following.