Guarded Successor: A Novel Temporal Logic

Abstract

We present GS (Guarded Successor), a novel decidable temporal logic with several unique distinctive features. Among those, it allows infinitely many data values that come not only with equality but with a somehow rich theory too: the first-order theory of atomless Boolean algebras. The language also distinguishes between inputs and outputs, and has a decision procedure for determining whether for all inputs exist outputs, at each point of time. Moreover, and maybe most surprisingly, the data values can be nothing but sentences in GS itself. We also present a non-temporal fragment called NSO (Nullary Second Order) that enjoys merely this last property. These results are crucial necessary ingredients in any meaningful design of safe AI. Finally, all those results are obtained from a novel treatment of the first-order theory of atomless Boolean algebras.

1 Introduction

Traditional computation is temporal manipulation of bits. Bits, are the elements of the smallest possible Boolean algebra. The construction here can be seen as a generalization of this into working over certain infinite Boolean algebras. Decidability of a specification language in this model is of course much less trivial. Further, we will show how this generalization can support some very surprising abilities.

In the following, GS (Guarded Successor) is introduced, an innovative and decidable temporal logic that offers several distinctive features. First, it accommodates infinitely many data values, enhanced by a complex theory: the first-order theory of atomless Boolean algebras. Second, the language differentiates between input and output variables, and allows a decision procedure to prove that for all inputs there exist outputs, at each point in time. Third, and perhaps most surprisingly, the data values can be sentences in GS itself. The language is closed under Boolean combinations and allows quantification over both data values and time points. Its decision procedure is of a uniquely simplistic and elegant nature, and differs very much from other common decision procedures. It relies on the ability to enhance certain languages with recurrence relations (a form of fixed-point operators), in particular an extension of the first order theory of atomless Boolean algebras.

We also present NSO (Nullary Second Order logic, a name that was selected during the first incarnations of the idea and perhaps has to be reconsidered), a non-temporal fragment that maintains the above third property. All of these findings stem from a new approach to the first-order theory of atomless Boolean algebras. These results are essential for designing safe AI systems.

For ease of understanding we will first introduce the non-temporal NSO logic and afterwards we will introduce GS, which can be seen as a temporal extension of NSO. All nontrivial proofs not appearing in the main text, appear in the appendix.

Ongoing implementation of the languages described in this paper appears at the repository
https://github.com/idni/tau-lang.

Intellectual Property

The methods described here are protected from unauthorized use by IDNI Inc intellectual property rights, including patents. However, IDNI Inc. grants permission to use the methods for free in the following specific enumerated non-commercial instances: personal use, educational use , and academic purposes. The enumerated non-commercial instances do not include creation of open-source software that is distributed to others (whether for free or otherwise).

Acknowledgements

I would like to thank Enrico Franconi, Paweł Parys, and Lucca Tiemens for their review of this material and plenty of useful discussions.

1.1 NSO

The goal of NSO is to have a language that can speak about its own sentences in a consistent and decidable way. Tarski’s Undefinability of Truth has shown that this is impossible under a certain broad setting. The key of NSO is to abstract sentences, so much so, that they make merely Boolean algebra (BA) elements. In particular, there is no access to the syntax of the sentences (in contrast to Tarski’s setting which relies on Gödel numbers), and logically equivalent sentences are identified.

Any classical logic closed under Boolean combinations makes a BA called the Lindenbaum-Tarski Algebra (LTA) of that logic. Recall that this is only up to logical equivalence. Now observe two important points: 1. Any such logic that has an infinite signature (whether constant, relation, or function symbols), makes an atomless BA. 2. All countable atomless BAs are isomorphic (which is a well known theorem), and moreover, all atomless BAs are elementarily equivalent, as proved by Tarski. Clearly all sentences in languages of interest are finite strings over a finite alphabet, hence countable. The countable atomless BA is therefore the LTA of major logics of interest.

When we say “the theory of BA interpreted in a fixed BA $\mathcal{B}$” we mean not only the first order theory of BA interpreted in $\mathcal{B}$ (recall that an interpretation is a mapping taking symbols from the signature to actual objects in a structure), but we also mean that its signature is equipped with constants that are interpreted in each element of $\mathcal{B}$, so each element has a unique constant assigned to it. We will refer to those constants as the interpreted constants.

Fix a language $\mathcal{L}$ that its LTA makes an atomless BA. Let NSO$\left[\mathcal{L}\right]$ be the first-order theory of BA interpreted in that LTA, so each sentence in $\mathcal{L}$ is a constant symbol in NSO$\left[\mathcal{L}\right]$. So far, NSO$\left[\mathcal{L}\right]$ is a language that speaks about $\mathcal{L}$, but still not about itself. To this end, first we make the LTA of NSO$\left[\mathcal{L}\right]$ be an atomless BA as well (as currently it is only the two-element BA, as any logic that is interpreted in a fixed structure). This can be done by adding infinitely many uninterpreted constant symbols (the uninterpreted constants), or any other such trick. Then, the interpreted constants are extended to include sentences in NSO$\left[\mathcal{L}\right]$ (this is well-founded by introducing curly brackets as below). Since both $\mathcal{L}$ and NSO$\left[\mathcal{L}\right]$ make an atomless BA, they are elementarily equivalent under the signature of BA. By that we can make NSO$\left[\mathcal{L}\right]$ speak (including quantify) over its own sentences. Further, NSO$\left[\mathcal{L}\right]$ is decidable iff $\mathcal{L}$ is decidable.

1.2 GS

As an intuitive starting point, any formula with two free variables, in any logic, can be seen as defining a set of sequences: we say that a sequence $s$ models $\phi\left(x,y\right)$ iff any two consecutive elements $s_{i-1},s_{i}$ in the sequence satisfy $\phi\left(s_{i-1},s_{i}\right)$ (we can interpret $\phi$ in a fixed model, or one may appeal to any suitable notion of satisfiability). We then write $s\models\phi\left(x,y\right)$. Now consider the class of logics having the following property: fix a finite set of constant and variable symbols. Then the set of formulas making use only of those constant and free variable symbols (we allow, and require, arbitrarily many quantified variables), up to logical equivalence, is finite. Here, the most relevant such logic is the theory of atomless BA (whether or not interpreted in a fixed BA, and when it does, it is equipped with infinitely many interpreted constants as above, and this is the nontrivial case).

Denote by $\left|s\right|$ the length of $s$. Given $\phi\left(x,y\right)$, consider the following process: ask whether exists $s$ s.t. $\left|s\right|=2$ and $s\models\phi$, then whether exists $s$ s.t. $\left|s\right|=3$ and $s\models\phi$, and so on. This series of questions may take the form of a recurrence relation $\phi_{n}\left(x\right):=\exists y.\phi_{n-1}\left(y\right)\wedge\phi\left(x,y\right)$ with base-case $\phi_{2}\left(x\right):=\exists y.\phi\left(x,y\right)$ (though our preferred form will be slightly different). Then $\phi_{n}\left(x\right)$ means “exists a sequence of length $n$ starting with $x$”, and then to get a final answer (per each $n$) we of course need to consider $\exists x.\phi_{n}\left(x\right)$. Due to the finiteness property above, this series of questions is going to loop (i.e. at one point, a logically equivalent formula will occur), and even reach a fixed point due to the monotonic nature of the setting. We obtain a result of the form: “if a sequence of length $N$ exists, then a sequence of any larger length exists”. It is easy to see that this implies the existence of an infinite sequence as well.

For now we mention only two additional points, which are apparently unique to this language in the landscape of decidable temporal logics:

1. Seen as a program specification language, those sequences are actually outputs or states, however we’d like to support inputs as well. This means that we’d like to prove that for each input, at each point of time, exists an output, that does not depend on future inputs (time-compatible). So we can deal with formulas of the form $\phi\left(x_{n},x_{n-1},y_{n},y_{n-1}\right)$ where $x_{n},x_{n-1}$ are the current and previous inputs, respectively, and similarly for the outputs $y_{n},y_{n-1}$. Observe the bounded lookback in this formula, and observe that $n$ can be seen as a free variable of sort $\mathbb{N}$ which is implicitly quantified universally. The quantifier pattern for the inputs and outputs would look like $\forall x_{1}\exists y_{1}\forall x_{2}\exists y_{2}\dots$. It is easy to express it as a recurrence relation similar to the above, and again use the finiteness property as above.

2. Allowining GS to operate over the LTA of its own sentences, in the fashion of NSO, so $\phi$ is in the language of atomless BA, gives us a software specification language where inputs and outputs may be sentences in this very same language. This allows, for the first time, support implementations of the form: “reject a software update if it doesn’t satisfy certain desired properties” where the currently running program is written in the same language as the update, as well as those “desired properties”. It is therefore a crucial ingredient in AI safety. Fortunately, even decidability is preserved.

2 The Theory of Atomless Boolean Algebras

We assume that the reader is familiar with the definition of atomless BAs. We will denote the Boolean operations by $\cup,\cap,^{\prime}$ (disjunction, conjunction, complementation, respectively) to distinguish them from logical connectives.

2.1 Boolean Functions and Equations

We follow Rudeanu’s terminology.

Definition 1.

A Boolean Function (BF) is a Boolean combination of variables and constants (from some chosen BA). A Simple Boolean Function (SBF) is a BF where all constants appearing in it are either $0$ or $1$.

Note that this is in contrast to the common definition of BF, which is typically defined as what we call SBF. Also note that SBFs can be understood over any BA, but a BF which is not an SBF has to be understood over a single BA.

In the standard fashion, constant symbols in the theory of BA are only $0,1$. However we deal with a much richer setting in which the theory of BA is interpreted in a fixed BA (typically atomless). The theory is then enhanced with infinitely many constant symbols in its signature, each uniquely corresponding to each of the BA elements. We’ll refer to those as the interpreted constants.

An atomic formula in the language of BA (interpreted in some fixed BA) is therefore of the form $f\left(X\right)=g$$\left(X\right)$ where $f,g$ are BFs and $X$ is a tuple of variables. Note that this is equivalent to $f\left(X\right)+g\left(X\right)=0$ where $+$ denotes the ring sum, alternatively XOR, or symmetric difference (similarly we’ll use juxtaposition for conjunction, mixing set-theoretic and ring-theoretic notations). So we assume that atomic formulas are of the form $f\left(X\right)=0$. In the standard formulation of the theory of BA (that do not involve the interpreted constants), atomic formulas would have to involve merely SBFs, and not BFs in general.

Any BA induces a partial order defined by $a\leq b\leftrightarrow ab^{\prime}=0$. Recall that a BA is atomless iff the following holds: $\forall x.\left(x\neq 0\right)\rightarrow\left(\exists y.0<y<x\right)$ For ease of understanding of virtually any BA material, it is useful to bear in mind Stone’s representation theorem for BAs: any BA is isomorphic to a field of sets. Practically, it means that BA elements can always be realized as sets, where the Boolean operations coincide with the usual set operations, and $\leq$ coincides with $\subseteq$. This justifies our set-theoretic notation. In LTAs, a formula can be seen as a set of models, which automatically entails identification under logical equivalence.

Given a system of equations $\bigwedge_{i}\left[f_{i}\left(X\right)=0\right]$ (we shall always assume that all systems contain finitely many equations), it is easy to see that it is equivalent to a single equation $\left[\bigcup_{i}f_{i}\left(X\right)\right]=0$. We shall refer to this as squeezing the positives. Similar “squeezing” is generally not possible for a system of the form $\bigwedge_{i}\left[f_{i}\left(X\right)\neq 0\right]$, at least not in infinite BAs.

Definition 2.

A minterm in $n$ variables, denoted by $X^{A}$, is a product $x_{1}^{a_{1}}x_{2}^{a_{2}}\dots x_{n}^{a_{n}}$ where $A\in\left\{0,1\right\}^{n}$ and $x_{i}^{1}=x_{i};x_{i}^{0}=x_{i}^{\prime}$.

If $xy=0$ then we say that $x,y$ are disjoint. So $0$ is disjoint from all elements including itself. Observe that $X^{A}X^{B}=0$ iff $A\neq B$ or $X^{A}=0$ or $X^{B}=0$.

Definition 3.

A function $\mathcal{B}^{n}\rightarrow\mathcal{B}$ is in minterm normal form if it is written as $f\left(X\right)=\bigcup_{A\in\left\{0,1\right\}^{n}}c_{A}X^{A}$ where $c_{A}\in\mathcal{B}$.

Clearly any function in minterm normal form is a BF. The converse is also true. cf. [4] for the proof of the following theorem:

Theorem 4.

A function $\mathcal{B}^{n}\rightarrow\mathcal{B}$ is a BF iff it can be written in minterm normal form $f\left(X\right)=\bigcup_{A\in\left\{0,1\right\}^{n}}f\left(A\right)X^{A}$.

Note that we use $X,A$ as tuples of variables, so the notation $f\left(X\right),f\left(A\right)$ should be clear.

Corollary 5.

A BF is uniquely determined by its values over the two-element BA.

Now we describe Boole’s consistency condition. A system of equations is consistent if it has a solution. The following was discovered by Boole and is actually a case of quantifier elimination:

Theorem 6.

Let $f:\mathcal{B}^{n}\rightarrow\mathcal{B}$ be a BF, then $\exists X.f\left(X\right)=0$ iff $\left[\bigcap_{A\in\left\{0,1\right\}^{n}}f\left(A\right)\right]=0$ and $\exists X.\left[f\left(X\right)\neq 0\right]$ iff $\left[\bigcup_{A\in\left\{0,1\right\}^{n}}f\left(A\right)\right]\neq 0$

Theorem 7.

Let $f:\mathcal{B}\rightarrow\mathcal{B}$ be a BF s.t. $f\left(0\right)f\left(1\right)=0$, or equivalently, $\exists x.f\left(x\right)=0$. Then f$\left(x\right)=0$ iff $x=t+f\left(t\right)$ for some $t$, iff $f\left(0\right)\leq x\leq f^{\prime}\left(1\right)$.

Our main proof of correctness for our quantifier elimination method in atomless BA will involve Hall’s marriage theorem. We present it here in its set-theoretical version:

Definition 8.

Let $A_{1},\dots,A_{n}$ be sets, not necessarily distinct. A choice of elements $a_{1}\in A_{1},\dots,a_{n}\in A_{n}$ such that $a_{i}\neq a_{j}$ for all $i\neq j$ is called a system of distinct representatives.

Theorem 9.

Let $\mathcal{A}=A_{1},\dots,A_{n}$ be a sequence of sets, not necessarily distinct. Then $\mathcal{A}$ does not have a system of distinct representatives, iff there exists a subsequence $\mathcal{B}=B_{1},\dots,B_{m}$ of $\mathcal{A}$ s.t. $\left|\bigcup_{i}B_{i}\right|<m$.

Remark 10.

A simple observation which we shall make use of later on is that a system of distinct representatives exists iff it exists for the subsequence in which all infinite $A$’s are removed from it. In other words, infinite sets in a finite collection of sets don’t influence the existence of distinct representatives.

The following theorem is our key step towards quantifier elimination:

Theorem 11.

Let $X^{A_{1}},\dots,X^{A_{m}}$ be minterms in $n$ variables, and $b_{1},\dots,b_{m}$ elements in some BA. Then $\exists X.\bigwedge_{i=1}^{m}\left[X^{A_{i}}\geq b_{i}\right]$ iff $b_{i}b_{j}=0$ whenever $A_{i}\neq A_{j}$.

Corollary 12.

The system $\bigwedge_{i\in I}\left[b_{i}X^{A_{i}}\neq 0\right]$ has a solution iff there exists $0<c_{i}\leq b_{i}$ s.t. $c_{i}c_{j}=0$ whenever $A_{i}\neq A_{j}$.

The corollary reduces the problem of determining consistency of the above system to a case of 9, once treating each $b_{i}$ as follows: if it can be written as a disjunction of distinct atoms, then we treat it as a set whose elements are those atoms, and each $c_{i}$ is a choice of one or more atoms. If $b_{i}\neq 0$ and cannot be written as a union of atoms, then we treat it as an infinite set and by that it is eliminated from the problem as we have pointed out in remark 10.

Remark 13.

Observe that $a\cup b=0\leftrightarrow a=0\wedge b=0$ and recall that each BF can be written as a sum of minterms, or in DNF (note that writing a BF in DNF is not the same thing as writing a formula in DNF). This allows an alternative syntax for theories of BA where atomic formulas are of the form $cX^{A}=0$. We call this form minterm normal form. Note that this is not the same minterm normal form of BFs, as here it applies to forms of atomic formulas.

The above results together with the last remark imply with the following key conclusion:

Corollary 14.

Finitely many multivariate BFs over an atomless BA have a common nonzero iff none of them is identically zero.

2.2 Quantifier Elimination in Atomless BAs

Given a formula in the language of BA, we can write it in a way such that some chosen innermost quantifier is existential. We then convert everything under that quantifier to DNF. Distributing the existential over the DNF clauses, and squeezing the positives (as above) in each clause, we see that if we can eliminate the existential quantifier from a formula of the form $\exists x.f\left(x\right)=0\wedge\bigwedge_{i}g_{i}\left(x\right)\neq 0$, then we can eliminate all quantifiers. We’ll therefore deal only with such a case.

Lemma 15.

In any BA, the system $f\left(x\right)=0\wedge g\left(x\right)\neq 0$ has a solution iff $f\left(0\right)f\left(1\right)=0\wedge\bigwedge_{i}g_{i}\left(x+f\left(x\right)\right)\neq 0$ has a solution.

Theorem 16.

In atomless BA, the system $f\left(x\right)=0\wedge\bigwedge_{i\in I}g_{i}\left(x\right)\neq 0$ has a solution iff $f\left(0\right)f\left(1\right)=0\wedge\bigwedge_{i\in I}g_{i}\left(f\left(0\right)\right)\cup g_{i}\left(f^{\prime}\left(1\right)\right)\neq 0$.

Proof.

Using the last lemma and 6. ∎

Proposition 17.

For any BF $f$ we have $xf\left(x\right)=xf\left(1\right)$ and $x^{\prime}f\left(x\right)=x^{\prime}f\left(0\right)$.

Proof.

Exercise. ∎

Lemma 18.

In any BA, $f\left(x\right)=0\wedge\bigwedge_{i}g_{i}\left(x\right)\neq 0$ has a solution iff $f\left(x\right)=0\wedge\bigwedge_{i}xf^{\prime}\left(1\right)g_{i}\left(1\right)\neq 0\vee x^{\prime}f^{\prime}\left(0\right)g_{i}\left(0\right)\neq 0$ has a solution, iff

$$\begin{array}[]{c}f\left(0\right)f\left(1\right)=0\wedge f\left(x\right)=0\wedge\\ \bigwedge_{i}g_{i}\left(0\right)g_{i}\left(1\right)\neq 0\vee xf^{\prime}\left(1\right)g_{i}\left(1\right)\neq 0\vee x^{\prime}f^{\prime}\left(0\right)g_{i}\left(0\right)\neq 0\end{array}$$

has a solution.

Corollary 19.

In atomless BA, the system $f\left(x\right)=0\wedge\bigwedge_{i}g_{i}\left(x\right)\neq 0$ has a solution iff $f\left(0\right)f\left(1\right)=0\wedge\bigwedge_{i}f^{\prime}\left(1\right)g_{i}\left(1\right)\cup f^{\prime}\left(0\right)g_{i}\left(0\right)\neq 0$.

Proof.

Applying corollary to 18. ∎

By that we conclude the final form of our quantifier elimination algorithm over atomless BAs. Basically we generalized proposition 5.5 in [5] from SBF to BF over atomless BA (or over any BA whenever the cardinalities of the constants appearing in the equations are large enough), using very different and much more direct methods, and we indicated how and why it fails in non-atomless BA. More treatment for the non-atomless case can be found at [1].

2.3 Number of Formulas

We presented minterm normal form of formulas in remark 13. This normal form puts a bound on the number of quantifier-free logically equivalent formulas with $n$ free variables and $k$ constants. The accounting is as follows: the formula is itself an SBF of atomic formulas, and there are $2^{2^{N}}$ different SBFs in $N$ variables. In our case $N$ is the number of possible minterms which is readily $k2^{n}$. We therefore end up with a triple exponential $2^{2^{k2^{n}}}$ upper bound. This extends to quantified formulas as well due to quantifier elimination.

2.4 Recurrence Relations

We propose the notion of weakly $\omega$-categorical theories. Recall that an $\omega$-categorical theory is a first order theory in which all of its countable models are isomorphic. The Ryll-Nardzewski theorem says that this definition is equivalent to another definition: that up to logical equivalence, there are only finitely many formulas with free variables taken from a fixed finite set. This gives rise to defining weakly $\omega$-categorical theories: those are theories for which the number of formulas using a fixed finite set of free variables and where the constants appearing in them are taken from a fixed finite subset of all constants in the language, up to logical equivalence, is finite. For the sake of this section, it does not matter whether or not the theory is partially interpreted in a fixed structure. The concept of partial interpretation is as follows: suppose we interpret a logic in a fixed structure. Now we enhance this logic with additional uninterpreted constants. Then a partial interpretation would be assigning domain elements to those newly added constant symbols.

It is easy to see, in light of the previous section and the quantifier elimination results, that the theory of atomless BA and of fixed finite BA, are both weakly $\omega$-categorical. In what follows we shall deal only with those BA theories. However many of the constructions in this section and across this paper can be carried out into any weakly $\omega$-categorical theory.

We are now ready to define formulas in the language of BA (or any weakly $\omega$-categorical theory) enhanced with recurrence relations. Any such formula takes the form $\phi_{n}\left(X\right)=\Phi\left(X,\phi_{n-1}\left(X\right),\phi_{n-2}\left(X\right),\dots\right)$ together with suitable base conditions $\phi_{1}\left(X\right)=\dots,\phi_{2}\left(X\right)=\dots,\dots$. Clearly those recurrence relations do not always have a fixed point, but weakly $\omega$-categoricity guarantees a partial fixed point. It is easy to pin down all cases in which a given formula has a fixed point (in case the theory we start with is decidable), as well as apply basic remedies for the case that only a partial fixed point exists, but those are omitted here for the sake of brevity.

We have shown that any formula in the theory of atomless BA enhanced with recurrence relations can be written in an equivalent form without recurrence relations.

3 NSO: Nullary Second Order Logic

Building on what we said in the introduction, we shall not merely present a language, but a language-extension mechanism, altough this extension is not in the standard sense, namely formulas in the base language are not, as for themselves, formulas in the extended language, but constant symbols in it, as we shall see. This extension preserves decidability, let alone consistency. We further consider extending many languages at once, and it is indeed yet another feature of our construction to allow languages to co-exist in one unified language, albeit, of course, the interaction between those languages is very limited. Referring to many BAs at once is easily done by considering the many-sorted theory of BA, alternatively the product algebra.

Fix arbitrary languages (the base logics) in which their formulas (or sentences), up to logical equivalence, make a BA. Then we can consider the many-sorted BA theory interpreted in those BAs. Constants in that languge are formulas in the base logics. Quantification takes the same semantics of quantification over arbitrary BA elements. If the base logics make an atomless BA, then the extended language has decidable satisfiability iff the base logics have. Otherwise decidable model counting is required, or more precisely, when seen as a BA, to tell whether an element is a disjunction of at least $n$ distinct atoms.

Denote the extended language by $NSO\left[\mathcal{L}_{1},\dots,\mathcal{L}_{n}\right]$. We show that $NSO\left[\mathcal{L}_{1},\dots,\mathcal{L}_{n}\right]$ can have itself as a base-logic. So far, each NSO formula is either true or false, because it is interpreted in a fixed model (being the BA which is the LTA of the base logic), and therfore makes a small BA (only two elements). To obtain an atomless BA from formulas in $NSO\left[\mathcal{L}_{1},\dots,\mathcal{L}_{n}\right]$ we can simply enhance it with infinitely many uninterpreted constant symbols. After doing so, we now let interepreted constants to be formulas in $NSO\left[\mathcal{L}_{1},\dots,\mathcal{L}_{n}\right]$ appearing inside curly brackets (in order to avoid syntactic ambiguity), and handling of quantifiers for the sake of a decision procedure can be done by means of the atomless BA quantifier elimination algorithm. The basic syntax of $NSO\left[\mathcal{L}_{1},\dots,\mathcal{L}_{n}\right]$ is therefore

$$\phi:=\exists var:sort.\phi|\phi\wedge\phi|\neg\phi|bf=0$$

$$sort:=\mathcal{L}_{1}|\dots|\mathcal{L}_{n}|NSO\left[\mathcal{L}_{1},\dots,L_{n}\right]$$

$$bf:=var|\left\{\phi^{sort}\right\}|const|0|1|bf\cap bf|bf^{\prime}$$

where $\phi^{\mathcal{L}}$ means any formula in the language $\mathcal{L}$. Clearly, each $bf$ may only contain variables and constants from the same sort. $const$ refers to an uninterpreted constant. The deep-most level of formulas in [nested] curly brackets will be either a formula in $\mathcal{L}_{1},\dots,\mathcal{L}_{n}$ or a formula in the language of BA in which the only constants appearing in it are $0,1$. It is then interpreted as a formula over arbitrary atomless BA since they’re all elementarily equivalent. It is easy to see that going inductively over the depth of curly brackets, gives a decision procedure as well as semantics to this language.

4 GSSOTC: A Temporal Logic

We devise a new, decidable, family of temporal logics over infinite data values, where those values come with theories much richer than merely equality, in particular with the theory of atomless Boolean Algebras (as well as fixed finite ones though such a case does not amount to a significant novelty). Further, this language enjoys the distinctive ability to verify statements of the form “at each point of time, for all inputs exist a well-defined output/state, possibly depending on the previous output/state”. It also presents a new kind of decision procedure, unrelated to automata, tableaux, or to any other decision method known to the author.

To describe the language in simple intuitive terms: fix an atomless BA and consider the theory of BA interpreted in this structure (with interpreted constants as above so the LTA of this logic is the countable atomless BA). Consider formulas with free variables $x_{n-k},\dots,x_{n},y_{n-k},\dots,y_{n}$ where the $x$’s are understood as inputs and the $y$’s are understood as outputs, and $n$ is any time point (so it can be seen as a free variable of sort $\mathbb{N}$). So it describes connection between current and previous inputs and outputs at each point of time. This is basically almost the full language.

This technique works for any weakly $\omega$-categorical language, as long as it supports conjunction and quantification. However in the atomless BA case we get the unique property of a language that can speak of its own sentences, in the spirit of NSO.

4.1 Time-Compatible Structures

A sequence of elements from some domain $\mathcal{D}$ can be seen as a function $\mathbb{N}\rightarrow\mathcal{D}$. A function between sequences is therefore of type $\left(\mathbb{N}\rightarrow\mathcal{D}\right)\rightarrow\left(\mathbb{N}\rightarrow\mathcal{D}\right)$. As customary in many texts, $\left[k\right]$ will denote the set $\left\{1,\dots,k\right\}$.

Definition 20.

A function $f:\left(\mathbb{N}\rightarrow\mathcal{D}\right)\rightarrow\left(\mathbb{N}\rightarrow\mathcal{D}\right)$ between sequences is prefix-preserving (alternatively time-compatible, TC) if for all sequences $p,s$, if $p$ is a strict prefix of $s$, then $f\left(p\right)$ is a strict prefix $f\left(s\right)$. We extend this notion also for $f:\left(\left[n\right]\rightarrow\mathcal{D}\right)\rightarrow\left(\left[n\right]\rightarrow\mathcal{D}\right)$.

Definition 21.

A Time-Compatible (TC) Structure of length $N\in\mathbb{N}\cup\left\{\infty\right\}$ is a domain $\mathcal{D}$ with prefix-preserving functions $\mathcal{D}^{N}\rightarrow\mathcal{D}^{N}$.

It should be clear that any computer program is a TC structure: at each point of time it takes an input and outputs an output, while the output may depend only on past and present inputs and outputs, not future ones. This is why we refer to prefix-preservation as TC.

Remark 22.

Due to the “lookback” ability, namely the dependence on previous inputs and outputs, we don’t need to refer to the concept of state, as it is subsumed by the concept of output.

Remark 23.

In what follows we will deal only with infinite-time TC structures (so $N=\infty$ in the above definition) unless stated otherwise.

Remark 24.

We will eventually be interested with functions from tuples of sequences to tuples of sequences (all tuples of fixed finite size, but the input tuple may be of different size than of the output tuple). All definitions and results should apply mutatis-mutandis.

Remark 25.

The setting can easily be extended to trees rather sequences. It is done by allowing more than one successor relation, and the same methods apply.

Definition 26.

A TC function has bounded lookback (BL) of length $k\in\mathbb{N}$ (or simply BL$\left[k\right]$) if exists $m\geq k$ (the recurrence point), s.t. for each $n>m$, the output sequence at point $n$ depends only on the input and output sequences at points $n-1,\dots,n-k$, as well as the input at point $n$.

Corollary 27.

If $f$ is BL$\left[k\right]$ then it can be expressed as a pair of functions, one of type $\mathcal{D}^{2k+1}\rightarrow\mathcal{D}$ and another of type $\mathcal{D}^{m}\rightarrow\mathcal{D}^{m}$ which is required to be TC.

Proof.

By definition of BL functions, we can write $f$ as a recurrence relation

$$\left[f\left(x\right)\right]_{n}=g\left(x_{n},x_{n-1},\dots,x_{n-k},\left[f\left(x\right)\right]_{n-1},\dots,\left[f\left(x\right)\right]_{n-k}\right)$$

(where $x$ is the input sequence) with initial conditions of the form $\left[f\left(x\right)\right]_{i}=\dots$ for $1\leq i\leq k$. This $g$ is of type $\mathcal{D}^{2k+1}\rightarrow\mathcal{D}$ and together with the initial conditions (which specify the behavior up until the recurrence point), fully encodes $f$. ∎

Corollary 28.

Given a pair of functions, one of type $\mathcal{D}^{2k+1}\rightarrow\mathcal{D}$, and another, which is TC, of type $\mathcal{D}^{m}\rightarrow\mathcal{D}^{m}$, we can uniquely assign to it a function of BL$\left[k\right]$.

4.2 Bounded Lookback and Recurrence Relations

Corollary 29.

Any formula (in virtually any logic) with $2k+2$ free variables defines a [possibly empty] set of BL$\left[k\right]$ functions.

Note that in the infinitary expression obtained in the proof, quantifiers can be pushed inside. This is a property of being TC, and this ability is one cruicial point in the upcoming construction. Also note that skolemization of this expression will yield something similar to the type in .

Remark 30.

The initial conditions are not expressed in the latter corollary. But the corollary still holds. It defines a set of functions that include functions per each possible initial condtions. This is not an inherent limitation. We used this form only for simplicity at this stage.

Fix a lookback parameter $k\geq 0$. $X_{j}$ will denote a tuple of variables of lookback $k$, so it’s a tuple of $k+1$ variables of the form $x_{j-k},x_{j-k+1},\dots,x_{j}$. We assume that the first time coordinate is $0$.

Definition 31.

Given formula $\phi$ (in virtually any logic) with $2k+2$ free variables $x_{n-k},\dots,x_{n},y_{n-k},\dots,y_{n}$, define a recurrence relation $\phi_{n}$ by $\phi_{n+1}\left(X_{k},Y_{k}\right):=\phi\left(X_{k},Y_{k}\right)\wedge\forall x_{k+1}\exists y_{k+1}.\phi_{n}\left(X_{k+1},Y_{k+1}\right)$ with base case $\phi_{1}:=\phi\left(X_{k},Y_{k}\right)$.

Remark 32.

Observe that $\phi_{n}\left(X_{k},Y_{k}\right)$ actually says that exists a BL$\left[k\right]$ function between sequences of length $n+k$, where the $k$ initial positions in the sequences are left as free variables.

Note that $\phi_{n}$ has a form of monotonicity wrt $n$: if exists a TC function between sequences of lentgh $n+1$, and the function satisfies $\phi$, then clearly exist such a function for sequences of length $n$.

Clearly, if $\forall x_{0}\exists y_{0}\dots\forall x_{k}\exists y_{k}.\phi_{n}\left(X_{k},Y_{k}\right)$ for all $n$, then $\phi$ defines a nonempty set of functions in the spirit of corollary . The crux of our construction is the observation that if the underlying logic is weakly $\omega$-categorical, then there are only finitely many $\phi_{n}$’s up to logical equivalence, hence decidability and decision procedure are immediate.

4.3 Guarded Successor

Observe that a formula of the form $\phi\left(X_{k},Y_{k}\right)$ can be given a direct BL$\left[k\right]$ semantics also by adding a sort of natural numbers with the successor relation $s$, and function symbols $f:\mathbb{N}\rightarrow\mathcal{D}$ and $F:\left(\mathbb{N}\rightarrow\mathcal{D}\right)\rightarrow\left(\mathbb{N}\rightarrow\mathcal{D}\right)$, where $F$ is required to be prefix-preserving, and writing $\phi$ as

$$\forall t_{0},\dots t_{k}.\left[\bigwedge_{i=0}^{k-1}s\left(t_{i},t_{i+1}\right)\right]\rightarrow\phi\left(f\left(t_{0}\right),\dots,f\left(t_{k}\right),F\left(f\right)\left(t_{0}\right),\dots,F\left(f\right)\left(t_{k}\right)\right)$$

Definition 33.

Fix a logic ${\cal L}$ and let $\mathcal{D}$ be the sort it operates over. First extend it with function symbols $f_{i}:\mathbb{N}\rightarrow\mathcal{D}$ and $F_{j}:\left(\mathbb{N}\rightarrow\mathcal{D}\right)\rightarrow\left(\mathbb{N}\rightarrow\mathcal{D}\right)$, where $F$ is required to be prefix-preserving. If $\psi$ is any formula in this extended language, then

$$\phi:=\psi|\phi\wedge\phi|\neg\phi|\forall t_{1},\dots,t_{m}.\left[\bigwedge_{\left(i,j\right)\in I}s\left(t_{i},t_{j}\right)\right]\rightarrow\phi$$

defines a second extension to the language which we shall refer to as the guarded successor extention of ${\cal L}$. The sublanguage of the form

$$\phi:=\psi|\phi\wedge\phi|\neg\phi|\forall t_{1},\dots,t_{m}.\left[\bigwedge_{\left(i,j\right)\in I}s\left(t_{i},t_{j}\right)\right]\rightarrow\psi$$

will be called the collapsed fragment. Its sublanguage of the form

$$\phi:=\bigvee_{k}\left(\forall t_{1},\dots,t_{m}.\left[\bigwedge_{\left(i,j\right)\in I_{k}}s\left(t_{i},t_{j}\right)\right]\rightarrow\psi_{k}^{1}\right)\wedge\left(\exists t_{1},\dots,t_{m}.\left[\bigwedge_{\left(i,j\right)\in J_{k}}s\left(t_{i},t_{j}\right)\right]\wedge\psi_{k}^{2}\right)$$

will be called the normalized fragment. In all cases, the guard $\bigwedge_{\left(i,j\right)\in I}s\left(t_{i},t_{j}\right)$ is required to uniquely determines the relative position between each $t_{i},t_{j}$, and $\psi,\psi_{k}^{1},\psi_{k}^{2}$ involve $t_{1},\dots,t_{m}$ only through application of $f,F$ (or several such functions), while $f,F$ may also be applied to constants from $\mathbb{N}$.

Remark 34.

Applying $f,F$ to constants from $\mathbb{N}$ corresponds to the above initial conditions.

Theorem 35.

Any formula in a guarded successor extension can be written as an equisatisfiable formula in the normalized fragment.

Remark 36.

Note that here we had to use the assumption that we are dealing with infinite-time structures, namely $N=\infty$. In the finite-time case we will also need the end-of-sequence predicate $\sharp$, resulting with a slightly more complicated quantifier collapse. We omit this simple derivation here for sake of brevity.

Corollary 37.

Any formula in a guarded successor extension without temporal existential quantifiers can be written in a free-variable BL$\left[k\right]$ form $\phi\left(X_{k},Y_{k}\right)$.

We of course bear in mind that if some language is decidable and is weakly $\omega$-categorical, then its extension with recurrence relations is also decidable. Together with a method to handle the existential part as described in the next section, we’ll conclude that:

Corollary 38.

Satisfiability of a formula in a guarded successor extension is decidable if this fragment is obtained from a decidable language $\mathcal{L}$ which is weakly $\omega$-categorical, enhanced with the sort $\mathbb{N}$, guarded successors, $\mathbb{N}\rightarrow\mathcal{D}$ function symbols, and BL$\left[k\right]$ function symbols.

We refer to this extended language as GSSOTC$\left[\mathcal{L}\right]$, where GSSOTC stands for Guarded-Successor Second-Order Time-Compatible. The second-order part is due to the following: given two sequences $f,g:\mathbb{N}\rightarrow\mathcal{D}$, we can declare a non-standard quantifer alternation $\forall f\exists g$, which would translate into $\exists F\forall f$ (so far just standard higher-order skolemization), where $F$ is a TC function between sequences. Those function quantifiers are eliminated when converting the formula to the free-variable form, which is then converted to function-free recurrence-relation form.

Some easy extensions of this language were described above, we reiterate them and add more: the end-of-string predicate $\sharp$, having multiple successor relations and by that considering trees rather sequences, having constant positions, so instead of e.g. $\phi\left(x_{n},x_{n-1},y_{n}\right)$, we have e.g. $\phi\left(x_{1},x_{2},x_{n},x_{n-1},y_{n}\right)$, having explicit second-order quantifiers that are eliminated by reduction to recurrence relations, and finally, having richer quantifier alternation, e.g. for all keyboard input at time $n$, exists a memory state at time $n$, s.t. for all network input at time $n$, and so on, resulting in quantification of the form $\forall x_{1}\exists y_{1}\forall z_{1}\forall x_{2}\exists y_{2}\forall z_{2}\forall x_{3}\exists y_{3}\forall z_{3}\dots$.

4.4 Decision Methods and Execution

In the spirit of remark 24, we shall have several input and output sequences, each referred to as a stream.

Theorem 39.

Given $\phi\left(X_{j}^{i},Y_{j}^{i}\right)$ where $X$ are inputs and $Y$ are outputs, and $i$ denoting the stream number, define the recurrence relation

$$\phi_{0}\left(X_{k}^{i},Y_{k}^{i}\right):=\phi\left(X_{k}^{i},Y_{k}^{i}\right)$$

$$\phi_{n}\left(X_{k}^{i},Y_{k}^{i}\right):=\phi\left(X_{k}^{i},Y_{k}^{i}\right)\wedge\forall x_{k+1}\exists y_{k+1}.\phi_{n-1}\left(X_{k+1}^{i},Y_{k+1}^{i}\right)$$

so $\phi_{n}$ means that exists a model with time points $0,\dots,n+k$ starting with $X_{k}^{i},Y_{k}^{i}$. Then the reccurence relation is monotonic, namely $\forall n\forall X_{k}^{i}Y_{k}^{i}.\phi_{n+1}\left(X_{k}^{i},Y_{k}^{i}\right)\rightarrow\phi_{n}\left(X_{k}^{i},Y_{k}^{i}\right)$ and therefore has a fixed point. Denote it by $\phi_{\infty}\left(X_{k}^{i},Y_{k}^{i}\right)$. Given a model of $\phi$ with $m$ time points, and given each input $X^{i}$ at point $m+1$, then an output $Y^{i}$ will have an unbounded continuation satisfying $\phi$ iff $\phi_{\infty}\left(X_{m+1}^{i},Y_{m+1}^{i}\right)$.

Remark 40.

The above formulation suggests that $\phi_{\infty}$ is a normal form of $\phi$ when understood as defining TC models.

Remark 41.

A TC structure is a model of $\phi$ iff any subsequence satisfies $\phi_{\infty}$ when understood as a formula in the language of BA.

Remark 42.

Given inputs at each point of time, satsifying outputs can be computed by substituting the known variables into $\phi_{\infty}$, and solving for the missing outputs. This is an execution method for software specification in this language. Software specification in this language is therefore directly executable as-is, using an oracle to determine satisfying assignments to formulas in the language of atomless BA. Finding satisfying assignments to a formula in the language of atomless BA is a topic by its own, and is omitted here for sake of brevity.

Corollary 43.

Given two formula $\phi\left(X_{j}^{i},Y_{j}^{i}\right),\psi\left(X_{j}^{i},Y_{j}^{i}\right)$, then the set of TC models of $\phi$ is a subset of those of $\psi$, iff $\forall x_{0}y_{0}\dots x_{k}y_{k}.\phi_{\infty}\left(X_{k}^{i},Y_{k}^{i}\right)\rightarrow\psi_{\infty}\left(X_{k}^{i},Y_{k}^{i}\right)$.

This gives us an algorithm to decide whether $\phi\psi^{\prime}=0$ where $\phi,\psi$ are seen as sets of TC models.

Remark 44.

Combined with 35 and its proof, this corollary gives us a decision procedure for the full language GS. Each DNF clause will have a single universal and a single existential (which is a negated universal), so deciding emptiness for each clause comes down to the last corollary.

Remark 45.

Since $\phi_{\infty}$ refers only to the universal parts, while the existential parts may of course restrict the models, therefore we should, at execution time, check at each point of time whether we can satisfy the existential parts. If so, we satisfy them indeed, just once. If the formula is satisfiable then such point in time is guaranteed to exist. If there are multiple existential parts in a DNF clause, then for execution, we have to squeeze them into one using the flags as in the proof of 35, since those existential parts may depend on each other.

Remark 46.

When $\phi\left(X_{n},Y_{n}\right)$ is understood as a GS formula, and $\phi$ is in the language of atomless BA intepreted in this very BA of GS formulas (possibly with more algebras as the consrtuction is closed under products), then NSO is a sublanguage of this language. That’d be a software specification language where inputs and outputs are nothing but sentences in this very language. This way we can support the software update mechanism described in the introduction as a crucial component for safe AI. Another way to look at it: a robot is programmed in a language ${\cal L}$ and accepts commands form the user in the very same language ${\cal L}$. Now its internal program has to ask whether the command is consistent with, say, safety conditions. It couldn’t do so unless ${\cal L}$ is a temporal logic with inputs in ${\cal L}$ equipped with the theory of BA.

4.5 Complexity

Quantifier elimination in theories of BA where constants are either 0,1 were studied by Tarski by introducing his so-called invariants. Kozen [3] extended this notion of invariants and by that derived a complexity characterization for the decision problem. For infinite BAs, it is complete for $\bigcup_{c}$STA$\left(*,c^{n},n\right)$. Roughly, this means anything that can be done in exponential time by an alternating Turing machine with linearly many alternations. For the two-element BA, it is simply QBF which is maybe the most famous PSPACE-complete problem. For GS, we saw that the number of formulas with fixed number of free variables and constants, is triple-exponential in the number of the free variables. This gives an upper bound for GS over atomless BA.

5 Conclusion

We have presented new methods in the theory of atomless BA that extend existing results from SBF to BF in general. We also presented the concept of weakly $\omega$-categorical theories and how they relate to decidable recurrence relation extensions. We used those ideas to construct a language that can speak of its own sentences by abstracting them to merely BA elements. We further extended this construction to a novel temporal logic with several distinguishing abilities.

Appendix: Proofs

Proof of 6.

We prove the first statement and the second is analogous. Further we prove it only for the univariate case, and the multivariate case follows immediately by induction. Any univariate BF can be written in Boole’s normal form (sometimes mistakingly called Shannon’s normal form) as $f\left(x\right)=ax\cup bx^{\prime}$ where $a=f\left(1\right)$ and $b=f\left(0\right)$. Now $f\left(x\right)=0$ iff $ax=bx^{\prime}=0$, which reads $b\leq x\leq a^{\prime}$, so a solution exists iff $b\leq a^{\prime}$ equivalently $ab=0$. ∎

Proof of 7.

The second equivalence follows immediately from the proof of the previous theorem. For the first equivalence, write $f\left(x\right)=ax+b$ (this is the algebraic normal form). Then

$$f\left(x+f\left(x\right)\right)=a\left(x+ax+b\right)+b=ax+ax+ab+b=ab+b=f\left(0\right)f\left(1\right)=0$$

and for the other direction, if $f\left(x\right)=0$, just put $t=x$. ∎

Proof of 11.

First assume that $X^{A_{1}},\dots.X^{A_{m}}$ are all distinct and therefore the nonzero $b$’s are all disjoint, otherwise convert any two equations of the form $\begin{matrix}X^{A_{i}}\geq s\\ X^{A_{i}}\geq t\end{matrix}$ into the equivalent form $X^{A_{i}}\geq s\cup t$. Necessity is now immediate recalling that two different minterms are always disjoint and that subsets of disjoint sets must also be disjoint. For sufficiency and $n=1$ the equations take the form $x\geq b_{1}$ and $x^{\prime}\geq b_{2}$ which indeed has a solution iff $b_{1}b_{2}=0$. Assume for $n$ and consider a distinguished variable $x$. Then we can split the equations into $p+q=m$ equations and rewrite them as $\begin{matrix}\bigwedge_{i=1}^{p}xX^{A_{i}}\geq b_{i}\\ \bigwedge_{j=1}^{q}x^{\prime}X^{B_{j}}\geq c_{j}\end{matrix}$ and let $X$ be a solution of $\begin{matrix}\bigwedge_{i=1}^{p}X^{A_{i}}\geq b_{i}\\ \bigwedge_{j=1}^{q}X^{B_{j}}\geq c_{j}\end{matrix}$ using the induction hypothesis after making sure that all $A_{i},B_{i}$ are distinct (while if $p+q=1$ then a solution trivially exists). If $p\neq 0$, set $x=\bigcup_{k}b_{k}$. Then $\bigcup_{k}c_{k}\leq x^{\prime}$ due to the disjointness assumption. Therefore

$$xX^{A_{i}}=\left(\bigcup_{k}b_{k}\right)\wedge X^{A_{i}}\geq b_{i}X^{A_{i}}=b_{i}$$

$$x^{\prime}X^{B_{j}}\geq\left(\bigcup_{k}c_{k}\right)X^{B_{j}}\geq c_{j}X^{B_{j}}=c_{j}$$

Similarly set $x=\bigcap_{k}c_{k}^{\prime}$ if $p=0$, or simply $x=0$. ∎

Proof of 15.

If $f$ has a zero, then all such zeros are precisely the range of $x+f\left(x\right)$ by 7. So we can write the system as $f\left(x+f\left(x\right)\right)=0\wedge\bigwedge_{i}g_{i}\left(x+f\left(x\right)\right)\neq 0$. Now $f$ has a zero iff $f\left(0\right)f\left(1\right)=0$ by Boole’s consistency condition, in which case $f\left(x+f\left(x\right)\right)$ is identically zero. ∎

Proof of 18.

First substitute the general solution $x+f\left(x\right)$ of the positive part into the negative parts and obtain:

$$f\left(x\right)=0\wedge\bigwedge_{i}g_{i}\left(x+f\left(x\right)\right)\neq 0$$

and since $f\left(x\right)=0$ there is no harm in multiplying the negative part with $f^{\prime}\left(x\right)$:

$$f\left(x\right)=0\wedge\bigwedge_{i}f^{\prime}\left(x\right)g_{i}\left(x+f\left(x\right)\right)\neq 0$$

now for any $h\left(x\right)$ we have $h\left(x\right)\neq 0$ iff $xh\left(1\right)\neq 0\vee x^{\prime}h\left(0\right)\neq 0$ (by Boole’s normal form), so we can rewrite the negative part as:

$$f\left(x\right)=0\wedge\bigwedge_{i}xf^{\prime}\left(1\right)g_{i}\left(f^{\prime}\left(1\right)\right)\neq 0\vee x^{\prime}f^{\prime}\left(0\right)g_{i}\left(f\left(0\right)\right)\neq 0$$

and using proposition 17 for the parts $f^{\prime}\left(1\right)g_{i}\left(f^{\prime}\left(1\right)\right)$ and $f^{\prime}\left(0\right)g_{i}\left(f\left(0\right)\right)$ we obtain the first result. Now simply account for the conditions of $f,g_{i}$ having zeros at all, and obtain the second result. ∎

Proof of corollary .

Assume $k=1$ for simplicity. Consider $\phi\left(x_{n-1},x_{n},y_{n-1},y_{n}\right)$. We understand $\phi$ as defining a relation between inputs and outputs at current time ($x_{n},y_{n}$ respectively) and in the previous time $x_{n-1},y_{n-1}$. Intuitively, it defines at least one BL$\left[k\right]$ function if the infinitary expression $\forall x_{1}\exists y_{1}\forall x_{2}\exists y_{2}\dots.\bigwedge_{n=2}^{\infty}\phi\left(x_{n-1},x_{n},y_{n-1},y_{n}\right)$ is satisfiable, alternatively if it is true in a model of choice. This infinitary expression can be given a concrete meaning by considering the first order theory containing all formulas of the form $\forall x_{1}\exists y_{1}\dots\forall x_{N}\exists y_{N}.\bigwedge_{n=2}^{N}\phi\left(x_{n-1},x_{n},y_{n-1},y_{n}\right)$ for all $N$. ∎

Proof of 35.

It is easy to see that we can always reduce into the collapsed fragment: this is immediate from the uniqueness of successor, for example $\forall n\exists k.s\left(n,k\right)\wedge\dots$ is same as $\forall nk.s\left(n,k\right)\rightarrow\dots$. For the normalized form, first convert the formula to DNF at its outermost level, so each literal may be a complex quantified formula, then collapse the quantifier alternation as above, so each quantified formula is either universal or existential. Moving to NNF we can consider universal and existential literals instead of positive and negative literals. In each DNF clause we can collapse the universal parts into a single one since universals distribute over conjunctions. Given an existential literal $\exists T.\gamma\left(T\right)\wedge\phi$ while denoting $T=t_{1},\dots,t_{k}$, we introduce a flag $e$ which is an additional output variable, and write

$$\left[\exists t.e\left(t\right)=0\right]\wedge\forall T.\gamma\left(T\right)\rightarrow\left[e\left(k-1\right)=1\wedge\left(e\left(t_{k}\right)=0\leftrightarrow\left(\psi\vee e\left(t_{k}-1\right)=0\right)\right)\right]$$

where $t_{k}=\max\left\{t_{1},\dots,t_{k}\right\}$ is assumed. The existential part is therefore reduced into a single atom at the expense of introducing a new output stream, and with introducing new universal literals which can then be collapsed into a single one as above. Given multiple single-atom existential parts $\bigwedge_{k}\exists t.e_{k}\left(t\right)=0$ we can easily see that they are equivalent to $\exists t.\left[\bigcup_{k}e_{k}\left(t\right)\right]=0$ because each flag remains zero once it becomes zero, so there is a point in time where all flags are eventually zero, so the existential part can be merely a single $\exists n.e\left(n\right)=0$ by defining this additional flag in the universal part. By that we reduced both the universal and the existential parts into a single one each. ∎

Proof of 39.

A model of size $n+1$ exists iff $\forall x_{0}\exists y_{0}\dots\forall x_{n}\exists y_{n}.\bigwedge_{m=k}^{n}\phi\left(X_{m}^{i},Y_{m}^{i}\right)$. Leaving free the first $k+1$ time points we can write

$$\phi_{n-k}\left(X_{k}^{i},Y_{k}^{i}\right):=\forall x_{k+1}\exists y_{k+1}\dots\forall x_{n}\exists y_{n}.\bigwedge_{m=k}^{n}\phi\left(X_{m}^{i},Y_{m}^{i}\right)$$

$$=\forall x_{k+1}\exists y_{k+1}\dots\forall x_{n}\exists y_{n}.\phi\left(X_{k}^{i},Y_{k}^{i}\right)\wedge\bigwedge_{m=k+1}^{n}\phi\left(X_{m}^{i},Y_{m}^{i}\right)$$

$$=\phi\left(X_{k}^{i},Y_{k}^{i}\right)\wedge\forall x_{k+1}\exists y_{k+1}\dots\forall x_{n}\exists y_{n}.\bigwedge_{m=k+1}^{n}\phi\left(X_{m}^{i},Y_{m}^{i}\right)$$

$$=\phi\left(X_{k}^{i},Y_{k}^{i}\right)\wedge\forall x_{k+1}\exists y_{k+1}.\phi_{n-k-1}\left(X_{k+1}^{i},Y_{k+1}^{i}\right)$$

since replacing $k$ with $k+1$ in $\phi_{n-k}\left(X_{k}^{i},Y_{k}^{i}\right):=\forall x_{k+1}\exists y_{k+1}\dots\forall x_{n}\exists y_{n}.\bigwedge_{m=k}^{n}\phi\left(X_{m}^{i},Y_{m}^{i}\right)$ results with $\phi_{n-k-1}\left(X_{k+1}^{i},Y_{k+1}^{i}\right):=\forall x_{k+2}\exists y_{k+2}\dots\forall x_{n}\exists y_{n}.\bigwedge_{m=k+1}^{n}\phi\left(X_{m}^{i},Y_{m}^{i}\right)$. In case that $\forall x_{0}\exists y_{0}\dots\forall x_{k}\exists y_{k}.\phi_{\infty}\left(X_{k}^{i},Y_{k}^{i}\right)$ then due to monotonicity, every $k+1$ subsequence of time points will have to satisfy $\phi_{\infty}\left(X_{k}^{i},Y_{k}^{i}\right)$, and any such subsequence can be extended arbitrarily due to the fact that it is a fixed point indeed. ∎