1 Introduction
Gödel’s first and second incompleteness theorems are some of the most important and profound results in the foundations of mathematics and have had wide influence on the development of logic, philosophy, mathematics, computer science, as well as other fields. Intuitively speaking, Gödel’s incompleteness theorems express that any rich enough logical system cannot prove its own consistency, i.e., that no contradiction like $0=1$ can be derived within this system.
Gödel [Reference Gödel43] proves his first incompleteness theorem $(\textsf {G1})$ for a certain formal system $\mathbf {P}$ related to Russell–Whitehead’s Principia Mathematica based on the simple theory of types over the natural number series and the Dedekind–Peano axioms (see [Reference Beklemishev8, p. 3]). Gödel announces the second incompleteness theorem $(\textsf {G2})$ in an abstract published in October 1930: no consistency proof of systems such as Principia, Zermelo–Fraenkel set theory, or the systems investigated by Ackermann and von Neumann is possible by methods which can be formulated in these systems (see [Reference Zach147, p. 431]).
Gödel comments in a footnote of [Reference Gödel43] that $\textsf {G2}$ is corollary of $\textsf {G1}$ (and in fact a formalized version of $\textsf {G1}$ ): if T is consistent, then the consistency of T is not provable in T where the consistency of T is formulated as the arithmetic formula which says that there exists an unprovable sentence in T. Gödel [Reference Gödel43] sketches a proof of $\textsf {G2}$ and promises to provide full details in a subsequent publication. This promise is not fulfilled, and a detailed proof of $\textsf {G2}$ for first-order arithmetic only appears in a monograph by Hilbert and Bernays [Reference Hilbert and Bernays59]. Abstract logic-free formulations of Gödel’s incompleteness theorems have been given by Kleene [Reference Kleene76] (“symmetric form”), Smullyan [Reference Smullyan121] (“representation systems”), and others. The following is a modern reformulation of Gödel’s incompleteness theorems.
Theorem 1.1 Gödel [Reference Gödel43]
Let T be a recursively axiomatized extension of $\mathbf {PA}$ .
-
G1 If T is $\omega $ -consistent, then T is incomplete.
-
G2 If T is consistent, then the consistency of T is not provable in T.
Gödel’s incompleteness theorems $\textsf {G1}$ and $\textsf {G2}$ are of a rather different nature and scope. In this paper, we will discuss different versions of $\textsf {G1}$ and $\textsf {G2}$ , from incompleteness for extensions of $\mathbf {PA}$ to incompleteness for systems weaker than $\mathbf {PA}$ w.r.t. interpretation. We will freely use $\textsf {G1}$ and $\textsf {G2}$ to refer to both Gödel’s first and second incompleteness theorems, and their different versions. The meaning of $\textsf {G1}$ and $\textsf {G2}$ will be clear from the context in which we refer to them.
Gödel’s incompleteness theorems exhibit certain weaknesses and limitations of a given formal system. For Gödel, his incompleteness theorems indicate the creative power of human reason. In Emil Post’s celebrated words: mathematical proof is an essentially creative activity (see [Reference Murawski97, p. 339]). The impact of Gödel’s incompleteness theorems is not confined to the community of mathematicians and logicians; popular accounts are well-known within the general scientific community and beyond. Gödel’s incompleteness theorems raise a number of philosophical questions concerning the nature of logic and mathematics as well as mind and machine. For the impact of Gödel’s incompleteness theorems, Feferman said:
their relevance to mathematical logic (and its offspring in the theory of computation) is paramount; further, their philosophical relevance is significant, but in just what way is far from settled; and finally, their mathematical relevance outside of logic is very much unsubstantiated but is the object of ongoing, tantalizing efforts (see [Reference Feferman34, p. 434]).
From the literature, there are some good textbooks and survey papers on Gödel’s incompleteness theorems. For textbooks, we refer to [Reference Boolos14, Reference Enderton29, Reference Franzen37, Reference Friedman40, Reference Hájek and Pudlák53, Reference Lindström91, Reference Murawski97, Reference Smith118, Reference Smullyan120, Reference Smullyan121]. For survey papers, we refer to [Reference Beklemishev8, Reference Blanck12, Reference Buldt16, Reference Cheng23, Reference Kotlarski79, Reference Smoryński and Barwise119, Reference Visser, Horsten and Welch133]. In the last 20 years, there have been a lot of advances in the study of incompleteness. We felt that a comprehensive survey paper for the current state-of-art of this research field is missing from the literature. The motivation of this paper is four-fold:
-
• Give the reader an overview of the current state-of-art of research on incompleteness.
-
• Classify these new advances on incompleteness under some important themes.
-
• Propose some new questions not covered in the literature.
-
• Set the direction for the future research of incompleteness.
Due to space limitations and our personal taste, it is impossible to cover all research results from the literature related to incompleteness in this survey. Therefore, we will focus on three aspects of new advances in research on incompleteness:
-
• classifications of different proofs of Gödel’s incompleteness theorems;
-
• the limit of the applicability of $\textsf {G1}$ ;
-
• the limit of the applicability of $\textsf {G2}$ .
We think these are the most important three aspects of research on incompleteness and reflect the depth and breadth of the research on incompleteness after Gödel. In this survey, we will focus on logical and mathematical aspects of research on incompleteness.
An important and interesting topic concerning incompleteness is missing in this paper: philosophy of Gödel’s incompleteness theorems. For us, the widely discussed and most important philosophical questions about Gödel’s incompleteness theorems are: the relationship between $\textsf {G1}$ and the mechanism thesis, the status of Gödel’s disjunctive thesis, and the intensionality problem of $\textsf {G2}$ . We leave a survey of philosophical discussions of Gödel’s incompleteness theorems for a future philosophy paper.
This paper is structured as follows. In Section 1, we introduce the motivation, the main content, and the structure of this paper. In Section 2, we list the preliminary notions and definitions used in this paper. In Section 3, we examine different proofs of Gödel’s incompleteness theorems and classify these proofs based on nine criteria. In Section 4, we examine the limit of the applicability of $\textsf {G1}$ both for extensions of $\mathbf {PA}$ , and for theories weaker than $\mathbf {PA}$ w.r.t. interpretation. In Section 5, we examine the limit of the applicability of $\textsf {G2}$ , and discuss sources of indeterminacy in the formulation of the consistency statement.
2 Preliminaries
2.1 Definitions and notations
We list the definitions and notations required below. These are standard and used throughout the literature.
Definition 2.1 Basic notions
-
• A language consists of an arbitrary number of relation and function symbols of arbitrary finite arity.Footnote 1 For a given theory T, we use $L(T)$ to denote the language of T, and often equate $L(T)$ with the list of non-logical symbols of the language.
-
• For a formula $\phi $ in $L(T)$ , ‘ $T\vdash \phi $ ’ denotes that $\phi $ is provable in T: i.e., there is a finite sequence of formulas $\langle \phi _0, \ldots ,\phi _n\rangle $ such that $\phi _n=\phi $ , and for any $0\leq i\leq n$ , either $\phi _i$ is an axiom of T, or $\phi _i$ follows from some $\phi _j\, (j<i)$ by using one inference rule.
-
• A theory T is consistent if no contradiction is provable in T.
-
• We say a sentence $\phi $ is independent of T if $T\not\vdash \phi $ and $T\not\vdash \neg \phi $ .
-
• A theory T is incomplete if there is a sentence $\phi $ in $L(T)$ which is independent of T; otherwise, T is complete (i.e., for any sentence $\phi $ in $L(T)$ , either $T\vdash \phi $ or $T\vdash \neg \phi $ ).
In this paper, we focus on first-order theories based on a countable language, and always assume the arithmetization of the base theory with a recursive set of non-logical symbols. For the technical details of arithmetization, we refer to [Reference Buss18, Reference Murawski97]. Arithmetization means that any formula or finite sequence of formulas can be coded by a natural number, called the Gödel number. This representation of syntax was pioneered by Gödel.
Definition 2.2 Basic notions following arithmetization
-
• We say a set of sentences $\Sigma $ is recursive if the set of Gödel numbers of sentences in $\Sigma $ is recursive.
-
• A theory T is decidable if the set of sentences provable in T is recursive; otherwise it is undecidable.
-
• A theory T is recursively axiomatizable if it has a recursive set of axioms (i.e., the set of Gödel numbers of axioms of T is recursive).
-
• A theory T is finitely axiomatizable if it has a finite set of axioms.
-
• A theory T is locally finitely satisfiable if every finitely axiomatized subtheory of T has a finite model.
-
• A theory T is recursively enumerable (r.e.) if it has a recursively enumerable set of axioms.
-
• A theory T is essentially undecidable if any recursively axiomatizable consistent extension of T in the same language is undecidable.
-
• A theory T is essentially incomplete if any recursively axiomatizable consistent extension of T in the same language is incomplete.Footnote 2
-
• A theory T is minimal essentially undecidable if T is essentially undecidable, and if deleting any axiom of T, the remaining theory is no longer essentially undecidable.
Definition 2.3 Basic notations
-
• We denote by $\overline {n}$ the numeral representing $n\in \omega $ in $L(\mathbf {PA})$ .
-
• We denote by $\ulcorner \phi \urcorner $ the numeral representing the Gödel number of $\phi $ .
-
• We denote by $\ulcorner \phi (\dot {x})\urcorner $ the numeral representing the Gödel number of the sentence obtained by replacing x with the value of x.Footnote 3
Definition 2.4 Representations, translations, and interpretations
-
• A n-ary relation $R(x_1, \ldots , x_n)$ on $\omega ^n$ is representable in T if there is a formula $\phi (x_1, \ldots , x_n)$ such that $T\vdash \phi (\overline {m_1}, \ldots , \overline {m_n})$ when $R(m_1, \ldots , m_n)$ holds, and $T\vdash \neg \phi (\overline {m_1}, \ldots , \overline {m_n})$ when $R(m_1, \ldots , m_n)$ does not hold.
-
• We say that a total function $f(x_1, \ldots , x_n)$ on $\omega ^n$ is representable in T if there is a formula $\varphi (x_1, \ldots , x_n,y)$ such that $T\vdash \forall y(\varphi (\overline {a_1}, \ldots , \overline {a_{n}},y)\leftrightarrow y=\overline {m})$ whenever $a_1, \ldots , a_n, m\in \omega $ are such that $f(a_1, \ldots , a_n)=m$ .
-
• Let T be a theory in a language $L(T)$ , and S a theory in a language $L(S)$ . In its simplest form, a translation I of language $L(T)$ into language $L(S)$ is specified by the following:
-
– an $L(S)$ -formula $\delta _I(x)$ denoting the domain of I;
-
– for each relation symbol R of $L(T)$ , as well as the equality relation =, an $L(S)$ -formula $R_I$ of the same arity;
-
– for each function symbol F of $L(T)$ of arity k, an $L(S)$ -formula $F_I$ of arity $k + 1$ .
-
-
• If $\phi $ is an $L(T)$ -formula, its I-translation $\phi ^I$ is an $L(S)$ -formula constructed as follows: we rewrite the formula in an equivalent way so that function symbols only occur in atomic subformulas of the form $F(\overline {x}) = y$ , where $x_i, y$ are variables; then we replace each such atomic formula with $F_I(\overline {x}, y)$ , we replace each atomic formula of the form $R(\overline {x})$ with $R_I(\overline {x})$ , and we restrict all quantifiers and free variables to objects satisfying $\delta _I$ . We take care to rename bound variables to avoid variable capture during the process.
-
• A translation I of $L(T)$ into $L(S)$ is an interpretation of T in S if S proves the following:
-
– for each function symbol F of $L(T)$ of arity k, the formula expressing that $F_I$ is total on $\delta _I$ :
$$ \begin{align*}&\forall x_0, \ldots, \forall x_{k-1} (\delta_I(x_0) \wedge \cdots \wedge \delta_I(x_{k-1}) \rightarrow \exists y (\delta_I(y)\\&\quad \wedge F_I(x_0, \ldots, x_{k-1}, y)));\end{align*} $$ -
– the I-translations of all axioms of T, and axioms of equality.
-
The simplified picture of translations and interpretations above actually describes only one-dimensional, parameter-free, and one-piece translations. For precise definitions of a multi-dimensional interpretation, an interpretation with parameters, and a piece-wise interpretation, we refer to [Reference Visser131–Reference Visser, Horsten and Welch133] for more details.
The notion of interpretation provides us with a method for comparing different theories in different languages, as follows.
Definition 2.5 Interpretations II
-
• A theory T is interpretable in a theory S if there exists an interpretation of T in S. If T is interpretable in S, then all sentences provable (refutable) in T are mapped, by the interpretation function, to sentences provable (refutable) in S.
-
• We say that a theory U weakly interprets a theory V (or V is weakly interpretable in U) if V is interpretable in some consistent extension of U in the same language (or equivalently, for some interpretation $\tau $ , the theory $U + V^{\tau }$ is consistent).
-
• Given theories S and T, let ‘ $S\unlhd T$ ’ denote that S is interpretable in T (or T interprets S); let ‘ $S\lhd T$ ’ denote that T interprets S but S does not interpret T; we say S and T are mutually interpretable if $S\unlhd T$ and $T\unlhd S$ .
Interpretability provides us with one measure of comparing strength of different theories. If theories S and T are mutually interpretable, then T and S are equally strong w.r.t. interpretation. In this paper, whenever we say that theory S is weaker than theory T w.r.t. interpretation, this means that $S\lhd T$ .
A general method for establishing the undecidability of theories is developed in [Reference Tarski, Mostowski and Robinson125]. The following theorem provides us with two methods for proving the essentially undecidability of a theory respectively via interpretation and representability.
Theorem 2.6 [Reference Tarski, Mostowski and Robinson125, Theorem 7 and Corollary 2]
-
• Let $T_1$ and $T_2$ be two consistent theories such that $T_2$ is interpretable in $T_1$ . If $T_2$ is essentially undecidable, then $T_1$ is essentially undecidable.
-
• If all recursive functions are representable in a consistent theory T, then T is essentially undecidable.
We shall also need some basic notions from recursion theory, as follows.
Definition 2.7 Basic recursion theory
-
• Let $\phi _0, \phi _1, \ldots $ be a list of all unary computable (partial recursive) functions such that $\phi _i(j)$ , if it exists, can be computed from i and j.
-
• A recursively enumerable set (r.e. for short) is the domain of $\phi _i$ for some $i \in \omega $ , which is denoted by $W_i$ .
-
• The notation $\phi _i(j)\uparrow $ means that the function $\phi _i$ is not defined at j, or $j \notin W_i$ ; and $\phi _i(j)\downarrow $ means that $\phi _i$ is defined at j, or $j \in W_i$ .
Provability logic provides us with an important tool to study the meta-mathematics of arithmetic and incompleteness. A good reference on the basics of provability logic is [Reference Boolos14].
Definition 2.8 Modal logic
-
• The modal system $\mathbf {K}$ consisting of the following axiom schemes:
-
– All tautologies;
-
– $\boxed{} (A\rightarrow B)\rightarrow (\boxed{} A\rightarrow \boxed{} B)$ ;
as well as two inference rules:
-
– if $\vdash A$ and $\vdash A\rightarrow B$ , then $\vdash B$ ;
-
– if $\vdash A$ , then $\vdash \boxed{} A$ .
-
-
• We denote by $\mathbf {GL}$ the modal system consisting of all axioms of $\mathbf {K}$ , all instances of the scheme $\boxed{} (\boxed{} A\rightarrow A)\rightarrow \boxed{} A$ , and the same inference rules with $\mathbf {K}$ .
-
• We denote by $\mathbf {GLS}$ the modal system consisting of all theorems of $\mathbf {GL}$ , and all instances of the scheme $\boxed{} A\rightarrow A$ . However, $\mathbf {GLS}$ has only one inference rule: Modus Ponens.
2.2 Logical systems
In this section, we introduce some well-known theories weaker than $\mathbf {PA}$ w.r.t. interpretation from the literature. In Section 4, we will show that these theories are essentially incomplete.
Robinson Arithmetic $\mathbf {Q}$ is introduced in [Reference Tarski, Mostowski and Robinson125] by Tarski, Mostowski, and Robinson as a base axiomatic theory for investigating incompleteness and undecidability.
Definition 2.9 Robinson Arithmetic $\mathbf {Q}$
Robinson Arithmetic $\mathbf {Q}$ is defined in the language $\{\mathbf {0}, \mathbf {S}, +, \times \}$ with the following axioms:
-
$\mathbf {Q}_1{:}$ $\forall x \forall y(\mathbf {S}x=\mathbf {S} y\rightarrow x=y)$ ;
-
$\mathbf {Q}_2{:}$ $\forall x(\mathbf {S} x\neq \mathbf {0})$ ;
-
$\mathbf {Q}_3{:}$ $\forall x(x\neq \mathbf {0}\rightarrow \exists y (x=\mathbf {S} y))$ ;
-
$\mathbf {Q}_4{:}$ $\forall x\forall y(x+ \mathbf {0}=x)$ ;
-
$\mathbf {Q}_5{:}$ $\forall x\forall y(x+ \mathbf {S} y=\mathbf {S} (x+y))$ ;
-
$\mathbf {Q}_6{:}$ $\forall x(x\times \mathbf {0}=\mathbf {0})$ ;
-
$\mathbf {Q}_7{:}$ $\forall x\forall y(x\times \mathbf {S} y=x\times y +x)$ .
Robinson Arithmetic $\mathbf {Q}$ is very weak: it cannot even prove that addition is associative.
Definition 2.10 Peano Arithmetic $\mathbf {PA}$
The theory $\mathbf {PA}$ consists of the axioms $\mathbf {Q}_1$ and $\mathbf {Q}_2$ , $\mathbf {Q}_4$ – $\mathbf {Q}_7$ in Definition 2.9, and the following axiom scheme of induction:
where $\phi $ is a formula with at least one free variable x. Let $\mathfrak {N}=\langle \mathbb {N}, +, \times \rangle $ denote the standard model of arithmetic.
We now introduce a well-known hierarchy of $L(\mathbf {PA})$ -formulas called the arithmetical hierarchy (see [Reference Hájek and Pudlák53, Reference Murawski97]).
Definition 2.11 Arithmetical hierarchy
-
• Bounded formulas ( $\Sigma ^0_0$ , or $\Pi ^0_0$ , or $\Delta ^0_0$ formulas) are built from atomic formulas using only propositional connectives and bounded quantifiers (in the form $\forall x\leq y$ or $\exists x\leq y$ ).
-
• A formula is $\Sigma ^0_{n+1}$ if it has the form $\exists x\phi $ where $\phi $ is $\Pi ^0_{n}$ .
-
• A formula is $\Pi ^0_{n+1}$ if it has the form $\forall x\phi $ where $\phi $ is $\Sigma ^0_{n}$ . Thus, a $\Sigma ^0_{n}$ -formula has a block of n alternating quantifiers, the first one being existential, and this block is followed by a bounded formula. Similarly for $\Pi ^0_{n}$ -formulas.
-
• A formula is $\Delta ^0_n$ if it is equivalent to both a $\Sigma ^0_{n}$ formula and a $\Pi ^0_{n}$ formula.
We can now formally introduce the notion of consistency in its various guises, as well as the various fragments of Peano Arithmetic $\mathbf {PA}$ .
Definition 2.12 Formal consistency and systems
-
• A theory T is said to be $\omega $ -consistent if there is no formula $\varphi (x)$ such that $T \vdash \exists x \varphi (x)$ , and for any $n\in \omega $ , $T \vdash \neg \varphi (\overline {n})$ .
-
• A theory T is $1$ -consistent if there is no such $\Delta ^0_1$ formula $\varphi (x)$ .
-
• We say a theory T is $\Sigma ^0_1$ -sound if for any $\Sigma ^0_1$ sentences $\phi $ , if $T\vdash \phi $ , then $\mathfrak {N}\models \phi $ .
-
• The collection axiom for $\Sigma ^0_{n+1}$ formulas is the following principle: $(\forall x< u)(\exists y) \varphi (x,y)\rightarrow (\exists v)(\forall x< u)(\exists y< v) \varphi (x,y)$ where $\varphi (x,y)$ is a $\Sigma ^0_{n+1}$ formula possibly containing parameters distinct from $u,v$ .
-
• The theory $I\Sigma _n$ is $\mathbf {Q}$ plus induction for $\Sigma ^0_n$ formulas, and $B\Sigma _{n+1}$ is $I\Sigma _0$ plus collection for $\Sigma ^0_{n+1}$ formulas.
-
• The theory $I\Delta _0$ is $\mathbf {Q}$ plus induction for $\Delta ^0_0$ formulas.
-
• The theory $\mathbf {PA}$ is the union of all $I\Sigma _{n}$ .
It is well-known that the following form a strictly increasing hierarchy:
Moreover, there are weak fragments of $\mathbf {PA}$ that play an important role in computer science, namely in complexity theory [Reference Buss17, Reference Buss18]. These systems are based on the following concept.
By [Reference Ferreira and Ferreira35, Proposition 2, p. 299], there is a bounded formula $\textsf {Exp(x, y, z)}$ such that $I\Sigma _0$ proves that $\textsf {Exp(x, 0, z)} \leftrightarrow z= 1$ , and $\textsf {Exp(x, Sy, z)} \leftrightarrow \exists t (\textsf {Exp(x, y, t)} \wedge z=t\cdot x)$ . However, $I\Sigma _0$ cannot prove the totality of $\textsf {Exp(x, y, z)}$ .
Definition 2.13 Sub-exponential functions
-
• Let $\mathbf {exp}$ denote the statement postulating the totality of the exponential function $\forall x\forall y\exists z \textsf {Exp(x, y, z)}$ .
-
• Elementary Arithmetic ( $\mathbf {EA}$ ) is $I\Delta _0 +\mathbf {exp}$ .
-
• Define $\omega _1(x)=x^{\mid x\mid }$ , and $\omega _{n+1}(x)=2^{\omega _{n}(\mid x\mid )}$ where $|x|$ is the length of the binary expression of x.
-
• Let $\Omega _{n}\equiv (\forall x)(\exists y) (\omega _{n}(x) =y)$ express that $\omega _{n}(x)$ is total.
Theorem 2.14 [Reference Ferreira and Ferreira35, Reference Hájek, Clote and Krajícek51]
-
• The theory $I\Sigma _0+\Omega _n$ is interpretable in $\mathbf {Q}$ for any $n\geq 1\ ($ see [Reference Ferreira and Ferreira35, Theorem 3, p. 304] $)$ .
-
• The theory $I\Sigma _0+ \mathbf {exp}$ is not interpretable in $\mathbf {Q}$ .Footnote 4
-
• The theory $I\Sigma _1$ is not interpretable in $I\Sigma _0+ \mathbf {exp}\ ($ see [Reference Hájek, Clote and Krajícek51, Theorem 1.1, p. 186] $)$ .
-
• The theory $I\Sigma _{n+1}$ is not interpretable in $B\Sigma _{n+1}\ ($ see [Reference Hájek, Clote and Krajícek51, Theorem 1.2, p. 186] $)$ .
-
• The theory $B\Sigma _{1}+ exp$ is interpretable in $I\Sigma _{0}+ \mathbf {exp}\ ($ see [Reference Hájek, Clote and Krajícek51, Theorem 2.4, p. 188] $)$ .
-
• The theory $B\Sigma _{1}+\Omega _{n}$ is interpretable in $I\Sigma _{0}+\Omega _{n}$ for each $n\geq 1\ ($ see [Reference Hájek, Clote and Krajícek51, Theorem 2.5, p. 189] $)$ .
-
• The theory $B\Sigma _{n+1}$ is interpretable in $I\Sigma _{n}$ for each $n\geq 0\ ($ see [Reference Hájek, Clote and Krajícek51, Theorem 2.6, p. 189] $)$ .
The theory $\mathbf {PA}^{-}$ is the theory of commutative, discretely ordered semi-rings with a minimal element plus the subtraction axiom. The theory $\mathbf {PA}^{-}$ has the following axioms, where the language $L(\mathbf {PA}^{-})$ is $L(\mathbf {PA})\cup \{\leq \}$ :
Definition 2.15 The system $\mathbf {PA}^{-}$
-
• $x + 0 = x$ ;
-
• $x + y = y + x$ ;
-
• $(x + y) + z = x + (y + z)$ ;
-
• $x \times 1 = x$ ;
-
• $x \times y = y \times x$ ;
-
• $(x \times y) \times z = x \times (y \times z)$ ;
-
• $x \times (y + z) = x \times y + x \times z$ ;
-
• $x \leq y \vee y \leq x$ ;
-
• $(x \leq y \wedge y \leq z)\rightarrow x \leq z$ ;
-
• $x + 1 \not\leq x$ ;
-
• $x \leq y \rightarrow (x = y \vee x + 1 \leq y)$ ;
-
• $x \leq y \rightarrow x + z \leq y + z$ ;
-
• $x \leq y \rightarrow x \times z \leq y \times z$ ;
-
• $x \leq y \rightarrow \exists z (x + z = y)$ .
The theory $\mathbf {Q}^+$ is the extension of $\mathbf {Q}$ in the language $L(\mathbf {Q}^+)=L(\mathbf {Q})\cup \{\leq \}$ with the following extra axioms:
Definition 2.16 The system $\mathbf {Q}^{+}$
The system $\mathbf {Q}^{+}$ is $\mathbf {Q}$ plus:
-
$\mathbf {Q}_8{:}$ $(x + y) + z = x + (y + z)$ ;
-
$\mathbf {Q}_9{:}$ $x \times (y + z) = x \times y + x \times z$ ;
-
$\mathbf {Q}_{10}{:}$ $(x \times y) \times z = x\times (y\times z)$ ;
-
$\mathbf {Q}_{11}{:}$ $x + y = y + x$ ;
-
$\mathbf {Q}_{12}{:}$ $x\times y = y \times x$ ;
-
$\mathbf {Q}_{13}{:}$ $x\leq y\leftrightarrow \exists z (x + z = y)$ .
Andrzej Grzegorczyk considers a theory $\mathbf {Q}^{-}$ in which addition and multiplication satisfy natural reformulations of the axioms of $\mathbf {Q}$ but are possibly non-total functions. More exactly, the language of $\mathbf {Q}^{-}$ is $\{\mathbf {0}, \mathbf {S}, A, M\}$ where A and M are ternary relations.
Definition 2.17 The system $\mathbf {Q}^{-}$
The axioms of $\mathbf {Q}^{-}$ are the axioms $\mathbf {Q}_1$ – $\mathbf {Q}_3$ of $\mathbf {Q}$ plus the following six axioms about A and M:
-
A: $\forall x\forall y\forall z_1\forall z_2(A(x, y, z_1)\wedge A(x, y, z_2) \rightarrow z_1 = z_2)$ ;
-
M: $\forall x\forall y\forall z_1\forall z_2(M(x, y, z_1) \wedge M(x, y, z_2) \rightarrow z_1 = z_2)$ ;
-
G4: $\forall x \, A(x, 0, x)$ ;
-
G5: $\forall x\forall y\forall z(\exists u(A(x, y, u) \wedge z = S(u)) \rightarrow A(x, S(y), z))$ ;
-
G6: $\forall x\, M(x, 0, 0)$ ;
-
G7: $\forall x\forall y\forall z(\exists u(M(x, y, u) \wedge A(u, x, z)) \rightarrow M(x, S(y), z))$ .
Samuel R. Buss [Reference Buss17] introduces $\mathbf {S^1_2}$ , a finitely axiomatizable theory, to study polynomial time computability. The theory $\mathbf {S^1_2}$ provides what is needed for formalizing the proof of $\textsf {G2}$ in a natural and effortless way: this process is actually easier in Buss’ theory than in full $\mathbf {PA}$ , since the restrictions present in $\mathbf {S^1_2}$ prevent one from making wrong turns and inefficient choices (see [Reference Visser132]).
Next, we introduce adjunctive set theory $\mathbf {AS}$ which has a language with only one binary relation symbol ‘ $\in $ ’.
Definition 2.18 Adjunctive set theory $\mathbf {AS}$ , [Reference Nelson98]
The axioms of $\mathbf {AS}$ consist of the following:
-
AS1: $\exists x \forall y (y \notin x)$ .
-
AS2: $\forall x\forall y \exists z \forall u (u \in z \leftrightarrow (u = x \vee u = y))$ .
We now consider the theory $\mathbf {R}$ introduced by A. Tarski, A. Mostowski, and R. Robinson in [Reference Tarski, Mostowski and Robinson125], and some variants of it.
Definition 2.19 The theory $\mathbf {R}$
Let $\mathbf {R}$ be the theory consisting of schemes $\textsf {Ax1}$ – $\textsf {Ax5}$ with $L(\mathbf {R})=\{\overline {0}, \ldots , \overline {n}, \ldots , +, \times , \leq \}$ where $m, n \in \omega $ .
-
Ax1: $\overline {m}+\overline {n}=\overline {m+n}$ ;
-
Ax2: $\overline {m}\times \overline {n}=\overline {m\times n}$ ;
-
Ax3: $\overline {m}\neq \overline {n}$ if $m\neq n$ ;
-
Ax4: $\forall x(x\leq \overline {n}\rightarrow x=\overline {0}\vee \cdots \vee x=\overline {n})$ ;
-
Ax5: $\forall x(x\leq \overline {n}\vee \overline {n}\leq x)$ .
As it happens, the system $\mathbf {R}$ contains all key properties of arithmetic for the proof of $\textsf {G1}$ . Unlike $\mathbf {Q}$ , the theory $\mathbf {R}$ is not finitely axiomatizable.
Definition 2.20 Variations of $\mathbf {R}$
-
• Let $\mathbf {R}_0$ be $\mathbf {R}$ without $\mathbf {Ax5}$ .
-
• Let $\mathbf {R}_1$ be the system consisting of schemes $\mathbf {Ax1}, \mathbf {Ax2}, \mathbf {Ax3}$ , and $\mathbf {Ax4^{\prime }}$ where the latter is as follows:
-
Ax4′ $\forall x(x\leq \overline {n}\leftrightarrow x=\overline {0}\vee \cdots \vee x=\overline {n}).$
-
-
• Let $\mathbf {R}_2$ be the system consisting of schemes $\mathbf {Ax2}, \mathbf {Ax3}$ , and $\mathbf {Ax4^{\prime }}$ .
The ‘concatenation’ theory $\mathbf {TC}$ has the language $\{\frown , \alpha , \beta \}$ with a binary function symbol and two constants.
Definition 2.21 The system $\mathbf {TC}$
-
TC1: $\forall x\forall y\forall z(x\frown (y\frown z) = (x\frown y)\frown z)$ ;
-
TC2: $\forall x\forall y\forall u\forall v(x\frown y = u\frown v \rightarrow ((x = u \wedge y = v) \vee \exists w((u = x\frown w \wedge w\frown v= y) \vee (x = u\frown w \wedge w\frown y = v))))$ ;
-
TC3: $\forall x\forall y(\alpha \neq x\frown y)$ ;
-
TC4: $\forall x\forall y(\beta \neq x\frown y)$ ;
-
TC5: $\alpha \neq \beta $ .
Primitive recursive arithmetic ( $\mathbf {PRA}$ ) is a quantifier-free formalization of the natural numbers, and the language of $\mathbf {PRA}$ can express arithmetic statements involving natural numbers and any primitive recursive function. Weak Konig’s Lemma ( $\mathbf {WKL_0}$ ) states that every infinite binary tree has an infinite branch. We refer to [Reference Hájek and Pudlák53, Reference Simpson117] for the definitions of $\mathbf {PRA}$ and $\mathbf {WKL_0}$ . In a nutshell, the former system allows us to perform ‘iteration of functions $f:\mathbb {N} \rightarrow \mathbb {N}$ ’, while the latter expresses a basic compactness argument for Cantor space.
Theorem 2.22 Friedman’s conservation theorem [Reference Kikuchi and Tanaka74, Theorem 2.1]
For any $\Pi ^0_2$ sentence $\phi $ in $L(\mathbf {PA})$ , if $\mathbf {WKL_0}\vdash \phi $ , then $\mathbf {PRA}\vdash \phi $ .
Finally, diagnolisation, in one form or other, forms the basis for the proof of $\textsf {G2}$ . The following lemma is crucial in this regard.
Lemma 2.23 The Diagnolisation Lemma
Let T be a consistent r.e. extension of $\mathbf {Q}$ . For any formula $\phi (x)$ with exactly one free variable, there exists a sentence $\theta $ such that $T\vdash \theta \leftrightarrow \phi (\ulcorner \theta \urcorner )$ .
Lemma 2.23 is the simplest and most often used version of the Diagnolisation Lemma. For a generalized version of the Diagnolisation Lemma, we refer to [Reference Boolos14]. In this paper, we use the term “Diagnolisation Lemma” to refer to Lemma 2.23 and some variants of the generalized version.
3 Proofs of Gödel’s incompleteness theorems
3.1 Introduction
In this section, we discuss different proofs of Gödel’s incompleteness theorems from the literature, and propose nine criteria for classifying them.
First of all, there are no requirements on the independent sentence in $\textsf {G1}$ . In particular, such a sentence need not have any mathematical meaning. This is often the case when meta-mathematical (proof-theoretic or recursion-theoretic or model-theoretic) methods are used to construct the independent sentence. In Sections 3.2–3.4, we will discuss proofs of Gödel’s incompleteness theorems via pure logic. In Section 3.5, we will give an overview of the “concrete incompleteness” research program which seeks to identify natural independent sentences with real mathematical meaning.
Secondly, we say that a proof of $\textsf {G1}$ is constructive if it explicitly constructs the independent sentence from the base theory by algorithmic means. A non-constructive proof of $\textsf {G1}$ only proves the mere existence of the independent sentence and does not show its existence algorithmically. We say that a proof of $\textsf {G1}$ for theory T has the Rosser property if the proof only assumes that T is consistent instead of assuming that T is $\omega $ -consistent or 1-consistent or $\Sigma ^0_1$ -sound; all these notions are introduced in Section 2.2.
After Gödel, many different proofs of Gödel’s incompleteness theorems have been found. These proofs can be classified using the following criteria:
-
• proof-theoretic proof;
-
• recursion-theoretic proof;
-
• model-theoretic proof;
-
• proof via arithmetization;
-
• proof via the Diagnolisation Lemma;
-
• proof based on “logical paradox”;
-
• constructive proof;
-
• proof having the Rosser property;
-
• the independent sentence has natural and real mathematical content.Footnote 5
However, these aspects are not exclusive: a proof of $\textsf {G1}$ or $\textsf {G2}$ may satisfy several of the above criteria.
Thirdly, there are two kinds of proofs of Gödel’s incompleteness theorems via pure logic: one based on logical paradox and one not based on logical paradox. In Section 3.2, we first provide an overview of the modern reformulation of proofs of Gödel’s incompleteness theorems. We discuss proofs of Gödel’s incompleteness theorems not based on logical paradox in Section 3.3. We discuss proofs of Gödel’s incompleteness theorems based on logical paradox in Section 3.4.
3.2 Overview and modern formulation
In a nutshell, the three main ideas in the (modern/standard) proofs of $\textsf {G1}$ and $\textsf {G2}$ are arithmetization, representability, and self-reference, as discussed in detail in Section 3.2.1. Interesting properties of $\textsf {G1}$ and $\textsf {G2}$ are discussed in Sections 3.2.2 and 3.2.4, while the formalized notions of ‘proof’ and ‘truth’ are discussed in Section 3.2.3. Finally, we formulate a blanket caveat for the rest of this section:
Unless stated otherwise, we will always assume that T is a recursively axiomatizable consistent extension of $\mathbf {Q}$ .
Other sections shall contain similar caveats and we sometimes stress these.
3.2.1 Three steps towards G1 and G2
Intuitively speaking, Gödel’s incompleteness theorems can be proved based on the following key ingredients.
-
• Arithmetization: since G1 and G2 are theorems about properties of the syntax of logic, we need to somehow represent the latter, which is done via a coding scheme called arithmetization.
-
• Representations: the notion of ‘proof’ and related concepts in G1 and G2 are then expressed (‘represented’) via arithmetization.
-
• Self-reference: given a representation of ‘proof’ and related concepts, one can write down formal statements that intuitively express ‘self-referential’ things like ‘this sentence does not have a proof’.
As we will see, the intuitively speaking ‘self-referential’ statements are the key to proving G1 and G2. We now discuss these three notions in detail.
First of all, arithmetization has the following intuitive content: it establishes a one-to-one correspondence between expressions of $L(T)$ and natural numbers. Thus, we can translate metamathematical statements about the formal theory T into statements about natural numbers. Furthermore, fundamental metamathematical relations can be translated in this way into certain recursive relations, hence into relations representable in T. Consequently, one can speak about a formal system of arithmetic, and about its properties as a theory in the system itself (see [Reference Murawski97])! This is the essence of Gödel’s idea of arithmetization, which was revolutionary at a time when computer hardware and software did not exist yet.
Secondly, in light of the previous, we can define certain relations on natural numbers that express or represent crucial metamathematical concepts related to the formal system T, like ‘proof’ and ‘consistency’. For example, modulo plenty of technical details, we can readily define a binary relation on $\omega ^2$ expressing what it means to prove a formula in T, namely as follows:
$\textit{Proof}_T(m,n)$ if and only if n is the Gödel number of a proof in T of the formula with Gödel number m.
Moreover, we can show that the relation $\textit{Proof}_T(m,n)$ is recursive. In addition, Gödel proves that every recursive relation is representable in ${\mathbf {PA}}$ .
Next, let $\mathbf {Proof}_T(x,y)$ be the formula which represents $\textit {Proof}_T(m,n)$ in $\mathbf {PA}$ .Footnote 6 From the formula $\mathbf {Proof}_T(x,y)$ , we can define the ‘provability’ predicate $\mathbf {Prov}_T(x)$ as $\exists y \mathbf {Proof}_T(x,y)$ . The provability predicate $\mathbf {Prov}_T(x)$ satisfies the following conditions which show that formal and intuitive provability have the same properties.
-
(1) If $T \vdash \varphi $ , then $T \vdash \mathbf {Prov}_T(\ulcorner \varphi \urcorner )$ ;
-
(2) $T \vdash \mathbf {Prov}_T(\ulcorner \varphi \rightarrow \psi \urcorner )\rightarrow (\mathbf {Prov}_T(\ulcorner \varphi \urcorner )\rightarrow \mathbf {Prov}_T(\ulcorner \psi \urcorner ))$ ;
-
(3) $T\vdash \mathbf {Prov}_T(\ulcorner \varphi \urcorner ) \rightarrow \mathbf {Prov}_T(\ulcorner \mathbf {Prov}_T(\ulcorner \varphi \urcorner )\urcorner )$ .
For the proof of $\textsf {G1}$ , Gödel defines the Gödel sentence $\mathbf {G}$ which asserts its own unprovability in T via a self-reference construction. Gödel shows that if T is consistent, then $T\not\vdash \mathbf {G}$ , and if T is $\omega $ -consistent, then $T\not\vdash \neg \mathbf {G}$ . One way of obtaining such a Gödel sentence is the Diagnolisation Lemma which intuitively speaking implies that the predicate $\neg \mathbf {Prov}_T(x)$ has a fixed point, i.e., there is a sentence $\theta $ in $L(T)$ such that
Clearly, $T\not\vdash \theta $ while $\theta $ intuitively expresses its own unprovability, i.e., the aforementioned self-referential nature.
For the proof of $\textsf {G2}$ , we first define the arithmetic sentence $\mathbf {Con}(T)$ in $L(T)$ as $\neg \mathbf {Prov}_T(\ulcorner \mathbf {0}\neq \mathbf {0}\urcorner )$ which says that for all x, x is not a code of a proof of a contradiction in T. Gödel’s second incompleteness theorem ( $\textsf {G2}$ ) states that if T is consistent, then the arithmetical formula $\mathbf {Con}(T)$ , which expresses the consistency of T, is not provable in T. In Section 5.3, we will discuss some other ways of expressing the consistency of T.
Finally, from the above conditions (1)–(3), one can show that $T\vdash \mathbf {Con}(T)\leftrightarrow \mathbf {G}$ . Thus, $\textsf {G2}$ holds: if T is consistent, then $T\not\vdash \mathbf {Con}(T)$ . For more details on these proofs of $\textsf {G1}$ and $\textsf {G2}$ , we refer to Chapter 2 in [Reference Murawski97].
3.2.2 Properties of $\mathsf {G1}$
In this section, we discuss some (sometimes subtle) comments on $\textsf {G1}$ .
First of all, Gödel’s proof of $\textsf {G1}$ is constructive as follows: given a consistent r.e. extension T of $\mathbf {PA}$ , the proof constructs, in an algorithmic way, a true arithmetic sentence which is unprovable in T. In fact, one can effectively find a true $\Pi ^0_1$ sentence $G_T$ of arithmetic such that $G_T$ is independent of T. Gödel calls this the “incompletability or inexhaustability of mathematics”.
Secondly, for Gödel’s proof of $\textsf {G1}$ , only assuming that T is consistent does not suffice to show that Gödel sentence is independent of T. In fact, the optimal condition to show that Gödel sentence is independent of T is: $T+\mathbf {Con}(T)$ is consistent (see Theorems 35 and 36 in [Reference Isaacson, DeVidi, Hallett and Clark62]).Footnote 7
Thirdly, in summary, Gödel’s proof of $\textsf {G1}$ has the following properties:
-
• uses proof-theoretic method with arithmetization;
-
• does not directly use the Diagnolisation Lemma;
-
• the proof formalizes the liar paradox;
-
• the proof is constructive;
-
• the proof does not have the Rosser property;
-
• Gödel’s sentence has no real mathematical content.
All these characteristics of Gödel’s proof of $\textsf {G1}$ are not necessary conditions for proving $\textsf {G1}$ . For example, $\textsf {G1}$ can be proved using recursion-theoretic or model-theoretic method, using the Diagnolisation Lemma, using other logical paradoxes, using non-constructive methods, only assuming that T is consistent (i.e., having the Rosser property), and can be proved without arithmetization.
Fourth, $\textsf {G1}$ does not tell us that any consistent theory is incomplete. In fact, there are many consistent complete first-order theories. For example, the following first-order theories are complete: the theory of dense linear orderings without endpoints ( $\mathbf {DLO}$ ), the theory of ordered divisible groups ( $\mathbf {ODG}$ ), the theory of algebraically closed fields of given characteristic ( $\mathbf {ACF_p}$ ), and the theory of real closed fields ( $\mathbf {RCF}$ ). We refer to [Reference Enderton and Szczerba31] for details of these theories. In fact, $\textsf {G1}$ only tells us that any consistent first-order theory containing a large enough fragment of $\mathbf {PA}$ (such as $\mathbf {Q}$ ) is incomplete: there is then a true $\Pi ^0_1$ sentence which is independent of the initial theory. Turing’s work in [Reference Turing126] shows that any true $\Pi ^0_1$ -sentence of arithmetic is provable in some transfinite iteration of ${\mathbf {PA}}$ . Feferman’s work in [Reference Feferman33] extends Turing’s work and shows that any true sentence of arithmetic is provable in some transfinite iteration of $\mathbf {PA}$ .
Fifth, whether a theory of arithmetic is complete depends on the language of the theory. There are respectively recursively axiomatized complete arithmetic theories in the language of $\, L(\mathbf {0}, \mathbf {S})$ , $L(\mathbf {0}, \mathbf {S}, <)$ , and $L(\mathbf {0}, \mathbf {S}, <, +)$ (see Sections 3.1 and 3.2 in [Reference Enderton29]). Containing enough information of arithmetic is essential for a consistent arithmetic theory to be incomplete. For example, Euclidean geometry is not about arithmetic but only about points, circles, and lines in general; but Euclidean geometry is complete as Tarski has proved (see [Reference Tarski and Givant124]). If the theory contains only information about the arithmetic of addition without multiplication, then it can be complete. For example, Presburger arithmetic is a complete theory of the arithmetic of addition in the language of $L(\mathbf {0}, \mathbf {S}, +)$ (see [Reference Murawski97, Theorem 3.2.2, p. 222]). Finally, containing the arithmetic of multiplication is not sufficient for a theory to be incomplete. For example, there exists a complete recursively axiomatized theory in the language of $L(\mathbf {0}, \times )$ (see [Reference Murawski97, p. 230]).
Finally, it is well-known that $Th(\mathbb {N}, +, \times )$ is interpretable in $Th(\mathbb {Z}, +, \times )$ and $Th(\mathbb {Q}, +, \times )$ .Footnote 8 Since $Th(\mathbb {N}, +, \times )$ is undecidable and has a finitely axiomatizable incomplete sub-theory $\mathbf {Q}$ , by Theorem 2.6, $Th(\mathbb {Z}, +, \times )$ and $Th(\mathbb {Q}, +, \times )$ are undecidable, and hence not recursively axiomatizable, but they respectively have a finitely axiomatizable incomplete sub-theory of integers and rational numbers. But $Th(\mathbb {R}, +, \times )$ is decidable and recursively axiomatizable (even if not finitely axiomatizable). In fact, $Th(\mathbb {R}, +, \times )=\mathbf {RCF}$ (the theory of real closed field) (see [Reference Enderton and Szczerba31, p. 320–321]). Note that this fact does not contradict $\textsf {G1}$ since none of $\mathbb {N}, \mathbb {Z}$ , and $\mathbb {Q}$ is definable in $(\mathbb {R}, +, \times )$ .
3.2.3 Between truth and provability
In this paper, unless stated otherwise, we equate a set of sentences with the set of Gödel’s numbers of these sentences. We discuss the formalized notions of ‘truth’ and ‘proof’, and how they relate to incompleteness.
Definition 3.1 We define $\textbf {Truth}=\{\phi \in L(\mathbf {PA}): \mathfrak {N}\models \phi \}$ and $\textbf {Prov}=\{\phi \in L(\mathbf {PA}): \mathbf {PA}\vdash \phi \}$ , i.e., the formalized notions of ‘proof’ and ‘truth’.
First of all, truth and provability are the same for purely existential statements. Put another way, incompleteness does not arise at the level of $\Sigma ^0_1$ sentences. Indeed, we have $\Sigma ^0_1$ -completeness for T: for any $\Sigma ^0_1$ sentences $\phi $ , $T\vdash \phi $ if and only if $\mathfrak {N}\models \phi $ . Thus, Gödel’s sentence is a true $\Pi ^0_1$ sentence in the form $\forall x \phi (x)$ such that $T\nvdash \forall x \phi (x)$ but ‘ $T\vdash \phi (\bar {n})$ ’ holds for any $n\in \omega $ .
Secondly, the properties of $\textbf {Truth}$ are essentially different from that of $\textbf {Prov}$ . Before Gödel’s work, it was thought that $\textbf {Truth}=\textbf {Prov}$ . Thus, Gödel’s first incompleteness theorem( $\textsf {G1}$ ) reveals the difference between the notion of provability in ${\mathbf {PA}}$ and the notion of truth in the standard model of arithmetic $\mathfrak {N}$ . There are some differences between $\textbf {Truth}$ and $\textbf {Prov}$ :
-
• $\textbf {Prov}\subsetneq \textbf {Truth}$ , i.e., there is a true arithmetic sentence which is unprovable in $\mathbf {PA}$ ;
-
• Tarski proves that: $\textbf {Truth}$ is not definable in $\mathfrak {N}$ but $\textbf {Prov}$ is definable in $\mathfrak {N}$ ;
-
• $\textbf {Truth}$ is not arithmetic but $\textbf {Prov}$ is recursive enumerable.
However, both $\textbf {Truth}$ and $\textbf {Prov}$ are not recursive and not representable in $\mathbf {PA}$ . For more details on $\textbf {Truth}$ and $\textbf {Prov}$ , we refer to [Reference Murawski97, Reference Tarski, Mostowski and Robinson125].
Thirdly, the differences between $\textbf {Truth}$ and $\textbf {Prov}$ can also be expressed in terms of arithmetical interpretations, defined as follows.
Definition 3.2 Arithmetical interpretations
A mapping from the set of all modal propositional variables to the set of $L(\mathbf {PA})$ -sentences is called an arithmetical interpretation .
Every arithmetical interpretation f is uniquely extended to the mapping $f^{\ast }$ from the set of all modal formulas to the set of $L(T)$ -sentences so that $f^{\ast }$ satisfies the following conditions:
-
• $f^{\ast }(p)=f(p)$ for each propositional variable p;
-
• $f^{\ast }$ commutes with every propositional connective;
-
• $f^{\ast }(\boxed{} A)$ is $\mathbf {Prov}_T(\ulcorner f^{\ast }(A)\urcorner )$ for every modal formula A.
In the following, we equate arithmetical interpretations f with their unique extensions $f^{\ast }$ defined on the set of all modal formulas. In this way, Solovay’s Arithmetical Completeness Theorems for $\mathbf {GL}$ and $\mathbf {GLS}$ characterize the difference between $\textbf {Prov}$ and $\textbf {Truth}$ via provability logic.
Theorem 3.3 Solovay [Reference Solovay122]
-
Arithmetical Completeness Theorem for GL Let T be a $\Sigma ^0_1$ -sound r.e. extension of $\mathbf {Q}$ . For any modal formula $\phi $ in $L(\mathbf {GL})$ , $\mathbf {GL}\vdash \phi $ if and only if $T\vdash f(\phi )$ for every arithmetic interpretation f.
-
Arithmetical Completeness Theorem for GLS For any modal formula $\phi , \mathbf {GLS}\vdash \phi $ if and only if $\mathfrak {N}\models f(\phi )$ for every arithmetic interpretation f.
Finally, one can study the notion of ‘proof predicate’ as given by $\mathbf {Proof}_T(x, y)$ in an abstract setting, namely as follows. Recall that T is a recursively axiomatizable consistent extension of $\mathbf {Q}$ . We introduce general notions of proof predicate and provability predicate which generalize the proof predicate $\mathbf {Proof}_T(x, y)$ and the provability predicate $\mathbf {Prov}_T(x)$ defined above in Gödel’s proof of $\textsf {G1}$ .
Definition 3.4 Proof predicate
We say a formula $\mathbf {Prf}_T(x, y)$ is a proof predicate of T if it satisfies the following conditions:Footnote 9
-
• $\mathbf {Prf}_T(x, y)$ is $\Delta ^0_1(\mathbf {PA})$ ;Footnote 10
-
• $\mathbf {PA} \vdash \forall x(\mathbf {Prov}_T(x) \leftrightarrow \exists y \mathbf {Prf}_T(x, y))$ ;
-
• for any $n \in \omega $ and formula $\phi , \mathbb {N}\models \mathbf {Proof}_T(\ulcorner \phi \urcorner , \overline {n}) \leftrightarrow \mathbf {Prf}_T(\ulcorner \phi \urcorner , \overline {n})$ ;
-
• $\mathbf {PA} \vdash \forall x\forall x^{\prime } \forall y (\mathbf {Prf}_T(x, y) \wedge \mathbf {Prf}_T(x^{\prime }, y) \rightarrow x = x^{\prime })$ .
Definition 3.5 Provability and consistency
We define the provability predicate $\mathbf {Pr}_T(x)$ from a proof predicate $\mathbf {Prf}_T(x,y)$ by $\exists y\, \mathbf {Prf}_T(x,y)$ , and the consistency statement $\mathbf {Con}(T)$ from a provability predicate $\mathbf {Pr}_T(x)$ by $\neg \mathbf {Pr}_T (\ulcorner \mathbf {0}\neq \mathbf {0}\urcorner )$ .
The items $\mathbf {D1}$ – $\mathbf {D3}$ below are called the Hilbert–Bernays–Löb derivability conditions. Note that $\mathbf {D1}$ holds for any provability predicate $\mathbf {Pr}_T(x)$ .
Definition 3.6 Standard proof predicate
We say that provability predicate $\mathbf {Pr}_T(x)$ is standard if it satisfies $\mathbf {D2}$ and $\mathbf {D3}$ as follows.
-
D1: If $T \vdash \phi $ , then $T \vdash \mathbf {Pr}_T(\ulcorner \phi \urcorner )$ ;
-
D2: If $T \vdash \mathbf {Pr}_T(\ulcorner \phi \rightarrow \varphi \urcorner ) \rightarrow (\mathbf {Pr}_T(\ulcorner \phi \urcorner )\rightarrow \mathbf {Pr}_T(\ulcorner \varphi \urcorner ))$ ;
-
D3: $T \vdash \mathbf {Pr}_T(\ulcorner \phi \urcorner )\rightarrow \mathbf {Pr}_T(\ulcorner \mathbf {Pr}_T(\ulcorner \phi \urcorner )\urcorner )$ .
We say that $\mathbf {Prf}_T(x, y)$ is a standard proof predicate if the induced provability predicate from it is standard.
The previous definition leads to another blanket caveat:
Unless stated otherwise, we always assume that $\mathbf {Pr}_T(x)$ is a standard provability predicate, and $\mathbf {Con}(T)$ is the canonical consistency statement defined as $\neg \mathbf {Pr}_T(\ulcorner \mathbf {0}\neq \mathbf {0}\urcorner )$ via the standard provability predicate $\mathbf {Pr}_T(x)$ .
3.2.4 Properties of $\textsf {G2}$
In this section, we discuss some (sometimes subtle) comments on $\textsf {G2}$ .
First of all, we examine a somewhat delicate mistake in the argument which claims that, by an easy application of the compactness theorem, we can show that for any recursive axiomatization of a consistent theory T, T cannot prove its own consistency. Visser presents this argument in [Reference Visser136] as an interesting dialogue between Alcibiades and Socrates:
Suppose a consistent theory T can prove its own consistency under some axiomatization. By compactness theorem, there must be a finitely axiomatized sub-theory S of T such that S already proves the consistency of T. Since S proves the consistency of T, it must also prove the consistency of S. So, we have a finitely axiomatized theory which proves its own consistency. But $\textsf {G2}$ applies to the finite axiomatization and we have a contradiction. It follows that T cannot prove its own consistency.
The mistake in this argument is: from the fact that S can prove the consistency of T we cannot infer that S can prove the consistency of S. Some may argue that since S is a sub-theory of T and S can prove the consistency of T, then of course S can prove the consistency of S.
However, as Visser correctly points out in [Reference Visser136], we should carefully distinguish three perspectives of the theory T: our external perspective, the internal perspective of S, and the internal perspective of T. From each perspective, the consistency of the whole theory implies the consistency of its sub-theory. From T’s perspective, S is a sub-theory of T. But from S’s perspective, S may not be a sub-theory of T. From the fact that T knows that S is a sub-theory of T, we cannot infer that S also knows that S is a sub-theory of T since S is a finite sub-theory of T and may not know any information that T knows, leading to the following (dramatic) conclusion:
the sub-theory relation between theories is not absolute.
Similarly, the notion of consistency is not absolute. For example, let $S=\mathbf {PA}+ \neg \mathbf {Con}(\mathbf {PA})$ . From $\textsf {G2}$ , S is consistent from the external perspective. But since $S\vdash \neg \mathbf {Con}(S)$ , the theory S is not consistent from the internal perspective of S. Note that $\mathbf {PA}\vdash \mathbf {Pr}_{\mathbf {PA}}(\mathbf {0}\neq \mathbf {0})\rightarrow \mathbf {Pr}_{\mathbf {PA}}(\mathbf {Pr}_{\mathbf {PA}}(\mathbf {0}\neq \mathbf {0})\rightarrow \mathbf {0}\neq \mathbf {0})$ . Thus, a theory may be consistent from the external perspective but inconsistent from the internal perspective.
From Gödel’s proof of $\textsf {G2}$ , we cannot infer that if T is a consistent r.e. extension of $\mathbf {Q}$ , then $\mathbf {Con}(T)$ is independent of T. The key point is: it is not enough to show that $T\nvdash \neg \mathbf {Con}(T)$ only assuming that T is consistent. However, we can show that $\mathbf {Con}(T)$ is independent of T assuming that T is 1-consistent.Footnote 11 In fact, the formalized version of “if T is consistent, then $\mathbf {Con}(T)$ is independent of T” is not provable in T.Footnote 12
Definition 3.7 Reflexivity
-
• A first-order theory T containing $\mathbf {PA}$ is said to be reflexive if $T \vdash \mathbf {Con}(S)$ for each finite sub-theory S of T where $\mathbf {Con}(S)$ is similarly defined as $\mathbf {Con}(\mathbf {PA})$ .
-
• We say the theory T is essentially reflexive if any consistent extension of T in $L(T)$ is reflexive.
-
• Let $\mathbf {Con}(T)\!\!\upharpoonright x$ denote the finite consistency statement “there are no proofs of contradiction in T with $\leq x$ symbols”.
Mostowski proves that $\mathbf {PA}$ is essentially reflexive (see [Reference Murawski97, Theorem 2.6.12]). In fact one can show that for every $n \in \mathbb {N}$ , $I\Sigma ^0_{n+1}\vdash \mathbf {Con}(I\Sigma ^0_{n})$ .Footnote 13 For a large class of natural theories U, Pudlák [Reference Pudlák109] shows that the lengths of the shortest proofs of $\mathbf {Con}(U)\!\!\upharpoonright n$ for $n\in \omega $ in the theory U itself are bounded by a polynomial in n. Pudlák conjectures [Reference Pudlák109] that U does not have polynomial proofs of the finite consistency statements $\mathbf {Con}(U + \mathbf {Con}(U))\upharpoonright n$ for $n\in \omega $ .
Finally, a big open question about $\textsf {G2}$ is: can we find a genuinely self-reference free proof of $\textsf {G2}$ ? As far as we know, at present there is no convincing essentially self-reference-free proofs of either $\textsf {G2}$ or of Tarski’s Theorem of the Undefinability of Truth. In [Reference Visser135], Visser gives a self-reference-free proof of $\textsf {G2}$ from Tarski’s Theorem of the Undefinability of Truth, which is a step in a program to find self-reference-free proofs of both $\textsf {G2}$ and Tarski’s Theorem (see [Reference Visser135]). Visser’s argument in [Reference Visser135] is model-theoretic and the main tool is the Interpretation Existence Lemma.Footnote 14 Visser’s proof in [Reference Visser135] is not constructive. An interesting question is then whether Visser’s argument can be made constructive.
3.3 Proofs of $\textsf{G1}$ and $\textsf{G2}$ from mathematical logic
In this section, we discuss various different proofs of $\textsf {G1}$ and $\textsf {G2}$ . We mention Jech’s [Reference Jech63] short proof of $\textsf {G2}$ for $\mathbf {ZF}$ : if $\mathbf {ZF}$ is consistent, then it is unprovable in $\mathbf {ZF}$ that there exists a model of $\mathbf {ZF}$ . Jech’s proof uses the Completeness Theorem, and also yields $\textsf {G2}$ for $\mathbf {PA}$ (see [Reference Jech63]). Other (lengthier) proofs are discussed in Sections 3.3.1–3.3.5.
3.3.1 Rosser’s proof
Rosser [Reference Rosser111] proves a “stronger” version of $\textsf {G1}$ , called Rosser’s first incompleteness theorem, which only assumes the consistency of T: if T is a consistent r.e. extension of $\mathbf {Q}$ , then T is incomplete. Gödel’s proof of $\textsf {G1}$ assumes that T is $\omega $ -consistent. Note that $\omega $ -consistency implies consistency. But the converse does not hold and the notion of $\omega $ -consistency is stronger than consistency since we can find examples of theories that are consistent but not $\omega $ -consistent.Footnote 15 Rosser’s proof is constructive and algorithmically constructs the Rosser sentence that is independent of T. Gödel’s proof of $\textsf {G1}$ uses a standard provability predicate but Rosser’s proof of $\textsf {G1}$ uses a Rosser provability predicate which is a kind of non-standard provability predicate, giving rise to the following.
Definition 3.8. Let T be a recursively axiomatizable consistent extension of $\mathbf {Q}$ , and $\mathbf {Prf}_T(x, y)$ be any proof predicate of T. Define the Rosser provability predicate $\mathbf {Pr}_T^R(x)$ to be the formula $\exists y(\mathbf {Prf}_T(x, y) \wedge \forall z \leq y \neg \mathbf {Prf}_T(\dot {\neg } x, z))$ where $\dot {\neg }$ is a function symbol expressing a primitive recursive function calculating the code of $\neg \phi $ from the code of $\phi $ . The fixed point of the predicate $\neg \mathbf {Pr}_T^R(x)$ is called the Rosser sentence of $\mathbf {Pr}_T^R(x)$ , i.e., a sentence $\theta $ satisfying $\mathbf {PA} \vdash \theta \leftrightarrow \neg \mathbf {Pr}_T^R(\ulcorner \theta \urcorner )$ .
In general, one can show that each Rosser sentence based on any Rosser provability predicate of T is independent of T. In particular, this independence does not rely on the choice of the proof predicate.
3.3.2 Recursion-theoretic proofs
Gödel’s first incompleteness theorem ( $\textsf {G1}$ ) is well-known in the context of recursion theory. Recall that $W_e=\{n\in \omega : \phi _e(n)\!\!\downarrow \}$ . Let $\langle W_e: e\in \omega \rangle $ be the list of recursive enumerable subsets of $\mathbb {N}$ . The following is an example of an ‘effective’ version of $\textsf {G1}$ (see [Reference Enderton30]):
there exists a recursive function f such that for any $e\in \omega $ , if $W_e\subseteq \textbf {Truth}$ , then $f(e)$ is defined and $f(e)\in \textbf {Truth}\setminus W_e$ .
Similarly, Avigad [Reference Avigad4] proves $\textsf {G1}$ and $\textsf {G2}$ in terms of the undecidability of the halting problem (see Theorems 3.1 and 3.2 in [Reference Avigad4]). Another related result due to Kleene is as follows.
Theorem 3.9 Kleene’s theorem [Reference Salehi and Seraji114, Theorem 2.2]
For any consistent r.e. theory T that contains $\mathbf {Q}$ , there exists some $t \in \omega $ such that $\varphi _t(t)\!\uparrow $ holds but .
Kleene’s proof of his theorem uses recursion theory, and is not constructive. Salehi and Seraji [Reference Salehi and Seraji114] show that there is a constructive proof of Kleene’s theorem, but this constructive proof does not have the Rosser property. Salehi and Seraji [Reference Salehi and Seraji114] comment that there could be a ‘Rosserian’ version of this constructive proof of Kleene’s theorem.
3.3.3 Proofs based on Arithmetic Completeness
Hilbert and Bernays [Reference Hilbert and Bernays60] present the Arithmetic Completeness Theorem expressing that any recursively axiomatizable consistent theory has an arithmetically definable model. Later, Kreisel [Reference Kreisel80] and Wang [Reference Wang138] adapt the Arithmetic Completeness Theorem and use paradoxes to obtain undecidability results.
Now, the Arithmetic Completeness Theorem is an important tool in model-theoretic proofs of the incompleteness theorems. For more details, we refer to [Reference Kaye and Kotlarski67, Reference Kotlarski79, Reference Lindström91]. Walter Dean [Reference Dean28] gives a detailed discussion on how the Arithmetized Completeness Theorem provides a tool for obtaining formal incompleteness results from some certain paradoxes.
Theorem 3.10 Arithmetic Completeness [Reference Kikuchi68, Theorem 3.1]
Let T be a recursively axiomatized consistent extension of $\mathbf {Q}$ . There exists a formula $\mathbf {Tr}_T(x)$ in $L(\mathbf {PA})$ that defines a model of T in $\mathbf {PA}+\mathbf {Con}(T)$ .
Lemma 3.11 is a corollary of the Arithmetized Completeness Theorem, and is essential for model-theoretic proofs of the incompleteness theorems.
Lemma 3.11 [Reference Kikuchi68, Reference Kikuchi and Tanaka74]. Let T be a recursively axiomatized consistent extension of $\mathbf {Q}$ , and $\mathbf {Tr}_T(x)$ is the formula as asserted in Theorem 3.10. For any model $M_0$ of $\mathbf {PA}+\mathbf {Con}(T)$ , there exists a model $M_1$ of T such that for any sentence $\phi $ , $M_0\models \mathbf {Tr}_T(\ulcorner \phi \urcorner )$ if and only if $M_1\models \phi $ .
Kreisel first applies the Arithmetized Completeness Theorem to establish model-theoretic proofs of $\textsf {G2}$ (cf. Kreisel [Reference Kreisel82], Smoryński [Reference Smoryński and Barwise119], and Kikuchi [Reference Kikuchi68]). Kikuchi–Tanaka [Reference Kikuchi and Tanaka74], Kikuchi [Reference Kikuchi68, Reference Kikuchi69], and Kotlarski [Reference Kotlarski77] use the Arithmetized Completeness Theorem to give model-theoretic proofs of $\textsf {G2}$ . For example, Kikuchi [Reference Kikuchi68] proves $\textsf {G2}$ model theoretically via the Arithmetized Completeness Theorem (Lemma 3.11): if $\mathbf {PA}$ is consistent, then $\mathbf {Con}(\mathbf {PA})$ is not provable in $\mathbf {PA}$ (see [Reference Kikuchi68, Theorem 3.4]).Footnote 16
Proofs of $\textsf {G2}$ by Kreisel [Reference Kreisel82] and Kikuchi [Reference Kikuchi68] do not directly yield the formalized version of $\textsf {G2}$ . Kikuchi’s proof of $\textsf {G2}$ in [Reference Kikuchi69] is not formalizable in $\mathbf {PRA}$ . Kikuchi and Tanaka [Reference Kikuchi and Tanaka74] prove in $\mathbf {WKL_0}$ that $\mathbf {Con}(\mathbf {PA})$ implies $\neg \mathbf {Pr}_{\mathbf {PA}}(\ulcorner \mathbf {Con}(\mathbf {PA})\urcorner )$ , since the Completeness Theorem is provable in $\mathbf {WKL_0}$ , and the key Lemma 3.11 used in Kikuchi’s proof [Reference Kikuchi68] is provable in $\mathbf {RCA_0}$ .Footnote 17 Using Theorem 2.22, Tanaka [Reference Kikuchi and Tanaka74] proves the formalized version of $\textsf {G2}$ : $\mathbf {PRA}\vdash \mathbf {Con}(\mathbf {PA})\rightarrow \mathbf {Con}(\mathbf {PA}+\neg \mathbf {Con}(\mathbf {PA}))$ .
One can give a simple proof of $\textsf {G1}$ via the Diagnolisation Lemma (see [Reference Murawski97]). Kotlarski [Reference Kotlarski77] proves the formalized version of $\textsf {G1}$ and $\textsf {G2}$ via model-theoretic arguments (e.g., using the Arithmetized Completeness Theorem and some quickly growing functions). Kotlarski [Reference Kotlarski77] proves the following version of $\textsf {G1}$ assuming that $\mathbf {PA}$ is $\omega $ -consistent, and shows that the following sentence is provable in $\mathbf {PA}$ :
if $\forall \varphi , x \{[\varphi \in \Delta _0\wedge \forall y \mathbf {Pr}_{\mathbf {PA}}(\neg \varphi (S^x 0, S^y 0))]\rightarrow \neg \mathbf {Pr}_{\mathbf {PA}}(\exists y \varphi (S^x 0, y))\}$ , then $\exists \varphi \in \Delta _0\exists x [\neg \mathbf {Pr}_{\mathbf {PA}}(\exists y \varphi (S^x 0, y))\wedge \neg \mathbf {Pr}_{\mathbf {PA}}(\neg \exists y \varphi (S^x 0, y))]$ .
However, it is unknown whether the method in [Reference Kotlarski77] can also give a new proof of Rosser’s first incompleteness theorem. Kotlarski [Reference Kotlarski77] proves the following formalized version of $\textsf {G2}$ : $\mathbf {PA}\vdash \mathbf {Con(\mathbf {PA})} \rightarrow \mathbf {Con}(\mathbf {PA}+\neg \mathbf {Con}(\mathbf {PA}))$ . Later, Kotlarski [Reference Kotlarski78] transforms the proof of the formalized version of $\textsf {G2}$ in [Reference Kotlarski77] to a proof-theoretic version without the use of the Arithmetized Completeness Theorem.
3.3.4 Proofs based on Kolmogorov complexity
Intuitively, Kolmogorov complexity is a measure of the quantity of information in finite objects. Roughly speaking, the Kolmogorov complexity of a number n, denoted by $K(n)$ , is the size of a program which generates n.
Definition 3.12 Kolmogorov–Chaitin Complexity [Reference Salehi and Seraji114]
For any natural number $n \in \omega $ , the Kolmogorov complexity for n, denoted by $K(n)$ , is defined as $\min \{i \in \omega \mid \varphi _i(0)\!\downarrow = n\}$ .
If $n \leq K(n)$ , then n is called random. Kolmogorov shows in 1960s that the set of non-random numbers is recursively enumerable but not recursive (cf. Odifreddi [Reference Odifreddi100]). Relations between $\textsf {G1}$ and Kolmogorov complexity have been intensively discussed in the literature (cf. Li and Vitányi [Reference Li, Vitányi and van Leeuwen90]). Chaitin [Reference Chaitin19] gives an information-theoretic formulation of $\textsf {G1}$ , and proves the following weaker version of $\textsf {G1}$ in terms of Kolmogorov complexity.
Theorem 3.13 Chaitin [Reference Chaitin19, Reference Salehi and Seraji114]
For any consistent r.e. extension T of $\mathbf {Q}$ , there exists a constant $c_T \in \mathbb {N}$ such that for any $e \geq c_T$ and any $w \in \mathbb {N}$ we have .
Salehi and Seraji [Reference Salehi and Seraji114] show that we can algorithmically construct the Chaitin constant $c_T$ in Theorem 3.13. I.e., for a given consistent r.e. extension T of $\mathbf {Q}$ , one can algorithmically construct a constant $c_T\in \mathbb {N}$ such that for all $e \geq c_T$ and all $w \in \mathbb {N}$ , we have (see Theorem 3.4 in [Reference Salehi and Seraji114]). From Theorem 3.13, it is not clear whether “ $K(w)> e$ ” holds (or whether “ $K(w)> e$ ” is independent of T). Salehi and Seraji [Reference Salehi and Seraji114] show that Chaitin’s proof of $\textsf {G1}$ is non-constructive: there is no algorithm such that given any consistent r.e. extension T of $\mathbf {Q}$ we can compute some $w_T$ such that $K(w_T)> c_T$ holds where $c_T$ is the Chaitin constant we can compute as in Theorem 3.13 (see [Reference Salehi and Seraji114, Theorem 3.5]). If such an algorithm exists, then for any consistent r.e. extension T of $\mathbf {Q}$ , we can compute some $c_T$ and $w_T$ such that $K(w_T)> c_T$ is true but unprovable in T.
Salehi and Seraji [Reference Salehi and Seraji114] also strengthen Chaitin’s Theorem 3.13 assuming T is $\Sigma ^0_1$ -sound: if T is a $\Sigma ^0_1$ -sound r.e. theory extending $\mathbf {Q}$ , then there exists some $c_T$ (which is computable from T) such that for any $e \geq c_T$ there are cofinitely many w’s such that “ $K(w)> e$ ” is independent of T (see [Reference Salehi and Seraji114, Corollary 3.7]). Using a version of the Pigeonhole Principle in $\mathbf {Q}$ , Salehi and Seraji [Reference Salehi and Seraji114] also prove the Rosserian form of Chaitin’s Theorem: for any consistent r.e. extension T of $\mathbf {Q}$ , there is a constant $c_T$ (which is computable from T) such that for any $e \geq c_T$ there are cofinitely many w’s such that “ $K(w)> e$ ” is independent of T (see [Reference Salehi and Seraji114, Theorem 3.9]).
Kikuchi [Reference Kikuchi69] proves the following formalized version of $\textsf {G1}$ via Kolmogorov complexity for any consistent r.e. extension T of $\mathbf {PA}$ : there exists $e \in \omega $ with:
-
• $T \vdash \mathbf {Con}(T) \rightarrow \forall x(\neg \mathbf {Pr}_T(\ulcorner K (x)>e\urcorner ))$ ;
-
• $T \vdash \omega $ - $\mathbf {Con}(T) \rightarrow \forall x(e < K(x) \rightarrow \neg \mathbf {Pr}_T(\ulcorner K (x) \leq e\urcorner ))$ .
However, this proof is not constructive. Moreover, Kikuchi [Reference Kikuchi69] proves $\textsf {G2}$ via Kolmogorov complexity and the Arithmetic Completeness Theorem: if T is a consistent r.e. extension of $\mathbf {PA}$ , then $T\nvdash \mathbf {Con}(T)$ . Kikuchi’s proof of $\textsf {G2}$ in [Reference Kikuchi69] cannot be formalized in $\mathbf {PRA}$ but can be carried out within $\mathbf {WKL_0}$ . Thus we can also obtain a formalized version of $\textsf {G2}$ in $\mathbf {WKL_0}$ by Theorem 2.22.
3.3.5 Model-theoretic proofs
Adamowicz and Bigorajska [Reference Adamowicz and Bigorajska1] prove $\textsf {G2}$ via model-theoretic method using the notion of 1-closed models and existentially closed models.
-
• A model M of a theory T is called 1-closed (w.r.t. T) if for any $a_1,\ldots , a_n$ in M, any $\Sigma _1$ formula $\phi $ , and any $M^{\prime }$ such that $M \prec _0 M^{\prime }$ and $M^{\prime }\models T$ , we have: if $M^{\prime }\models \phi (a_1,\cdots , a_n)$ , then $M \models \phi (a_1,\ldots , a_n)$ . In other words, we can say that M is 1-closed if for any $M^{\prime }$ such that $M \prec _0 M^{\prime }$ , we have $M \prec _1 M^{\prime }$ .
-
• Let $\mathcal {K}$ be a class of structures in the same language. A model $M\in \mathcal {K}$ is existentially closed in $\mathcal {K}$ if for every model $N\supseteq M$ such that $N\in \mathcal {K}$ , we have $M\preceq _1 N$ : every existential formula with parameters from M which is satisfied in N is already satisfied in M.
Adamowicz and Bigorajska [Reference Adamowicz and Bigorajska1] first prove $\textsf {G2}$ without the use of the Arithmetized Completeness Theorem: every 1-closed model of any subtheory T of $\mathbf {PA}$ extending $I\Delta _0 + \mathbf {exp}$ satisfies $\neg \mathbf {Con}(\mathbf {PA})$ . Then Adamowicz and Bigorajska [Reference Adamowicz and Bigorajska1] prove the formalized version of $\textsf {G2}$ via the idea of existentially closed models and the Arithmetized Completeness Theorem: $\mathbf {PA} \vdash \mathbf {Con}(\mathbf {PA}) \rightarrow \mathbf {Con}(\mathbf {PA} + \neg \mathbf {Con}(\mathbf {PA}))$ (see [Reference Adamowicz and Bigorajska1, Theorem 2.1]). This is proved by showing that an arbitrary model of $\mathbf {PA} + \mathbf {Con}(\mathbf {PA})$ satisfies $\mathbf {Con}(\mathbf {PA} + \neg \mathbf {Con(PA)})$ .
3.4 Proofs of $\textsf {G1}$ and $\textsf {G2}$ based on logical paradox
We provide a survey of proofs of incompleteness theorems based on ‘logical paradox’.
3.4.1 Introduction
As noted in Section 3.2.1, Gödel’s incompleteness theorems are closely related to paradox and self-reference. In fact, Gödel comments in his famous paper that “any epistemological antinomy could be used for a similar proof of the existence of undecidable propositions”.
Now, the Liar Paradox is an old and most famous paradox in modern science. In Gödel’s proof of $\textsf {G1}$ , we can view Gödel’s sentence as the formalization of the Liar Paradox. Gödel’s sentence concerns the notion of provability but the liar sentence in the Liar Paradox concerns the notion of truth in the standard model of arithmetic. There is a big difference between the notion of provability and truth. Gödel’s sentence does not lead to a contradiction as the Liar sentence does.
Besides the Liar Paradox, many other paradoxes have been used to give new proofs of incompleteness theorems: for example, Berry’s Paradox in [Reference Boolos13, Reference Chaitin19, Reference Kikuchi68, Reference Kikuchi, Kurahashi and Sakai73, Reference Kikuchi and Tanaka74, Reference Vopěnka137], Grelling–Nelson’s Paradox in [Reference Cieśliński25], the Unexpected Examination Paradox in [Reference Fitch36, Reference Kritchman and Raz83], and Yablo’s Paradox in [Reference Cieśliński and Urbaniak26, Reference Kikuchi and Kurahashi70, Reference Kurahashi85, Reference Priest106]. We now discuss some of these paradoxes in detail.
3.4.2 Berry’s Paradox
Berry’s Paradox introduced by Russell [Reference Russell112] is the paradox that “the least integer not nameable in fewer than 19 syllables” is itself a name consisting of 18 syllables. Informally, we say that an expression names a natural number n if n is the unique natural number satisfying the expression. Berry’s Paradox can be formalized in formal systems by interpreting the concept of “name” suitably. The following is Boolos’s formulation of the concept of “name” in [Reference Boolos13].
Definition 3.15 Boolos [Reference Boolos13]
Let $n \in \omega $ and $\varphi (x)$ be a formula with only one free variable x. We say that $\varphi (x)$ names n if $\mathfrak {N}\models \varphi (\overline {n})\wedge \forall v_0\forall v_1(\varphi (v_0)\wedge \varphi (v_1)\rightarrow v_0= v_1)$ .
Proofs of the incompleteness theorems based on Berry’s Paradox have been given by Vopěnka [Reference Vopěnka137], Chaitin [Reference Chaitin19], Boolos [Reference Boolos13], Kikuchi–Kurahashi–Sakai [Reference Kikuchi, Kurahashi and Sakai73], Kikuchi [Reference Kikuchi68], and Kikuchi–Tanaka [Reference Kikuchi and Tanaka74]. In fact, Robinson first uses Berry’s Paradox in [Reference Robinson110] to prove Tarski’s theorem on the undefinability of truth, which anticipates the later use of Berry’s Paradox to obtain incompleteness results by Vopěnka [Reference Vopěnka137], Boolos [Reference Boolos13], and Kikuchi [Reference Kikuchi69].
Boolos [Reference Boolos13] proves a weak form of $\textsf {G1}$ in the 1980s by formalizing Berry’s Paradox in arithmetic via considering the length of formulas that name natural numbers in the standard model of arithmetic. Using this formulation of the concept of “name”, Boolos [Reference Boolos13] first shows that Berry’s Paradox leads to a proof of $\textsf {G1}$ in the following form: there is no algorithm whose output contains all true statements of arithmetic and no false ones (i.e., the theory of true arithmetic is not recursively axiomatizable). Barwise praises Boolos’s proof as “very lovely and the most straightforward proof of Gödel’s incompleteness theorem that I have ever seen”. The optimal sufficient and necessary condition for the independence of a Boolos sentence from $\mathbf {PA}$ is that $\mathbf {PA}+\mathbf {Con}(\mathbf {PA})$ is consistent (see [Reference Salehi and Seraji114]).
Boolos’s theorem is different from Gödel’s theorem in the following way:
-
• Boolos’s theorem refers to the concept of truth but Gödel’s theorem does not;
-
• Boolos’s proof is not constructive, and we can prove that there is no algorithm for computing the true but unprovable sentence;
-
• Boolos’s theorem is weaker than Gödel’s first incompleteness theorem, and hence we cannot obtain the second incompleteness theorem from Boolos’s theorem in the standard way (see [Reference Kikuchi, Kurahashi and Sakai73]).
Boolos’s proof is modified by Kikuchi and Tanaka in [Reference Kikuchi68, Reference Kikuchi and Tanaka74]. The difference between Kikuchi’s proof and Boolos’s proof lies in the interpretation of the word “name”. Kikuchi [Reference Kikuchi68] modifies Boolos’s formulation of the concept of “name” by replacing “truth” with “provability” in the definition.
Definition 3.16 Kikuchi [Reference Kikuchi68, Definition 3.1]
Let $n \in \omega $ and $\varphi (x)$ be a formula with only one free variable x. We say that $\varphi (x)$ names n if $\mathbf {PA}\vdash \varphi (\overline {n})\wedge \forall v_0\forall v_1(\varphi (v_0)\wedge \varphi (v_1)\rightarrow v_0= v_1)$ .
Using this formulation of the concept of “name”, Kikuchi [Reference Kikuchi68] gives a proof-theoretic proof of $\textsf {G1}$ by formalizing Berry’s Paradox without the use of the Diagnolisation Lemma. Kikuchi [Reference Kikuchi68] constructs a sentence $\theta $ and shows that if $\mathbf {PA}$ is consistent, then $\neg \theta $ is not provable in $\mathbf {PA}$ ; if $\mathbf {PA}$ is $\omega $ -consistent, then $\theta $ is not provable in $\mathbf {PA}$ (see [Reference Kikuchi68, Theorem 2.2]). Note that Kikuchi’s proof of $\textsf {G1}$ in [Reference Kikuchi68] is constructive. Kikuchi and Tanaka [Reference Kikuchi and Tanaka74] reformulate Kikuchi’s proof of $\textsf {G1}$ in [Reference Kikuchi68], and show in $\mathbf {WKL_0}$ that if $\mathbf {PA}+\mathbf {Con}(\mathbf {PA})$ is consistent, then $\theta $ is independent of $\mathbf {PA}$ . By Theorem 2.22, Kikuchi and Tanaka [Reference Kikuchi and Tanaka74] prove the formalized version of $\textsf {G1}$ : $\mathbf {PRA}\vdash \mathbf {Con}(\mathbf {PA}+\mathbf {Con}(\mathbf {PA}))\rightarrow \neg \mathbf {Pr}_{\mathbf {PA}}(\ulcorner \theta \urcorner )\wedge \neg \mathbf {Pr}_{\mathbf {PA}}(\ulcorner \neg \theta \urcorner )$ . An interesting question not covered in [Reference Kikuchi68, Reference Kikuchi and Tanaka74] is whether we can improve Kikuchi’s proof of $\textsf {G1}$ by only assuming that $\mathbf {PA}$ is consistent.
Vopěnka [Reference Vopěnka137] proves $\textsf {G2}$ for $\mathbf {ZF}$ by formalizing Berry’s Paradox, via adopting Kikuchi’s definition of the concept of “name” in [Reference Kikuchi68] over models of $\mathbf {ZF}$ Footnote 18 : $\mathbf {Con(ZF)}$ is not provable in $\mathbf {ZF}$ . Vopěnka’s proof uses the Completeness Theorem but does not use the Arithmetic Completeness Theorem. Kikuchi, Kurahashi, and Sakai [Reference Kikuchi, Kurahashi and Sakai73] show that Vopěnka’s method can be adapted to prove $\textsf {G2}$ for $\mathbf {PA}$ based on Kikuchi’s formalization of Berry’s Paradox in [Reference Kikuchi68] with an application of the Arithmetic Completeness Theorem.
Proofs of $\textsf {G1}$ and $\textsf {G2}$ based on Berry’s Paradox by Vopěnka [Reference Vopěnka137], Chaitin [Reference Chaitin19], Boolos [Reference Boolos13], and Kikuchi [Reference Kikuchi68] do not use the Diagnolisation Lemma. We can also prove $\textsf {G1}$ based on Berry’s Paradox using the Diagnolisation Lemma. For example, Kikuchi, Kurahashi, and Sakai [Reference Kikuchi, Kurahashi and Sakai73] adopt Kikuchi’s definition of the concept of “name” in [Reference Kikuchi68], and show that the independent statement in Kikuchi’s proof in [Reference Kikuchi68] can be obtained by using the Diagnolisation Lemma.
In summary, the distinctions between using and not using the Diagnolisation Lemma, and between using and not using the Arithmetic Completeness Theorem are not essential for proofs of $\textsf { G1}$ and $\textsf {G2}$ based on Berry’s Paradox. From the above discussions, we can characterize different proofs of $\textsf {G1}$ and $\textsf {G2}$ based on Berry’s Paradox by the method of interpreting the word “name”: Boolos [Reference Boolos13] uses the standard model of arithmetic; Kikuchi [Reference Kikuchi68] uses provability in arithmetic; Chaitin [Reference Chaitin19] and Kikuchi [Reference Kikuchi69] use Kolmogorov complexity; Kikuchi and Tanaka [Reference Kikuchi and Tanaka74] use nonstandard models of arithmetic; and Vopěnka [Reference Vopěnka137] uses models of $\mathbf {ZF}$ (see [Reference Kikuchi, Kurahashi and Sakai73]).
3.4.3 Unexpected Examination and Grelling–Nelson’s Paradox
First of all, Kritchman and Raz [Reference Kritchman and Raz83] give a new proof of $\textsf {G2}$ based on Chaitin’s incompleteness theorem and an argument that resembles the Unexpected Examination ParadoxFootnote 19 (for more details, we refer to [Reference Kritchman and Raz83]): for any consistent r.e. extension T of $\mathbf {PA}$ , if T is consistent, then $T\nvdash \mathbf {Con}(T)$ .
Secondly, we say a one-place predicate is “heterological” if it does not apply to itself (e.g., “long” is heterological, since it’s not a long expression). Consider the question: is the predicate “heterological” we have just defined heterological? If “heterological” is heterological, then it isn’t heterological; and if “heterological” isn’t heterological, then it is heterological. This contradiction is called Grelling–Nelson’s Paradox.
Cieśliński [Reference Cieśliński25] presents semantic proofs of $\textsf {G2}$ for $\mathbf {ZF}$ and $\mathbf {PA}$ based on Grelling–Nelson’s Paradox. For a theory T containing $\mathbf {ZF}$ , Cieśliński defines the sentence $\mathbf {HET_T}$ which says intuitively that the predicate “heterological” is itself heterological, and then shows that $T\nvdash \mathbf {HET_T}$ and $T\vdash \mathbf {HET_T}\leftrightarrow \mathbf {Con}(T)$ . Finally, Cieśliński shows how to adapt the proof of $\textsf {G2}$ for $\mathbf {ZF}$ to a proof of $\textsf {G2}$ for $\mathbf {PA}$ . In fact, Cieśliński [Reference Cieśliński25] proves the semantic version of $\textsf {G2}$ : if T has a model, then $T + \neg \mathbf {Con}(T)$ has a model (i.e., $T\nvdash \mathbf {Con}(T)$ ).
3.4.4 Yablo’s Paradox
We discuss proofs of $\textsf {G1}$ and $\textsf {G2}$ based on Yablo’s Paradox in the literature. Yablo’s Paradox is an infinite version of the Liar Paradox proposed in [Reference Yablo146]: consider an infinite sequence $Y_1, Y_2, \ldots $ of propositions such that each $Y_i$ asserts that $Y_j$ are false for all $j> i$ . Different proofs of $\textsf {G1}$ and $\textsf {G2}$ based on Yablo’s Paradox have been given by some authors (e.g., Priest [Reference Priest106], Cieśliński–Urbaniak [Reference Cieśliński and Urbaniak26], and Kikuchi–Kurahashi [Reference Kikuchi and Kurahashi70]).
Recall that we assume by default that T is a consistent r.e. extension of $\mathbf {Q}$ . Priest [Reference Priest106] first points out that $\textsf { G1}$ can be proved by formalizing Yablo’s Paradox. Priest defines a formula $Y(x)$ as follows which says that for any $y> x, Y (y)$ is not provable in T.
Definition 3.17 [Reference Cieśliński and Urbaniak26, Reference Kurahashi85]
A formula $Y(x)$ is called a Yablo formula of T if $T \vdash \forall x (Y(x)\leftrightarrow \forall y> x \neg \mathbf {Pr}_T(\ulcorner Y(\dot {y})\urcorner ))$ .
Cieśliński and Urbaniak originally prove the following version of $\textsf {G1}$ , and show that each instance $Y(\overline {n})$ of the Yablo formula is independent of T if T is $\Sigma ^0_1$ -sound (or 1-consistent).
Theorem 3.18 [Reference Cieśliński and Urbaniak26, Theorem 19]; see also [Reference Kurahashi85, Theorem 4]
Let $Y(x)$ be a Yablo formula.
-
• If T is consistent, then $T\nvdash Y(\overline {n})$ .
-
• If T is $\Sigma ^0_1$ -sound, then $T \nvdash \neg Y(\overline {n})$ .
Cieśliński and Urbaniak originally prove that $T \vdash \forall x(Y(x)\leftrightarrow \mathbf {Con}(T))$ (see [Reference Cieśliński and Urbaniak26, Theorems 21 and 22]). As a corollary, we have $T \vdash \forall x\forall y(Y(x)\leftrightarrow Y(y))$ , and $\textsf {G2}$ holds: if T is consistent, then $T\nvdash \mathbf {Con}(T)$ .
Definition 3.19 [Reference Cieśliński and Urbaniak26, Reference Kurahashi85]
A formula $Y^R(x)$ is called a Rosser-type Yablo formula of $\mathbf {Prf}_T(x, y)$ if $\mathbf {PA} \vdash \forall x (Y^R(x) \leftrightarrow \forall y> x\neg \mathbf {Pr}_T^R(x)(\ulcorner Y^R(\dot {y})\urcorner ))$ .
Theorem 3.20 shows that the Rosser-type Yablo formula is independent of any $\Sigma ^0_1$ -sound theory T.
Theorem 3.20 [Reference Kurahashi85, Theorem 10]
Let $\mathbf {Prf}_T(x, y)$ be any standard proof predicate of T, and $Y^R(x)$ be any Rosser-type Yablo formula of $\mathbf {Prf}_T(x, y)$ . Given $n \in \omega $ , if T is consistent, then $T \nvdash Y^R(\overline {n})$ ; if T is $\Sigma ^0_1$ -sound, then $T \nvdash \neg Y^R(\overline {n})$ .
The independence of $Y^R(\overline {n})$ for T which is not $\Sigma ^0_1$ -sound is discussed in [Reference Kurahashi85]. For a consistent but not $\Sigma ^0_1$ -sound theory, the situation of Rosser-type Yablo formulas is quite different from that of Rosser sentences. Kurahashi [Reference Kurahashi85] shows that for any consistent but not $\Sigma ^0_1$ -sound theory, the independence of each instance of a Rosser-type Yablo formula depends on the choice of standard proof predicates (see Theorems 12 and 25 in [Reference Kurahashi85]). Kurahashi [Reference Kurahashi85] shows that for any consistent but not $\Sigma ^0_1$ -sound theory T, there is a standard proof predicate of T such that each instance $Y^R(\overline {n})$ of the Rosser-type Yablo formula $Y^R(x)$ based on this proof predicate is provable in T for any $n \in \omega $ . Moreover, Kurahashi [Reference Kurahashi85] constructs a standard proof predicate of T and a Rosser-type Yablo formula $Y^R(x)$ based on this proof predicate such that each instance of $Y^R(x)$ is independent of T. Proofs of these results use the technique of Guaspari and Solovay in [Reference Guaspari and Solovay50].
Cieśliński and Urbaniak [Reference Cieśliński and Urbaniak26] conjecture that any two distinct instances $Y^R(\overline {m})$ and $Y^R(\overline {n})$ of a Rosser-type Yablo formula $Y^R(x)$ based on a standard proof predicate are not provably equivalent. Leach–Krouse [Reference Leach-Krouse84] and Kurahashi [Reference Kurahashi85] construct a standard proof predicate, and a Rosser-type Yablo formula $Y^R(x)$ based on this proof predicate such that $T\vdash \forall x\forall y (Y^R(x) \leftrightarrow Y^R(y))$ (see [Reference Leach-Krouse84, Theorem 9] and [Reference Kurahashi85, Corollary 21]).
Kurahashi [Reference Kurahashi85] constructs a partial counterexample to Cieśliński and Urbaniak’s conjecture: a standard proof predicate, and a Rosser-type Yablo formula $Y^R(x)$ based on this proof predicate such that $\forall x\forall y (Y^R(x) \leftrightarrow Y^R(y))$ is not provable in T ([Reference Kurahashi85, Corollary 20]). Thus the provability of the sentence $\forall x\forall y (Y^R(x) \leftrightarrow Y^R(y))$ also depends on the choice of standard proof predicates (see [Reference Kurahashi85, Corollaries 20 and 21]). Proofs of these results by Leach-Krouse and Kurahashi also use the technique of Guaspari and Solovay in [Reference Guaspari and Solovay50]. An interesting open question is: whether there is a standard proof predicate such that $Y^R(\overline {n})$ and $Y^R(\overline {n+1})$ are not provably equivalent for some $n \in \omega $ (see [Reference Kurahashi85])?
3.4.5 Beyond arithmetization
All the proofs of $\textsf {G1}$ we have discussed use arithmetization. Andrzej Grzegorczyk proposes the theory $\mathbf {TC}$ in as a possible alternative theory for studying incompleteness and undecidability, and shows that $\mathbf {TC}$ is essentially incomplete and mutually interpretable with $\mathbf {Q}$ without arithmetization.
Now, in $\mathbf {PA}$ we have numbers that can be added or multiplied; while in $\mathbf {TC}$ , one has strings (or texts) that can be concatenated. In Gödel’s proof, the only use of numbers is coding of syntactical objects. The motivations for accepting strings rather than numbers as the basic notion are as follows: on metamathematical level, the notion of computability can be defined without reference to numbers; for Grzegorczyk, dealing with texts is philosophically better justified since intellectual activities like reasoning, communicating, or even computing involve working with texts not with numbers. Thus, it is natural to define notions like undecidability directly in terms of texts instead of natural numbers. Grzegorczyk only proves the incompleteness of $\mathbf {TC}$ in. Later, Grzegorczyk and Zdanowski [Reference Grzegorczyk, Zdanowski, Ehrenfeucht, Marek and Srebrny48] prove that $\mathbf {TC}$ is essentially incomplete.
3.5 Concrete incompleteness
3.5.1 Introduction
All proofs of Gödel’s incompleteness theorems we have discussed above make use of meta-mathematical or logical methods, and the independent sentence constructed has a clear meta-mathematical or logical flavour which is devoid of real mathematical content. To be blunt, from a purely mathematical point of view, Gödel’s sentence is artificial and not mathematically interesting. Gödel’s sentence is constructed not by reflecting about arithmetical properties of natural numbers, but by reflecting about an axiomatic system in which those properties are formalized (see [Reference Isaacson, Barwise, Kaplan, Keisler, Suppes and Troelstra61]). A natural question is then: can we find true sentences not provable in $\mathbf {PA}$ with real mathematical content? The research program concrete incompleteness is the search for natural independent sentences with real mathematical content.
This program has received a lot of attention because despite Gödel’s incompleteness theorems, one can still cherish the hope that all natural and mathematically interesting sentences about natural numbers are provable or refutable in $\mathbf {PA}$ , and that elementary arithmetic is complete w.r.t. natural and mathematically interesting sentences. However, after Gödel, many natural independent sentences with real mathematical content have been found. These independent sentences have a clear mathematical flavor, and do not refer to the arithmetization of syntax and provability.
In this section, we provide an overview of the research on concrete incompleteness. The survey paper [Reference Bovykin15] provides a good overview on the state-of-the-art up to Autumn 2006. For more detailed discussions about concrete incompleteness, we refer to Cheng [Reference Cheng21], Bovykin [Reference Bovykin15], and Friedman [Reference Friedman40].
3.5.2 Paris–Harrington and beyond
Paris and Harrington [Reference Paris, Harrington and Barwise103] propose the first mathematically natural statement independent of $\mathbf {PA}$ : the Paris–Harrington Principle ( $\textsf {PH}$ ) which generalizes the finite Ramsey theorem. Gödel’s sentence is a pure logical construction (via the arithmetization of syntax and provability predicate) and has no relevance with classic mathematics (without any combinatorial or number-theoretic content). On the contrary, Paris–Harrington Principle is an independent arithmetic sentence from classic mathematics with combinatorial content as we will show. We refer to [Reference Isaacson, Barwise, Kaplan, Keisler, Suppes and Troelstra61] for more discussions about the distinction between mathematical arithmetic sentences and meta-mathematical arithmetic sentences.
Definition 3.21 Paris–Harrington Principle ( $\textsf {PH}$ ) [Reference Paris, Harrington and Barwise103]
-
• For set X and $n\in \omega $ , let $[X]^n$ be the set of all n-elements subset of X. We identify n with $\{0, \ldots , n-1\}$ .
-
• For all $m, n, c\in \omega $ , there is $N\in \omega $ such that for all $f : [N]^m\rightarrow c$ , we have: $(\exists H \subseteq N)(|H| \geq n \wedge H \text { is homogeneous for } f \wedge |H|> \min (H))$ .
Theorem 3.22 Paris–Harrington [Reference Paris, Harrington and Barwise103]
The principle $\textsf {PH}$ is true but not provable in $\mathbf {PA}$ .
Now, $\textsf {PH}$ has a clear combinatorial flavor, and is of the form $\forall x \exists y\psi (x, y)$ where $\psi $ is a $\Delta ^0_0$ formula. It can be shown that for any given natural number n, $\mathbf {PA} \vdash \exists y\psi (\overline {n}, y)$ , i.e., all particular instances of $\textsf {PH}$ are provable in $\mathbf {PA}$ .
Following $\textsf {PH}$ , many other mathematically natural statements independent of $\mathbf {PA}$ with combinatorial or number-theoretic content have been formulated: the Kanamori–McAloon principle, the Kirby–Paris sentence [Reference Paris and Kirby104], the Hercules–Hydra game [Reference Paris and Kirby104], the Worm principle [Reference Beklemishev7, Reference Hamano and Okada56], the flipping principle [Reference Kirby75], the arboreal statement [Reference Mills93], the kiralic and regal principles [Reference Clote and Mcaloon27], and the Pudlák’s principle [Reference Hájek and Paris52, Reference Pudlák107] (see [Reference Bovykin15, p. 40]). In fact, all these principles are equivalent to $\textsf {PH}$ (see [Reference Bovykin15, p.40]).
An interesting and amazing fact is: all the above mathematically natural principles are in fact provably equivalent in $\mathbf {PA}$ to a certain meta-mathematical sentence. Consider the following reflection principle for $\Sigma ^0_1$ sentences: for any $\Sigma ^0_1$ sentence $\phi $ in $L(\mathbf {PA})$ , if $\phi $ is provable in $\mathbf {PA}$ , then $\phi $ is true. Using the arithmetization of syntax, one can write this principle as a sentence of $L({\mathbf {PA}})$ , and denote it by $\textsf {Rfn}_{\Sigma ^0_1}(\mathbf {PA})$ (see [Reference Murawski97, p. 301]). McAloon has shown that $\mathbf {PA} \vdash \textsf {PH}\leftrightarrow \textsf {Rfn}_{\Sigma ^0_1}(\mathbf {PA})$ (see [Reference Murawski97, p. 301]), and similar equivalences can be established for the other independent principles mentioned above. Equivalently, all these principles are equivalent to so-called 1-consistency of $\mathbf {PA}$ (see [Reference Beklemishev8, p. 36], [Reference Beklemishev7, p. 3], and [Reference Murawski97, p. 301]).
The above phenomenon indicates that the difference between mathematical and meta-mathematical statements is perhaps not as huge as we might have expected. Moreover, the above principles are provable in fragments of Second-Order Arithmetic and are more complex than Gödel’s sentence: Gödel’s sentence is equivalent to $\mathbf {Con}(\mathbf {PA})$ in $\mathbf {PA}$ ; but all these principles are not only independent of $\mathbf {PA}$ but also independent of $\mathbf {PA}+ \mathbf {Con}(\mathbf {PA})$ (see [Reference Beklemishev8, p. 36] and [Reference Murawski97, p. 301]).
3.5.3 Harvey Friedman’s contributions
Incompleteness would not be complete without mentioning the work of Harvey Friedman who is a central figure in research on the foundations of mathematics after Gödel. He has made many important contributions to concrete mathematical incompleteness. The following quote is telltale:
the long range impact and significance of ongoing investigations in the foundations of mathematics is going to depend greatly on the extent to which the Incompleteness Phenomena touches normal concrete mathematics (see [Reference Friedman40, p. 7]).
In the following, we give a brief introduction to H. Friedman’s work on concrete mathematical incompleteness. In his early work, H. Friedman examines how one uses large cardinals in an essential and natural way in number theory, as follows.
the quest for a simple meaningful finite mathematical theorem that can only be proved by going beyond the usual axioms for mathematics has been a goal in the foundations of mathematics since Gödel’s incompleteness theorems (see [Reference Friedman39, p. 805]).
H. Friedman shows in [Reference Friedman38, Reference Friedman39] that there are many mathematically natural combinatorial statements in $L(\mathbf {PA})$ that are neither provable nor refutable in $\mathbf {ZFC}$ or $\mathbf {ZFC}$ + large cardinals. H. Friedman’s more recent monograph [Reference Friedman40] is a comprehensive study of concrete mathematical incompleteness. H. Friedman studies concrete mathematical incompleteness over different systems, ranging from weak subsystems of $\mathbf {PA}$ to higher-order arithmetic and $\mathbf {ZFC}$ . H. Friedman lists many concrete mathematical statements in $L(\mathbf {PA})$ that are independent of subsystems of $\mathbf {PA}$ , or stronger theories like higher-order arithmetic and set theory.
The theories $\mathbf {RCA_0}$ (Recursive Comprehension), $\mathbf {WKL_0}$ (Weak Konig’s Lemma), $\mathbf {ACA_0}$ (Arithmetical Comprehension), $\mathbf {ATR_0}$ (Arithmetic Transfinite Recursion), and $\Pi ^1_1$ - $\mathbf {CA_0}$ ( $\Pi ^1_1$ -Comprehension) are the most famous five subsystems of Second-Order Arithmetic ( $\mathbf {SOA}$ ), and are called the ‘Big Five’. For the definition of $\mathbf {SOA}$ and the ‘Big Five’, we refer to [Reference Simpson117].
To give the reader a better sense of H. Friedman’s work, we list some sections dealing with concrete mathematical incompleteness in [Reference Friedman40].
-
• Section 0.5 on Incompleteness in Exponential Function Arithmetic.
-
• Section 0.6 on Incompleteness in Primitive Recursive Arithmetic, Single Quantifier Arithmetic, $\mathbf {RCA_0}$ , and $\mathbf {WKL_0}$ .
-
• Section 0.7 on Incompleteness in Nested Multiply Recursive Arithmetic and Two Quantifier Arithmetic.
-
• Section 0.8 on Incompleteness in Peano Arithmetic and $\mathbf {ACA_0}$ .
-
• Section 0.9 on Incompleteness in Predicative Analysis and $\mathbf {ATR_0}$ .
-
• Section 0.10 on Incompleteness in Iterated Inductive Definitions and $\Pi ^1_1$ - $\mathbf {CA_0}$ .
-
• Section 0.11 on Incompleteness in Second-Order Arithmetic and $\mathbf {ZFC}^{-}$ .Footnote 20
-
• Section 0.12 on Incompleteness in Russell Type Theory and Zermelo Set Theory.
-
• Section 0.13 on Incompleteness in $\mathbf {ZFC}$ using Borel Functions.
-
• Section 0.14 on Incompleteness in $\mathbf {ZFC}$ using Discrete Structures.
H. Friedman [Reference Friedman40] provides us with examples of concrete mathematical theorems not provable in subsystems of Second-Order Arithmetic stronger than $\mathbf {PA}$ , and a number of concrete mathematical statements provable in Third-Order Arithmetic but not provable in Second-Order Arithmetic.
Related to Friedman’s work, Cheng [Reference Cheng21, Reference Cheng and Schindler24] gives an example of concrete mathematical theorems based on Harrington’s principle which is isolated from the proof of the Harrington’s Theorem (the determinacy of $\Sigma ^1_1$ games implies the existence of zero sharp), and shows that this concrete theorem saying that Harrington’s principle implies the existence of zero sharp is expressible in Second-Order Arithmetic, not provable in Second-Order Arithmetic or Third-Order Arithmetic, but provable in Fourth-Order Arithmetic (i.e., the minimal system in higher-order arithmetic to prove this concrete theorem is Fourth-Order Arithmetic).
Many other examples of concrete mathematical incompleteness, and the discussion of this subject in 1970s–1980s can be found in the four volumes [Reference Berline, Mcaloon and Ressayre10, Reference Pacholski and Wierzejewski101, Reference Simpson115, Reference Simpson116]. Weiermann’s work in [Reference Weiermann139–Reference Weiermann143] provides us with more examples of naturally mathematical independent sentences. We refer to [Reference Friedman40] for new advances in Boolean Relation Theory and for more examples of concrete mathematical incompleteness.
4 The limit of the applicability of $\textsf {G1}$
4.1 Introduction
In this section, we discuss the limit of the applicability of $\textsf {G1}$ based on the following two questions.
-
• To what extent does $\textsf {G1}$ apply to extensions of $\mathbf {PA}$ ?
-
• To what extent does $\textsf {G1}$ apply to theories weaker than $\mathbf {PA}$ w.r.t. interpretation?
Definition 4.1 Conservativity
-
• Let $\Gamma $ denote either $\Sigma ^0_n$ or $\Pi ^0_n$ for some $n \geq 1$ , and $\Gamma ^d$ denote either $\Pi ^0_n$ or $\Sigma ^0_n$ .
-
• We say a sentence $\varphi $ is $\Gamma $ -conservative over theory T if for any $\Gamma $ sentence $\psi $ , $T \vdash \psi $ whenever $T + \varphi \vdash \psi $ .
We list some generalizations of $\textsf {G1}$ needed below.
Fact 4.2 Guaspari [Reference Guaspari49]
Let T be a consistent r.e. extension of $\mathbf {Q}$ . Then there is a $\Gamma ^d$ sentence $\phi $ such that $\phi $ is $\Gamma $ -conservative over T and $T\nvdash \phi $ .
If $T\vdash \neg \phi $ , then $\phi $ is not $\Gamma $ -conservative over T because T is consistent. Thus, we can view Fact 4.2 as an extension of Rosser’s first incompleteness theorem. Solovay improves this fact and shows that there is a $\Gamma ^d$ sentence $\phi $ such that $\phi $ is $\Gamma $ -conservative over T, $\neg \phi $ is $\Gamma ^d$ -conservative over T, but $\phi $ is independent of T.
Fact 4.3 Let $\{T_n: n\in \omega \}$ be an r.e. sequence of consistent theories extending $\mathbf {Q}$ . Then there is a $\Pi ^0_1$ sentence $\phi $ such that for any $n\in \omega $ , $T_n\nvdash \phi $ , and $T_n\nvdash \neg \phi $ .
4.2 Generalizations of $\textsf {G1}$ beyond $\mathbf {PA}$
We study generalization of $\textsf {G1}$ for extensions of $\mathbf {PA}$ w.r.t. interpretation. We know that $\textsf {G1}$ applies to all consistent r.e. extensions of $\mathbf {PA}$ . A natural question is then: whether $\textsf {G1}$ can be extended to non-r.e. arithmetically definable extensions of $\mathbf {PA}$ ?
Kikuchi–Kurahashi [Reference Kikuchi and Kurahashi71] and Salehi–Seraji [Reference Salehi and Seraji113] make contributions to generalize Gödel–Rosser’s first incompleteness theorem to non-r.e. arithmetically definable extensions of $\mathbf {PA}$ .
Definition 4.4 [Reference Kikuchi and Kurahashi71]
Let T be a consistent extension of $\mathbf {Q}$ .
-
• T is $\Sigma ^0_n$ -definable if there is a $\Sigma ^0_n$ formula $\phi (x)$ such that n is the Gödel number of some sentence of T if and only if $\mathfrak {N}\models \phi (\overline {n})$ .Footnote 21
-
• T is $\Sigma ^0_n$ -sound if for all $\Sigma ^0_n$ sentences $\phi $ , $T\vdash \phi $ implies $\mathfrak {N}\models \phi $ ; T is sound if T is $\Sigma ^0_{n}$ -sound for any $n \in \omega $ .
-
• T is $\Sigma ^0_n$ -consistent if for all $\Sigma ^0_{n}$ formulas $\phi $ with $\phi =\exists x\theta (x)$ and $\theta \in \Pi ^0_{n-1}$ , if $T \vdash \neg \theta (\overline {n})$ for all $n\in \omega $ , then $T \nvdash \phi $ .
-
• T is $\Pi ^0_n$ -decisive if for all $\Pi ^0_n$ sentences $\phi $ , either $T\vdash \phi $ or $T \vdash \neg \phi $ holds.
From $\textsf {G1}$ , we have: if T is a $\Sigma ^0_{1}$ -definable and $\Sigma ^0_{1}$ -sound extension of $\mathbf {Q}$ , then T is not $\Pi ^0_{1}$ -decisive. Kikuchi and Kurahashi [Reference Kikuchi and Kurahashi71] generalize $\textsf {G1}$ to arithmetically definable theories via the notion of “ $\Sigma ^0_{n}$ -sound”.
Theorem 4.5 [Reference Kikuchi and Kurahashi71, Theorem 4.8] and [Reference Salehi and Seraji113, Theorem 2.5]
If T is a $\Sigma ^0_{n+1}$ -definable and $\Sigma ^0_{n}$ -sound extension of $\mathbf {Q}$ , then T is not $\Pi ^0_{n+1}$ -decisive.
Salehi and Seraji [Reference Salehi and Seraji113] point out that Theorem 4.5 has a constructive proof: given a $\Sigma ^0_{n+1}$ -definable and $\Sigma ^0_{n}$ -sound extension T of $\mathbf {Q}$ , one can effectively construct a $\Pi ^0_{n+1}$ sentence which is independent of T. The optimality of Theorem 4.5 is shown by Salehi and Seraji in [Reference Salehi and Seraji113]: there exists a $\Sigma ^0_{n-1}$ -sound and $\Sigma ^0_{n+1}$ -definable complete extension of $\mathbf {Q}$ for any $n\geq 1$ ([Reference Salehi and Seraji113, Theorem 2.6]).
Salehi and Seraji [Reference Salehi and Seraji113] generalize $\textsf {G1}$ to arithmetically definable theories via the notion of “ $\Sigma ^0_n$ -consistent”.
Theorem 4.6 [Reference Kikuchi and Kurahashi71, Theorem 4.9] and [Reference Salehi and Seraji113, Theorem 4.3]
If T is a $\Sigma ^0_{n+1}$ -definable and $\Sigma ^0_n$ -consistent extension of $\mathbf {Q}$ , then T is not $\Pi ^0_{n+1}$ -decisive.
Theorem 4.6 is also optimal: the complete $\Sigma ^0_{n-1}$ -sound and $\Sigma ^0_{n+1}$ -definable theory constructed in the proof of Theorem 2.6 in [Reference Salehi and Seraji113] is also $\Sigma ^0_{n-1}$ -consistent since if a theory is $\Sigma ^0_{n}$ -sound, then it is $\Sigma ^0_{n}$ -consistent. The proof of Theorem 4.6 cannot be constructive as the following theorem shows.
Theorem 4.7 Non-constructivity of $\Sigma ^0_{n}$ -consistency incompleteness [Reference Salehi and Seraji113, Theorem 4.4]
For $n\geq 3$ , there is no (partial) recursive function f (even with the oracle $0^{n}$ ) such that if m codes (the Gödel code) a $\Sigma ^0_{n+1}$ -formula which defines an $\Sigma ^0_{n}$ -consistent extension T of $\mathbf {Q}$ , then $f(m)$ halts and codes a $\Pi ^0_{n+1}$ sentence which is independent of T.Footnote 22
In summary, $\textsf {G1}$ can be generalized to the incompleteness of $\Sigma ^0_{n+1}$ -definable and $\Sigma ^0_{n}$ -sound extensions of $\mathbf {Q}$ constructively; and to the incompleteness of $\Sigma ^0_{n+1}$ -definable and $\Sigma ^0_{n}$ -consistent extensions of $\mathbf {Q}$ non-constructively (when $n>2$ ).
4.3 Generalizations of $\textsf {G1}$ below $\mathbf {PA}$
We study generalizations of $\textsf {G1}$ for theories weaker than $\mathbf {PA}$ w.r.t. interpretation.
4.3.1 Generalizations of $\textsf {G1}$ via interpretability
We show that $\textsf {G1}$ can be generalized to theories weaker than $\mathbf {PA}$ via interpretability. Indeed, there exists a weak recursively axiomatizable consistent subtheory T of $\mathbf {PA}$ such that each recursively axiomatizable theory S in which T is interpretable is incomplete (see [Reference Tarski, Mostowski and Robinson125]). To generalize this fact further, we propose a new notion “ $\textsf {G1}$ holds for T”, as follows.
Definition 4.8 Let T be a consistent r.e. theory. We say $\textsf {G1}$ holds for T if for any recursively axiomatizable consistent theory S, if T is interpretable in S, then S is incomplete.
First of all, for a consistent r.e. theory T, it is not hard to show that the followings are equivalent (see [Reference Cheng22]):
-
• $\textsf {G1}$ holds for T.
-
• T is essentially incomplete.
-
• T is essentially undecidable.
It is well-known that $\textsf {G1}$ holds for many weaker theories than $\mathbf {PA}$ w.r.t. interpretation (e.g., Robinson Arithmetic $\mathbf {Q}$ ).
Secondly, we mention theories weaker than $\mathbf {PA}$ w.r.t. interpretation for which $\textsf {G1}$ holds. We first review some essentially undecidable theories weaker than $\mathbf {PA}$ w.r.t. interpretation from the literature (i.e., $\textsf {G1}$ holds for these theories). For the definition of theory $\mathbf {Q}$ , $I\Sigma _n$ , $B\Sigma _n$ , $\mathbf {PA}^{-}$ , $\mathbf {Q}^+$ , $\mathbf {Q}^{-}$ , $\mathbf {S^1_2}$ , $\mathbf {AS}$ , $\mathbf {EA}$ , $\mathbf {PRA}$ , $\mathbf {R}$ , $\mathbf {R}_0$ , $\mathbf {R}_1$ , and $\mathbf {R}_2$ , we refer to Section 2.
Robinson shows that any consistent r.e. theory that interprets $\mathbf {Q}$ is undecidable, and hence $\mathbf {Q}$ is essentially undecidable. The fact that $\mathbf {Q}$ is essentially undecidable is very useful and can be used to prove the essentially undecidability of other theories via Theorem 2.6. Since $\mathbf {Q}$ is finitely axiomatized, it follows that any theory that weakly interprets $\mathbf {Q}$ is also undecidable.
The Lindenbaum algebras of all r.e. theories that interpret $\mathbf {Q}$ are recursively isomorphic (see Pour-El and Kripke [Reference Pour-El and Kripke105]). In fact, $\mathbf {Q}$ is minimal essentially undecidable in the sense that if deleting any axiom of $\mathbf {Q}$ , then the remaining theory is not essentially undecidable and has a complete decidable extension (see [Reference Tarski, Mostowski and Robinson125, Theorem 11, p. 62]).
Thirdly, Nelson [Reference Nelson98] embarks on a program of investigating how much mathematics can be interpreted in Robinson’s Arithmetic $\mathbf {Q}$ : what can be interpreted in $\mathbf {Q}$ , and what cannot be interpreted in $\mathbf {Q}$ ? In fact, $\mathbf {Q}$ represents a rich degree of interpretability since a lot of stronger theories are interpretable in it as we will show in the following passages. For example, using Solovay’s method of shortening cuts (see [Reference Guaspari and Solovay50]), one can show that $\mathbf {Q}$ interprets fairly strong theories like $I\Delta _0 + \Omega _1$ on a definable cut.
Fourth, we discuss some prominent fragments of $\mathbf {PA}$ extending $\mathbf {Q}$ from the literature. As a corollary of Theorem 2.14, we have:
-
• The theories $\mathbf {Q}, I\Sigma _0, I\Sigma _0+\Omega _{1}, \ldots , I\Sigma _0+\Omega _{n}, \ldots , B\Sigma _1, B\Sigma _1+\Omega _{1}, \ldots $ , $B\Sigma _1+\Omega _{n}, \ldots $ are all mutually interpretable;
-
• $I\Sigma _0+\mathbf {exp}$ and $B\Sigma _1+\mathbf {exp}$ are mutually interpretable;
-
• For $n\geq 1$ , $I\Sigma _n$ and $B\Sigma _{n+1}$ are mutually interpretable;
-
• $\mathbf {Q}\lhd I\Sigma _0+\mathbf {exp}\lhd I\Sigma _1\lhd I\Sigma _2\lhd \cdots \lhd I\Sigma _n\lhd \cdots \lhd \mathbf {PA}$ .
Since any consistent r.e. theory which interprets $\mathbf {Q}$ is essentially undecidable, $\textsf {G1}$ holds for all these fragments of $\mathbf {PA}$ extending $\mathbf {Q}$ .
Fifth, we discuss some weak theories mutually interpretable with $\mathbf {Q}$ from the literature. It is interesting to compare $\mathbf {Q}$ with its bigger brother $\mathbf {PA}^{-}$ . From [Reference Visser132], $\mathbf {PA}^{-}$ is interpretable in $\mathbf {Q}$ , and hence $\mathbf {Q}$ is mutually interpretable with $\mathbf {PA}^{-}$ . The theory $\mathbf {Q}^+$ is interpretable in $\mathbf {Q}$ (see [Reference Ferreira and Ferreira35, Theorem 1, p. 296]), and thus mutually interpretable with $\mathbf {Q}$ . A. Grzegorczyk asks whether $\mathbf {Q}^{-}$ is essentially undecidable. Švejdar [Reference Švejdar123] provides a positive answer to Grzegorczyk’s original question by showing that $\mathbf {Q}$ is interpretable in $\mathbf {Q}^{-}$ using the Solovay’s method of shortening cuts. Thus $\mathbf {Q}^{-}$ is essentially undecidable and mutually interpretable with $\mathbf {Q}$ .
Sixth, by [Reference Ferreira and Ferreira35], $I\Sigma _0$ is interpretable in $\mathbf {S^1_2}$ , and $\mathbf {S^1_2}$ is interpretable in $\mathbf {Q}$ . Hence $\mathbf {S^1_2}$ is essentially undecidable and mutually interpretable with $\mathbf {Q}$ . The theory $\mathbf {AS}$ interprets Robinson’s Arithmetic $\mathbf {Q}$ , and hence is essentially undecidable. Nelson [Reference Nelson98] shows that $\mathbf {AS}$ is interpretable in $\mathbf {Q}$ . Thus, $\mathbf {AS}$ is mutually interpretable with $\mathbf {Q}$ .
Seventh, Grzegorczyk, and Zdanowski [Reference Grzegorczyk, Zdanowski, Ehrenfeucht, Marek and Srebrny48] formulate but leave unanswered an interesting problem: are $\mathbf {TC}$ and $\mathbf {Q}$ mutually interpretable? M. Ganea [Reference Ganea42] proves that $\mathbf {Q}$ is interpretable in $\mathbf {TC}$ using the detour via $\mathbf {Q}^{-}$ (i.e., first show that $\mathbf {Q}^{-}$ is interpretable in $\mathbf {TC}$ ; since $\mathbf {Q}$ is interpretable in $\mathbf {Q}^{-}$ , then we have $\mathbf {Q}$ is interpretable in $\mathbf {TC}$ ). Sterken and Visser [Reference Visser129] give a proof of the interpretability of $\mathbf {Q}$ in $\mathbf {TC}$ not using $\mathbf {Q}^{-}$ . Note that $\mathbf {TC}$ is easily interpretable in the bounded arithmetic $I\Sigma _0$ . Thus, $\mathbf {TC}$ is mutually interpretable with $\mathbf {Q}$ .
Note that $\mathbf {R}\lhd \mathbf {Q}$ since $\mathbf {Q}$ is not interpretable in $\mathbf {R}$ (if $\mathbf {Q}$ is interpretable in $\mathbf {R}$ , then $\mathbf {Q}$ is interpretable in some finite fragment of $\mathbf {R}$ ; however $\mathbf {R}$ is locally finitely satisfiable and any model of $\mathbf {Q}$ is infinite). Visser [Reference Visser131] provides us with a unique characterization of $\mathbf {R}$ .
Theorem 4.9 Visser [Reference Visser131, Theorem 6]
For any consistent r.e. theory T, T is interpretable in $\mathbf {R}$ if and only if T is locally finitely satisfiable.Footnote 23
Since relational $\Sigma _2$ sentences have the finite model property, by Theorem 4.9, any consistent theory axiomatized by a recursive set of $\Sigma _2$ sentences in a finite relational language is interpretable in $\mathbf {R}$ . Since all recursive functions are representable in $\mathbf {R}$ (see [Reference Tarski, Mostowski and Robinson125, Theorem 6, p. 56]), as a corollary of Theorem 2.6, $\mathbf {R}$ is essentially undecidable. Cobham shows that $\mathbf {R}$ has a stronger property than essential undecidability. Vaught gives a proof of Cobham’s Theorem 4.10 via existential interpretation in [Reference Vaught, Nagel, Suppes and Tarski127].
Theorem 4.10 Cobham [Reference Vaught, Nagel, Suppes and Tarski127]
Any consistent r.e. theory that weakly interprets $\mathbf {R}$ is undecidable.
Eighth, we discuss some variants of $\mathbf {R}$ in the same language as $L(\mathbf {R})=\{\overline {0}, \ldots , \overline {n}, \ldots , +, \times , \leq \}$ . The theory $\mathbf {R}_0$ is no longer essentially undecidable in the same language as $\mathbf {R}$ .Footnote 24 In fact, whether $\mathbf {R}_0$ is essentially undecidable depends on the language of $\mathbf {R}_0$ : if $L(\mathbf {R}_0)=\{\mathbf {0}, \mathbf {S}, + , \times , \leq \}$ with $\leq $ defined in terms of $+$ , then $\mathbf {R}_0$ is essentially undecidable (Cobham first observed that $\mathbf {R}$ is interpretable in $\mathbf {R}_0$ in the same language $\{\mathbf {0}, \mathbf {S}, + , \times \}$ , and hence $\mathbf {R}_0$ is essentially undecidable (see [Reference Jones and Shepherdson65, Reference Vaught, Nagel, Suppes and Tarski127]). The theory $\mathbf {R}_1$ is essentially undecidable since $\mathbf {R}$ is interpretable in $\mathbf {R}_1$ (see [Reference Jones and Shepherdson65, p. 62]).
However $\mathbf {R}_1$ is not minimal essentially undecidable. From [Reference Jones and Shepherdson65], $\mathbf {R}$ is interpretable in $\mathbf {R}_2$ , and hence $\mathbf {R}_2$ is essentially undecidable.Footnote 25 The theory $\mathbf {R}_2$ is minimal essentially undecidable in the sense that if we delete any axiom scheme of $\mathbf {R}_2$ , then the remaining system is not essentially undecidable.Footnote 26 By essentially the same argument as in [Reference Visser132], we can show that any consistent r.e. theory that weakly interprets $\mathbf {R}_2$ is undecidable.
Kojiro Higuchi and Yoshihiro Horihata introduce the theory of concatenation $\mathbf {WTC}^{-\epsilon }$ , which is a weak subtheory of Grzegorczyk’s theory $\mathbf {TC}$ , and show that $\mathbf {WTC}^{-\epsilon }$ is minimal essentially undecidable and $\mathbf {WTC}^{-\epsilon }$ is mutually interpretable with $\mathbf {R}$ (see [Reference Higuchi and Horihata58]).
In summary, we have the following pictures:
-
• Theories $\mathbf {PA}^{-}, \mathbf {Q}^{+}, \mathbf {Q}^{-}, \mathbf {TC}, \mathbf {AS},\mathbf {S^1_2}$ , and $\mathbf {Q}$ are all mutually interpretable, and hence $\textsf {G1}$ holds for them;
-
• Theories $\mathbf {R}, \mathbf {R}_1, \mathbf {R}_2$ , and $\mathbf {WTC}^{-\epsilon }$ are mutually interpretable, and hence $\textsf {G1}$ holds for them;
-
• $\mathbf {R}\lhd \mathbf {Q}\lhd \mathbf {EA}\lhd \mathbf {PRA}\lhd \mathbf {PA}$ .
4.3.2 The limit of $\textsf {G1}$ w.r.t. interpretation and Turing reducibility
We first discuss the limit of $\textsf {G1}$ for theories weaker than $\mathbf {PA}$ w.r.t. interpretation, i.e., finding a theory with minimal degree of interpretation for which $\textsf {G1}$ holds.
First of all, a natural question is: is $\mathbf {Q}$ the weakest finitely axiomatized essentially undecidable theory w.r.t. interpretation such that $\mathbf {R}\lhd \mathbf {Q}$ ? The following theorem tells us that the answer is no: for any finitely axiomatized subtheory A of $\mathbf {Q}$ that extends $\mathbf {R}$ , we can find a finitely axiomatized subtheory B of A such that B extends $\mathbf {R}$ and B does not interpret A.
Theorem 4.11 Visser [Reference Visser132, Theorem 2]
Suppose A is a finitely axiomatized consistent theory and $\mathbf {R} \subseteq A$ . Then there is a finitely axiomatized theory B such that $\mathbf {R} \subseteq B \subseteq A$ and $B\lhd A$ .
Define $X=\{S: \mathbf {R}\unlhd S\lhd \mathbf {Q}$ and S is finitely axiomatized}. Theorem 4.11 shows that the structure $\langle X, \lhd \rangle $ is not well-founded.
Theorem 4.12 Visser [Reference Visser132, Theorem 12]
Suppose A and B are finitely axiomatized theories that weakly interpret $\mathbf {Q}$ . Then there are finitely axiomatized theories $\overline {A}\supseteq A$ and $\overline {B}\supseteq B$ such that $\overline {A}$ and $\overline {B}$ are incomparable (i.e., $\overline {A}\ntrianglelefteq \overline {B}$ and $\overline {B}\ntrianglelefteq \overline {A}$ ).
Theorem 4.12 shows that there are incomparable theories extending $\mathbf {Q}$ w.r.t. interpretation.
Up to now, we do not have an example of essentially undecidable theory that is weaker than $\mathbf {R}$ w.r.t. interpretation. To this end, we introduce Jeřábek’s theory $\mathbf {Rep}_{\textsf {PRF}}$ .
Definition 4.13 The system $\mathbf {Rep}_{\textsf {PRF}}$
-
• Let $\textsf {PRF}$ denote the sets of all partial recursive functions.
-
• The language $L(\mathbf {Rep}_{\textsf {PRF}})$ consists of constant symbols $\overline {n}$ for each $n \in \omega $ , and function symbols $\overline {f}$ of appropriate arity for each partial recursive function f.
-
• The theory $\mathbf {Rep_{\textsf {PRF}}}$ has axioms:
-
– $\overline {n} \neq \overline {m}$ for $n \neq m \in \omega $ ;
-
– $\overline {f}(\overline {n_0}, \ldots , \overline {n_{k-1}}) = \overline {m}$ for each k-ary partial recursive function f such that $f(n_0, \ldots , n_{k-1}) = m$ where $n_0, \ldots , n_{k-1}, m\in \omega $ .
-
The theory $\mathbf {Rep}_{\textsf {PRF}}$ is essentially undecidable since all recursive functions are representable in it. Since $\mathbf {Rep}_{\textsf {PRF}}$ is locally finitely satisfiable, by Theorem 4.9, $\mathbf {Rep}_{\textsf {PRF}}\unlhd \mathbf {R}$ . Jeřábek proves that $\mathbf {R}$ is not interpretable in $\mathbf {Rep}_{\textsf {PRF}}$ . Thus $\mathbf {Rep}_{\textsf {PRF}}\lhd \mathbf {R}$ .
Cheng [Reference Cheng22] provides more examples of a theory S such that $\textsf {G1}$ holds for S and $S\lhd \mathbf {R}$ , and shows that we can find many theories T such that $\textsf {G1}$ holds for T and $T \lhd \mathbf {R}$ based on Jeřábek’s work which uses model theory.
Theorem 4.14 Cheng [Reference Cheng22]
For any recursively inseparable pair $\langle A,B\rangle $ , there is an r.e. theory $U_{\langle A,B\rangle }$ such that $\textsf {G1}$ holds for $U_{\langle A,B\rangle }$ , and $U_{\langle A,B\rangle }\lhd \mathbf {R}$ .
Define $\textsf {D}=\{S: S\lhd \mathbf {R}$ and $\textsf {G1}$ holds for theory S}. Theorem 4.14 shows that we could find many witnesses for $\textsf {D}$ . Naturally, we could ask the following questions:
-
• Is $\langle \textsf {D}, \lhd \rangle $ well-founded?
-
• Are any two elements of $\langle \textsf {D}, \lhd \rangle $ comparable?
-
• Does there exist a minimal theory w.r.t. interpretation such that $\textsf {G1}$ holds for it?
We conjectured the following answers to these questions: $\langle \textsf {D}, \lhd \rangle $ is not well founded, $\langle \textsf {D}, \lhd \rangle $ has incomparable elements, and there is no minimal theory w.r.t. interpretation for which $\textsf {G1}$ holds.
Finally, we discuss the limit of applicability of $\textsf {G1}$ w.r.t. Turing reducibility. We have discussed the limit of applicability of $\textsf {G1}$ w.r.t. interpretation. A natural question is: what is the limit of applicability of $\textsf {G1}$ w.r.t. Turing reducibility?
Definition 4.16 Turing reducibility, the structure $\overline {\textsf {D}}$
-
• Let $\mathcal {R}$ be the structure of the r.e. degrees with the ordering $\leq _{T}$ induced by Turing reducibility with the least element $\mathbf {0}$ and the greatest element $\mathbf {0}^{\prime }$ .
-
• Let $\overline {\textsf {D}}=\{S: S<_{T} \mathbf {R}$ and $\textsf {G1}$ holds for theory S} where $S<_{T} \mathbf {R}$ stands for $S\leq _{T} \mathbf {R}$ but $\mathbf {R} \not\leq _{T} S$ .
Cheng [Reference Cheng22] shows that for any Turing degree $\mathbf {0}< \mathbf {d}<\mathbf {0}^{\prime }$ , there is a theory U such that $\textsf {G1}$ holds for U, $U<_{T} \mathbf {R}$ , and U has Turing degree $\mathbf {d}$ . As a corollary of this result and known results about the degree structure of $\langle \mathcal {R}, <_{T}\rangle $ in recursion theory, we can answer above questions for the structure $\langle \overline {\textsf {D}}, <_{T}\rangle $ :
Theorem 4.17 Cheng [Reference Cheng22]
-
• $\langle \overline {\textsf {D}}, <_{T}\rangle $ is not well-founded;
-
• $\langle \overline {\textsf {D}}, <_{T}\rangle $ has incomparable elements;
-
• There is no minimal theory w.r.t. Turing reducibility such that $\textsf {G1}$ holds for it.
Moreover, Cheng [Reference Cheng22] shows that for any Turing degree $\mathbf {0}< \mathbf {d}<\mathbf {0}^{\prime }$ , there is a theory U such that $\textsf {G1}$ holds for U, $U\unlhd \mathbf {R}$ , and U has Turing degree $\mathbf {d}$ . Thus, examining the limit of applicability of $\textsf {G1}$ w.r.t. interpretation is much harder than that w.r.t. Turing reducibility. The structure of $\langle \textsf {D}, \lhd \rangle $ is a deep and interesting open question for future research.
5 The limit of the applicability of $\textsf {G2}$
5.1 Introduction
In our view, $\textsf {G2}$ is fundamentally different from $\textsf {G1}$ . In fact, both mathematically and philosophically, $\textsf {G2}$ is more problematic than $\textsf {G1}$ for the following reason. On one hand, in the case of $\textsf {G1}$ , we can construct a natural independent sentence with real mathematical content without referring to arithmetization and provability predicates. On the other hand, the meaning of $\textsf {G2}$ strongly depends on how we exactly formulate the consistency statement.
Similar to [Reference Halbach and Visser54], we call a result intensional if it depends on (the details of) the representation used. Thus, $\textsf {G1}$ can be called extensional (that is, non-intensional), while $\textsf {G2}$ is (highly) intensional. We refer to Section 5.3 for more discussion on the intensionality of $\textsf {G2}$ . In this section, we discuss the limit of applicability of $\textsf {G2}$ : under what conditions $\textsf {G2}$ holds, and under what conditions $\textsf {G2}$ fails. In Section 5.2, we discuss generalizations of $\textsf {G2}$ .
5.2 Some generalizations of $\textsf {G2}$
After Gödel, generalizations of $\textsf {G2}$ are the subject of extensive studies. We know that $\textsf {G2}$ holds for any consistent r.e. extension of $\mathbf {PA}$ . However, it is not true that $\textsf {G2}$ holds for any extension of $\mathbf {PA}$ . For example, Karl-Georg Niebergall [Reference Niebergall, Baaz, Friedman and Krajĺček99] shows that the theory ( $\mathbf {PA}+\mathbf {RFN}(\mathbf {PA}))\cap (\mathbf {PA}$ + all true $\Pi ^0_1$ sentences) can prove its own canonical consistency sentence.Footnote 27
Similarly to $\textsf {G1}$ , one can generalise $\textsf {G2}$ to arithmetically definable non-r.e. extensions of $\mathbf {PA}$ . Kikuchi and Kurahashi [Reference Kikuchi and Kurahashi71] reformulate $\textsf {G2}$ as: if S is a $\Sigma ^0_{1}$ -definable and consistent extension of $\mathbf {PA}$ , then for any $\Sigma ^0_{1}$ definition $\sigma (u)$ of S, $S \nvdash \mathbf {Con}_{\sigma }(S)$ (see [Reference Kikuchi and Kurahashi71, Fact 5.1]). Kikuchi and Kurahashi [Reference Kikuchi and Kurahashi71] generalize $\textsf {G2}$ to arithmetically definable non-r.e. extensions of $\mathbf {PA}$ and prove that if S is a $\Sigma ^0_{n+1}$ -definable and $\Sigma ^0_{n}$ -sound extension of $\mathbf {PA}$ , then there exists a $\Sigma ^0_{n+1}$ definition $\sigma (u)$ of some axiomatization of $Th(S)$ such that $\mathbf {Con}_{\sigma }(S)$ is independent of S. This corollary shows that the witness for the generalized version of $\textsf {G1}$ can be provided by the appropriate consistency statement.
Chao–Seraji [Reference Chao and Seraji20] and Kikuchi–Kurahashi [Reference Kikuchi and Kurahashi71] give another generalization of $\textsf {G2}$ to arithmetically definable non-r.e. extensions of $\mathbf {PA}$ : for each $n \in \omega $ , any $\Sigma ^0_{n+1}$ -definable and $\Sigma ^0_{n}$ -sound extension of $\mathbf {PA}$ cannot prove its own $\Sigma ^0_{n}$ -soundness (see [Reference Chao and Seraji20, Theorem 2] and [Reference Kikuchi and Kurahashi71, Theorem 5.6]). The optimality of this generalization is shown in [Reference Chao and Seraji20]: there is a $\Sigma ^0_{n+1}$ -definable and $\Sigma ^0_{n-1}$ -sound extension of $\mathbf {PA}$ that proves its own $\Sigma ^0_{n-1}$ -soundness for $n>0$ (see [Reference Chao and Seraji20, Theorem 3]).
Let T be a consistent r.e. extension of $\mathbf {Q}$ . Kreisel [Reference Kreisel81] shows that $\neg \mathbf {Con}(T)$ is $\Pi ^0_1$ -conservative over T which is a generalization of $\textsf {G2}$ . We can also generalize $\textsf {G2}$ via the notion of standard provability predicate.
Theorem 5.1 Let T be any consistent r.e. extension of $\mathbf {Q}$ . If $\mathbf {Pr}_T(x)$ is a standard provability predicate, then $T\nvdash \mathbf {Con}(T)$ .
Lev Beklemishev and Daniyar Shamkanov [Reference Beklemishev, Shamkanov and Alberti9] prove that in an abstract setting that presupposes the presence of Gödel’s fixed point (instead of directly constructing it, as in the case of formal arithmetic), the Hilbert–Bernays–Löb conditions imply $\textsf {G2}$ even with fairly minimal conditions on the underlying logic. The following two theorems, due to Feferman and Visser, generalize $\textsf {G2}$ in terms of the notion of interpretation.
Theorem 5.2 Feferman’s theorem on the interpretability of inconsistency [Reference Feferman32]
If T is a consistent r.e. extension of $\mathbf {Q}$ , then $T + \neg \mathbf {Con}(T)$ is interpretable in T.
Theorem 5.3 Pudlák [Reference Hájek and Pudlák53, Reference Pudlák108]
There is no consistent r.e. theory S such that ( $\mathbf {Q} + \mathbf {Con}(S)) \unlhd S$ .Footnote 28
As a corollary of Theorem 5.3, for any consistent r.e. theory S that interprets $\mathbf {Q}$ , $\textsf {G2}$ holds for S: $S\nvdash \mathbf {Con}(S)$ . The Arithmetic Completeness Theorem tells us that $S\unlhd (\mathbf {Q} + \mathbf {Con}(S))$ (see [Reference Visser130] for the details). As a corollary, we have the following version of $\textsf {G2}$ which highlights the interpretability power of consistency statements.
Corollary 5.4 For any consistent r.e. theory S, we have $S\lhd (\mathbf {Q} + \mathbf {Con}(S))$ .
Definition 5.5 Let T be a consistent extension of $\mathbf {Q}$ . A formula $I(x)$ with one free variable (understood as a number variable) is a definable cut in T (in short, a T-cut) if:
-
• $T\vdash I(\mathbf {0})$ ;
-
• $T\vdash \forall x (I(x) \rightarrow I(x+1))$ ;
-
• $T\vdash \forall x\forall y (y < x \wedge I(x) \rightarrow I(y))$ .
Definition 5.6 Let $T\supseteq I\Sigma _1$ , let J be a T-cut, and let $\tau $ be a $\Sigma _0^{\mathbf {exp}}$ -definition of T.Footnote 29
-
• $\mathbf {Pr}_{\tau }^I(x)$ is the formula $\exists y(I(y)\wedge \mathbf {Proof}_{\tau }^I(x,y))$ (saying that there is a $\tau $ -proof of x in I).
-
• $\mathbf {Con}_{\tau }^I$ is the formula $\neg \exists y(I(y)\wedge \mathbf {Proof}_{\tau }^I(\mathbf {0}\neq \mathbf {0},y))$ .
The following theorem generalizes $\textsf {G2}$ to definable cuts.
Theorem 5.7 [Reference Hájek and Pudlák53, Theorem 3.11]
Let $T\supseteq I\Sigma _1$ , let J be a T-cut, and let $\tau $ be a $\Sigma _0^{\mathbf {exp}}$ -definition of T. Then $T\nvdash \mathbf {Con}_{\tau }^I$ .
Next, consider a theory U and an interpretation N of the Tarski–Mostowski–Robinson theory $\mathbf {R}$ in U. A U-predicate $\triangle $ is an L-predicate for $U, N$ if it satisfies the following Löb conditions. We write $\triangle A$ for $\triangle (\ulcorner A \urcorner )$ , where $\ulcorner A \urcorner $ is the numeral of the Gödel number of A and we interpret the numbers via N. The Gödel numbering is supposed to be fixed and standard.
Definition 5.8 Löb conditions
-
L1: $\vdash A \Rightarrow \ \vdash \triangle A$ .
-
L2: $\triangle A, \triangle (A \rightarrow B) \vdash \triangle B$ .
-
L3: $\triangle A \vdash \triangle \triangle A$ .
Proposition 5.9 Löb’s theorem [Reference Visser, Horsten and Welch133, Theorem 3.3.2]
Suppose that U is a theory, N is an interpretation of the theory $\mathbf {R}$ in U, and $\triangle $ is a U-predicate that is an L-predicate for $U, N$ . Then:
-
• For all U-sentences A we have: if $U \vdash \triangle A \rightarrow A$ , then $U \vdash A$ .
-
• For all U-sentences A we have: $U \vdash \triangle (\triangle A \rightarrow A) \rightarrow \triangle A$ .
As a corollary of Proposition 5.9, we formulate a general version of $\textsf {G2}$ which does not mention the notion of provability predicate.
Theorem 5.10 Visser [Reference Visser, Horsten and Welch133]
For all consistent theories U and all interpretations N of $\mathbf {R}$ in U and all L-predicates $\triangle $ for $U, N$ , we have $U \nvdash \neg \triangle \perp $ .
5.3 The intensionality of $\textsf {G2}$
In this section, we discuss the intensionality of $\textsf {G2}$ which reveals the limit of the applicability of $\textsf {G2}$ .
5.3.1 Introduction
For a consistent theory T, we say that $\textsf {G2}$ holds for T if the consistency of T is not provable in T. However, this definition is vague, and whether $\textsf {G2}$ holds for T depends on how we formulate the consistency statement. We refer to this phenomenon as the intensionality of $\textsf {G2}$ . In fact, $\textsf {G2}$ is essentially different from $\textsf {G1}$ due to the intensionality of $\textsf {G2}$ : “whether $\textsf {G2}$ holds for the base theory” depends on how we formulate the consistency statement in the first place.
Both mathematically and philosophically, $\textsf {G2}$ is more problematic than $\textsf {G1}$ . In the case of $\textsf {G1}$ , we are mainly interested in the fact that some sentence is independent of $\mathbf {PA}$ . We make no claim to the effect that that sentence “really” expresses what we would express by saying “ $\mathbf {PA}$ cannot prove this sentence”. But in the case of $\textsf {G2}$ , we are also interested in the content of the consistency statement. We can say that $\textsf {G1}$ is extensional in the sense that we can construct a concrete independent mathematical statement without referring to arithmetization and provability predicate. However, $\textsf {G2}$ is intensional and “whether $\textsf {G2}$ holds for T” depends on varied factors as we will discuss.
In this section, unless stated otherwise, we assume the following:
-
• T is a consistent r.e. extension of $\mathbf {Q}$ ;
-
• the canonical arithmetic formula to express the consistency of the base theory T is $\mathbf {Con}(T)\triangleq \neg \mathbf {Pr}_T(\mathbf {0}\neq \mathbf {0})$ ;
-
• the canonical numbering we use is Gödel’s numbering;
-
• the provability predicate we use is standard;
-
• the formula representing the set of axioms is $\Sigma ^0_1$ .
The intensionality of Gödel sentence and the consistency statement has been widely discussed from the literature (e.g., Halbach–Visser [Reference Halbach and Visser54, Reference Halbach and Visser55] and Visser [Reference Visser130]). Halbach and Visser examine the sources of intensionality in the construction of self-referential sentences of arithmetic in [Reference Halbach and Visser54, Reference Halbach and Visser55], and argue that corresponding to the three stages of the construction of self-referential sentences of arithmetic, there are at least three sources of intensionality: coding, expressing a property, and self-reference. The three sources of intensionality are not independent of each other, and a choice made at an earlier stage will have influences on the availability of choices at a later stage. Visser [Reference Visser130] locates three sources of indeterminacy in the formalization of a consistency statement for a theory T:
-
• the choice of a proof system;
-
• the choice of a way of numbering;
-
• the choice of a specific formula representing the set of axioms of T.
In summary, the intensional nature ultimately traces back to the various parameter choices that one has to make in arithmetizing the provability predicate. That is the source of both the intensional nature of the Gödel sentence and the consistency sentence.
Based on this and other works from the literature, we argue that “whether $\textsf {G2}$ holds for the base theory” depends on the following factors:
-
(1) the choice of the provability predicate (Section 5.3.2);
-
(2) the choice of the formula expressing consistency (Section 5.3.3);
-
(3) the choice of the base theory (Section 5.3.4);
-
(4) the choice of the numbering (Section 5.3.5);
-
(5) the choice of the formula representing the set of axioms (Section 5.3.6).
These factors are not independent, and a choice made at an earlier stage may have effects on the choices available at a later stage. In the following, unless stated otherwise, when we discuss how $\textsf {G2}$ depends on one factor, we always assume that other factors are fixed, and only the factor we are discussing is varied. For example, Visser [Reference Visser130] rests on fixed choices for (1)–(2) and (4)–(5) but varies the choice of (3); Grabmayr [Reference Grabmayr46] rests on fixed choices for (1)–(3) and (5) but varies the choice of (4); Feferman [Reference Feferman32] rests on fixed choices for (1)–(4) but varies the choice of (5).
5.3.2 The choice of provability predicate
In this section, we show that “whether $\textsf {G2}$ holds for the base theory” depends on the choice of the provability predicate we use.
As Visser argues in [Reference Visser, Horsten and Welch133], being a consistency statement is not an absolute concept but a role w.r.t. a choice of provability predicate (see Visser [Reference Visser, Horsten and Welch133]). From Theorem 5.1, $\textsf {G2}$ holds for standard provability predicates. However, $\textsf {G2}$ may fail for non-standard provability predicates.
Mostowski [Reference Mostowski96] gives an example of a $\Sigma ^0_1$ provability predicate for which $\textsf {G2}$ fails. Let $\mathbf {Pr}_T^M(x)$ be the $\Sigma ^0_1$ formula “ $\exists y(\mathbf {Prf}_T(x, y) \wedge \neg \mathbf {Prf}_T(\ulcorner \mathbf {0}\neq \mathbf {0}\urcorner , y))$ ” where $\mathbf {Prf}_T(x, y)$ is a $\Delta ^0_1$ formula saying that “y is a proof of x”. Then $\neg \mathbf {Pr}_T^M(\ulcorner 0\neq 0\urcorner )$ is trivially provable in $\mathbf {PA}$ . We know that $\textsf {G2}$ holds for provability predicates satisfying $\mathbf {D1}$ – $\mathbf {D3}$ . Since the formula $\mathbf {Pr}_T^M(x)$ satisfies $\mathbf {D1}$ and $\mathbf {D3}$ , it does not satisfy $\mathbf {D2}$ . Mostowski’s example [Reference Mostowski96] shows that $\textsf {G2}$ may fail for $\Sigma ^0_1$ provability predicates satisfying $\mathbf {D1}$ and $\mathbf {D3}$ .
One important non-standard provability predicate is Rosser provability predicate $\mathbf {Pr}_T^R(x)$ introduced by Rosser [Reference Rosser111] to improve Gödel’s first incompleteness theorem. Recall that we have defined the Rosser provability predicate in Definition 3.8. The consistency statement $\mathbf {Con}^R(T)$ based on a Rosser provability predicate $\mathbf {Pr}_T^R(x)$ is naturally defined as $\neg \mathbf {Pr}^R_{T}(\ulcorner \mathbf {0}\neq \mathbf {0}\urcorner )$ .
It is an easy fact that for any sentence $\phi $ and Rosser provability predicate $\mathbf {Pr}^R_{T}(x)$ , if $T\vdash \neg \phi $ , then $T\vdash \neg \mathbf {Pr}^R_{T}(\ulcorner \phi \urcorner )$ (see [Reference Kikuchi and Kurahashi72, Proposition 2.1]). As a corollary, the consistency statement $\mathbf {Con}^R(T)$ based on Rosser provability predicate $\mathbf {Pr}^R_{T}(x)$ is provable in T. In this sense, we can say that $\textsf {G2}$ fails for the consistency statement constructed from Rosser provability predicates.
We can construct different Rosser provability predicates with varied properties. We know that each Rosser provability predicate $\mathbf {Pr}_{T}^R(x)$ does not satisfy at least one of conditions $\mathbf {D2}$ and $\mathbf {D3}$ . Guaspari and Solovay [Reference Guaspari and Solovay50] establish a very powerful method of constructing a new proof predicate with required properties from a given proof predicate by reordering nonstandard proofs. Applying this tool, Guaspari and Solovay [Reference Guaspari and Solovay50] construct a Rosser provability predicate for which both $\mathbf {D2}$ and $\mathbf {D3}$ fail. Arai [Reference Arai2] constructs a Rosser provability predicate with condition $\mathbf {D2}$ , and a Rosser provability predicate with condition $\mathbf {D3}$ .
Slow provability, introduced by S. D. Friedman, M. Rathjen, and A. Weiermann [Reference Friedman, Rathjen and Weiermann41], is another notion of nonstandard provability for $\mathbf {PA}$ from the literature. The slow consistency statement $\mathbf {Con}^{\ast }(\mathbf {PA})$ asserts that a contradiction is not slow provable in $\mathbf {PA}$ (for the definition of $\mathbf {Con}^{\ast }(\mathbf {PA})$ , we refer to [Reference Friedman, Rathjen and Weiermann41]). In fact, $\textsf {G2}$ holds for slow provability: Friedman, Rathjen, and Weiermann show that $\mathbf {PA} \nvdash \mathbf {Con}^{\ast }(\mathbf {PA})$ (see [Reference Friedman, Rathjen and Weiermann41, Proposition 3.3]). Moreover, Friedman, Rathjen, and Weiermann [Reference Friedman, Rathjen and Weiermann41] show that $\mathbf {PA} + \mathbf {Con}^{\ast }(\mathbf {PA}) \nvdash \mathbf {Con}(\mathbf {PA})$ (see [Reference Friedman, Rathjen and Weiermann41, Theorem 3.10]), and the logical strength of the theory $\mathbf {PA}+\mathbf {Con}^{\ast }(\mathbf {PA})$ lies strictly between $\mathbf {PA}$ and $\mathbf {PA}+ \mathbf {Con}(\mathbf {PA})$ : $\mathbf {PA} \varsubsetneq \mathbf {PA}+ \mathbf {Con}^{\ast }(\mathbf {PA}) \varsubsetneq \mathbf {PA}+ \mathbf {Con}(\mathbf {PA})$ . Henk and Pakhomov [Reference Henk and Pakhomov57] study three variants of slow provability, and show that the associated consistency statement of each of these notions of provability yields a theory that lies strictly between $\mathbf {PA}$ and $\mathbf {PA}+\mathbf {Con}(\mathbf {PA})$ in terms of logical strength.
5.3.3 The choice of the formula expressing consistency
We show that “whether $\textsf {G2}$ holds for the base theory” depends on the choice of the arithmetic formula used to express consistency. In the literature, an arithmetic formula is usually used to express the consistency statement. Artemov [Reference Artemov3] argues that in Hilbert’s consistency program, the original formulation of consistency “no sequence of formulas is a derivation of a contradiction” is about finite sequences of formulas, not about arithmetization, proof codes, and internalized quantifiers.
The canonical consistency statement, the arithmetical formula $\mathbf {Con}(\mathbf {PA})$ , says that for all x, x is not a code of a proof of a contradiction in $\mathbf {PA}$ . In a nonstandard model of $\mathbf {PA}$ , the universal quantifier “for all x” ranges over both standard and nonstandard numbers, and hence $\mathbf {Con}(\mathbf {PA})$ expresses the consistency of both standard and nonstandard proof codes (see [Reference Artemov3]). Thus, $\mathbf {Con}(\mathbf {PA})$ is stronger than the original formulation of consistency which only talks about sequences of formulas and such sequences have only standard codes. Hence, Artemov [Reference Artemov3] concludes that $\textsf {G2}$ , saying that $\mathbf {PA}$ cannot prove $\mathbf {Con}(\mathbf {PA})$ , does not actually exclude finitary consistency proofs of the original formulation of consistency (see [Reference Artemov3]).
Artemov shows that the original formulation of consistency admits a direct proof in informal arithmetic, and this proof is formalizable in $\mathbf {PA}$ (see [Reference Artemov3]).Footnote 30 Artemov’s work establishes the consistency of $\mathbf {PA}$ by finitary means, and vindicates Hilbert’s consistency program to some extent.
In the following, we use a single arithmetic sentence to express the consistency statement. Among consistency statements defined via arithmetization, there are three candidates of arithmetic formulas to express consistency as follows:
-
• $\mathbf {Con}^0(T) \triangleq \forall x(\mathbf {Fml}(x) \wedge \mathbf {Pr}_T(x) \rightarrow \neg \mathbf {Pr}_T(\dot {\neg } x))$ ;Footnote 31
-
• $\mathbf {Con}(T) \triangleq \neg \mathbf {Pr}_T(\ulcorner \mathbf {0}\neq \mathbf {0}\urcorner )$ ;
-
• $\mathbf {Con}^1(T) \triangleq \exists x(\mathbf {Fml}(x) \wedge \neg \mathbf {Pr}_T(x))$ .
Gödel originally formulates $\textsf {G2}$ with the consistency statement $\mathbf {Con}^1(T)$ : if T is consistent, then $T\nvdash \mathbf {Con}^1(T)$ . From the literature, $\mathbf {Con}(T)$ is the widely used canonical consistency statement. Note that $\mathbf {Con}^0(T)$ implies $\mathbf {Con}(T)$ , and $\mathbf {Con}(T)$ implies $\mathbf {Con}^1(T)$ . However the converse implications do not hold in general (see [Reference Kurahashi88]). Kurahashi [Reference Kurahashi88] proposes different sets of derivability conditions (local version, uniform version, and global version), and examines whether they are sufficient to show the unprovability of these consistency statements (e.g., $\mathbf {Con}(T), \mathbf {Con}^0(T),$ and $\mathbf {Con}^1(T)$ ).
-
HB1: If $T \vdash \phi \rightarrow \varphi $ , then $T \vdash \mathbf {Pr}_T(\ulcorner \phi \urcorner ) \rightarrow \mathbf {Pr}_T(\ulcorner \varphi \urcorner )$ .
-
HB2: $T \vdash \mathbf {Pr}_T(\ulcorner \neg \phi (x)\urcorner ) \rightarrow \mathbf {Pr}_T(\ulcorner \neg \phi (\dot {x})\urcorner )$ .
-
HB3: $T \vdash f(x) = 0 \rightarrow \mathbf {Pr}_T(\ulcorner f(\dot {x}) = 0\urcorner )$ for every primitive recursive term $f(x)$ .
$\mathbf {HB1}$ – $\mathbf {HB3}$ are called the Hilbert–Bernays derivability conditions. If a provability predicate $\mathbf {Pr}_T(x)$ satisfies $\mathbf {HB1}$ – $\mathbf {HB3}$ , then $T\nvdash \mathbf {Con}^0(T)$ (see [Reference Hilbert and Bernays59]). Kurahashi [Reference Kurahashi89] constructs two Rosser provability predicates satisfying $\mathbf {HB1}$ – $\mathbf {HB3}$ . Thus, $\mathbf {HB1}$ – $\mathbf {HB3}$ are not sufficient to prove that $T\nvdash \mathbf {Con}(T)$ .
Löb [Reference Löb92] proves that if $\mathbf {Pr}_T(x)$ satisfies the Hilbert–Bernays–Löb derivability conditions $\mathbf {D1}$ – $\mathbf {D3}$ (see Definition 3.6), then Löb’s theorem holds: for any sentence $\phi $ , if $T \vdash \mathbf {Pr}_T(\ulcorner \phi \urcorner ) \rightarrow \phi $ , then $T\vdash \phi $ . It is well-known that Löb’s theorem implies $\textsf {G2}$ : $T\nvdash \mathbf {Con}(T)$ (see [Reference Murawski97]). Thus, if a provability predicate $\mathbf {Pr}_T(x)$ satisfies $\mathbf {D1}$ – $\mathbf {D3}$ , then $T\nvdash \mathbf {Con}(T)$ . Kurahashi [Reference Kurahashi88, Proposition 4.11] constructs a provability predicate $\mathbf {Pr}_T(x)$ with conditions $\mathbf {D1}$ – $\mathbf {D3}$ , but $T\vdash \mathbf {Con}^1(T)$ . Thus, $\mathbf {D1}$ – $\mathbf {D3}$ are not sufficient to prove that $T\nvdash \mathbf {Con}^1(T)$ .
Montagna [Reference Montagna94] proves that if a provability predicate $\mathbf {Pr}_T(x)$ satisfies the following two conditions, then $T\nvdash \mathbf {Con}^1(T)$ :
-
• $T \vdash \forall x$ (“x is a logical axiom” $\rightarrow \mathbf {Pr}_T(x)$ );
-
• $T \vdash \forall x\forall y(\mathbf {Fml}(x) \wedge \mathbf {Fml}(y) \rightarrow (\mathbf {Pr}_T(x\rightarrow y) \rightarrow (\mathbf {Pr}_T(x) \rightarrow \mathbf {Pr}_T(y))))$ .
5.3.4 The choice of base theory
We show that “whether $\textsf {G2}$ holds for the base theory” depends on the base theory we choose. A foundational question about $\textsf {G2}$ is: how much information about arithmetic is required for the proof of $\textsf {G2}$ ? If the base system does not contain enough information about arithmetic, then $\textsf {G2}$ may fail. The widely used notion of consistency is consistency in proof systems with cut elimination. However, notions like cutfree consistency, Herbrand consistency, tableaux consistency, and restricted consistency for different base theories behave differently (see [Reference Visser130]). We do have proof systems that prove their own cutfree consistency: for example, finitely axiomatized sequential theories prove their own cut-free consistency on a definable cut (see [Reference Visser136, p. 25]).
A natural question is: whether $\textsf {G2}$ can be generalized to base systems weaker than $\mathbf {PA}$ w.r.t. interpretation? As a corollary of Theorem 5.3, we have $\mathbf {Q}\nvdash \mathbf {Con(Q)}$ and hence $\mathbf {Q}\nvdash \mathbf {Con}^0(\mathbf {Q})$ . Bezboruah and Shepherdson [Reference Bezboruah and Shepherdson11] define the consistency of $\mathbf {Q}$ as the sentence $\mathbf {Con}^0(\mathbf {Q})$ Footnote 32 , and prove that $\textsf {G2}$ holds for $\mathbf {Q}: \mathbf {Q}\nvdash \mathbf {Con}^0(\mathbf {Q})$ . However, the method used by Bezboruah and Shepherdson in [Reference Bezboruah and Shepherdson11] is quite different from Theorem 5.3. Bezboruah–Shepherdson’s proof depends on some specific assumptions about the coding, does not easily generalize to stronger theories, and tells us nothing about the question whether $\mathbf {Q}$ can prove its consistency on some definable cut (see [Reference Visser132]). The next question is: whether $\textsf {G2}$ holds for other theories weaker than $\mathbf {Q}$ w.r.t. interpretation (e.g., $\mathbf {R}$ )? In a forthcoming paper, we will show that $\textsf {G2}$ holds for $\mathbf {R}$ via the canonical consistency statement. However, we can find weak theories mutually interpretable with $\mathbf {R}$ for which $\textsf {G2}$ fails.
Willard [Reference Willard145] explores the generality and boundary-case exceptions of $\textsf {G2}$ over some base theories. Willard constructs examples of r.e. arithmetical theories that cannot prove the totality of their successor functions but can prove their own canonical consistencies (see [Reference Willard144, Reference Willard145]). However, the theories Willard constructs are not completely natural since some axioms are constructed using Gödel’s Diagnolisation Lemma. Pakhomov [Reference Pakhomov102] constructs a more natural example of this kind. Pakhomov [Reference Pakhomov102] defines a theory $H_{<\omega }$ , and shows that it proves its own canonical consistency. Unlike Willard’s theories, $H_{<\omega }$ isn’t an arithmetical theory but a theory formulated in the language of set theory with an additional unary function. From [Reference Pakhomov102], $H_{<\omega }$ and $\mathbf {R}$ are mutually interpretable. Hence, the theory $H_{<\omega }$ can be regarded as the set-theoretic analogue of $\mathbf {R}$ from the interpretability theoretic point of view.
From Theorem 5.3, $\textsf {G2}$ holds for any consistent r.e. theory interpreting $\mathbf {Q}$ . However, it is not true that $\textsf {G2}$ holds for any consistent r.e. theory interpreting $\mathbf {R}$ since $H_{<\omega }$ interprets $\mathbf {R}$ , but $\textsf {G2}$ fails for $H_{<\omega }$ . We know that if $S\unlhd T$ and $\textsf {G1}$ holds for S, then $\textsf {G1}$ holds for T. However, it is not true that if $S\unlhd T$ and $\textsf {G2}$ holds for S, then $\textsf {G2}$ holds for T since $\mathbf {R}\unlhd H_{<\omega }$ , $\textsf {G2}$ holds for $\mathbf {R}$ but $\textsf {G2}$ fails for $H_{<\omega }$ . This shows the difference between $\mathbf {Q}$ and $\mathbf {R}$ , and the difference between $\textsf {G1}$ and $\textsf {G2}$ .
One way to eliminate the intensionality of $\textsf {G2}$ is to uniquely characterize the consistency statement. In [Reference Visser130], Visser proposes the interesting question of a coordinate-free formulation of $\textsf {G2}$ and a unique characterization of the consistency statement. Visser [Reference Visser130] shows that consistency for finitely axiomatized sequential theories can be uniquely characterized modulo $\mathbf {EA}$ -provable equivalence (see [Reference Visser130, p. 543]). But characterizing the consistency of infinitely axiomatized r.e. theories is more delicate and a big open problem in the current research on the intensionality of $\textsf {G2}$ .
After Gödel, Gentzen constructs a theory $\mathbf {T}^{\ast }$ (primitive recursive arithmetic with the additional principle of quantifier-free transfinite induction up to the ordinal $\epsilon _0$ ),Footnote 33 and proves the consistency of $\mathbf {PA}$ in $\mathbf {T}^{\ast }$ . Gentzen’s theory $\mathbf {T}^{\ast }$ contains $\mathbf {Q}$ but does not contain $\mathbf {PA}$ since $\mathbf {T}^{\ast }$ does not prove the ordinary mathematical induction for all formulas. By the Arithmetized Completeness Theorem, $\mathbf {Q}+\mathbf {Con}(\mathbf {PA})$ interprets $\mathbf {PA}$ . Since Gentzen’s theory $\mathbf {T}^{\ast }$ contains $\mathbf {Q}$ and $\mathbf {T}^{\ast }\vdash \mathbf {Con}(\mathbf {PA})$ , Gentzen’s theory $\mathbf {T}^{\ast }$ interprets $\mathbf {PA}$ . By Pudlák’s result that no consistent r.e. extension T of $\mathbf {Q}$ can interpret $\mathbf {Q}+\mathbf {Con}(T)$ , $\mathbf {PA}$ does not interpret Gentzen’s theory $\mathbf {T}^{\ast }$ . Thus $\mathbf {PA}\lhd \mathbf {T}^{\ast }$ . Gentzen’s work has opened a productive new direction in proof theory: finding the means necessary to prove the consistency of a given theory. More powerful subsystems of Second-Order Arithmetic have been given consistency proofs by Gaisi Takeuti and others, and theories that have been proved consistent by these methods are quite strong and include most ordinary mathematics.
5.3.5 The choice of numbering
We show that “whether $\textsf {G2}$ holds for the base theory” depends on the choice of the numbering encoding the language.
For the influence of different numberings on $\textsf {G2}$ , we refer to [Reference Grabmayr46]. Any injective function $\gamma $ from a set of $L(\mathbf {PA})$ -expressions to $\omega $ qualifies as a numbering. Gödel’s numbering is a special kind of numberings under which the Gödel number of the set of axioms of $\mathbf {PA}$ is recursive. In fact, $\textsf {G2}$ is sensitive to the way of numberings. Let $\gamma $ be a numbering and $\ulcorner \varphi ^{\gamma }\urcorner $ denote $\overline {\gamma (\varphi )}$ , i.e., the standard numeral of the $\gamma $ -code of $\varphi $ .
Definition 5.11 Relativized Löb conditions
A formula $\mathbf {Pr}_T^{\gamma }(x)$ is said to satisfy Löb’s conditions relative to $\gamma $ for the base theory T if for all $L(\mathbf {PA})$ -sentences $\varphi $ and $\psi $ we have that:
-
$\mathbf {D1^{\ast }}\!\!:$ If $T \vdash \varphi $ , then $\mathbf {PA} \vdash \mathbf {Pr}_T^{\gamma }(\ulcorner \varphi ^{\gamma }\urcorner )$ ;
-
$\mathbf {D2^{\ast }}\!\!:$ $T \vdash \mathbf {Pr}_T^{\gamma }(\ulcorner (\varphi \rightarrow \psi )^{\gamma }\urcorner )\rightarrow (\mathbf {Pr}_T^{\gamma }(\ulcorner \varphi ^{\gamma }\urcorner )\rightarrow \mathbf {Pr}_T^{\gamma }(\ulcorner \psi ^{\gamma }\urcorner ))$ ;
-
$\mathbf {D3^{\ast }}\!\!:$ $T \vdash \mathbf {Pr}_T^{\gamma }(\ulcorner \varphi ^{\gamma }\urcorner )\rightarrow \mathbf {Pr}_T^{\gamma }(\ulcorner (\mathbf {Pr}_T^{\gamma }(\ulcorner \varphi ^{\gamma }\urcorner ))^{\gamma }\urcorner )$ .
Grabmayr [Reference Grabmayr46] examines different criteria of acceptability, and proves the invariance of $\textsf {G2}$ with regard to acceptable numberings (for the definition of acceptable numberings, we refer to [Reference Grabmayr46]).
Theorem 5.12 Invariance of $\textsf {G2}$ under acceptable numberings [Reference Grabmayr46, Theorem 4.8]
Let $\gamma $ be an acceptable numbering and T be a consistent r.e. extension of $\mathbf {Q}$ . If $\mathbf {Pr}_T^{\gamma }(x)$ satisfies Löb’s conditions $\mathbf {D1^{\ast }}$ – $\mathbf {D3^{\ast }}$ relative to $\gamma $ for T, then $T \nvdash \neg \mathbf {Pr}_T^{\gamma }(\ulcorner (\mathbf {0}\neq \mathbf {0})^{\gamma }\urcorner )$ .
Theorem 5.12 shows that $\textsf {G2}$ holds for acceptable numberings. But $\textsf {G2}$ may fail for non-acceptable numberings. Grabmayr [Reference Grabmayr46] gives some examples of deviant numberings $\gamma $ such that $\textsf {G2}$ fails w.r.t. $\gamma $ : $T\vdash \mathbf {Pr}_T^{\gamma }(\ulcorner (\mathbf {0}\neq \mathbf {0})^{\gamma }\urcorner )$ .
Definition 5.13 We say that $\alpha (x)$ is a numeration of T if for any n, we have $\mathbf {PA} \vdash \alpha (\overline {n})$ if and only if n is the Gödel number of some $\phi \in T$ .
5.3.6 The choice of the formula representing the set of axioms
We show that “whether $\textsf {G2}$ holds for T” depends on the way the axioms of T are represented.
First of all, Definition 5.14 gives a more general definition of provability predicate and consistency statement for T w.r.t. the numeration of T.
Definition 5.14 Let T be any consistent r.e. extension of $\mathbf {Q}$ and $\alpha (x)$ be a formula in $L(T)$ .
-
• Define the formula $\mathbf {Prf}_{\alpha }(x,y)$ saying “y is the Gödel number of a proof of the formula with Gödel number x from the set of all sentences satisfying $\alpha (x)$ ”.
-
• Define the provability predicate $\mathbf {Pr}_{\alpha }(x)$ of $\alpha (x)$ as $\exists y \mathbf {Prf}_{\alpha }(x,y)$ and the consistency statement $\mathbf {Con}_{\alpha }(T)$ as $\neg \mathbf {Pr}_{\alpha }(\ulcorner \mathbf {0}\neq \mathbf {0}\urcorner )$ .
For each formula $\alpha (x)$ , we have:
If $\alpha (x)$ is a numeration of T, then $\mathbf {Pr}_{\alpha }(x)$ satisfies the following properties (see [Reference Kurahashi86, Fact 2.2]):
-
$\mathbf {D1^{\prime }}\!\!:$ If $T \vdash \varphi $ , then $\mathbf {PA} \vdash \mathbf {Pr}_{\alpha }(\ulcorner \varphi \urcorner )$ ;
-
$\mathbf {D3^{\prime }}\!\!:$ If $\varphi $ is $\Sigma ^0_1$ , then $\mathbf {PA} \vdash \varphi \rightarrow \mathbf {Pr}_{\alpha }(\ulcorner \varphi \urcorner )$ .
Now we give a new reformulation of $\textsf {G2}$ via numerations.
Theorem 5.15 Let T be any consistent r.e. extension of $\mathbf {Q}$ . If $\alpha (x)$ is any $\Sigma ^0_1$ numeration of T, then $T\nvdash \mathbf {Con}_{\alpha }(T)$ .
In fact, $\textsf {G2}$ holds for any $\Sigma ^0_1$ numeration of T, but fails for some $\Pi ^0_1$ numeration of T. Feferman [Reference Feferman32] constructs a $\Pi ^0_1$ numeration $\pi (x)$ of T such that $\textsf {G2}$ fails, i.e., $\mathbf {Con}_{\pi }(T)\triangleq \neg \mathbf {Pr}_{\pi }(\ulcorner \mathbf {0}\neq \mathbf {0}\urcorner )$ is provable in T. Feferman’s construction keeps the proof predicate and its numbering fixed but varies the formula representing the set of axioms. Notice that Feferman’s predicate satisfies $\mathbf {D1}$ and $\mathbf {D2}$ , but does not satisfy $\mathbf {D3}$ . Feferman’s example shows that $\textsf {G2}$ may fail for provability predicates satisfying $\mathbf {D1}$ and $\mathbf {D2}$ .
Generally, Feferman [Reference Feferman32] shows that if T is a $\Sigma ^0_{1}$ -definable extension of $\mathbf {Q}$ , then there is a $\Pi ^0_{1}$ definition $\tau (u)$ of T such that $T \vdash \mathbf {Con}_{\tau }(T)$ . In summary, $\textsf {G2}$ is not coordinate-free (it is dependent on numerations of $\mathbf {PA}$ ). An important question is how to formulate $\textsf {G2}$ in a general way such that it is coordinate-free (independent of numerations of T).
The properties of the provability predicate are intensional and depend on the numeration of the theory. I.e., under different numerations of T, the provability predicate may have different properties. It may happen that T has two numerations $\alpha (x)$ and $\beta (x)$ such that $\mathbf {Con}_{\alpha }(T)$ is not equivalent to $\mathbf {Con}_{\beta }(T)$ . For example, under Gödel’s recursive numeration $\tau (x)$ and Feferman’s $\Pi ^0_1$ numeration $\pi (x)$ of T, the corresponding consistency statements $\mathbf {Con}_{\tau }(T)$ and $\mathbf {Con}_{\pi }(T)$ are not equivalent. But ${\mathbf {PA}}$ does not know this fact, i.e., ${\mathbf {PA}}\nvdash \neg (\mathbf {Con}_{\tau }(T)\leftrightarrow \mathbf {Con}_{\pi }(T))$ since ${\mathbf {PA}}\nvdash \neg \mathbf {Con}_{\tau }(T)$ .
Generally, Kikuchi and Kurahashi prove in [Reference Kikuchi and Kurahashi71, Corollary 5.11] that if T is $\Sigma ^0_{n+1}$ -definable and not $\Sigma ^0_{n}$ -sound, then there are $\Sigma ^0_{n+1}$ definitions $\sigma _1(x)$ and $\sigma _2(x)$ of T such that $T\vdash \mathbf {Con}_{\sigma _1}(T)$ and $T \vdash \neg \mathbf {Con}_{\sigma _2}(T)$ .
Provability logic is an important tool for the study of incompleteness and meta-mathematics of arithmetic. The origins of provability logic (e.g., Henkin’s problem, the isolation of derivability conditions, and Löb’s theorem) are all closely tied to Gödel’s incompleteness theorems historically. In this sense, we can say that Gödel’s incompleteness theorems play a unifying role between first order arithmetic and provability logic.
Provability logic is the logic of properties of provability predicates. Note that $\textsf {G2}$ is very sensitive to the properties of the provability predicate used in its formulation. Provability logic provides us with a new viewpoint and an important tool that can be used to understand incompleteness. Provability logic based on different provability predicates reveals the intensionality of provability predicates which is one source of the intensionality of $\textsf {G2}$ .
Let T be a consistent r.e. extension of $\mathbf {Q}$ , and $\tau (u)$ be any numeration of T. Recall that an arithmetical interpretation f is a mapping from the set of all modal propositional variables to the set of $L(T)$ -sentences. Every arithmetical interpretation f is uniquely extended to the mapping $f_{\tau }$ from the set of all modal formulas to the set of $L(T)$ -sentences such that $f_{\tau }$ satisfies the following conditions:
-
• $f_{\tau }(p)$ is $f(p)$ for each propositional variable p;
-
• $f_{\tau }(\bot )$ is $\mathbf {0} \neq \mathbf {0}$ ;
-
• $f_{\tau }$ commutes with every propositional connective;
-
• $f_{\tau }(\boxed{} A)$ is $\mathbf {Pr}_{\tau }(\ulcorner f_{\tau }(A)\urcorner )$ for every modal formula A.
Provability logic provides us with a new way of examining the intensionality of provability predicates. Under different numerations of T, the provability predicate may have different properties, and hence may correspond to different modal principles under different arithmetical interpretations.
Definition 5.16 Given a numeration $\tau (u)$ of T, the provability logic $\mathbf {PL}_{\tau }(T)$ of $\tau (u)$ is defined to be the set of modal formulas A such that $T \vdash f_{\tau }(A)$ for all arithmetical interpretations f.
Note that the provability logic $\mathbf {PL}_{\tau }(T)$ of a $\Sigma ^0_n$ numeration $\tau (x)$ of T is a normal modal logic. A natural and interesting question is: which normal modal logic can be realized as a provability logic $\mathbf {PL}_{\tau }(T)$ of some $\Sigma ^0_n$ numeration $\tau (x)$ of T? An interesting research program is to classify the provability logic $\mathbf {PL}_{\alpha }(T)$ according to the numeration $\alpha (x)$ of T. We first discuss $\Sigma ^0_1$ numerations of T.
Theorem 5.17 Generalized Solovay’s Arithmetical Completeness Theorem [Reference Kurahashi86, Theorem 2.5]
Let T be any consistent r.e. extension of $\mathbf {PA}$ . If T is $\Sigma ^0_1$ -sound, then for any $\Sigma ^0_1$ numeration $\alpha (x)$ of T, the provability logic $\mathbf {PL}_{\alpha }(T)$ is precisely $\mathbf {GL}$ .Footnote 34
Moreover, Visser [Reference Visser128] examines all possible provability logics for $\Sigma ^0_1$ numerations of $\Sigma ^0_1$ -unsound theories. To state Visser’s result, we need some definitions.
Definition 5.18 [Reference Kurahashi87, Definitions 3.5 and 3.6]
We define the sequence $\{\mathbf {Con}_{\tau }^n: n\in \omega \}$ recursively as follows: $\mathbf {Con}_{\tau }^0$ is $\mathbf {0} = \mathbf {0}$ , and $\mathbf {Con}_{\tau }^{n+1}$ is $\neg \mathbf {Pr}_{\tau }(\ulcorner \neg \mathbf {Con}_{\tau }^{n}\urcorner )$ . The height of $\tau (u)$ is the least natural number n such that $T \vdash \neg \mathbf {Con}_{\tau }^{n}$ if such an n exists. If not, the height of $\tau (u)$ is $\infty $ .
For $\Sigma ^0_1$ -unsound theories, Visser proves that $\mathbf {PL}_{\tau }(T)$ is determined by the height of the numeration $\tau (u)$ . Visser [Reference Visser128, Theorem 3.7] shows that the height of $\tau (u)$ is $\infty $ if and only if $\mathbf {PL}(\tau ) = \mathbf {GL}$ ; and the height of $\tau (u)$ is n if and only if $\mathbf {PL}(\tau ) = \mathbf {GL} + \boxed{} ^n \bot $ . Beklemishev [Reference Beklemishev6, Lemma 7] shows that if T is $\Sigma ^0_1$ -unsound, then the height of $\Sigma ^0_1$ numerations of T can take any values except $0$ .
Let U be any consistent theory of arithmetic. Based on the previous work by Artemov, Visser, and Japaridze, Beklemishev [Reference Beklemishev6] proves that for $\Sigma ^0_1$ numeration $\tau $ of U, $\mathbf {PL}_{\tau }(U)$ coincides with one of the logics $\mathbf {GL}_{\alpha }, \mathbf {D}_{\beta }, \mathbf {S}_{\beta }$ , and $\mathbf {GL}^-_{\beta }$ where $\alpha $ and $\beta $ are subsets of $\omega $ and $\beta $ is cofinite (for definitions of $\mathbf {GL}_{\alpha }, \mathbf {D}_{\beta }, \mathbf {S}_{\beta }$ and $\mathbf {GL}^-_{\beta }$ , we refer to [Reference Beklemishev6]).
Feferman [Reference Feferman32] constructs a $\Pi ^0_1$ numeration $\pi (x)$ of T such that the consistency statement $\mathbf {Con}_{\pi }(T)$ defined via $\mathbf {Pr}_{\pi }(x)$ is provable in T. Thus, the provability logic $\mathbf {PL}_{\pi }(T)$ of $\mathbf {Pr}_{\pi }(x)$ contains the formula $\neg \boxed{} \bot $ , and is different from $\mathbf {GL}$ . However, the exact axiomatization of the provability logic $\mathbf {PL}_{\pi }(T)$ under Feferman’s numeration $\pi (x)$ is not known. Kurahashi [Reference Kurahashi86] proves that for any recursively axiomatized consistent extension T of $\mathbf {PA}$ , there exists a $\Sigma ^0_2$ numeration $\alpha (x)$ of T such that the provability logic $\mathbf {PL}_{\alpha }(T)$ is the modal system $\mathbf {K}$ . As a corollary, the modal principles commonly contained in every provability logic $\mathbf {PL}_{\alpha }(T)$ of T is just $\mathbf {K}$ .
It is often thought that a provability predicate satisfies $\mathbf {D1}$ – $\mathbf {D3}$ if and only if $\textsf {G2}$ holds (i.e., for the induced consistency statement $\mathbf {Con}(T)$ from the provability predicate, $T\nvdash \mathbf {Con}(T)$ ). But this is not true. From Definition 5.14, conditions $\mathbf {D1}$ and $\mathbf {D2}$ hold for any numeration of T. Whether the provability predicate satisfies condition $\mathbf {D3}$ depends on the numeration of T. For any $\Sigma ^0_1$ -numeration $\alpha (x)$ of T, $\mathbf {D3}$ holds for $\mathbf {Pr}_{\alpha }(x)$ . From Kurahashi [Reference Kurahashi86], there is a $\Sigma ^0_2$ -numeration $\alpha (x)$ of T such that the provability logic for that numeration is precisely $\mathbf {K}$ . Since $\mathbf {K}\nvdash \neg \square \bot $ , as a corollary, $\textsf {G2}$ holds for T, i.e., $\mathbf {Con}_{\alpha }(T)$ defined as $\neg \mathbf {Pr}_{\alpha }(\ulcorner \mathbf {0}\neq \mathbf {0}\urcorner )$ is not provable in T. But the Löb condition $\mathbf {D3}$ does not hold since $\mathbf {K}\nvdash \square A\rightarrow \square \square A$ . This gives us an example of a $\Sigma ^0_2$ numeration $\alpha (x)$ of T such that $\mathbf {D3}$ does not hold for $\mathbf {Pr}_{\alpha }(x)$ but $\textsf {G2}$ holds for T. Thus, $\textsf {G2}$ may hold for a provability predicate which does not satisfy the Löb condition $\mathbf {D3}$ .
Moreover, Kurahashi [Reference Kurahashi87] proves that for each $n \geq 2$ , there exists a $\Sigma ^0_2$ numeration $\tau (x)$ of T such that the provability logic $\mathbf {PL}_{\tau }(T)$ is just the modal logic $\mathbf {K} + \square (\square ^n p \rightarrow p) \rightarrow \square p$ . Hence there are infinitely many normal modal logics that are provability logics for some $\Sigma ^0_2$ numeration of T. A good question from Kurahashi [Reference Kurahashi87] is: for $n \geq 2$ , is the class of provability logics $\mathbf {PL}_{\tau }(T)$ for $\Sigma ^0_n$ numerations $\tau (x)$ of T the same as the class of provability logics $\mathbf {PL}_{\tau }(T)$ for $\Sigma ^0_{n+1}$ numerations $\tau (x)$ of T? However, this question is still open as far as we know. Define that $\mathbf {KD}=\mathbf {K}+\neg \boxed{} \bot $ . A natural and interesting question, which is also open as far as we know, is: can we find a numeration $\tau (x)$ of T such that $\mathbf {PL}_{\tau }(T)=\mathbf {KD}$ ?
In summary, $\textsf {G2}$ is intensional with respect to the following parameters: the formalization of consistency, the base theory, the method of numbering, the choice of a provability predicate, and the representation of the set of axioms. Current research on incompleteness reveals that $\textsf {G2}$ is a deep and profound theorem both mathematically and philosophically in the foundations of mathematics, and there is a lot more to be explored about the intensionality of $\textsf {G2}$ .
6 Conclusion
We conclude this paper with some personal comments. To the author, the research on concrete incompleteness is very deep and important.
After Gödel, people have found many different proofs of incompleteness theorems via pure logic, and many concrete independent statements with real mathematical contents. As Harvey Friedman comments, the research on concrete mathematical incompleteness shows how the Incompleteness Phenomena touch normal concrete mathematics, and reveal the impact and significance of the foundations of mathematics.
Harvey Friedman’s research project on concrete incompleteness plans to show that we will be able to find, in just about any subject of mathematics, many natural looking statements that are independent of $\mathbf {ZFC}$ . Harvey Friedman’s work is very profound and promising, and will reveal that incompleteness is everywhere in mathematics, which, if it is true, may be one of the most important discoveries after Gödel in the foundations of mathematics.
Acknowledgment
I would like to thank Matthias Baaz, Ulrich Kohlenbach, Taishi Kurahashi, Zachiri McKenzie, Fedor Pakhomov, Michael Rathjen, Saeed Salehi, Sam Sanders, and Albert Visser for their valuable comments on this work. Especially, I would like to thank Sam Sanders and Zachiri McKenzie for their proofreading of my English writing of this paper. I would like to thank the referees for providing detailed and helpful comments for improvements. I thank the lamp for my feet, and the light on my path. This paper is the research result of the Humanities and Social Sciences of Ministry of Education Planning Fund project “Research on Gödel’s incompleteness theorem” (project no: 17YJA72040001). I would like to thank the fund support by the Humanities and Social Sciences of Ministry of Education Planning Fund.
Some materials of my old paper “Note on some misinterpretations of Gödel’s incompleteness theorems” have been incorporated into this paper.