The Surprise Examination Paradox and the Second Incompleteness Theorem^††thanks: First published in Notices of the AMS volume 57 number 11 (December 2010), published by the American Mathematical Society.

G$\ddot{\mbox{o}}$del’s original proof for the first incompleteness theorem [Gödel31] is based on the liar paradox.

The liar paradox: consider the statement “this statement is false”. The statement can be neither true nor false.

G$\ddot{\mbox{o}}$del considered the related statement “this statement has no proof”. He showed that this statement can be expressed in any theory that is capable of expressing elementary arithmetic. If the statement has a proof, then it is false; but since in a consistent theory any statement that has a proof must be true, we conclude that if the theory is consistent the statement has no proof. Since the statement has no proof, it is true (over ${\mathbb{N}}$). Thus, if the theory is consistent, we have an example for a true statement (over ${\mathbb{N}}$) that has no proof.

The main conceptual difficulty in G$\ddot{\mbox{o}}$del’s original proof is the self-reference of the statement “this statement has no proof”. A conceptually simpler proof of the first incompleteness theorem, based on Berry’s paradox, was given by Chaitin [Chaitin71].

Berry’s paradox: consider the expression “the smallest positive integer not definable in under eleven words”. This expression defines that integer in under eleven words.

To formalize Berry’s paradox, Chaitin uses the notion of Kolmogorov complexity. The Kolmogorov complexity $K(x)$ of an integer $x$ is defined to be the length (in bits) of the shortest computer program that outputs $x$ (and stops). Formally, to define $K(x)$ one has to fix a programming language, such as LISP, Pascal or C++. Alternatively, one can define $K(x)$ by considering any universal Turing machine.

Chaitin’s incompleteness theorem states that for any rich enough consistent mathematical theory, there exists a (large enough) integer $L$ (depending on the theory and on the programming language that is used to define Kolmogorov complexity), such that, for any integer $x$, the statement “$K(x)>L$” cannot be proved within the theory.

The proof given by Chaitin is as follows. Let $L$ be a large enough integer. Assume for a contradiction that for some integer $x$, there is a proof for the statement “$K(x)>L$”. Let $w$ be the first proof (say, according to the lexicographic order) for a statement of the form “$K(x)>L$”. Let $z$ be the integer $x$ such that $w$ proves “$K(x)>L$”. It is easy to give a computer program that outputs $z$: the program enumerates all possible proofs $w$, one by one, and for the first $w$ that proves a statement of the form “$K(x)>L$”, the program outputs $x$ and stops. The length of this program is a constant + $\log L$. Thus, if $L$ is large enough, the Kolmogorov complexity of $z$ is less than $L$. Since $w$ is a proof for “$K(z)>L$” (which is a false statement), we conclude that the theory is inconsistent.

Note that the number of computer programs of length $L$ bits is at most $2^{L+1}$. Hence, for any integer $L$, there exists an integer $0\leq x\leq 2^{L+1}$, such that $K(x)>L$. Thus, for some integer $x$, the statement “$K(x)>L$” is a true statement (over ${\mathbb{N}}$) that has no proof.

A different proof for G$\ddot{\mbox{o}}$del’s first incompleteness theorem, also based on Berry’s paradox, was given by Boolos [Boolos89] (see also [Vopenka66, Kikuchi94]). Other proofs for the first incompleteness theorem are also known (for a recent survey, see [Kotlarski04]).

The second incompleteness theorem follows directly from G$\ddot{\mbox{o}}$del’s original proof for the first incompleteness theorem. As described above, G$\ddot{\mbox{o}}$del expressed the statement “this statement has no proof” and showed that, if the theory is consistent, this is a true statement (over ${\mathbb{N}}$) that has no proof. Informally, since the proof that this is a true statement can be obtained within any rich enough theory, such as Peano Arithmetic (PA) or ZFC, if the consistency of the theory itself can also be proved within the theory, then the statement can be proved within the theory, which is a contradiction. Hence, if the theory is rich enough, the consistency of the theory cannot be proved within the theory.

Thus, the second incompleteness theorem follows directly from G$\ddot{\mbox{o}}$del’s original proof for the first incompleteness theorem. However, the second incompleteness theorem doesn’t follow from Chaitin’s and Boolos’ simpler proofs for the first incompleteness theorem. The problem is that these proofs only show the existence of a true statement (over ${\mathbb{N}}$) that has no proof, without giving an explicit example of such a statement.

A different proof for the second incompleteness theorem, based on Berry’s paradox, was given by Kikuchi [Kikuchi97]. This proof is model theoretic, and seems to us somewhat less intuitive for people who are less familiar with model theory. For previous model theoretic proofs for the second incompleteness theorem see [Kreisel50] (see also [Smoryński77]).

We give a new proof for the second incompleteness theorem, based on Chaitin’s incompleteness theorem and an argument that resembles the surprise examination paradox, (also known as the unexpected hanging paradox).

The surprise examination paradox: the teacher announces in class: “next week you are going to have an exam, but you will not be able to know on which day of the week the exam is held until that day”. The exam cannot be held on Friday, because otherwise, the night before the students will know that the exam is going to be held the next day. Hence, in the same way, the exam cannot be held on Thursday. In the same way, the exam cannot be held on any of the days of the week.

Let $T$ be a (rich enough) mathematical theory, such as PA or ZFC. For simplicity, the reader can assume that $T$ is ZFC, the theory of all mathematics; thus, any mathematical proof, and in particular any proof in this paper, is obtained within $T$.

Let $L$ be the integer guaranteed by Chaitin’s incompleteness theorem. Thus, for any integer $x$, the statement “$K(x)>L$” cannot be proved (in the theory $T$), unless the theory is inconsistent. Note, however, that for any integer $x$, such that, $K(x)\leq L$, there is a proof (in $T$) for the statement “$K(x)\leq L$”, simply by giving the computer program of length at most $L$ that outputs $x$ and stops, and by describing the running of that computer program until it stops.

Let $m$ be the number of integers $0\leq x\leq 2^{L+1}$, such that, $K(x)>L$. (The number $m$ is analogous to the day of the week on which the exam is held in the surprise examination paradox). Recall that since the number of computer programs of length $L$ bits is at most $2^{L+1}$, there exists at least one integer $0\leq x\leq 2^{L+1}$, such that, $K(x)>L$. Hence, $m\geq 1$.

Assume that $m=1$. Thus, there exists a single integer $x\in\{0,\ldots,2^{L+1}\}$ such that $K(x)>L$, and every other integer $y\in\{0,\ldots,2^{L+1}\}$ satisfies $K(y)\leq L$. In this case, one can prove that $x$ satisfies $K(x)>L$ by proving that every other integer $y\in\{0,\ldots,2^{L+1}\}$ satisfies $K(y)\leq L$ (and recall that there is a proof for every such statement). Since we proved that $m\geq 1$, the only $x$ for which we didn’t prove $K(x)\leq L$ must satisfy $K(x)>L$.

Thus, if $m=1$ then for some integer $x$, the statement “$K(x)>L$” can be proved (in $T$). But we know that for any integer $x$, the statement “$K(x)>L$” cannot be proved (in $T$), unless the theory is inconsistent. Hence, if the theory is consistent, $m\geq 2$. Since we assume that $T$ is a rich enough theory, we can prove the last conclusion in $T$. That is, we can prove in $T$ that: if $T$ is consistent then $m\geq 2$.

Assume for a contradiction that the consistency of $T$ can be proved within $T$. Thus, we can prove in $T$ the statement “$m\geq 2$”. In the same way, we can work our way up and prove that $m\geq i+1$, for every $i\leq 2^{L+1}+1$. In particular, $m>2^{L+1}+1$, which is a contradiction, since $m\leq 2^{L+1}+1$ (by the definition of $m$).

To present the proof formally, one needs to be able to express provability within $T$, in the language of $T$. The standard way of doing that is by assuming that the language of $T$ contains the language of arithmetics and by encoding every formula and every proof in $T$ by an integer, usually referred to as the G$\ddot{\mbox{o}}$del number of that formula or proof. For a formula $A$, let $\ulcorner A\urcorner$ be its G$\ddot{\mbox{o}}$del number. Let $\mbox{Pr}_{T}(\ulcorner A\urcorner)$ be the following formula: there exists $w$ that is the G$\ddot{\mbox{o}}$del number of a $T$-proof for the formula $A$. Intuitively, $\mbox{Pr}_{T}(\ulcorner A\urcorner)$ expresses the provability of the formula $A$. Formally, the formulas $\mbox{Pr}_{T}(\ulcorner A\urcorner)$ satisfy the so-called Hilbert-Bernays derivability conditions (see, for example, [Mendelson97]):

1. 1.

   If $T$ proves $A$ then $T$ proves $\mbox{Pr}_{T}(\ulcorner A\urcorner)$.

2. 2.

   $T$ proves: $\mbox{Pr}_{T}(\ulcorner A\urcorner)\rightarrow\mbox{Pr}_{T}(\ulcorner\mbox{Pr}_{T}(\ulcorner A\urcorner)\urcorner)$.

3. 3.

   $T$ proves: $\mbox{Pr}_{T}(\ulcorner A\rightarrow B\urcorner)\rightarrow\left(\mbox{Pr}_{T}(\ulcorner A\urcorner)\rightarrow\mbox{Pr}_{T}(\ulcorner B\urcorner)\right)$

The consistency of $T$ is usually expressed as the formula $\mbox{Con}(T)\equiv\neg\mbox{Pr}_{T}(\ulcorner 0=1\urcorner)$. In all that comes below, $T\vdash A$ denotes “$T$ proves $A$”. We will prove that $T\not\vdash\mbox{Con}(T)$, unless $T$ is inconsistent.

For our proof, we will need two facts about provability of claims concerning Kolmogorov complexity. First, we need to know that $\mbox{Con}(T)\rightarrow\neg\mbox{Pr}_{T}(\ulcorner K(x)>L\urcorner)$. We will use the following form of Chaitin’s incompleteness theorem (see, for example, [Kikuchi97], Theorem 3.3).

$\displaystyle T\vdash\mbox{Con}(T)\rightarrow\forall x\in\{0,\ldots,2^{L+1}\}\>\neg\mbox{Pr}_{T}(\ulcorner K(x)>L\urcorner)$

(1)

Second, we need to know that $(K(y)\leq L)\rightarrow\mbox{Pr}_{T}(\ulcorner K(y)\leq L\urcorner)$. We will use the following form (formally, this follows since $K(y)\leq L$ is a $\Sigma_{1}$ formula; see, for example, [Kikuchi97], Theorem 1.2 and Section 2).

$\displaystyle T\vdash\forall y\in\{0,\ldots,2^{L+1}\}\>\left((K(y)\leq L)\rightarrow\mbox{Pr}_{T}(\ulcorner K(y)\leq L\urcorner)\right)$

(2)

Assume for a contradiction that $T$ is consistent and $T\vdash\mbox{Con}(T)$. Then, by Equation 1,

$\displaystyle T\vdash\forall x\in\{0,\ldots,2^{L+1}\}\;\neg\mbox{Pr}_{T}(\ulcorner K(x)>L\urcorner)$

(3)

We will derive a contradiction by proving by induction that, for every $i\leq 2^{L+1}+1$, $T\vdash(m\geq i+1)$, where $m$ is defined as in the previous section. Since obviously $T\vdash(m\leq 2^{L+1}+1)$, this is a contradiction to the assumption that $T$ is consistent and $T\vdash\mbox{Con}(T)$. Since we already know that $T\vdash(m\geq 1)$, we already have the base case of the induction. Assume (the induction hypothesis) that for some $1\leq i\leq 2^{L+1}+1$,

$\displaystyle T\vdash(m\geq i)$

We will show that $T\vdash(m\geq i+1)$, as follows. Let $r=2^{L+1}+1-i$.

1. 1.

   By the definition of $m$,

   $T\vdash(m=i)\rightarrow\exists\;\mbox{different}\;y_{1},\ldots,y_{r}\in\{0,\ldots,2^{L+1}\}\;\bigwedge_{j=1}^{r}(K(y_{j})\leq L)$

2. 2.

   Hence, by Equation 2,

   $T\vdash(m=i)\rightarrow\exists\;\mbox{different}\;y_{1},\ldots,y_{r}\in\{0,\ldots,2^{L+1}\}\;\bigwedge_{j=1}^{r}\mbox{Pr}_{T}(\ulcorner K(y_{j})\leq L\urcorner)$

3. 3.

   For every $\mbox{different}\;y_{1},\ldots,y_{r}\in\{0,\ldots,2^{L+1}\}$, and every $x\in\{0,\ldots,2^{L+1}\}\setminus\{y_{1},\ldots,y_{r}\}$,

   $T\vdash(m\geq i)\rightarrow\left(\bigwedge_{j=1}^{r}(K(y_{j})\leq L)\rightarrow(K(x)>L)\right),$

   (by the definition of $m$), and hence by Hilbert-Bernays derivability conditions,

   $T\vdash\mbox{Pr}_{T}(\ulcorner m\geq i\urcorner)\rightarrow\left(\bigwedge_{j=1}^{r}\mbox{Pr}_{T}(\ulcorner K(y_{j})\leq L\urcorner)\rightarrow\mbox{Pr}_{T}(\ulcorner K(x)>L\urcorner)\right)$

4. 4.

   By the previous two items,

   $T\vdash((m=i)\wedge\mbox{Pr}_{T}(\ulcorner m\geq i\urcorner))\rightarrow\exists x\in\{0,\ldots,2^{L+1}\}\;\mbox{Pr}_{T}(\ulcorner K(x)>L\urcorner)$

5. 5.

   Since $T\vdash(m\geq i)$ (by the induction hypothesis), $T\vdash\mbox{Pr}_{T}(\ulcorner m\geq i\urcorner)$. Hence,

   $T\vdash(m=i)\rightarrow\exists x\in\{0,\ldots,2^{L+1}\}\;\mbox{Pr}_{T}(\ulcorner K(x)>L\urcorner)$

6. 6.

   Hence, by Equation 3,

   $T\vdash\neg(m=i)$

7. 7.

   Hence, since $T\vdash(m\geq i)$ ,

   $T\vdash(m\geq i+1)$

∎

In the previous sections we gave a proof for G$\ddot{\mbox{o}}$del’s second incompleteness theorem by an argument that resembles the surprise examination paradox. In this section we go the other way around and suggest that the second incompleteness theorem gives a possible resolution of the surprise examination paradox. Roughly speaking, we argue that the flaw in the derivation of the paradox is that it contains a hidden assumption that one can prove the consistency of the mathematical theory in which the derivation is done; which is impossible by the second incompleteness theorem.

The important step in analyzing the paradox is the translation of the teacher’s announcement into a mathematical language. The key point lies in the formalization of the notions of surprise and knowledge.

As before, let $T$ be a rich enough mathematical theory (say, ZFC). Let $\left\{1,\dots,5\right\}$ be the days of the week and let $m$ denote the day of the week on which the exam is held. Recall the teacher’s announcement: “next week you are going to have an exam, but you will not be able to know on which day of the week the exam is held until that day”. The first part of the announcement is formalized as $m\in\left\{1,\dots,5\right\}$. A standard way that appears in the literature to formalize the second part is by replacing the notion of knowledge by the notion of provability [Shaw58, Fitch64] (for a recent survey see [Chow98]). The second part is rephrased as “on the night before the exam you will not be able to prove, using this statement, that the exam is tomorrow”, or, equivalently, “for every $1\leq i\leq 5$, if you are able to prove, using this statement, that $(m\geq i)\rightarrow(m=i)$, then $m\neq i$”. This can be formalized as the following statement that we denote by $S$ (the statement $S$ contains both parts of the teacher’s announcement):

$$S\;\;\equiv\;\;[m\in\left\{1,\dots,5\right\}]\bigwedge_{1\leq i\leq 5}\left[\mbox{Pr}_{T,S}\left(\ulcorner m\geq i\to m=i\urcorner\right)\to(m\neq i)\right]$$

where $\mbox{Pr}_{T,S}(\ulcorner A\urcorner)$ expresses the provability of a formula $A$ from the formula $S$ in the theory $T$, (formally, $\mbox{Pr}_{T,S}(\ulcorner A\urcorner)$ is the formula: there exists $w$ that is the G$\ddot{\mbox{o}}$del number of a $T$-proof for the formula $A$ from the formula $S$). Note that the formula $S$ is self-referential. Nevertheless, it is well known that this is not a real problem and that such a formula $S$ can be formulated (see [Shaw58, Chow98]; for more about this issue, see below).

Let us try to analyze the paradox when the teacher’s announcement is formalized as the above statement $S$. We will start from the last day. The statement $m\geq 5$ together with $m\in\left\{1,\dots,5\right\}$ imply $m=5$. Hence, $\mbox{Pr}_{T,S}\left(\ulcorner m\geq 5\to m=5\urcorner\right)$ and by $S$ we can conclude $m\neq 5$. Thus, $S$ implies $m\in\left\{1,\dots,4\right\}$. In the same way, working our way down, we can prove $\mbox{Pr}_{T,S}\left(\ulcorner m\geq 4\to m=4\urcorner\right)$ and by $S$ we can conclude $m\neq 4$. In the same way, $m\neq 3$, $m\neq 2$, and $m\neq 1$. In other words, $S$ implies $m\notin\left\{1,\dots,5\right\}$. Thus, $S$ contradicts itself.

The fact that $S$ contradicts itself gives a certain explanation for the paradox; the teacher’s announcement is just a contradiction. On the other hand, we feel that this formulation doesn’t fully explain the paradox: Note that since $S$ is a contradiction it can be used to prove any statement. So, for example, on Tuesday night the students can use $S$ to prove that the exam will be held on Wednesday. Is it fair to say that this means that they know that the exam will be held on Wednesday? No, because they can also use $S$ to prove that the exam will be held on Thursday. Thus, we conclude that since $S$ is a contradiction, provability from $S$ doesn’t imply knowledge. Recall, however, that the very intuition behind the formalization of the teacher’s announcement as $S$ was that the notion of knowledge can be replaced by the notion of provability. But if provability from $S$ doesn’t imply knowledge, the statement $S$ doesn’t seem to be an accurate translation of the teacher’s announcement into a mathematical language.

Is there a better way to formalize the teacher’s announcement? To answer this question, let us analyze the situation from the students’ point of view on Tuesday night. There are three possibilities:

1. 1.

   On Tuesday night, the students are not able to prove that the exam will be held on Wednesday.

2. 2.

   On Tuesday night, the students are able to prove that the exam will be held on Wednesday, but they are also able to prove for some other day that the exam will be held on that day.
   (Note that this possibility can only occur if the system is inconsistent, and is in fact equivalent to the inconsistency of the system).

3. 3.

   On Tuesday night, the students are able to prove that the exam will be held on Wednesday, and they are not able to prove for any other day that the exam will be held on that day.

We feel that only in the third case is it fair to say that the students know that the exam will be held on Wednesday. They know that the exam will be held on Wednesday only if they are able to prove that the exam will be held on Wednesday, and they are not able to prove for any other day that the exam will be held on that day.

We hence rephrase the second part of the teacher’s announcement as “for every $1\leq i\leq 5$, if one can prove (using this statement) that $(m\geq i)\rightarrow(m=i)$, and there is no $j\neq i$ for which one can prove (using this statement) $(m\geq i)\rightarrow(m=j)$, then $m\neq i$”. Thus, the teacher’s announcement is the following statement³³3This statement is equivalent to one of the suggestions (the statement $I_{5}$) made by Halpern and Moses [HM86]. However, the analysis of the paradox there is different from the one shown here and makes no use of G$\ddot{\mbox{o}}$del’s second incompleteness theorem.:

$$S\;\;\equiv\;\;[m\in\left\{1,\dots,5\right\}]\bigwedge_{1\leq i\leq 5}\left[\left(\mbox{Pr}_{T,S}\left(\ulcorner m\geq i\to m=i\urcorner\right)\bigwedge_{1\leq j\leq 5,j\neq i}\neg\mbox{Pr}_{T,S}\left(\ulcorner m\geq i\to m=j\urcorner\right)\right)\to(m\neq i)\right]$$

Let us try to analyze the paradox when the teacher’s announcement is formalized as the new statement $S$. As before, $m\geq 5$ together with $m\in\left\{1,\dots,5\right\}$ imply $m=5$. Hence, $\mbox{Pr}_{T,S}\left(\ulcorner m\geq 5\to m=5\urcorner\right)$. However, this time one cannot use $S$ to conclude $m\neq 5$, since it is possible that for some $j\neq 5$ we also have $\mbox{Pr}_{T,S}\left(\ulcorner m\geq 5\to m=j\urcorner\right)$. This happens iff the system $T+S$ is inconsistent. Formally, this time one cannot use $S$ to deduce $m\neq 5$, but rather the formula

$$\mbox{Con}(T,S)\rightarrow(m\neq 5),$$

where $\mbox{Con}(T,S)\equiv\neg\mbox{Pr}_{T,S}(\ulcorner 0=1\urcorner)$ expresses the consistency of $T+S$. Since by the second incompleteness theorem one cannot prove $\mbox{Con}(T,S)$ within $T+S$, we cannot conclude that $S$ implies $m\neq 5$ and hence cannot continue the argument.

More precisely, since $S$ doesn’t imply $m\in\{1,\ldots,4\}$, but rather $\mbox{Con}(T,S)\rightarrow m\in\{1,\ldots,4\},$ when we try to work our way down we do not get the desired formula $\mbox{Pr}_{T,S}\left(\ulcorner m\geq 4\to m=4\urcorner\right)$, but rather the formula

$$\mbox{Pr}_{T,S}\left(\ulcorner\mbox{Con}(T,S)\wedge(m\geq 4)\to m=4\urcorner\right),$$

which is not enough to continue the argument.

Thus, our conclusion is that if the students believe in the consistency of $T+S$ the exam cannot be held on Friday, because on Thursday night the students will know that if $T+S$ is consistent the exam will be held on Friday. However, the exam can be held on any other day of the week because the students cannot prove the consistency of $T+S$.

Finally, for completeness, let us address the issue of the self-reference of the statement $S$. The issue of self-referentiality of a statement goes back to G$\ddot{\mbox{o}}$del’s original proof for the first incompleteness theorem. The self-reference is what makes G$\ddot{\mbox{o}}$del’s original proof conceptually difficult, and what makes the teacher’s announcement in the surprise examination paradox paradoxical.

To solve this issue, G$\ddot{\mbox{o}}$del introduced the technique of diagonalization. The same technique can be used here. To formalize $S$, we will use the notation $a\Rightarrow b$ to indicate implication between G$\ddot{\mbox{o}}$del numbers $a$ and $b$. That is, $a\Rightarrow b$ is a statement indicating that $a$ is a G$\ddot{\mbox{o}}$del number of a statement $A$, and $b$ is a G$\ddot{\mbox{o}}$del number of a statement $B$, such that, $A\rightarrow B$. We will also need the function $\mbox{Sub}(a,b)$ that represents substitution of $b$ in the formula with G$\ddot{\mbox{o}}$del number $a$. That is, if $a$ is a G$\ddot{\mbox{o}}$del number of a formula $A(x)$ with free variable $x$, and $b$ is a number, then $\mbox{Sub}(a,b)$ is the G$\ddot{\mbox{o}}$del number of the statement $A(b)$.

Let $v_{ij}\equiv\ulcorner m\geq i\to m=j\urcorner$. Denote by $Q(x)$ the formula

$$Q(x)\;\;\equiv\;\;[m\in\left\{1,\dots,5\right\}]\bigwedge_{1\leq i\leq 5}\left[\left(\mbox{Pr}_{T}\left(\mbox{Sub}(x,x)\Rightarrow v_{ii}\right)\bigwedge_{1\leq j\leq 5,j\neq i}\neg\mbox{Pr}_{T}\left(\mbox{Sub}(x,x)\Rightarrow v_{ij}\right)\right)\to(m\neq i)\right]$$

Let $q$ be the G$\ddot{\mbox{o}}$del number of the formula $Q(x)$. The statement $S$ is formalized as $S\equiv Q(q)$. To see that this statement is the one that we are interested in, denote by $s$ the G$\ddot{\mbox{o}}$del number of $S$ and note that $s=\mbox{Sub}(q,q)$. Thus,

$$S\;\;\equiv\;\;[m\in\left\{1,\dots,5\right\}]\bigwedge_{1\leq i\leq 5}\left[\left(\mbox{Pr}_{T}\left(s\Rightarrow v_{ii}\right)\bigwedge_{1\leq j\leq 5,j\neq i}\neg\mbox{Pr}_{T}\left(s\Rightarrow v_{ij}\right)\right)\to(m\neq i)\right]$$