Proving the Completeness of Propositional Logic

The completeness of a logic is a really nice property to establish. For a logic to be complete, it must be that every semantic entailment is also syntactically entailed. Said more simply, it must be that every truth in the language is provable. Gödel’s incompleteness theorems showed us that we cannot have such high hopes for mathematics in general, but we can still establish completeness for some simple logics, such as propositional and first order logic.

I want to post a proof of the completeness of propositional logic here in full for future reference. Roughly the first half of what’s below is just establishing some necessary background, so that this post is fairly self-contained and doesn’t reference lemmas that are proved elsewhere.

The only note I’ll make before diving in is that the notation I(A,P) is a way to denote the smallest set that contains A and is closed under the operations in P. It’s a handy way to inductively define sets that would be enormously complicated to define otherwise. With that out of the way, here we go!

First we define the proof system for propositional logic.

Axiom 1: α→(β→α)
Axiom 2: (α→(β→γ)) → ((α→β)→(α→γ))
Axiom 3: ((¬α)→(¬β)) → (β→α)

The α, β, and γ symbols in these axioms are meant to stand for any well-formed formula. What this means is that we actually have a countable infinity of axioms that fall into the three categories above. For simplicity, I’ll keep calling them “Axioms 1, 2, and 3”, and assume you don’t find it too confusing.

You might also notice that the axioms only involve the symbols → and ¬, but neglect ∧ and ∨. This is okay because → and ¬ are adequate connectives for the semantics of propositional logic (which is to say that any truth function can be expressed in terms of them).

Axioms = {Axiom 1, Axiom 2, Axiom 3}
Deduction rule = Modus Ponens (MP)
….. MP(α, α→β) = β

The set of all provable sentences is just the set of all sentences that includes the axioms and is closed under modus ponens.

Theorems = I(Axioms, MP)

We can also easily talk about the set of sentences that can be proven from assumptions in a set Σ:

Th(Σ) = I(Axioms ∪ Σ, MP)
Notation: Σ ⊢ α iff α ∈ Th(Σ)

With that out of the way, let’s establish some basic but important results about the propositional proof system.

Monotonicity: If Σ ⊆ Σ’, then Th(Σ) ⊆ Th(Σ’).

Strong monotonicity: If Σ ⊢ Σ’, then Th(Σ’) ⊆ Th(Σ).

Intuitively, monotonicity says that if you expand the set of assumptions, you never shrink the set of theorems. Strong monotonicity says that if Σ can prove everything in Σ’, then Σ’ cannot be stronger than Σ. Both of these follow pretty directly from the definition of Th(Σ).

Soundness: If ⊢ α, then ⊨ α.
Proof by structural induction
….. Each axiom is a tautology.
….. Tautology is closed under MP (if ⊨ α and ⊨ (α→β), then ⊨ β).

Extended Soundness: If Σ ⊢ α, then Σ ⊨ α.
Proof by structural induction
….. Σ ⊨ α ∈ Axioms and Σ ⊨ α ∈ Σ.
….. If Σ ⊨ α and Σ ⊨ (α→β), then Σ ⊨ β.

Law of identity: ⊢ (α→α)
….. α→((α→α)→α), Axiom 1
….. (α→((α→α)→α))→((α→(α→α))→(α→α)), Axiom 2
….. (α→(α→α))→(α→α), MP
….. α→(α→α), Axiom 1
….. α→α, MP

Principle of Explosion: If Σ ⊢ α and Σ ⊢ (¬α), then Σ ⊢ β.
….. By strong monotonicity, it suffices to show that Σ ∪ {α} ∪ {¬α} ⊢ β.
……….. (¬α)→((¬β)→(¬α)), Axiom 1
……….. ¬α, Assumption
……….. (¬β)→(¬α), MP
……….. ((¬β)→(¬α))→(α→β), Axiom 3
.………. α→β, MP
……….. α, Assumption
……….. β, MP

And finally, our most important background theorem:

Deduction Theorem: Σ ⊢ (α→β) iff Σ ∪ {α} ⊢ β.
Proof =>
….. Suppose Σ ⊢ (α→β).
….. By monotonicity, Σ ∪ {α} ⊢ (α→β).
….. Also, clearly Σ ∪ {α} ⊢ α.
….. So Σ ∪ {α} ⊢ β.
Proof <=
….. Suppose Σ ∪ {α} ⊢ β.
….. Base cases
………. β ∈ Axioms. (β, β→(α→β), α→β).
………. β ∈ Σ. (β, β→(α→β), α→β).
………. β = α. (⊢ (α→α), so by monotonicity Σ ⊢ (α→α)).
….. Inductive step
………. Suppose Σ ⊢ (α→γ) and Σ ⊢ (α→(γ→𝛿)).
………. By strong monotonicity, suffices to show Σ ∪ {α→γ, α→(γ→𝛿)} ⊢ (α→𝛿)
……………. (α→(γ→𝛿)) → ((α→γ)→(α→𝛿)), Axiom 2
……………. α→(γ→𝛿), Assumption
……………. (α→γ)→(α→𝛿), MP
……………. (α→γ), Assumption
……………. (α→𝛿). MP

Now, let’s go into the main body of the proof. The structure of the proof is actually quite similar to the proof of the compactness theorem I gave previously. First we show that every consistent set of sentences Σ has a maximally consistent extension Σ’. Then show that Σ’ is satisfiable. Now since Σ’ is satisfiable and it’s an extension of Σ, Σ must also be satisfiable. From there it’s a simple matter to show that the logic is complete.

So, let’s define some of the terms I just used.

Σ is consistent iff for no α does Σ ⊢ α and Σ ⊢ (¬α)
….. Equivalently: iff for some α, Σ ⊬ α

Σ is maximally consistent iff Σ is consistent and for every α, either Σ ∪ {α} is inconsistent or Σ ⊢ α.

One final preliminary result regarding consistency before diving into the main section of the proof:

If Σ is satisfiable, then Σ is consistent.
….. Suppose Σ is inconsistent.
….. Then there’s an α such that Σ ⊢ α and Σ ⊢ (¬α).
….. By extended soundness, Σ ⊨ α and Σ ⊨ (¬α).
….. So Σ is not satisfiable.

This is the converse of the result we actually want, but it’ll come in handy. Now, let’s begin to construct our extension!

Any consistent Σ can be extended to a maximally consistent Σ’
….. Choose any ordering {αn} of well-formed-formulas.
….. Define Σ0 = Σ.
….. Σn+1 = Σn if Σn ⊢ (¬αn+1), and Σn ∪ {αn+1} otherwise.
….. For each n, (i) Σn is consistent and (ii) either Σn ⊢ αn or Σn ⊢ (¬αn)
……….. Base case: Σ0 is consistent by assumption, and (ii) doesn’t apply.
……….. Inductive step: Suppose Σn satisfies (i) and (ii). Two cases:
…………….. If Σn ⊢ (¬αn+1), then Σn+1 = Σn. Clearly consistent and satisfies (ii).
…………….. If Σn ⊬ (¬αn+1), then Σn+1 = Σn ∪ {αn+1}. Clearly satisfies (ii), but is it consistent?
………………….. Suppose not. Then Σn+1 ⊢ (¬αn+1), by explosion.
………………….. So Σn+1 ∪ {αn+1} ⊢ (¬αn+1).
………………….. So Σn+1 ⊢ (αn+1 → (¬αn+1)).
………………….. ⊢ ((α→¬α)→¬α), so Σn+1 ⊢ (¬αn+1). Contradiction!

….. Define Σ’ = ∪ Σn. Σ’ is maximally consistent.
……….. Maximality
…………….. Suppose not. Then for some αn, Σ’ ⊬ αn and Σ’ ∪ {αn} is consistent.
…………….. But Σn ⊆ Σ’, and either Σn ⊢ αn or Σn ⊢ (¬αn).
…………….. If Σn ⊢ αn, by monotonicity Σ ⊢ αn. Contradiction. So Σn ⊢ (¬αn).
…………….. By monotonicity, Σ ⊢ (¬αn), so Σ ∪ {αn} ⊢ (¬αn).
…………….. But Σ ∪ {αn} ⊢ αn. So Σ ∪ {αn} is inconsistent. Contradiction!
……….. Consistency
…………….. Suppose Σ’ is inconsistent. Then for some α, Σ’ ⊢ α and Σ’ ⊢ (¬α).
…………….. So there are proofs of α and (¬α) from Σ’.
…………….. Proofs are finite, so each proof uses only a finite number of assumptions from Σ’.
…………….. So we can choose an n such that Σn contains all the needed assumptions.
…………….. Now both proofs from Σ’ are also proofs from Σn.
…………….. So Σn ⊢ αn and Σ’ ⊢ (¬αn).
…………….. So Σn is inconsistent. Contradiction!

Alright, we’re almost there! So now we have that for any consistent Σ, there’s an extension Σ’ that is maximally consistent. We’ll take it a little further and prove that not only is Σ’ maximally consistent, it’s also complete! (This is the purely syntactic sense of completeness, which is that for every sentence α, either Σ’ proves α or refutes α. This is different from the sense of logical completeness that we’re establishing with the proof.)

Σ’ is complete.
….. Σn ⊆ Σ’, and Σn ⊢ αn or Σn ⊢ (¬αn).
….. So by monotonicity Σ’ ⊢ αn or Σ’ ⊢ (¬αn).

Now we have everything we need to show that Σ’, and thus Σ, is satisfiable.

If Σ is consistent, then Σ is satisfiable.
….. Let Σ’ be a maximally consistent extension of Σ.
….. Define vΣ’(p) over propositional variables p:
….. VΣ’(p) = T if Σ’ ⊢ p and F if Σ’ ⊬ p
….. ṼΣ’(α) = T iff Σ’ ⊢ α
……….. Base case: Let α be a propositional variable. Then ṼΣ’(α) = T iff Σ’ ⊢ α by definition of VΣ’.
……….. Inductive steps:
……….. (¬α)
…………….. If Σ’ ⊢ (¬α), then by consistency Σ’ ⊬ α, so ṼΣ’(α) = F, so ṼΣ’(¬α) = T.
…………….. If Σ’ ⊬ (¬α), then by completeness Σ’ ⊢ α. So ṼΣ’(α) = T, so ṼΣ’(¬α) = F.
……….. (α→β)
…………….. Suppose Σ’ ⊢ (α→β). By completeness Σ’ ⊢ α or Σ’ ⊢ (¬α).
………………….. If Σ’ ⊢ α, then Σ’ ⊢ β, so ṼΣ’(β) = T, so ṼΣ’(α→β) = T.
………………….. If Σ’ ⊢ (¬α), then ṼΣ’(α) = F, so ṼΣ’(α→β) = T.
……………. Suppose Σ’ ⊬ (α→β).
………………….. By completeness Σ’ ⊢ ¬(α→β).
………………….. ⊢ (β→(α→β)), so Σ’ ⊬ β on pain of contradiction. So ṼΣ'(β) = F.
………………….. Suppose ṼΣ’(α→β) = T. Then ṼΣ’(α) = F, so Σ’ ⊢ (¬α).
………………….. ⊢ (¬α→(α→β)). So Σ’ ⊢ (α→β). Contradiction.
………………….. So ṼΣ’(α→β) = F.
….. So vΣ’ satisfies Σ’.
….. Σ ⊆ Σ’, so vΣ’ satisfies Σ.
….. So Σ is satisfiable!

Now our final result becomes a four-line proof.

If Σ ⊨ α, then Σ ⊢ α.
….. Suppose Σ ⊬ α.
….. Then Σ ∪ {¬α} is consistent.
….. So Σ ∪ {¬α} is satisfiable.
….. So Σ ⊭ α.

And we’re done! We’ve shown that if any sentence α is semantically entailed by a set of sentences Σ, then it must also be provable from Σ! If you’ve followed this proof all the way, pat yourself on the back.

With the Completeness Theorem in hand, the proof of the Compactness Theorem goes from several pages to a few lines. It’s so nice and simple that I just have to include it here.

If Σ is finitely satisfiable, then Σ is satisfiable.

….. Suppose Σ is not satisfiable.
….. Then Σ is not consistent.
….. So there is some α for which Σ ⊢ α and Σ ⊢ (¬α).
….. Since proofs are finite, there must be some finite subset Σ* of Σ such that Σ* ⊢ α and Σ* ⊢ (¬α).
….. By soundness, Σ* ⊨ α and Σ* ⊨ (¬α).
….. So Σ* is not satisfiable!

In other words, if Σ is not satisfiable, then there’s some finite subset of Σ that’s also not satisfiable. This is the Compactness Theorem! Previously we proved it entirely based off of the semantics of propositional logic, but now we can see that it is also provable as a consequence of the finite nature of our proof system!

Sum and Product Puzzle

X and Y are two integers.

X < Y
X > 1
Y > 1
X + Y < 100

S and P are two perfect logicians. S knows X + Y, and P knows X × Y.

Everything I’ve just said is common knowledge. S and P have the following conversation:

S: “P, you don’t know X and Y”
P: “Now I do know X and Y!”
S: “And now so do I!”

What are X and Y?

Once you figure out that, here’s a question: If instead of saying that X + Y < 100, we say X + Y < N, then what’s the range of values of N for which this puzzle has a unique solution?

Four Pre-Gödelian Limitations on Mathematics

Even prior to the devastating Incompleteness Theorems there were hints of what was to come. I want to describe and prove four results in mathematical logic that don’t depend on Incompleteness at all, but establish some rather serious limitations on the project of mathematics.

Here are the four. I’ll go through them in order of increasing level of sophistication required to prove them.

  • Indescribable sets of possible worlds
  • Noncompossibility theorem
  • Inevitable nonstandard numbers
  • Mysterious missing subsets

1. Indescribable sets of possible worlds

I already talked about this one here. The basic idea is that even in our safest and least troublesome logic, propositional calculus, it turns out that the language is insufficient to fully capture all the semantic notions. It’s the first hint at something going awry with syntax and semantics, where the semantics can outpace the syntax and leave axiomatic mathematics behind.

So to recap: the result is that in propositional logic there are sets of truth assignments that can not be “described” by any set of propositional sentences (even allowing infinite sets!). A set of sentences is said to “describe” a set of truth assignments if that set of truth assignments is the unique set of truth assignments consistent with all those sentences being true. If we think of truth assignments as possible worlds, and sets of sentences as descriptions of sets of possible worlds, then this result says that there are sets of possible worlds in propositional semantics that cannot be described by any propositional syntax.

The proof of this is astoundingly simple: just look at the cardinality of the set of descriptions and the cardinality of the set of sets of possible worlds. The second is strictly larger than the first, so any mapping from descriptions to sets of possible worlds will of necessity leave some sets of possible worlds out. In fact, it also tells us that virtually all sets of possible worlds are not describable!

2. Noncompossibility Theorem

The noncompossibility theorem is a little-known theorem that establishes a serious limitation on our ability to describe mathematical structures. Here’s what it says. Suppose that you have a description of a countably infinite structure (like, say, the natural numbers, which have a countable infinity of objects) that has the following three properties:

(1) The language has a term for denoting every object in the structure (like 0, 1, 2, 3, 4, and so on)
(2) The axioms in your description are weakly complete: if something is inconsistent with the axioms, it can be proven false.
(3) There is some algorithm for determining whether any given sentence is an axiom.

The noncompossibility theorem tells us that if you have all three of these properties, then your axioms will fail to uniquely pick out your intended structure, and will include models that have extra objects that aren’t in the structure.

Let’s prove this.

We’ll denote the mathematical structure that we’re trying to describe as M and our language as L. We choose L to have sufficient syntactic structure to express the truths of M. From L, we select a decidable set of sentences X with the goal that all these sentences be true of M. We now select a proof system F in L such that for any finite extension L* to L involving only new constant terms, and for any Y ⊆ L*, if X ∪ Y is not satisfiable, then F refutes some finite subset of X ∪ Y.

(As an aside: Why care about this strange weak form of completeness? Well, intuitively all that it’s saying is that our axioms should be able to rule out any set of sentences that are inconsistent with them using some finite proof, as long as those sentences only use finitely many additional constant symbols. This is relatively weak compared to the usual notions of completeness that logicians talk about, which makes it an even better choice for our purposes, as the weaker the axiom the harder to deny.)

Our assumptions can now be written:

(0) |M| is countably infinite.
(1) ∀m ∈ M, ∃tm ∈ terms(X) such that (m = tm)
(2) If we extend L to L* by adding finitely many constant symbols, then for any Y ⊆ L*, if X ∪ Y is not satisfiable, then F refutes some finite subset of X ∪ Y.
(3) X is recursively enumerable.

Our proof starts by adding a new constant term c to our language and constructing an extension of X:

Y = X ∪ {c ≠ tm | m ∈ M}

In other words, Y is X but supplemented with the assertion that there exists an object that isn’t in M. If we can prove that Y is satisfiable, then this entails that X is also satisfiable by the same truth assignment. And this means that there is a model of X in which there are extra objects that aren’t in M.

We proceed with proof by contradiction. Suppose that Y is not satisfiable. Then, by assumption (2), we must be able to refute some finite subset Z of X ∪ Y. But since Z is finite, it involves only finitely many terms tm. And since M is countably infinite, there will always be objects in M that are not equal to any of the chosen terms! So we can’t refute any finite subset of Z! Thus Y is satisfiable.

And if Y is satisfiable, then so must be X, as Y is a superset of X. And since Y is satisfiable, then there’s some truth assignment v that satisfies all of v. But then v also satisfies X, as X is a subset of Y and removing axioms cannot rule out models, only add more! So we’ve proven that X has a model in which there is an object that is not equal to any of the objects in M. That is, X is not categorical: it does not uniquely describe M.

Tennant described this theorem as saying that “in countably infinite realms, you cannot know both where you are and where you are going.” More dully, we cannot have a satisfactory theory of a countably infinite mathematical structure that is both categorical and weakly complete. This isn’t super shocking by today’s standards, but it’s quite cool when you consider how little elaborate theoretical apparatus is required to prove it.

3. Inevitable nonstandard numbers

Suppose we have some first-order theory T that models the natural numbers. Take this theory and append to it a new constant symbol c, as well as an infinite axiom schema saying “c > 0” , “c > 1” , “c > 2”, and so on forever. Call this new theory T*.

Does T* have a model? Well by the compactness theorem, it has a model as long as all its finite subsets have a model. And for every finite subset of T*s axioms, the natural numbers are a model! So T* does have a model. Could this model be the natural numbers? Clearly not, because to satisfy T*, there must be a number greater than all the natural numbers. So whatever the model of T* is, it’s not the standard natural numbers. Let’s call it a nonstandard model, and label it ℕ*.

Here’s the final step of the proof: ℕ* is a model of T*, and T* is a superset of T, so ℕ* must also be a model of T! And thus we find that in any logic with a compactness theorem, a theory of the natural numbers will have models with nonstandard numbers that are greater than all of ℕ.

It’s one of my favorite proofs, because it’s so easy to describe and has such a devastating conclusion. It’s also an example of the compactness theorem using the existence of one type of model (ℕ for each of the finite cases) to prove the existence of something entirely different (ℕ* for the infinite case).

4. Mysterious missing subsets

The Löwenheim-Skolem theorem tells us that if a first-order theory has a model with an infinite cardinality, then it has models with every infinite cardinality. This places a major restriction on our ability to describe an infinite mathematical structure using first order logic. For if we were to try to single out the natural numbers, say, we would inevitably end up failing to rule out models of our axioms that are the cardinality of the real numbers, or worse, or the set of functions from real numbers to real numbers, and so on for all other possible cardinalities.

When applied to set theory, this implies a result that seems on its face to be a straightforward contradiction. Namely, Löwenheim-Skolem tells us that any first order axiomatization of sets will inevitably have a model that contains only a countable infinity of sets. But this seems bizarre, as all we appear to need to rule out countably infinite universes of sets is one axiom that asserts the existence of a countably infinite set, and another that asserts that admits the power-set of any set to the universe of sets. Then we will be forced to admit that there is a set which is the power set of a countably infinite set, and as Cantor’s famous diagonal argument shows, that this set is uncountably large.

So on the one hand, Cantor tells us that there are sets that contain uncountably many objects. And on the other hand, Löwenheim-Skolem tell us that there is a model of set theory with only countably many objects. This dichotomy is known by the name Skolem’s paradox. It appears to be a straightforward contradiction, but it’s not.

What Skolem realized was that the formal notion of a power set, which is something like “the set P(X) such that for all sets Y, if Y is a subset of X, then Y is an element P(X)”, relies on a quantification over all sets, and that in a countable universe of sets, that quantification ranges over only a countable number of objects. In other words, P(X) is only uncountable if our quantifier ranges over all possible sets, but for a countable model, there are sets that are not describable within the model. This means that the notion of a power set is relative to your model of set theory! In fact, there’s no way in first order logic to unambiguously pin down what you mean by “power set” in such a way that all models will agree on what P(X) actually contains. It also means that the notions of cardinality and countability are relative to your model! In Skolem’s words, “even the notions ‘finite’, ‘infinite’, ’simply infinite sequence’ and so forth turn out to be merely relative within axiomatic set theory.”

A proof of the Compactness Theorem

The Compactness Theorem is one of the most powerful results in mathematical logic, and I want to prove it for you. Fair warning, the proof is not a walk in the park, and you probably won’t comprehend it if you speed through. Take your time! This is the easiest proof of Compactness that I know of which doesn’t rely on the Completeness Theorem, which is itself not easy to prove.

First, some background concepts:

Interpretation An interpretation of a set of sentences Σ is a consistent assignment of truth values to all well-formed-formulas such that every sentence in Σ is assigned true.

Logical Entailment Σ logically entails α (Σ ⊧ α) iff in every interpretation v of Σ, v(α) = True.

Completeness A set of sentences Σ is complete iff for every well-formed-formula α, Σ ⊧ α or Σ ⊧ ¬α.

Satisfiability A set of sentences Σ is satisfiable iff there is at least one interpretation under which all sentences in Σ are true.

Finite Satisfiability A set of sentences Σ is finitely satisfiable iff every finite subset of Σ is satisfiable.

Compactness Theorem Σ is satisfiable iff every finite subset of Σ is satisfiable.

With that aside, here’s an outline of the proof. I’ve adapted the outline from this lecture series, and added in some minor improvements.

There are five steps, and most of the work is in the middle three.

  1. Establish that there is an ordering on the set of well-formed-formulas.
  2. Prove that if Σ is finitely satisfiable, then so is either Σ ∪ {α} or Σ ∪ {¬α}.
  3. Construct Δ, an extension of Σ, which is complete and finitely satisfiable.
  4. Prove that Δ is satisfiable.
  5. Prove that Σ is satisfiable.

This will prove that Σ is satisfiable if it’s finitely satisfiable. The other direction (if it’s satisfiable then it’s finitely satisfiable) is trivial, as any subset of a satisfiable set is also satisfiable.

So let’s dive in!

1. There is an ordering on the set of well-formed-formulas.

This is a fun fact that is basically established by Gödel numbering. Assign numbers to all the symbols of your language. An example assignment for propositional logic:

“(” assigned 0
“)” assigned 1
“¬” assigned 2
“∧” assigned 3
“∨” assigned 4
“→” assigned 5
“P0” assigned 6
“P1” assigned 7
“P2” assigned 8
And so on…

Now we translate a string by going through symbol by symbol and making the number assigned to the nth symbol in the string the exponent of the nth prime. Here are a few translations to illustrate:

Initial string: “(¬P0)”
Codes for each character: [0, 2, 6, 1]
Gödel number: 20 32 56 71 = 984,375

Initial string: “(P0 → P2)”
Codes for each character: [0, 6, 5, 8, 1]
Gödel number: 20 36 55 78 111 = 144,462,310,059,375

Initial string: “((P0 ∨ P2) → (¬P1))”
Codes for each character: [0, 0, 6, 4, 8, 1, 5, 0, 2, 7, 1]
Gödel number: 20 30 56 74 118 131 175 190 232 297 311 ≈ 4.2 × 1037

Since the prime factorization of a number is unique, we have a unique number for every possible string, which means we can go backwards from any number to its corresponding string. Now to construct our order, we just start from 2 and work our way up, discarding any sentences we get that are not well-formed-formulas. This will give us an ordered list of all well-formed formulas!

2. If Σ is finitely satisfiable, then so is either Σ ∪ {α} or Σ ∪ {¬α}.

Suppose not. Then neither Σ ∪ {α} nor Σ ∪ {¬α} are finitely satisfiable. So there must exist some finite sets Σ0 and Σ1 such that Σ0 ⊆ Σ ∪ {α} and Σ1 ⊆ Σ ∪ {¬α}, where neither Σ0 and Σ1 are satisfiable.

But now consider Σ’ = (Σ0 ∪ Σ1) \ {α, ¬α}. Σ’ is a finite subset of Σ, so it must be satisfiable (remember, Σ is finitely satisfiable by assumption). This means that there is some interpretation which satisfies Σ’. But this interpretation must assign to α either T or F.

If it assigns T to α, then Σ’ ∪ {α} is satisfiable. But Σ’ ∪ {α} is a superset of Σ0! So Σ0 must also be satisfiable. But this contradicts our assumption!

But if it assigns F to α, then Σ’ ∪ {¬α} is satisfiable. And Σ’ ∪ {¬α} is a superset of Σ1! So Σ1 must also be satisfiable. Contradiction!

Either way we get a contradiction, so this proves our theorem.

3. Extend Σ to Δ, which is complete and finitely satisfiable.

We inductively define Δ as follows:

Δ0 = Σ
Δn+1 = Δn ∪ {αn} if Δn ∪ {αn} is finitely satisfiable, otherwise Δn ∪ {¬αn}
Δ = ∪ Δn

Here we’re using our ordering that we constructed in Step 1 to go through every well-formed-formula in order. This establishes that Δ is complete, as we construct it by appending to Σ either α or ¬α for every well-formed-formula. I.e. we construct it by taking Σ and adding to it an explicit “opinion” on every possible sentence, ensuring at each stage that we remain finitely satisfiable.

How do we know that the final result Δ is also finitely satisfiable? This is due to our proof in Step 2, which allows us to say that Δn is always finitely satisfiable for any n. But any finite subset of Δ is also a subset of Δn for some n! So any finite subset of Δ is satisfiable. Thus Δ is finitely satisfiable.

4. Δ is satisfiable.

To prove this, we’ll hand-make a truth assignment for our purposes. First we define a truth assignment v over atomic propositions as follows:

v(p) = T if p ∈ Δ, otherwise F

Now we extend v to the rest of the set of well-formed-formulas in the usual way, ensuring consistency of the truth-assignment.

We’ll prove now that under this extension, v(α) = T if and only if α ∈ Δ. The proof is by structural induction.

Base case

  • α is an atomic proposition. Then v(α) = T iff α ∈ Δ, by definition of v.

Inductive step(s)

  • If α = ¬β, then v(α) = v(¬β) = T iff ¬β ∈ Δ iff α ∈ Δ.
  • If α = (β ∧ γ), then v(α) = v(β ∧ γ) = T iff (v(β) = T and v(γ) = T) iff (β ∈ Δ and γ ∈ Δ) iff (β ∧ γ) ∈ Δ iff α ∈ Δ
    • If β ∈ Δ and γ ∈ Δ, then (β ∧ γ) ∈ Δ, since Δ is complete and finitely satisfiable.
      • If (β ∧ γ) ∉ Δ, then ¬(β ∧ γ) ∈ Δ by completeness. But then {β, γ, ¬(β ∧ γ)} is a finite subset of Δ that’s not satisfiable, so Δ isn’t finitely satisfiable. Proof by contradiction.
    • If β ∉ Δ or γ ∉ Δ, then (β ∧ γ) ∉ Δ, since Δ is complete and finitely satisfiable.
      • Suppose (β ∧ γ) ∈ Δ. Since β ∉ Δ or γ ∉ Δ, at least one of ¬β and ¬γ must be in Δ. Without loss of generality, assume ¬β ∈ Δ. Then {¬β, (β ∧ γ)} is a finite subset of Δ that’s not satisfiable, so Δ isn’t finitely satisfiable. Proof by contradiction.

Since ¬ and ∧ are a complete set of truth functions (all other well-formed formulas can be converted into a form that uses only ¬ and ∧), this proves the proposition for all well-formed formulas.

We’ve shown that v(α) = T if and only if α ∈ Δ. This proves that v satisfies Δ!

5. Σ is satisfiable.

The final step is the easiest one. Δ is satisfied by v, and Δ is a superset of Σ, so Σ must also be satisfied by v. So we’re done!

A challenge to constructivists

Constructive mathematicians do not accept a proof of existence unless it provides a recipe for how to construct the thing whose existence is being asserted. Constructive mathematics is quite interesting, but it also appears to have some big problems. Here’s a challenge for constructivists:

Suppose that I hand you some complicated function f from a set A to another set B. I ask you: “Can every element in B be reached by applying the function to an element in A?” In other words, is f surjective?

Now, it so happens that the cardinality of B is greater than the cardinality of A. That’s sufficient to tell us that f can’t be surjective, as however it maps elements there will always be some left over. So we know that the answer is “no, we can’t reach every element in B.” But we proved this without explicitly constructing the particular element in B that can’t be reached! So a constructivist will be left unsatisfied.

The trick is that I’ve made this function extremely complicated, so that there’s no clever way for them to point to exactly which element is missing. Would they say that even though |B| is strictly larger than |A|, it could still be somehow that every element in B is in the image of f? Imagine asking them to bet on this proposition. I don’t think any sane person would put any money on the proposition that f is onto.

And as a final kicker, our sets don’t even have to be infinite! Let |A| = 20 and |B| = 21. I describe a function from A to B, such that actually computing the “missing element” involves having to calculate the 21st Busy Beaver number or something. And the constructivist gets busy searching for the particular element in B that doesn’t get mapped to, instead of just saying “well of course we can’t map 20 elements to 21 elements!”

Even simpler, let F map {1} to {1,2} as follows: F(1) = 1 if the last digit of the 20th busy beaver number is 0, 1, 2, 3, or 4, and F(1) = 2 otherwise. Now to prove constructively that there is an element in {1, 2} that isn’t in the image of F requires knowing the last digit of the 20th busy beaver number, which humans will most likely never be able to calculate (we’re stuck on the fifth one now). So a constructivist will be remain uncertain on the question of if F is surjective.

But a sane person would just say “look, of course F isn’t surjective; it maps one object to two objects. You can’t do this without leaving something out! It doesn’t matter if we don’t know which element is left out, it has to be one of them!”

And if humanity is about to meet an alien civilization with immense computational power that knows all the digits of the 20th busy beaver number, the standard mathematician could bet their entire life savings on F not being surjective at any odds whatsoever, and the constructive mathematician would bet in favor of F being surjective at some odds. And of course, the constructivist would be wrong and lose money! So this also means that you have a way to make money off of any constructivist mathematicians you encounter, so long as we’re about to make contact with advanced aliens.

Describing the world

Wittgenstein starts his Tractatus Philosophicus with the following two sentences.

1. The world is everything that is the case.

1.1 The world is the totality of facts, not of things.

Let’s take him up on this suggestion and see how far we get. In the process, we’ll discover some deep connections to theorems in mathematical logic, as well as some fascinating limitations on the expressive powers of propositional and first order logic.

We start out with a set of atomic propositions. For a very simple world, we might only need a finite number of these: “Particle 1 out of 3 has property 1 out of 50”, “Particle 2 of 3 has property 17 out of 50”, and so on. More realistically, the set of atomic propositions will be infinite (countable if the universe doesn’t have any continuous properties, and uncountable otherwise).

For simplicity, we’ll imagine labeling our set of atomic propositions P1, P2, P3, and so on (even though this entails that there are at most countably many, nothing important will rest on this assumption.) We combine these atomic propositions with the operators of propositional logic {(, ), ¬, ∧, ∨, →}. This allows us to build up more complicated propositions, like ((P7∧P2)→(¬P13)). This will be the language that we use to describe the world.

Now, the way that the world is is just a consistent assignment of truth values to the set of all grammatical sentences in our language. For example, one simple assignment of truth values is the one that assigns “True” to all atomic propositions. Once we’ve assigned truth values to all the atomic propositions, we get the truth values for the rest of the set of grammatical sentences for free, by the constraint that our truth assignment be consistent. (For instance, if P1 and P2 are both true, then (P1∧P2) must also be true.)

Alright, so the set of ways the world could be corresponds to the set of truth assignments over our atomic propositions. The final ingredient is the notion that we can encode our present knowledge of the world as a set of sentences. Maybe we know by observation that P5 is true, and either P2 or P3 is true but not both. Then to represent this state of knowledge, we can write the following set of sentences:

{P5, (P2∨P3), ¬(P2∧P3)}

Any set of sentences picks out a set of ways the world could be, such that each of these possible worlds is compatible with that knowledge. If you know nothing at all, then the set of sentences representing your knowledge will be the empty set {}, and the set of possible worlds compatible with your knowledge will be the set of all possible worlds (all possible truth assignments). On the other extreme, you might know the truth values of every atomic proposition, in which case your state of knowledge uniquely picks out one possible world.

In general, as you add more sentences to your knowledge-set, you cut out more and more possible worlds. But this is not always true! Ask yourself what the set of possible worlds corresponding to the set {(P1∨¬P1), (P2∨¬P2), (P3∨¬P3)} is. Since each of these sentences is a tautology, no possible worlds are eliminated! So our set of possible worlds is still the set of all worlds.

Now we get to an interesting question: clearly for any knowledge-set of sentences, you can express a set of possible worlds consistent with that knowledge set. But is it the case that for any set of possible worlds, you can find a knowledge-set that uniquely picks it out? If I hand you a set of truth assignment functions and ask you to tell me a set of propositions which are consistent with that set of worlds and ONLY that set, is that always possible? Essentially, what we’re asking is if all sets of possible worlds are describable.

We’ve arrived at the main point of this essay. Take a minute to ponder this and think about whether it’s possible, and why/why not! For clarification, each sentence can only be finitely long. But! You’re allowed to include an infinity of sentences.


(Spoiler-hiding space…)


If there were only a finite number of atomic propositions, then you could pick out any set of possible worlds with just a single sentence in conjunctive normal form. But when we start talking about an infinity of atomic propositions, it turns out that it is not always possible! There are sets of possible worlds that are literally not describable, even though our language includes the capacity to describe each of those words and we’re allowed to include an infinite set of sentences.

There’s a super simple proof of this. Let’s give a name to the cardinality of the set of sentences: call it K. (We’ve been tacitly acting as if the cardinality is countable this whole time, but that doesn’t actually matter.) What’s the cardinality of the set of all truth assignments?

Well, each truth assignment is a function from all sentences to {True, False}. And there are 2K such assignments. 2K is strictly larger than K, so there are more possible worlds than there are sentences. Now, the cardinality of the set of sets of sentences is also 2K. But the set of SETS of truth of assignments is 22^K!

What this means is that we can’t map sets of sentences onto sets of truth assignments without leaving some things out! This proof carries over to predicate logic as well. The language for both propositional and predicate logic is unable to express all sets of possible worlds corresponding to that language!

I love this result. It’s the first hint in mathematical logic that syntax and semantics can come apart.

That result is the climax of this post. What I want to do with the rest of this post is to actually give an explicit example of a set of truth assignments that are “indescribable” by any set of sentences, and to prove it. Warning: If you want to read on, things will get a bit more technical from here.

Alright, so we’ll use a shortcut to denote truth assignments. A truth assignment will be written as a string of “T”s and “F”s, where the nth character corresponds to how the truth assignment evaluates Pn. So the all-true truth assignment will just be written “TTTTTT…” and the all-false truth assignment will be written “FFFFF…”. The truth assignment corresponding to P1 being false and everything else true will be written “FTTTTT…”. And so on.

Now, here’s our un-describable set of truth assignments. {“FFFFFF…”, “TFFFFF…”, “TTFFFF…”, “TTTFFF…”, …}. Formally, define Vn to be the truth assignment that assigns “True” to every atomic proposition up to and including Pn, and “False” to all others. Now our set of truth assignments is just {Vn | n ∈ ℕ}.

Let’s prove that no set of sentences uniquely picks out this set of truth assignments. We prove by contradiction. Suppose that we could find a set of sentences that uniquely pick out these truth assignments and none other. Let’s call this set A. Construct a new set of sentences A’ by appending all atomic propositions A: A’ = A ∪ {P1, P2, P3, …}.

Is there any truth assignment that is consistent with all of A’? Well, we can answer this by using the Compactness Theorem: A’ has a truth assignment if and only if every finite subset of A’ has a truth assignment. But every finite subset of A’ involves sentences from A (which are consistent with Vn for each n by assumption), and a finite number of atomic propositions. Since each finite subset of A’ is only asserting the truth of a finite number of atomic sentences, we can always find a truth assignment Vk in our set that is consistent with it, by choosing one that switches to “False” long after the last atomic proposition that is asserted by our finite subset.

This means that each finite subset of A’ is consistent with at least one of our truth assignments, which means that A’ is consistent with at least one of our truth assignments. But A’ involves the assertion that all atomic propositions are true! The only truth assignment that is consistent with this assertion is the all-true assignment! And is that truth assignment in our set? No! And there we have it, we’ve reached our contradiction!

We cannot actually describe a set of possible worlds in which either all atomic propositions are false, or only the first is true, or only the first two are true, or only the first three are true, and so on forever. But this might prompt the question: didn’t you just describe it? How did you do that, if it’s impossible? Well, technically I didn’t describe it. I just described the first four possibilities and then said “and so on forever”, assuming that you knew what I meant. To have actually fully pinned down this set of possible worlds, I would have had to continue with this sentence forever. And importantly, since this sentence is a disjunction, I could not split this infinite sentence into an infinite set of finite sentences. This fundamental asymmetry between ∨ and ∧ is playing a big role here: while an infinite conjunction can be constructed by simply putting each clause in the conjunction as a separate sentence, an infinite disjunction cannot be. This places a fundamental limit on the ability of a language with only finite sentences to describe the world.

The full solution to the dog puzzle

A couple of days ago I posted A logic puzzle about dogs. Read that post first and try solving it before reading on!

Below is the full explanation of the puzzle. Heavy spoilers, obviously.

The dog society is structured identically to the Robinson axioms for natural number arithmetic. Dogs are numbers, Spot is 0, the alpha of n is n + 1, the referee for n and m is n + m, the counselor for n and m is n × m, and the strength relation is the < relation. This means that “the marriage counselor for Spot’s alpha and the referee of a fight between Spot’s alpha and Spot’s alpha” is translated to 1 × (1 + 1) = 2, which is Spot’s alpha’s alpha. In Robinson arithmetic, you can also prove that ∀n (n < n + 1).

As for the question of if it’s possible for a dog to be stronger than Spot, Spot’s alpha, Spot’s alpha’s alpha, and so on: The primary difference between Robinson arithmetic and Peano arithmetic (the usual axiomatization of natural numbers) is that the Robinson axioms have no induction axiom (which would be something like “If Spot has a property, and if the alpha of any dog with the property also has the property, then all dogs have the property”). The induction axiom serves to rule out many models of the axioms that are not actually the natural numbers.

If the induction axiom is stated using second-order logic, then the axiom system uniquely pins down (ℕ,+,×,>) and there are no dogs besides those in Spot’s hierarchy. But the induction axiom cannot be stated as a single axiom in first order logic, since it involves quantifying over all properties. For first-order Peano arithmetic, we instead have an infinite axiom schema, one for each property that is definable within the first-order language. This turns out to be strictly weaker than the single second-order axiom, as there are some properties of numbers that cannot be described in a first-order language (like being larger than a finite number of numbers).

What this amounts to is that first-order Peano arithmetic with its infinite axiom schema is too weak to pin down the natural numbers as a unique model. There are what’s called nonstandard models of first order PA, which contains the ordinary numbers but also an infinity of weird extra numbers.(In fact, there exist models of first-order PA with every infinite cardinality!) Some of these numbers have the property that they are bigger than all the regular natural numbers.And since Robinson arithmetic is strictly weaker than first-order PA (lacking an induction axiom as it does), this means that Robinson arithmetic is also not strong enough to rule out numbers greater than all elements of ℕ. Which means that we cannot prove that there are no dogs stronger than every dog in Spot’s hierarchy!

I made this puzzle to illustrate three things: First, how the same axioms, and even the same models of the same axioms, can have wildly different interpretations, and that a shift in how you think about these axioms can make seemingly impossible tasks trivial. Second, how axiom systems for structures like ℕ can (and inevitably do) fail to capture important and intuitively obvious features of the structure. And third, how logic is so interesting! Just trying to create a simple rule system to describe one of the most natural and ordinary mathematical structures that we ALL use ALL THE TIME for reasoning about the world, turns out to be so nontrivial, and in fact impossible to do perfectly!

The Surprise-Response Heuristic

Often we judge if somebody else is understanding something that we do not understand by whether the things they say in response to our questions are surprising.

When somebody understands it about as well as you do, the things they say about it will generally be fairly understandable and expected (as they mesh with your current insufficient level of understanding). But if they actually understand it and you don’t, then you should expect to be surprised by the things they say, since you couldn’t have produced those responses yourself or predicted them coming.


Q: “In this step of the proof, are they talking about extending the model or the language?”

A1: “They’re talking about extending the model. Look at the way that they worded the description of the extension in the previous step, it specifically describes adding a character to the model, not the language.”

A2: “No, they couldn’t be extending the model even though the wording suggests that, because then the proof wouldn’t even work; it’s required that we just change the language or else we end up working with a different model and failing to prove that the original model had the desired property. Also, it doesn’t even make sense to talk about adding a character to a model, the characters are a property of the language.”

Even with no context to judge whether the claims are true, I imagine that the second response feels much more convincing than the first, even though it’s probably less likely to be understood. The first is the type of response that is unsurprising and easy to see coming, and indicates only that the person is understanding the grammar of the English sentences they’re reading. It doesn’t strongly discriminate between a person that understands what’s going on and a person that doesn’t. The second is certainly surprising; it suggests that the person objects to the specific wording of the proof because of their understanding of the way it misrepresents the logical structure of the argument. They aren’t just comprehending the grammar, they are comprehending the actual content. Ordinarily, a person wouldn’t be able to off-the-cuff make up a response like that without actually understanding what’s going on.

This is a problem when people are good at saying surprising things without understanding. I’ve met a few people that are very good “contrarians”; they are good at coming up with strange and creative ways to say things that ultimately shed very little insight on the topic at hand. I often found myself in a weird position with such people where I feel like they understand the topic at hand better than me, and yet simultaneously I’m deeply suspicious of every word coming out of their mouth.

Three paradoxes of self-reference

I’ve spent a lot of time in the past describing the paradoxes that arise when we try to involve infinities into our reasoning. Another way to get paradoxes aplenty is by invoking self-reference. Below are three of the best paradoxes of self-reference for you to sort out.

In each case, I want you to avoid the temptation to just say “Ah, it’s just self-reference that’s causing the problem” and feel that the paradox is thus resolved. After all, there are plenty of benign cases of self-reference. Self-modifying code, flip-flops in computing hardware, breaking the fourth wall in movies, and feeding a function as input to itself are all examples. Self-referential definitions in mathematics are also often unobjectionable: as an example, we can define the min function by saying y = min(X) iff y is in X and for all elements x of X, y ≤ x (the definition quantifies over a group of objects that includes the thing being defined). So if we accept that self-reference is not necessarily paradoxical (just as infinity is sometimes entirely benign), then we must do more to resolve the below paradoxes than just say “self-reference.”

1. Berry’s Paradox

Consider the set of integers definable in an English sentence of under eighty letters. This set is finite, because there are only a finite number of possible strings of English characters of under eighty letters. So since this set is finite and there are an infinity of integers, there must be a smallest integer that’s not in the set.

But hold on: “The smallest positive integer not definable in under eighty letters” appears to now define this integer, and it does so with only 67 letters! So now it appears that there is no smallest positive integer not definable in under eighty letters. And that means that our set cannot be finite! But of course, the cardinality of the set of strings cannot be less than the cardinality of the set of numbers those strings describe. So what’s going on here?

2. Curry’s paradox

“If this sentence is true, then time is infinite.”

Curry’s paradox tells us that just from the existence of this sentence (assuming nothing about its truth value), we can prove that time is infinite.

Proof 1

Let’s call this sentence P. We can then rewrite P as “If P is true, then time is infinite.” Now, let’s suppose that the sentence P is true. That means the following:

Under the supposition that P is true, it’s true that “If P is true, then time is infinite.”

And under the supposition that P is true, P is true.

So under the supposition that P is true, time is infinite (by modus ponens within the supposition).

But this conclusion we’ve just reached is just the same thing as P itself! So we’ve proven that P is true.

And therefore, since P is true and “If P is true, then time is infinite” is true, time must be infinite!

If you’re suspicious of this proof, here’s another:

Proof 2

If P is false, then it’s false that “If P is true then time is infinite.” But the only way that sentence can be false is if the antecedent is true and the consequent false, i.e. P is true and time is finite. So from P’s falsity, we’ve concluded P’s truth. Contradiction, so P must be true.

Now, if P is true, then it’s true that “If P is true, then time is infinite”. But then by modus ponens, time must be infinite.

Nothing in our argument relied on time being infinite or finite, so we could just as easily substitute “every number is prime” for “time is infinite”, or anything we wanted. And so it appears that we’ve found a way to prove the truth any sentence! Importantly, our conclusion doesn’t rest on the assumption of the truth of the sentence we started with! All it requires is the *existence of the sentence*. Is this a proof of the inconsistency of propositional logic? And if not, then where exactly have we gone wrong?

3. Curry’s paradox, Variant

Consider the following two sentences:

1) At least one of these two sentences is false.
2) Not all numbers are prime.

Suppose that (1) is false. Well then at least one of the two sentences is false, which makes (1) true! This is a contradiction, so (1) must be true.

Since (1) is true, at least one of the two sentences must be false. But since we already know that (1) is true, (2) must be false. Which means that all numbers are prime!

Just like last time, the structure of the argument is identical no matter what we put in place of premise 2, so we’ve found a way to disprove any statement! And again, we didn’t need to start out by assuming anything about the truth values of sentences (1) and (2), besides that they have truth values.

Perhaps the right thing to say, then, is that we cannot always be sure that self-referential statements actually have truth values. But then we have to answer the question of how we are to distinguish between self-referential statements that are truth-apt and those that are not! And that seems very non-trivial. Consider the following slight modification:

1) Both of these two sentences are true.
2) Not all numbers are prime.

Now we can just say that both (1) and (2) are true, and there’s no problem! And this seems quite reasonable; (1) is certainly a meaningful sentence, and it seems clear what the conditions for its truth would be. So what’s the difference in the case of our original example?

A logic puzzle about dogs

Spot is a dog. Every dog has one alpha (also a dog), and no two dogs have the same alpha. But Spot, alone amongst dogs, isn’t anybody’s alpha. 🙁

For any two dogs, there is an assigned referee in case they get into a fight and an assigned marriage counselor in case they get married. The dogs have set up the following rules for deciding who will be the referees and counselors for who:

If Spot fights with any dog, the other dog gets to be the referee. The referee of a fight between dog 1’s alpha and dog 2 has to be the alpha of the referee of a fight between dogs 1 and 2.

Spot has to be his own marriage counselor, no matter who he marries. The marriage counselor for dog 1’s alpha and dog 2 has to referee any fight between dog 2 and the marriage counselor for dog 1 and dog 2.

Finally, dog 1 is stronger then dog 2 if and only if dog 1 is the referee for dog 2 and some other dog. Strength is transitive, and no dog is stronger than itself.

Question 1: Who’s the marriage counselor for Spot’s alpha and the referee of a fight between Spot’s alpha and Spot’s alpha?

Question 2: How many dogs are there?

Question 3: Is the referee for dog 1’s alpha and dog 2 always the same as the referee for dog 2’s alpha and dog 1? What if we asked the same question about marriage counselors?

Question 4: Is any dog stronger than their own alpha?

Bonus Question: Is it possible for there to be a dog that’s stronger than Spot, Spot’s alpha, Spot’s alpha’s alpha, and so on?

A challenge: Try to figure out a trick that allows you to figure out the above questions in your head. I promise, it’s possible!