Gödel part 3: Meta-Logic with Arithmetic

On to the next part of Gödel’s proof of incompleteness. To refresh your memory, here’s a sketch of the proof:

  1. Take a simple logic. We’ve been using a variant of the Principia Mathematica’s logic, because that’s what Gödel used.
  2. Show that any statement in the logic can be encoded as a number using an arithmetic process based on the syntax of the logic. The process of encoding statements numerically is called Gödel numbering.
  3. Show that you can express meta-mathematical properties of logical statements in terms of arithemetic properties of their Gödel numbers. In particular, we need to build up the logical infrastructure that we need to talk about whether or not a statement is provable.
  4. Using meta-mathematical properties, show how you can create an unprovable statement encoded as a Gödel number.

What we’ve done so far is the first two steps, and part of the third. In this post, we saw the form of the Principia’s logic that we’re using, and how to numerically encode it as a Gödel numbering. We’ve start started on the third point in this post, by figuring out just what it means to say that things are encoded arithmetically. Now we can get to the part where we see how to encode meta-mathematical properties in terms of arithmetic properties of the Gödel numbering. In this post, we’re going to build up everything we need to express syntactic correctness, logical validity, and provability in terms of arithmetical properties of Gödel numbers. (And, as a reminder, I’ve been using this translation on Gödel’s original paper on incompleteness.)

This is the most complex part of the incompleteness proof. The basic concept of what we’re doing is simple, but the mechanics are very difficult. What we want to do is define a set of predicates about logical statements, but we want those predicates to be expressed as arithmetic properties of the numerical representations of the logical statements.

The point of this is that we’re showing that done in the right way, arithmetic is logic – that doing arithmetic on the Gödel numbers is doing logical inference. So what we need to do is build up a toolkit that shows us how to understand and manipulate logic-as-numbers using arithmetic. As we saw in the last post, primitive recursion is equivalent to arithmetic – so if we can show how all of the properties/predicates that we define are primitive recursive, then they’re arithmetic.

This process involves a lot of steps, each of which is building the platform for the steps that follow it. I struggled quite a bit figuring out how to present these things in a comprehensible way. What I ended up with is writing them out as code in a pseudo-computer language. Before inventing this language, I tried writing actual executable code, first in Python and then in Haskell, but I wasn’t happy with the clarity of either one.

Doing it in an unimplemented language isn’t as big a problem as you might think. Even if this was all executable, you’re not going to be able to actually run any of it on anything real – at least not before you hair turns good and gray. The way that this stuff is put together is not what any sane person would call efficient. But the point isn’t to be efficient: it’s to show that this is possible. This code is really all about searching; if we wanted to be efficient, this could all be done in a different representation, with a different search method that was a lot faster – but that wolud be harder to understand.

So, in the end, I threw together a simple language that’s easy to read. This language, if it were implemented, wouldn’t really even be Turing complete – it’s a primitive recursive language.

Basics

We’ll start off with simple numeric properties that have no obvious connection to the kinds of meta-mathematical statements that we want to talk about, but we’ll use those to define progressively more and more complex and profound properties, until we finally get to our goal.

# divides n x == True if n divides x without remainder.
pred divides(n, x) = x mod n == 0

pred isPrime(0) = False
pred isPrime(1) = False
pred isPrime(2) = True
pred isPrime(n) = {
  all i in 2 to n {
    not divides(i, n)
  }
}

fun fact(0) = 1
fun fact(n) = n * fact(n - 1)

Almost everything we’re going to do here is built on a common idiom. For anything we want to do arithmetically, we’re going to find a bound – a maximum numeric value for it. Then we’re going to iterate over all of the values smaller than that bound, searching for our target.

For example, what’s the nth prime factor of x? Obviously, it’s got to be smaller than x, so we’ll use x as our bound. (A better bound would be the square root of x, but it doesn’t matter. We don’t care about efficiency!) To find the nth prime factor, we’ll iterate over all of the numbers smaller than our bound x, and search for the smallest number which is prime, which divides x, and which is larger than the n-1th prime factor of x. We’ll translate that into pseudo-code:

fun prFactor(0, x) = 0
fun prFactor(n, x) = {
  first y in 1 to x {
    isPrime(y) and divides(y, x) and prFactor(n - 1, x) < y
  }
}

Similarly, for extracting values from strings, we need to be able to ask, in general, what's the nth prime number? This is nearly identical to prFactor above. The only difference is that we need a different bound. Fortunately, we know that the nth prime number can't be larger than the factorial of the previous prime plus 1.

fun nthPrime(0) = 0
fun nthPrime(n) = {
  first y in 1 to fact(nthPrime(n - 1)) + 1  {
    isPrime(y) and y > nthPrime(n - 1))
  }
}

In composing strings of Gödel numbers, we use exponentiation. Given integers x and n, xn, we can obviously compute them via primitive recursion. I'll define them below, but in the rest of this post, I'll write them as an operator in the language:

fun pow(n, 0) = 1
fun pow(n, i) = n * pow(n, i - 1)

String Composition and Decomposition

With those preliminaries out of the way, we can get to the point of defining something that's actually about one of the strings encoded in these Gödel numbers. Given a number n encoding a string, item(n, x) is the value of the nth character of x. (This is slow. This is really slow! We're getting to the limit of what a very powerful computer can do in a reasonable amount of time. But this doesn't matter. The point isn't that this is a good way of doing these things: it's that these things are possible. To give you an idea of just how slow this is, I started off writing the stuff in this post in Haskell. Compiled with GHC, which is a very good compiler, using item to extract the 6th character of an 8 character string took around 10 minutes on a 2.4Ghz laptop. In the stuff that follows, we'll be using this to extract characters from strings that could be hundreds of characters long!)

fun item(n, x) = {
  first y in 1 to x {
    divides(prFactor(n, x) ** y, y) and
      not divides(prFactor(n, x)**(y+1), x)
  }
}

Given a string, we want to be able to ask how long it is; and given two strings, we want to be able to concatenate them.

fun length(x) = {
  first y in 1 to x {
    prFactor(y, x) > 0 and prFactor(y + 1, x) == 0
  }
}

fun concat(x, y) = {
  val lx = length(x)
  val ly = length(y)

  first z in 1 to nthprime(lx + ly)**(x + y) {
    (all n in 1 to lx {
        item(n, z) == item(n, x)
     }) and (all n in 1 to ly {
        item(n + lx, z) == item(n, y)
      })
  }
}

fun concatl([]) = 0
fun concatl(xs) = {
  concat(head(xs), concatl(tail(xs)))
}

fun seq(x) = 2**x

We want to be able to build statements represented as numbers from other statements represented as numbers. We'll define a set of functions that either compose new strings from other strings, and to check if a particular string is a particular kind of syntactic element.

# x is a variable of type n.
pred vtype(n, x) = {
  some z in 17 to x {
    isPrime(z) and x == n**z
  }
}

# x is a variable
pred isVar(x) = {
  some n in 1 to x {
    vtype(n, x)
  }
}

fun paren(x) =
  concatl([gseq(11), x, gseq(13)])

# given the Gödel number for a statement x, find
# the Gödel number for not x.
fun gnot(x) =
  concat(gseq(5), paren(x))

# Create the number for x or y.
fun gor(x, y) =
  concatl([paren(x), seq(7), paren(y)])

# Create the number for 'forall x(y)'.
fun gforall(x, y) =
  concatl([seq(9), seq(x), paren(y)])

# Create the number for x with n invocations of the primitive
# successor function.
fun succn(0, x) = x
fun succn(n, x) = concat(seq(3), succn(n - 1, x))

# Create the number n using successor and 0.
fun gnumber(n) = succn(n, seq(1))

# Check if a statement is type-1.
pred stype_one(x) = {
  some m in 1 to x {
     m == 1 or (vtype(1, m) and x == succn(n, seq(m))
  }
}

# Check if a statement is type n.
pred fstype(1, x) = stype_one(x)
pred fstype(n, x) =
  some v in 1 to x {
    vtype(n, v) and R(v)
  }
}

That last function contains an error: the translation of Gödel that I'm using says R(v) without defining R. Either I'm missing something, or the translator made an error.

Formulae

Using what we've defined so far, we're now ready to start defining formulae in the basic Principia logic. Forumlae are strings, but they're strings with a constrained syntax.

pred elFm(x) = {
  some y in 1 to x {
    some z in 1 to x {
      some n in 1 to x {
        stype(n, y) and stype(n+1, z) and x == concat(z, paren(y))
      }
    }
  }
}

All this is doing is expressing the grammar rule in arithmetic form: an elementary formula is a predicate: P(x), where x is a variable on level n, and P is a variable of level x + 1.

The next grammar rule that we encode this way says how we can combine elementary formulae using operators. There are three operators: negation, conjunction, and universal quantification.

pred op(x, y, z) = {
  x == gnot(y) or
  x == gor(y, z) or
  (some v in 1 to x { isVar(v) and x == gforall(v, y) })
}

And now we can start getting complex. We're going to define the idea of a valid sequence of formulae. x is a valid sequence of formulae when it's formed from a collection of formulae, each of which is either an elementary formula, or is produced from the formulae which occured before it in the sequence using either negation, logical-or, or universal quantification.

In terms of a more modern way of talking about it, the syntax of the logic is a grammar. A formula sequence, in this system, is another way of writing the parse-tree of a statement: the sequence is the parse-tree of the last statement in the sequence.

pred fmSeq(x) = {
  all p in 0 to length(x) {
    elFm(item(n, x)) or
      some p in 0 to (n - 1) {
        some q in 0 to (n - 1) {
          op(item(n,x), item(p, x), item(q, x))
        }
      }
  }
}

The next one bugs me, because it seems wrong, but it isn't really! It's a way of encoding the fact that a formula is the result of a well-defined sequence of formulae. In order to ensure that we're doing primitive recursive formulae, we're always thinking about sequences of formulae, where the later formulae are produced from the earlier ones. The goal of the sequence of formula is to produce the last formula in the sequence. What this predicate is really saying is that a formula is a valid formula if there is some sequence of formulae where this is the last one in the sequence.

Rephrasing that in grammatical terms, a string is a formula if there is valid parse tree for the grammar that produces the string.

pred isFm(x) = {
  some n in 1 to nthPrime(length(x)**2)**(x*length(x)**2) {
    fmSeq(n)
  }
}

So, now, can we say that a statement is valid because it's parsed according to the grammar? Not quite. It's actually a familiar problem for people who write compilers. When you parse a program in some language, the grammar doesn't usually specify variables must be declared before they're used. It's too hard to get that into the grammar. In this logic, we've got almost the same problem: the grammar hasn't restricted us to only use bound variables. So we need to have ways to check whether a variable is bound in a Gödel-encoded formula, and then use that to check the validity of the formula.

# The variable v is bound in formula x at position n.
pred bound(v, n, x) = {
  isVar(v) and isFm(x) and
  (some a in 1 to x {
    some b in 1 to x {
      some c in 1 to x {
        x == concatl([a, gforall(v, b), c]) and
        isFm(b) and
        length(a) + 1 ≤ n ≤ length(a) + length(forall(v, b))
      }
    }
  })
}

# The variable v in free in formula x at position n
pred free(v, n, x) = {
  isVar(v) and isFm(x) and
  (some a in 1 to x {
    some b in 1 to x {
      some c in 1 to x {
        v == item(n, x) and n ≤ length(x) and not bound(v, n, x)
      }
    }
  })
}

pred free(v, x) = {
  some n in 1 to length(x) {
    free(v, n, x)
  }
}

To do logical inference, we need to be able to do things like replace a variable with a specific infered value. We'll define how to do that:

# replace the item at position n in x with y.
fun insert(x, n, y) = {
  first z in 1 to nthPrime(length(x) + length(y))**(x+y) {
    some u in 1 to x {
      some v in 1 to x {
        x == concatl([u, seq(item(n, x)), v]) and
        z == concatl([u, y, v]) and
        n == length(u) + 1
      }
    }
  }
}

There are inference operations and validity checks that we can only do if we know whether a particular variable is free at a particular position.

# freePlace(k, v, k) is the k+1st place in x (counting from the end)
# where v is free.
fun freePlace(0, v, x) = {
  first n in 1 to length(x) {
    free(v, n, x) and
    not some p in n to length(x) {
      free(v, p, x)
    }
  }
}

fun freePlace(k, v, x) = {
  first n in 1 to freePlace(n, k - 1, v) {
    free(v, n, x) and
    not some p in n to freePlace(n, k - 1, v) {
      free(v, p, x)
    }
  }
}

# number of places where v is free in x
fun nFreePlaces(v, x) = {
  first n in 1 to length(x) {
    freeplace(n, v, x) == 0
  }
}

In the original logic, some inference rules are defined in terms of a primitive substitution operator, which we wrote as subst[v/c](a) to mean substitute the value c for the variable c in the statement a. We'll build that up on a couple of steps, using the freePlaces function that we just defined.

# Subst1 replaces a single instance of v with y.
fun subst'(0, x, v, y) = x
fun subst1(0k, x, v, y) =
  insert(subst1(k, x, v, y), freePlace(k, v, x), y)

# subst replaces all instances of v with y
fun subst(x, v, y) = subst'(nFreePlaces(v, x), x, v, y)

The next thing we're going to do isn't, strictly speaking, absolutely necessary. Some of the harder stuff we want to do will be easier to write using things like implication, which aren't built in primitive of the Principia logic. To write those as clearly as possible, we'll define the full suite of usual logical operators in terms of the primitives.

# implication
fun gimp(x, y) = gor(gnot(x), y)

# logical and
fun gand(x, y) = gnot(gor(gnot(x), gnot(y)))

# if/f
fun gequiv(x, y) = gand(gimp(x, y), gimp(y, x))

# existential quantification
fun gexists(v, y) = not(gforall(v, not(y)))

Axioms

The Peano axioms are valid logical statements, so they have Gödel numbers in this system. We could compute their value, but why bother? We know that they exist, so we'll just give them names, and define a predicate to check if a value matches them.

The form of the Peano axioms used in incompleteness are:

  1. Zero: ¬(succ(x1) = 0)
  2. Uniqueness: succ(x1) = succ(y1) Rightarrow x = y
  3. Induction: x2(0) ∧ ∀x1(x2(x1)⇒ x2(succ(x1))) ⇒ ∀x1(x2(x1))
const pa1 = ...
const pa2 = ...
const pa3 = ...

pred peanoAxiom(x) =
  (x == pa1) or (x == pa2) or (x == pa3)

Similarly, we know that the propositional axioms must have numbers. The propositional
axioms are:

  1. p \lor p \Rightarrow p
  2. p \Rightarrow p \lor q
  3. p \lor q \Rightarrow p \lor q
  4. (p \Rightarrow q) \Rightarrow (r \lor p \Rightarrow r \lor q)

I'll show the translation of the first - the rest follow the same pattern.

# Check if x is a statement that is a form of propositional
# axiom 1: y or y => y
pred prop1Axiom(x) =
  some y in 1 to x {
    isFm(x) and x == imp(or(y, y), y)
  }
}

pred prop2Axiom(x) = ...
pred prop3Axiom(x) = ...
pred prop4Axiom(x) = ...
pred propAxiom(x) = prop2Axiom(x) or prop2Axiom(x) or
    prop3Axiom(x) or prop4Axiom(x)

Similarly, all of the other axioms are written out in the same way, and we add a predicate isAxiom to check if something is an axiom. Next is quantifier axioms, which are complicated, so I'll only write out one of them - the other follows the same basic scheme.

The two quantifier axioms are:

  1. \forall v(a) \Rightarrow \text{subst}[v/c](a)
  2. \forall v(b \lor a) \Rightarrow (b \lor \forall v(a))
quantifier_axiom1_condition(z, y, v) = {
  not some n in 1 to length(y) {
    some m in 1 to length(z) {
      some w in 1 to z {
         w == item(m, z) and bound(w, n, y) and free(v, n, y)
      }
    }
  }
}

pred quantifier1Axiom(x) = {
  some v in 1 to x {
    some y in 1 to x {
      some z in 1 to x {
        some n in 1 to x {
          vtype(n, v) and stype(n, z) and
          isFm(y) and
          quantifier_axiom1_condition(z, y, v) and
          x = gimp(gforall(v, y), subst(y, v, z))
        }
      }
    }
  }
}

quanitifier_axiom2 = ...
isQuantifierAxiom = quantifier1Axiom(x) or quantifier2Axiom(x)

We need to define a predicate for the reducibility axiom (basically, the Principia's version of the ZFC axiom of comprehension). The reducibility axiom is a schema: for any predicate a, \exists u (\forall v (u(v) \Leftrightarrow a. In our primitive recursive system, we can check if something is an instance of the reducibility axiom schema with:

pred reduAxiom(x) =
  some u in 1 to x {
    some v in 1 to x {
      some y in 1 to x {
        some n in 1 to x {
          vtype(n, v) and
          vtype(n+1, u) and
          not free(u, y) and
          isFm(y) and
          x = gexists(u, gforall(v, gequiv(concat(seq(u), paren(seq(v))), y)))
        }
      }
    }
  }
}

Now, the set axiom. In the logic we're using, this is the axiom that defines set equality. It's written as \forall x_1 (x_2(x_1) \Leftrightarrow y_2(y_1) \Rightarrow x_2 = x_1). Set equality is defined for all types of sets, so we need to have one version of axiom for each level. We do that using type-lifting: we say that the axiom is true for type-1 sets, and that any type-lift of the level-1 set axiom is also a version of the set axiom.

fun typeLift(n, x) = {
  first y in 1 to x**(x**n) {
    all k in 1 to length(x) {
      item(k, x) ≤ 13 and item(k, y) == item(k, v) or
      item(k, x) > 13 and item(k, y) = item(k, x) * prFactor(1, item(k, x))**n
    }
  }
}

We haven't defined the type-1 set axiom. But we just saw the axiom above, and it's obviously a simple logical statement. That mean that it's got a Gödel number. Instead of computing it, we'll just say that that number is called sa1. Now we can define a predicate to check if something is a set axiom:

val sa1 = ...
pred setAxiom(x) =
  some n in 1 to x {
    x = typeLift(n, sa)
  }
}

We've now defined all of the axioms of the logic, so we can now create a general predicate to see if a statement fits into any of the axiom categories:

pred isAxiom(x) =
  peanoAxiom(x) or propAxiom(x) or quantifierAxom(x) or
  reduAxiom(x) or setAxiom(x)

Proofs and Provability!

With all of the axioms expressible in primitive recursive terms, we can start on what it means for something to be provable. First, we'll define what it means for some statement x to be an immediate consequence of some statements y and z. (Back when we talked about the Principia's logic, we said that x is an immediate consequence of y and z if either: y is the formula z ⇒ x, or if c is the formula ∀v.x).

pred immConseq(x, y, z) = {
  y = imp(z, x) or
  some v in 1 to x {
    isVar(v) and x = forall(v, y)
  }
}

Now, we can use our definition of an immediate consequence to specify when a sequence of formula is a proof figure. A proof figure is a sequence of statements where each statement in it is either an axiom, or an immediate consequence of two of the statements that preceeded it.

pred isProofFigure(x) = {
  (all n in 0 to length(x) {
    isAxiom(item(n, x)) or
    some p in 0 to n {
      some q in 0 to n {
        immConseq(item(n, x), item(p, x), item(q, x))
      }
    }
  }) and
  length(x) > 0
}

We can say that x is a proof of y if x is proof figure, and the last statement in x is y.

pred proofFor(x, y) =
  isProofFigure(x) and
  item(length(x), x) == y

Finally, we can get to the most important thing! We can define what it means for something to be provable! It's provable if there's a proof for it!

pre provable(x) =
  some y {
    proofFor(y, x)
  }
}

Note that this last one is not primitive recursive! There's no way that we can create a bound for this: a proof can be any length.

At last, we're done with these definition. What we've done here is really amazing: now, every logical statement can be encoded as a number. Every proof in the logic can be encoded as a sequence of numbers: if something is provable in the Principia logic, we can encode that proof as a string of numbers, and check the proof for correctness using nothing but (a whole heck of a lot of) arithmetic!

Next post, we'll finally get to the most important part of what Gödel did. We've been able to define what it means for a statement to be provable - we'll use that to show that there's a way of creating a number encoding the statement that something is not provable. And we'll show how that means that there is a true statement in the Principia's logic which isn't provable using the Principia's logic, which means that the logic isn't complete.

In fact, the proof that we'll do shows a bit more than that. It doesn't just show that the Principia's logic is incomplete. It shows that any consistent formal system like the Principia, any system which is powerful enough to encode Peano arithmetic, must be incomplete.

Godel part 2: Arithmetic and Logic

In the last post, we saw how to take statements written in the logic of the Principia Mathematica, and convert them into numerical form using Gödel numbering. For the next step in Gödel’s proof, we need to go meta-mathematical.

Ultimately, we want to write first-order statements that can reason about first order statements. But the entire structure of the principia and its logic is designed to make
that impossible. First order statements can only reason about numbers and their properties.

But now, we’ve got the ability to represent statements – first order, second order, third order, any order. What we still need is a way of describing the properties of those numerical statements in terms of operations that can be expressed using nothing but first order statements.

The basic trick to incompleteness is that we’re going to use the numerical encoding of statements to say that a predicate or relation is represented by a number. Then we’re going to write predicates about predicates by defining predicates on the numerical representations of the first-order predicates. That’s going to let us create a true statement in the logic that can’t be proven with the logic.

To do that, we need to figure out how to take our statements and relations represented as numbers, and express properties of those statements and relations in terms of arithmetic. To do that, we need to define just what it means to express something arithmetically. Gödel did that by defining “arithmetically” in terms of a concept called primitive recursion.

I learned about primitive recursion when I studied computational complexity. Nowadays, it’s seen as part of theoretical computer science. The idea, as we express it in modern terms, is that there are many different classes of computable functions. Primitive recursion is one of the basic complexity classes. You don’t need a Turing machine to compute primitive recursive functions – they’re a simpler class.

The easiest way to understand primitive recursion is that it’s what you get in a programming language with integer arithmetic, and simple for-loops. The only way you can iterate is by repeating things a bounded number of times. Primitive recursion has a lot of interesting properties: the two key ones for our purposes here are: number theoretic proofs are primitive recursive, and every computation of a primitive recursive function is guaranteed to complete within a bounded amount of time.

The formal definition of primitive recursion, the way that Gödel wrote it, is quite a bit more complex than that. But it means the same thing.

We start with what it means to define a formula via primitive recursion. (Note the language that I used there: I’m not explaining what it means for a function to be primitive recursive; I’m explaining what it means to be defined via primitive recursion.) And I’m defining formulae, not functions. In Gödel’s proof, we’re always focused on numerical reasoning, so we’re not going to talk about programs or algorithms, we’re going to about the definition of formulae.

A formula phi(x_1, x_2, ..., x_n) is defined via primitive recursion if, for some other formulae \psi and \mu:

  • Base: \phi(0, x_2, ..., x_n) = \psi(x_2, ..., x_n)
  • Recursive: \phi(i+1, x_2, ..., x_n) = \mu(i, \phi(i, x_2, ..., x_n), x_2, ..., x_n).

So, basically, the first parameter is a bound on the number of times that phi can invoked recursively. When it’s 0, you can’t invoke \phi any more.

A formula is primitive recursive if it defined from a collection of formulae \phi_1, ..., \phi_n where any formula \phi_i is defined via primitive recursion from \phi_1, ..., \phi_{i-1}, or the primitive succ function from Peano arithmetic.

For any formula phi_i in that sequence, the degree of the formula is the number of other primitive recursive formulae used in its definition.

Now, we can define a primitive recursive property: R(x_1, ..., x_n) is primitive recursive if and only if there exists a primitive recursive function \phi such that \phi(x_1, ..., x_n) = 0.

With primitive recursive formulae and relations defined, there’s a bunch of theorems about how you can compose primitive recursive formulae and relations:

  1. Every function or relation that you get by substituting a primitive recursive function for a variable in a primitive recursive function/relation is primitive recursive.
  2. If R and S are primitive relations, then ¬R, R∧S, R∨S are all primitive recursive.
  3. If \phi(x_1, ..., x_n) and \psi(x_1, ..., x_n) are primitive recursive functions, then the relation R(x_1, ..., x_n) \Leftrightarrow (\phi(x_1, ..., x_n) = \psi(x_1, ..., x_n) is also primitive recursive.
  4. Let xv and zv be finite-length tuples of variables. If the function \phi(xv) and the relation R(y, zv) are primitive recursive, then so are the relations:
    • S(xv, zv) \Leftrightarrow  (\exists y \le \phi(xv). R(y, zv))
    • T(xv, zv) \Leftrightarrow  (\forall y \le A(xv). R(y, zv))
  5. Let xv and zv be finite-length tuples of variables. And let text{argmin}[y le f(x).R(x)] be the smallest value of x for which y le f(x) and R(x) is true, or 0 if there is no such value. Then if the function phi(xv) and the relation R(y, zv) are primitive recursive, then so is the function P(xv, zv) = (\text{argmin}[y \le A(xv). R(y, zv))].

By these definitions, addition, subtraction, multiplication, and integer division are all primitive recursive.

Ok. So, now we’ve got all of that out of the way. It’s painful, but it’s important. What we’ve done is come up with a good formal description of what it means for something to be an arithmetic property: if we can write it as a primitive recursive relation or formula, it’s arithmetic.

Godel Numbering

The first step in Gödel’s incompleteness proof was finding a way of taking logical statements and encoding them numerically. Looking at this today, it seems sort-of obvious. I mean, I’m writing this stuff down in a text file – that text file is a stream of numbers, and it’s trivial to convert that stream of numbers into a single number. But when Gödel was doing it, it wasn’t so obvious. So he created a really clever mechanism for numerical encoding. The advantage of Gödel’s encoding is that it makes it much easier to express properties of the encoded statements arithmetically. (Arithmetically means something very specific here; we’ll see what in a later post.

Before we can look at how Gödel encoded his logic into numbers, we need to look at the logic that he used. Gödel worked with the specific logic variant used by the Principia Mathematica. The Principia logic is minimal and a bit cryptic, but it was built for a specific purpose: to have a minimal syntax, and a complete but minimal set of axioms.

The whole idea of the Principia logic is to be purely syntactic. The logic is expected to have a valid model, but you shouldn’t need to know anything about the model to use the logic. Russell and Whitehead were deliberately building a pure logic where you didn’t need to know what anything meant to use it. I’d really like to use Gödel’s exact syntax – I think it’s an interestingly different way of writing logic – but I’m working from a translation, and the translator updated the syntax. I’m afraid that switching between the older Gödel syntax, and the more modern syntax from the translation would just lead to errors and confusion. So I’m going to stick with the translation’s modernization of the syntax.

The basic building blocks of the logic are variables. Already this is a bit different from what you’re probably used to in a logic. When we think of logic, we usually consider predicates to be a fundamental thing. In this logic, they’re not. A predicate is just a shorthand for a set, and a set is represented by a variable.

Variables are stratified. Again, it helps to remember where Russell and Whitehead were coming from when they were writing the Principia. One of their basic motivations was avoiding self-referential statements like Russell’s paradox. In order to prevent that, they thought that they could create a stratified logic, where on each level, you could only reason about objects from the level below. A first-order predicate would be a second-level object could only reason about first level objects. A second-order predicate would be a third-level object which could reason about second-level objects. No predicate could ever reason about itself or anything on its on level. This leveling property is a fundamental property built into their logic. The way the levels work is:

  • Type one variables, which range over simple atomic values, like specific single natural numbers. Type-1 variables are written as a_1, b_1.
  • Type two variables, which range over sets of atomic values, like sets of natural numbers. A predicate, like IsOdd, about specific natural numbers would be represented as a type-2 variable. Type-2 variables are written a_2, b_2, …
  • Type three variables range over sets of sets of atomic values. The mappings of a function could be represented as type-3 variables: in set theoretic terms, a function is set of ordered pairs. Ordered pairs, in turn, can be represented as sets of sets – for example, the ordered pair (1, 4) would be represented by the set { {1}, {1, 4} }. A function, in turn, would be represented by a type-4 variable – a set of ordered pairs, which is a set of sets of sets of values.

Using variables, we can form simple atomic expressions, which in Gödel’s terminology are called signs. As with variables, the signs are divided into stratified levels:

  • Type-1 signs are variables, and successor expressions. Successor expressions are just Peano numbers written with “succ”: 0, succ(0), succ(succ(0)), succ(a1), etc.
  • Signs of any type greater than 1 are just variables of that type/level.

Once you have signs, you can assemble the basic signs into formulae. Gödel explained how to build formulae in a classic logicians form, which I think is hard to follow, so I’ve converted it into a grammar:

 elementary_formula → signn+1(signn)
 formula → ¬(elementary_formula)
 formula → (elementary_formula) or (elementary_formula)
 formula → ∀ signn (elementary_formula)

That’s the entire logic! It’s tiny, but it’s enough. Everything else from predicate logic can be defined in terms of combinations of these basic formulae. For example, you can define logical “and” in terms of negation and logical “or”: (a ∧ b) ⇔ ¬ (¬ a ∨ ¬ b).

With the syntax of the system set, the next thing we need is the basic axioms of logical inference in the system. In terms of logic the way I think of it, these axioms include both “true” axioms, and the inference rules defining how the logic works. There are five families of axioms.

  • First, there’s the Peano axioms, which define the natural numbers.
    1. \lnot \text{succ}(x_1) = 0): 0 is a natural number, and it’s not the successor of anything.
    2. \text{succ}(x_1) = \text{succ}(y_1) \Rightarrow x_1 = y_1: Successors are unique.
    3. (x_2(0) \land \forall x_1 (x_2(x_1) \Rightarrow x_2(succ(x_1)) \Leftrightarrow \forall x_1 (x_2(x_1)): induction works on the natural numbers.
  • Next, we’ve got a set of basic inference rules about simple propositions. These are defined as axiom schemata, which can be instantiated for any set of formalae p, q, and r.
    1. p \lor p \Rightarrow p
    2. p \Rightarrow p \lor q
    3. p \lor q \Rightarrow q \lor p
    4. (p \Rightarrow q) \rightarrow (p \lor r) \Rightarrow q \lor r
  • Axioms that define inference over quantification. v is a variable, a is any formula, b is any formula where v is not a free variable, and c is a sign of the same level as v, and which doesn’t have any free variables that would be bound if it were inserted as a replacement for v.
    1. \forall v(a) \Rightarrow \text{subst}[v/c](a): if formula a is true for all values of v, then you can substitute any specific value c for v in a, and a must still be true.
    2. (\forall v (b \lor a)) \Rightarrow (b \lor \forall v(a))
  • The Principia’s version of the set theory axiom of comprehension:
    \exists u (\forall v (u(v) \Rightarrow a))
  • And last but not least, an axiom defining set equivalence:
    \forall x_i (x_{i+1}(x_i) \Rightarrow y_{i+1}(y_i)) \Rightarrow x_{i+1} = y_{i+1}

So, now, finally, we can get to the numbering. This is quite clever. We’re going to use the simplest encoding: for every possible string of symbols in the logic, we’re going to define a representation as a number. So in this representation, we are not going to get the property that every natural number is a valid formula: lots of natural numbers won’t be. They’ll be strings of nonsense symbols. (If we wanted to, we could make every number be a valid formula, by using a parse-tree based numbering, but it’s much easier to just let the numbers be strings of symbols, and then define a predicate over the numbers to identify the ones that are valid formulae.)

We start off by assigning numbers to the constant symbols:

Symbols Numeric Representation
0 1
succ 3
¬ 5
7
9
( 11
) 13

Variables will be represented by powers of prime numbers, for prime numbers greater that 13. For a prime number p, p will represent a type one variable, p2 will represent a type two variable, p3 will represent a type-3 variable, etc.

Using those symbol encodings, we can take a formula written as symbols x1x2x3…xn, and encode it numerically as the product 2x13x25x2…pnxn, where pn is the nth prime number.

For example, suppose I wanted to encode the formula: ∀ x1 (y2(x1)) ∨ x2(x1).

First, I’d need to encode each symbol:

  1. “∀” would be 9.
  2. “x1“” = 17
  3. “(” = 11
  4. “y2” = 192 = 361
  5. “(” = 11
  6. “x1” = 17
  7. “)” = 13
  8. “∨” = 7
  9. “x2” = 172 = 289
  10. “(” = 11
  11. “x1” = 17
  12. “)” = 13
  13. “)” = 13

The formula would thus be turned into the sequence: [9, 17, 11, 361, 11, 17, 13, 7, 289, 11, 17, 13, 13]. That sequence would then get turned into a single number 29 317 511 7361 1111 1317 1713 197 23289 2911 3117 3713 4113, which according to Hugs is the number (warning: you need to scroll to see it. a lot!):

1,821,987,637,902,623,701,225,904,240,019,813,969,080,617,900,348,538,321,073,935,587,788,506,071,830,879,280,904,480,021,357,988,798,547,730,537,019,170,876,649,747,729,076,171,560,080,529,593,160,658,600,674,198,729,426,618,685,737,248,773,404,008,081,519,218,775,692,945,684,706,455,129,294,628,924,575,925,909,585,830,321,614,047,772,585,327,805,405,377,809,182,961,310,697,978,238,941,231,290,173,227,390,750,547,696,657,645,077,277,386,815,869,624,389,931,352,799,230,949,892,054,634,638,136,137,995,254,523,486,502,753,268,687,845,320,296,600,343,871,556,546,425,114,643,587,820,633,133,232,999,109,544,763,520,253,057,252,248,982,267,078,202,089,525,667,161,691,850,572,084,153,306,622,226,987,931,223,193,578,450,852,038,578,983,945,920,534,096,055,419,823,281,786,399,855,459,394,948,921,598,228,615,703,317,657,117,593,084,977,371,635,801,831,244,944,526,230,994,115,900,720,026,901,352,169,637,434,441,791,307,175,579,916,097,240,141,893,510,281,613,711,253,660,054,258,685,889,469,896,461,087,297,563,191,813,037,946,176,250,108,137,458,848,099,487,488,503,799,293,003,562,875,320,575,790,915,778,093,569,309,011,025,000,000,000.

Next, we’re going to look at how you can express interesting mathematical properties in terms of numbers. Gödel used a property called primitive recursion as an example, so we’ll walk through a definition of primitive recursion, and show how Gödel expressed primitive recursion numerically.

Godel (Reposts)

I’m going to be on vacation this week, which means that I won’t have time to write new posts. But my friend Dr. SkySkull was just talking about Gödel on twitter, and chatting with him, I realized that this would be a good time to repost some stuff that I wrote about Gödel’s incompleteness proof.

Incompleteness is one of the most beautiful and profound proofs that I’ve ever seen. If you’re at all interested in mathematics, it’s something that’s worth taking the effort to understand. But it’s also pretty on-topic for what I’ve been writing about. The original incompleteness proof is written for a dialect of math based on ST type theory!

It takes a fair bit of effort to work through the incompleteness proof, so it’ll be a weeks worth of reposts. What I’m going to do is work with this translation of the original paper where Gödel published his first incompleteness proof. Before we can get to the actual proof, we need to learn a bit about the particular kind of logic that he used in his proof.

It goes right back to the roots of type theory. Set theory was on the rocks, due to Russell’s paradox. Russell’s paradox did was show that there was a foundational problem in math. You could develop what appeared to be a valid mathematicial structure and theory, only to later discover that all the work you did was garbage, because there was some non-obvious fundamental inconsistency in how you defined it. But the way that foundations were treated simple wasn’t strong or formal enough to be able to detect, right up front, whether you’d hard-wired an inconsistency into your theory. So foundations had to change, to prevent another incident like the disaster of early set theory.

In the middle of this, along came two mathematicians, Russell and Whitehead, who wanted to solve the problem once and for all. They created an amazing piece of work called the Principia Mathematica. The principia was supposed to be an ideal, perfect mathematical foundation. It was designed to have to key properties: it was supposed to consistent, and it was supposed to be complete.

  • Consistent meant that the statement would not allow any inconsistencies of any kind. If you used the logic and the foundantions of the Principia, you couldn’t even say anything like Russell’s paradox: you couldn’t even write it as a valid statement.
  • Complete meant that every true statement was provably true, every false statement was provably false, and every statement was either true or false.

A big part of how it did this was by creating a very strict stratification of logic – the stratification that we discussed in ST type theory. The principia’s type-theoretic logic could reason about a specific number, using level-0 statements. You could reason about sets of numbers using level-1 statements. And you could reason about sets of sets of numbers using level-2 statements. In a level-1 statement, you could make meta-statements about level-0 properties, but you couldn’t reason about level-1. In level-2, you could reason about level-1 and level-0, but not about level-2. This meant that you couldn’t make a set like the set of all sets that don’t contain themselves – because that set is formed using a predicate – and when you’re forming a set like “the set of all sets that don’t contain themselves”, the sets you can reason about are a level below you. You can form a level-1 set using a level-1 statement about level-0 objects. But you can’t make a level-1 statement about level-1 objects! So self-referential systems would be completely impossible in the Principia’s logic. You just can’t even state Russell’s paradox. It’s inexpressible.

As a modern student of math, it’s hard to understand what a profound thing they were trying to do. We’ve grown up learning math long after incompleteness became a well-known fact of life. (I read “Gödel Escher Bach” when I was a college freshman – well before I took any particularly deep math classes – so I knew about incompleteness before I knew enough to really understand what completeness woud have meant!) The principia would have been the perfection of math, a final ultimate perfect system. There would have been nothing that we couldn’t prove, nothing in math that we couldn’t know!

What Gödel did was show that using the Principia’s own system, and it’s own syntax, that not only was the principia itself flawed, but that any possible effort like the principia would inevitably be flawed!

With the incompleteness proof, Gödel showed that even in the Principia, even with all of the effort that it made to strictly separate the levels of reasoning, that he could form self-referential statements, and that those self-referential statements were both true and unprovable.

The way that he did it was simply brilliant. The proof was a sequence of steps.

  1. He showed that using Peano arithmetic – that is, the basic definition of natural numbers and natural number arithmetic – that you could take any principia-logic statement, and uniquely encode it as a number – so that every logical statement was a number, and ever number was a specific logical statement.
  2. Then using that basic mechanic, he showed how you could take any property defined by a predicate in the principia’s logic, and encode it as a arithmetic property of the numbers. So a number encoded a statement, and the property of a number could be encoded arithmetically. A number, then, could be both a statement, and a definition of an arithmetic property of a stament, and a logical description of a logical property of a statement – all at once!
  3. Using that encoding, then – which can be formed for any logic that can express Peano arithmetic – he showed that you could form a self-referential statement: a number that was a statement about numbers including the number that was statement itself. And more, it could encode a meta-property of the statement in a way that was both true, and also unprovable: he showed how to create a logical property “There is no proof of this statement”, which applied to its own numeric encoding. So the statement said, about itself, that it couldn’t be proven.

The existence of that statement meant that the Principia – and any similar system! – was incomplete. Completeness means that every true statement is provably true within the system. But the statement encodes the fact that it cannot be proven. If you could prove it, the system would be inconsistent. If you can’t, it’s consistent, but incomplete.

We’re going to go through all of that in detail.

Types and Lambda Calculus

Most programmers are familiar with λ-calculus. It’s one of the most widely used tools in theoretical computer science and logic. It’s the basis of many programming languages, including Lisp, Haskell, ML, and Scala. It’s had a strong influence on many other programming languages that aren’t strictly based on it, including Java, C++, and Javascript.

Motivation: the Problem

Modern programmers love the λ-calculus. A lot of the mathematicians who worked in what became computer science also loved it. But if you go back to its roots, there was a problem.

The λ-calculus in its non-typed form, as it was originally defined, was inconsistent.

Haskell Curry eventually reduced the paradox to something quite simple:

 r = (\lambda x. ((x x) \Rightarrow y))

That is a function which returns y if (x x) is true.

To see the paradox, we need to change how we think of λ-calculus. Most people today who know λ-calculus are programmers, and we think of it primarily as something like a programming language. But the roots of λ-calculus were laid before computers as we know them existed. At the time, λ-calculus was a tool used by logicians. To them it wasn’t a computing machine, it was a logical model of computation.

When we look at the expression (r r) \Rightarrow y, what we see is executable logic that reads as “If applying r to r returns true, then return y”. And by using computational reasoning, we’d conclude that (r r) is a non-terminating computation, because to get the value of (r r), we need to evaluate (r r), and then apply r to that result. But to logicians like Haskell Curry, it read is “The statement (r r) implies y”. Rendered into simple english, it’s a statement like: “If this statement is true, then my hair is purple”. It’s purely a logical implication, and so even though we can’t actually evaluate (r r), in logical terms, we can still say “This is a well-formed logical statement, so is it true?”.

Is that true? Suppose it is. If it’s true, then that means that it says that whatever y is must be true. Without knowing what y is, we can’t be sure if this is a true statement.

Suppose that it’s false. If (r r) is false, that means that the implication (r r) must be true, because FOPL says that if the antecedent of an implication is false, the entire implication is true.

It’s clearer when you look at the english: If this sentence is true, then my hair is purple”. My hair isn’t purple, which means that the statement can’t be true. But if the statement isn’t true, then the implication is true, so the statement is true. We’re caught in a self-reference loop, just like what we saw in Russell’s paradox.

This was considered a very bad thing. It’s a bit subtler than the problem with naive set theory. In set theory, we had an unambiguous inconsistency: we had a set that was well-defined under the axioms of set theory, and without any question, that set was inconsistent. Here, we’ve got an expression which might be consistent, and then again, it might not. It depends on what the consequent – the “my hair is purple” part – of the implication is. If it’s something true, then we’re fine. If it’s not, then we’re stuck.

The problem is, no matter what you put into that “my hair is purple” part, you can use this to produce a proof that it’s true, even if it isn’t. And that’s a fundamental inconsistency, which can’t be tolerated.

Curry’s paradox meant that the logicians working with λ-calculus had a problem that they needed to solve. And following the fashion of the time, they solved it using types. Much like ST type theory attempted to preserve as much of set theory as possible while fixing the inconsistency, typed λ-calculus tried to fix the inconsistency of λ-calculus. The basic approach was also similar: the typed λ-calculus introduced a stratification which made it impossible to build the structures that led to inconsistency.

In ST, the solution was to build a logical system in which it was impossible to express self-reference. In λ-calculus, the basic goal was the same: to eliminate self-reference. But in λ-calculus, that restriction is stated differently. What the type system in λ-calculus does is make it impossible to construct a statement that is a fixed-point combinator.

If we look carefully at the paradoxical statement up above, it’s not really pure λ-calculus. It relies on the fact that we’re defining a function named r, and then applying r to itself using the name r. But that’s just syntactic sugar: in fact, r can’t reference r. You need some tool to let you apply an expression to itself: that tool is called a fixed-point combinator. The most common fixed-point in λ-calculus is the Y combinator, which underlies how recursive computations usually work in λ-calculus.

The way that the simply typed λ-calculuss gets around the Curry paradox is by making it impossible to build a well-typed fixed-point combinator. Without that, you can’t build the self-referential constructs that cause the inconsistency. The downside is that the simply typed λ-calculus, without a fixed point combinator, is not Turing complete. The evaluation of every simple typed λ-calculus expression will eventually terminate.

(As an aside: this isn’t really a problem in practice. The self-referential expressions that cause the Curry paradox turn into non-terminating computations. So they don’t produce a paradox; they just don’t produce anything. Logical inconsistencies don’t produce results: they’re still an error, instead of terminating with an inconsistent result, they just never terminate. Again, to the logicians at the time, the idea of non-termination was, itself, a deep problem that needed a solution.)

The Solution: Stratification by Types

The way that the simply typed λ-calculus fixed things was by creating a stratification using types. The type system created a restriction on the set of statements that were valid, well-formed logical statements. That restriction made it impossible to express a fixed point combinator or a general recursive computation of any kind.

It’s helpful to once again refer back to ST. In ST, we started the type of atoms at level 0. To get to level 1, we defined predicates over level-0 objects, and the set of objects that matched the predicate was a level-1 type. Then we could define predicates over level-1 objects, and the set of level-1 types that satisfied the predicate was a level-2 type, and so on. In the simply typed λ-calculus, we do the same thing, but with functions: we can build functions that operate on primitive atoms (also called base values), or on other functions. When we define a function, it must be assigned a type, and that type must be something at a lower level than the function being defined. You can’t ever pass a function to itself, because the function is an object at a higher level than the type of its parameter.

We start with base types. Every simply-typed lambda calculus starts with a collection of primitive atomic values. The set of atomic values is partitioned into a collection of types, which are called the base types. Base types are usually named by single lower-case greek letters: So, for example, we could have a type \sigma which consists of the set of natural numbers; a type \tau which corresponds to boolean true/false values; and a type \gamma which corresponds to strings.

Once we have basic types, then we can talk about the type of a function. A function maps from a value of one type (the type of parameter) to a value of a second type (the type of the return value). For a function that takes a parameter of type \gamma, and returns a value of type \delta, we write its type as “\gamma \rightarrow \delta“. The “\rightarrow” is called the function type constructor; it associates to the right, so \gamma \rightarrow \delta \rightarrow \epsilon is equivalent to \gamma \rightarrow (\delta \rightarrow \epsilon).

In every function declaration, we need to specify the type of the parameters and the type of the result. So:

  1. \sigma \rightarrow \sigma is a function which takes natural number as a parameter, and returns a natural number as a result.
  2. (\sigma \rightarrow \sigma) \rightarrow \sigma is a function which takes a \sigma \rightarrow \sigma function as a parameter, and produces a natural number as a result.

As usual in λ-calculus, we don’t have multi-parameter functions – all functions are curried, so a function like addNatural would be a function that takes a natural number as a paramater, and returns a function that takes a natural number and returns a natural number. So the type of addNatural is \sigma \rightarrow \sigma \rightarrow \sigma.

How does this get around the self-reference problem? A function like the one in the Curry paradox takes an object of its own type as a parameter. There’s no way to write that in a type system. It’s a significant restriction which makes it impossible to write general recursive expressions – it limits us to something close to primitive recursion, but it avoids the inconsistency. All valid expressions written with this system of types in place is guaranteed to terminate with a consistent result.

Extending λ-calculus with types

Now, it’s time to get at least a little bit formal, to see how we integrate a stratified type system into the lambda calculus. There’s two facets to that: the syntactic, and the analytic. The syntactic part shows how we extend λ-calculus to include type declarations, and the analytic part shows how to determine whether or not an expression with type declarations is valid.

The syntax part is easy. We add a “:” to the notation; the colon has an expression or variable binding on its left, and a type specification on its right. It asserts that whatever is on the left side of the colon has the type specified on the right side. A few examples:

\lambda x: \nu . x + 3
This asserts that the parameter, x has type nu, which we’ll use as the type name for the natural numbers. (In case it’s hard to tell, that’s a greek letter “nu” for natural.) There is no assertion of the type of the result of the function; but since we know that “+” is a function with type \nu  \rightarrow \nu \rightarrow \nu, we can infer that the result type of this function will be \nu.
(\lambda x . x + 3): \nu \rightarrow \nu
This is the same as the previous, but with the type declaration moved out, so that it asserts the type for the lambda expression as a whole. This time we can infer that x : \nu because the function type is \nu \rightarrow \nu, which means that the function parameter has type \nu.
\lambda x: \nu, y:\delta . \text{if}\hspace{1ex} y\hspace{1ex} \text{then}\hspace{1ex} x * x \hspace{1ex}\text{else} \hspace{1ex} x
This is a two parameter function; the first parameter has type ν, and the second has type δ. We can infer the return type, which is ν. So the type of the full function is ν → δ → ν. This may seem surprising at first; but remember that λ-calculus really works in terms of single parameter functions; multi-parameter functions are a shorthand for currying. So really, this function is: λ x : ν . (λ y : δ . if y then x * x else x); the inner lambda is type δ → ν; the outer lambda is type ν → (δ → ν).

To talk about whether a program is valid with respect to types (aka well-typed), we need to introduce a set of rules for checking the validity of the type declarations. Using those rules, we can verify that the program is type-consistent.

In type analysis, we’ll talk about judgements. When we can infer the type of an expression using an inference rule, we call that inference a type judgement. Type analysis allows us to use inference and judgements to reason about types in a lambda expression. If any part of an expression winds up with an inconsistent type judgement, then the expression is invalid. If we can show that all of the components of an expression have consistent type judgements, then we can conclude that the expression is well-typed, meaning that it’s valid with respect to the type system.

Type judgements are usually written in a sequent-based notation, which looks like a fraction where the numerator consists of statements that we know to be true; and the denominator is what we can infer from the numerator. In the numerator, we normally have statements using a context, which is a set of type judgements that we already know;it’s usually written as an uppercase greek letter. If a type context includes the judgement that x : \alpha, I’ll write that as \Gamma :- x : \alpha.

Rule 1: Type Identity

\frac{\mbox{}}{x : \alpha :- x: \alpha}

This is the simplest rule: if we have no other information except a declaration of the type of a variable, then we know that that variable has the type it was declared with.

Rule 2: Type Invariance

\frac{ \Gamma :- x:\alpha, x != y}{\Gamma + y:\beta :- x:\alpha}

This rule is a statement of non-interference. If we know that x:\alpha, then inferring a type judgement about any other term cannot change our type judgement for x.

Rule 3: Function Type Inference

\frac{\Gamma + x:\alpha :- y:\beta}{\Gamma :- (\lambda x:\alpha . y):\alpha \rightarrow \beta}

This statement allows us to infer function types given parameter types. Ff we know the type of the parameter to a function is \alpha; and if, with our knowledge of the parameter type, we know that the type of term that makes up the body of the function is \beta, then we know that the type of the function is \alpha \rightarrow \beta.

Rule 4: Function Application Inference

\frac{\Gamma :- x: \alpha \rightarrow \beta, \Gamma :- y:\alpha}{\Gamma :- (x y): \beta}

This one is easy: if we know that we have a function that takes a parameter of type \alpha and returns a value of type \beta, then if we apply that function to a value of type \alpha, we’ll get a value of type \beta.

These four rules are it. If we can take a lambda expression, and come up with a consistent set of type judgements for every term in the expression, then the expression is well-typed. If not, then the expression is invalid.

So let’s try taking a look at a simple λ-calculus expression, and seeing how inference works on it.

\lambda x y . y x

Without any type declarations or parameters, we don’t know the exact type. But we do know that “x” has some type; we’ll call that “α”; and we know that “y” is a function that will be applied with “x” as a parameter, so it’s got parameter type α, but its result type is unknown. So using type variables, we can say “x:α,y:α→β”. We can figure out what “α” and “β” are by looking at a complete expression. So, let’s work out the typing of it with x=”3″, and y=”λ a:ν.a*a”. We’ll assume that our type context already includes “*” as a function of type “ν→ν→ν”, where ν is the type of natural numbers.

  • “λ x y . y x) 3 (λ a:ν . a * a)”: Since 3 is a literal integer, we know its type: 3:ν.
  • By rule 4, we can infer that the type of the expression “a*a” where “a:ν” is “ν”, and *:ν→ν→ν so therefore, by rule 3 the lambda expression has type “ν→ν”. So with type labelling, our expression is now: “(λ x y . y x) (3:ν) (λ a:ν.(a*a):ν) : ν→ν”.
  • So – now, we know that the parameter “x” of the first lambda must be “ν”; and “y” must be “ν→ν”; so by rule 4, we know that the type of the application expression “y x” must be “ν”; and then by rule 3, the lambda has type: “ν→(ν→ν)→ν”.
  • So, for this one, both α and β end up being “ν”, the type of natural numbers.

So, now we have a simply typed λ-calculus. The reason that it’s simply typed is because the type treatment here is minimal: the only way of building new types is through the unavoidable \rightarrow constructor. Other typed lambda calculi include the ability to define parametric types, which are types expressed as functions ranging over types.

Programs are Proofs

Now we can get to the fun part. The mantra of type theory is the program is the proof. Here’s where we get our first glimpse of just what that really means!

Think about the types in the simple typed language calculus. Anything which can be formed from the following grammar is a λ-calculus type:

  • type ::= primitive | function | ( type )
  • primitive ::= α | β | δ | …
  • function ::= type→type

The catch with that grammar is that you can create type expressions which, while they are valid type definitions, you can’t write a single, complete, closed expression which will actually have that type. (A closed expression is one with no free variables.) When there is an expression that has a type, we say that the expression inhabits the type; and that the type is an inhabited type. If there is no expression that can inhabit a type, we say it’s uninhabitable. Any expression which either can’t be typed using the inference rules, or which is typed to an uninhabitable type is a type error.

So what’s the difference between inhabitable type, and an uninhabitable type?

The answer comes from something called the Curry-Howard isomorphism. For a typed λ-calculus, there is a corresponding intuitionistic logic. A type expression is inhabitable if and only if the type is a provable theorem in the corresponding logic.

The type inference rules in λ-calculus are, in fact, the same as logical inference rules in intuitionistic logic. A type \alpha \rightarrow \beta can be seen both as a statement that this is a function that maps from a value of type \alpha to a value of type \beta, and as a logical statement that if we’re given a fact \alpha, we could use that to infer the truth of a fact \beta.

If there’s a logical inference chain from an axiom (a given type assignment) to an inferred statement, then the inferred statement is an inhabitable type. If we have a type \alpha \rightarrow \alpha, then given a inhabited type \alpha, we know that \alpha \rightarrow \alpha is inhabitable, because if \alpha is a fact, then \alpha \rightarrow \alpha is also a fact.

On the other hand, think of a different case \alpha \rightarrow \beta. That’s not a theorem, unless there’s some other context that proves it. As a function type, that’s the type of a function which, without including any context of any kind, can take a parameter of type α, and return a value of a different type β. You can’t do that – there’s got to be some context which provides a value of type β – and to access the context, there’s got to be something to allow the function to access its context: a free variable. Same thing in the logic and the λ-calculus: you need some kind of context to establish “α→β” as a theorem (in the logic) or as an inhabitable type (in the λ-calculus).

What kind of context would make a type \alpha \rightarrow \beta inhabitable? A definition of a valid function that takes an α, and returns a β. If such a function exists, then that function is a proof of the inhabitility of the type. Literally, the program is the proof.

The Constructivist Roots of Type Theory

This post started off as an introduction to a post about the simply typed lambda calculus. It got a bit out of control – but reading it over, I think that it’s valuable background. So I’ve made it into its own post.

ST type theory is an interesting idea, but it’s mostly interesting as a historical curiosity and background. Type theory started with things like ST, but it turned into something very different. You can see the roots in ST, but modern type theory takes a completely different approach. ST is an axiomatic foundation for a stratified version of type theory. Modern type theory, in contrast, has constructed a universe of mathematics that, rather than being based on an axiomatic foundation of first order predicate logic, is based almost entirely on ideas of computability and computation.

The core of the modern version of type theory is the intuitionistic type theory of Per Martin-Lof, and it’s part of the constructionist school of math.

Like ST type theory, constructivism can also been seen as a reaction to paradox and inconsistency. Some mathematicians decided that the right way to approach the problem of paradox was to build a better, stronger axiomatic foundation. That effort let to things like ST type theory, ZFG/NBG set theory, and the principia mathematica. But others thought that that approach was wrong.

The axiomatic approach ends up with things like the axiom of choice in its foundation. And the axiom of choice leads directly to results that seem to be completely nonsensical. For example, the classic Banach-Tarski paradox.. Banach-Tarski says that there’s a way of taking a sphere the size of an orange, cutting it into some very odd slices, and then reassembling those slices into two oranges exactly the same size as the original. On the face of it, that’s ridiculous!

The key to Banach-Tarski is that it’s built on an operation that’s impossible in reality. It’s exploiting the boundaries of a theory. You start with a sphere. That’s an object on which metrics like “volume” are well-defined. You’re “cutting” it – not really cutting, but actually subdividing it – into pieces. Those pieces are objects where the metrics no longer work. You’ve broken the metric. And then, you’ve got objects for which the entire concept of volume no longer makes any sense, and you reassemble them into new objects – and those new objects are, once again, objects where the volume metric works.

From one point of view, you’ve done something that makes sense. You started with a topological space with meaningful metrics. A topological space is just a set with a bunch of cute structural properties – the neighborhood relationships. Those structural properties in turn have higher order structure properties, which can be used to define metrics like distance and volume. So the original sphere has a meaningful volume. But still, mathematically, it’s just a set with neighborhood relationships.

Since it’s just a topological space, you can divide it into sub-spaces – that is, subsets which retain structural properties. They’re still topological spaces, but now, the neighborhood relations, because of the way that you divided the original set, no longer have the properties that you need to define metrics. They’re not metric spaces.

When you recombine them, you can combine them into two new topological spaces that do have the properties that you need to define metrics – so volumes are, once again, meaningful. But in that process, you’re redefining the metrics, and so they can be redefined in any way that you want, producing any results that you want.

There’s a pretty strong argument that that entire process is an abuse of the underlying axioms. It only works because you can destroy the metric, and then recreate it. When you state what’s going on clearly, it seems like it’s cheating. But according to the standard axiomatic set-theory based math, everything that happens is completely legal. It’s ridiculous and non-sensical, and it produces a result that looks like it must be obviously wrong, but according to the axioms, it’s just fine.

The problem of cheating on the metrics is bad, but it’s not all that’s wrong with Banach-Tarski according to the constructivists. B-T says that you can cut a spherical metric space into pieces. Only you really can’t – it’s all just a clever sleight-of-hand! You can’t describe or specify how to cut the original sphere into pieces. There is no shape, no structure, no process that you can use to actually cut the sphere. The axioms say that such a cut exists, but the actual method of cutting is not just unknown, but essentially unknowable.

The constructivists argue that that’s rubbish. How can you argue that it exists, but we can’t produce it, we can’t find it, we can’t study it. According to the constructivists, it exists only as an artifact of a flawed system. It’s meaningless to say that something exists if it only exists in an unreachable theoretical ether.

The constructivists say that things like Banach-Tarski and the Russell paradox are part of the same basic problem: defining things not by how they can be created, but by how they can be described using some fuzzy airy-fairy gibberish derived from axioms. The solution isn’t to find a better set of axioms, but to find a fundamentally different way of doing things.

The constructivist method says that the way to do math is to build things. That the only way you can really consider a statement to be true is if you have concrete proof: the only way to prove that something exists is to show a concrete example of it. The only way to prove that something is false is with a specific counterexample.

That leads to an absolutely fascinatingly different way of doing math, which is what we’ll see in Martin-Lof’s intuitionistic type theory. When you talk about typed programming languages, you’ll hear a mantra: “The program is the proof”. That is literally true. In Martin-Lof’s intuitionistic type theory, the way that you prove things is by, essentially, writing programs! A proof of a fact is a computation. A proof of an existential statement isn’t just some abstract reasoning that shows that it exists – it’s a computation that produces a concrete example. A proof of a universal statement is a metaproof – a program that, given any value, produces a proof specific to that value.

It avoids the basic problem with the axiomatic problem. You can’t have something ridiculous like the set of all sets that don’t include themselves, because the only way to show that a set exists is to show a process to create it – and there’s no way to create a paradoxical set! You don’t get nonsense like Banch-Tarski, because the division of the spherical topological space into the non-metric subspaces is impossible: you can’t show a process that produces it. In the world of intuitionism and constructivism, existence proofs that don’t produce concrete examples don’t exist!

It’s an impressive thing, and it’s really a mind-blowingly different way of understanding what it is to do mathematics.

For introducing Martin-Lof’s type theory, I’m going to be mostly working from the Nordstrom/Petersson/Smith text, with supplementation from many different sources. For the programming language oriented part, I highly recommend Types and Programming Languages by Benjamin Pierce.

Before we really get started, I’m going to back and do some review. I’m going to go start with an introduction to the simply typed lambda calculus, and intuitionistic logic, because they’re both really essentials for understanding what’s coming next. So that will be the next couple of posts.

Administrivia

Quick note: I’m experimenting with showing ads on the blog. In my time with scientopia, I dumped a frankly ridiculous amount of money into keeping things online – it ended up coming to around $4K/year. Obviously, this site is a lot cheaper, since I only need to support the one blog, but still, I’d like to get to the point where it’s self-supporting.

If the ads are annoying, let me know. I’m trying to strike a compromise. I’d like to get some money coming in to cover hosting costs, but I don’t want to degrade the experience of my readers. The way I’ve set it up, the ads shouldn’t have popups, and they shouldn’t have any interactive animations. If you see anything like that, please let me know as soon as possible, and I’ll do my best to get it fixed. At the moment, the visual appearance of the ads is less than idea, but I’ll be tweaking it until it’s reasonable. (For now, things are still in process with Google, so all you see is a big yellowish box as a placeholder; as soon as someone at Google gets around to approving it, that should be replaced by ads.

If ads are offensive, again, let me know. I don’t want to be giving exposure to blatantly offensive rubbish.

If ads are pseudoscience or that sort of thing, I don’t care. I believe that the people who read this blog are intelligent enough to realize what a sham that crap is, and I have no problem taking money from them – every cent they waste buying ads on my blog is a cent that they’re not using for something a lot worse!

The beauty of math; the humor of stupidity.

%d bloggers like this: