HoTT for Dummies

Posted on by Chris Warburton

Based on a Hacker News thread, about a talk by Thorsten Altenkirch, I ended up writing a rather long explanation of some Type Theory and Homotopy Type Theory, which I thought I might as well reproduce here:

Let’s say you’re creating a new programming language. You think it’s a good idea to use static types, but which ones? Well, everyone uses booleans, so let’s include them:

True : Boolean
False : Boolean

(Foo : Bar means “Foo has type Bar”, or “Foo is a Bar”). OK, that seemed pretty easy. But wait, you’ve not given a type to Boolean. For the sake of completeness:

Boolean : Type

Uh oh, now you need to give a type to Type. What should it be? It turns out (via Girard’s paradox) that simply saying Type : Type would make things inconsistent, ie. we would be able to trick the compiler into accepting incorrect programs.

Instead, we use a series of “levels”:

Boolean : Type 0

Type n : Type (n+1)

So far so good. Now let’s say we want function types, for example:

identity : Boolean -> Boolean
identity x = x

not : Boolean -> Boolean
not True  = False
not False = True

So what’s this -> thing? We can think of it as a type-level operator: it takes two types and returns a function type. In the syntax of natural deduction, we can say:

a : Type n         b : Type m
-----------------------------
 a -> b : Type (1 + max n m)

ie. given a of type Type n, and b of type Type m, then a -> b has type Type (1 + max n m). Because the type a -> b somehow ‘contains’ the types a and b, we need to ensure it’s at a higher level than either of them, which is why we do 1 + max n m.

In fact, there’s no reason for the identity function to only work on Booleans. We can replace it with an “identity function factory”, which accepts a type and returns an identity function for that type:

identity : (t : Type n) -> t -> t
identity x y = y

Here we’ve re-used the foo : bar notation: rather than just giving the type of the first argument, we’ve also introduced a variable t representing its value (this is known as a dependent function). Notice that the definition of identity actually ignores the type it’s been given (x); the implementation doesn’t care what it is, it’ll just return the second argument no matter what; yet we need that argument in order to type-check. When we compile this program, we can “erase” the first argument, since it has no “computational content”.

We can recover our old identity function, of type Boolean -> Boolean, by applying this new identity function to the Boolean type:

identity Boolean : Boolean -> Boolean

OK, what next? Well, since we have functions, we might as well have function composition:

compose : (a : Type x) -> (b : Type y) -> (c : Type z) -> (b -> c) -> (a -> b) -> a -> c
compose t1 t2 t3 g f x = f (g x)

It would also be useful to have equality. Here it is for Booleans:

equal : Boolean -> Boolean -> Boolean
equal True  x = x
equal False x = not x

However, just like the identity function, this isn’t very satisfying. We’d like an “equality function factory”, with this type:

equal : (t : Type n) -> t -> t -> Boolean

Except, how would we ever implement such a “factory”? It was easy for identity: we just return whatever we’re given. In the case of equality, we need to inspect our arguments, to see whether they’re actually equal or not. We can’t do this in a way which works for all types (eg. what if we allow user-defined types?).

However, there’s a trick. By returning a Boolean, we’re defining equality (True) and disequality (False). That’s hard. Instead, we can ignore the disequality, and only focus on equality, using a different return type; let’s call it Equal x y.

What does it mean for two things to be equal? It means that they’re the same thing. In which case, we don’t need both of them! Every value is equal to itself (a property known as “reflexivity”), so that’s all we need!

refl : (t : Type n) -> (x : t) -> Equal x x

For example, here’s equality for the Booleans:

refl Boolean True  : Equal True  True
refl Boolean False : Equal False False

Note that refl isn’t actually a function, it’s a data constructor. You can think of a value like refl Boolean True as being a piece of data, similar to something like pair Int String 10 'foo'; it doesn’t reduce to anything, it just gets passed around as-is. (These are often called “proof objects”, but that’s a bit arbitrary; a value like pair Int String 10 'foo' is a “proof” of “Int AND String”).

If we allow computation in our types, then two different values which compute (technically: beta reduce) to the same thing are still equal by reflexivity:

refl Boolean True : Equal True (not False)

Here, the not False will compute to True, and refl will type-check. Different values, eg. False and True, are never equal, since they don’t reduce to the same thing. We can make our computations as complex as we like, for example:

refl Boolean True : Equal (identity Boolean True) (compose Boolean Boolean Boolean not not True)

We can even have equality between functions and equality between types:

-- "identity" is equal to "identity"
refl ((t : Type n) -> t -> t) identity : Equal identity identity

-- "Boolean" is equal to "Boolean"
refl (Type 0) Boolean : Equal Boolean Boolean

We can even have equalities between equalities!

-- "refl Boolean True" is equal to "refl Boolean True"
refl (Equal True True) (refl Boolean True) (refl Boolean True) : Equal (refl Boolean True) (refl Boolean True)

Most of this predates Homotopy Type Theory, so what are the points being made in the slides?

One point is to give a topological perspective for types: a type is like a space, values in the type are like points in the space. Equalities between values are paths in the space (eg. refl Boolean True is a trivial path from the point True to itself, in the Boolean space). Interestingly, equalities between equalities are homotopies (smooth transformations between paths).

One question we might ask is whether all equality values are the same; ie. are they all just refl? That’s known as the “Uniqueness of Identity Proofs” (UIP), and it’s an assumption that many people have been making for decades. However, if we think of equalities as paths through a space, then UIP says that all those paths can be transformed into each other. Yet that’s not the case if the space contains a hole! Consider two paths going from a point X back to itself; if one of those paths loops around a hole, and the other doesn’t, then there’s no way to smoothly transform between the two (without “cutting and sticking”):

Type containing a hole

The topological perspective also gives us some intuition about the “levels” of types: Type 0 contains spaces with distinct points, eg. Boolean containing True and False. These are essentially sets, from Set Theory. Although HoTT doesn’t assume UIP for all types, those which do just-so-happen to “collapse” down to one value (ie. there are equalities between every point) actually occupy a level below sets; ie. they end up at Type -1 (there’s no significance to the negative number; it’s just a historical accident caused by definitions like Boolean : Type 0). Likewise, those which add more structure occupy higher levels.

One important question is how function equality behaves. It’s useful to have equality for, say, Booleans, since we can compute their value and check whether they’re the same. Functions are trickier: we have intensional equality (eg. Equal identity identity) but we’d like extensional equality (eg. that identity and not . not are equal, ie. Equal (identity Boolean) (compose Boolean Boolean Boolean not not)). This is tricky. If we assume extensionality as an axiom (ie. we assume a built-in function of type ((x : t) -> Equal (f x) (g x)) -> Equal f g), like in NuPRL, then we lose the ability to execute our programs (axioms are basically “built in primitives”; we don’t know how to implement such an extensionality primitive). Observational Type Theory gets us most of the way there, but relies on UIP, which would collapse all of the higher-level structure in HoTT.

Univalence is also a nice feature of HoTT. It’s incompatible with UIP, but allows us to reason “up to isomorphism”. For example, in set theory the sets {foo, bar} and {x, y} are isomorphic: eg. we can switch foo with x and bar with y. However, results concerning {foo, bar} may be invalid for {x, y}, since set theory lets us say things like “is foo a member of S?”, which is true when S = {foo, bar} but false when S = {x, y}. HoTT doesn’t let us say things like “is foo a member of type T?”, hence we don’t get these kind of “abstraction leaks”, so our programs and proofs can be automatically lifted from one representation to another. For example, if you define a fast, distributed datastructure which is isomorphic to a linked-list, then you can automatically lift every linked-list program to your new datastructure using univalence. Unfortunately, we haven’t figured out how to implement univalence yet (but many people think it’s possible).

Some of the results listed at the end of the slides concern the use of HoTT to prove topological results, eg. that the integers are isomorphic to paths around a hole (you can go around the hole any number of times, clockwise or anticlockwise).