Continuations are a criminally underappreciated language feature. Very few languages (off the top of my head: most Scheme dialects, some Standard ML implementations, and Ruby) support the---already very expressive---undelimited continuations---the kind introduced by `call/cc`{.scheme}---and even fewer implement the more expressive delimited continuations. Continuations (and tail recursion) can, in a first-class, functional way, express _all_ local and non-local control features, and are an integral part of efficient implementations of algebraic effects systems, both as language features and (most importantly) as libraries.
Continuations are a criminally underappreciated language feature. Very few languages (off the top of my head: most Scheme dialects, some Standard ML implementations, and Ruby) support the---already very expressive---undelimited continuations---the kind introduced by `call/cc`{.scheme}---and even fewer implement the more expressive delimited continuations. Continuations (and tail recursion) can, in a first-class, functional way, express _all_ local and non-local control features, and are an integral part of efficient implementations of algebraic effects systems, both as language features and (most importantly) as libraries.
<divstyle="display: flex">
<divclass="text-image">
<div>
<div>
Continuations, however, are notoriously hard to understand. In an informal way, we can say that a continuation is a first-class representation of the "future" of a computation. By this I do not mean a future in the sense of an asynchronous computation, but in a temporal sense. While descriptions like this are "correct" in some sense, they're also not useful. What does it mean to store the "future" of a computation in a value?
Continuations, however, are notoriously hard to understand. In an informal way, we can say that a continuation is a first-class representation of the "future" of a computation. By this I do not mean a future in the sense of an asynchronous computation, but in a temporal sense. While descriptions like this are "correct" in some sense, they're also not useful. What does it mean to store the "future" of a computation in a value?
@ -14,7 +14,7 @@ Operationally, we can model continuations as just a segment of control stack. Th
With no offense to the authors of these articles, some of which I greatly respect as programming language designers and implementers, I do not think that this approach is very productive in a functional programming context. This reductionist, imperative view would almost be like explaining the _concept_ of a proper tail call by using a trampoline, rather than presenting trampolines as one particular implementation strategy for the tail-call-ly challenged.
With no offense to the authors of these articles, some of which I greatly respect as programming language designers and implementers, I do not think that this approach is very productive in a functional programming context. This reductionist, imperative view would almost be like explaining the _concept_ of a proper tail call by using a trampoline, rather than presenting trampolines as one particular implementation strategy for the tail-call-ly challenged.
title: "This Sentence is False, or: On Natural Language, Typing and Proof"
date: September 9th, 2020
---
The Liar's paradox is often the first paradox someone dealing with logic, even in an informal setting, encounters. It is _intuitively_ paradoxical: how can a sentence be both true, and false? This contradicts (ahem) the law of non-contradiction, that states that "no proposition is both true and false", or, symbolically, $\neg (A \land \neg A)$. Appealing to symbols like that gives us warm fuzzy feelings, because, _of course, the algebra doesn't lie!_
There's a problem with that the appeal to symbols, though. And it's nothing to do with non-contradiction: It's to do with well-formedness. How do you accurately translate the "this sentence is false" sentence into a logical formula? We can try by giving it a name, say $L$ (for liar), and state that $L$ must represent some logical formula. Note that the equality symbol $=$ here is _not_ a member of the logic we're using to express $L$, it's a symbol of this discourse. It's _meta_logical.
$$ L = \dots $$
But what should fill in the dots? $L$ is the sentence we're symbolising, so "this sentence" must mean $L$. Saying "X is false" can be notated in a couple of equivalent ways, such as $\neg X$ or $X \to \bot$. We'll go with the latter: it's a surprise tool that will help us later. Now we know how to fill in the dots: It's $L \to \bot$.
<details>
<summary>Truth tables demonstrating the equivalence between $\neg A$ and $A \to \bot$, if you are classically inclined.</summary>
<divclass="mathpar">
<table>
<tr>
<th> $A$ </th>
<th> $\neg A$ </th>
</tr>
<tr><td>$\top$</td><td>$\bot$</td></tr>
<tr><td>$\bot$</td><td>$\top$</td></tr>
</table>
<table>
<tr>
<th> $A$ </th>
<th> $A\to\bot$ </th>
</tr>
<tr><td>$\top$</td><td>$\bot$</td></tr>
<tr><td>$\bot$</td><td>$\top$</td></tr>
</table>
</div>
</details>
But wait. If $L = L \to \bot$, then $L = (L \to \bot) \to \bot$, and also $L = ((L \to \bot) \to \bot) \to \bot$, and so... forever. There is no finite, well-formed formula of first-order logic that represents the sentence "This sentence is false", thus, assigning a truth value to it is meaningless: Saying "This sentence is false" is true is just as valid as saying that it's false, both of those are as valid as saying "$\neg$ is true".
Wait some more, though: we're not done. It's known, by the [Curry-Howard isomorphism], that logical systems correspond to type systems. Therefore, if we can find a type-system that assigns a meaning to our sentence $L$, then there _must_ exist a logical system that can express $L$, and so, we can decide its truth!
Even better, we don't need to analyse the truth of $L$ logically, we can do it type-theoretically: if we can build an inhabitant of $L$, then it is true; If we can build an inhabitant of $\neg L$, then it's false; And otherwise, I'm just not smart enough to do it.
So what is the smallest type system that lets us assign a meaning to $L$?
# A system of equirecursive types: $\lambda_{\text{oh no}}$[^1]
[^1]: The reason for the name will become obvious soon enough.
We do not need a complex type system to express $L$: a simple extension over the basic simply-typed lambda calculus $\lambda_{\to}$ will suffice. No fancy higher-ranked or dependent types here, sorry!
As a refresher, the simply-typed lambda calculus has _only_:
* A set of base types $\mathbb{B}$,
* Function types $\tau \to \sigma$,
* For each base type $b \in \mathbb{B}$, a set of base terms $\mathbb{T}_b$,
* Variables $v$,
* Lambda abstractions $\lambda v. e$, and
* Application $e\ e'$.
<details>
<summary>Type assignment rules for the basic $\lambda_{\to}$ calculus.</summary>
<divclass="math-paragraph">
<div>
$$\frac{x : \tau \in \Gamma}{\Gamma \vdash x : \tau}$$
</div>
<div>
$$\frac{b \in \mathbb{B} \quad x \in \mathbb{T}_{b}}{\Gamma \vdash x : b}$$
</div>
<div>
$$\frac{\Gamma, x : \sigma \vdash e : \tau}{\Gamma \vdash \lambda x. e : \sigma \to \tau}$$
</div>
<div>
$$\frac{\Gamma, e : \sigma \to \tau \quad \Gamma \vdash e' : \sigma}{\Gamma \vdash e\ e' : \tau}$$
</div>
</div>
</details>
First of all, we'll need a type to represent the logical proposition $\bot$. This type is empty: It has no type formers. Its elimination rule corresponds to the principle of explosion, and we write it $\mathtt{absurd}$. The inference rule:
<divclass="math-paragraph">
$$\frac{\Gamma \vdash e : \bot}{\mathtt{absurd}\ e : A}$$
</div>
We're almost there. What we need now is a type former that serves as a solution for equations of the form $v = ... v ...$. That's right: we're just _inventing_ a solution to this class of equations---maths!
These are the _equirecursive_ types, $\mu a. \tau$. The important part here is _equi_: these types are entirely indistinguishable from their unrollings. Formally, we extend the set of type formers with type variables $a$ and $\mu$-types $\mu a. \tau$, where $\mu a$ acts as a binder for $a$.
Since we invented $\mu$ types as a solution for equations of the form $a = \tau$, we have that $\mu a. \tau = \tau[\mu a.\tau/a]$, where $\tau[\sigma{}/a]$ means "substitute $\sigma{}$ everywhere $a$ occurs in $\tau$". The typing rules express this identity, saying that anywhere a term might have one as a type, the other works too:
<divclass="math-paragraph">
<div>
$$\frac{\Gamma \vdash e : \tau[\mu a.\tau / a]}{\Gamma \vdash e : \mu a. \tau}$$
</div>
<div>
$$\frac{\Gamma \vdash e : \mu a.\tau}{\Gamma \vdash e : \tau[\mu a. \tau / a]}$$
</div>
</div>
Adding these rules, along with the one for eliminating $\bot$, to the $\lambda_{\to}$ calculus nets us the system $\lambda_{\text{oh no}}$. With it, one can finally formulate a representation for our $L$-sentence: it's $\mu a. a \to \bot$.
There exists a closed term of this type, namely $\lambda k. k\ k$, which means: The "this sentence is false"-sentence is true. We can check this fact ourselves, or, more likely, use a type checker that supports equirecursive types. For example, OCaml with the `-rectypes` compiler option does.
We'll first define the empty type `void` and the type corresponding to $L$:
<divclass="math-paragraph">
~~~~{.ocaml}
type void ;;
type l = ('a -> void) as 'a ;;
~~~~
</div>
Now we can define our proof of $L$, called `yesl`, and check that it has the expected type:
<divclass="math-paragraph">
~~~~{.ocaml}
let yesl: l = fun k -> k k ;;
~~~~
</div>
However. This same function is also a proof that... $\neg L$. Check it out:
<divclass="math-paragraph">
~~~~{.ocaml}
let notl (x : l) : void = x x ;;
~~~~
</div>
# I am Bertrand Russell
Bertrand Russell (anecdotally) once proved, starting from $1 = 0$, that he was the Pope. I am also the Pope, as it turns out, since I have on hand a proof that $L$ and $\neg L$, in violation of non-contradiction; By transitivity, I am Bertrand Russell. <spanstyle="float: right; display: inline-block;"> $\blacksquare$ </span>
Alright, maybe I'm not Russell (drat). But I am, however, a trickster. I tricked you! You thought that this post was going to be about a self-referential sentence, but it was actually about typed programming language design (not very shocking, I know). It's a demonstration of how recursive types (in any form) are logically inconsistent, and of how equirecursive types _are wrong_.
The logical inconsistency, we all deal with, on a daily basis. It comes with Turing completeness, and it annoys me to no end every single time I accidentally do `let x = ... x ...`{.haskell}. I _really_ wish I had a practical, total functional programming language to use for my day-to-day programming, and this non-termination _everywhere_ is a great big blotch on Haskell's claim of purity.
The kind of recursive types you get in Haskell is _fine_. They're not _great_ if you like the propositions-as-types interpretation, since it's trivial to derive a contradiction from them, but they're good enough for programming that implementing a positivity checker to ensure your definitions are strictly inductive isn't generally worth the effort.
Unless your language claims to have "zero runtime errors", in which case, if you implement isorecursive types instead of inductive types, you are _wrong_. See: Elm. God damn it.
<details>
<summary>So much for "no runtime errors"... I guess spinning forever on the client side is acceptable.</summary>
<divclass="flex-list">
```elm
-- Elm
type Void = Void Void
type Omega = Omega (Omega -> Void)
yesl : Omega
yesl = Omega (\(Omega x) -> x (Omega x))
notl : Omega -> Void
notl (Omega x) = x (Omega x)
```
</div>
</details>
Equirecursive types, however, are a totally different beast. They are _basically_ useless. Sure, you might not have to write a couple of constructors, here and there... at the cost of _dramatically_ increasing the set of incorrect programs that your type system accepts. Suddenly, typos will compile fine, and your program will just explode at runtime (more likely: fail to terminate). Isn't this what type systems are meant to prevent?
Thankfully, very few languages implement equirecursive types. OCaml is the only one I know of, and it's gated behind a compiler flag. However, that's a footgun that should _not_ be there.
@ -95,9 +95,9 @@ The elements of $\mathrm{Prop}$ are taken to be _propositions_, not in the sense
Given $A : u$, and some $B : v$ with one variable $x : A$ free (where $u$ and $v$ are possibly distinct universes), we can form the type $\prod_{x : A} B$ of _dependent products_ from $A$ to $B$. If $v$ is in $\mathrm{Prop}$, then the dependent product also lives in $\mathrm{Prop}$. Otherwise, it lives in $\mathrm{Set}$. Moreover, we can also form the type $\sum_{x : A} B$ of _dependent sums_ of $A$ and $B$, which always lives in $\mathrm{Set}$.
Given $A : u$, and some $B : v$ with one variable $x : A$ free (where $u$ and $v$ are possibly distinct universes), we can form the type $\prod_{x : A} B$ of _dependent products_ from $A$ to $B$. If $v$ is in $\mathrm{Prop}$, then the dependent product also lives in $\mathrm{Prop}$. Otherwise, it lives in $\mathrm{Set}$. Moreover, we can also form the type $\sum_{x : A} B$ of _dependent sums_ of $A$ and $B$, which always lives in $\mathrm{Set}$.
Given a type $A$ and two elements $a, b : A$, we can form the _proposition_ $a \equiv_{A} b$. Note the emphasis on the word proposition here! Since equivalence is a proposition, we have uniqueness of equality proofs by definition: there's at most one way for things to be equal, conflicting with univalence. So we get _some_ extensionality, namely of functions, but not for arbitrary types. Given types <spanclass=together>$A$ and $B$,</span> a proof $p : A \equiv B$ and $x : A$, we have the term $\mathrm{coe}(A, B, P, x) : B$, which represents the *coe*rcion of $x$ along the path $p$.
Given a type $A$ and two elements $a, b : A$, we can form the _proposition_ $a \equiv_{A} b$. Note the emphasis on the word proposition here! Since equivalence is a proposition, we have uniqueness of equality proofs by definition: there's at most one way for things to be equal, conflicting with univalence. So we get _some_ extensionality, namely of functions, but not for arbitrary types. Given types <spanclass=together>$A$ and $B$,</span> a proof $p : A \equiv B$ and $x : A$, we have the term $\mathrm{coe}(A, B, p, x) : B$, which represents the *coe*rcion of $x$ along the path $p$.
Here is where my presentation of observational equality starts to differentiate from the paper's: McBride et al use _heterogeneous_ equality, i.e. a 4-place relation $(x : A) \equiv (y : B)$, where $A$ and $B$ are potentially distinct types. But! Their system only allows you to _use_ an equality when $A \equiv B$. The main motivation for heterogeneous equality is to "bunch up" as many equalities as possible to be eliminated all in one go, since coercion in their system does not compute. However, if coercion computes normally, then we don't need to do this "bunching": one can just use coercion normally.
Here is where my presentation of observational equality starts to differentiate from the paper's: McBride et al use _heterogeneous_ equality, i.e. a 4-place relation $(x : A) \equiv (y : B)$, where $A$ and $B$ are potentially distinct types. But! Their system only allows you to _use_ an equality when $A \equiv B$. The main motivation for heterogeneous equality is to "bunch up" as many equalities as possible to be eliminated all in one go, since coercion in their system does not compute. However, if coercion computes normally, then we don't need to (and, in fact, can't) do this "bunching": one just uses coercion normally.
The key idea of OTT is to identify as "equal" objects which support the same _observations_: for functions, observation is _application_; for pairs, it's projection, etc. This is achieved by making the definition of equality acts as a "pattern matching function" on the structure of terms and types. For example, there is a rule which says an equality between functions is a function that returns equalities:
The key idea of OTT is to identify as "equal" objects which support the same _observations_: for functions, observation is _application_; for pairs, it's projection, etc. This is achieved by making the definition of equality acts as a "pattern matching function" on the structure of terms and types. For example, there is a rule which says an equality between functions is a function that returns equalities:
@ -35,11 +35,11 @@ Before we consider the full induction principle, it's useful to consider a simpl
- A _motive_, the type we elminate into;
- A _motive_, the type we elminate into;
- A _method_ for the constructor `zero`, that is, an element of `A`, and
- A _method_ for the constructor `zero`, that is, an element of `A`, and
- A _method_ for the constructor `succ`, which is a function from `A → A`.
- A _method_ for the constructor `suc`, which is a function from `A → A`.
To put it into code:
To put it into code:
```
```agda
foldNat : (A : Type) → A → (A → A) → Nat → A
foldNat : (A : Type) → A → (A → A) → Nat → A
```
```
@ -72,7 +72,7 @@ A morphism between algebras $(N_0, z_0, s_0)$ and $(N_1, z_1, s_1)$ consists of
Suppose for a second that we have a "special" `Nat`{.dt}-algebra, denoted `Nat*`{.dt}. The statement that there exists a recursion principle for the `Nat`{.dt}ural numbers is then a function with the type boxed below.
Suppose for a second that we have a "special" `Nat`{.dt}-algebra, denoted `Nat*`{.dt}. The statement that there exists a recursion principle for the `Nat`{.dt}ural numbers is then a function with the type boxed below.
_∘_ : {ℓ : _} {A : Set ℓ} {x y z : A} → y ≡ z → x ≡ y → x ≡ z
_∘_ : {ℓ : _} {A : Set ℓ} {x y z : A} → y ≡ z → x ≡ y → x ≡ z
refl ∘ refl = refl
refl ∘ refl = refl
@ -429,7 +430,7 @@ Complicating it further: Induction-induction
We've seen, in the past 2 thousand-something words, how to build algebras, algebra homomorphisms, displayed algebras and algebra sections for inductive types, both indexed and not. Now, we'll complicate it further: Induction-induction allows the formation of two inductive families of types $A : \mathrm{Set}$ and $B : A \to \mathrm{Set}$ _together_, such that the constructors of A can refer to those of B and vice-versa.
We've seen, in the past 2 thousand-something words, how to build algebras, algebra homomorphisms, displayed algebras and algebra sections for inductive types, both indexed and not. Now, we'll complicate it further: Induction-induction allows the formation of two inductive families of types $A : \mathrm{Set}$ and $B : A \to \mathrm{Set}$ _together_, such that the constructors of A can refer to those of B and vice-versa.
The classic example is defining a type together with an inductively-defined predicate on it, for instance defining sorted lists together with a "less than all elements" predicate in one big induction. However, in the interest of brevity, I'll consider a simpler example: A syntax for contexts and Π-types in type theory. We have one base type, `ι`, which is valid in any context, and a type for dependent products `Π`. If `σ` is a type in `Γ` and `τ` a type in the context `Γ, σ`, then `Π σ τ` is also a type in `Γ`.
The classic example is defining a type together with an inductively-defined predicate on it, for instance defining sorted lists together with a "less than all elements" predicate in one big induction. However, in the interest of brevity, I'll consider a simpler example: A syntax for contexts and Π-types in type theory. We have one base type, `ι`, which is valid in any context, and a type for dependent products `Π`. Its type expresses the rule for $\prod$-introduction, which says that given $\Gamma \vdash \sigma\ \mathrm{type}$ and $\Gamma, \sigma \vdash \tau\ \mathrm{type}$, then $\Gamma \vdash (\prod(\sigma) \tau)\ \mathrm{type}$.
```agda
```agda
data Ctx : Set
data Ctx : Set
@ -444,7 +445,7 @@ data Ty where
Π : {Γ : Ctx} (σ : Ty Γ) (τ : Ty (pop Γ σ)) → Ty Γ
Π : {Γ : Ctx} (σ : Ty Γ) (τ : Ty (pop Γ σ)) → Ty Γ
```
```
Here I'm using Agda's forward-declarationthe definition of their algebras feature to specify the signatures of `Ctx` and `Ty` before its constructors. This is because the constructors of `Ctx` mention the type `Ty`, and the type _of_`Ty` mentions `Ctx`, so there's no way to untangle them other than this.
Here I'm using Agda's forward-declaration feature to specify the signatures of `Ctx` and `Ty` before their constructors. This is because the constructors of `Ctx` mention the type `Ty`, and the type _of_`Ty` mentions `Ctx`, so there's no good way to untangle them.
When talking about our standard set of categorical gizmos for inspecting values of inductive type, it's important to note that, just like we can't untangle the definitions of `Ctx` and `Ty`, we can't untangle their algebras either. Instead of speaking of `Ctx`-algebras or `Ty`-algebras, we can only talk of `Ctx-Ty`-algebras.[^2] Let's think through the algebras first, since the rest are derived from that.
When talking about our standard set of categorical gizmos for inspecting values of inductive type, it's important to note that, just like we can't untangle the definitions of `Ctx` and `Ty`, we can't untangle their algebras either. Instead of speaking of `Ctx`-algebras or `Ty`-algebras, we can only talk of `Ctx-Ty`-algebras.[^2] Let's think through the algebras first, since the rest are derived from that.
Now the type of the fourth component is interesting. We want `seg0` and `seg1` to be "equal" in some appropriate sense of equality. But they have totally different types! `seg0 : l0 ≡I0 r0` and `seg1 : l1 ≡I1 r1`. It's easy enough to fix the type of `seg0` to be a path in `I0`:`ap I seg0` is a path in `I1` between `I l0` and `I r0`. Out of the pan, into the fire, though. Now the endpoints don't match up!
Now the type of the fourth component is interesting. We want `seg0` and `seg1` to be "equal" in some appropriate sense of equality. But they have totally different types! `seg0 : l0 ≡I0 r0` and `seg1 : l1 ≡I1 r1`. It's easy enough to fix the type of `seg0` to be a path in `I1`: Since `I : I0 → I1`, its action on paths`ap I seg0` is a path in `I1` between `I l0` and `I r0`. Out of the pan, into the fire, though. Now the endpoints don't match up!
Can we fix `I l0` and `I r0` in the endpoints of `seg0`? The answer, it turns out, is yes. We can use path _transitivity_ to alter both endpoints. `r` correctly lets us go from `I r0` to `r1`, but `l` is wrong—it takes `l1` to `I r0`. We want that backwards, so we apply _symmetry_ here.
Can we fix `I l0` and `I r0` in the endpoints of `seg0`? The answer, it turns out, is yes. We can use path _transitivity_ to alter both endpoints. `r` correctly lets us go from `I r0` to `r1`, but `l` is wrong—it takes `l1` to `I r0`. We want that backwards, so we apply _symmetry_ here.
