gallais / sott

Simplified Observational Type Theory

Simplified Observational Type Theory

This is an implementation of Observational Type Theory. It is "simplified" because the definitions of equality of types and terms do not reduce as in the original definition. Instead, the theory provides ways to construct proofs of equality, and to project out consequences of an equality (e.g., equal pairs implies the components are equal). Not reducing equality propositions stops (this source of) explosion of the sizes of types.

How to build it

1. Install OCaml and opam. You will need OCaml version 4.03.0 or newer, because I've used inline records in algebraic datatypes.

2. opam install jbuilder cmdliner menhir

3. jbuilder build cmdline/sott.exe

How to typecheck a file

./_build/default/cmdline/sott.exe typecheck <file.sott>

Examples

1. test1.sott Derivation of transport from coerce; and a demonstration that Hofmann's counterexample to canonicity in the presence of a functional extensionality axiom fails in OTT.
2. test1.5.sott Shortened version of test1.sott.
3. test2.sott An equality-proof irrelevance test; transivity and symmetry; coherence.
4. test3.sott Encoding of sum types via booleans and sigma types; demonstration that the counterexample to canonicity involving constructors when adding functional extensionality fails in OTT [https://coq.inria.fr/cocorico/extensional_equality].
5. arith.sott Construction of and some proofs for the integers constructed as a quotient of pairs of natural numbers.

Features

1. Type equalities, written [A = B], where A and B are proper types. All proofs of the same equality are definitionally equal.

2. Heterogeneous term equality, written [a : A = b : B], where A and B are types. As a shorthand, if A is the same as B, then this can be written as [a = b : A]. All proofs of the same equality are definitionally equal.

3. Reflexivity for both type and term equality, written refl.

4. Coercion, written as coerce(a,A,B,p), where A and B are types, a : A, and p : [A = B]. If A and B are in normal form, then the coercion will compute as appropriate to the type. This is how we maintain canonicity even when the equality type is arbitrarily extended.

5. Coherence, coherence(e) where e : [A = B], has type [x : A = coerce(x,A,B,e) : B] (the x is inferred from the type being pushed in).

6. Pi types: (x : A) -> B, \x -> e and e1 e2. Function extensionality is witnessed by funext(x1 x2 x_e. XXX) : [f1 : (x : A1) -> B1 = f2 : (x : A2) -> B2], where x1 : A1, x2 : A2, x_e : [x1 : A1 = x2 : A2] and XXX : [f1 x1 : B1[x1] = f2 x2 : B2[x2]].

7. Sigma types (x : A) * B, (x,y) and x #fst, x #snd.

8. Booleans: Bool, True, False and x for x. T { True -> e1; False -> e2 }.

9. Natural numbers: Nat, Zero, Succ n and x for y. T { Zero -> e1; Succ n p -> e2 }.

10. A Russell-style universe Set, which includes equalities, booleans, naturals, Pi, Sigma.

11. Quotient types:

    • Formation: A / R, where A is a type and R : A -> A -> Set
    • Introduction: [a] : A / R, where a : A
    • Elimination: x for z. T { [z] -> e1; z1 z2 zr -> e2 }. The expression e1 computes the answer, and e2 is a proof that e1 doesn't depend on the equivalence class.
    • Propositional equality: same-class(e) : [[a1] : A/R = [a2] : A/R] when e : R a1 a2.

Bugs/TODO

1. The parser reports a line number on error, but otherwise the error messages are useless. One day, I hope to use Menhir's fancy error message support.

2. Not every useful combinator for equalities is present. This is a matter of implementing them in the lexer/parser/typechecker.

3. The normaliser is a bit too keen and unfolds all definitions whether it needs to for equality checking or not. This leads to massive terms during type checking. Error messages are now reported without unfolded definitions though, which makes them a bit easier to read.

4. Type error messages don't include the context, and may output names that shadow names in the context, leading to incorrect terms.