Skip to content

logsem/fw-rec-inj

Repository files navigation

Prerequisites

  • coq 8.16.1

Installation

  • with nix:
    nix-shell
    
  • with opam:
    opam pin add coq 8.16.1
    eval $(opam env)
    
  • with docker:
    docker build -t artifact .
    docker run --rm -it --entrypoint /bin/bash artifact
    eval $(opam env)
    
  • alternatively, you can install it with a package manager of your choice

Typechecking the development

  • make all --- typecheck the project.
  • ./check_admits.sh --- count used admits.
  • ./check_axioms.sh --- count declared axioms.

Cleanup

  • make clean --- remove generated build files.
  • make cleanfull --- remove Coq bytecode files.

Mapping the paper onto the development

In this section, we walk through the main claims of the paper, section by section, and refer to the appropriate definitions/theorems in the Coq formalisation, as well as explain some representation choices and formalisation decisions. The inverse mapping (in the form of annotations for the particular files of the project) and a glossary of names used in the paper/formalisation may be helpful, and can be found below.

Section 2, definition of the calculus

The definition of the calculus can be found in the Lang/FOmegaEq subdirectory of the projects, in the files Kinds.v, Types.v, Syntax.v, Typing.v and OpSemantics.v. The syntax of kinds and type constructors (Fig. 1 in the paper) is defined in the file Kinds.v as inductive families kind and tctor respectively; note that we represent type-constructors as intrinsically well-kinded (thus incorporating part of Fig. 2). The syntax of types and constraints (the remainder of Fig. 1) is defined in the file Types.v as the mutually inductive families type and constr; as with the type constructors these are intrinsically well-kinded (or well-formed in the case of constraints; thus, they incorporate the remainder of Fig. 2). This design choice is rather helpful in the remainder of formalisation, as it incorporates additional information (e.g., about the type of argument in type-level beta redexes) that allows for building the interpretation by induction on the structure of the type. At the same type (in contrast to well-behaved but ill-typed programs), ill-kinded types are not practically useful, so we do not lose any important information.

The discriminability and provability of equality constraints (Fig. 3) judgments are also defined in the file Types.v (as inductive judgments tdiscr and tproves_i, respectively), together with a number of lemmas (closure under renaming/substitution, weakening, cut admissibility).

The term-level syntax of the calculus (Fig. 4), comprising values and expressions, is defined in the file Syntax.v as the mutually inductive types value and expr. These are intrinsically well-scoped, but untyped. Instead of defining evaluation contexts explicitly, together with a reduction semantics (Fig. 6) we define a structural operational semantics, as the judgment head_step (for a single-step relation), found in the file OpSemantics.v. Particularly for a fine-grain call-by-value, equivalence of the two formats is trivial.

Finally, the type system (Fig. 5) is presented in the file Typing.v, as the mutually inductive predicates typing_expr and typing_val (respectively for typing of expressions and values). The file also provides numerous lemmas (closure under renaming/substitution, both with respect to type- and term-level variables, as well as weakening and cut lemmas for the equality assumptions).

The example of a closed value of the empty type can be found in the final section of the file Typing.v, with the final lemma, tnt, building a proof that the constructed value has indeed type void. Since the simpler sub-calculus without recursive types and injectivity of quantifiers is only explicitly used in the (un-formalised) Theorem 3.2, we do not provide this variant explicitly. We also do not provide the encodings of examples of GADTs found in Section 2.1.

Section 3, non-expressibility of injective constraints

The formalised part of this section comprises the propositional model of Fω, and can be found in the subdirectory Lang/FOmega. The calculus is structured similarly to the larger system and is split across three files: the file Types.v provides the structure of kinds, (intrinsically well-kinded) types and type equality, while the file Syntax.v provides the term-level syntax of the calculus. The file Interp.v provides the propositional interpretation of the calculus, written tint, as defined in Fig. 11 in the paper. Note that we work in the category of setoids and equality-preserving arrows, in order to construct the interpretation together with the proof that it preserves equality of contexts. While other options are available (e.g., assuming functional extensionality), this construction is more in keeping with the NbE interpretation presented in Section 5, which relies on a similar approach (in a more complicated setting) to build functoriality into the construction. Theorem 3.1, which ensures that the truth of the interpretation of a type follows from the truth of the interpretation of the context is located at the end of the file Interp.v, under the name of etypes_true. As noted in the paper, Theorem 3.2 is left un-mechanised, since its proof is simple and provided in the paper, while a mechanisation would be rather convoluted.

Section 4, type-soundness of the calculus

The syntactic proof of type-soundness can be found in the file Soundness.v in the subdirectory Lang/FOmegaEq of the development. While it depends on our NbE interpretation, we defer its discussion till the next section; note only that the Lemma 4.1 (Consistency), on which the remainder of this section depends, can be found at the end of the file Interp.v. From there, we proceed via a series of canonical-forms lemmas for particular type constructors, through the usual formulations of the progress and preservation lemmas, up to the final soundness theorem stating that well-typed programs are safe (def'n in OpSemantics.v, above). Note that, while that definition is not identical to the paper, it is equivalent, since irreducibility is trivially decidable.

Section 5, NbE and relational models

The structure of the formalisation of this section differs slightly from the paper. This is because, in order not to duplicate the formalisation effort, the NbE procedures used in Sections 5.2 and 5.3 are joined together, with a type of "open universes" (Fig. 18), found in the functor NF in the file Universe.v parameterised by the type of predicates. Taking this parameter to be empty we regain the usual neutral and normal forms (Fig 12.), while taking it to be step-indexed relations on values gives us the open universes we use in Section 5.3. The other ingredient necessary for our implementation of the NbE algorithm is the notion of presheaves, and in particular of the exponential presheaf: these can be found in the file Presheaves.v. Note that, although we use these notions and concepts, we cannot fully utilise the internal language of presheaves within Coq; we will point out the discrepancies in the following. The interpretation of kinds (Definition 5.2, written kint in the formalisation), as well as its point-wise extension to contexts and proofs that the universes form presheaves can be found in the file Universe.v, after the inductive universes are defined.

We can now move on to the actual normalisation procedure, which is defined in the file Interp.v. The reflection and reification functions (Fig. 13) are defined by mutual induction on the structure of the kind reflectI and reifyI, respectively. For practical reasons, rather than at an arbitrary index, we define them at the empty indexing context (picking an indexing context is necessary, since the type of the exponential presheaf PArr (PNeu κ) (kint κ) is not a Coq universe, but a type of universe-valued functions). This does not prevent us from using the functions at other indexes, however: the wrappers reify and reflect provide the mappings as Coq functions at the arbitrary index. The interpretation of types and constraints (Fig. 14) follows the same pattern, with functions tint and constrint building an instance of the appropriate presheaf, and wrappers interp and interpC used for practical convenience. We also define the identity environment (Def. 5.3) as η_id.

Since the universes we use allow for non-syntactic elements (like step-indexed relations over values), we replay the interpretation for the universes (as discussed in Section 5.3): this re-interpretation is defined as the three functions neuint, nfint and nfconstrint, which interpret neutral universes, normal universes and normal constraints back into the appropriate semantics spaces.

With these definitions, we can define the logical relation (Fig. 15), which is written sub_rel in the development and precisely matches the paper definition, and its extension to environments, written sub_relS. With these two, we can prove the necessary lemmas: Lemma 5.5 corresponds to interp_reify, Lemma 5.6 — to interp_rel and Lemma 5.4 — to reify_refl_inj. Finally, Lemma 5.7, which expresses validity of provability and discriminability judgment corresponds to tproves_int (for provability) and tdiscr_int (for discriminability) — and together, they imply the consistency lemma of the previous section. Before the elements of the universe are realized as predicates/relations, we provide a more convenient representation for its closed, ground fragment in the file GroundView.v.

At this point, the development forks into two branches, corresponding to the simpler, unary realization of Section 5.2, and the more sophisticated binary realization of Section 5.3. The first can be found in the subdirectory Lang/FOmegaEq/Unary. The definition of realization (Fig. 16) can be found in the file Realize.v in that directory, with the evaluation closure written ECL, and the realization (defined as a guarded fixpoint via a step-indexed logic library) written RP. The logical relation (Fig. 17) corresponds to definitions logrelE and logrelV for the versions for expressions and values, respectively. This leaves us with the compatibility lemmas and fundamental theorem, found in the file Compat.v, and its corollary, semantic proof of soundness, found in the file Soundness.v.

As already mentioned, the binary model follows the same pattern, although with a slightly more sophisticated universe of normal forms. However, this has little bearing on realization: the definition is very similar, with the only difference stemming from the usual asymmetry of a step-indexed binary model. The files can be found in the Lang/FOmegaEq/Binary subdirectory, with the realization defined in the file Realize.v, compatibility and fundamental lemma in the file Compat.v, contextual approximation and adequacy in the file Soundess.v, and the examples in the file Examples.v.

Project layout

.
+-- README
+-- Makefile
+-- _CoqProject
+-- Binding/ (substitution framework)
+-- IxFree/ (step-indexed logic framework)
+-- Lang/
|   +-- FOmega/ (propositional model of Fω)
|   +-- FOmegaEq/Kinds.v (syntax for kinds, syntax for type constructors with their corresponding kinds)
|   +-- FOmegaEq/Types.v (intrinsicly well-kinded types, discriminability and provability rules, renaming and substitution lemmas for kinding, cut lemma for provability)
|   +-- FOmegaEq/Syntax.v (syntax for well-scoped expressions and values, renaming and substitution functions)
|   +-- FOmegaEq/Typing.v (typing of expressions and values, renaming and substitution lemmas for typing, non-terminating example, derived using injectivity of forall)
|   +-- FOmegaEq/OpSemantics.v (operational semantics, safety definition, reflexive-transitive closure, n-step relation)
|   +-- FOmegaEq/Presheaves.v (presheaves, presheaf exponential)
|   +-- FOmegaEq/Universe.v (inductively defined universe of extended normal and neutral forms, presheaf structure of Nf k and Neu k)
|   +-- FOmegaEq/Interp.v (reflect, reify and eval for NbE, injectivity of reify)
|   +-- FOmegaEq/GroundView.v (closed ground types)
|   +-- FOmegaEq/Soundness.v (syntactic soundness)
|   +-- FOmegaEq/Unary/ (unary realizability model)
|   |   +-- Realize.v
|   |   +-- Compat.v
|   |   +-- Soundness.v
|   |   +-- SoundnessNbE.v (proof of soundness for NbE)
|   +-- FOmegaEq/UnaryPredicate (unary realizability model with predicates)
|   |   +-- Realize.v
|   |   +-- Compat.v
|   |   +-- Soundness.v
|   |   +-- Examples.v
|   +-- FOmegaEq/Binary/ (binary model)
|   |   +-- Realize.v
|   |   +-- Compat.v
|   |   +-- Soundness.v
|   |   +-- Examples.v

Paper-formalization glossary

Paper entry Coq qualified identifier
fig. 1 (kinds) Kinds.kind
fig. 2 (well-kinded constructors) Kinds.tctor
fig. 1 (types) Types.type
fig. 2 (well-kinded types) Types.type
fig. 3 (provability, discriminability, auxiliary lemmas) Types.tdiscr, Types.tproves_i
fig. 4 (values, expressions, evaluation context, auxiliary lemmas) Syntax.value, Syntax.expr
fig. 5 (type system) Typing.typing_expr, Typing.typing_val
non-termination paragraph in sec. 2 Typing.ExampleClosedVoid
fig. 6 (operational semantics) OpSemantics.head_step
Auxiliary definitions for presheaves in sec. 5 (used for def. 5.1, and for computations taking place in Pr(K)) Presheaves.Functors, Presheaves.Arrows
fig. 12 Module Universe.NF, in particular neu, nf, nfc (note, that it's already has an additional constructor that is introduced in fig. 18, but if the module is instantiated with False, then the definition matches the one in the fig. 12)
fig. 18 Module Universe.NF, in particular neu, nf, nfc, if instantiated with val ∅ -> val ∅ -> iProp
def. 5.2 Universe.NF.kint
lemma 4.1 Interp.Interp.consistency
lemma 5.4 Interp.Interp.reify_refl_inj, Interp.Interp.reify_refl_inj'
lemma 5.5 Interp.Interp.reinterp_nf
lemma 5.6 Interp.Interp.interp_rel
lemma 5.7 Interp.Interp.tproves_int, Interp.Interp.tdiscr_int
def. 5.3 Interp.Interp.η_id
fig. 13 Interp.Interp.reifyI, Interp.Interp.reflectI
fig. 14 Interp.Interp.tint
fig. 15 Interp.Interp.sub_rel
lemma 4.2 Soundness.canArr
lemma 4.3 Soundness.preservation
lemma 4.4 Soundness.progress
lemma 4.6 Soundness.soundness
def. 4.5 Soundness.safety
fig. 16 Unary.Realize.RPR, Unary.Realize.RP
fig. 17 Unary.Realize.logrelV, Unary.Realize.logrelE
theorem 5.8 Unary.Compat.fl_expr, Unary.Compat.fl_val
fig. 19 Binary.Realize.RPR, Binary.Realize.RP
Representation independence example from sec. 5.3 Binary.Examples.Vec
Free theorem example from sec. 5.3 Binary.Examples.FreeThm

Notation glossary

Substitution framework

Paper Coq formalization
Renaming type Δ → Δ' Δ [→] Δ'
Substitution in terms 𝑒 [𝑣/𝑥] subst e v

Step-indexed framework

Paper Coq formalization
Steped-indexed proposition iProp *ₛ
Step-indexed relation A →ₛ B
Embedding of intuitionistic logic into step-indexed logic ( P )ᵢ
Step-indexed proposition P is valid at n n ⊨ P
Step-indexed proposition P is valid at any n ⊨ P

FOmegaEq

Paper Coq formalization
Discriminability τ # σ ⊮ τ ≃ σ
Provability Δ ∣ Ψ ⊩ τ ≡ σ Ψ ⊩ τ ≃ σ
Kinding Δ ⊢ τ :: κ τ : typ κ Δ
Presheaf category on context renamings Pr(K) PShf
Presheaf exponential PArr
Neutral forms U♮, Neu Neu
Normal forms U, Nf Nf
Kind interpretation 〚κ〛 kint κ
Context interpretation 〚Δ〛 NSub Δ
Type interpretation (first stage) 〚τ〛 tint τ
Good context interpretation good(μ) sub_rel η_id μ μ
Type interpretation (second stage) R RP
Evaluation closure E ECL

Tactic glossary

Substitution framework

Coq identifier Explanation
term_simpl move substitutions at the outermost level, move renamings at the intermost level, compose them, if possible, simplify

Step-indexed framework

Coq identifier Explanation
later_shift eliminate ▷ in front of the goal, and all the hypothesis
i<tactic> step-indexed counterparts of Coq tactics

Additional comments

  • In our implementation, we use two styles of representation of terms with binders. Our types are intrinsicly well-kinded, parametrized over contexts and kinds:
    Inductive type {Δ : Ctx kind} : kind → Type :=
    | t_var κ (x : dom Δ) (EQ : Δ x = κ) : type κ
    | t_lam {κₐ κᵣ} (τ : @type (Δ ▹ κₐ) κᵣ) : type (κₐ ⇒ κᵣ)
    | t_app {κₐ κᵣ} (σ : type (κₐ ⇒ κᵣ)) (τ : type κₐ) : type κᵣ
    ...
    .
    
    Our expressions and values are well-scoped:
    Inductive value {V : Set} :=
    | v_lam (e : @expr (inc V)) : value
    ...
    with expr {V : Set} :=
    | e_bind (e₁ : expr) (e₂ : @expr (inc V)) : expr
    ...
    .
    
  • We allow recursive types with arbitrary arity, but require them to be fully applied for programs. To enforce this condition with use the predicate Types.rel_app.
  • For the purpose of de-duplication, we implemented the first stage (denoted by 〚 〛 in the paper), for the unary and binary cases in one module, parametrized by a type, used as a parameter for the additional constructor. This type is given by signature Point (Universe.v):
    Module Type Point.
        Parameter P : Type.
    End Point.
    
    Module NF (PM : Point).
        Definition Point := PM.P.
    
        Inductive neu {Δ : Ctx kind} : kind → Type :=
        | neu_rel (P : Point) : neu ⋆
        ...
        with nf {Δ : Ctx kind} : kind → Type :=
        ...
        with nfc {Δ : Ctx kind} : Type :=
        ...
        .
    
    If Point.P := False, and, hence, the set of normal/neutral types is a subset of types, we obtain a model for beta-short, eta-long types (used to prove syntactic soundness, and unary semantical soundness in the paper). If Point.P := val ∅ -> iProp, we obtain a model that allows embedding of predicates for types at the base kind (used to construct a unary relational model), which is not used in the paper. If Point.P := val ∅ -> val ∅ -> iProp, we obtain a model that allows embedding of relations for types at the base kind (used to construct a binary model in the paper).
  • As for program-level reasoning it suffices to study their behavior at closed types at the base kind, and after NbE these types have a simplier structure (beta-short, eta-long), we can use gtyp (and viewG : Neu ⋆ ε -> gtyp), defined in GroundView.v, to define realizability semantics.
  • Unary/Example.v contains a trivial example of instantiating forall with a predicate to check a program with a presence of semantically correct (but probably syntactically not well-typed) expression.
  • Unary/Soundness.v contains a proof that the unary logical relation implies safety (adequacy).
  • Binary/Soundness.v contains a proof that the binary logical relation implies contextual approximation (adequacy).
  • Binary/Compat.v contains a fundamental lemma for the binary logical relation.
  • As any value of Point.P, apart from False prevents the set of neutral and normal forms to be a subset of types, most lemmas in Interp.v require duplicate versions for types, for neutrals, and for normals.
  • Syntax includes fixpoint combinator (v_fix) to shorten the examples.
  • We disable the following warnings, which don't affect consistency of the mechanization: notation-overridden uniform-inheritance ambiguous-paths implicit-core-hint-db require-in-module unexpected-implicit-declaration undeclared-scope deprecated-hint-without-locality deprecated-instance-without-locality deprecated-syntactic-definition.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages