``` {-# OPTIONS --rewriting #-} module rewriting.examples.BlogTypeSafety10Easy4Med1Hard where open import Data.Bool using (true; false) renaming (Bool to 𝔹) open import Data.Empty using (⊥; ⊥-elim) open import Data.Nat open import Data.Nat.Properties using (≤-refl) open import Data.List using (List; []; _∷_) open import Data.Product using (_,_;_×_; proj₁; proj₂; Σ-syntax; ∃-syntax) open import Data.Sum using (_⊎_; inj₁; inj₂) open import Data.Unit using (⊤; tt) open import Data.Unit.Polymorphic renaming (⊤ to topᵖ; tt to ttᵖ) open import Relation.Binary.PropositionalEquality as Eq using (_≡_; _≢_; refl; sym; cong; subst; trans) open import Relation.Nullary using (¬_; yes; no) ``` Ok, so logical relations are overkill for proving type safety. The proof technique is better suited to proving more interesting properties such as parametricity, program equivalence, and the gradual guarantee. Nevertheless, understanding a proof of type safety via logical relations is a helpful stepping stone to understanding these more complex use cases, especially when the logical relations employ more advanced techniques, such as step indexing. In this blog post I prove type safety of a cast calculus (an intermediate language of the gradually typed lambda calculus). The proof is in Agda and the proof uses step-indexed logical relations because the presence of the unknown type (aka. dynamic type) prevents the use of logical relations that are only indexed by types. To reduce the clutter of reasoning about step indexing, we conduct the proof using a temporal logic, in the spirit of the LSLR logic of Dreyer, Ahmed, and Birkedal (LMCS 2011), that we embed in Agda. This is a literate Agda file, so most of the details are here, but it imports several items whose root on github is here: https://github.com/jsiek/abstract-binding-trees/tree/master/src This post is based on work with Philip Wadler and Peter Thiemann. ## Review of the Cast Calculus ``` open import Var open import rewriting.examples.Cast ``` We review the syntax and reduction rules of this cast calculus. Just like the lambda calculus, types include base types (Booleans and natural numbers), and function types. To support gradual typing, we include the unknown type ★. ι ::= 𝔹 | ℕ A,B,C,G,H ::= ι | A ⇒ B | ★ The ground types are G,H ::= ι | ★⇒★ Just like the lambda calculus, there are variables (de Bruijn indices), lambdas, and application. We throw in literals (Booleans and natural numbers). Also, to support gradual typing, we include a term `M ⟨ G !⟩` for injecting from a ground type `G` to the unknown type, and a term `M ⟨ H ?⟩` for projecting from the unknown type back out to a ground type. Finally, we include the `blame` term to represent trapped runtime errors. L,M,N ::= ` x | ƛ N | L · M | $ k | M ⟨ G !⟩ | M ⟨ H ?⟩ | blame This cast calculus is somewhat unusual in that it only includes injections and projections but not the other kinds of casts that one typically has in a cast calculus, such as a cast from one function type `★ ⇒ ℕ` to another function type `ℕ ⇒ ℕ`. That is OK because those other casts can still be expressed in this cast calculus. The values include lambdas, literals, and injected values. V,W ::= ƛ N | $ c | V ⟨ G !⟩ The reduction rules make use of frames, which are defined as follows. F ::= □· M | V ·□ | □⟨ G !⟩ | □⟨ H ?⟩ The operation `F ⟦ M ⟧` plugs a term into a frame. The reduction rules of the cast calculus are as follows: (ξ) If M —→ N, then F ⟦ M ⟧ —→ F ⟦ N ⟧ (ξ-blame) F ⟦ blame ⟧ —→ blame (β) (ƛ N) · W —→ N [ W ] (collapse) V ⟨ G !⟩ ⟨ G ?⟩ —→ V (collide) If G ≢ H, then V ⟨ G !⟩ ⟨ H ?⟩ —→ blame. ## A First Attempt at a Logical Relation for Type Safety The following is a first attempt to define a logical relation for type safety for the cast calculus. The predicate ℰ expresses the semantic notion of a term being well typed at a given type A. Here "semantic" means "runtime behavior". We define that a term M is semantically well typed at type A if it satisfies "progress" and "preservation". The progress part says that M is either (1) a semantic value at type `A`, (2) reducible, or (3) an error. The preservation part says that if M reduces to N, then N is also semantically well typed at A. ℰ⟦_⟧ : (A : Type) → Term → Set ℰ⟦ A ⟧ M = (𝒱 ⟦ A ⟧ M ⊎ reducible M ⊎ Blame M) × (∀ N → (M —→ N) → ℰ⟦ A ⟧ N) The predicate 𝒱 expresses the semantic notion of a value being well typed at some type A. For a base type `ι` (𝔹 or ℕ), the value must be the appropriate kind of literal (Boolean or natural number). For a function type `A ⇒ B`, the value must be a lambda expression `ƛ N`, and furthermore, substituting any value `W` that is semantically well typed at `A` into the body `N` produces a term that is semantically well typed at `B`. For the unknown type `★`, the value must be an injection of a value `V` from some ground type `G`, and `V` must be semantically well typed at `G`. 𝒱⟦_⟧ : (A : Type) → Term → Set 𝒱⟦ ι ⟧ ($ c) = ι ≡ typeof c 𝒱⟦ A ⇒ B ⟧ (ƛ N) = ∀ W → 𝒱⟦ A ⟧ W → ℰ⟦ B ⟧ (N [ W ]) 𝒱⟦ ★ ⟧ (V ⟨ G !⟩) = Value V × 𝒱⟦ gnd⇒ty G ⟧ V 𝒱⟦ _ ⟧ _ = ⊥ Note that the definitions of ℰ and 𝒱 are recursive. Unfortunately they are not proper definitions of (total) functions because there is no guarantee of their termination. For simple languages, like the Simply Typed Lambda Calculus, 𝒱 can be defined by recursion on the type `A`. However, here we have the unknown type `★` and the recursion in that clause invokes `𝒱⟦ gnd⇒ty G ⟧ V`, but `gnd⇒ty G` is not a structural part of `★` (nothing is). (The definition of ℰ above is also problematic, but one could reformulate ℰ to remove the recursion in ℰ.) ## An Explicitly Step-indexed Logical Relation for Type Safety We can force the definitions of ℰ and 𝒱 to terminate using step-indexing (aka. the "gasoline" technique), which was first applied to logical relations by Appel and McAllester (TOPLAS 2001). We add a parameter k (a natural number) to ℰ and 𝒱, and decrement k on each recursive call. When k is zero, ℰ and 𝒱 accept all terms. Thus, the meaning of `ℰ⟦ A ⟧ M k` is that term `M` is guaranteed to behave according to type `A` for `k` reduction steps, but after that there are no guarantees. ℰ⟦_⟧ : (A : Type) → Term → ℕ → Set ℰ⟦ A ⟧ M 0 = ⊤ ℰ⟦ A ⟧ M (suc k) = (𝒱 ⟦ A ⟧ M k ⊎ reducible M ⊎ Blame M) × (∀ N → (M —→ N) → ℰ⟦ A ⟧ N k) 𝒱⟦_⟧ : (A : Type) → Term → ℕ → Set 𝒱⟦ A ⟧ M 0 = ⊤ 𝒱⟦ ι ⟧ ($ ι′ c) (suc k) = ι ≡ ι′ 𝒱⟦ A ⇒ B ⟧ (ƛ N) (suc k) = ∀ W → 𝒱⟦ A ⟧ W k → ℰ⟦ B ⟧ (N [ W ]) k 𝒱⟦ ★ ⟧ (V ⟨ G !⟩) (suc k) = Value V × 𝒱⟦ gnd⇒ty G ⟧ V k 𝒱⟦ _ ⟧ _ (suc k) = ⊥ We now have proper definitions of ℰ and 𝒱 but proving theorems about these definitions involves a fair bit of reasoning about the step indices, which is tedious, especially in Agda because it's support for automating proofs about arithmetic is cumbersome to use. To streamline the definitions and proofs that involve step indexing, Dreyer, Ahmed, and Birkedal (2011) propose the use of a temporal logic that hides the step indexing. Next we discuss the embedding of such a logic in Agda. ## Step-indexed Logic ``` open import rewriting.examples.StepIndexedLogic2 ``` Our Step-indexed Logic (SIL) includes first-order logic (i.e., a logic with "and", "or", "implies", "for all", etc.). To distinguish its connectives from Agda's, we add a superscript "o". So "and" is written `×ᵒ`, "implies" is written `→ᵒ`, and so on. SIL also includes a notion of time in which there is a clock counting down. The logic is designed in such a way that if a formula `P` is true at some time then `P` stays true in the future (at lower counts). So formulas are downward closed. When the clock reaches zero, every formula becomes true. Furthermore, the logic includes a "later" operator, written `▷ᵒ P`, meaning that `P` is true one clock tick in the future. When we use SIL to reason about the cast calculus, one clock tick will correspond to one reduction step. Just as `Set` is the type of true/false formulas in Agda, `Setᵒ` is the type of true/false formulas in SIL. It is a record that bundles the formula itself, represented with a function of type `ℕ → Set`, with proofs that the formula is downward closed and true at zero. record Setᵒ : Set₁ where field # : ℕ → Set down : downClosed # tz : # 0 -- tz short for true at zero open Setᵒ public For example, the "false" proposition is false at every time except zero. ⊥ᵒ : Setᵒ ⊥ᵒ = record { # = λ { zero → ⊤ ; (suc k) → ⊥ } ; down = ... ; tz = ... } The "and" proposition `P ×ᵒ Q` is true at a given time `k` if both `P` and `Q` are true at time `k`. _×ᵒ_ : Setᵒ → Setᵒ → Setᵒ P ×ᵒ Q = record { # = λ k → # P k × # Q k ; down = ... ; tz = ... } The "for all" proposition `∀ᵒ[ a ] P` is true at a given time `k` if the predicate `P` is true for all `a` at time `k`. ∀ᵒ : ∀{A : Set} → (A → Setᵒ) → Setᵒ ∀ᵒ{A} P = record { # = λ k → ∀ (a : A) → # (P a) k ; down = ... ; tz = ... } The "exists" proposition `∃ᵒ[ a ] P` is true at a given time `k` if the predicate `P` is true for some `a` at time `k`. However, we must require that the type `A` is inhabited so that this proposition is true at time zero. ∃ᵒ : ∀{A : Set}{{_ : Inhabited A}} → (A → Setᵒ) → Setᵒ ∃ᵒ{A} P = record { # = λ k → Σ[ a ∈ A ] # (P a) k ; down = ... ; tz = ... } We embed arbitrary Agda formulas into the step-indexed logic with the following constant operator, written `S ᵒ`, which is true if and only if `S` is true, except at time zero, when `S ᵒ` has to be true. _ᵒ : Set → Setᵒ S ᵒ = record { # = λ { zero → ⊤ ; (suc k) → S } ; down = ... ; tz = ... } Next we discuss the most important and interesting of the propositions, the one for defining a recursive predicate. The following is a first attempt at writing down the type of this proposition. The idea is that this constructor of recursive predicates works like the Y-combinator in that it turns a non-recursive predicate into a recursive one. μᵒ : ∀{A} → (A → (A → Setᵒ) → Setᵒ) ----------------------- → A → Setᵒ The non-recursive predicate has type `A → (A → Setᵒ) → Setᵒ`. It has an extra parameter `(A → Setᵒ)` that will be bound to the recursive predicate itself. To clarify, lets look at an example. Suppose we want to define multi-step reduction according to the following rules: M —→ L L —→* N ------- ------------------ M —→* M M —→* N We would first define a non-recursive predicate that has an extra parameter, let us name it `R` for recursion. Inside the definition of `mreduce`, we use `R` is the place where we would recursively use `mreduce`, as follows. mreduce : Term × Term → (Term × Term → Setᵒ) → Setᵒ mreduce (M , N) R = (M ≡ N)ᵒ ⊎ᵒ (∃ᵒ[ L ] (M —→ L)ᵒ ×ᵒ R (L , N)) Because we use `∃ᵒ` with a Term, we need to prove that Term is inhabited. ``` instance TermInhabited : Inhabited Term TermInhabited = record { elt = ` 0 } ``` We then apply the `μᵒ` proposition to `mreduce` to obtain the desired recursive predicate `—→*`. _—→*_ : Term → Term → Setᵒ M —→* N = μᵒ mreduce (M , N) The problem with the above story is that it's not possible in Agda (to my knowledge) to construct a recursive predicate from an arbitrary function of type `A → (A → Setᵒ) → Setᵒ`. Instead, we need to place restrictions on the function. In particular, if we make sure that the recursion never happens "now", but only "later", then it becomes possible to construct `μᵒ`. We define the `Setˢ` type in Agda to capture this restriction. (The superscript "s" stands for step indexed.) Furthermore, to allow the nesting of recursive definitions, we must generalize from a single predicate parameter to an environment of predicates. The type of the environment is given by a `Context`: Context : Set₁ Context = List Set We represent an environment of recursive predicates with a tuple of the following type. RecEnv : Context → Set₁ RecEnv [] = topᵖ RecEnv (A ∷ Γ) = (A → Setᵒ) × RecEnv Γ We use de Bruijn indices to represent the variables that refer to the recursive predicates, which we define as follows. data _∋_ : Context → Set → Set₁ where zeroˢ : ∀{Γ}{A} → (A ∷ Γ) ∋ A sucˢ : ∀{Γ}{A}{B} → Γ ∋ B → (A ∷ Γ) ∋ B For each variable, we track whether it has been used "now" or not. So we define `Time` as follows. The `Later` constructor does double duty to mean a predicate has either been used "later" or not at all. data Time : Set where Now : Time Later : Time The following defines a list of times, one for each variable in `Γ`. data Times : Context → Set₁ where ∅ : Times [] cons : ∀{Γ}{A} → Time → Times Γ → Times (A ∷ Γ) The `Setˢ` type is a record indexed by the type of the environment and by the time for each variable. The representation of `Setˢ` (the `#` field) is a function that maps an environment of predicates (one predicate for each in-scope μ) to a `Setᵒ`. record Setˢ (Γ : Context) (ts : Times Γ) : Set₁ where field # : RecEnv Γ → Setᵒ ... open Setˢ public We define variants of all the propositional connectives to work on Setˢ. The "later" operator `▷ˢ` asserts that `P` is true in the future, so the predicate `▷ˢ P` can safely say that any use of recursive predicate in `P` happens `Later`. laters : ∀ (Γ : Context) → Times Γ laters [] = ∅ laters (A ∷ Γ) = cons Later (laters Γ) ▷ˢ : ∀{Γ}{ts : Times Γ} → Setˢ Γ ts ----------------- → Setˢ Γ (laters Γ) The "and" operator, `P ×ˢ Q` is categorized as `Later` for a variable only if both `P` and `Q` are `Later` for that variable. Otherwise it is `Now`. We use the following function to make this choice: choose : Kind → Kind → Kind choose Now Now = Now choose Now Later = Now choose Later Now = Now choose Later Later = Later We define `combine` to apply `choose` to a list of times. combine : ∀{Γ} (ts₁ ts₂ : Times Γ) → Times Γ combine {[]} ts₁ ts₂ = ∅ combine {A ∷ Γ} (cons x ts₁) (cons y ts₂) = cons (choose x y) (combine ts₁ ts₂) Here's the type of the "and" operator: _×ˢ_ : ∀{Γ}{ts₁ ts₂ : Times Γ} → Setˢ Γ ts₁ → Setˢ Γ ts₂ → Setˢ Γ (combine ts₁ ts₂) The other propositions follow a similar pattern. The membership formula `a ∈ x` is true when `a` is in the predicate bound to variable `x` in the environment. The time for `x` is required to be `Now`. var-now : ∀ (Γ : Context) → ∀{A} → (x : Γ ∋ A) → Times Γ var-now (B ∷ Γ) zeroˢ = cons Now (laters Γ) var-now (B ∷ Γ) (sucˢ x) = cons Later (var-now Γ x) _∈_ : ∀{Γ}{A} → A → (x : Γ ∋ A) → Setˢ Γ (var-now Γ x) a ∈ x = record { # = λ δ → (lookup x δ) a ; ... } The `μˢ` formula defines a (possibly nested) recursive predicate. μˢ : ∀{Γ}{ts : Times Γ}{A} → (A → Setˢ (A ∷ Γ) (cons Later ts)) ---------------------------------- → (A → Setˢ Γ ts) It takes a non-recursive predicate from `A` to `Setˢ` and produces a recursive predicate in `A`. Note that the variable `zeroˢ`, the one introduced by this `μˢ`, is required to have time `Later`. If the recursive predicate is not nested inside other recursive predicates, then you can directly use the following `μᵒ` operator. μᵒ : ∀{A} → (A → Setˢ (A ∷ []) (cons Later ∅)) ---------------------------------- → (A → Setᵒ) Let's revisit the example of defining multi-step reduction. The non-recursive `mreduce` predicate is defined as follows. ``` mreduce : Term × Term → Setˢ ((Term × Term) ∷ []) (cons Later ∅) mreduce (M , N) = (M ≡ N)ˢ ⊎ˢ (∃ˢ[ L ] (M —→ L)ˢ ×ˢ ▷ˢ (((L , N) ∈ zeroˢ))) ``` Note that the `R` parameter has become implicit; it has moved into the environment. Also the application `R (L , N)` is replaced by `▷ˢ ((L , N) ∈ zeroˢ)`, where the de Bruijn index `zeroˢ` refers to the predicate `R` in the environment. We define the recursive predicate `M —→* N` by applying `μᵒ` to `mreduce`. ``` infix 2 _—→*_ _—→*_ : Term → Term → Setᵒ M —→* N = μᵒ mreduce (M , N) ``` Here are a couple uses of the multi-step reduction relation. ``` X₀ : #($ (Num 0) —→* $ (Num 0)) 1 X₀ = inj₁ refl X₁ : #((ƛ ($ (Num 1))) · $ (Num 0) —→* $ (Num 1)) 2 X₁ = inj₂ (_ , (β ($̬ _) , inj₁ refl)) ``` ## Proofs in Step-indexed Logic Just like first-orderd logic, SIL comes with rules of deduction for carrying out proofs. The judgement form is `𝒫 ⊢ᵒ P`, where `𝒫` is a list of assumptions and `P` is a formula. The judgement `𝒫 ⊢ᵒ P` is true iff for every time `k`, all of `𝒫` are true at `k` implies that `P` is true at `k`. So in Agda we have the following definition. Πᵒ : List Setᵒ → Setᵒ Πᵒ [] = ⊤ᵒ Πᵒ (P ∷ 𝒫) = P ×ᵒ Πᵒ 𝒫 _⊢ᵒ_ : List Setᵒ → Setᵒ → Set 𝒫 ⊢ᵒ P = ∀ k → # (Πᵒ 𝒫) k → # P k Many of the deduction rules are the same as in first order logic. For example, here are the introduction and elimination rules for conjunction. We use the same notation as Agda, but with a superscript "o". _,ᵒ_ : ∀{𝒫 : List Setᵒ }{P Q : Setᵒ} → 𝒫 ⊢ᵒ P → 𝒫 ⊢ᵒ Q ------------ → 𝒫 ⊢ᵒ P ×ᵒ Q proj₁ᵒ : ∀{𝒫 : List Setᵒ }{P Q : Setᵒ} → 𝒫 ⊢ᵒ P ×ᵒ Q ------------ → 𝒫 ⊢ᵒ P proj₂ᵒ : ∀{𝒫 : List Setᵒ }{P Q : Setᵒ} → 𝒫 ⊢ᵒ P ×ᵒ Q ------------ → 𝒫 ⊢ᵒ Q The introduction rule for a constant formula `S ᵒ` is straightforward. A proof of `S` in regular Agda is sufficient to build a proof of `S ᵒ` in SIL. constᵒI : ∀{𝒫}{S : Set} → S → 𝒫 ⊢ᵒ S ᵒ On the other hand, given a proof of `S ᵒ` in SIL, one cannot obtain a proof of `S` directly in Agda. That is, the following rule is invalid because `𝒫` could be false at every index. bogus-constᵒE : ∀ {𝒫}{S : Set}{R : Setᵒ} → 𝒫 ⊢ᵒ S ᵒ → S Instead, we have an elimination rule in continuation-passing style. That is, if we have a proof of `S ᵒ` and need to prove some arbitrary goal `R`, then it suffices to prove `R` under the assumption that `S` is true. constᵒE : ∀ {𝒫}{S : Set}{R : Setᵒ} → 𝒫 ⊢ᵒ S ᵒ → (S → 𝒫 ⊢ᵒ R) → 𝒫 ⊢ᵒ R Analogous to `subst` in Agda's standard library, SIL has `substᵒ` which says that if `P` and `Q` are equivalent, then a proof of `P` gives a proof of `Q`. substᵒ : ∀{𝒫}{P Q : Setᵒ} → P ≡ᵒ Q ------------------- → 𝒫 ⊢ᵒ P → 𝒫 ⊢ᵒ Q The deduction rules also include ones for the "later" operator. As we mentioned earlier, if a proposition is true now it will also be true later. monoᵒ : ∀ {𝒫}{P} → 𝒫 ⊢ᵒ P ----------- → 𝒫 ⊢ᵒ ▷ᵒ P One can transport induction on natural numbers into SIL to obtain the following Löb rule, which states that when proving any property `P`, one is allowed to assume that `P` is true later. lobᵒ : ∀ {𝒫}{P} → (▷ᵒ P) ∷ 𝒫 ⊢ᵒ P ----------------------- → 𝒫 ⊢ᵒ P For comparison, here's induction on natural numbers P 0 → (∀ k → P k → P (suc k)) → ∀ n → P n In the world of SIL, propositions are always true at zero, so the base case `P 0` is not necessary. The induction step `(∀ k → P k → P (suc k))` is similar to the premise `(▷ᵒ P) ∷ 𝒫 ⊢ᵒ P` because `▷ᵒ` subtracts one. The following is a handy proof rule that turns a proof of `P` in SIL into an assumption in Agda that `P` is true for some positive natural number. ⊢ᵒ-sucP : ∀{𝒫}{P Q : Setᵒ} → 𝒫 ⊢ᵒ P → (∀{n} → # P (suc n) → 𝒫 ⊢ᵒ Q) → 𝒫 ⊢ᵒ Q As usual for temporal logics (or more generally, for modal logics), there are distribution rules that push "later" through the other logical connectives. For example, the following rule distributes "later" through conjunction. ▷× : ∀{𝒫} {P Q : Setᵒ} → 𝒫 ⊢ᵒ (▷ᵒ (P ×ᵒ Q)) ---------------------- → 𝒫 ⊢ᵒ (▷ᵒ P) ×ᵒ (▷ᵒ Q) This project was the first time for me conducting nontrivial proofs in a modal logic, and it took some getting use to! ## Defining a Logical Relation for Type Safety With the Step-indexed Logic in hand, we are ready to define a logical relation for type safety. The two predicates ℰ and 𝒱 are mutually recursive, so we combine them into a single recursive predicate named `ℰ⊎𝒱` that takes a sum type, where the left side is for ℰ and the right side is for 𝒱. We shall define `ℰ⊎𝒱` by an application of `μᵒ`, so we first need to define the non-recursive version of `ℰ⊎𝒱`, which we call `pre-ℰ⊎𝒱`, defined below. It simply dispatches to the non-recursive `pre-ℰ` and `pre-ℰ` which we define next. ``` ℰ⊎𝒱-type : Set ℰ⊎𝒱-type = (Type × Term) ⊎ (Type × Term) ℰ⊎𝒱-ctx : Context ℰ⊎𝒱-ctx = ℰ⊎𝒱-type ∷ [] pre-ℰ : Type → Term → Setˢ ℰ⊎𝒱-ctx (cons Later ∅) pre-𝒱 : Type → Term → Setˢ ℰ⊎𝒱-ctx (cons Later ∅) pre-ℰ⊎𝒱 : ℰ⊎𝒱-type → Setˢ ℰ⊎𝒱-ctx (cons Later ∅) pre-ℰ⊎𝒱 (inj₁ (A , V)) = pre-𝒱 A V pre-ℰ⊎𝒱 (inj₂ (A , M)) = pre-ℰ A M ``` To improve the readability of our definitions, we define the following notation for recursive applications of the ℰ and 𝒱 predicates. ``` ℰˢ⟦_⟧ : Type → Term → Setˢ ℰ⊎𝒱-ctx (cons Now ∅) ℰˢ⟦ A ⟧ M = (inj₂ (A , M)) ∈ zeroˢ 𝒱ˢ⟦_⟧ : Type → Term → Setˢ ℰ⊎𝒱-ctx (cons Now ∅) 𝒱ˢ⟦ A ⟧ V = (inj₁ (A , V)) ∈ zeroˢ ``` The definition of `pre-ℰ` and `pre-𝒱` below are of similar form to the explicitly step-indexed definition of ℰ and 𝒱 above, however the parameter `k` is gone and all of the logical connectives have a superscript `s`, indicating that we're building a `Setˢ`. Also, note that all the uses of `ℰˢ` and `𝒱ˢ` are guarded by the later operator `▷ˢ`. Finally, in the definition of `pre-ℰ`, we do not use `▷ˢ (𝒱⟦ A ⟧ M)` but instead use `pre-𝒱 A M` because we need to say in that spot that `M` is a semantic value now, not later. ``` pre-ℰ A M = (pre-𝒱 A M ⊎ˢ (reducible M)ˢ ⊎ˢ (Blame M)ˢ) ×ˢ (∀ˢ[ N ] (M —→ N)ˢ →ˢ ▷ˢ (ℰˢ⟦ A ⟧ N)) pre-𝒱 ★ (V ⟨ G !⟩ ) = (Value V)ˢ ×ˢ ▷ˢ (𝒱ˢ⟦ gnd⇒ty G ⟧ V) pre-𝒱 ($ₜ ι) ($ c) = (ι ≡ typeof c)ˢ pre-𝒱 (A ⇒ B) (ƛ N) = ∀ˢ[ W ] ▷ˢ (𝒱ˢ⟦ A ⟧ W) →ˢ ▷ˢ (ℰˢ⟦ B ⟧ (N [ W ])) pre-𝒱 A M = ⊥ ˢ ``` We define ℰ and 𝒱 by creating a recursive predicate (apply `μᵒ` to `pre-ℰ⊎𝒱`) and then apply it to an argument injected with either `inj₁` for 𝒱 or `inj₂` for ℰ. ``` ℰ⊎𝒱 : ℰ⊎𝒱-type → Setᵒ ℰ⊎𝒱 X = μᵒ pre-ℰ⊎𝒱 X ℰ⟦_⟧ : Type → Term → Setᵒ ℰ⟦ A ⟧ M = ℰ⊎𝒱 (inj₂ (A , M)) 𝒱⟦_⟧ : Type → Term → Setᵒ 𝒱⟦ A ⟧ V = ℰ⊎𝒱 (inj₁ (A , V)) ``` To succinctly talk about the two aspects of ℰ, we define semantic `progress` and `preservation` as follows. ``` progress : Type → Term → Setᵒ progress A M = 𝒱⟦ A ⟧ M ⊎ᵒ (reducible M)ᵒ ⊎ᵒ (Blame M)ᵒ preservation : Type → Term → Setᵒ preservation A M = ∀ᵒ[ N ] ((M —→ N)ᵒ →ᵒ ▷ᵒ (ℰ⟦ A ⟧ N)) ``` We can prove that ℰ is indeed equivalent to progress and preservation by use of the `fixpointᵒ` theorem in SIL. ``` ℰ-stmt : ∀{A}{M} → ℰ⟦ A ⟧ M ≡ᵒ progress A M ×ᵒ preservation A M ℰ-stmt {A}{M} = ℰ⟦ A ⟧ M ⩦⟨ ≡ᵒ-refl refl ⟩ μᵒ pre-ℰ⊎𝒱 (inj₂ (A , M)) ⩦⟨ fixpointᵒ pre-ℰ⊎𝒱 (inj₂ (A , M)) ⟩ # (pre-ℰ⊎𝒱 (inj₂ (A , M))) (ℰ⊎𝒱 , ttᵖ) ⩦⟨ cong-×ᵒ (cong-⊎ᵒ (≡ᵒ-sym (fixpointᵒ pre-ℰ⊎𝒱 (inj₁ (A , M)))) (≡ᵒ-refl refl)) (≡ᵒ-refl refl) ⟩ progress A M ×ᵒ preservation A M ∎ ``` For convenience, we define introduction and elimination rules for ℰ. ``` ℰ-intro : ∀ {𝒫}{A}{M} → 𝒫 ⊢ᵒ progress A M → 𝒫 ⊢ᵒ preservation A M ---------------------- → 𝒫 ⊢ᵒ ℰ⟦ A ⟧ M ℰ-intro 𝒫⊢prog 𝒫⊢pres = substᵒ (≡ᵒ-sym ℰ-stmt) (𝒫⊢prog ,ᵒ 𝒫⊢pres) ℰ-progress : ∀ {𝒫}{A}{M} → 𝒫 ⊢ᵒ ℰ⟦ A ⟧ M → 𝒫 ⊢ᵒ progress A M ℰ-progress 𝒫⊢ℰM = proj₁ᵒ (substᵒ ℰ-stmt 𝒫⊢ℰM ) ℰ-preservation : ∀ {𝒫}{A}{M} → 𝒫 ⊢ᵒ ℰ⟦ A ⟧ M → 𝒫 ⊢ᵒ preservation A M ℰ-preservation 𝒫⊢ℰM = proj₂ᵒ (substᵒ ℰ-stmt 𝒫⊢ℰM ) ``` Similarly, we can derive the expected equations for 𝒱. ``` 𝒱-base : ∀{ι}{c : Lit} → (𝒱⟦ $ₜ ι ⟧ ($ c)) ≡ᵒ (ι ≡ typeof c)ᵒ 𝒱-base = ≡ᵒ-intro λ k → (λ x → x) , (λ x → x) 𝒱-dyn : ∀{G}{V} → 𝒱⟦ ★ ⟧ (V ⟨ G !⟩) ≡ᵒ ((Value V)ᵒ ×ᵒ ▷ᵒ (𝒱⟦ gnd⇒ty G ⟧ V)) 𝒱-dyn {G}{V} = let X = (inj₁ (★ , V ⟨ G !⟩)) in 𝒱⟦ ★ ⟧ (V ⟨ G !⟩) ⩦⟨ ≡ᵒ-refl refl ⟩ ℰ⊎𝒱 X ⩦⟨ fixpointᵒ pre-ℰ⊎𝒱 X ⟩ # (pre-ℰ⊎𝒱 X) (ℰ⊎𝒱 , ttᵖ) ⩦⟨ ≡ᵒ-refl refl ⟩ (Value V)ᵒ ×ᵒ ▷ᵒ (𝒱⟦ gnd⇒ty G ⟧ V) ∎ 𝒱-fun : ∀{A B}{N} → 𝒱⟦ A ⇒ B ⟧ (ƛ N) ≡ᵒ (∀ᵒ[ W ] ((▷ᵒ (𝒱⟦ A ⟧ W)) →ᵒ (▷ᵒ (ℰ⟦ B ⟧ (N [ W ]))))) 𝒱-fun {A}{B}{N} = let X = (inj₁ (A ⇒ B , ƛ N)) in 𝒱⟦ A ⇒ B ⟧ (ƛ N) ⩦⟨ ≡ᵒ-refl refl ⟩ ℰ⊎𝒱 X ⩦⟨ fixpointᵒ pre-ℰ⊎𝒱 X ⟩ # (pre-ℰ⊎𝒱 X) (ℰ⊎𝒱 , ttᵖ) ⩦⟨ ≡ᵒ-refl refl ⟩ (∀ᵒ[ W ] ((▷ᵒ (𝒱⟦ A ⟧ W)) →ᵒ (▷ᵒ (ℰ⟦ B ⟧ (N [ W ]))))) ∎ ``` We have defined `𝒱` such that it only accepts terms that are syntactic values. (We included `Value V` in the case for `★` of `pre-𝒱`.) ``` 𝒱⇒Value : ∀ {k} A M → # (𝒱⟦ A ⟧ M) (suc k) --------------------- → Value M 𝒱⇒Value ★ (M ⟨ G !⟩) (v , _) = v 〈 G 〉 𝒱⇒Value ($ₜ ι) ($ c) 𝒱M = $̬ c 𝒱⇒Value (A ⇒ B) (ƛ N) 𝒱M = ƛ̬ N ``` A value `V` in 𝒱 is also in ℰ. The definition of `progress` includes values, and to prove preservation we note that a value is irreducible. ``` 𝒱⇒ℰ : ∀{A}{𝒫}{V} → 𝒫 ⊢ᵒ 𝒱⟦ A ⟧ V --------------- → 𝒫 ⊢ᵒ ℰ⟦ A ⟧ V 𝒱⇒ℰ {A}{𝒫}{V} 𝒫⊢𝒱V = ℰ-intro prog pres where prog = inj₁ᵒ 𝒫⊢𝒱V pres = Λᵒ[ N ] →ᵒI (constᵒE Zᵒ λ V—→N → ⊢ᵒ-sucP (⊢ᵒ-weaken 𝒫⊢𝒱V) λ 𝒱V → ⊥-elim (value-irreducible (𝒱⇒Value A V 𝒱V ) V—→N)) ``` ## Semantic Type Safety for Open Terms The `ℰ` predicate applies to closed terms, that is, terms without any free variables, such as a whole program. However, we'll need a notion of semantic type safety that also includes open terms. The standard way to define safety for an open term `M` is to substitute the free variables for values and then use `ℰ`. That is, we apply a substitution `γ` to `M` where all the values in `γ` must be semantically well typed. The following `𝓖` expresses this contraint on `γ`. ``` 𝓖⟦_⟧ : (Γ : List Type) → Subst → List Setᵒ 𝓖⟦ [] ⟧ σ = [] 𝓖⟦ A ∷ Γ ⟧ σ = (𝒱⟦ A ⟧ (σ 0)) ∷ 𝓖⟦ Γ ⟧ (λ x → σ (suc x)) ``` A term `M` is semantically well typed at `A` in context `Γ` if, for any well-typed substitution `γ`, we have `ℰ⟦ A ⟧ (⟪ γ ⟫ M)`. ``` _⊨_⦂_ : List Type → Term → Type → Set Γ ⊨ M ⦂ A = ∀ (γ : Subst) → 𝓖⟦ Γ ⟧ γ ⊢ᵒ ℰ⟦ A ⟧ (⟪ γ ⟫ M) ``` ## The Fundamental Lemma via Compatibility Lemmas The main lemma on our way to proving type safety is the Fundamental Lemma, which states that well-typed programs are semantically type safe. That is, well-typed programs behave as expected according to their types. fundamental : ∀ {Γ A} → (M : Term) → Γ ⊢ M ⦂ A ---------- → Γ ⊨ M ⦂ A The proof of `fundamental` is by induction on the typing derivation, with each case dispatching to a compatibility lemma. The compatibility lemma for number literals is proved by showing that `$ (Num n)` is in `𝒱⟦ $ₜ ′ℕ ⟧` via the definition of `𝒱` and then apply the `𝒱⇒ℰ` lemma. ``` compatible-nat : ∀{Γ}{n : ℕ} ----------------------- → Γ ⊨ $ (Num n) ⦂ ($ₜ ′ℕ) compatible-nat {Γ}{n} γ = 𝒱⇒ℰ (substᵒ (≡ᵒ-sym 𝒱-base) (constᵒI refl)) ``` The compability lemma for Boolean literals is the same. ``` compatible-bool : ∀{Γ}{b : 𝔹} -------------------------- → Γ ⊨ ($ (Bool b)) ⦂ ($ₜ ′𝔹) compatible-bool {Γ}{b} γ = 𝒱⇒ℰ (substᵒ (≡ᵒ-sym 𝒱-base) (constᵒI refl)) ``` The compatibility lemma for the `blame` term is similar to the `𝒱⇒ℰ` lemma in that `blame` is one of the alternatives allowed in `progress` and `blame` is irreducible. ``` ℰ-blame : ∀{𝒫}{A} → 𝒫 ⊢ᵒ ℰ⟦ A ⟧ blame ℰ-blame {𝒫}{A} = ℰ-intro prog pres where prog = inj₂ᵒ (inj₂ᵒ (constᵒI isBlame)) pres = Λᵒ[ N ] →ᵒI (constᵒE Zᵒ λ blame→ → ⊥-elim (blame-irreducible blame→)) compatible-blame : ∀{Γ}{A} ------------- → Γ ⊨ blame ⦂ A compatible-blame {Γ}{A} γ = ℰ-blame ``` The compatibility lemma for variables makes use of the premise that the values in the environment are semantically well typed. The following lemma proves that for any variable `y` in `Γ`, `γ` in `𝓖⟦ Γ ⟧` imples that `γ y` in `𝒱⟦ A ⟧`. ``` lookup-𝓖 : (Γ : List Type) → (γ : Subst) → ∀ {A}{y} → (Γ ∋ y ⦂ A) → 𝓖⟦ Γ ⟧ γ ⊢ᵒ 𝒱⟦ A ⟧ (γ y) lookup-𝓖 (B ∷ Γ) γ {A} {zero} refl = Zᵒ lookup-𝓖 (B ∷ Γ) γ {A} {suc y} ∋y = Sᵒ (lookup-𝓖 Γ (λ x → γ (suc x)) ∋y) ``` Once we have `γ y` in `𝒱⟦ A ⟧`, we conclude by applying the `𝒱⇒ℰ` lemma. (The `sub-var` lemma just says that `⟪ γ ⟫ (` x) ≡ γ x`.) ``` compatibility-var : ∀ {Γ A x} → Γ ∋ x ⦂ A ----------- → Γ ⊨ ` x ⦂ A compatibility-var {Γ}{A}{x} ∋x γ rewrite sub-var γ x = 𝒱⇒ℰ (lookup-𝓖 Γ γ ∋x) ``` The next compatibility lemma is for lambda abstraction. To show that `ƛ N` is in `ℰ⟦A ⇒ B⟧` we shows that `ƛ N` is in `𝒱⟦A ⇒ B⟧`. According to that definition, we need to show that for any argument value `W` in `𝒱⟦ A ⟧` (later), we have `(⟪ ext γ ⟫ N) [ W ]` in `ℰ⟦ B ⟧` (also later). But that follows almost directly from the premise that `N` is semantically type safe. From that premise we have ▷ᵒ ℰ ⟦ B ⟧ (⟪ W • γ ⟫ N) and the Abstract Binding Tree library provides rewrites for the following equation ⟪ W • γ ⟫ N = (⟪ ext γ ⟫ N) [ W ] which gives us what we need: ▷ᵒ ℰ ⟦ B ⟧ (⟪ ext γ ⟫ N) [ W ] Here's all the details in Agda: ``` compatible-lambda : ∀{Γ}{A}{B}{N} → (A ∷ Γ) ⊨ N ⦂ B ------------------- → Γ ⊨ (ƛ N) ⦂ (A ⇒ B) compatible-lambda {Γ}{A}{B}{N} ⊨N γ = 𝒱⇒ℰ ⊢𝒱λN where ⊢𝒱λN : 𝓖⟦ Γ ⟧ γ ⊢ᵒ 𝒱⟦ A ⇒ B ⟧ (ƛ (⟪ ext γ ⟫ N)) ⊢𝒱λN = (substᵒ (≡ᵒ-sym 𝒱-fun) (Λᵒ[ W ] →ᵒI ▷𝓔N[W])) where ▷𝓔N[W] : ∀{W} → ▷ᵒ 𝒱⟦ A ⟧ W ∷ 𝓖⟦ Γ ⟧ γ ⊢ᵒ ▷ᵒ ℰ⟦ B ⟧ ((⟪ ext γ ⟫ N) [ W ]) ▷𝓔N[W] {W} = appᵒ (Sᵒ (▷→ (monoᵒ (→ᵒI (⊨N (W • γ)))))) Zᵒ ``` The next few compatibility lemmas, for application, injection, and projection all involve reasoning about the reduction of one or two subexpressions. Instead of duplicating this reasoning, the standard approach is to put that reasoning in the "bind" lemma, which we discuss next. ## Interlude: the "Bind" Lemma The bind lemma says that if we have an expression `N` with a subexpression `M` (so `N` is equal to plugging `M` into an appropriate frame `F`, i.e. `N = F ⟦ M ⟧`), if `M` is semantically safe, then to prove `ℰ⟦ A ⟧ (F ⟦ M ⟧)` it suffices to prove that `ℰ⟦ A ⟧ (F ⟦ V ⟧))` for some semantically safe value `V` that `M` reduced to. ℰ-bind : ∀{𝒫}{A}{B}{F}{M} → 𝒫 ⊢ᵒ ℰ⟦ B ⟧ M → 𝒫 ⊢ᵒ (∀ᵒ[ V ] (M —↠ V)ᵒ →ᵒ 𝒱⟦ B ⟧ V →ᵒ ℰ⟦ A ⟧ (F ⟦ V ⟧)) ---------------------------------------------------------- → 𝒫 ⊢ᵒ ℰ⟦ A ⟧ (F ⟦ M ⟧) In the title of this blog post I alluded to one hard lemma. This is the one! We begin by creating some names for parts of the statement of this lemma. First we have a name for the second premise. ``` 𝒱V→ℰF[V] : Type → Type → Frame → Term → Setᵒ 𝒱V→ℰF[V] A B F M = ∀ᵒ[ V ] (M —↠ V)ᵒ →ᵒ 𝒱⟦ B ⟧ V →ᵒ ℰ⟦ A ⟧ (F ⟦ V ⟧) ``` Then we have a name for the two premises and the conclusion, with the implications expressed in SIL. ``` ℰ-bind-M : Type → Type → Frame → Term → Setᵒ ℰ-bind-M A B F M = ℰ⟦ B ⟧ M →ᵒ 𝒱V→ℰF[V] A B F M →ᵒ ℰ⟦ A ⟧ (F ⟦ M ⟧) ``` The following adds universal quantification (in SIL) over the term `M`. ``` ℰ-bind-prop : Type → Type → Frame → Setᵒ ℰ-bind-prop A B F = ∀ᵒ[ M ] ℰ-bind-M A B F M ``` We shall need the `𝒱V→ℰF[V]` property to be preserved under reverse reduction, i.e., expansion. The proof is as follows. We need to show that `ℰ⟦ A ⟧ (F ⟦ V ⟧)` under the assumption that `M′ —↠ V` and `𝒱⟦ B ⟧ V`. With the first premise `M —→ M′`, we obtain `M —↠ V`. Then we apply the second premise to conclude that `ℰ⟦ A ⟧ (F ⟦ V ⟧)`. ``` 𝒱V→ℰF[V]-expansion : ∀{𝒫}{A}{B}{F}{M}{M′} → M —→ M′ → 𝒫 ⊢ᵒ 𝒱V→ℰF[V] A B F M ----------------------- → 𝒫 ⊢ᵒ 𝒱V→ℰF[V] A B F M′ 𝒱V→ℰF[V]-expansion {𝒫}{A}{B}{F}{M}{M′} M→M′ 𝒱V→ℰF[V][M] = Λᵒ[ V ] let M′→V→ℰFV : 𝒱⟦ B ⟧ V ∷ (M′ —↠ V)ᵒ ∷ 𝒫 ⊢ᵒ ℰ⟦ A ⟧ (F ⟦ V ⟧) M′→V→ℰFV = ⊢ᵒ-sucP (Sᵒ Zᵒ) λ M′→V → let M—↠V = constᵒI (M —→⟨ M→M′ ⟩ M′→V) in let M→V→ℰFV = ⊢ᵒ-weaken(⊢ᵒ-weaken(instᵒ 𝒱V→ℰF[V][M] V)) in appᵒ (appᵒ M→V→ℰFV M—↠V) Zᵒ in →ᵒI (→ᵒI M′→V→ℰFV) ``` We now proceed to prove the `ℰ-bind` lemma by way of an auxilliary lemma `ℰ-bind-aux` that restates the lemma so that the term `M` is universally quantified in SIL (instead of Agda), so that we can do the proof by Löb induction, that is, by use of the `lobᵒ` rule of SIL. So after the use of `lobᵒ`, it remains to prove that `ℰ⟦ A ⟧ (F ⟦ M ⟧)`, but now we have the additional assumption that we can apply the bind lemma in the future to any term, i.e., we have `▷ᵒ ℰ-bind-prop A B F`. From the premise `ℰ⟦ B ⟧ M` we have that `M` satisfies progress, so either (1) it is a semantic value, (2) it can reduce, or (3) it is blame. We proceed by reasoning about each of these three cases. * `M` is already a value, so it can multi-step reduce to itself in zero steps, and then we apply the `𝒱V→ℰF[V]` premise to immediately conclude. * `M` is reducible. Now to prove `ℰ⟦ A ⟧ (F ⟦ M ⟧)` we need to prove progress and preservation. The progress part is immediate, because by rule `ξ` we have `F ⟦ M ⟧ —→ F ⟦ M′ ⟧` because `M —→ M′ for some `M′`. The preservation part is more involved. We are given that `F ⟦ M ⟧ —→ N` and need to prove that `▷ᵒ (ℰ⟦ A ⟧ N)`. By the `frame-inv2` lemma, we obtain an `M′` such that `M —→ M′` and `N ≡ F ⟦ M′ ⟧`. So we need to prove that `▷ᵒ (ℰ⟦ A ⟧ (F ⟦ M′ ⟧))` We shall obtain this via the induction hypothesis, and for that we need to prove (1) `▷ᵒ ℰ⟦ B ⟧ M′` and (2) `▷ᵒ (𝒱V→ℰF[V] A B F M′)`. We obtain (1) from the preservation part of `ℰ⟦ B ⟧ M`. We obtain (2) by the `𝒱V→ℰF[V]-expansion` lemma and shift it to later using `monoᵒ`. * `M` is blame. We need to show `ℰ⟦ A ⟧ (F ⟦ blame ⟧)`. For the progress part, we have the reduction `F ⟦ blame ⟧ —→ blame` by rule `ξ-blame`. For preservation, we have `F ⟦ blame ⟧ —→ N` and need to prove that `▷ᵒ (ℰ⟦ A ⟧ N)`. The `blame-frame` lemma tells us that `N ≡ blame`, so we conclude by use of `ℰ-blame` and then `monoᵒ`. ``` open import rewriting.examples.CastDeterministic using (frame-inv2; deterministic) ℰ-bind-aux : ∀{𝒫}{A}{B}{F} → 𝒫 ⊢ᵒ ℰ-bind-prop A B F ℰ-bind-aux {𝒫}{A}{B}{F} = lobᵒ (Λᵒ[ M ] →ᵒI (→ᵒI Goal)) where Goal : ∀{M} → (𝒱V→ℰF[V] A B F M) ∷ ℰ⟦ B ⟧ M ∷ ▷ᵒ ℰ-bind-prop A B F ∷ 𝒫 ⊢ᵒ ℰ⟦ A ⟧ (F ⟦ M ⟧) Goal{M} = case3ᵒ (ℰ-progress (Sᵒ Zᵒ)) Mval Mred Mblame where 𝒫′ = (𝒱V→ℰF[V] A B F M) ∷ ℰ⟦ B ⟧ M ∷ ▷ᵒ ℰ-bind-prop A B F ∷ 𝒫 Mval : 𝒱⟦ B ⟧ M ∷ 𝒫′ ⊢ᵒ ℰ⟦ A ⟧ (F ⟦ M ⟧) Mval = let 𝒱V→ℰF[V][M] = λ V → (M —↠ V)ᵒ →ᵒ 𝒱⟦ B ⟧ V →ᵒ ℰ⟦ A ⟧ (F ⟦ V ⟧) in appᵒ (appᵒ (instᵒ{P = 𝒱V→ℰF[V][M]} (Sᵒ Zᵒ) M) (constᵒI (M END))) Zᵒ Mred : (reducible M)ᵒ ∷ 𝒫′ ⊢ᵒ ℰ⟦ A ⟧ (F ⟦ M ⟧) Mred = ℰ-intro progressMred preservationMred where progressMred : (reducible M)ᵒ ∷ 𝒫′ ⊢ᵒ progress A (F ⟦ M ⟧) progressMred = inj₂ᵒ (inj₁ᵒ (constᵒE Zᵒ λ {(M′ , M→M′) → constᵒI (_ , (ξ F M→M′))})) preservationMred : (reducible M)ᵒ ∷ 𝒫′ ⊢ᵒ preservation A (F ⟦ M ⟧) preservationMred = (constᵒE Zᵒ λ redM → ⊢ᵒ-weaken (Λᵒ[ N ] →ᵒI (constᵒE Zᵒ λ FM→N → ⊢ᵒ-weaken (redM⇒▷ℰN redM FM→N)))) where redM⇒▷ℰN : ∀{N} → reducible M → (F ⟦ M ⟧ —→ N) → 𝒫′ ⊢ᵒ ▷ᵒ (ℰ⟦ A ⟧ N) redM⇒▷ℰN {N} rM FM→N = let finv = frame-inv2{M}{N}{F} rM FM→N in let M′ = proj₁ finv in let M→M′ = proj₁ (proj₂ finv) in let N≡ = proj₂ (proj₂ finv) in let ▷ℰM′ : 𝒫′ ⊢ᵒ ▷ᵒ ℰ⟦ B ⟧ M′ ▷ℰM′ = appᵒ (instᵒ{P = λ N → (M —→ N)ᵒ →ᵒ ▷ᵒ (ℰ⟦ B ⟧ N)} (ℰ-preservation (Sᵒ Zᵒ)) M′) (constᵒI M→M′) in let ▷M′→V→𝒱V→ℰFV : 𝒫′ ⊢ᵒ ▷ᵒ (𝒱V→ℰF[V] A B F M′) ▷M′→V→𝒱V→ℰFV = monoᵒ (𝒱V→ℰF[V]-expansion{𝒫′}{A}{B} M→M′ Zᵒ) in let IH : 𝒫′ ⊢ᵒ ▷ᵒ ℰ-bind-prop A B F IH = Sᵒ (Sᵒ Zᵒ) in let ▷ℰFM′ : 𝒫′ ⊢ᵒ ▷ᵒ (ℰ⟦ A ⟧ (F ⟦ M′ ⟧)) ▷ℰFM′ = frame-prop-lemma IH ▷ℰM′ ▷M′→V→𝒱V→ℰFV in subst (λ N → 𝒫′ ⊢ᵒ ▷ᵒ ℰ⟦ A ⟧ N) (sym N≡) ▷ℰFM′ where frame-prop-lemma : ∀{𝒫}{A}{B}{M}{F} → 𝒫 ⊢ᵒ ▷ᵒ ℰ-bind-prop A B F → 𝒫 ⊢ᵒ ▷ᵒ ℰ⟦ B ⟧ M → 𝒫 ⊢ᵒ ▷ᵒ 𝒱V→ℰF[V] A B F M → 𝒫 ⊢ᵒ ▷ᵒ (ℰ⟦ A ⟧ (F ⟦ M ⟧)) frame-prop-lemma{𝒫}{A}{B}{M}{F} IH ℰM V→FV = appᵒ(▷→ (appᵒ(▷→ (instᵒ(▷∀{P = λ M → ℰ-bind-M A B F M} IH) M)) ℰM)) V→FV Mblame : (Blame M)ᵒ ∷ 𝒫′ ⊢ᵒ ℰ⟦ A ⟧ (F ⟦ M ⟧) Mblame = ℰ-intro progressMblame (constᵒE Zᵒ λ blameM → ⊢ᵒ-weaken (Λᵒ[ N ] →ᵒI (constᵒE Zᵒ λ FM→N → ⊢ᵒ-weaken (blameM⇒▷ℰN blameM FM→N)))) where progressMblame : (Blame M)ᵒ ∷ 𝒫′ ⊢ᵒ progress A (F ⟦ M ⟧) progressMblame = inj₂ᵒ (inj₁ᵒ (constᵒE Zᵒ λ {isBlame → constᵒI (_ , (ξ-blame F))})) blameM⇒▷ℰN : ∀{N} → Blame M → (F ⟦ M ⟧ —→ N) → 𝒫′ ⊢ᵒ ▷ᵒ (ℰ⟦ A ⟧ N) blameM⇒▷ℰN {N} isBlame FM→N = let eq = blame-frame FM→N in subst (λ N → 𝒫′ ⊢ᵒ ▷ᵒ ℰ⟦ A ⟧ N) (sym eq) (monoᵒ ℰ-blame) ``` The `ℰ-bind` lemma follows as a corollary of `ℰ-bind-aux`. ``` ℰ-bind : ∀{𝒫}{A}{B}{F}{M} → 𝒫 ⊢ᵒ ℰ⟦ B ⟧ M → 𝒫 ⊢ᵒ (∀ᵒ[ V ] (M —↠ V)ᵒ →ᵒ 𝒱⟦ B ⟧ V →ᵒ ℰ⟦ A ⟧ (F ⟦ V ⟧)) ---------------------------------------------------------- → 𝒫 ⊢ᵒ ℰ⟦ A ⟧ (F ⟦ M ⟧) ℰ-bind {𝒫}{A}{B}{F}{M} ⊢ℰM ⊢𝒱V→ℰFV = appᵒ (appᵒ (instᵒ{𝒫}{P = λ M → ℰ-bind-M A B F M} ℰ-bind-aux M) ⊢ℰM) ⊢𝒱V→ℰFV ``` ## More Compatibility Lemmas The next compatibility lemma to prove is the one for function application. For that we'll need the following elimination lemma for a value `V` in `𝒱⟦ A ⇒ B ⟧`. ``` safe-body : List Setᵒ → Term → Type → Type → Set safe-body 𝒫 N A B = ∀{W} → 𝒫 ⊢ᵒ (▷ᵒ (𝒱⟦ A ⟧ W)) →ᵒ (▷ᵒ (ℰ⟦ B ⟧ (N [ W ]))) 𝒱-fun-elim : ∀{𝒫}{A}{B}{V}{R} → 𝒫 ⊢ᵒ 𝒱⟦ A ⇒ B ⟧ V → (∀ N → V ≡ ƛ N → safe-body 𝒫 N A B → 𝒫 ⊢ᵒ R) ------------------------------------------------ → 𝒫 ⊢ᵒ R 𝒱-fun-elim {𝒫}{A}{B}{V}{R} ⊢𝒱V cont = ⊢ᵒ-sucP ⊢𝒱V λ { 𝒱Vsn → G {V} 𝒱Vsn ⊢𝒱V cont} where G : ∀{V}{n} → # (𝒱⟦ A ⇒ B ⟧ V) (suc n) → 𝒫 ⊢ᵒ 𝒱⟦ A ⇒ B ⟧ V → (∀ N → V ≡ ƛ N → safe-body 𝒫 N A B → 𝒫 ⊢ᵒ R) → 𝒫 ⊢ᵒ R G{ƛ N}{n} 𝒱V ⊢𝒱V cont = cont N refl λ {W} → instᵒ{P = λ W → (▷ᵒ (𝒱⟦ A ⟧ W)) →ᵒ (▷ᵒ (ℰ⟦ B ⟧ (N [ W ])))} (substᵒ 𝒱-fun ⊢𝒱V) W ``` The proof of compatibility for application, given below, starts with two uses of the `ℰ-bind` lemma, once for subexpression `L` and again for `M`. So we obtain that `L` reduces to value `V` and `M` reduces to `W` and that `𝒱⟦ A ⇒ B ⟧ V` and `𝒱⟦ A ⟧ W`. At this point, our goal is to show that `ℰ⟦ B ⟧ (V · W)`. Next we use the elimination lemma on `𝒱⟦ A ⇒ B ⟧ V` which tells us that `V` is a lambda abstraction `ƛ N` with a semantically safe body `N`. We thus obtain the `progress` part of `ℰ⟦ B ⟧ (V · W)` because `(ƛ N) · W —→ N [ W ]`. For the preservation part, we need to show that `ℰ⟦ B ⟧ (N [ W ])`, but that follows from `𝒱⟦ A ⟧ W` and that `N` is a semantically safe body. ``` compatible-app : ∀{Γ}{A}{B}{L}{M} → Γ ⊨ L ⦂ (A ⇒ B) → Γ ⊨ M ⦂ A ------------------- → Γ ⊨ L · M ⦂ B compatible-app {Γ}{A}{B}{L}{M} ⊨L ⊨M γ = ⊢ℰLM where ⊢ℰLM : 𝓖⟦ Γ ⟧ γ ⊢ᵒ ℰ⟦ B ⟧ (⟪ γ ⟫ (L · M)) ⊢ℰLM = ℰ-bind {F = □· (⟪ γ ⟫ M)} (⊨L γ) (Λᵒ[ V ] →ᵒI (→ᵒI ⊢ℰVM)) where 𝒫₁ = λ V → 𝒱⟦ A ⇒ B ⟧ V ∷ (⟪ γ ⟫ L —↠ V)ᵒ ∷ 𝓖⟦ Γ ⟧ γ ⊢ℰVM : ∀{V} → 𝒫₁ V ⊢ᵒ ℰ⟦ B ⟧ (V · ⟪ γ ⟫ M) ⊢ℰVM {V} = ⊢ᵒ-sucP Zᵒ λ 𝒱Vsn → let v = 𝒱⇒Value (A ⇒ B) V 𝒱Vsn in let 𝒫₁⊢ℰM : 𝒫₁ V ⊢ᵒ ℰ⟦ A ⟧ (⟪ γ ⟫ M) 𝒫₁⊢ℰM = Sᵒ (Sᵒ (⊨M γ)) in ℰ-bind {F = v ·□} 𝒫₁⊢ℰM (Λᵒ[ V ] →ᵒI (→ᵒI ⊢ℰVW)) where 𝒫₂ = λ V W → 𝒱⟦ A ⟧ W ∷ (⟪ γ ⟫ M —↠ W)ᵒ ∷ 𝒱⟦ A ⇒ B ⟧ V ∷ (⟪ γ ⟫ L —↠ V)ᵒ ∷ 𝓖⟦ Γ ⟧ γ ⊢ℰVW : ∀{V W} → 𝒫₂ V W ⊢ᵒ ℰ⟦ B ⟧ (V · W) ⊢ℰVW {V}{W} = let ⊢𝒱V : 𝒫₂ V W ⊢ᵒ 𝒱⟦ A ⇒ B ⟧ V ⊢𝒱V = Sᵒ (Sᵒ Zᵒ) in let ⊢𝒱W : 𝒫₂ V W ⊢ᵒ 𝒱⟦ A ⟧ W ⊢𝒱W = Zᵒ in ⊢ᵒ-sucP ⊢𝒱W λ 𝒱Wsn → let w = 𝒱⇒Value A W 𝒱Wsn in 𝒱-fun-elim ⊢𝒱V λ {N′ refl 𝒱W→ℰNW → let prog : 𝒫₂ (ƛ N′) W ⊢ᵒ progress B (ƛ N′ · W) prog = (inj₂ᵒ (inj₁ᵒ (constᵒI (_ , (β w))))) in let pres : 𝒫₂ (ƛ N′) W ⊢ᵒ preservation B (ƛ N′ · W) pres = Λᵒ[ N ] →ᵒI (constᵒE Zᵒ λ {r → let ⊢▷ℰN′W = appᵒ 𝒱W→ℰNW (monoᵒ ⊢𝒱W) in let eq = deterministic r (β w) in ⊢ᵒ-weaken (subst (λ N → 𝒫₂ (ƛ N′) W ⊢ᵒ ▷ᵒ ℰ⟦ B ⟧ N) (sym eq) ⊢▷ℰN′W)}) in ℰ-intro prog pres } ``` The compability lemma for an injection cast also begins with applying the bind lemma to subexpression `M`, taking us from `ℰ⟦ gnd⇒ty G ⟧ M` to `𝒱⟦ gnd⇒ty G ⟧ V`. This also gives us that `V` is a syntactic value via `𝒱⇒Value`. So we have `𝒱⟦ ★ ⟧ (V ⟨ G !⟩)` and then conclude using `𝒱⇒ℰ`. ``` compatible-inject : ∀{Γ}{G}{M} → Γ ⊨ M ⦂ gnd⇒ty G -------------------- → Γ ⊨ M ⟨ G !⟩ ⦂ ★ compatible-inject {Γ}{G}{M} ⊨M γ = ℰMg! where ℰMg! : 𝓖⟦ Γ ⟧ γ ⊢ᵒ ℰ⟦ ★ ⟧ ((⟪ γ ⟫ M) ⟨ G !⟩) ℰMg! = ℰ-bind {F = □⟨ G !⟩} (⊨M γ) (Λᵒ[ V ] →ᵒI (→ᵒI ⊢ℰVg!)) where 𝒫₁ = λ V → 𝒱⟦ gnd⇒ty G ⟧ V ∷ (⟪ γ ⟫ M —↠ V)ᵒ ∷ 𝓖⟦ Γ ⟧ γ ⊢ℰVg! : ∀{V} → 𝒫₁ V ⊢ᵒ ℰ⟦ ★ ⟧ (V ⟨ G !⟩) ⊢ℰVg!{V} = ⊢ᵒ-sucP Zᵒ λ 𝒱Vsn → let v = 𝒱⇒Value (gnd⇒ty G) V 𝒱Vsn in 𝒱⇒ℰ (substᵒ (≡ᵒ-sym 𝒱-dyn) (constᵒI v ,ᵒ monoᵒ Zᵒ)) ``` The last compatibility lemma is for a projection cast. Here we also need an elimination lemma, this time for a value `V` of type `★`. ``` 𝒱-dyn-elim : ∀{𝒫}{V}{R} → 𝒫 ⊢ᵒ 𝒱⟦ ★ ⟧ V → (∀ W G → V ≡ W ⟨ G !⟩ → 𝒫 ⊢ᵒ ((Value W)ᵒ ×ᵒ ▷ᵒ (𝒱⟦ gnd⇒ty G ⟧ W)) → 𝒫 ⊢ᵒ R) ---------------------------------------------- → 𝒫 ⊢ᵒ R 𝒱-dyn-elim {𝒫}{V}{R} ⊢𝒱V cont = ⊢ᵒ-sucP ⊢𝒱V λ { 𝒱Vsn → G 𝒱Vsn ⊢𝒱V cont } where G : ∀{V}{n} → # (𝒱⟦ ★ ⟧ V) (suc n) → 𝒫 ⊢ᵒ 𝒱⟦ ★ ⟧ V → (∀ W G → V ≡ W ⟨ G !⟩ → 𝒫 ⊢ᵒ ((Value W)ᵒ ×ᵒ ▷ᵒ (𝒱⟦ gnd⇒ty G ⟧ W)) → 𝒫 ⊢ᵒ R) → 𝒫 ⊢ᵒ R G {W ⟨ G !⟩}{n} 𝒱Vsn ⊢𝒱V cont with 𝒱⇒Value ★ (W ⟨ G !⟩) 𝒱Vsn ... | w 〈 _ 〉 = let ⊢▷𝒱W = proj₂ᵒ (substᵒ (𝒱-dyn{V = W}) ⊢𝒱V) in cont W _ refl (constᵒI w ,ᵒ ⊢▷𝒱W) ``` The compatibility lemma for a projection `M ⟨ H ?⟩` begins by using `ℰ-bind` on the subexpression `M` to obtain a value `V` where `⟪ γ ⟫ M —↠ V` and `𝒱⟦ ★ ⟧ V`. We then apply lemma `𝒱-dyn-elim` to compose `V` into an injection `W ⟨ G !⟩` of a value `W` where `▷ᵒ 𝒱⟦ G ⟧ W`. We need to show `ℰ⟦ H ⟧ (W ⟨ G !⟩ ⟨ H ?⟩)`. The progress part comes from showing that it reduces to `W` (if `G ≡ H`) or to `blame`. The preservation part is from `▷ᵒ 𝒱⟦ G ⟧ W` (in the `G ≡ H` case) or because `ℰ⟦ H ⟧ blame`. ``` compatible-project : ∀{Γ}{H}{M} → Γ ⊨ M ⦂ ★ ----------------------------- → Γ ⊨ M ⟨ H ?⟩ ⦂ gnd⇒ty H compatible-project {Γ}{H}{M} ⊨M γ = ℰMh? where ℰMh? : 𝓖⟦ Γ ⟧ γ ⊢ᵒ ℰ⟦ gnd⇒ty H ⟧ ((⟪ γ ⟫ M) ⟨ H ?⟩) ℰMh? = ℰ-bind {F = □⟨ H ?⟩} (⊨M γ) (Λᵒ[ V ] →ᵒI (→ᵒI ⊢ℰVh?)) where 𝒫₁ = λ V → 𝒱⟦ ★ ⟧ V ∷ (⟪ γ ⟫ M —↠ V)ᵒ ∷ 𝓖⟦ Γ ⟧ γ ⊢ℰVh? : ∀{V} → 𝒫₁ V ⊢ᵒ ℰ⟦ gnd⇒ty H ⟧ (V ⟨ H ?⟩) ⊢ℰVh?{V} = let ⊢𝒱V : 𝒫₁ V ⊢ᵒ 𝒱⟦ ★ ⟧ V ⊢𝒱V = Zᵒ in 𝒱-dyn-elim ⊢𝒱V λ { W G refl ⊢w×▷𝒱W → let ⊢w = proj₁ᵒ ⊢w×▷𝒱W in let ▷𝒱W = proj₂ᵒ ⊢w×▷𝒱W in ⊢ᵒ-sucP ⊢w λ{n} w → let prog : 𝒫₁ (W ⟨ G !⟩) ⊢ᵒ progress (gnd⇒ty H) ((W ⟨ G !⟩) ⟨ H ?⟩) prog = inj₂ᵒ (inj₁ᵒ (constᵒI (reduce-inj-proj w))) in let pres : 𝒫₁ (W ⟨ G !⟩) ⊢ᵒ preservation (gnd⇒ty H)((W ⟨ G !⟩) ⟨ H ?⟩) pres = Λᵒ[ N ] →ᵒI (constᵒE Zᵒ λ r → ⊢ᵒ-weaken (Goal r w ▷𝒱W)) in ℰ-intro prog pres } where reduce-inj-proj : ∀{G}{H}{W} → Value W → reducible ((W ⟨ G !⟩) ⟨ H ?⟩) reduce-inj-proj {G} {H} {W} w with G ≡ᵍ H ... | yes refl = W , (collapse w refl) ... | no neq = blame , (collide w neq refl) Goal : ∀{W}{G}{H}{N} → (W ⟨ G !⟩ ⟨ H ?⟩) —→ N → Value W → 𝒫₁ (W ⟨ G !⟩) ⊢ᵒ ▷ᵒ 𝒱⟦ gnd⇒ty G ⟧ W → 𝒫₁ (W ⟨ G !⟩) ⊢ᵒ ▷ᵒ ℰ⟦ gnd⇒ty H ⟧ N Goal (ξξ □⟨ H ?⟩ refl refl r) w ▷𝒱W = ⊥-elim (value-irreducible (w 〈 _ 〉) r) Goal {W} (ξξ-blame □⟨ H ?⟩ ()) Goal {W}{G}{G}{W} (collapse{H} w′ refl) w ▷𝒱W = ▷→▷ ▷𝒱W (𝒱⇒ℰ Zᵒ) Goal {W} (collide x x₁ x₂) w ▷𝒱W = monoᵒ ℰ-blame ``` ## Fundamental Lemma The Fundamental Lemma states that a syntactically well-typed term is also a semantically well-typed term. Or given how we have defined the logical relations, it means that a well-typed term satisfies progress and preservation. ``` fundamental : ∀ {Γ A} → (M : Term) → Γ ⊢ M ⦂ A ---------- → Γ ⊨ M ⦂ A fundamental {Γ} {A} .(` _) (⊢` ∋x) = compatibility-var ∋x fundamental {Γ} {.($ₜ ′ℕ)} .($ (Num _)) (⊢$ (Num n)) = compatible-nat fundamental {Γ} {.($ₜ ′𝔹)} .($ (Bool _)) (⊢$ (Bool b)) = compatible-bool fundamental {Γ} {A} (L · M) (⊢· ⊢L ⊢M) = compatible-app{L = L}{M} (fundamental L ⊢L) (fundamental M ⊢M) fundamental {Γ} {.(_ ⇒ _)} (ƛ N) (⊢ƛ ⊢N) = compatible-lambda {N = N} (fundamental N ⊢N) fundamental {Γ} {.★} (M ⟨ G !⟩) (⊢⟨!⟩ ⊢M) = compatible-inject {M = M} (fundamental M ⊢M) fundamental {Γ} {A} (M ⟨ H ?⟩) (⊢⟨?⟩ ⊢M H) = compatible-project {M = M} (fundamental M ⊢M) fundamental {Γ} {A} .blame ⊢blame = compatible-blame ``` ## Proof of Type Safety For the Type Safety theorem, we need to consider multi-step reduction. So we first prove the following lemma which states that if `M —↠ N` and `M` is in `ℰ⟦ A ⟧`, then `N` satisfies progress. The lemma is by induction on the multi-step reduction, using the preservation part of `ℰ⟦ A ⟧` at each step. ``` sem-type-safety : ∀ {A} → (M N : Term) → (r : M —↠ N) → # (ℰ⟦ A ⟧ M) (suc (len r)) --------------------------------------------- → Value N ⊎ (∃[ N′ ] (N —→ N′)) ⊎ N ≡ blame sem-type-safety {A} M .M (.M END) (inj₁ 𝒱M , presM) = inj₁ (𝒱⇒Value A M 𝒱M) sem-type-safety {A} M .M (.M END) (inj₂ (inj₁ r) , presM) = inj₂ (inj₁ r) sem-type-safety {A} M .M (.M END) (inj₂ (inj₂ isBlame) , presM) = inj₂ (inj₂ refl) sem-type-safety {A} M N (_—→⟨_⟩_ .M {M′} M→M′ M′→N) (_ , presM) = let ℰM′ : # (ℰ⟦ A ⟧ M′) (suc (len M′→N)) ℰM′ = presM M′ (suc (suc (len M′→N))) ≤-refl M→M′ in sem-type-safety M′ N M′→N ℰM′ ``` The Type Safety theorem is then a corollary of the Fundamental Lemma together with the above lemma regarding multi-step reduction. ``` type-safety : ∀ {A} → (M N : Term) → [] ⊢ M ⦂ A → M —↠ N --------------------------------------------- → Value N ⊎ (∃[ N′ ] (N —→ N′)) ⊎ N ≡ blame type-safety M N ⊢M M→N = let ℰM = ⊢ᵒ-elim ((fundamental M ⊢M) id) (suc (len M→N)) tt in sem-type-safety M N M→N ℰM ```
Monday, April 17, 2023
Type Safety in 10 Easy, 4 Medium, and 1 Hard Lemma using Step-indexed Logical Relations
Subscribe to:
Post Comments (Atom)
No comments:
Post a Comment