Example 1: Addition of ordinal numbers in SML

Slide 1A Semantic Analysis of Structural RecursionAndreas AbelMay 10, 1999Example 1: Addition of ordinal numbers in SMLdatatype Nat = . . .Slide 2datatype Ord = O| S of Ord| Lim of Nat –> Ord;fun addord x O = x| addord x (S y ′ ) = S (addord x y ′ )| addord x (Lim f) = Lim (fn z => addord x (f z))WST ’99 Andreas Abel 1

Example 2: A pattern matching proof in LEGOSlide 3$[leRefl: {T|ClTy}{t|ClTm T}{v:Val T Ts0 t}vLe v v];[[S,T:ClTy][t:ClTm T][v:Val T Ts0 t][s:ClTm S][w:Val S Ts0 s][R:Ty one][r:ClTm (UnfoldRec R)][x:ClV r]leRefl vUnit ==> leUnit|| leRefl (vInl S v) ==> leInl S S (leRefl v)|| leRefl (vInr S v) ==> leInr S S (leRefl v)|| leRefl (vPair v w) ==> lePair (leRefl v) (leRefl w)|| leRefl (vFold R x) ==> leFoldl R (leFoldr R (leRefl x))];Goal: From structural recursiveness . . .∀v. (∀w < v. f(w) ⇓) → f(v) ⇓... infer terminationOutline:∀v. f(v) ⇓Slide 41. Def. of the foetus system: types σ ∈ Ty( ⃗ X), terms t ∈ Tm σ [Γ]2. Def. of the evaluation strategy: syntactic values v ∈ Val σ , closures 〈t; e〉 ∈ Cl σ ,op. sem. ⇓⊆ Cl σ × Val σ3. Def. of the semantics: “good” values v ∈ [[σ]]4. Def. of the structural ordering < σ,τ ⊆ [[σ]] × [[τ]]5. Proof of the wellfoundedness [[σ]] w.r.t.

The foetus systemType Terms / Values Explanation(Unit) 1 () unit set(Var) X, Y, Z, . . . – type variablesSlide 5(Sum) σ + τ inl, inr, case disjoint sum(Prod) σ × τ (–,–), fst, snd product(Arr) σ → τ( ⃗ X) λ, rec, – – (app) function space(Rec) Rec X.σ(X) fold, unfold recursive (fixed-point) typeσ(Rec X.σ(X))fold−−−−−→←−−−−−unfoldRec X.σ(X)Example 3: Recursor for Nat in foetusNat ≡ Rec X. 1 + XSlide 6O ≡ fold(inl())S(v) ≡ fold(inr(v))R σ ≡ rec R σ→(Nat→σ→σ)→Nat→σ . λfO σ Nat→σ→σ. λfS . λn Nat .case(unfold(n),1 . f O ,n ′Nat . f S n ′ (R f O f S n ′ ))WST ’99 Andreas Abel 3

Example 4: addord in foetusOrd ≡ Rec X.(1 + X) + (Nat → X)Slide 7O ≡ fold(inl(inl()))S(v) ≡ fold(inl(inr(v)))Lim(f) ≡ fold(inr(f))addOrd ≡ rec addOrd Ord→Ord→Ord . λx Ord . λy Ord . case(unfold(y),n 1+Ord . case(n,1 . x,y ′Ord . S(addOrd x y ′ ))f Nat→Ord . Lim(λz Nat . addOrd x (f z)))Operational semanticsClosures Cl σ :〈t σ ; e〉 t term, e environmentSlide 8f ρ→σ @u ρf function value, u argument valueEvaluation relation ⇓ σ ⊆ Cl σ × Val σ :• big step• call-by-value• fixed evaluation strategyWST ’99 Andreas Abel 4

SemanticsLet ⃗ V ⊆ Val ⃗τ . Define [[σ( ⃗ X)]] ⃗Vinductively:(Unit) [[1]] := {()}(Var)[[X n ]] ⃗V := V nSlide 9(Sum) [[(σ + τ)( ⃗ X)]] ⃗V := {inl(v) : v ∈ [[σ( ⃗ X)]] ⃗V } ∪ {inr(v) : v ∈ [[τ( ⃗ X)]] ⃗V }(Arr) [[σ → τ( ⃗ X)]] ⃗V := {f ∈ Val σ→τ(⃗τ) : ∀u ∈ [[σ]]. ∃v ∈ [[τ( ⃗ X)]] ⃗V .f@u ⇓ v}(Rec)[[Rec Y.σ( ⃗ X, Y )]] ⃗V := lfp F, where we define F as(F : PVal Rec Y.σ(⃗τ,Y )) → PW ↦→ fold([[σ( X, ⃗ )Y )]] ⃗V ,WRec Y.σ(⃗τ,Y(Val))Fixed-pointLet (U, ⊆) be a complete lattice, F : U → U an operator. The least fixed-pointF = lfp F is characterized by:(ispfp)(ismpfp)F(F ) ⊆ F∀A ∈ U. F(A) ⊆ A → F ⊆ ASlide 10Monotonicity∀σ(X). A ⊆ B → [[σ(X)]] A⊆ [[σ(X)]] BProof: Induction on σ:(Arr)Show: For all f ∈ [[σ → τ(X)]] Aand u ∈ [[σ]] there is av ∈ [[τ(X)]] Bsatisfying f@u ⇓ v.(Rec) Show: [[Rec Z.σ(X, Z)]] A⊆ [[Rec Z.σ(X, Z)]] B.WST ’99 Andreas Abel 5

SubstitutionSlide 11[[σ(X)]] [τ ] = [[σ(τ)]]For V ⊆ [[τ]] we get[[σ(X)]] V⊆ [[σ(τ)]]Example 5: All numerals are goodVal Nat = {(fold ◦ inr) n (inl()) : n ∈ N} ! = [[Nat]]Slide 12Show: Val Nat is smallest fixed-point of(F : P Val Nat) → P(Val Nat)F(W ) := {fold(v) : v ∈ [[1 + X]] W}= {fold(inl()), fold(inr(v)) : v ∈ W }We must only show (ismpfp):(i)(ii)⋃n∈NF n (∅) ⊆ WVal Nat ⊆ ⋃ n∈NF n (∅)WST ’99 Andreas Abel 6

Structural orderingIdea: Values are trees, “

Structural recursive termsSR σ→τ [Γ] := {rec g.t ∈ Tm σ→τ [Γ] : ∀e ∈ [[Γ]], v ∈ [[σ]].(∀[[σ]] ∋ w < v.〈rec g.t; e〉@w ⇓) → 〈rec g.t; e〉@v ⇓}Slide 15Induction principle for wellfounded sets:(accind)∀v ∈ [[σ]]. (∀[[σ]] ∋ w < v. P (w)) → P (v)∀v ∈ Acc σ . P (v)All structural recursive terms are good:t ∈ SR σ→τ [Γ], e ∈ [[Γ]] → 〈t; e〉 ∈ [[σ → τ]]Example 6: addord is structural recursiveRecursive calls:Slide 16. . . S(addOrd x ′ y) . . .lereflv ′ ≤ v ′ltinrv ′ < inr(v ′ )leltv ′ ≤ inr(v ′ )ltinlv ′ < inl(inr(v ′ ))ltfoldv ′ < S(v ′ ) ≡ fold(inl(inr(v ′ ))). . . Lim(λz Nat . addOrd (f z) y) . . .lereflw ≤ w∃I∃v ′ ∈ CoDom(f). w ≤ v ′ learrw ≤ fltinrw < inr(f)ltfoldw < Lim(f) ≡ fold(inr(f))WST ’99 Andreas Abel 8

NormalizationDefine the good terms TM σ [Γ] ⊂ Tm σ [Γ] inductively in the same way as Tm withthe exceptionSlide 17(REC)t ∈ TM σ→τ [Γ, g σ→τ ]rec g.t ∈ TM σ→τ [Γ]rec g.t ∈ SR σ→τ [Γ]Show normalization∀σ, Γ, t ∈ TM σ [Γ], e ∈ [[Γ]]. 〈t; e〉 ⇓by induction on t using the operational sematics.Extensions and open questionsSlide 18• positive types (?)• polymorphic types √• dependent types• coinductive typesWST ’99 Andreas Abel 9