Linearity in the non-deterministic call-by-value setting

Linearity in the non-deterministic
call-by-value setting

Alejandro Díaz-Caro Barbara Petit
LIPN, Université Paris 13, Sorbonne Paris Cité Focus (INRIA) – Università di Bologna

19th WoLLIC
Buenos Aires, September 3–6, 2012

Non-determinism (or parallelism)
t + u: non-deterministic superposition between t and u
(it may run t or u, non-deterministically)

Alejandro Díaz-Caro & Barbara Petit Linearity in the non-deterministic call-by-value setting 1 / 20

(t + u)s may run ts or us.


Hence (t + u)s → ts + us


Sum of functions: (f + g)(x) = f(x) + g(x)
(this is the interpretation of the algebraic calculi [Vaux, Arrighi-Dowek])


t(u + s)

(λx.t)(u + s) → t[(u + s)/x] (CBN)
(λx.t)(u + s) → (λx.t)u + (λx.t)s (CBV)

Generically in CBV t(u + s) → tu + ts


t(u + s)

(λx.t)(u + s) → t[(u + s)/x] (CBN)

In algebraic terms, functions are treated as linear in CBV:
f(x + y ) = f(x) + f(y )


t(u + s)

(λx.t)(u + s) → t[(u + s)/x] (CBN)

In algebraic terms, functions are treated as linear in CBV:
f(x + y ) = f(x) + f(y )

We want to understand this ‘linearity’


Outline

The untyped calculus

The Additive type system capturing the CBV behaviour of +

Logical interpretation: translation into System F with pairs


The untyped CBV non-deterministic λ-calculus

t, u, s ::= v | tu | t + u | 0
v ::= x | λx.t
Intuitions
t + u = non-deterministic superposition between t and u
0 = impossible computation



t, u, s ::= v | tu | t + u | 0
v ::= x | λx.t
Intuitions

Hence, t + 0 → t because 0 is impossible!



t, u, s ::= v | tu | t + u | 0
v ::= x | λx.t
Intuitions

Also 0t → 0 t0 → 0
Because non of them reduces (0 not a value) and there is an impossible
computation



t, u, s ::= v | tu | t + u | 0
v ::= x | λx.t
Intuitions

Also 0t → 0 t0 → 0
Because non of them reduces (0 not a value) and there is an impossible
computation

Running t or u non-deterministically, is the
Finally t + u = u + t
same as running u or t non-deterministically
Also t + (u + s) = (t + u) + s



t, u, s ::= v | tu | t + u | 0
v ::= x | λx.t

βv -reduction
(λx.t)v → t[v/x]

Distributivity rules Zero rules
(t + u)s → ts + us t+0→t
t(u + s) → tu + ts 0t → 0
t0 → 0

with + associative and commutative

Remark: in the algebraic case, 0 is the sum of zero terms


The Additive type system
Objective: capture as much as possible the behaviour of +



Γ t:T Γ s:S
Γ t+s :T +S Γ 0:¯
0



Γ t:T Γ s:S Γ t:T T ≡S
Γ t+s :T +S Γ 0:¯
0 Γ t:S

T +R ≡ R +T ; T +(R +S) ≡ (T +R)+S ; T +¯ ≡ T
0



Γ t+s :T +S Γ 0:¯
0 Γ t:S

T +R ≡ R +T ; T +(R +S) ≡ (T +R)+S ; T +¯ ≡ T
0

T → R: functions from T to R



Γ t+s :T +S Γ 0:¯
0 Γ t:S

T +R ≡ R +T ; T +(R +S) ≡ (T +R)+S ; T +¯ ≡ T
0

But (T + R) → S does not exists
there is no function taking a non-deterministic superposition as argument
Recall t(v1 + v2 ) → tv1 + tv2
If v1 : T and v2 : R then t needs to be both T → S1 and R → S2



Γ t+s :T +S Γ 0:¯
0 Γ t:S

T +R ≡ R +T ; T +(R +S) ≡ (T +R)+S ; T +¯ ≡ T
0

But (T + R) → S does not exists
there is no function taking a non-deterministic superposition as argument
Recall t(v1 + v2 ) → tv1 + tv2
If v1 : T and v2 : R then t needs to be both T → S1 and R → S2
Polymorphism & unit types “U” (atomic types w.r.t. +)
Arrows: U → T

The Additive type system (cont.)
Examples

Concrete example

v1 : U1 ; v2 : U2 ; I : ∀X .X → X

Hence I (v1 + v2 ) → I v1 + I v2 has type U1 + U2

A more generic example

v1 : U[W1 /X ] ; v2 : U[W2 /X ] ; t : ∀X .U → T

Hence t(v1 + v2 ) → tv1 + tv2 has type T [W1 /X ] + T [W2 /X ]


Examples

Γ t : ∀X .U → T Γ v1 + v2 : U[W1 /X ] + U[W2 /X ]
Γ t(v1 + v2 ) : T [W1 /X ] + T [W2 /X ]

A more generic example

v1 : U[W1 /X ] ; v2 : U[W2 /X ] ; t : ∀X .U → T

Hence t(v1 + v2 ) → tv1 + tv2 has type T [W1 /X ] + T [W2 /X ]


Examples

t:U→T ; u:V →R

(t + u)v → tv + uv
Hence U and V needs to “polymorph” to the type of v

v : U[W1 /X ] = V [W2 /X ] t : ∀X .(U → T ) u : ∀X .(V → R)

Hence tv + uv has type T [W1 /X ] + R[W2 /X ]


Examples

t:U→T ; u:V →R

(t + u)v → tv + uv
Hence U and V needs to “polymorph” to the type of v

v : U[W1 /X ] = V [W2 /X ] t : ∀X .(U → T ) u : ∀X .(V → R)

Hence tv + uv has type T [W1 /X ] + R[W2 /X ]

Γ t + u : ∀X .(U → T ) + ∀X .(V → R) Γ v : U[W1 /X ] = V [W2 /X ]
Γ (t + u)v : T [W1 /X ] + R[W2 /X ]


Arrow elimination
Γ t : ∀X .U → T Γ v1 + v2 : U[W1 /X ] + U[W2 /X ]
Γ t(v1 + v2 ) : T [W1 /X ] + T [W2 /X ]

Γ (t + u)v : T [W1 /X ] + R[W2 /X ]

combined...
Γ t + u : ∀X .(U → T ) + ∀X .(U → R) Γ v1 + v2 : U[W1 /X ] + U[W2 /X ]
Γ (t + u)(v1 + v2 ) : T [W1 /X ] + T [W2 /X ] + R[W1 /X ] + R[W2 /X ]


Arrow elimination
Γ t : ∀X .U → T Γ v1 + v2 : U[W1 /X ] + U[W2 /X ]
Γ t(v1 + v2 ) : T [W1 /X ] + T [W2 /X ]

Γ (t + u)v : T [W1 /X ] + R[W2 /X ]

combined...
Γ t + u : ∀X .(U → T ) + ∀X .(U → R) Γ v1 + v2 : U[W1 /X ] + U[W2 /X ]
Γ (t + u)(v1 + v2 ) : T [W1 /X ] + T [W2 /X ] + R[W1 /X ] + R[W2 /X ]

generalising...
n m
Γ t: ∀X .(U → Ti ) Γ u: U[Wj /X ]
i=1 j=1
n m
Γ tu : Ti [Wj /X ]
i=1 j=1


Examples

V1 = U[W1 /X ] ; V2 = U[W2 /X ]

Γ λx.t + λy .u : ∀X .(U → T ) + ∀X .(U → R) Γ v1 + v2 : V1 + V2
Γ (λx.t + λy .u)(v1 + v2 ) : T [W1 /X ] + T [W2 /X ] + R[W1 /X ] + R[W2 /X ]

(λx.t)v1 + (λx.t)v2 + (λy .u)v1 + (λy .u)v2
T [W1 /X ] T [W2 /X ] R[W1 /X ] R[W2 /X ]

Alejandro Díaz-Caro Barbara Petit Linearity in the non-deterministic call-by-value setting 8 / 20

Examples

V1 = U[W1 /X ] ; V2 = U[W2 /X ]

Γ λx.t + λy .u : ∀X .(U → T ) + ∀X .(U → R) Γ v1 + v2 : V1 + V2
Γ (λx.t + λy .u)(v1 + v2 ) : T [W1 /X ] + T [W2 /X ] + R[W1 /X ] + R[W2 /X ]

(λx.t)v1 + (λx.t)v2 + (λy .u)v1 + (λy .u)v2
T [W1 /X ] T [W2 /X ] R[W1 /X ] R[W2 /X ]

Simpler example
Γ λx.x : ∀X .X → X Γ v1 + v2 : U + V
Γ (λx.x)(v1 + v2 ) : U + V
without simultaneous arrow/forall elimination it is not possible to type it!

Summarising

T , R, S := U | T + R | ¯
0 general types
U, V , W := X | U → T | ∀X .U unit types
T +¯ ≡ T
0 ; T +R ≡ R +T ; T +(R +S) ≡ (T +R)+S

ax Γ t:T Γ u:R
ax¯
0 +I
Γ, x : U x :U Γ 0:¯
0 Γ t+u:T +R
n m
Γ t: ∀X .(U → Ti ) Γ u: U[Wj /X ]
i=1 j=1 Γ, x : U t:T
→E →I
n m
Γ λx.t : U → T
Γ tu : Ti [Wj /X ]
i=1 j=1
Γ t : ∀X .U Γ t:U X ∈ FV (Γ)
/ Γ t:T T ≡R
∀E ∀I ≡
Γ t : U[V /X ] Γ t : ∀X .U Γ t:R

• Strong normalisation
• Subject reduction


Outline

The untyped calculus

The Additive type system capturing the CBV behaviour of +

Logical interpretation: translation into System F with pairs


System F with pairs

t, u ::= x | λx.t | tu | | t, u | π1 (t) | π2 (t)
A, B ::= X | A ⇒ B | ∀X .A | 1 | A × B

(λx.t)u → t[u/x] π1 ( t1 , t2 ) → t1 π2 ( t1 , t2 ) → t2

Additive System F with pairs
X X
U→T |U| ⇒ |T |
∀X .U ∀X .|U|
¯
0 1
T +S |T | × |S|


Sums as Pairs
+, ¯
0 ×, 1

T +S ≡ S +T A×B = B ×A
T + (S + R) ≡ (T + S) + R A × (B × C ) = (A × B) × C
T +¯ ≡ T
0 A×1 = A


Sums as Pairs
+, ¯
0 ×, 1

T +S ≡ S +T A×B = B ×A
T + (S + R) ≡ (T + S) + R A × (B × C ) = (A × B) × C
T +¯ ≡ T
0 A×1 = A
T , R, S ::= U | T + R | ¯ 0
U, V , W ::= X | U → T | ∀X .U
Type Binary tree (leaf: U or ¯
0)

Example:
T = (U1 + U2 ) + ¯
0 ¯
0
T [r → ¯ lr → U2 , ll → U1 ]
0,
U1 U2

We keep structured sum types


Structured Arrow-elimination
n m
Γ t: ∀X .(U → Ti ) Γ u: U[Wj /X ]
i=1 j=1
n m
Γ tu : Ti [Wj /X ]
i=1 j=1
Γ t : T [ → ∀X .(U → T )] Γ u : T [ → U[W /X ]]
Γ tu : T ◦ T [ → T [W /X ]]

tu:
t: u:

U[W1 /X ] U[W2 /X ]
∀X .U→T1 ∀X .U→T2

T1 [W1 /X ] T1 [W2 /X ] T2 [W1 /X ] T2 [W2 /X ]


Structured Arrow-elimination
An example

t = (t1 + t2 ) + 0 tu →∗ t1 u1 + (u2 + u3 ) +
t2 u1 + (u2 + u3 ) + 0
→∗ t1 u1 + (t1 u2 + t1 u3 )+
t2 u1 + (t2 u2 + t2 u3 ) + 0
¯
0

U → T1 U → T2
¯
0
u = u1 + (u2 + u3 )

U T1 T2

U U T1 T1 T2 T2


Equivalence in System F

What about associativity, commutativity and neutral element
in system F with pairs?

Lemma
T ≡T implies |T | ↔ |T |

Where A ↔ B means

F εA,B : A ⇒ B and F εB,A : B ⇒ A

for some terms εA,B , εB,A s.t.

εA,B ◦ εB,A ≈ idA and εB,A ◦ εA,B ≈ idB


Translation of terms
What happens with the distributivity?

t+u [t], [u]

(t1 + t2 )(u1 + u2 ) →∗ but t1 , t2 r1 , r2
t1 u1 + t1 u2 + t2 u1 + t2 u2 t1 r1 , t1 r2 , t2 r1 , t2 r2
No distributivity in System F with pairs



t+u [t], [u]

(t1 + t2 )(u1 + u2 ) →∗ but t1 , t2 r1 , r2

Main ideas:
The sum is distributed during the translation of application
[tr] = [t1 ][u1 ], [t1 ][u2 ] , [t2 ][u1 ], [t2 ][u2 ]
if t = t1 + t2 and u = u1 + u2



t+u [t], [u]

(t1 + t2 )(u1 + u2 ) →∗ but t1 , t2 r1 , r2

Main ideas:
The sum is distributed during the translation of application
[tr] = [t1 ][u1 ], [t1 ][u2 ] , [t2 ][u1 ], [t2 ][u2 ]
if t = t1 + t2 and u = u1 + u2
The “sum structure” of a term is known thanks to its type
Γ t : (T1 + T2 ) + T3 t ∼ (t1 + t2 ) + t3
with Γ ti : Ti



Γ t:T |Γ| F [t]D : |T |

Γ, x : T x : T [x]D = x
Γ 0:¯0 [0]D =



Γ t:T |Γ| F [t]D : |T |

Γ, x : T x : T [x]D = x
Γ 0:¯0 [0]D =
Γ t:T T ≡T
[t]D = ε|T |,|T | [t]D
Γ t:T



Γ t:T |Γ| F [t]D : |T |

Γ, x : T x : T [x]D = x
Γ 0:¯0 [0]D =
Γ t:T T ≡T
[t]D = ε|T |,|T | [t]D
Γ t:T
Γ t:T Γ u:S
[t+r]D = [t]D1 , [u]D2
Γ t+u:T +S
Γ, x : U t : T
[λx.t]D = λx.[t]D
Γ λx.t : U → T



Γ t:T |Γ| F [t]D : |T |

Γ, x : T x : T [x]D = x
Γ 0:¯0 [0]D =
Γ t:T T ≡T
[t]D = ε|T |,|T | [t]D
Γ t:T
Γ t:T Γ u:S
[t+r]D = [t]D1 , [u]D2
Γ t+u:T +S
Γ, x : U t : T
[λx.t]D = λx.[t]D
Γ λx.t : U → T
Γ t : T [ →∀X .(U→T )] Γ u : T [ →U[W /X ]]
Γ tu : T ◦ T [ →T [W /X ]]
[tu]D = T ◦ T [ →π ([t]D1 )π ([u]D2 )]


An example

Γ t : (U → T1 ) + (U → T2 ) + (U → T3 ) Γ u:U +U

Γ tu : (T1 + T1 ) + (T2 + T2 ) + (T3 + T3 )

t1 =π11 ([t]);t2 =π12 ([t]);t3 =π2 ([t]);
r1 =π1 ([u]);r2 =π2 ([u]);

tu : [tu]D =

T3 T3
T1 T1T2 T2 t3 r1 t3 r2

t1 r1 t1 r2 t2 r1 t2 r2


Correctness

Theorem (Correctness w.r.t. typing)
Γ t:T implies |Γ| F [t]D : |T |

We provide a partial inverse translation (| · |) and prove

Theorem (Inverse translation)
Γ t:T = (||Γ||) F (|[t]D |) : (||T ||)

Theorem (Reduction preservation)
Γ t : T and t → t implies [t]D →+ [t ]D for some D
(except for t + 0 → t)


Summarising and concluding

+ “like linear functions”: f(x + y ) = f(x) + f(y )
Additive type system tight to the behaviour of +
Typed system translate to System F with pairs




System F with pairs correspond to the non-linear fragment of
IMELL




System F with pairs correspond to the non-linear fragment of
IMELL

Linear Logic CBV nd/alg
Force unique use of the argument Ban sum terms substitutions
e.g. e.g.

f(x) = x 2 no linear f(x + y ) → f(x) + f(y )
f(x) = x + 1 linear

In the CBV non-deterministic setting (or algebraic) it is enough to treat
functions as linear, even if they are not. (In LL all functions are linear).


Linearity in the non-deterministic call-by-value setting

More Related Content

What's hot (10)

Similar to Linearity in the non-deterministic call-by-value setting (20)

More from Alejandro Díaz-Caro (13)

Recently uploaded (20)

Linearity in the non-deterministic call-by-value setting