Sémantique des jeux asynchrones et réécriture 2-dimensionnelle

Transcription

1 Sémantique des jeux asynchrones et réécriture 2-dimensionnelle Soutenance de thèse de doctorat Samuel Mimram Laboratoire PPS (CNRS Université Paris Diderot) 1 er décembre / 64

2 A program is a text in a programming language. 2 / 64

3 3 / 64

4 A program is a text in a programming language. We want to give a meaning to this language! 4 / 64

5 The Curry-Howard correspondence Typing programs can avoid some errors 1 : int 5 / 64

6 The Curry-Howard correspondence Typing programs can avoid some errors fun x x+1 : int int 5 / 64

7 The Curry-Howard correspondence Typing programs can avoid some errors fun x not x : bool bool 5 / 64

8 The Curry-Howard correspondence Typing programs can avoid some errors fun x not x : bool bool The Curry-Howard correspondence: type of a program = formula program = proof of its type Studying programs is the same as studying proofs. 5 / 64

9 Denotational semantics A model interprets a type A as a computation space A a program f : A B as a transformation f : A B A B : f f : A B 6 / 64

10 Denotational semantics A denotational model interprets a type A as a computation space A a program f : A B as a transformation f : A B A B : f β f f : A B f denotational semantics = program invariants 6 / 64

11 Here, a program will be modeled by its interactive behavior i.e. by the way it reacts to information provided by its environment. 7 / 64

12 Here, a program will be modeled by its interactive behavior i.e. by the way it reacts to information provided by its environment. (fun x not x)false true (fun x not x)true false Game Semantics! 7 / 64

13 Outline Programs do actions in a particular order... We study the causality induced by programs in two frameworks: 1 Asynchronous Games 2 2-Dimensional Rewriting 8 / 64

14 Part I Causality in Asynchronous Games 9 / 64

15 Game semantics A type A is interpreted as a game. A program f : A is interpreted as a strategy playing on the game associated to A. 10 / 64

16 Event structures Definition An event structure (E,, #) is a set E of events (or moves) equipped with a partial order of causal dependency and a binary relation # of incompatibility. q V # F 11 / 64

17 Games Definition A game is an event structure (M,, #) equipped with a polarisation function λ : M { 1, +1} which indicates if a move is Opponent or Player. bool = q V # F 12 / 64

18 The game bool : q V # F 13 / 64

19 The game bool bool : q q V # F V # F 13 / 64

20 Strategies Definition A play is a sequence of moves which respects the partial order and the incompatibility relation. Definition A strategy is a set of plays. 14 / 64

21 The strategy not: bool bool 15 / 64

22 The strategy not: bool bool q q 15 / 64

23 The strategy not: bool bool q q V F 15 / 64

24 The strategy not: bool bool q q F V 15 / 64

25 The characterizations of definable strategies capture the interactive behavior of programs. Two series of works laid the foundations of game semantics: fully complete models of MLL [AJ,HO] fully abstract models of PCF [AJM,HON] imposing conditions on strategies: innocence, bracketing,... extended later on to model various programming features: references, control, non-determinism, / 64

26 How can we extend these results to concurrent languages? 17 / 64

27 Concurrent computations Computations are more and more performed in parallel: networks, multi-core processors, etc. 18 / 64

28 Concurrent computations Computations are more and more performed in parallel: networks, multi-core processors, etc. which raises new problems: synchronization shared resources 18 / 64

29 Concurrent computations Computations are more and more performed in parallel: networks, multi-core processors, etc. which raises new problems: synchronization shared resources which raises new questions: how can we describe these computations? (π-calculus, bigraphs, interaction nets,... ) how can we type them? how can we model them? 18 / 64

30 We are going to define a game semantics which captures concurrency in programs and proofs. 19 / 64

31 Revisiting game semantics Our games are asynchronous: sequential plays are replaced by Mazurkiewicz traces non-alternating. 20 / 64

32 Linear Logic We consider here MALL formulas (without units): Γ, A, B Γ, A ` B (`) Γ 1, A Γ 2, B Γ 1, Γ 2, A B ( ) Γ, A Γ, B (&) Γ, A & B Γ, A Γ, A B ( ) 21 / 64

33 Mixing points of view Asynchronous games provide a model in which we can recover trace semantics (games), causal semantics (event structures), relational semantics, concurrent semantics (closure operators), 22 / 64

34 Mixing points of view Asynchronous games provide a model in which we can recover trace semantics (games), causal semantics (event structures), relational semantics, concurrent semantics (closure operators), in which we characterize a fully complete model of MLL (without units), HO innocent strategies in LL, 22 / 64

35 Mixing points of view Asynchronous games provide a model in which we can recover trace semantics (games), causal semantics (event structures), relational semantics, concurrent semantics (closure operators), in which we characterize a fully complete model of MLL (without units), HO innocent strategies in LL, by using ideas coming from game semantics, concurrency theory, linear logic, category theory. 22 / 64

36 The asynchronous graph of a game bool q V # F V q {q} {q, V } {q, F } F Position: downward-closed set of compatible moves. Play: path from the initial position. Strategy: prefix-closed set of plays. 23 / 64

37 The asynchronous graph of a game o q q {q} 23 / 64

38 The asynchronous graph of a game o o q L q R {q L } q L q R {q R } q R {q L, q R } q L 23 / 64

39 The asynchronous graph of a game o o q L q R {q L } q L q R {q R } q R {q L, q R } q L 23 / 64

40 Implementations of conjunction Left implementation of conjunction: bool bool bool q q L V L q R F R F 24 / 64

41 Implementations of conjunction Right implementation of conjunction: bool bool bool q q R F R q L V L F 24 / 64

42 Strategies of conjunction The game bool bool bool contains eight subgraphs: q q q q q L q R q q V q V L q R q q q q L F R F q q R V q q V L F R q F q q L F R V F q V L F V F F 25 / 64

43 Strategies of conjunction Left implementation of conjunction: q L q q q R q q q q V L q R q L V q q q q F R q R F q V q q V L F R q L F R q F q V F q V L F V F F 25 / 64

44 Strategies of conjunction Right implementation of conjunction: V q V L q R q q V q q q L q R F R V L q q q R q q F q R L q q q F q F R q L q F q V L V F q F V F F 25 / 64

45 Strategies of conjunction Parallel implementation of conjunction: q q q L q R q q q q V L q R q L F R V q q q q F q V q L F R R q L V q q q F q F R V L V F q F V F F 25 / 64

46 We want to understand the behavior of strategies generated by proofs in linear logic. We are going to enforce a series of diagrammatic axioms on our strategies. 26 / 64

47 An MLL proof.. A, A B, B A B, A, B ( ) A B, A ` B (`). C, C A B, C ( ), (A ` B) C (A B ) ` C (`), (A ` B) C ((A B ) ` C (`) ) ` ((A ` B) C) A formula induces a partial order on its connectives. A proof is a way to explore the formula i.e. a partial order on the moves which refines the partial order of the game. 27 / 64

48 The Cube property Every partial order induces an asynchronous graph. Conversely, every asynchronous graph such that m x 1 x o x 2 n m x 1 x o x 2 y 3 n n x 3 y 1 n y 1 y 2 o y m y 2 o y m has its homotopy classes are characterized by a partial order on the moves appearing in the paths. 28 / 64

49 Asynchronous games Definition A game is a pointed asynchronous graph satisfying the Cube property. 29 / 64

50 We are going to consider strategies σ which are positional (or history-free): σ s x and s t x and t x σ implies x t σ u y y u i.e. characterized by the subgraph of the game they explore. satisfy the Cube property. 30 / 64

51 These properties are not preserved by composition of strategies! 31 / 64

52 Definition A strategy σ : A is deterministic when Ingenuous strategies x σ m n σ y 1 y 2 implies x σ m n σ y 1 y 2 σ n m σ z where m is a Player move. Definition A strategy σ is ingenuous when 1 it satisfies the preceding conditions, 2 it is deterministic. 32 / 64

53 Property Ingenuous strategies compose and form a -autonomous category (which is compact closed). 33 / 64

54 This category still has too many strategies! A B = A ` B 34 / 64

55 Halting positions In the spirit of the relational model, a strategy σ should be characterized by its set σ of halting positions. Definition A halting position of a strategy σ is a position x such that there is no Player move m : x y that σ can play. 35 / 64

56 The game bool bool contains the subgraph: q L q R q q V V L q R q q q L F R F q R V q V L F R q F q L F R V F V L 36 / 64

57 The pair true false: q L q V L q R V q q V q L R V q qr q q L F R FR q L q F F F R V L V F 36 / 64

58 The left biased pair true false: q L q R q q V L q R q L V q q V q L R F R V q q F FR q L F F R V F V L 36 / 64

59 Courteous strategies Definition An ingenuous strategy σ is courteous when it satisfies σ m y 1 x n y 2 implies x σ m n σ y 1 y 2 σ n z m σ n m σ z where m is a Player move. Theorem A courteous ingenuous strategy σ is characterized by its set σ of halting positions. 37 / 64

60 Concurrent strategies The halting positions of such a strategy σ : A are precisely the fixpoints of a closure operator on the positions of A. We thus recover the model of concurrent strategies. A semantical counterpart of the focalization property: strategies can play all their Player moves in one cluster of moves. q L q R q q V L q R q L F R V q q F V q L F R R q L V q q F F R V L V F 38 / 64

61 The operation ( ) from the category of games to the category of relations is not functorial! Games Proofs Rel 39 / 64

62 This mismatch is essentially due to deadlock situations occurring during the interaction. 40 / 64

63 Scheduled strategies: avoiding deadlocks Composing the right implementation of the conjunction with the left biased pair true false leads to a deadlock. 41 / 64

64 Scheduled strategies: avoiding deadlocks Composing the right implementation of the conjunction with the left biased pair true false leads to a deadlock. 41 / 64

65 Scheduled strategies: avoiding deadlocks Composing the right implementation of the conjunction with the left biased pair true false leads to a deadlock. The culprit is the pair, since it induces dependencies between components of the tensor product bool bool. This can be detected by a dynamic scheduling criterion, which enforces an oriented version of the correctness criterion. A B = (A B A B) 41 / 64

66 Outcome We reconstruct the concurrent game model (closure operators). We thus obtain a model of MALL, fully complete for MLL (without units). Two diagrammatic axioms enable us to characterize HO innocent strategies in this model. Asynchronous games constitute a rich and unifying framework in which we can study concurrent situations and compare various models. 42 / 64

67 Part II Causality in String Diagrams 43 / 64

68 The structure of causality We have studied the structure of interactive traces generated by proofs, but what is the structure of causality between moves? We have given an external characterization of strategies generated by proofs (by restricting the space of strategies), can we give an internal characterization by generating those strategies? We are going to give a presentation of a category of games and strategies in a simple case. 44 / 64

69 First-order propositional logic Formulas: Rules: A ::= x.a x.a P(x)... Γ P, Γ x.p, ( ) (with x FV(Γ, )) Γ P[t/x], Γ x.p, ( ) 45 / 64

70 Causality in proofs π Γ A, B, Γ A, y.b, ( ) Γ x.a, y.b, ( ) O O O O 46 / 64

71 Causality in proofs π Γ A, B, Γ A, y.b, ( ) Γ x.a, y.b, ( ) π Γ A, B, Γ x.a, B, ( ) Γ x.a, y.b, ( ) O O implies O O O O O O 46 / 64

72 Causality in proofs π Γ A[t/x], B[t /y], Γ A[t/x], y.b, ( ) Γ x.a, y.b, ( ) π Γ A[t/x], B[t /y], Γ x.a, B[t /y], ( ) Γ x.a, y.b, ( ) P P implies P P P P P P 46 / 64

73 Causality in proofs π Γ A[t/x], B, Γ A[t/x], y.b, ( ) Γ x.a, y.b, ( ) π Γ A[t/x], B, Γ x.a, B, ( ) Γ x.a, y.b, ( ) P O implies P O O P O P 46 / 64

74 Causality in proofs π Γ A, B[t/y], Γ A, y.b, ( ) Γ x.a, y.b, ( ) π Γ A, B[t/y], Γ x.a, B[t/y], ( ) Γ x.a, y.b, ( ) O P P O 46 / 64

75 Causality in proofs π Γ A, B[t/y], Γ A, y.b, ( ) Γ x.a, y.b, ( ) π Γ A, B[t/y], Γ x.a, B[t/y], ( ) Γ x.a, y.b, ( ) O P P O Only when x FV(t)! 46 / 64

76 Causality in proofs Essential causal dependencies induced by proofs are x y where the witness t given for y admits x as free variable. 47 / 64

77 Games A formula x. y. z. t.p will be interpreted by the game (= polarized poset) m 1 m 2 m 3 m 4 48 / 64

78 Games A sequent x.p(x) y. z.p(y) P(z) will be interpreted by the game (= polarized poset) m n 1 n 2 48 / 64

79 A proof. P(x) P(x) P(x) P(x) z.p(x) P(z) ( ) P(x) y. z.p(y) P(z) ( ) will be interpreted by the strategy x.p(x) y. z.p(y) P(z) ( ) m n 1 n 2 49 / 64

80 Causal strategies game A = partial order on the moves strategy σ = relation on moves 50 / 64

81 Causal strategies game A = partial order on the moves strategy σ = relation on moves A strategy σ : A is causal when 1 if m σ n then m Opponent and n Player 2 the relation A σ is acyclic 50 / 64

82 Causal strategies game A = partial order on the moves strategy σ = relation on moves A strategy σ : A is causal when 1 if m σ n then m Opponent and n Player 2 the relation A σ is acyclic Forbids: Admits: 50 / 64

83 Causal strategies game A = partial order on the moves strategy σ = relation on moves A strategy σ : A is causal when 1 if m σ n then m Opponent and n Player 2 the relation A σ is acyclic Forbids: and Allows: 50 / 64

84 Definition A presentation of a monoid M is given by a set G of generators, a set R G G of relations, such that Presentations of monoids M = G R = G / Example N = a N/2N = a aa = 1 N N = a, b ab = ba etc. 51 / 64

85 Presentations of categories More generally, a polygraph presents an n-category by giving typed generators of dimension i (for 0 i n), typed relations of dimension n / 64

86 Illustration: the simplicial category The simplicial category has as objects: sets [n] = {0, 1,..., n 1} where n N, as morphisms: weakly increasing functions. 53 / 64

87 Illustration: the simplicial category The simplicial category has as objects: sets [n] = {0, 1,..., n 1} where n N, as morphisms: weakly increasing functions : [4] [2] / 64

88 The simplicial category has Illustration: the simplicial category as objects: sets [n] = {0, 1,..., n 1} where n N, as morphisms: weakly increasing functions. It is a category: horizontal composition ( ) / 64

89 Illustration: the simplicial category The simplicial category has as objects: sets [n] = {0, 1,..., n 1} where n N, as morphisms: weakly increasing functions. It is a category: horizontal composition ( ) / 64

90 Illustration: the simplicial category The simplicial category has as objects: sets [n] = {0, 1,..., n 1} where n N, as morphisms: weakly increasing functions. This category is monoidal: vertical composition ( ) : [4] [2] / 64

91 The simplicial category has Illustration: the simplicial category as objects: sets [n] = {0, 1,..., n 1} where n N, as morphisms: weakly increasing functions. This category is monoidal: vertical composition ( ) : [4] [2] : [1] [2] 1 53 / 64

92 Illustration: the simplicial category The simplicial category has as objects: sets [n] = {0, 1,..., n 1} where n N, as morphisms: weakly increasing functions. This category is monoidal: vertical composition ( ) : [5] [4] / 64

93 A theory of monoids The category contains two morphisms: µ : [2] [1] and η : [0] [1] / 64

94 A theory of monoids The category contains two morphisms: µ : [2] [1] and η : [0] [1] which satisfy = and = = 54 / 64

95 A theory of monoids Property The morphisms µ and η generate / 64

96 A theory of monoids Property The morphisms µ and η generate / 64

97 A presentation of the category The category is isomorphic to the free monoidal category on the two generators µ : [2] [1] and η : [0] [1] quotiented by the relations 0 = and = = 56 / 64

98 A theory of monoids strict monoidal functor C = monoid in C Mon(C) = StrMonCat(, C) 57 / 64

99 A theory of games strict monoidal functor Games C =????? in C The corresponding theory is a polarized variant of relations. 57 / 64

100 The category Games The category Games is the category whose objects are integers [n] = {0, 1, 2,..., n 1} equipped with a polarization function λ : [n] {, } / 64

101 The category Games The category Games is the category whose objects are integers [n] = {0, 1, 2,..., n 1} equipped with a polarization function λ : [n] {, } 58 / 64

102 The category Games is the category whose objects are integers The category Games [n] = {0, 1, 2,..., n 1} equipped with a polarization function λ : [n] {, } morphisms are causal strategies. 58 / 64

103 The structure of strings 59 / 64

104 The structure of strings 59 / 64

105 Presentation of the category Games The category Games is presented by the polygraph with one 0-cell, 60 / 64

106 Presentation of the category Games The category Games is presented by the polygraph with one 0-cell, two 1-cells and, 60 / 64

107 Presentation of the category Games The category Games is presented by the polygraph with one 0-cell, two 1-cells and, six 2-cells 60 / 64

108 Presentation of the category Games The category Games is presented by the polygraph with one 0-cell, two 1-cells and, six 2-cells 3-cells (relations) ensuring that is a bicommutative bialgebra, is left dual to. 60 / 64

109 Technical byproducts From this presentation we can deduce that causal strategies compose, causal strategies are definable: we only have to show that generators are. 61 / 64

110 2-dimensional rewriting The proof is done by showing that every diagram is in relation with a diagram in canonical form and that these canonical forms are in bijection with the morphisms of this category. 62 / 64

111 2-dimensional rewriting The proof is done by showing that every diagram is in relation with a diagram in canonical form and that these canonical forms are in bijection with the morphisms of this category. This proof is very repetitive and requires to handle numerous cases: it should be automated. 62 / 64

112 2-dimensional rewriting The proof is done by showing that every diagram is in relation with a diagram in canonical form and that these canonical forms are in bijection with the morphisms of this category. This proof is very repetitive and requires to handle numerous cases: it should be automated. We have oriented the presentation of Mat(N) into a confluent rewriting system ( 5.5) = 62 / 64

113 2-dimensional rewriting The proof is done by showing that every diagram is in relation with a diagram in canonical form and that these canonical forms are in bijection with the morphisms of this category. This proof is very repetitive and requires to handle numerous cases: it should be automated. We have oriented the presentation of Mat(N) into a confluent rewriting system ( 5.5) = We have introduced an unification algorithm in order to compute critical pairs of such rewriting systems ( 5.4). 62 / 64

114 Contributions We defined an asynchronous non-alternating game semantics which takes the concurrency of proofs in account, which unifies preexisting semantics of linear logic ( 2.3), in which we extend the notion of HO innocence ( 2.4), in which we give an interactive reformulation of the correctness criterion ( 2.5). We gave a presentation of a category of games and strategies which reveals the algebraic structure of first-order causal dependencies ( 4.2), which lays the foundations for a 2-dimensional extension of rewriting theory ( 5). 63 / 64

115 Thanks! 64 / 64