Duality in Logic and Computation

Transcription

1 Duality in Logic and Computation Prakash Panangaden 1 1 School of Computer Science McGill University IEEE Symposium On Logic In Computer Science, June 2013 Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

2 Heros of duality Pictures Israel M. Gelfand Marshall H. Stone Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

3 Pictures Machine Learning Collaborators Chris Hundt Monica Dinculescu Doina Precup Joelle Pineau Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

4 Pictures Collaborators on today s talk This is joint work with many people: Nick Bezahanishvili, Filippo Bonchi, Marcello Bonsangue, Helle Hvid Hansen, Dexter Kozen, Clemens Kupke, Jan Rutten and Alexandra Silva. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

5 Pictures Other Stone-duality collaborators A related paper on Stone Duality for Markov Processes with Dexter Kozen, Kim G. Larsen and Radu Mardare will be presented tomorrow. Dexter Kozen Kim G. Larsen Radu Mardare Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

6 Pictures Collaborators on LMPs and duality Previous work was with Josée Desharnais, Vincent Danos, François Laviolette (Info. Comp 2006) and Philippe Chaput, Vincent Danos and Gordon Plotkin (ICALP 2009). Josée Desharnais François Laviolette Philippe Chaput Vincent Danos Gordon Plotkin Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

7 Introduction Examples of duality principles Linear programming. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

8 Introduction Examples of duality principles Linear programming. Electric and magnetic fields, Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

9 Introduction Examples of duality principles Linear programming. Electric and magnetic fields, now vastly generalized to geometric Langlands duality. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

10 Introduction Examples of duality principles Linear programming. Electric and magnetic fields, now vastly generalized to geometric Langlands duality. Controllability and observability in control theory, Kalman. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

11 Introduction Examples of duality principles Linear programming. Electric and magnetic fields, now vastly generalized to geometric Langlands duality. Controllability and observability in control theory, Kalman. Many examples from semantics and logic. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

12 Introduction What is Stone-type duality? Two types of structures: Shiv and Vish. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

13 Introduction What is Stone-type duality? Two types of structures: Shiv and Vish. Every Shiv has an associated Vish and vice versa. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

14 Introduction What is Stone-type duality? Two types of structures: Shiv and Vish. Every Shiv has an associated Vish and vice versa. V S, S V ; V and V are isomorphic. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

15 Introduction What is Stone-type duality? Two types of structures: Shiv and Vish. Every Shiv has an associated Vish and vice versa. V S, S V ; V and V are isomorphic. Two apparently different structures are actually two different descriptions of the same thing. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

16 Introduction What is Stone-type duality? Two types of structures: Shiv and Vish. Every Shiv has an associated Vish and vice versa. V S, S V ; V and V are isomorphic. Two apparently different structures are actually two different descriptions of the same thing. More importantly given a map: f : S 1 S 2 we get a map ˆf : V 2 V 1 and vice versa; Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

17 Introduction What is Stone-type duality? Two types of structures: Shiv and Vish. Every Shiv has an associated Vish and vice versa. V S, S V ; V and V are isomorphic. Two apparently different structures are actually two different descriptions of the same thing. More importantly given a map: f : S 1 S 2 we get a map ˆf : V 2 V 1 and vice versa; note the reversal in the direction of the arrows. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

18 Introduction What is Stone-type duality? Two types of structures: Shiv and Vish. Every Shiv has an associated Vish and vice versa. V S, S V ; V and V are isomorphic. Two apparently different structures are actually two different descriptions of the same thing. More importantly given a map: f : S 1 S 2 we get a map ˆf : V 2 V 1 and vice versa; note the reversal in the direction of the arrows. The two mathematical universes are mirror images of each other. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

19 Introduction What is Stone-type duality? Two types of structures: Shiv and Vish. Every Shiv has an associated Vish and vice versa. V S, S V ; V and V are isomorphic. Two apparently different structures are actually two different descriptions of the same thing. More importantly given a map: f : S 1 S 2 we get a map ˆf : V 2 V 1 and vice versa; note the reversal in the direction of the arrows. The two mathematical universes are mirror images of each other. Two completely different sets of theorems that one can use. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

20 Introduction Examples of Stone-type dualities Vector spaces and vector spaces. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

21 Introduction Examples of Stone-type dualities Vector spaces and vector spaces. Boolean algebras and Stone spaces. [Stone] Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

22 Introduction Examples of Stone-type dualities Vector spaces and vector spaces. Boolean algebras and Stone spaces. [Stone] Modal logics and boolean algebras with operators. [Jonsson, Tarski] Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

23 Introduction Examples of Stone-type dualities Vector spaces and vector spaces. Boolean algebras and Stone spaces. [Stone] Modal logics and boolean algebras with operators. [Jonsson, Tarski] State transformer semantics and weakest precondition semantics. [DeBakker,Plotkin,Smyth] Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

24 Introduction Examples of Stone-type dualities Vector spaces and vector spaces. Boolean algebras and Stone spaces. [Stone] Modal logics and boolean algebras with operators. [Jonsson, Tarski] State transformer semantics and weakest precondition semantics. [DeBakker,Plotkin,Smyth] Logics and Transition systems. [Bonsangue, Kurz,...] Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

25 Introduction Examples of Stone-type dualities Vector spaces and vector spaces. Boolean algebras and Stone spaces. [Stone] Modal logics and boolean algebras with operators. [Jonsson, Tarski] State transformer semantics and weakest precondition semantics. [DeBakker,Plotkin,Smyth] Logics and Transition systems. [Bonsangue, Kurz,...] Measures and random variables. [Kozen] Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

26 Introduction Examples of Stone-type dualities Vector spaces and vector spaces. Boolean algebras and Stone spaces. [Stone] Modal logics and boolean algebras with operators. [Jonsson, Tarski] State transformer semantics and weakest precondition semantics. [DeBakker,Plotkin,Smyth] Logics and Transition systems. [Bonsangue, Kurz,...] Measures and random variables. [Kozen] Commutative unital C*-algebras and compact Hausdorff spaces. [Gelfand, Stone] Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

27 Introduction Examples of Stone-type dualities Vector spaces and vector spaces. Boolean algebras and Stone spaces. [Stone] Modal logics and boolean algebras with operators. [Jonsson, Tarski] State transformer semantics and weakest precondition semantics. [DeBakker,Plotkin,Smyth] Logics and Transition systems. [Bonsangue, Kurz,...] Measures and random variables. [Kozen] Commutative unital C*-algebras and compact Hausdorff spaces. [Gelfand, Stone] Labelled Markov processes and C -algebras with operators. [Mislove, Ouaknine, Pavlovic, Worrell] Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

28 Maps matter! The need for Category Theory An essential aspect of mathematics: structure-preserving maps between objects. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

29 The need for Category Theory Maps matter! An essential aspect of mathematics: structure-preserving maps between objects. Interesting constructions on objects (usually) have corresponding constructions on the maps. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

30 The need for Category Theory Maps matter! An essential aspect of mathematics: structure-preserving maps between objects. Interesting constructions on objects (usually) have corresponding constructions on the maps. Compositions are preserved or reversed. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

31 The need for Category Theory Maps matter! An essential aspect of mathematics: structure-preserving maps between objects. Interesting constructions on objects (usually) have corresponding constructions on the maps. Compositions are preserved or reversed. This is functoriality. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

32 The need for Category Theory Maps matter! An essential aspect of mathematics: structure-preserving maps between objects. Interesting constructions on objects (usually) have corresponding constructions on the maps. Compositions are preserved or reversed. This is functoriality. From this one can often conclude invariance properties. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

33 The need for Category Theory Duality categorically F C D op G Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

34 The need for Category Theory Duality categorically Given A C f B C Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

35 The need for Category Theory Duality categorically We get A C f B C F(A) D F(B) D Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

36 The need for Category Theory Duality categorically and A C f F(f ) F(A) D B C F(B) D. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

37 The need for Category Theory Duality categorically Similarly, given C D g D D Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

38 The need for Category Theory Duality categorically We get G(C) C G(D) C C D g D D Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

39 The need for Category Theory Duality categorically and G(C) C G(g) G(D) C C D g D D. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

40 Isomorphisms The need for Category Theory We have isomorphisms A G(F(A)) and C F(G(C)). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

41 The need for Category Theory Duality categorically Stone-type Duality We have a (contravariant) adjunction between categories C and D, which is an equivalence of categories. Often obtained by looking at maps into an object living in both categories: a schizophrenic object. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

42 Vector space self-duality Duality for high school students I Finite-dimensional vector space V over, say R, Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

43 Vector space self-duality Duality for high school students I Finite-dimensional vector space V over, say R, Dual space V of linear maps from V to R. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

44 Vector space self-duality Duality for high school students I Finite-dimensional vector space V over, say R, Dual space V of linear maps from V to R. V has the same dimension as V and a (basis-dependent) isomorphism between V and V. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

45 Vector space self-duality Duality for high school students I Finite-dimensional vector space V over, say R, Dual space V of linear maps from V to R. V has the same dimension as V and a (basis-dependent) isomorphism between V and V. The double dual V is also isomorphic to V with a nice canonical isomorphism: v V λσ V.σ(v). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

46 Vector space self-duality Duality for high school students II U θ V U Given a linear maps θ between vector spaces U and V we get a map θ in the opposite direction between the dual spaces: θ V θ (σ V )(u U) = σ(θ(u)). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

47 Classical Stone duality Boolean algebras A Boolean algebra is a set A equipped with two constants, 0, 1, a unary operation ( ) and two binary operations, which obey the following axioms, p, q, r are arbitrary members of A: 0 = 1 1 = 0 p 0 = 0 p 1 = 1 p 1 = p p 0 = p p p = 0 p p = 1 p p = p p p = p Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

48 Classical Stone duality Boolean algebras II p = p (p q) = p q (p q) = p q p q = q p p q = q p p (q r) = (p q) r p (q r) = (p q) r p (q r) = (p q) (p r) p (q r) = (p q) (p r) The operation is called join, is called meet and ( ) is called complement. Maps are Boolean algebra homomorphisms. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

49 (Toy) Stone duality Classical Stone duality P FinSet FinBoolAlg op Here P is power-set and A takes the atoms of a Boolean algebra. A Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

50 Stone spaces Classical Stone duality A Stone space is a compact Hausdorff space with a base of clopen sets: zero-dimensional space. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

51 Classical Stone duality Stone spaces A Stone space is a compact Hausdorff space with a base of clopen sets: zero-dimensional space. Totally disconnected: the only connected sets are singletons. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

52 Classical Stone duality Stone spaces A Stone space is a compact Hausdorff space with a base of clopen sets: zero-dimensional space. Totally disconnected: the only connected sets are singletons. Many, but not all, Stone spaces are Polish. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

53 Classical Stone duality Classical Stone duality Stone space S, the clopens form a Boolean algebra: Cl(S). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

54 Classical Stone duality Classical Stone duality Stone space S, the clopens form a Boolean algebra: Cl(S). Boolean algebra B, Stone space U(B) is the space U of maximal filters (ultrafilters). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

55 Classical Stone duality Classical Stone duality Stone space S, the clopens form a Boolean algebra: Cl(S). Boolean algebra B, Stone space U(B) is the space U of maximal filters (ultrafilters). b B, a basic open is defined by U b := {u U b u}. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

56 Classical Stone duality Classical Stone duality Stone space S, the clopens form a Boolean algebra: Cl(S). Boolean algebra B, Stone space U(B) is the space U of maximal filters (ultrafilters). b B, a basic open is defined by U b := {u U b u}. Ultrafilters correspond precisely to maps from B to 2: Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

57 Classical Stone duality Classical Stone duality Stone space S, the clopens form a Boolean algebra: Cl(S). Boolean algebra B, Stone space U(B) is the space U of maximal filters (ultrafilters). b B, a basic open is defined by U b := {u U b u}. Ultrafilters correspond precisely to maps from B to 2: {b B f (b) = 1} is always an ultrafilter. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

58 Classical Stone duality Classical Stone duality Stone space S, the clopens form a Boolean algebra: Cl(S). Boolean algebra B, Stone space U(B) is the space U of maximal filters (ultrafilters). b B, a basic open is defined by U b := {u U b u}. Ultrafilters correspond precisely to maps from B to 2: {b B f (b) = 1} is always an ultrafilter. Continuous maps f : S 1 S 2 between Stone spaces give Boolean algebra homomorphisms f 1 : Cl(S 2 ) Cl(S 1 ). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

59 Classical Stone duality Classical Stone duality Stone space S, the clopens form a Boolean algebra: Cl(S). Boolean algebra B, Stone space U(B) is the space U of maximal filters (ultrafilters). b B, a basic open is defined by U b := {u U b u}. Ultrafilters correspond precisely to maps from B to 2: {b B f (b) = 1} is always an ultrafilter. Continuous maps f : S 1 S 2 between Stone spaces give Boolean algebra homomorphisms f 1 : Cl(S 2 ) Cl(S 1 ). Boolean algebra homomorphisms h : B 1 B 2 give rise to continuous functions ( ) f : U(B 2 ) U(B 1 ) between the Stone spaces. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

60 Classical Stone duality Classical Stone duality Stone space S, the clopens form a Boolean algebra: Cl(S). Boolean algebra B, Stone space U(B) is the space U of maximal filters (ultrafilters). b B, a basic open is defined by U b := {u U b u}. Ultrafilters correspond precisely to maps from B to 2: {b B f (b) = 1} is always an ultrafilter. Continuous maps f : S 1 S 2 between Stone spaces give Boolean algebra homomorphisms f 1 : Cl(S 2 ) Cl(S 1 ). Boolean algebra homomorphisms h : B 1 B 2 give rise to continuous functions ( ) f : U(B 2 ) U(B 1 ) between the Stone spaces. Everything that can, and should be, an isomorphism is an isomorphism. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

61 Duality in semantics State-transformer semantics Operational semantics: states, transitions. What are the next states? Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

62 Duality in semantics State-transformer semantics Operational semantics: states, transitions. What are the next states? Elegant and (almost) compositional version: Plotkin s structured operational semantics. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

63 Duality in semantics State-transformer semantics Operational semantics: states, transitions. What are the next states? Elegant and (almost) compositional version: Plotkin s structured operational semantics. Denotational semantics: compositional, equivalent to operational semantics. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

64 Duality in semantics Predicate transformers: Dijkstra Predicate transformers: if after the execution of a command a property P holds, what must have been true before? Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

65 Duality in semantics Predicate transformers: Dijkstra Predicate transformers: if after the execution of a command a property P holds, what must have been true before? Weakest precondition (wp). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

66 Duality in semantics Predicate transformers: Dijkstra Predicate transformers: if after the execution of a command a property P holds, what must have been true before? Weakest precondition (wp). Backward flow in wp semantics. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

67 Duality in semantics Predicate transformers: Dijkstra Predicate transformers: if after the execution of a command a property P holds, what must have been true before? Weakest precondition (wp). Backward flow in wp semantics. D and E domains, viewed as topological spaces, open sets: O D and O E. A predicate transformer is a strict, continuous and multiplicative map p : O E O D. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

68 Duality in semantics Predicate transformers: Dijkstra Predicate transformers: if after the execution of a command a property P holds, what must have been true before? Weakest precondition (wp). Backward flow in wp semantics. D and E domains, viewed as topological spaces, open sets: O D and O E. A predicate transformer is a strict, continuous and multiplicative map p : O E O D. Relate predicate-transformer semantics to state-transformer semantics: Jaco De Bakker (1978). Duality: The category of state transformers is equivalent to the (opposite of) the category of predicate transformers: Plotkin (1979). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

69 Duality in semantics Duality for probabilistic programs: Kozen Probabilistic programs and expectation transformers: Kozen (1981) Logic Probability States s Distributions µ Formulas P Random variables f Satisfaction s = P Integration f dµ Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

70 Duality in semantics Domain theory in logical form Abramsky s program based on Stone duality. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

71 Duality in semantics Domain theory in logical form Abramsky s program based on Stone duality. Metalanguage for types and terms (programs). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

72 Duality in semantics Domain theory in logical form Abramsky s program based on Stone duality. Metalanguage for types and terms (programs). Usual denotational semantics: types are domains, terms are elements. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

73 Duality in semantics Domain theory in logical form Abramsky s program based on Stone duality. Metalanguage for types and terms (programs). Usual denotational semantics: types are domains, terms are elements. DTLF: types are propositional theories, finite elements are propositions. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

74 Duality in semantics Domain theory in logical form Abramsky s program based on Stone duality. Metalanguage for types and terms (programs). Usual denotational semantics: types are domains, terms are elements. DTLF: types are propositional theories, finite elements are propositions. Terms are described by axiomatizing satisfaction. A modal logic of programs. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

75 Duality in semantics Domain theory in logical form Abramsky s program based on Stone duality. Metalanguage for types and terms (programs). Usual denotational semantics: types are domains, terms are elements. DTLF: types are propositional theories, finite elements are propositions. Terms are described by axiomatizing satisfaction. A modal logic of programs. The two interpretations are Stone duals. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

76 Duality in semantics Domain theory in logical form Abramsky s program based on Stone duality. Metalanguage for types and terms (programs). Usual denotational semantics: types are domains, terms are elements. DTLF: types are propositional theories, finite elements are propositions. Terms are described by axiomatizing satisfaction. A modal logic of programs. The two interpretations are Stone duals. Ties together semantics, logic and verification. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

77 Chu spaces: Pratt Duality in semantics Vaughan Pratt Chu spaces: general construction for -autonomous categories. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

78 Chu spaces: Pratt Duality in semantics Vaughan Pratt Chu spaces: general construction for -autonomous categories. Generalized matrices. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

79 Chu spaces: Pratt Duality in semantics Vaughan Pratt Chu spaces: general construction for -autonomous categories. Generalized matrices. Pratt s observation: many interesting categories embed in Chu categories. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

80 Chu spaces: Pratt Duality in semantics Vaughan Pratt Chu spaces: general construction for -autonomous categories. Generalized matrices. Pratt s observation: many interesting categories embed in Chu categories. Stone duality is transposition of matrices. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

81 Duality in semantics Many other contributors Bart Jacobs,... Achim Jung and Drew Moshier Mai Gehrke, Jean-Eric Pin,... Bezhanishvilis Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

82 Brzozowski s strange algorithm Brzozowski s Algorithm 1964 Start with DFA. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

83 Brzozowski s strange algorithm Brzozowski s Algorithm 1964 Start with DFA. Reverse transitions, interchange initial and final states. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

84 Brzozowski s strange algorithm Brzozowski s Algorithm 1964 Start with DFA. Reverse transitions, interchange initial and final states. Determinize the result. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

85 Brzozowski s strange algorithm Brzozowski s Algorithm 1964 Start with DFA. Reverse transitions, interchange initial and final states. Determinize the result. Take the reachable states. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

86 Brzozowski s strange algorithm Brzozowski s Algorithm 1964 Start with DFA. Reverse transitions, interchange initial and final states. Determinize the result. Take the reachable states. Repeat. This gives the minimal DFA recognizing the same language! Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

87 Brzozowski s strange algorithm Brzozowski s Algorithm 1964 Start with DFA. Reverse transitions, interchange initial and final states. Determinize the result. Take the reachable states. Repeat. This gives the minimal DFA recognizing the same language! The intermediate step can blow up the size of the automaton exponentially before minimizing it. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

88 Brzozowski s strange algorithm Brzozowski s Algorithm 1964 Start with DFA. Reverse transitions, interchange initial and final states. Determinize the result. Take the reachable states. Repeat. This gives the minimal DFA recognizing the same language! The intermediate step can blow up the size of the automaton exponentially before minimizing it. But experimental results seem to indicate that it often works well in practice. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

89 Thinking logically Deterministic Automata M = (S, A, O, δ, γ): a deterministic finite (Moore) automaton. S is the set of states, A an input alphabet (actions), O is a set of observations. δ : S A S is the state transition function. γ : S 2 O or γ : S O 2 is a labeling function. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

90 Thinking logically Deterministic Automata M = (S, A, O, δ, γ): a deterministic finite (Moore) automaton. S is the set of states, A an input alphabet (actions), O is a set of observations. δ : S A S is the state transition function. γ : S 2 O or γ : S O 2 is a labeling function. If O = {accept} we have ordinary deterministic finite automata, Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

91 Thinking logically Deterministic Automata M = (S, A, O, δ, γ): a deterministic finite (Moore) automaton. S is the set of states, A an input alphabet (actions), O is a set of observations. δ : S A S is the state transition function. γ : S 2 O or γ : S O 2 is a labeling function. If O = {accept} we have ordinary deterministic finite automata, except that we do not have a start state, Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

92 Thinking logically Deterministic Automata M = (S, A, O, δ, γ): a deterministic finite (Moore) automaton. S is the set of states, A an input alphabet (actions), O is a set of observations. δ : S A S is the state transition function. γ : S 2 O or γ : S O 2 is a labeling function. If O = {accept} we have ordinary deterministic finite automata, except that we do not have a start state, which means that reachability makes no sense. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

93 Thinking logically Deterministic Automata M = (S, A, O, δ, γ): a deterministic finite (Moore) automaton. S is the set of states, A an input alphabet (actions), O is a set of observations. δ : S A S is the state transition function. γ : S 2 O or γ : S O 2 is a labeling function. If O = {accept} we have ordinary deterministic finite automata, except that we do not have a start state, which means that reachability makes no sense. We will worry about that in a minute. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

94 Thinking logically A Simple Modal Logic View O as propositions, define a simple modal logic. A formula ϕ is: ϕ ::== ω O (a)ϕ where a A. We say s = ω, if ω γ(s) (or γ(s, ω) = T). We say s = (a)ϕ if δ(s, a) = ϕ. Now we define [ϕ] M = {s S s = ϕ}. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

95 Thinking logically An Equivalence Relation on Formulas We write sa as shorthand for δ(s, a). Define M between formulas as ϕ M ψ if [ϕ] M = [ψ ] M. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

96 Thinking logically An Equivalence Relation on Formulas We write sa as shorthand for δ(s, a). Define M between formulas as ϕ M ψ if [ϕ] M = [ψ ] M. Equivalence class for ϕ same as of states [ϕ] M that satisfy ϕ. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

97 Thinking logically A Dual Automaton Given a finite automaton M = (S, A, O, δ, γ). Let T be the set of M -equivalence classes of formulas on M. We define M = (S, A, O, δ, γ ) as follows: S = T = {[ϕ] M } O = S δ ([ϕ] M, a) = [(a)ϕ] M γ ([ϕ] M ) = [ϕ] M or γ ([ϕ] A, s) = (s = ϕ). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

98 The intuition Thinking logically Interchange states and observations. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

99 Minimality Properties Thinking logically In general, the double dual is the minimal machine with the same behaviour! Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

100 Thinking logically Minimality Properties In general, the double dual is the minimal machine with the same behaviour! For deterministic machines bisimulation is the same as trace equivalence. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

101 Thinking logically Minimality Properties In general, the double dual is the minimal machine with the same behaviour! For deterministic machines bisimulation is the same as trace equivalence. This gives an intuition for why Brzozowski s algorithm works, Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

102 Thinking logically Minimality Properties In general, the double dual is the minimal machine with the same behaviour! For deterministic machines bisimulation is the same as trace equivalence. This gives an intuition for why Brzozowski s algorithm works, but it does not really address the role of reachability properly. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

103 Probabilistic systems Thinking logically In joint work with Chris Hundt, Joelle Pineau and Doina Precup (AAAI 2006): duals for various kinds of probabilistic transition systems like probabilistic Moore automata and partially observable Markov decision processes. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

104 Thinking logically Probabilistic systems In joint work with Chris Hundt, Joelle Pineau and Doina Precup (AAAI 2006): duals for various kinds of probabilistic transition systems like probabilistic Moore automata and partially observable Markov decision processes. Dual automaton from tests: probabilistic analogues of modal formulas. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

105 Thinking logically Probabilistic systems In joint work with Chris Hundt, Joelle Pineau and Doina Precup (AAAI 2006): duals for various kinds of probabilistic transition systems like probabilistic Moore automata and partially observable Markov decision processes. Dual automaton from tests: probabilistic analogues of modal formulas. Main point: not minimization, but can learn systems from data even when the state is not directly observable Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

106 Thinking logically Probabilistic systems In joint work with Chris Hundt, Joelle Pineau and Doina Precup (AAAI 2006): duals for various kinds of probabilistic transition systems like probabilistic Moore automata and partially observable Markov decision processes. Dual automaton from tests: probabilistic analogues of modal formulas. Main point: not minimization, but can learn systems from data even when the state is not directly observable because the double-dual serves as a substitute for the original machine. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

107 Application to learning Thinking logically One can plan when one has the model: value iteration etc., Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

108 Thinking logically Application to learning One can plan when one has the model: value iteration etc., but quite often one does not have the model. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

109 Thinking logically Application to learning One can plan when one has the model: value iteration etc., but quite often one does not have the model. In the absence of a model, one is forced to learn from data. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

110 Thinking logically Application to learning One can plan when one has the model: value iteration etc., but quite often one does not have the model. In the absence of a model, one is forced to learn from data. Learning is hopeless when one has no idea what the state space is. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

111 Thinking logically Application to learning One can plan when one has the model: value iteration etc., but quite often one does not have the model. In the absence of a model, one is forced to learn from data. Learning is hopeless when one has no idea what the state space is. There should be no such thing as absolute state! Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

112 Thinking logically Application to learning One can plan when one has the model: value iteration etc., but quite often one does not have the model. In the absence of a model, one is forced to learn from data. Learning is hopeless when one has no idea what the state space is. There should be no such thing as absolute state! State is just a summary of past observations that can be used to make predictions. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

113 Thinking logically Application to learning One can plan when one has the model: value iteration etc., but quite often one does not have the model. In the absence of a model, one is forced to learn from data. Learning is hopeless when one has no idea what the state space is. There should be no such thing as absolute state! State is just a summary of past observations that can be used to make predictions. Double dual: state can be regarded as the summary of the outcomes of experiments. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

114 The categorical picture of automata duality What is the right categorical description? Is this is any kind of familiar Stone-type duality? Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

115 The categorical picture of automata duality What is the right categorical description? Is this is any kind of familiar Stone-type duality? We know that machines are co-algebras and logics are algebras but Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

116 The categorical picture of automata duality What is the right categorical description? Is this is any kind of familiar Stone-type duality? We know that machines are co-algebras and logics are algebras but why is the dual another automaton? Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

117 The categorical picture of automata duality Automata as Coalgebras Our automata are coalgebras of the following functor: F(S) = S A 2 O, F(f : S S ) = λ α : A S, O O. f α, O. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

118 The categorical picture of automata duality Automata as Coalgebras Our automata are coalgebras of the following functor: F(S) = S A 2 O, F(f : S S ) = λ α : A S, O O. f α, O. The category of these coalgebras is called PODFA. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

119 The categorical picture of automata duality Automata as Coalgebras Our automata are coalgebras of the following functor: F(S) = S A 2 O, F(f : S S ) = λ α : A S, O O. f α, O. The category of these coalgebras is called PODFA. In ordinary language they are quintuples (S, A, O, δ : S A S, γ : S O) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

120 The categorical picture of automata duality Automata as Coalgebras Our automata are coalgebras of the following functor: F(S) = S A 2 O, F(f : S S ) = λ α : A S, O O. f α, O. The category of these coalgebras is called PODFA. In ordinary language they are quintuples (S, A, O, δ : S A S, γ : S O) with S: states, A: actions, O: observations, δ is a transition function and γ is an observation function. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

121 The categorical picture of automata duality Automata as Coalgebras Our automata are coalgebras of the following functor: F(S) = S A 2 O, F(f : S S ) = λ α : A S, O O. f α, O. The category of these coalgebras is called PODFA. In ordinary language they are quintuples (S, A, O, δ : S A S, γ : S O) with S: states, A: actions, O: observations, δ is a transition function and γ is an observation function. They are well known as Moore machines. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

122 The categorical picture of automata duality Homomorphisms A homomorphism for these coalgebras is a function f : S S such that the following diagram commutes: S f S δ, γ δ, γ S A 2 O S A 2 O f A id where f A (α) = f α. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

123 The categorical picture of automata duality This translates to the following conditions: s S, ω O, ω γ(s) ω γ (f (s)) (1) and s S, a A, f (δ(s, a)) = δ (f (s), a). (2) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

124 The categorical picture of automata duality The Dual Category The category of finite boolean algebras with operators (FBAO) has as objects finite boolean algebras B with Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

125 The categorical picture of automata duality The Dual Category The category of finite boolean algebras with operators (FBAO) has as objects finite boolean algebras B with the usual operations, and constants T and and, in addition, Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

126 The categorical picture of automata duality The Dual Category The category of finite boolean algebras with operators (FBAO) has as objects finite boolean algebras B with the usual operations, and constants T and and, in addition, together with unary operators (a) and constants ω. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

127 The categorical picture of automata duality The Dual Category The category of finite boolean algebras with operators (FBAO) has as objects finite boolean algebras B with the usual operations, and constants T and and, in addition, together with unary operators (a) and constants ω. We denote an object by B = (B, {(a) a A}, {ω ω O}, T,, ). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

128 The categorical picture of automata duality Equations The following three equations hold: (a)(b 1 b 2 ) = (a)b 1 (a)b 2, (a)t = T, (a) b = (a)b. (3a) (3b) (3c) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

129 Morphisms The categorical picture of automata duality The morphisms are the usual boolean homomorphisms preserving, in addition, the constants and the unary operators. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

130 The categorical picture of automata duality Duality Theorem There is a dual equivalence of categories PODFA op = FBAO. One functor P is just the contravariant power set functor and the other one H maps a boolean algebra to its set of atoms. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

131 Minimization? The categorical picture of automata duality Obviously, if we have an equivalence of categories we get the same machine back when we go back and forth. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

132 The categorical picture of automata duality Minimization? Obviously, if we have an equivalence of categories we get the same machine back when we go back and forth. So how do we explain the minimization? Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

133 The categorical picture of automata duality Definable Subsets Define a logic L by φ ::== T φ 1 φ 2 φ (a)φ ω and define the definable subsets D(S) of a machine M = (S, δ, γ) as sets of the form φ. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

134 The categorical picture of automata duality D(S) is a subobject of P(M) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

135 The categorical picture of automata duality D(S) is a subobject of P(M) in fact it is the smallest possible subalgebra and Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

136 The categorical picture of automata duality D(S) is a subobject of P(M) in fact it is the smallest possible subalgebra and any other subalgebra must contain D(S). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

137 In Pictures The categorical picture of automata duality M P(M) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

138 The categorical picture of automata duality M P(M) 1 D(S) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

139 The categorical picture of automata duality M 2 H(D(S)) P(M) 1 D(S) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

140 The categorical picture of automata duality M 2 3 H(D(S)) P(M) 1 D(S) S 4 P(S ) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

141 The categorical picture of automata duality M 2 3 H(D(S)) P(M) 1 D(S) 5 S 4 P(S ) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

142 The categorical picture of automata duality M 2 3 H(D(S)) P(M) 1 D(S) 5 S 4 6 P(S ) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

143 The categorical picture of automata duality The Secret of Minimization M P(M) 3 2 H(D(S)) 1 D(S) 5 S P(S ) Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

144 A Simpler Logic The categorical picture of automata duality Why did the minimization work with just the logic φ ::== ω (a)φ? Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

145 A Simpler Logic The categorical picture of automata duality Why did the minimization work with just the logic φ ::== ω (a)φ? With this logic the definable subsets E(S) do not form a boolean algebra, Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

146 The categorical picture of automata duality A Simpler Logic Why did the minimization work with just the logic φ ::== ω (a)φ? With this logic the definable subsets E(S) do not form a boolean algebra, it is just a set with operations Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

147 The categorical picture of automata duality A Simpler Logic Why did the minimization work with just the logic φ ::== ω (a)φ? With this logic the definable subsets E(S) do not form a boolean algebra, it is just a set with operations in other words, it can be viewed as an automaton! Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

148 The categorical picture of automata duality Why the simpler logic works For deterministic automata we can flatten formulas like (a)(ω 1 (b)ω 2 ) to (a)ω 1 (a)(b)ω 2. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

149 The categorical picture of automata duality Why the simpler logic works For deterministic automata we can flatten formulas like (a)(ω 1 (b)ω 2 ) to (a)ω 1 (a)(b)ω 2. Thus for deterministic automata the boolean algebra generated by E(S) is just the same as D(S) so the minimization picture works with boolean algebra generated by E(S). Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

150 The categorical picture of automata duality Why the simpler logic works For deterministic automata we can flatten formulas like (a)(ω 1 (b)ω 2 ) to (a)ω 1 (a)(b)ω 2. Thus for deterministic automata the boolean algebra generated by E(S) is just the same as D(S) so the minimization picture works with boolean algebra generated by E(S). For nondeterministic automata the story is different. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

151 Duality of reachability and observability Minimality = fewest states? The minimal machines defined above are really the most quotiented versions of a system. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

152 Duality of reachability and observability Minimality = fewest states? The minimal machines defined above are really the most quotiented versions of a system. To really get the system with the fewest states one needs to deal with reachability. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80

153 Duality of reachability and observability Minimality = fewest states? The minimal machines defined above are really the most quotiented versions of a system. To really get the system with the fewest states one needs to deal with reachability. The following discussion is a rapid version of what Jan Rutten discussed in his beautiful MFPS talk on Monday. Panangaden (McGill University) Duality in Logic and Computation New Orleans, June / 80