Date post: | 18-Dec-2014 |
Category: |
Education |
Upload: | valeria-de-paiva |
View: | 768 times |
Download: | 0 times |
Dialectica Categories...and Lax Topological Spaces?
Valeria de Paiva
Rearden CommerceUniversity of Birmingham
March, 2012
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 1 / 38
LLI nasslli2012.com
June 18–22, 2012
Johan van Benthem Logical Dynamics of Information and Interaction
Chris Po�s Stanford UniversityExtracting Social meaning and Sentiment
Craige Roberts Ohio State UniversityQuestions in Discourse
Mark Steedman University of EdinburghCombinatory Categorial Grammar: �eory and Practice
Noah Goodman Stanford UniversityStochastic Lambda Calculus
and its Applications in Semantics and Cognitive Science
University of Amsterdam / Stanford University
�e Fi�h North American Summer School of Logic, Language, and Information
Jonathan Ginzburg - University of ParisRobin Cooper - Göteborg UniversityType theory with records for natural language semantics
Jeroen Groenendijk - University of AmsterdamFloris Roelofsen - University of AmsterdamInquisitive semantics
Shalom Lappin - King’s College LondonAlternative Paradigms for Computational Semantics
Tandy Warnow - University of TexasEstimating phylogenetic trees in linguistics and biology
Hans Kamp - University of TexasMark Sainsbury - University of TexasVagueness and context
Steve Wechsler - University of Texas Eric McCready - Osaka UniversityWorkshop on Meaning as Use: Indexicality and Expressives (Speakers: Eric McCready, Steve Wechsler, Hans Kamp, Chris Po�s, Pranav Anand, and Sarah Murray)
Catherine Legg - University of WaikatoPossible Worlds: A Course in Metaphysics(for Computer Scientists and Linguists)
Adam Lopez - Johns Hopkins UniversityStatistical Machine Translation
Eric Pacuit - Stanford UniversitySocial Choice �eory for Logicians
Valeria de Paiva - Rearden CommerceUlrik Buchholtz - Stanford UniversityIntroduction to Category �eory
Adam Pease - Rearden CommerceOntology Development and Application with SUMO
Ede Zimmermann - University of FrankfurtIntensionality
�omas Icard - Stanford UniversitySurface Reasoning
Nina Gierasimczuk - University of GroningenBelief Revision Meets Formal Learning �eory
at the University of Texas at Austin
Registration: $175 (academic rate) / $400 (professional rate)Student scholarships available, see website for application instructions
Accommodation provided for $70 / night (single) or $35 / night (double)
June 16–17: Bootcamp Session June 23–24: Texas Linguistic Society Conference Special sessions on American Sign Language, Semantics, and Computational Linguistics Details regarding Call for Papers at nasslli2012.com June 23: Turing Centennial Symposium
Additional Courses: Special Events:
Poster by Derya Kadipasaoglu (dkadipas.weebly.com) and Christopher BrownNASSLLI is sponsored by the NSF (BCS 1019206), the UT College of Liberal Arts, and the UT Departments of Linguistics, Philosophy, and Psychology.
Twenty interdisciplinary graduate-level courses, 90
minutes each, every day for �ve days, on the UT campus,
taught by leading international scholars, and crossing the interfaces of philosophy,
linguistics, computer science, psychology, statistics, and
logic.
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 2 / 38
Goals
Describe two kinds of dialectica categoriesShow they are models of ILL and CLL, respectivelyCompare and contrast dialectica categories with Chu spacesDiscuss how we might be able to use these ideas for topological spaces
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 3 / 38
Functional Interpretations and Gödel’s Dialectica
Starting with Gödel’s Dialectica interpretation(1958) a series of“translation functions" between theoriesAvigad and Feferman on the Handbook of Proof Theory:
This approach usually follows Gödel’s original example: first,one reduces a classical theory C to a variant I based onintuitionistic logic; then one reduces the theory I to aquantifier-free functional theory F.
Examples of functional interpretations:Kleene’s realizabilityKreisel’s modified realizabilityKreisel’s No-CounterExample interpretationDialectica interpretationDiller-Nahm interpretation, etc...
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 4 / 38
Gödel’s Dialectica Interpretation
Gödel’s Aim: liberalized version of Hilbert’s programme – to justifyclassical systems in terms of notions as intuitively clear as possible.
Gödel’s result: an interpretation of intuitionistic arithmetic HA in aquantifier-free theory of functionals of finite type Tidea: translate every formula A of HA to AD = ∃u∀x.AD where AD isquantifier-free.use: If HA proves A then T proves AD(t, y) where y is a string ofvariables for functionals of finite type, t a suitable sequence of terms notcontaining y.
Method of ‘functional interpretation’ extended and adapted to several othertheories systems, cf. Feferman.
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 5 / 38
Gödel’s Dialectica Interpretation
For each formula A of HA we associate a formula of the formAD = ∃u∀xAD(u, x) (where AD is a quantifier-free formula of Gödel’ssystem T) inductively as follows: when Aat is an atomic formula, then itsinterpretation is itself.Assume we have already defined AD = ∃u∀x.AD(u, x) andBD = ∃v∀y.BD(v, y).We then define:
(A ∧B)D = ∃u, v∀x, y.(AD ∧BD)(A→ B)D = ∃f : U → V, F : U ×X → Y, ∀u, y.( AD(u, F (u, y))→ BD(fu; y))(∀zA)D(z) = ∃f : Z → U∀z, x.AD(z, f(z), x)(∃zA)D(z) = ∃z, u∀x.AD(z, u, x)
The intuition here is that if u realizes ∃u∀x.AD(u, x) then f(u) realizes∃v∀y.BD(v, y) and at the same time, if y is a counterexample to∃v∀y.BD(v, y), then F (u, y) is a counterexample to ∀x.AD(u, x).
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 6 / 38
Gödel’s Dialectica Interpretation: logical implication
Most interesting aspect: If AD = ∃u∀x.AD and BD = ∃v∀y.BD
(A⇒ B)D = ∃V (u)∃X(u, y)∀u∀y[AD(u,X(u, y))⇒ BD(V (u), y)]
This can be explained (Spector’62) by the following equivalences:
(A⇒ B)D = [∃u∀xAD ⇒ ∃v∀yBD]D (1)= [∀u(∀xAD ⇒ ∃v∀yBD)D] (2)= [∀u∃v(∀xAD ⇒ ∀yBD)]D (3)= [∀u∃v∀y(∀xAD ⇒ BD)]D (4)= [∀u∃v∀y∃x(AD ⇒ BD)]D (5)= ∃V ∃X∀u, y(AD(u,X(u, y)⇒ BD(V (u), y)] (6)
All steps classically valid. But:Step (2) (moving ∃v out) is IP (not intuitionistic),Step (4) uses gener. of Markov’s principle (not intuitionistic)for (6) need to skolemize twice, using AC (not intuitionistic).
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 7 / 38
Basics of Categorical Semantics
Model theory using categories, instead of sets or posets.Two main uses “categorical semantics”:categorical proof theory, and categorical model theoryCategorical proof theory models derivations (proofs) not simplywhether theorems are true or not
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 8 / 38
Categorical Semantics: what?
Based on Curry-Howard Isomorphism:
Natural deduction (ND) proofs ⇔ λ-termsNormalization of ND proof ⇔ reduction in λ-calculus.
Originally: constructive prop. logic ⇔ simply typed λ-calculus.Applies to other (constructive) logics and λ-calculi, too.Cat Proof Theory:λ-calculus types ⇔ objects in (appropriate) categoryλ-calculus terms ⇔1−1 morphisms in categoryCat. structure models logical connectives (type operators).Isomorphism: transfers results/tools from one side to the otherλ-calculus basis of functional programming: logical view ofprogramming
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 9 / 38
Where is my category theory?
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 10 / 38
Categorical Logic?
Well known that we can do categorical models as well as settheoretical ones.Why would you?Because sometimes this is the only way to obtain models.E.g. Scott’sλ-calculus
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 11 / 38
Categorical Logic?
Well known that we can do categorical models as well as settheoretical ones.
Why would you?Because sometimes this is the only way to obtain models.E.g. Scott’sλ-calculus
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 11 / 38
Categorical Logic?
Well known that we can do categorical models as well as settheoretical ones.Why would you?
Because sometimes this is the only way to obtain models.E.g. Scott’sλ-calculus
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 11 / 38
Categorical Logic?
Well known that we can do categorical models as well as settheoretical ones.Why would you?Because sometimes this is the only way to obtain models.
E.g. Scott’sλ-calculus
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 11 / 38
Categorical Logic?
Well known that we can do categorical models as well as settheoretical ones.Why would you?Because sometimes this is the only way to obtain models.E.g. Scott’sλ-calculus
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 11 / 38
Categorical Semantics: some gains..
Functional Programming: optimizations that do not compromisesemantic foundations (e.g. referential transparency)Logic: new applications of old theorems (e.g. normalization)Category Theory: new concepts not from mathsPhilosophy: new criteria for identities of proofs
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 12 / 38
Basics of Linear Logic
A new(ish..) logic taking resource sensitivity into accountDeveloped mainly from proof theoretic perspective:— must keep track of where premises / assumptions are usedBut ability to ignore resource sensitivity whenever you want:(a) Modal ! allowing duplication and erasing of resources(b) A⇒ B = !A−◦B
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 13 / 38
Where is my category theory?
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 14 / 38
Linear categories
Model multiplicative-additive fragment is easymultiplicative fragment: need a a symmetric monoidal closedcategory (SMCC)additive fragment: need (weak) categorical products andcoproductslinear negation: need an involution
modalities (exponentials): the problem...several solutions, depending on linear λ-calculus chosen...
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 15 / 38
Symmetric Monoidal Closed Categories (SMCC)
A category L is a symmetric monoidal closed category ifit has a tensor product A⊗B for all object A,B of C.for each object B of C, the tensor-product functor _⊗B has aright-adjoint, written B−◦_(ie have isomorphism between (A⊗B)→ C and A→ (B−◦C))
Recap:Categorical product A×B Tensor product A⊗B
1 Projections onto A and B Not necessarily2 For any object C with maps Not necessarily
to A and B, there is aunique map from C to A×B
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 16 / 38
Modelling the Modality !
Objects A in a CCC have duplication A→ A×A and erasing A→ 1(satisfying equations of a comonoid).Objects A in a SMCC have no duplication A→ A⊗A nor erasingA→ I, in general. Neither have they projections A⊗B → A,B.But objects of the form !A should have duplication and erasing.Moreover they must satisfy S4-Box introduction and elimination rules.S4-axioms give you proofs !A→ A and !A→!!A, uniformly for any Aobject A in C. Hence want a functor (unary operator) on category,! : C → C such that there are such natural transformations. Such astructure already in category theory, a comonad.
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 17 / 38
Modelling the Modality !
An object !A must have four maps:
ε : !A→ A, δ : !A→!!A, er : !A→ I, dupl : !A→!A⊗!A
To make identity of proofs sensible, must have lots of equations. Theones relating to duplicating and erasing say !A is a comonoid withrespect to tensor; the ones relating to S4-axioms say ! is a comonad.First idea (Seely’87): Logical equivalence !(A&B) ≡ !A⊗!BIntroduce a comonad (!, δ, ε) defining ! such that
!(A×B) ∼= !A⊗!B
Say this comonad takes additive comonoid structure(A, A→ A×A, A→ 1) (if it exists) to desired comonoid(!A, dupl : !A→!A⊗!A, er : !A→ I).
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 18 / 38
Categorical Semantics of LL: problem
Modelling uses isomorphism !(A&B) ∼=!A⊗!BBut can’t model multiplicatives & modalities without additives.So, better idea:If we don’t have additive conjunction, ask for morphisms
mA,B : !(A⊗B)→ !A⊗!B
mI : !I → I
i.e. monoidal comonad to deal with contexts
Must also ask for (coalgebra) conditions relating comonad structure tocomonoid structure.Hence:
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 19 / 38
Categorical Semantics of LL: solution
A linear category comprises— an SMCC C, (with products and coproducts),— a symmetric monoidal comonad (!, ε, δ,mI ,mA,B)(a) For every free co-algebra (!A, δ) there are nat. transfns.erA : !A→ I and duplA : !A→!A⊗!A, forming a commutative comonoid,which are coalgebra maps.(b) Every map of free coalgebras is also a map of comonoids.Bierman 1995 (based on Benton et al’93)
long definition, lots of diagrams to check..Prove A⇒ B ∼=!A−◦B by constructing CCC out of SMCC plus comonad!. Extended Curry-Howard works!!
Classical linear category = linear category with involution,()⊥ : Cop → C, to model negation, same as ∗-autonomous category, with(weak) products and coproducts plus a symm monoidal comonad satisfying(a) and (b) above.
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 20 / 38
Alternative Categorical Semantics
(Benton 1995) A model of LNL (Linear/NonLinear Logic) consists of aSMC category L together with a cartesian closed category C and amonoidal adjunction between these categories.
Modelling two logics, intuitionistic and linear at the same time.Logics on (more) equal footing..
First definition builds the cartesian closed category out of the monoidalcomonad, this def asks for the adjunction
Easier to remember, easier to say, same diagrams to check..
Extended Curry-Howard works (Barber’s thesis)(reduction calculus (+explicit subst) needed for implementation, Wollic’98)
classical version: *-autonomous cat, plus monoidal adjunction, plusproducts
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 21 / 38
Categorical Semantics: Back to Dialectica
Have a definition of what a categorical model of ILL ought to be, canwe find some concrete ones?Dialectica categories (Boulder’87) one such model..Assume C is cartesian closed categoryDC objects are relations between U and X in C,monics AαU ×X written as (UαX).DC maps are pairs of maps of Cf : U → V , F : U × Y → Xmaking a certain pullback condition hold.Condition in Sets: uαF (u, y)⇒ f(u)βy.
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 22 / 38
Reformulating DC
C is Sets: objects are maps A = U ×Xα2, and B = V × Y β2.morphisms are pairs of maps f : U → V , F : U × Y → X such thatα(u, F (u, y)) ≤ β(f(u), y):
U × Y〈π1, F 〉- U ×X
V × Y
f × idY
?
β- 2
α
?
Usually write maps asU α X
⇓ ∀u ∈ U,∀y ∈ Y α(u, F (u, y)) ≤ β(f(u), y)
V
f
?β Y
F
6
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 23 / 38
Relationship to Dialectica
Objects of DC represent essentially AD.Maps some kind of normalisation class of proofs.abstractly: maps are realisation of the formula AD → BD.Expected: cat DC is ccc, as Gödel int. as constructive as possible.Surprise: DC is a model of ILL, not IL.But (very neat result!)Diller-Nahm variant gives CCC.Note: reformulation makes basic predicates AD decidable, as in Gödel’s work, but first version is more general.
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 24 / 38
Structure of DC
DC is a SMCC with categorical products:
A�B = (U × V α� βX × Y )
A−◦B = (V U ×XU×Y α−◦βU × Y )
A&B = (U × V αβX + Y )
only weak coproducts:
A+B = (U + V α+ βXU × Y V )
Morever!A = (Uα∗X∗)
is cofree comonoid on A.
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 25 / 38
Conjunction: neat detail
A 6` A�A as need a DC map (∆, δ) such that
U α X
⇓
U × U
∆
?α� αX ×X.
δ
6
∀u ∈ U,∀x, x′ ∈ X ×X α(u, δ(u, x, x′))⇒ (α� α)(∆(u), (x, x′)).But (α� α)(∆(u), (x, x′)) = α(u, x) ∧ α(u, x′)hence unless α decidable can’t do anything...If α decidable define:δ(u, x, x′) = x′ if α(u, x)δ(u, x, x′) = x otherwise. Same logical argument..
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 26 / 38
Dialectica categories DC: summary
DC pluses:structure needed to model ILL,including non-trivial modality !related to Diller-Nahm variantconjunction explainedindependent confirmation of LL’s worth;Dialectica is interesting (and mysterious).DC minuses:accounting of recursion?mathematics involved;terse style – AMS vol.92
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 27 / 38
Dialectica model GC
Girard’s suggestion ’87C cartesian closed category, say Sets.GC objects are relations between U and X, in C as before.GC maps are pairs of maps of C, f : U → V , F : Y → Xsuch that uαF (y)⇒ f(u)βy.As before, can ‘reformulate’ and say A is map U ×X → 2
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 28 / 38
Structure of GC
GC is ∗-autonomous with products and coproducts. Involution is given byimplication into ⊥, unit for extra monoidal structure .....................................................
.......................................... .
A−◦B = (V U ×XY α−◦βU × Y )A⊗B = (U × V α⊗ βXV × Y U )A
...............................................................................................B = (V X × UY α
............................................................................................... βX × Y )
A&B = (U × V α.βX + Y )A⊕B = (U + V α.βX × Y )
Moreover, GC has Exponentials
!A = (U !α(X∗)U )
?A = ((U∗)X?αX)
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 29 / 38
Dialectica-like categories GC
GC pluses:the best model of LL ?..no collapsing of unitiesmaths easier for modality-free fragmentmodalities adaptableshould think of Sets!GC minuses:MIX ruleC must be ccc
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 30 / 38
Chu construction ChuKC
Suppose C is cartesian closed category, (but could be smcc with pullbacks)and K any object of C.ChuKC construction in Barr’s book, 1972.ChuKC objects are (generalised) relations between U and X or maps ofthe form U ×XαK.ChuKC maps are pairs of maps of C f : U → V , F : Y → X such that
uαF (y) = f(u)βy
To make comparison with GC transparent read equality as
uαF (y)⇔ f(u)βy)
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 31 / 38
Comparing GC and ChuKC
ChuKC pluses: smcc instead of ccctopological spaces, plus other representabilities (see Pratt)ChuKC minuses:No K can make I and ⊥ different!!modalities Barr(1990) harder, also more pullbacks.Lafont and Streicher (LICS’91), ChuKSets is GAMEK , but with modalitybased on GCStructure:
A−◦ChuB = (L1α−◦CβU × Y )
where L1 = pullback ⊂ V U ×XY
similarly for par and tensor..
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 32 / 38
A mild generalisation of GC
Categories DialLC, where C is smcc with products and L a closed poset (orlineale).Objects of DialLC:generalised relations between U and X or maps of the form
U ⊗XαL
DialLC maps are pairs of maps of C f : U → V , F : Y → Xsuch that uαF (y)⇒ f(u)βy(Now implication inherited from L).
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 33 / 38
Vicker’s Topological Systems
A triple (U,α,X), where X is a frame, U is a set, and α : U ×X → 2 abinary relation, is called by Vicker’s a topological system if
when S is a finite subset of X, then
α(u,∧S) ≤ α(u, x) for all x ∈ S.
when S is any subset of X, then
α(u,∨S) ≤ α(u, x) for some x ∈ S.
α(u,>) = 1 and α(u,⊥) = 0 for all u ∈ U .
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 34 / 38
A Frame, did you say?
A partially ordered set (X,≤) is a frame if and only ifEvery subset of X has a join;
if S is a finite subset subset of X it has a meet;
Binary meets distribute over joins:x ∧
∨Y =
∨{x ∧ y : y ∈ Y }
Frames are sometimes called locales. They confuse me a lot.
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 35 / 38
Maps of Topological Systems
If (U,α,X) and (V, β, Y ) are topological systems, a map consists off : U → V and F → X such that uαFy = fuβy.As in Chu spaces this condition can be broken down into two halves:uαFy implies fuβy and fuβy implies uαFy.If we only have one half, we are back into a dialectica-like situation.
Call this the category of lax topological systems, TopSysL.What can we prove about TopSysL?We know it has an SMCC structure.Anything else? Would topologists be interested?
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 36 / 38
Further Work?
Presented 2 kinds of dialectica categories and four theorems aboutthem, explaining where they came fromHinted at some others...Not enough work done on these dialectica categories all togetherIteration,(co-)recursion not touched, spent three days thinking aboutit...We’re thinking about instances of the construction appropriate forcardinal invariantsAlso about doing primitive recursion categoricallyWork on logical extensions, model theory should be attempted (vanBenthem little paper)
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 37 / 38
Some References
The Dialectica Categories, V.C. V de Paiva, Categories in ComputerScience, Contemporary Maths, AMS vol 92, 1989.Dialectica-like Categories, V.C. V de Paiva, Proc. of CTCS’89, LNCSvol 389.The Dialectica categories, Ph D thesis, available as Technical Report213, Computer Lab, University of Cambridge, UK, 1991.Categorical Proof Theory and Linear Logic –preliminary notes forSummer SchoolLineales, Hyland, de Paiva, O que nos faz pensar? 1991A Dialectica Model of the Lambek Calculus, Proc. 8th AmsterdamColloquium’91.A linear specification language for Petri nets, Brown, Gurr & de Paiva,TR ’91Full Intuitionistic Linear Logic, Hyland & de Paiva APAL’93 273–291Abstract Games for Linear Logic, Extended Abstract, Hyland & SchalkCTCS’99.
Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 38 / 38