Dialectica Categories Surprising Application: mapping cardinal invariants

Dialectica Categories Surprising Application:
mapping cardinal invariants

Valeria de Paiva

Rearden Commerce
University of Birmingham

March, 2012

Valeria de Paiva (Rearden Commerce University of Birmingham ) March, 2012 1 / 46

Outline

Outline

1 Dialectica Categories

2 Computational Complexity

3 Cardinal characteristics

4 Structure of PV

5 More Dialecticas


Dialectica Categories

Introduction

I’m a logician professionally, but I feel like an outsider in this conference.
Some times being an outsider is a good thing.
I reckon that mathematicians should try some more of it,
especially with diﬀerent kinds of mathematics.

“There are two ways to do great mathematics. The ﬁrst way is to
be smarter than everybody else. The second way is to be stupider
than everybody else – but persistent." – Raoul Bott



A little bit of personal history...

Some twenty years ago I ﬁnished my PhD thesis "The Dialectica
Categories" in Cambridge. My supervisor was Dr Martin Hyland.



Dialectica categories came from Gödel’s Dialectica
Interpretation

The interpretation is named after the Swiss journal Dialectica where it
appeared in a special volume dedicated to Paul Bernays 70th birthday in
1958.
I was originally trying to provide an internal categorical model of the
Dialectica Interpretation. The categories I came up with proved (also) to
be a model of Linear Logic...



Dialectica categories are models of Linear Logic

Linear Logic was created by Girard (1987) as a proof-theoretic tool: the
dualities of classical logic plus the constructive content of proofs of
intuitionistic logic.
Linear Logic: a tool for semantics of Computing.



Dialectica Interpretation: motivations...

For Gödel (in 1958) the interpretation was a way of proving consistency of
arithmetic. Aim: liberalized version of Hilbert’s programme – to justify
classical systems in terms of notions as intuitively clear as possible.
Since Hilbert’s ﬁnitist methods are not enough (Gödel’s incompleteness theorem and experience with consistency

proofs) must admit some abstract notions. G’s approach: computable (or primitive recursive) functionals of ﬁnite

type.

For me (in 1988) an internal way of modelling Dialectica that turned out to
produce models of Linear Logic instead of models of Intuitionistic Logic,
which were expected...
For Blass (in 1995) a way of connecting work of Votjáš in Set Theory with
mine and his own work on Linear Logic and cardinalities of the continuum.



Dialectica categories: useful for proving what...?

Blass (1995) Dialectica categories as a tool for proving inequalities between
nearly countable cardinals.
Questions and Answers – A Category Arising in Linear Logic, Complexity
Theory, and Set Theory in Advances in Linear Logic (ed. J.-Y. Girard, Y.
Lafont, and L. Regnier) London Math. Soc. Lecture Notes 222 (1995).
Also Propositional connectives and the set theory of the continuum (1995)
and the survey Nearly Countable Cardinals.



Questions and Answers: The short story

Blass realized that my category for modelling Linear Logic was also used by
Peter Votjáš for set theory, more speciﬁcally for proving inequalities
between cardinal invariants and wrote Questions and Answers – A Category
Arising in Linear Logic, Complexity Theory, and Set Theory (1995).
When we discussed the issue in 1994/95 I simply did not read the sections
of the paper on Set Theory. or Computational Complexity.
Two years ago I learnt from Samuel about his and Charles work using
Blass/Votjáš’ ideas and got interested in understanding the cardinal
invariants connections.
Late last year we decided that a short visit would be a good way forward.
This is a visit to start a collaboration. Hence I will be talking about old
results...



Questions and Answers: The categories GSets and PV
Is this simply a coincidence?

The second dialectica category in my thesis GSets (for Girard’s sets) is the
dual of GT the Galois-Tukey connections category in Votjáš work.
Blass calls this category PV.
The objects of PV are triples A = (U, X, α) where U, X are sets and
α ⊆ U × X is a binary relation, which we usually write as uαx or α(u, x).
+
Blass writes it as (A− , A , A) but I get confused by the plus and minus signs.

Two conditions on objects in PV (not the case in GSets):
1. U, X are sets of cardinality at most |R|.
2. The condition in Moore, Hrusák and Dzamonja (MHD) holds:

∀u ∈ U ∃x ∈ X such that α(u, x)

and
∀x ∈ X∃u ∈ U such that ¬α(u, x)



Questions and Answers: The category PV
In Category Theory morphisms are more important than objects.
Given objects A = (U, X, α) and B = (V, Y, β)
a map from A to B in GSets is a pair of functions f : U → V and
F : Y → X such that α(u, F y) implies β(f u, y).
Usually I write maps as

U α X
6

f ⇓ F ∀u ∈ U, ∀y ∈ Y α(u, F y) implies β(f u, y)
?
V β Y
But a map in PV satisﬁes the dual condition that is β(f u, y) → α(u, F y).
Trust set-theorists to create morphisms A → B where the relations go in
the opposite direction!...

Computational Complexity

Can we give some intuition for these morphisms?

Blass makes the case for thinking of problems in computational complexity.
Intuitively an object of GSets or PV

(U, X, α)

can be seen as representing a problem.
The elements of U are instances of the problem, while the elements of X
are possible answers to the problem instances.
The relation α say whether the answer is correct for that instance of the
problem or not.



Which problems?

A decision problem is given by a set of instances of the problem, together
with a set of positive instances.
The problem is to determine, given an instance whether it is a positive one
or not.
Examples:
1. Instances are graphs and positive instances are 3-colorable graphs;
2. Instances are Boolean formulas and positive instances are satisﬁable
formulas.
A many-one reduction from one decision problem to another is a map
sending instances of the former to instances of the latter, in such a way
that an instance of the former is positive if and only its image is positive.
An algorithm computing a reduction plus an algorithm solving the latter
decision problem can be combined in an algorithm solving the former.



Computational Complexity 101

A search problem is given by a set of instances of the problem, a set of
witnesses and a binary relation between them.
The problem is to find, given an instance, some witness related to it.
The 3-way colorability decision problem (given a graph is it 3-way
colorable?) can be transformed into the 3-way coloring search problem:
Given a graph find me one 3-way coloring of it.
Examples:
1. Instances are graphs, witnesses are 3-valued functions on the vertices of
the graph and the binary relation relates a graph to its proper 3-way
colorings
2. Instances are Boolean formulas, witnesses are truth assignments and the
binary relation is the satisfiability relation.
Note that we can think of an object of PV as a search problem, the set of
instances is the set U , the set of witnesses the set X.



Computational Complexity 101

A many-one reduction from one search problem to another is a map
sending instances of the former to instances of the latter, in such a way
that an instance of the former is positive if and only its image is positive.
An algorithm computing a reduction plus an algorithm solving the latter
decision problem can be combined in an algorithm solving the former.
We can think of morphisms in PV as reductions of problems.
If this intuition is useful, great, if not, carry on thinking simply of sets and
a relation and complicated notion of how to map triples to other triples.



Questions and Answers

Objects in PV are only certain objects of GSets, as they have to satisfy the
two extra conditions. What happens to the structure of the category when
we restrict ourselves to this subcategory?
Fact: GSets has products and coproducts, as well as a terminal and an
initial object.
Given objects of PV, A = (U, X, α) and B = (V, Y, β)
the product A × B in GSets is the object (U × V, X + Y, choice)
where choice : U × V × (X + Y ) → 2 is the relation sending
(u, v, (x, 0)) to α(u, x) and (u, v, (y, 1)) to β(v, y).
The terminal object is T = (1, 0, e) where e is the empty relation,
e : 1 × 0 → 2 on the empty set.



Questions and Answers

Similarly the coproduct of A and B in GSets is the object
(U + V, X × Y, choice)
where choice : U + V × (X × Y ) → 2 is the relation sending
((u, 0), x, y) to α(u, x) and ((v, 1), x, y) to β(v, y)
The initial object is 0 = (0, 1, e) where e : 0 × 1 → 2 is the empty relation.
Now if the basic sets all U, V, X, Y have cardinality up to |R| then
(co-)products will do the same.
But the MHD condition is a diﬀerent story.



The structure of PV

Note that neither T or 0 are objects in PV,
as they don’t satisfy the MHD condition.

∀u ∈ U, ∃x ∈ X such that α(u, x) and ∀x ∈ X∃u ∈ U such that ¬α(u, x)

To satisfy the MHD condition neither U nor X can be empty.
Also the object I = (1, 1, id) is not an object of PV, as it satisfies the first
half of the MHD condition, but not the second. And the object
⊥ = (1, 1, ¬id) satisfies neither of the halves.
The constants of Linear Logic do not fare too well in PV.



The structure of PV

Back to morphisms of PV, the use that is made of the category in
applications is simply of the pre-order induced by the morphisms.

It is somewhat perverse that here, in contrast to usual categorical logic,

A ≤ B ⇐⇒ There is a morphism from B to A



Examples of objects in PV

1. The object (N, N, =) where n is related to m iff n = m.
To show MHD is satisfied we need to know that ∀n ∈ N∃m ∈ N(n = m),
can take m = n. But also that ∀m ∈ N∃k ∈ N such that ¬(m = k). Here
we can take k = suc(m).
2. The object (N, N, ≤) where n is related to m iff n ≤ m.
3. The object (R, R, =) where r1 and r2 are related iff r1 = r2 , same
argument as 1 but equality of reals is logically much more complicated.
4. The objects (2, 2, =) and (2, 2, =) with usual equality.



What about GSets?

GSets is a category for categorical logic, we have:

A ≤ B ⇐⇒ There is a morphism from A to B

Examples of objects in GSets:
“Truth-value" constants of Linear Logic as discussed T , 0, ⊥ and I.
All the PV objects are in GSets. Components of objects such as U, X are
not bound above by the cardinality of R.
Also the object 2 of Sets plays an important role in GSets, as our relations
α are maps into 2, but the objects of the form (2, 2, α) have played no
major role in GSets so far.


Cardinal characteristics

What have Set Theorists done with PV?
Cardinal Characteristics of the Continuum

Blass: “One of Set Theory’s first and most important contributions to
mathematics was the distinction between different infinite cardinalities,
especially countable infinity and non-countable one."
Write N for the natural numbers and ω for the cardinality of N.
Similarly R for the reals and 2ω for their cardinality.
All the cardinal characteristics considered will be smaller or equal to the
cardinality of the reals.
They are of little interest if the Continuum Hypothesis holds, as then there
are no cardinalities between the cardinality of the integers ω and the
cardinality of the reals 2ω .
But if the continuum hypothesis does NOT hold there are many interesting
connections (in general in the form of inequalities) between various
characteristics that Votjáš discusses.



Cardinals from Analysis
I recall the main definitions that Blass uses in Questions and Answers and
his main “theorem":
- If X and Y are two subsets of N we say that X splits Y if both
X ∩ Y and Y X are infinite.
- The splitting number s is the smallest cardinality of any family S of
subsets of N such that every infinite subset of N is split by some
element of S.
Recall the Bolzano-Weierstrass Theorem: Any bounded sequence of
real numbers (xn )n∈N has a convergent subsequence, (xn )n∈A .
One can extend the theorem to say:
For any countably bounded many sequences of real numbers
xk = (xkn )n∈N there is a single infinite set A ⊆ N such that the
subsequences indexed by A, (xkn )n∈A all converge.
If one tries to extend the result for uncountably many sequences s above is
the first cardinal for which the analogous result fail.


Cardinals from Analysis

- If f and g are functions N → N, we say that f dominates g if for all
except ﬁnitely many n’s in N, f (n) ≤ g(n).
- The dominating number d is the smallest cardinality of any family D
contained in NN such that every g in NN is dominated by some f in D.
- The bounding number b is the smallest cardinality of any family
B ⊆ NN such that no single g dominates all the members of B.



Connecting Cardinals to the Category

Blass “theorems” : We have the following inequalities

ω ≤ s ≤ d ≤ 2ω

ω ≤ b ≤ r ≤ rσ ≤ 2ω

b≤d
The proofs of these inequalities use the category of Galois-Tukey
connections and the idea of a "norm" of an object of PV.
(and a tiny bit of structure of the category...)



The structure of the category PV

Given an object A = (U, X, α) of PV its "norm" ||A|| is the smallest
cardinality of any set Z ⊆ X suﬃcient to contain at least one correct
answer for every question in U .
Blass again: “It is an empirical fact that proofs between cardinal
characteristics of the continuum usually proceed by representing the
characteristics as norms of objects in PV and then exhibiting explicit
morphisms between those objects.”
One of the aims of our proposed collaboration is to explain this empirical
fact, preferably using categorical tools.
Haven’t done it, yet. So will try to explain some of the easy instances.


Structure of PV

The structure of PV
Proposition[Rangel] The object (R, R, =) is maximal amongst objects of
PV.
Given any object A = (U, X, α) of PV we know both U and X have
cardinality small than |R|. In particular this means that there is an injective
function ϕ : U → R. (let ψ be its left inverse, i.e ψ(ϕu) = u)
Since α is a relation α ⊆ U × X over non-empty sets, if one accepts the
Axiom of Choice, then for each such α there is a map f : U → X such
that for all u in U , uαf (u).
Need a map Φ : R → X such that
U α X
6

ϕ ⇑ Φ ∀u ∈ U, ∀r ∈ R (ϕu = r) → α(u, Φr)
?
R = R

Structure of PV

The structure of PV
Axiom of Choice is essential

Given cardinality fn ϕ : U → R, let ψ be its left inverse, ψ(ϕu) = u, for all
u ∈ U . Need a map Φ : R → X such that
U α X
6

ϕ ⇑ Φ ∀u ∈ U, ∀r ∈ R (ϕu = r) → α(u, Φr)
?
R = R
Let u and r be such that ϕu = r and define Φ as ψ ◦ f .
Since ϕu = r can apply Φ to both sides to obtain Φ(ϕ(u)) = Φ(r).
Substituting Φ’s definition get f (ψ(ϕu))) = Φ(r). As ψ is left inverse of ϕ
(ψ(ϕu) = u) get f u = Φ(r).
Now the definition of f says for all u in U uαf u, which is uαΦr holds, as
desired. (Note that we did not need the cardinality function for X.)

Structure of PV

Fun meeting somewhere else


Structure of PV

The structure of PV

Proposition[Rangel] The object (R, R, =) is minimal amongst objects of
PV.
This time we use the cardinality function ϕ : X → R for X. We want a
map in PV of the shape:

R = R
6

Φ ⇑ ϕ ∀r ∈ R, ∀x ∈ X α(Φr, x) → (r = ϕx)
?
U α X
Now using Choice again, given the relation α ⊆ U × X we can ﬁx a
function g : X → U such that for any x in X g(x) is such that ¬g(x)αx.


Structure of PV

The structure of PV
Axiom of Choice is essential

Given cardinality fn ϕ : X → R, let ψ : R → X be its left inverse,
ψ(ϕx) = x, for all x ∈ X. Need a map Φ : R → U such that
R = R
6

Φ ⇑ ϕ ∀r ∈ R, ∀x ∈ X α(Φr, x) → ¬(ϕx = r)
?
U α X
Exactly the same argument goes through. Let r and x be such that
ϕx = r and define Φ as ψ ◦ g.
Since ϕx = r can apply Φ to both sides to obtain Φ(ϕ(x)) = Φ(r).
Substituting Φ’s definition get g(ψ(ϕx))) = Φ(r). As ψ is left inverse of ϕ,
(ψ(ϕx) = x) get gx = Φ(r).
But the definition of g says for all x in X ¬gxαx, which is ¬α(Φr, x)
holds. Not quite the desired, unless you’re happy with RAA.

Structure of PV

More structure of GSets

Given objects A and B of GSets we can consider their tensor products.
Actually two diﬀerent notions of tensor products were considered in GSets,
but only one has an associated internal-hom.
(Blass considered also a mixture of the two tensor products of GSets, that
he calls a sequential tensor product.)
Having a tensor product with associated internal hom means that we have
an equation like:

A ⊗ B → C ⇐⇒ A → (B → C)
Can we do the same for PV? Would it be useful?
The point is to check that the extra conditions on PV objects are satisﬁed.


Structure of PV

Tensor Products in GSets

Given objects A and B of GSets we can consider a preliminary tensor
product, which simply takes products in each coordinate. Write this as
A B = (U × V, X × Y, α × β) This is an intuitive construction to
perform, but it does not provide us with an adjunction.
To "internalize" the notion of map between problems, we need to consider
the collection of all maps from U to V , V U , the collection of all maps from
Y to X, X Y and we need to make sure that a pair f : U → V and
F : Y → X in that set, satisﬁes our dialectica (or co-dialectica) condition:

∀u ∈ U, y ∈ Y, α(u, F y) ≤ β(f u, y) (respectively ≥)
This give us an object (V U × X Y , U × Y, eval) where
eval : V U × X Y × (U × Y ) → 2 is the map that evaluates f, F on the pair
u, y and checks the implication between relations.


Structure of PV

More structure in GSets

By reverse engineering from the desired adjunction, we obtain the ‘right’
tensor product in the category.
The tensor product of A and B is the object (U × V, Y U × X V , prod),
where prod : (U × V ) × (Y U × X V ) → 2 is the relation that ﬁrst evaluates
a pair (ϕ, ψ) in Y U × X V on pairs (u, v) and then checks the
(co)-dialectica condition.
Blass discusses a mixture of the two tensor products, which hasn’t showed
up in the work on Linear Logic, but which was apparently useful in Set
Theory.


Structure of PV

An easy theorem of GSets/PV...

Because it’s fun, let us calculate that the reverse engineering worked...

A ⊗ B → C if and only if A → [B → C]

U × V α ⊗ βX V × Y U U α X
6 6

f ⇓ (g1 , g2 ) ⇓
? ?
V
W γ Z W × Y Zβ → γ V × Z


More Dialecticas

More Original Dialectica Categories

My thesis has four chapters, four main definitions and four main theorems.
The first two chapters are about the “original" dialectica categories.
Theorem (V de Paiva, 1987)
If C is a ccc with stable, disjoint coproducts, then Dial(C) has products,
tensor products, units and a linear function space (1, ×, ⊗, I, →) and
Dial(C) is symmetric monoidal closed.

This means that Dial(C) models Intuitionistic Linear Logic (ILL)
without modalities. How to get modalities? Need to define a special
comonad and lots of work to prove theorem 2...


More Dialecticas

Original Dialectica Categories

!A must satisfy !A →!A⊗!A, !A ⊗ B →!A, !A → A and !A →!!A,
together with several equations relating them.
The point is to define a comonad such that its coalgebras are commutative
comonoids and the coalgebra and the comonoid structure interact nicely.
Theorem (V de Paiva, 1987)
Given C a cartesian closed category with free monoids (satisfying certain
conditions), we can define a comonad T on Dial(C) such that its Kleisli
category Dial(C)T is cartesian closed.

Define T by saying A = (U, X, α) goes to (U, X ∗ , α∗ ) where X ∗ is the free
commutative monoid on X and α∗ is the multiset version of α.
Loads of calculations prove that the linear logic modalitiy ! is well-defined
and we obtain a full model of ILL and IL, a posteriori of CLL.
Construction generalized in many ways, cf. dePaiva, TAC, 2006.


More Dialecticas

What is the point of (these) Dialectica categories?

First, the construction ends up as a model of Linear Logic, instead of
constructive logic.This allows us to see where the assumptions in Godel’s
argument are used (Dialectica still a bit mysterious...)
It justiﬁes linear logic in terms of a more traditional logic tool and
conversely explains the more traditional work in terms of a ‘modern’ (linear,
resource conscious) decomposition of concerns.
Theorems(87/89): Dialectica categories provide models of linear logic as
well as an internal representation of the dialectica interpretation. Modeling
the exponential ! is hard, ﬁrst model to do it. Still (one of) the best ones.


More Dialecticas

Dialectica categories: 20 years later...

It is pretty: produces models of Intuitionistic and classical linear logic and
special connectives that allow us to get back to usual logic.
Extended it in a number of directions:
a robust proof can be pushed in many ways...

used in CS as a model of Petri nets (more than 3 phds),
it has a non-commutative version for Lambek calculus (linguistics),
it has been used as a model of state (with Correa and Hausler, Reddy ind.)
Also in Categorical Logic: generic models (with Schalk04) of Linear Logic,
Dialectica interp of Linear Logic and Games (Shirahata and Oliveira)
ﬁbrational versions (B. Biering and P. Hofstra).
Most recently and exciting:
Formalization of partial compilers: correctness of MapReduce in MS .net
frameworks via DryadLINQ, “The Compiler Forest", M. Budiu, J. Galenson,
G. Plotkin 2011.


More Dialecticas

Conclusions

Introduced you to dialectica categories GSets/PV.
Hinted at Blass and Votjáš use of them for mapping cardinal invariants.
Uses for Categorical Proof Theory very diﬀerent from uses in Set Theory
for cardinal invariants.
Showed one easy, but essential , theorem in categorical logic.
But haven’t even started looking at Parametrized Diamond Principles...
Haven’t even started talking about "lax topological systems" in the sense
of Vickers, a diﬀerent connection to Topology.

Believe it is not a simple coincidence that dialectica categories are useful in
these disparate areas.
We’re starting our collaboration, so hopefully real "new" theorems will
come up.


More Dialecticas

More Conclusions...

Working in interdisciplinary areas is hard, but rewarding.
The frontier between logic, computing, linguistics and categories is a fun
place to be.
Mathematics teaches you a way of thinking, more than specific theorems.
Fall in love with your ideas and enjoy talking to many about them...

Thanks Samuel, Marcelo and all locals for this lovely meeting.
Thanks to Charles and Samuel for mentioning my work in connection to
their stuff on parametrized Diamond principles and MHD’s work...
and thanks Samuel, Andreas and Thierry for all the effort to bring me here.


More Dialecticas

References

Categorical Semantics of Linear Logic for All, Manuscript.
Dialectica and Chu constructions: Cousins?Theory and Applications of
Categories, Vol. 17, 2006, No. 7, pp 127-152.
A Dialectica Model of State. (with Correa and Haeusler). In CATS’96,
Melbourne, Australia, Jan 1996.
Full Intuitionistic Linear Logic (extended abstract). (with Martin Hyland).
Annals of Pure and Applied Logic, 64(3), pp.273-291, 1993.
Thesis TR: The Dialectica Categories
http://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-213.html

A Dialectica-like Model of Linear Logic.In Proceedings of CTCS,
Manchester, UK, September 1989. LNCS 389
The Dialectica Categories. In Proc of Categories in Computer Science and
Logic, Boulder, CO, 1987. Contemporary Mathematics, vol 92, American
Mathematical Society, 1989
all available from http://www.cs.bham.ac.uk/~vdp/publications/papers.html


More Dialecticas

Functional Interpretations and Gödel’s Dialectica

Starting with Gödel’s Dialectica interpretation(1958) a series of
"translation functions" between theories
Avigad 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 on
intuitionistic logic; then one reduces the theory I to a
quantifier-free functional theory F.

Examples of functional interpretations:
Kleene’s realizability
Kreisel’s modified realizability
Kreisel’s No-CounterExample interpretation
Dialectica interpretation
Diller-Nahm interpretation, etc...


More Dialecticas

Gödel’s Dialectica Interpretation
For each formula A of HA we associate a formula of the form
AD = ∃u∀xAD (u, x) (where AD is a quantifier-free formula of Gödel’s
system T) inductively as follows: when Aat is an atomic formula, then its
interpretation is itself.
Assume we have already defined AD = ∃u∀x.AD (u, x) and
B D = ∃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 (f u; 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).

More Dialecticas

Where is my category theory?


Dialectica Categories Surprising Application: mapping cardinal invariants

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Dialectica Categories Surprising Application: mapping cardinal invariants (20)

More from Valeria de Paiva (16)

Recently uploaded (20)

Dialectica Categories Surprising Application: mapping cardinal invariants