Quantifiers and Measures, Part II Tom a s Jakl & Luca Reggio - - PowerPoint PPT Presentation

quantifiers and measures part ii
SMART_READER_LITE
LIVE PREVIEW

Quantifiers and Measures, Part II Tom a s Jakl & Luca Reggio - - PowerPoint PPT Presentation

Quantifiers and Measures, Part II Tom a s Jakl & Luca Reggio (j.w.w. Mai Gehrke) 26 March 2020 Taking over where Luca stopped... The image of R f : ( Mod n 1 ) V ( Typ n ) is the Stone dual of B x n = x n |


slide-1
SLIDE 1

Quantifiers and Measures, Part II

Tom´ aˇ s Jakl & Luca Reggio (j.w.w. Mai Gehrke) 26 March 2020

slide-2
SLIDE 2

Taking over where Luca stopped...

The image of Rf : β(Modn−1) → V(Typn) is the Stone dual of B∃xn = ∃xnϕ | ϕ ∈ FOn.

1

slide-3
SLIDE 3

Taking over where Luca stopped...

The image of Rf : β(Modn−1) → V(Typn) is the Stone dual of B∃xn = ∃xnϕ | ϕ ∈ FOn. The construction B B∃xn works for any B ֒ → P(Modn) dually given by f : β(Modn) ։ X then we build Rf : β(Modn−1) → V(X). And B∃xn can be identified with a subalgebra of P(Modn).

1

slide-4
SLIDE 4

The Boolean algebra of formulas

Inductively, B(n) = QF(x1, . . . , xn) B(n)

i+1 = the image of

B(n)

∃x1 + . . . + B(n) ∃xn + B(n) i

→ P(Modn) we build FO =

  • n=1

  • i=1

B(n)

i

as a Boolean subalgebra of P(Modω).

2

slide-5
SLIDE 5

Inductive constructions in domain theory

In DTLF operators +, ×, PP, PH, PS, → on the space side dually correspond to enrichments of logic. E.g. function space construction [E → D] adds a layer of implications to the logic.

3

slide-6
SLIDE 6

Inductive constructions in domain theory

In DTLF operators +, ×, PP, PH, PS, → on the space side dually correspond to enrichments of logic. E.g. function space construction [E → D] adds a layer of implications to the logic. The solution of a domain equation D ∼ = σ(D) computed as a bilimit, dually adds logical connectives, step by step.

3

slide-7
SLIDE 7

Vietoris as a space of measures

Closed subsets of a Stone space X ← → finitely additive measures on X → 2 functions µ: Clp(X) → 2 s.t.

  • µ(∅) = 0
  • A ∩ B = ∅ =

⇒ µ(A ∪ B) = µ(A) ∨ µ(B) (where 2 = ({0, 1}, ∧, ∨, 0, 1))

4

slide-8
SLIDE 8

Vietoris as a space of measures

Closed subsets of a Stone space X ← → finitely additive measures on X → 2 functions µ: Clp(X) → 2 s.t.

  • µ(∅) = 0
  • A ∩ B = ∅ =

⇒ µ(A ∪ B) = µ(A) ∨ µ(B) (where 2 = ({0, 1}, ∧, ∨, 0, 1))

4

slide-9
SLIDE 9

Vietoris as a space of measures

Closed subsets of a Stone space X ← → finitely additive measures on X → 2 functions µ: Clp(X) → 2 s.t.

  • µ(∅) = 0
  • A ∩ B = ∅ =

⇒ µ(A ∪ B) = µ(A) ∨ µ(B) (where 2 = ({0, 1}, ∧, ∨, 0, 1)) Via the correspondence C → µC such that µC(A) =    1 C ∩ A = ∅

  • therwise

This yields a homeomorphism V(X) ∼ = M(X, 2).

4

slide-10
SLIDE 10

Quantifiers ← → measures?

  • Existential quantifiers ←

→ space of measures X → 2

5

slide-11
SLIDE 11

Quantifiers ← → measures?

  • Existential quantifiers ←

→ space of measures X → 2

  • Semiring quantifiers ←

→ space of measures X → S (from Logic on Words, adaptable to arbitrary finite models)

5

slide-12
SLIDE 12

Quantifiers ← → measures?

  • Existential quantifiers ←

→ space of measures X → 2

  • Semiring quantifiers ←

→ space of measures X → S (from Logic on Words, adaptable to arbitrary finite models) e.g. for k ∈ S, ϕ(x) ∈ FO, A | = ∃k x.ϕ(x) iff 1 + · · · + 1

  • for every a∈A s.t. A|

=ϕ(a)

= k in S

5

slide-13
SLIDE 13

Quantifiers ← → measures?

  • Existential quantifiers ←

→ space of measures X → 2

  • Semiring quantifiers ←

→ space of measures X → S (from Logic on Words, adaptable to arbitrary finite models) e.g. for k ∈ S, ϕ(x) ∈ FO, A | = ∃k x.ϕ(x) iff 1 + · · · + 1

  • for every a∈A s.t. A|

=ϕ(a)

= k in S

  • “Probabilistic quantifiers” ←

→ probabilistic measures X → [0, 1] (from structural limits)

5

slide-14
SLIDE 14

Quantifiers ← → measures?

  • Existential quantifiers ←

→ space of measures X → 2

  • Semiring quantifiers ←

→ space of measures X → S (from Logic on Words, adaptable to arbitrary finite models) e.g. for k ∈ S, ϕ(x) ∈ FO, A | = ∃k x.ϕ(x) iff 1 + · · · + 1

  • for every a∈A s.t. A|

=ϕ(a)

= k in S

  • “Probabilistic quantifiers” ←

→ probabilistic measures X → [0, 1] (from structural limits)

  • more... ?

5

slide-15
SLIDE 15

Stone pairings [Neˇ setˇ ril, Ossona de Mendez, 2013]

For a formula ϕ(x1, . . . , xn) and a finite structure A, ϕ, A = |{ a ∈ An | A | = ϕ(a) }| |A|n (Stone pairing)

6

slide-16
SLIDE 16

Stone pairings [Neˇ setˇ ril, Ossona de Mendez, 2013]

For a formula ϕ(x1, . . . , xn) and a finite structure A, ϕ, A = |{ a ∈ An | A | = ϕ(a) }| |A|n (Stone pairing) Mapping A → −, A defines an embedding Fin ֒ → M(Typ, [0, 1]) Recall that Typ is dual to FO, i.e. clopens are of the form ϕ for ϕ ∈ FO.

6

slide-17
SLIDE 17

Stone pairings [Neˇ setˇ ril, Ossona de Mendez, 2013]

For a formula ϕ(x1, . . . , xn) and a finite structure A, ϕ, A = |{ a ∈ An | A | = ϕ(a) }| |A|n (Stone pairing) Mapping A → −, A defines an embedding Fin ֒ → M(Typ, [0, 1]) Recall that Typ is dual to FO, i.e. clopens are of the form ϕ for ϕ ∈ FO. which lifts uniquely to β(Fin) → M(Typ, [0, 1]) Motivation: The limit of (Ai)i is computed as lim

i→∞ −, A in M(Typ, [0, 1]). 6

slide-18
SLIDE 18

The dual space of the image?

What is the dual of X? β(Fin) ։ X ֒ → M(Typ, [0, 1]) M(Typ, [0, 1]) has no non-trivial clopen! = ⇒ Clp(X) ∼ = 2

7

slide-19
SLIDE 19

The dual space of the image?

What is the dual of X? β(Fin) ։ X ֒ → M(Typ, [0, 1]) M(Typ, [0, 1]) has no non-trivial clopen! = ⇒ Clp(X) ∼ = 2 Two possible solutions:

  • 1. Describe X in terms of geometric logic, logic of proximity lattices
  • r de Vries algebras, ...
  • 2. Replace [0, 1] to retain classical logic.

7

slide-20
SLIDE 20

The dual space of the image?

What is the dual of X? β(Fin) ։ X ֒ → M(Typ, [0, 1]) M(Typ, [0, 1]) has no non-trivial clopen! = ⇒ Clp(X) ∼ = 2 Two possible solutions:

  • 1. Describe X in terms of geometric logic, logic of proximity lattices
  • r de Vries algebras, ...
  • 2. Replace [0, 1] to retain classical logic.

Our choice today!

7

slide-21
SLIDE 21

The Stone space Γ (motivation)

Problem: We need to replace [0,1] by a similar space Γ s.t.

  • 1. we can define measures X → Γ
  • 2. the space M(X, Γ) is compact 0-dimensional
  • 3. Stone pairing −, − : Fin → M(Typ, Γ) definable

and is “comparable” with the original Stone pairing

8

slide-22
SLIDE 22

The Stone space Γ (motivation)

Problem: We need to replace [0,1] by a similar space Γ s.t.

  • 1. we can define measures X → Γ
  • 2. the space M(X, Γ) is compact 0-dimensional
  • 3. Stone pairing −, − : Fin → M(Typ, Γ) definable

and is “comparable” with the original Stone pairing Observe: For ϕ(v1, . . . , vk), the Stone pairing ϕ, A takes values in a finite chain In =

  • 0 < 1

n < 2 n < · · · < n − 1 n < 1

  • where n = |A|k.

8

slide-23
SLIDE 23

The Stone space Γ (motivation)

Problem: We need to replace [0,1] by a similar space Γ s.t.

  • 1. we can define measures X → Γ
  • 2. the space M(X, Γ) is compact 0-dimensional
  • 3. Stone pairing −, − : Fin → M(Typ, Γ) definable

and is “comparable” with the original Stone pairing Observe: For ϕ(v1, . . . , vk), the Stone pairing ϕ, A takes values in a finite chain In =

  • 0 < 1

n < 2 n < · · · < n − 1 n < 1

  • where n = |A|k.

= ⇒ Define Γ as an inverse limit of those (discrete) posets!

8

slide-24
SLIDE 24

The Stone space Γ (description)

Define Γ = lim{f n

nm : Inm ։ In}n,m∈N

where f n

nm( a

nm) = ⌊a/m⌋ n . Elements of Γ are vectors (xn)n ∈

  • n

In such that f n

nm(xnm) = xn, for every n, m ∈ N. 9

slide-25
SLIDE 25

The Stone space Γ (description)

Define Γ = lim{f n

nm : Inm ։ In}n,m∈N

where f n

nm( a

nm) = ⌊a/m⌋ n . Elements of Γ are vectors (xn)n ∈

  • n

In such that f n

nm(xnm) = xn, for every n, m ∈ N.

Intuitively: coordinates represent approximations of numbers in [0,1] from bottom. The larger the n the better the approximation. This gives   

  • ne representation of irrational numbers: r−

two representations of rational numbers: q−, q◦ r− q◦ q− 1◦ 1− 0◦ Γ =

9

slide-26
SLIDE 26

Properties of Γ

  • The subspace topology Γ ⊆

n In is compact 0-dimensional

  • Retraction-section maps Γ

[0, 1]

  • Semicontinuous partial operations − and ∼ on Γ

−, ∼: Γ × Γ ⇀ Γ

10

slide-27
SLIDE 27

Properties of Γ

  • The subspace topology Γ ⊆

n In is compact 0-dimensional

  • Retraction-section maps Γ

[0, 1]

  • Semicontinuous partial operations − and ∼ on Γ

−, ∼: Γ × Γ ⇀ Γ allow to define measures X → Γ monotone functions µ: Clp(X) → Γ s.t.

  • µ(∅) = 0◦, µ(X) = 1◦
  • µ(A) ∼ µ(A ∩ B) ≤ µ(A ∪ B) − µ(B)
  • µ(A) − µ(A ∩ b) ≥ µ(A ∪ B) ∼ µ(B)

10

slide-28
SLIDE 28

Properties of Γ

  • The subspace topology Γ ⊆

n In is compact 0-dimensional

  • Retraction-section maps Γ

[0, 1]

  • Semicontinuous partial operations − and ∼ on Γ

−, ∼: Γ × Γ ⇀ Γ allow to define measures X → Γ monotone functions µ: Clp(X) → Γ s.t.

  • µ(∅) = 0◦, µ(X) = 1◦
  • µ(A) ∼ µ(A ∩ B) ≤ µ(A ∪ B) − µ(B)
  • µ(A) − µ(A ∩ b) ≥ µ(A ∪ B) ∼ µ(B)
  • X → M(X, Γ) endofunctor on Stone spaces
  • We also have −, − : Fin → M(Typ, Γ) such that

M(Typ, Γ) Fin M(Typ, [0, 1])

−,− −,− 10

slide-29
SLIDE 29

Theorem [Gehrke, J., Reggio, 2019]. If X is dual to B then M(X, Γ) is dual to P(B), the free Boolean algebra on the set of generators P≥q ϕ (for ϕ ∈ D, q ∈ [0, 1] ∩ Q) and factored by the congruence | = given below Intuitively, A | = P≥q ϕ if the probability of A | = ϕ(a) is ≥q.

11

slide-30
SLIDE 30

Theorem [Gehrke, J., Reggio, 2019]. If X is dual to B then M(X, Γ) is dual to P(B), the free Boolean algebra on the set of generators P≥q ϕ (for ϕ ∈ D, q ∈ [0, 1] ∩ Q) and factored by the congruence | = given below Intuitively, A | = P≥q ϕ if the probability of A | = ϕ(a) is ≥q. (L1) p ≤ q implies P≥q ϕ | = P≥p ϕ If the probability of A | = ϕ(a) is ≥q then it is also ≥p (L2) ϕ ≤ ψ implies P≥q ϕ | = P≥q ψ (L3) P≥p f | = f for p > 0, t | = P≥0 f, and t | = P≥q t (L4) 0 ≤ p + q − r ≤ 1 implies P≥p+q−r (ϕ ∨ ψ) ∧ P≥r (ϕ ∧ ψ) | = P≥p ϕ ∨ P≥q ψ and P≥p ϕ ∧ P≥q ψ | = P≥p+q−r (ϕ ∨ ψ) ∨ P≥r (ϕ ∧ ψ)

11

slide-31
SLIDE 31

Theorem [Gehrke, J., Reggio, 2019]. If X is dual to B then M(X, Γ) is dual to P(B), the free Boolean algebra on the set of generators P≥q ϕ (for ϕ ∈ D, q ∈ [0, 1] ∩ Q) and factored by the congruence | = given below Intuitively, A | = P≥q ϕ if the probability of A | = ϕ(a) is ≥q. (L1) p ≤ q implies P≥q ϕ | = P≥p ϕ If the probability of A | = ϕ(a) is ≥q then it is also ≥p (L2) ϕ ≤ ψ implies P≥q ϕ | = P≥q ψ (L3) P≥p f | = f for p > 0, t | = P≥0 f, and t | = P≥q t (L4) 0 ≤ p + q − r ≤ 1 implies P≥p+q−r (ϕ ∨ ψ) ∧ P≥r (ϕ ∧ ψ) | = P≥p ϕ ∨ P≥q ψ and P≥p ϕ ∧ P≥q ψ | = P≥p+q−r (ϕ ∨ ψ) ∨ P≥r (ϕ ∧ ψ) (Think of ¬P≥p ϕ as P<p ϕ)

11

slide-32
SLIDE 32

Stone pairing logically

  • Then M(Typ, Γ) is the space of complete consistent theories for

the logic of P(FO) (i.e. an extension of FO).

12

slide-33
SLIDE 33

Stone pairing logically

  • Then M(Typ, Γ) is the space of complete consistent theories for

the logic of P(FO) (i.e. an extension of FO).

  • Stone pairing

Fin → M(Typ, Γ), A → −, A maps A ∈ Fin the theory containing {P≥p ϕ | ϕ, A ≥ p} ∪ {P<p ϕ | ϕ, A < p}

12

slide-34
SLIDE 34

Stone pairing logically

  • Then M(Typ, Γ) is the space of complete consistent theories for

the logic of P(FO) (i.e. an extension of FO).

  • Stone pairing

Fin → M(Typ, Γ), A → −, A maps A ∈ Fin the theory containing {P≥p ϕ | ϕ, A ≥ p} ∪ {P<p ϕ | ϕ, A < p}

  • The space X in β(Fin) ։ X ֒

→ M(Typ, Γ) is dual to P(FO)/∼ where P≥p ϕ ∼ P≥q ψ iff ∀A ∈ Fin ϕ, A ≥ p ↔ ψ, A ≥ q

12

slide-35
SLIDE 35

Conclusion

Topological or duality theoretical techniques elsewhere:

  • Duality-theoretical story in database theory (schema mappings)?
  • Can duality theory say something interesting about Pk, Ek, Mk?
  • Topological approach to 0–1 laws?
  • Logical approach to probabilistic powerdomains? Or replace [0,1]

by Γ as the valuation space?

  • Is the slogan “quantifiers ←

→ measures” justified? More examples? Counterexamples?

13

slide-36
SLIDE 36

Thank you! (see arXiv:1907.04036 for details about probabilistic quantifiers)

14