Download lecture16

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
Finite Model Theory
Lecture 16
Lw1w Summary
and 0/1 Laws
1
Outline
• Summary on Lw1w
– All you need to know in 5 slides !
• Start 0/1 Laws: Fagin’s theorem
– Will continue next time
New paper:
Infinitary Logics and 0-1 Laws, Kolaitis&Vardi, 1992
2
Summary on
w
L 1w
Notation Comes from in classical logic
• Lab = formulas where:
– Conjunctions/disjunctions of ordinal < a
Çi 2 g fi, Æi 2 g, where g < a
– Quantifier chains of ordinal < b
9i 2 g xi. f, where g < b
• Hence, L1w = [a Law
3
Summary on
w
L 1w
Motivation
• Any algorithmic computation that applies FO formulas
is expressible in Lw1w
• Relational machines
• While-programs with statements R := f
• Fixpoint logics: LFP, IFP, PFP, etc, etc
Consequence: cannot express EVEN, HAMILTONEAN
4
Summary on
w
L 1w
Canonical Structure
Any algorithmic computation on A can be decomposed
• Compute the ¼k equivalence relation on k-tuples, and
order the equivalence classes ) in LFP
[how do we choose k ???]
• Then compute on ordered structure ) any complexity
Consequence: PTIME=PSPACE iff IFP=PFP
But note that DTC  TC yet L ? NL [ why ?]
5
Summary on
w
L 1w
Pebble Games: with k pebbles
• Notation: A 1wk B if duplicator wins
Theorem 1. For any two structures A, B:
• A, B are Lk1w equivalent iff
• A 1wk B
Theorem 2. If A, B are finite:
• A, B are FOk equivalent iff
• A, B are Lk1w equivalent iff
• A 1wk B
6
Summary on
w
L 1w
Definability of FOk types
• FOk types are the same as Lk1w types [ why ?]
Theorem [Dawar, Lindell, Weinstein] The type of A
(or of (A, a)) can be expressed by some f 2 FOk
B ² f[b] iff Tpk(A,a) = Tpk(B,b)
Difficult result: was unknown to Kolaitis&Vardi
7
0/1 Laws in Logic
Motivation: random graphs
• 0/1 law for FO proven by Glebskii et al., then
rediscovered by Fagin (and with nicer proof)
– Only for constant probability distribution
• Later extended to other logics, and other
probability distributions
Why we care: applications in degrees of belief,
probabilistic databases, etc.
8
Definitions
• Let s = a vocabulary
• Let n ¸ 0, and An µ STRUCT[s] be all
models over domain {0, 1, …, n-1}
• Uniform probability distribution on An
• Given sentence f, denote mn(f) its
probability
9
Definition
• Denote m(f) = limn ! 1 mn(f) if it exists
Definition A logic L has a convergence law if
for every sentence f, m(f) exists
Definition A logic L has a 0/1 law if for every
sentence f, m(f) exists and is 0 or 1
10
Theorems
• Suppose s has no constants
Theorem [Fagin 76, Glebskii et al. 69]
FO admits a 0/1 law
Theorem [Kolaitis and Vardi 92]
Lw1w admits a 0/1 law
11
Application
• What does this tell us for database query
processing ?
• Don’t bother evaluating a query: it’s either
true or false, with high probability 
12
Examples [ in class ]
• Compute mn(f), then m(f):
R(0,1) /* I’m using constants here */
R(0,1) Æ R(0,3) Æ : R(1,3)
9 x.R(2,x)
: (9 x.9 y.R(x,y))
8 x.8 y.(9 z.R(x,z) Æ R(z,y))
13
Types
• We only need rank-0 types (i.e. no
quantifiers)
• Recall the definition
Definition A type t(x) over variables (x1, …,
xm) is conjunction of a maximally consistent
set of atomic formulas over x1, …, xm
14
Types
The type t(x) says:
• For each i, j whether xi = xj or xi  xj
• For each R and each xi1, …, xip whether
R(xi1, …, xip) or : R(xi1, …, xip)
15
Extension Axioms
Definition Type s(x, z) extends the type t(x) if
{s, t} is consistent;
Equivalently: every conjunct in t occurs in s
Definition The extension axiom for types t, s
is the formula
tt,s = 8 x1…8 xk (t(x) ) 9 z.s(x, z))
16
Example of Extension Axiom
t(x1, x2, x3) =
x1  x2 Æ x2  x3 Æ x1  x3 Æ
R(x1,x2) Æ R(x2,x3) Æ R(x2,x2) Æ
: R(x1, x1) Æ : R(x2, x1) Æ …
s(x1, x2, x3, z) =
t(x1, x2, x3) Æ
z  x1 Æ z  x2 Æ z  x3 Æ
R(z,x1) Æ R(x3,z) Æ R(z,z) Æ
: R(x1, z) Æ : (z, x2) Æ …
x1
z
x2
x3
17
Example of Extension Axiom
tt,s =
8 x1.8 x2.8 x3. (t(x1, x2, x3) ) 9 z. s(x1, x2, x3, z))
18
The Theory T
• Let T be the set of all extension axioms
– Studied by Gaifman
• Is T consistent ?
– In a model of T the duplicator always wins [ why ? ]
• Does it have finite models ?
• Does it have infinite models ?
19
The Theory T
• Let qk be the conjunction of all extension axioms
for types with up to k variables
• There exists a finite model for qk [why ?]
• Hence any finite subset of T has a model
• Hence T has a model. [can it be finite ?]
20
The Model(s) of T
• T has no finite models, hence it must have
some infinite model
• By Lowenheim-Skolem, it has a countable
model
21
The Theory T
Theorem T is w-categorical
Proof: let A, B be two countable model.
Idea: use a back-and-forth argument to find an
isomorphism f : A ! B
22
The Theory T
Theorem T is w-categorical
Proof: (cont’d)
A = {a1, a2, a3, ….} B = {b1, b2, b3, ….}
Build partial isomorphisms f1 µ f2 µ f3 µ …
such that:
8 n.9 m. an 2 dom(fm)
and
8 n.9 m. bn 2 rng(fm)
[in class]
Then f = ([m ¸ 1 fm) : A ! B is an isomorphism
23
The Theory T
Corollary T has a unique countable model R
• R = the Rado graph
= the “random” graph
Corollary The theory Th(T) is complete
24
0/1 Law for FO
Lemma
For every extension axiom t, m(t) = limn mn(t) = 1
Proof: later
Corollary For any m extension axioms t1, …, tm:
m(t1 Æ … Æ tm) = 1
Proof
mn(:(t1 Æ … Æ tm))
= mn(: t1 Ç … Ç : tm)
· mn(: t1) + … + mn(: tm) ! 0
25
Fagin’s 0/1 Law for FO
Theorem
For every f 2 FO, either m(f) = 0 or m(f) = 1.
Proof.
Case 1: R ² f. Then there exists m extension
axioms s.t. t1, …, tm ² f. Then mn(f) ¸ mn(t1 Æ
… Æ tm) ! 1
Case 2: R 2 f. Then R ² : f, hence m(: f) = 1, and
m(f) = 0
26
Proof for the Extension Axioms
• Let t = 8 x. t(x) ) 9 z.s(x, z)
• Assume wlog that t asserts xi  xj forall i  j.
Denote (x) the formula Æi < j xi  xj
– Hence t(x) = (x) Æ t’(x)
• Similarly, s asserts z  xi forall i.
Denote (x, z) = Æi xi  z
– Hence s(x, z) = t(x) Æ (x, z) Æ s’(x, z)
where all atomic predicates in s’(x, z) contain z
• Hence:
t = 8 x.((x) Æ t’(x) ) 9 z. (x,z) Æ s’(x, z))
27
Proof for the Extension Axioms
: t = 9 x.((x) Æ t’(x) Æ 8 z.((x, z) ) : s’(x, z)))
mn(: t) · mn(9 x.((x) Æ 8 z.((x, z) ) : s’(x, z))))
28
Proof for the Extension Axioms
mn(: t) · mn(9 x.((x) Æ 8 z.((x, z) ) :s’(x, z))))
· a1, ... , ak 2 {1, …, n} mn(8 z. ((x, z) ) :s’(a1, …, ak, z)))
= n(n-1)…(n-k+1) mn(8 z. (x, z) ) :s’(1, 2, …, k, z))
· nk mn(8 z. (x, z) ) :s’(1, 2, …, k, z)) =
= nk z=k+1, n : s’(1,2,…,k,z)
/* by independence !! */
= nk ( 1 - 1 / 22k+1 )n-k /* since s’ is about 2k+1 edges */
!0
when n ! 1
29
Complexity
Theorem [Grandjean] The problem whether
m(f) = 0 or 1 is PSPACE complete
30
Discussion
• Old way to think about formulas and
models: finite satsfiability/ validity
FO
f valid
f unsatisfiable
Undecidable
31
Discussion
• New way to think about formulas and
models: probability
m(f)=0
m(f)=1
FO
f valid
f unsatisfiable
PSPACE
32
Related documents