4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 1
On the Very Weak 01 Law for Random
Graphs with Orders
Saharon Shelah
The Hebrew University, Math Institute
Rutgers University, Math Department
Abstract: Let us draw a graph R on 0, 1, . . . , n1 by having an edge i, j with
probability p
ij
, where
i
p
i
< , and let M
n
= (n, <, R). For a rst
order sentence let a
n
be the probability of M
n
[= . We know that the
sequence a
1
, a
2
, . . . , a
n
a
n+1
=: Prob(M
n
[= ) of M
n
[= does not necessarily converge
(but if
i
ip
i
< then it converges, and the value of
p
i
is not enough,
in general, to determine convergence). The theorem in the abstract appears in
1 and is proved in 3, it says that the sequence of probabilities still behaves
(somewhat) nicely.
The rst results (in various probabilistic distributions) on the asymptomatic
behavior of a
n
(see Glebski et al [GKLT], Fagin [F] and survey [Sp]) say that it
converges to 0 or 1 hence the name zero one law. In other cases weaker results
were gotten: a
n
a
n
[ converges to zero. We also dene hvery weak zero one law (see
Denition 1.2(2)), but concentrate on the one above. Note that many examples
to nonconvergence are done by nding such that e.g. if log(n) = 1 mod 10
then a
n
a
n+1
n
nice enough, ensuring holds, while the probability changes little (see
3). Now M
n
is almost the sum of M
n
(m
i
, m
i+1
), precisely M
n
is determined
by M
n
[m
i
, m
i+2
), for i = 0, 1, 2, . . ., so we call it a distorted sum. Now a model
theoretic lemma from 2 on the ntheory of a distorted sum of models enables
us to prove the main theorem in 3 (on the model theoretic background see 2).
Later in 4, 5 we deal with some renements not needed for the main theorem
(1.4).
In 1 we also get the very weak 01 law for a random partial order suggested
by Luczak. In a subsequent paper [Sh 548] we prove the very weak zero law for
some other very natural cases: e.g. for a random 2place function and for (n, <, R)
with < the natural order, R a random graph (=symmetric irreexive relation)
with edge probability p. In another one, [Sh 467] we deal with zero one law for
the random model from the abstract with p
i
=
1
i
a
(mainly: no order, a (0, 1)
irrational). See also [Sh 550], [Sh 551], [Sh 581]. Spencer is continuing [Sh 548]
2
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 3
looking at the exact h for which hvery weak zero one law holds (see Denition
1.2(2) here). I thank Shmuel Lifsches for many corrections.
Notation
N is the set of natural numbers.
We identify n N with the set 0, . . . , n 1.
Z is the set of integers.
R is the set of reals, R
+
is the set of reals which are positive (i.e. > 0).
i, j, k, , m, n, r, s, t are natural numbers.
, are positive reals (or functions with values in R
+
).
f, g, h are functions.
denotes a vocabulary (for simplicity set of predicates), L
fo
is the set of
rst order formulas in the vocabulary . (Generally, if L is a logic L
the set
of sentences (or formulas) in L in the vocabulary and L
fo
is the rst order
logic).
For a rst order sentence (or formula) let d
i<n
M
i
is (assuming
for simplicity the universes are pairwise disjoint) the models dened by: universe
i<n
M
i
, R
M
=
i<n
R
M
i
for R except that if < then: x <
M
y
_
i<j<n
[x M
i
&y M
j
]
_
i<n
x <
M
i
y (similarly with any linear order I
instead of n). We write M
0
+ M
1
instead of
i<2
M
i
.
if()
is if is true, if is false.
We identify true, false with yes, no.
Note: t is a natural number, t kind of depth n theory, t is a truth value, is the
symmetric dierence, denote a set of formulas.
1 The Very Weak Zero One Law.
3
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
4 01 Law
1.1 Denition: 1) A 01 law context is a sequence
K = K
n
,
n
: n < ) such
that:
(a) for some vocabulary =
K
, for every n, K
n
is a family of models,
closed under isomorphism with the family of isomorphism types being a
set.
(b) for each n,
n
is a probability measure on the set of isomorphism types
(of models from K
n
).
2)
K is nitary (countable) if for each n the set M/
= : M K
n
is nite
(countable).
3) For a sentence (not necessarily f.o.) Prob
n
(M
n
[= ) or Prob
Kn
(M
n
[= )
or Prob
n
(M
n
[=
M
n
K
n
) means
n
M/
= : M K
n
, M [= .
4) Instead of clause (a) of (1), we may use K
n
a set of models,
n
a proba
bility measure on K
n
; particularly we introduce a random choice of M
n
; the
translation between the two contexts should be clear.
1.1A Discussion: A 01 law context is not necessarily a context which satises
a 01 law. It is a context in which we can formulate a 01 law, and also weaker
variants.
1.2 Denition: 0) A 01 context
K satises the 01 law for a logic L if for every
sentence L
(with =
K
of course) we have: a
n
def
= Prob
n
(M
n
[= )
converges to zero or converges to 1 when n .
1)
K satises the very weak 01 law for the logic L if for every sentence L
we have: a
n
def
= Prob
n+1
(M
n+1
[=
M
n+1
K
n+1
) Prob
n
(M
n
[=
M
n
K
n
) converges to zero as n .
2)
K satises the hvery weak 01 one law for the logic L if for every sentence
L
, max
m
1
,m
2
[n,h(n)]
_
Prob
m
1
(M
m
1
[=
M
m
1
K
m
1
) Prob
m
2
(M
m
2
[=
M
m
2
K
m
2
)
_
converge to zero as n . (We shall concentrate on part 1).
3)
K satises the convergence law for L if for every L
we have: Prob
n
(M
n
[=
M
n
K
n
) : n < ) converges (to some real [0, 1]
R
). (So if it always converges
to 0 or to 1 we say that
K satises the 01 law for L).
4) If L is the rst order logic we may omit it.
The following
K is from Luczak Shelah [LuSh 435].
1.3 Denition: Let p = p
i
: i N), p
i
a real number 0 p
i
1, p
0
= 0. We
dene
K
og
p
as follows:
the models from K
n
are of the form M = (n, <, R), so we are using the variant
4
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 5
from
M
n
K
n
) is
p
ji
: i < j < n and iRj
1 p
ji
: i < j < n and
iRj, i.e. for each i < j < n we decide whether iRj by ipping a coin with
probability p
ji
for yes, independently for the dierent pairs.
1.4 Theorem:
K
og
satises the very weak 01 law if
i
p
i
< .
This is our main result. We shall prove it later in 3. Luczak [Lu] suggested
another context:
1.5 Denition:
K
opo
pn
is the following 01 law context (p
n
is a function of n,
0 p
n
1; so more precisely we are dening K
opo
pn
: n < )). The models
are of the form, (n, <, <
the
following partial order: we draw a random graph on n (edge relation R) with edge
probability p
n
, x <
Rx
+1
. The probability is derived from the probability distribution
for R (which does not appear in the model.)
1.6 Theorem Assume p
n
=
1
(n+1)
a
, where 0 < a < 1 (like [ShSp 304]). Then
K
opo
pn
satises the very weak zero one law.
Proof: Similar to the previous theorem as easily
(*) for > 0 we have: Prob(M
n
[= (x < y)[x <
y&x + n
1
y) is very
small.
1.6
2 Model Theory: Distorted sum of Models
The main lemma (2.14) generalizes the addition theory and deals with models
with distances (both from model theory). Concerning the rst, see Feferman
Vaught [FV]. The method has its origin in Mostowski [Mo], who dealt with
reduced products. The rst work on other products is Beth [B] who dealt with
the ordinal sum of nitely many ordered systems. For a presentation and history
see Feferman Vaught [FV], pp 5759 and Gurevich [Gu] (particularly the th
n
s).
Concerning models with distance see Gaifman [Gf], a forerunner of which was
Marcus [M1] who deals with the case of M = (M, F, P
i
)
i<n
, F a unary function,
P
i
unary predicates and the distance is as in the graph which the function F
but no two models are isomorphic so the dierence is even more trivial
5
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
6 01 Law
denes (i.e. x, y connected by an edge if: x = F(y) or y = F(x); where for a
graph G the distance is d
G
(x, y) = min k : we can nd x
0
, . . . , x
k
such that:
x = x
0
, y = x
k
, and x
, x
+1
are connected). We may look at our subject here
as dealing with sums with only local disturbances, semi sums; distorted
sums. The connections are explained in 4, 5.
In 2.16 we draw the conclusion for linear order which we shall use in 3 to
prove theorem 1.4, in fact proving the main theorem 1.4 from 2 we use almost
only 2.16. Elsewhere we shall return to improving the numerical bound involved
in the proof see [Sh, F120]
Note: B is a (two sorted) model.
2.1 Denition: 1) We call a vocabulary of systems if =
1
,
2
),
1
,
2
are sets
of predicates (usually nite but not needed).
2) We call B a system if:
(A) B = (M, I, h, d) (= (M
B
, I
B
, h
B
, d
B
)).
(B) M is a
1
model and I is a
2
model, but we use M, I also for their universes.
(C) h is a function from M onto I,
(D) d is a distance function on I, i.e.
() d is a symmetric two place function from I to N = 0, 1, 2, 3, . . . , ,
() d(x, x) = 0
() d(x, z) d(x, y) + d(y, z).
(E) M, I are disjoint.
3) Let (B) = for B a system and
(B) =
if =
1
,
2
).
2.1A Discussion: The demands M, I are disjoint and h is onto I are not
essential, this is just for convenience of presentation. If h is not onto I, we should
allow relations on MI, but then if M, I are not disjoint then for each predicate,
for each of its places we should assign: does there appear a member of M or a
member of I; also we should have two kinds of variables not allowing equality
between variables of the dierent kinds. So in application we may use those
alternative presentations.
2.2 Conventions: (1) We may allow function symbols (and individual constants)
but then treat them as relations.
(2) We stipulate h(x) = x for x I (remember 2.1(2)(E)).
(3) a B means Rang ( a) M
B
I
B
.
6
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 7
(4) A model M will be identied with the system B = B
sim
[M] : M
B
= M,
I
B
= M, h
B
= id
M
,
d(x, y) =
_
0 x = y
x ,= y
(you may of course take two disjoint copies of M as M
B
and I
B
).
(5) From a model M we can derive another system B
dis
[M]: M
B
= M,I
B
= M,
h
B
= id
M
, and d(x, y)
def
= Minn : there are z
0
, . . . , z
n
M, x = z
0
, y = z
n
and
for < n for some R (M), and sequence a R
M
we have z
, z
+1
Rang a
(remember (M) is the vocabulary of M; this is the denition of distance in
Gaifman [Gf]).
2.2A Remark: 1) So the dierence between B
sim
[M] and B
dis
[M] is only in the
choice of the distance function d.
2) Below in L
0
the formula d(x, y) = 0 appears which normally means x = y.
If not we could have in Def. 2.3(3), (4) replaced k r, s n by k < r,
s < n respectively.
2.3 Denition: 1) For a system B = (M, I, h, d) :
for x I,
N
r
(x)
def
= y I : d(y, x) r
for x M I,
N
+
r
(x)
def
= y M I : d(h(y), h(x)) r.
2) L
= (M
B
, I
B
, h), B
r
= (M
B
, I
B
, h
B
, d(x, y) k)
kr
so B
0
= B
and
we let
r
=
r
(B) = (B
r
) (so, B
r
is a two sorted model but not a system).
4) So L
n
is dened as in part (2) but h(y) N
s
(h(x)) appear only for s n.
We usually apply our theorems in the following case:
2.4 Denition: A system B is simple if
(a) d(x, y) 1 (where x, y vary on I) is equivalent to a quantier free formula
in L
0
(B)
.
(b) for x, y I we have: d(x, y) r i there are x
0
, . . . , x
r
I such that x = x
0
,
y = x
r
, and d(x
, x
+1
) 1 for < r (i.e., like 2.2(5)).
7
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
8 01 Law
2.4A Remark: If B is simple, note that every formula in L
(B)
is equivalent to
some formula in L
0
, if we know just that clause (b) of Denition 2.4 is satised
then every formula in L
(B)
is equivalent to some formula in L
1
.
2.5 Convention: 1) We dene f: we let f
n
(r) = r + 3
n
for r, n N or more
generally, f a two place function (written f
n
(r)) from N to N satisfying: f
n
non
decreasing in n and in r, r < f
n
(r) N and f
(3)
n
(r) f
n+1
(r) where f
(0)
n
(r) = r,
f
(+1)
n
(r) = f
n
(f
()
n
(r)) and f
(2)
n
(r) f
n
(r) + f
n
(0).
2) We call f nice if in addition f
(4)
n
(r) f
n+1
(r).
3) For g a function from N to N, let g(r
) : < m).
2.6 Denition: 1) For a system B = (M, I, h, d) and m, n N, and a =
a
0
, . . . , a
m1
) B we dene th
n
r
( a, B), here r stands for a distance. We dene
it by induction
on n:
th
0
r
( a, B) is (x
0
, . . . , x
m1
) :(M, I, h) [= [a
0
, . . . , a
m1
],
a basic formula in L
(Br)
th
n+1
r
( a, B) = th
n
r
( ac), B) :c M I.
2) If is a vocabulary of systems, n, m N, then TH
n
r
(m, ) is the set of
formally possible th
n
r
( a, B) for B a system, a B and lg( a) = m; this is
dened naturally. Pedantically, we dene it by induction on n; for n = 0 it is the
family of sets t of basic formulas (x
0
, . . . , x
m1
) of L
r
such that for each atomic
(x
0
, . . . , x
m1
) exactly one of (x
0
, . . . , x
m1
), (x
0
, . . . , x
m1
) belongs to t.
TH
n+1
r
(m, ) is the family of subsets of TH
n
r
(m + 1, ).
3) If is a vocabulary of models (see 2.2(4)), n, m N then TH
n
(m, ) is the set
of formally possible th
n
0
( a, B), B a model, i.e. B = B
sim
[M] for some model
M (note: the value of r is immaterial as the distance function is trivial).
4) If r = 0 we may omit it, so for a model M, using I = M, h the identity we
get the usual th
n
( a, M) (but we do not assume knowledge about it).
5) If a is empty sequence we may omit it.
2.7 Claim: 1) For B, n, m, a as above, = (x
0
, . . . , x
m1
) a (rst order) for
mula in L
(Br)
of quantier depth n, we have:
from th
n
r
( a, B) we can compute the truth value of B
r
[= [ a].
Here and in later instance we mean:
for any t TH
n
r
(m, ) we can compute a truth value t such that: if t = th
n
r
( a, B)
= a
m
we can com
pute th
n
r
(
i
) = t
for = 1, 2 and for i < max n
1
, n
2
, a vocabulary, M
i
K
then
th
d
(),
i<n
1
M
1
i
) = th
d
(),
i<n
2
M
2
i
). If M
i
K
i<k
M
1
i
) = th
d
(),
i<k
M
2
i
).
7) For a given vocabulary , and d N there is an operation on TH
d
(0, )
such that th
d
(),
i<k
M
i
) = th
d
(), M
i
) : i < k), this operation is associative
(but in general not commutative).
Proof: 1) We prove this by induction on the formula. (It goes without saying that
the reasoning below does not depend on B.)
atomic: Thus n = d() = 0, and by the Denition of th
n
r
( a, B) the statement
is trivial.
= : Easy by the induction hypothesis.
=
1
2
(or
1
2
, or
1
2
): Easy by the induction hypothesis.
= (x)
1
: Without loss of generality = (x
m
)
1
(x
0
, . . . , x
m1
, x
m
). So
d(
1
) = n 1, and by the induction hypothesis for a
0
, . . . , a
m
B we have: the
truth value of B [=
1
[a
0
, . . . , a
m1
, a
m
] is computable from th
n1
r
(a
0
, . . . , a
m
), B).
Say it holds i th
n1
r
(a
0
, . . . , a
m
), B) T
1
(T
1
a subset of TH
n1
r
(m +
1, (B))).
Now B [= [a
0
, . . . , a
m1
] i for some a
m
, B [=
1
[a
0
, . . . , a
m1
, a
m
], i
th
n1
r
(a
0
, . . . , a
m1
, a
m
), B) T
1
(T
1
the subset of TH
n1
r
(m+1, (B)) from
above) for some a
m
B i for some c B, th
n1
r
(a
0
, . . . , a
m1
)c), B) T
1
i th
n
r
(a
0
, . . . , a
m1
), B) is not disjoint to T
1
( the rst i by the denition
of satisfaction, the second i by the choice of T
1
, the third i is trivial; the
last i by the induction step in the denition of th
n
r
). So we have completed
9
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
10 01 Law
the induction.
2) We dene =
t
for t TH
n
r
(m, ) as required by induction on n; check the
inductive denition of th
n
r
.
3) Read the denition. (2.6(2))
4) By induction on n, for n = 0 as th
n
r
speaks on more basic formulas. For
n + 1 using the induction hypothesis (and the denition of th
n+1
r
).
5) We prove it by induction on n. The step from n to n + 1 is very straight
forward. For n = 0, we prove the statement by induction on r. For r = 0
note th
n
2
r ( a, B) = th
0
2
0
( a, B) = th
0
1
( a, B) so there is nothing to prove. For
r = r(0) + 1 just note that for s
0
2
r
we have: d(x, y) s
0
is equivalent
to
_
s
1
+s
2
=s
0
s
1
,s
2
2
r(0)
(z)[d(x, z) s
1
&d(z, y) s
2
]
6) The rst phrase is a special case of the second (with M
i
a model with a single
element: i). Let n()
def
= n
i<n()
M
i
and 0 i
(1) < i
(k
1) < n(),
we stipulate i
(0) = 1 and i
(k
) = n() (possibly k
= 1), assume
further a
k
M
(k)
has length m(k) for k < k
1 we have th
d
( a
1
k
, M
1
i
1
(k)
) = th
d
( a
2
k
, M
2
i
2
(k)
) and (i
1
(k + 1)
i
1
(k) 1) and (i
2
(k + 1) i
2
(k) 1) are equal or both 2
d
1. Lastly
assume th
d
(M
1
i
) = th
d
(M
2
j
) at least when (m)[i (i
1
(m), i
1
(m + 1))&j
(i
2
(m), i
2
(m + 1)] (holds automatically when proving second phrase of (6)).
Then
th
d
( a
1
,
i<n(1)
M
1
i
) = th( a
2
,
i<n(2)
M
2
i
) where a
= a
0
a
1
a
2
. . . a
1
.
The proof is straightforward, and for the case k
= n() + 1, i
i
= ).
7) The proof is like that of (6), but in ()
d
we add
_
k
1
k=1
i
1
(k) = i
2
(k).
Remark: If we will want to quantify on sequence of elements (i.e. use c rather
than c) this helps.
2.7
10
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 11
2.7A Claim: Let B be a system, let a
i
: i < m) and r
i
: i < m) (where
a
i
B and r
i
N) f
n
: N N for n N (not necessarily as in 2.5) be given
then for some n
i
: i w) where w m (= 0, . . . , m1) and function g from
0, 1, . . . , m1 to w which is the identity on w we have:
iw
n
i
m[w[ and
the sets in N
+
fn
i
(
{r
j
:g(j)=i})
(a
i
) : i w) are pairwise disjoint and a is included
in their union provided that
() 2f
n
1
(r
1
) + f
n
2
(r
2
) f
n
1
+n
2
+1
(r
1
+ r
2
) and f non decreasing in r and in n
(considering f a two place function from N to N).
Proof: We call w, n, g) with n = n
i
: i w) candidate if it satises all the
requirements in the conclusion of 2.7A except possibly the pairwise disjoint.
Clearly there is a candidate: w = 0, . . . , m1,
_
iw
n
i
= 0, g the identity on
w. So there is a candidate w, n, g) with [w[ minimal. If the disjointness demand
holds, then we are done. So assume i(1) ,= i(2) are in w and there is x belonging
to N
+
fn
i(1)
(
{r
j
:g(j)=i(1)})
(a
i(1)
) and to N
+
fn
i(2)
(
{r
j
:g(j)=i(2)}
(a
i(2)
).
Let w
= w i(2), g
i
for i w
: n
i
= n
i
if i ,= i(1) and n
i
= n
i(1)
+ n
i(2)
+ 1 if i = i(1) and let
n
= n
i
: i w
, m
, g
) is a candidate thus
nishing the proof: for this we have to check the two relevant conditions. First
iw
i
=
iw
i=i(1)
n
i
+ n
i(1)
=
iw
i=i(1)
n
i
+ n
i(1)
+ n
i(2)
+ 1 =
iw
n
i
+ 1 m[w[ + 1 = m([w[ 1) = m[w
[.
Secondly, why
i
N
+
f
n
i
(
{r
j
:g
(j)=i})
(a
i
) includes a? if j() < m then for some
i w we have a
j()
N
+
fn
i
(
{r
j
:g(j)=i})
(a
i
); if i ,= i(1), i(2) then i w
, and
N
+
f
n
i
(
{r
j
:g
(j)=i})
(a
i
) = N
+
fn
i
(
{r
j
:g(j)=i})
(a
i
)
so we are done; if i = i(1) then n
i
= n
i(1)
+ n
i(2)
+ 1 n
i(1)
= n
i
and
r
j
: g
(j) = i
r
j
: g(j) = i
hence
N
+
f
n
i
(
{r
j
:g
(j)=i})
(a
i
) N
+
fn
i
(
{r
j
:g(j)=i})
(a
i
)
and we are done.
11
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
12 01 Law
We are left with the case i = i(2), so by the choice of i we have d(a
j()
, a
i(2)
)
f
n
i(2)
(
r
j
: g(j) = i(2)) and by the choice of x (and i(1), i(2)) above
d(a
i(2)
, x) f
n
i(2)
(
r
j
: g(j) = i(2)) and d(x, a
i(1)
) f
n
i(1)
(
r
j
:
g(j) = i(1)). So as d is a metric (i.e. the triangular inequality) d(a
j()
, a
i(1)
)
2f
n
i(2)
(
r
j
: g(j) = i(2)) + f
n
i(1)
(
r
j
: g(j) = i(1)). Now
r
j
: g
(j) =
i(1) =
r
j
: g(j) = i(1) +
r
j
: g(j) = i(2) (by the denition of g
) and
n
i(1)
= n
i(1)
+ n
i(2)
+ 1, hence what we need is
() 2f
n
1(r
1
) + f
n
2(r
2
) f
n
1
+n
2
+1
(r
1
+ r
2
)
which is assumed.
2.7A
2.7B Remark: 1) We can replace in 2.7A and its proof
r
j
: g(j) = i by
max r
j
: g(j) = i and (in ()) r
1
+ r
2
by max r
1
, r
2
.
(2) Concerning (), letting n = max n
1
, n
2
and r = max r
1
, r
2
it suces to
have
()
1
f
n
(r) is non decreasing in n and in r.
()
2
3f
n
(r) f
n+1
(r)
2.8 Denition: 1) For a system B, and r, n, m N, and a = a
0
, . . . , a
m1
) B
we call a a (B, r)component if a N
+
r
(a
0
). In this case we dene bth
n
r
( a, B)
(bth is for bounded theory). We do it by induction on n (for all r) (the function
f from 2.5 is an implicit parameter.)
() bth
0
r
( a, B) = th
0
( a, B
r
)
() bth
n+1
r
( a, B) is t
0
, t
1
, t
2
) where
(i) t
0
= bth
n
r
( a, B)
(ii) t
1
= bth
n
r
(c, B) : c N
+
f
(2)
n
(r)
(a
0
) N
+
fn(r)
(a
0
)
(iii) t
2
= bth
n
f
(2)
n
(r)
( ac), B) : c N
+
fn(r)
(a
0
)
2) If is a vocabulary of systems and r, n, m N then BTH
n
r
(m, ) is the set of
formally possible bth
n
r
( a, B), (B a system, a = a
0
, . . . , a
m1
) B and a is a
(B, r)component).
2.9 Claim: 1) For any (vocabulary of systems), numbers n, m, r N, and t
BTH
n
r
(m, ), there is = (x
0
, . . . , x
m1
) L
(even L
(B
fn(r)
)
) of quantier
depth n such that for any system B, and a = a
0
, . . . , a
m1
) B a (B, r)
component we have:
t = bth
n
r
( a, B) i B
fn(r)
N
+
fn(r)
(a
0
) [= [ a].
12
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 13
2) If n m,
b is the permutation of a by the function h or
b = a
: < k)
for some k g( a) and b
0
= a
0
then from t = bth
n
r
( a, B) we can compute
bth
m
r
(
t
= a : a is a (B, r
)component , t =
bth
n
r
( a, B), [ = 1 a M] and [ = 2 a I] for each t BTH
n
r
(m
, )
m
m + n, r
f
(3)
n
(r) and 1, 2. We let I
n
r,m
[B] = B
n
r,m
I. Writing
r = r
,m
r.
2.11 Claim: I
n+1
r, m
[B] essentially expands I
n
r
, m
[B] when
(ii) r
= r0), m
= m1) or
(ii) r
f
(2)
n
(r), m
m
i
+ 1 : i < lg( m))
(essentially expand means that every predicate in the latter is equivalent to a
quantier free formula in the former, the function giving this is the scheme of
expansion).
Proof: Should be clear by (i) of () of 2.8(1).
2.12 Denition: For a system B and n, m N and r = r
N and a = a
0
, . . . , a
k1
) I
B
1) We say a is (n, r)sparse for B if r = r
: < m) and
< k < m implies N
+
fn(r
)
(a
) N
+
fn(r
k
)
(a
k
) =
moreover (slightly stronger)
d(a
, a
k
) f
n
(r
) + f
n
(r
k
) + 1.
13
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
14 01 Law
2) We dene uth
n
r
( a, B) for an (n, r)sparse a B, by induction on n:
uth
0
r
( a, B) = th
0
( a, B) and uth
n+1
r
( a, B) = t
0
, t
1
) where:
t
0
= s, uth
n
s
( a, B)) : s f
(2)
n
(r)
(see 2.5, f
(2)
n
(r) = f
n
(f
n
(r
) and c I
B
and
ac) is (n, s0))sparse i.e.
N
+
fn(0)
(c) is disjoint from N
+
fn(s
)
(a
n
( r)
where f
n
(r) = min m : if
1
<
2
< g( a) then m > f
n
(r
1
) + f
n
(r
2
)
of quantier depth n such that: For a system B and (n, r)sparse a =
a
0
, . . . , a
m1
) I
B
, we have: B [= [a
0
, . . . , a
m1
] i t = uth
n
r
( a, B) (and
being (n, r)sparse is equivalent to some quantier free formula).
2) In Denition 2.12 only BI matters and in part (1), the quantications are
on I only.
3) If B
n
there are functions F
n,r, m
, for r = r
: < k), m = m
, m
N, such that:
() if B is a system, a = a
0
. . . a
k1
, a
an (B, r
)component, a
0
: <
k) is (n, r)sparse and m = g( a
def
= bth
n
r
( a
, B)
(for < k) and t
def
= uth
n
r
_
h(a
0
) : < k), I
n
r, m
[B]
_
we have th
n
0
( a, B) =
F
n,r, m
(t, t
0
, . . . , t
k1
).
() F
n,r, m
is recursive in its variables, n, r, m and the functions F
0,r
, m
where
m
= m
i
: i < k
), k
k, r
n
(r) (see 2.13(1)) and for i < k we have
m
i
m
i
and
i<k
(m
i
m
i
) +
1
i=k
m
i
n.
14
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 15
2.14A Remark: Why we use th
n
0
and not th
n
r
(in the conclusion of ())? We can
if we assume d(x, y) s &x I &y I is an atomic formula for B for s r.
Proof: We prove this by induction on n, (for all r, m), so for n = 0 clearly
n
holds. So assume
n
and we shall prove
n+1
. We shall now describe a value t
computed from bth
n+1
r
( a
0
) : < k), I
n+1
r, m
[B]).
Our intention is that t = th
n+1
( a, B). Remember t = th
n
( ac), B) : c B.
Now t will be the union of two sets.
We use an informal description as it is clearer.
First Part: The set of th
n
( ac), B), where for some () < k, c N
+
f
(2)
n
(r
()
)
(a
()
0
).
Why can we compute this (using the induction hypothesis of course), for each
such c we have: let
b
= a
is: r
for ,= () and f
(2)
n
(r
()
) if = (). Necessarily for
< k, m
i
is m
if ,= () and m
()
+ 1 if = (). Now:
() N
+
fn(r
)
(b
0
) for < k are pairwise disjoint, moreover for (1) < (2) < k,
d(b
(1)
0
, b
(2)
0
) > f
n
(r
(1)
) + f
n
(r
(2)
).
[Why? For = () remember that for every r: f
n
(f
(2)
n
(r)) = f
(3)
n
(r))
f
n+1
(r) and for ,= () remember that for every r: f
n
(r) f
n+1
(r) so
in all cases clearly f
n
(r
) f
n+1
(r
) and b
0
= a
0
. So if (1) ,= (2) are
< k, N
+
fn(r
(1)
)
(a
(1)
0
) N
+
fn(r
(2)
)
(b
(2)
0
) = as: if (1) ,= (), (2) ,= (),
trivial. Otherwise without loss of generality (2) = (), as a satises the
assumption of () (() is from the lemma) a is (n + 1, r)sparse hence (see
Def. 2.12(1)) d(a
(1)
0
, a
()
0
) > f
n+1
(r
(1)
) + f
n+1
(r
(2)
), hence d(b
(1)
0
, b
()
0
) =
d(a
(1)
0
, a
()
0
) > f
n+1
(r
(1)
) + f
n+1
(r
(2)
) f
n
(r
(1)
) + f
n
(f
(2)
n
(r
()
)) =
f
n
(r
(1)
) + f
n
(r
()
)), as required.]
()
b
N
+
r
(b
0
), so
b
is an r
component.
[Why? when = () as r
f
(2)
n
(r
) = r
, B) for ,= ()
[Why? by monotonicity properties of bth i.e. by 2.9(2) (that is clause (i) i.e.
t
0
from Def. 2.8(1)())];
() We can compute the set of possibilities of bth
n
r
()
(
b
()
, B).
[Why? those possibilities are listed in bth
n+1
(see t
2
in Denition 2.8(1)()
mainly clause(ii)).]
15
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
16 01 Law
() We can compute uth
n
r
(h(b
0
) : < k), I
n
r, m
[B]).
[As we can compute uth
n
s
(h(b
0
) : < k), I
n+1
r, m
[B]) for s f
(2)
n
(r) by
the denition of t
0
in Def. 2.12(2), choose s = r
; now, by 2.11, I
n+1
r, m
[B]
essentially expand I
n
r
, m
b
0
. . .
b
k1
, B) gotten as
above for xed () < k, all c N
+
f
(2)
n
(r
()
)
(a
()
0
).
[Why? by ()  () above and
n
.]
() We can compute th
n
( ac, B) : for some () < k, c N
+
f
(2)
n
(r
()
)
(a
()
0
)
Second Part: The set of th
n
( ac), B) where for each < k, c / N
+
f
(2)
n
(r
)+fn(0)
(a
0
).
Why can we compute this? for such a c we can let k
= k + 1, a
k
= c) (so
c = a
k
0
, m
k
= 1). Let r
= r
k
= 0. Let m
be m1). Now
() N
+
fn(r
)
(a
0
) for k are pairwise disjoint and d(a
(1)
0
, a
(2)
0
) > f
n
(r
(1)
) +
f
n
(r
(2)
).
[Why? as f
n
(r) f
n+1
(r) for (1) < (2) < k this is trivial, and for
(1) = < k = (2) we have d(a
(1)
0
, a
(2)
0
) = d(a
0
, a
k
0
) = d(a
0
, c) > f
(2)
n
(r
)
f
n
(r
) + f
n
(0) = f
n
(r
) + f
n
(r
k
), as required].
() a
N
+
r
(a
0
), so a
is an r
component.
[Why? for < k as r
= r
, for = k as a
k
= a
k
0
)(= c))].
() we can compute bth
n
r
( a
, B) for < k
[Why? by monotonicity properties of bth i.e. 2.9(2)].
() we can compute the possibilities for pairs (t
, t
) where t
= bth
n+1
0
(c), B)
and t
= uth
n
r
(h(a
0
) : < k)h(c)), I
n+1
r, m
[B]).
[Why? straightforward; by the denition of uth
n+1
, i.e. t
1
of Def. 2.12(2) and
I
n
r, m
[B]].
() in () we can replace I
n+1
r, m
[B] by I
n
r
, m
[B].
[Why? by 2.11 and 2.13(3).]
By the induction hypothesis this is enough.
Why the union of the two parts is th
n
( a, B)?
Both obviously give subsets, and if c fails the rst part then < k c /
N
+
f
(2)
n
(r
)
(a
0
). So N
+
fn(0)
(c) is disjoint to such N
+
fn(r
)
(a
0
) and moreover d(a
0
, c) >
f
n
(r
) + f
n
(0).
2.14
16
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 17
2.15 Conclusion: For any system vocabulary , and (rst order) sentence
of quantier depth n, given F
0,r, m
s satisfying
0
of 2.14 we can compute
numbers r, m and a sentence
.
Proof: By 2.14 and 2.13(1) and 2.7(1) (and see 2.14A(2)).
2.16 Conclusion: Let be a vocabulary (nite for simplicity) including a binary
relation < and a unary relation P, and L
1
with
d(
) d() where
1
def
=< P
i
: i < m (m N computable from , each
P
i
a unary predicate) satisfying the following:
() Assume M is a nite model, <
M
a linear order, P is unary, such that:
if R < and a = a
0
, . . . , a
n(R)1
) R
M
, then P
M
[ min a, max a]
M
has at most one member.
Dene I[M] = (<
M
, P
M
, . . . , P
. . .)
<m
where P
=
_
a P
M
:
[a] is
satised in Mx : x a, [[x, a]
M
P
M
[ 3
d[]
or a x, [[a, x]
M
P
M
[
3
d[]
_
(d[] in the quantier depth of ).
Then M [= I[M] [=
.
Proof: Should be clear from 2.15.
Remark: Concerning 2.16 we can deduce it also from 4.
3 Proof of the Main Theorem
3.1 Denition: 1) For a nite set J I let spr(I, J) be the set of pairs
(Q
no
, Q
yes
), where Q
no
I, Q
yes
I, Q
no
J = Q
yes
J, [Q
yes
Q
no
[ = 1 (AB
is the symmetric dierence). Let spr(I) = spr(I, I).
2) For nite J I let
(I) for
(I, I).
3.1A Remark: Note that the distribution
(k) on
spr(k) = (Q
no
, Q
yes
) : Q
no
, Q
yes
k, [Q
yes
Q
no
[ = 1 we have: if M
t
i
K, for
i < k, t yes no, and we choose
(k)randomly (Q
no
, Q
yes
) spr(k), then
() the probability of th
d
(
i<k
M
if (iQ
yes
)
i
) = th
d
(
i<k
M
if (iQ
no
)
i
) is at least
1 (th
d
is dened as in 2.6(1) considering a model as a system by 2.2(4)).
2) Also if we rst choose (Q
no
u
, Q
yes
u
) spr
_
[[
k+1
2
], k)
_
as above and then (possibly
depending on the result) make a decision on a choice of Q
no
d
= Q
yes
d
[0, [
k+1
2
])
and let Q
yes
= Q
yes
d
Q
yes
u
, Q
no
= Q
no
d
Q
no
u
then () above still holds.
3) If we choose Q
yes
k such that [Q
yes
[ = [
k+1
2
] and then Q
no
Q
yes
, [Q
no
[+1 =
[Q
yes
[ (all possibilities with the same probabilities) then with probability tending
to 1 with k going to we get that () (of 3.3(1) above) holds.
4) The parallel of (1) holds for
(k), npr(k).
Before proving 3.3 we dene and note:
3.4Denition: For the , K, d from 3.3, let
k
be the maximal real in [0, 1] such
that: if M
t
i
K for i < r, t yes, no (where r N), J i < r : th
d
(M
no
i
) ,=
th
d
(M
yes
i
) has k elements (and J I = r = 0, . . . , r 1), then
k
r
(1
k
) Prob
(I,J)
_
th
d
(
i<r
M
if (iQ
no
)
i
) ,= th
d
(
i<r
M
if (iQ
yes
)
i
)
(Q
no
, Q
yes
)
spr(I, J)
_
. Let
k
= 1
k
.
3.4A Observation: 1)
k
is well dened; and
k
k+1
2) An alternative denition of
k
is that it is the maximum real in [0, 1] satisfying:
(1
k
) is not smaller than the relativeprobability of th
d
_
i<r
M
if (iQ
no
)
i
_
,=
18
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 19
th
d
_
i<r
M
if(iQ
yes
)
i
_
for the probability distribution
i<r
M
if(iQ
no
)
i
) ,= th
d
(
i<r
M
if(iQ
yes
)
i
)
(Q
no
, Q
yes
) spr(I, J)
_
=
1
()
<()
2
(r1)
[Q Ii
: th
d
(
i<r
M
if(iQ)
i
) ,= th
d
(
i<r
M
if(iQ{i
})
i
)[.
Hence if () = k + 1 then
()
2
Prob
(I,J)
_
th
d
(
i<r
M
if(iQ
no
)
i
) ,= th
d
(
i<r
(M
if(iQ
yes
)
i
)
(Q
no
, Q
yes
) spr(I, J)
_
=
1
k + 1
k
Prob
(I,J
k+1,
)
_
c
(Q
no
, Q
yes
) spr(I, J
)
_
,
where
c
= (Q
no
, Q
yes
) spr(I, J) : th
d
(
i<r
M
if(iQ
no
)
i
) = th
d
(
i<r
M
if(iQ
yes
)
i
).
Now compute.
2) Should be clear.
3) By addition theory i.e. 2.7(7) and by 2).
3.4A
3.5 Denition: Let c N be the number of members in TH
d
(0, ) (the set of
formally possible th
d
(), M), M K, see 2.6(3)). Let k
0
N be such that
k
0
(3
d+8
)
2
c
2
(exists by Ramsey theorem).
3.6 Observation:
()
1
k
0
1
(k
0
2
k
0)
.
[Why? Let r, M
t
i
, and J, [J[ = k
0
be given. First draw Q
yes,no
(r J) and
19
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
20 01 Law
assume they are equal.
Now (by 3.4A(2)) it is enough to prove that now the probability of the equality
i.e. of
th
d
_
i<r
M
if(iQ
no
)
i
_
= th
d
_
i<r
M
if(iQ
yes
)
i
)
_
is 1/(k
0
2
k
0
), assuming Q
no
Q
yes
= j J. For i < j from J let t
no
i,j
, t
yes
i,j
) be
th
d
(M
no
i
+M
no
i+1
+. . . +M
no
j1
), th
d
(M
yes
i
+M
no
i+1
+M
no
i+2
+. . . +M
no
j1
)). So by
the choice of k
0
we can nd J
J, [J
[ = 3
d+8
and t
0
, t
1
) such that for i < j
in J
, we have t
no
i,j
, t
yes
i,j
) = t
0
, t
1
). Let J
= i
: < 3
d+8
. For each j 3
d+8
let Q
i
= i
m
: m < j. By addition theory for (rst order theory) linear order,
(that is 2.7(6)) for
def
= [3
d+8
/2], th
d
(
i
M
if (iQ
)
i
) = th
d
(
i
M
if (iQ
+1
)
i
). So
the probability for equality is at least the probability of Q
no
= Q
, Q
yes
= Q
+1
which is 1/(k
0
2
k
0
)].
3.6
3.7 Observation: For every , k > 0 we have
k
k
(
j1
[
_
1
j
_
j
k
(1
k
)
1j
]
j
).
Proof: Let us be given r, M
t
i
for i < r and J as in () so [J[ = k. Choose
I
j
: j < ) a partition of I
def
= r = 0, . . . , r 1 to intervals such that for j < ,
J
j
def
= J I
j
has exactly k members. Now rst draw Q
no
j
I
j
for j < (with equal
probabilities), second draw s
j
J
j
for j < (with equal probabilities) and third
draw R
no
0, . . . , 1 (with equal probabilities) and fourth draw j() <
(with equal probabilities). Let Q
yes
j
= Q
no
j
s
j
, and R
yes
= R
no
j(), and
Q
no
=
j<
Q
if(R
no
)
j
, Q
yes
=
j<
Q
if(R
yes
)
j
.
Easily (Q
no
j
, Q
yes
j
) was chosen by the distribution
(I
j
, J
j
) and (R
no
, R
yes
)
was chosen by the distribution
(0, . . . , 1) and (Q
no
, Q
yes
) was chosen by
the distribution
i<r
M
if(iQ
no
)
) ,= th
d
(
i<r
M
if(iQ
yes
)
i
)
_
k
(
j1
[
_
(1)
j
_
j
k
(1
k
)
1j
]
j
).
For j < let N
t
i
=
iI
j
M
if (iQ
t
j
)
and let p
i
[0, 1]
R
be Prob
_
th
d
(N
no
j
) ,=
th
d
(N
yes
j
)
_
and let A = j : th
d
(N
no
j
) ,= th
d
(N
yes
j
), so the events j A are
independent and p
i
= Prob(i A)
k
. Now if we make the rst and second
20
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 21
drawing only, we know A, N
t
i
: i < , t no, yes) and modulo this, by the
denition of
A
we know
Prob
_
th
d
(
i<r
M
if (iQ
no
)
i
) ,= th
d
(
i<r
M
if (iQ
yes
)
i
)
j<
N
if(jR
no
)
j
) ,= th
d
(
j<
N
if (jR
yes
)
j
)
(I,J)
_
th
d
(
i<r
M
if(iQ
no
)
i
) ,= th
d
(
i<r
M
if (iQ
yes
)
i
)
(Q
no
, Q
yes
) spr(I, J)
_
j
Prob([A[ = j)
j
j
=
j
_
u,u=j
mu
p
m
m<,m/ u
(1 p
m
)
_
j
.
Now looking at this as a function in p
m
[0,
k
]
R
for m < , for some p
m
: m < )
we get maximal values, and as the function is linear, p
m
0,
k
, and as
0
j
j+1
1, necessarily p
m
=
k
so
Prob
(I,J)
_
th
d
(
i<r
M
if(iQ
no
)
i
) ,= th
d
(
i<r
M
if (iQ
yes
)
i
)
(Q
no
, Q
yes
) spr(I, J)
_
j
(
u,u=1
mu
p
m<,m/ u
(1 p
m
))
j
j
=
=
j
_
n
j
_
(
k
)
j
(1
k
)
j
(
j
)
j
=
=
0<j
_
j
_
(
k
)
j
(1
k
)
j
j
j
=
=
j1
[
_
j + 1
_
(j + 1)
(
k
)
j+1
(1
k
)
j1
]
j
=
j1
[
_
1
j
_
(
k
)
j
(1
k
)
(1)j
]
j
.
3.7
3.8 Observation.
1) If , k > 0, j
0
k
then
k
k
(
1+
j
0
2
).
2) If
k
1
1
m
, > k/
k
then
k
k
(
1+
k
2
).
21
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
22 01 Law
Proof: 1) As
j
j+1
we have
k
k
(
j1
[
_
1
j
_
(
k
)
j
(1
k
)
1j
]
j
)
k
_
j<j
0
[
_
1
j
_
(
k
)
j
(1
k
)
1j
] 1+
j[j
0
,1)
[
( 1)!
j!( j 1)!
(
k
)
j
(1
k
)
1j
]
j
0
_
k
(
1 +
j
0
2
).
2) Follows.
3.8
3.8A Remark. Using the binomial distribution approach normal distribution
and 3.6, clearly we get e.g.:
for every > 0, for some
, for every
we have
k
k
(1)
.
3.9 Proof of 3.3: 1) By the denition of
k
and Observations 3.5, 3.6 we get that
lim
k
k
= 1 and we can nish easily.
2) Follows by (1) (and the addition theory see 2.7, particularly 2.7(7))
3) Similar proof and not used (e.g. imitate the proof of 3.6. First choose j() then
we have probability
(I
j()
, J
j()
), but
the induced distribution is very similar to it).
4) Follows very easily. For spr(k) with probability 1/2 we are choosing by npr(k).
3.3
3.10 Proof of 1.4: Let a real > 0 and a sentence L
fo
1
(a vocabulary) and
m N and
i
(x) for i < m be dened as in 2.16 (for the here). Let d be the
quantier depth of
(i.e. d = d[
]). Let k
1
). Let k be (2k
+ 2)(3
d[]
+ 1) (we could have waived
the 3
d[]
+ 1). Now choose by induction on r k, m
r
N such that
() (a) 0 = m
0
(b) m
r
< m
r+1
< . . .
(c) for any n,
22
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 23
/3 > Prob
n
_
M
n
[=
r
(x m
r
+ 1)(y m
r+1
1)[xRy]
M
n
K
n
_
.
[Why this is possible? We choose m
r
by induction on r. The probability in
question is, for each xed r, bounded from above by
i<mr
j>m
r+1
p
ji
, the
sum is the tail of an (absolutely) convergent innite sum so by increasing m
r+1
we can make it < /2
r+2
, this suces].
Next we try to draw the model M
n
in another way. Let n > m
k
be given; let
n
= n+k
+1. Let J = m
(3
d[]+1
)i
: 0 < i < 2k
, and I = m
i
: i < 2k
3
d[]+1
.
We rst draw A
n
, a drunkard model A
n
for n. Drawing A
n
means:
laziness case=rst case if i < j < n
,
_
r<k
[i m
r
+1&j m
r+1
1] then i, j
is non edge (no drawing).
normal case=second case: if i < j < n
1 Q (so M
Q
[A
n
] is n
or
k
+ 1)
< is the usual order on I
Q
P = m
r
: r < k
R
Q
is
(i, j) : i, j = i
, j
, i
< j
< n
, and:
(a) (i
, j
,j
= yes
or (b) (i
, j
,j
= yes
or (c) (i
, j
,j
= yes.
We also dene a model N[A
n
]:
the universe: 0, . . . , n
1
relations: < the usual order
R = i, j) : e
i,j
= yes, i < j
< j
.
23
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
24 01 Law
R
1
= i, j) : e
1
i,j
= yes, i < j
R
2
= i, j) : e
2
i,j
= yes, i < j
P = m
r
: r k (on see above, before ())
Observe
()
1
in (N[A
n
], Q) we can dene M
Q
[A
n
] by q.f. formulas.
So for some rst order depending on , but not on n (promised above in
the beginning of the proof):
()
2
M
Q
[A
n
] [= i (N[A
n
], Q) [= .
By 2.16 (where I[(N[A
n
], Q)] is dened in 2.16 with M there standing for
(N[A
n
], Q) here, so its set of elements is P) and the choice of
we have:
()
3
(N[A
n
], Q) [= i I[(N[A
n
], Q)] [=
.
Looking at the denition of I[N[A
n
], Q] in 2.16 without loss of generality
()
4
I[(N[A
n
], Q)] [=
i (I[N[A
n
]], Q) [=
.
Let J = J
d
J
u
, J
d
an initial segment, J
u
an end segment, [J
u
[ = k
, [J
d
[ =
k
[J, J
u
] be the distribution from
3.2 above on npr(J
u
), and choose (Q
u
0
, Q
u
1
) npr(J
u
) randomly by
(J
u
) then
choose Q
d
1
J
d
such that [Q
d
1
[ = k
+ 1 [Q
u
1
[ with equal probabilities, and let
Q
1
= Q
d
1
Q
u
1
, and let Q
d
0
= Q
d
1
, Q
0
= Q
d
0
Q
u
0
.
Now for 0, 1 we make a further drawing: if i < j is a pair from the
rst possibility (in the drawing of A
n
), we ip a coin for
e
i,j
yes, no with
probability p
[i,j)\Q

for yes.
Let M
be (A
Q
, <, R
Q
i,j
= yes) (it depends
on the choice of A
n
and on the further drawing).
Now reecting we see
()
5
for = 0, 1, the distribution of M
[= ) = Prob(M
n+1
[=
M
n+1
K
n+1
).
By the choice of m
r
s
()
7
Prob(M
= M
Q
[A
n
]) is 1 /3.
By 3.3 above (used above: the drawing of (Q
u
0
, Q
u
1
) was randomly by
(J
u
)).
()
8
the absolute value of the dierences between the following is /3:
Prob([N[A
n
]], Q
0
) [=
)
Prob([N[A
n
]], Q
1
) [=
).
So for = 0, 1:
(a) Prob(M
n+
[=
M
n+
K
n+
) = Prob(M
[= )
[Why? by ()
6
.]
24
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 25
(b) Prob(M
= M
Q
[A
n
]) 1 /3
[Why? by ()
7
.]
(c) M
Q
[A
n
] [= i (I[N[A
n
]], Q
) [=
[Why? by ()
1
+ ()
2
+ ()
3
+ ()
4
.]
By (a)+(b)+(c) it suces to prove that the probabilities of
(I[N[A
n
]], Q
) [=
tI
[M
t
[ (if the M
t
s are not pairwise disjoint:
(t, a) : t I, a M
t
) and we dene h
: M I, h(a) = t if a M
t
(if not
disjoint h(t, a)) = t),
(b) if t
1
, . . . , t
k
I are pairwise distinct, a = a
1
. . . a
k
and a
M
t
(nite
sequences) then
tp
qf
( a, , M) = F
_
tp
qf
(t
1
, . . . , t
k
), , I), tp
qf
( a
1
, , M
t
1
), . . . , tp
qf
( a
k
, , M
t
k
)
_
.
Another way to say it is:
(b)
if a
0
, . . . , a
k1
M then tp
qf
(a
0
, . . . , a
k1
), , M) =
F
_
tp
qf
(h(a
) : < k)), . . . , tp
qf
(a
: < k, t = h(a
)), , M
t
), . . .
_
t{h(a
):<k}
.
25
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
26 01 Law
4.2 Remark: 1) So the form of F is implicitly dened, it suces to look at
sequences a
0
. . . a
k
(in clause (b)) or a
0
, . . . , a
k
) of length the arity of
2
(if
it is nite) i.e. maximum numbers of places for P predicates P
2
.
2) We can consider a generalization where the universe of M and equality are
dened like any other relation.
4.3 The generalized Sum Theorem: In the notation above if
n
is the set
of formulas of quantier depth n then we can compute from F the following
function: like F in (b) (or (b)
) replacing tp
qf
by tp
n
.
4.4 Discussion: So looking at a sequence a from M, to nd its [quantier free]
types we need to know two things:
() the [quantier free] type of its restriction to each h
1
(t) = b M : h(b) =
t in M
t
() the [quantier free] type of the sequence of parts, h(b
) : < k) in I.
4.5 Denition: Now M is a ddistorted Fsum of M
t
: t I if
(a) d is a distance function on I.
(b) [M[ is the disjoint union of A
t
(t I) and for each t we have a model
M
t
with universe A
s
: s I, d(s, t) 1 and
(c)
+
if b
0
, . . . , b
k1
M, then
tp
qf
(b
0
, . . . , b
k1
), , M) =
F
_
tp
qf
(h(b
0
) . . . , h(b
k1
), , I), . . . ,
tp
qf
(b
: < k, d(h(b
), t) 1), , M
t
), . . .
_
t{h(bm):m<k}
.
4.6 Remark: Note: by b
).
4.7 Discussion: Our main Lemma 2.14, generalizes the generalized sum theorem,
to distorted sum but naturally the distortion expands with the quantier
depth.
4.8 The Distorted Sum Generalized Lemma: In the notation above, for m let
M
m
t
be the model with
universe: A
s
: d(s, t) m s I : d(s, t) m
26
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 27
relations: those of the M
s
s i.e. for R
1
, a kplace predicate we let
Q
t,m
R
= s, a
1
, . . . , a
k
) : s I, d(s, t) m + 1, a
1
, . . . , a
k
) R
Ms
(so a
1
, . . . , a
k
A
s
: d(s, t) 1)
Q
t,,m
d
= (s
1
, s
2
) : d(s
1
, t) and d(s
2
, t) , d(s
1
, s
2
) m,
Q
t
h
= (s, a) : a A
s
, s I, d(s, t) m.
We dene I
[m,n]
as the expansion of I by:
for m and
n
(
0
)
Q
= t I : M
t
[= .
Now there are functions F
n
and a number m(n) = 3
n
computable from F (and
n) such that:
for b
0
, . . . , b
k1
M
tp
n
(b
0
, . . . , b
k1
), , M) = F
n
_
tp
n
(h(b
0
), . . . , h(b
k1
)), , I
[m(n),n]
), . . . ,
tp
n
(b
: < k, d(h(b
), t) m(n)), , M
m(n)
t
), . . .
_
t{h(bm):m<k}
.
4.9 Discussion: 1) Now if d is trivial:
d(x, y) =
_
0 x = y
x ,= y
then M
n
t
= M
t
, and 2.8 (the disorted generalized sum Lemma) becomes degen
erated to 4.3 (the generalized sum Lemma), more exactly a variant.
2) Note that 4.8 improve on the result of 2 in m(n) not depending on k. We
can have this improvement in 2 and in 5.
3) To prove 4.8 given b
0
, . . . , b
k1
and looking for y with
t = tp
n+1
(b
0
, . . . , b
k1
, y), , M),
we x w = < k : d(h(b
), h(y)) m(n).
E.g. if w ,= and () = min (w), then the relevant properties of y are expressed
in the balls
z M : d(h(z), h(b
)) m(n)
27
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
28 01 Law
for w and
z M : d(h(z), h(y)) m(n),
all included in the ball
z M : d(h(z), h(b
()
)) 3m(n).
The case w = is simpler.
5 Models with Distance
5.1 Discussion: We try here to explain the results of 2 as concerning a model
with a distance function weakly suitable for the model and the connection to
models with a distance function for the whole vocabulary which is suitable for
the model. This is another variant of the distorted sums.
5.2 Context: Let be a xed vocabulary.
1) Let K be the class of A = (M, d), M a model, d a distance on M (i.e. a
two place symmetric function from [M[ to , d(x, x) = 0, satisfying the
triangular inequality) and for simplicity d(x, y) = 0 x = y.
2) K
sut
K is the class of (M, d) K such that d (which is a distance on M)
is suitable for the model, i.e.
1
a
0
, . . . , a
k1
) R
M
, R
_
<m<k
d(a
, a
m
) 1.
3) K
sim
K is the class of (M, d) K which are simple, i.e
2
d(x, y) = Minn : there are z
0
, . . . , z
n
such that x = z
0
, z
n
= y and
_
<n
_
R
( a R
M
)[z
, z
+1
Rang a].
4) K
ws
F
is the class of (M, d) K which are Fweakly suitable which means:
3
for every m
i
if a
i
= a
i
0
, . . . , a
i
n
i
1
) for i < k, a
i
M, d(a
i
0
, a
i
) m
i
and i < j < k d(a
i
0
, a
j
0
) > m
i
+ m
j
+ 1 then the quantier free type
of a
0
a
1
. . . a
k1
is computed by F from the quantier free types of
a
0
, a
1
, . . . , a
k1
and of a
0
0
, a
1
0
, . . . , a
k1
0
).
Note: we can strengthen the demands e.g. (for f as in 2.5)
() d(a
i
0
, a
i
) r
i
, d(a
i
0
, a
j
0
) > f
0
(r
i
) + f
0
(r
j
) + 1 or at least (x, y)[d(x, a
i
)
f
0
(r
i
) d(y, a
j
) f
0
(r
j
) d(x, y) 1].
5) K
as
is the family of B = (M, d) K which are almost simple: d(x, y) 1 is
dened by quantier free formula and d(x, y) = Minn: we can nd z
0
, . . . , z
n
, x =
z
0
, z
n
= y, d(z
, z
+1
) 1.
28
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
01 Law 29
5.3 Discussion: For (M, d) K
ws
F
we want to separate the quantication to
bounded ones and to distant ones. We can either note that it ts the context of
2 or repeat it.
5.4 Denition: 1) For B = (M, d) K, x M, m < let N
+
m
(x) = y M :
d(y, x) m.
2) We dene a is a (B, r)component and bth
n
r
( a, B) as in Denition 2.8(1),
and BTH
n
r
(m, ) as in 2.8(2).
3) We dene B
2
n,m
as expanding B by the relation R
t
= a : a is a (B, r)
component, t = bth
n
r
( a, B) for t BTH
n
r
(m, ).
4) We dene a is (B, r)sparse and uth
n
r
( a, B), UTH
n
r
(m, ) as in 2.12.
5.5 Theorem: the parallel of 2.14.
References
[B] E.W. Beth, Observations metamathematiques sur les structures simplement
ordonnes, Collection de Logique Mathematique, Serie A, Fasc 5, ParisLouvain
1954, 2935.
[CK] Chang and Keisler, Model theory, North Holland Publishing Co, 1973,
1977.
[Gf] H. Gaifman, On local and non local properties, Proc. Herbrand Symp. Logic
Colloquium, 81, ed. J. Stern; Studies in Logic Found. of Math vol 107 North Hol
land (1982) 105135.
[Gu] Y. Gurevich, Monadic second order theories, Chap XIII, Model theoretic
logics, ed. Barwise and Feferman Springer Verlag (1985); 479506.
[Lu] T. Luczak, notes.
[LcSh 435] T. Luczak and S. Shelah, Convergence in homogeneous random
graphs, Random Structures & Algorithms, 1995, in press.
[ShSp 304] S. Shelah J. Spencer, On zeroone laws for random graph, Journal of
A.M.S vol. 1 (1988) 97115.
[Mo] A. Mostowski, On direct products of theories, Journal of Symbolic Logic,
vol 17 (1952) 131.
[Mr] L. Marcus, Minimal models of theories of one function symbol, Israel J. of
Math vol 18 (1974) 117131.
[Sh 42] S. Shelah, The monadic theory of order, Annals of Mathematics, vol 102
(1975), 379419.
[Sh F120] S. Shelah, More on 01 laws, unpublished notes.
29
(
4
6
3
)
r
e
v
i
s
i
o
n
:
2
0
0
8

1
0

3
0
m
o
d
i
f
i
e
d
:
2
0
0
8

1
0

3
0
30 01 Law
[Sh 467] S. Shelah, More on random nite models, in preparation.
[Sh 548] S. Shelah, Very weak zero one law for random graphs with order and
random binary functions, Random Structures & Algorithms, submitted.
[Sp] J. Spencer, Survey/expository paper: zero one laws with variable probabili
ties, Journal of Symbolic Logic vol 58 (1993) 114.
[Sh 550] S. Shelah, 01 laws, in preparation.
[Sh 551] S. Shelah, In the random graph G(n, p), p = n
a
: if has probability
0(n