Saharon Shelah - Classification Theory For Elementary Classes With The Dependence Property - A Modest Beginning

(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

CLASSIFICATION THEORY FOR ELEMENTARY
CLASSES WITH THE DEPENDENCE PROPERTY
- A MODEST BEGINNING
SH715
Saharon Shelah
Institute of Mathematics
The Hebrew University
Jerusalem, Israel
Rutgers University
Mathematics Department
New Brunswick, NJ USA
Abstract. Our thesis is that for the family of classes of the form EC(T), T a com-
plete rst order theory with the dependence property (which is just the negation of
the independence property) there is a substantial theory which means: a substantial
body of basic results for all such classes and some complimentary results for the rst
order theories with the independence property, as for the family of stable (and the
family of simple) rst order theories. We examine some properties.
I would like to thank Alice Leonhardt for the beautiful typing.
The author would like to thank the United States Israel Binational Science Foundation for partially
supporting this research.
First Typed - 98/Nov/25; revised with proofreading 2004/Feb/12
Latest Revision - 06/Sept/16
Done 10/98, 9/99
Typeset by A
M
S-T
E
X
1
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

2 SAHARON SHELAH
Annotated Content
1 Indiscernible sequences and averages, p.4-23
[We consider indiscernible sequences

b =
b
t
: t I) wondering whether
they have an average type as in the stable case. We investigate for any
such

b the set stfor(
b) of formulas ( x, y) such that every instance ( x, c)

divide

b to a nite/co-nite sets. We also consider the set dpfor(
b) of
formulas ( x, y) which can divide

b only to nitely many intervals; this is
always the case if T has the dependence property, i.e., dpfor(
b) = L
(T)
. If
T has the dependence property, indiscernible sequences behave reasonably
while indiscernible sets behave nicely. Similar behavior occurs for p S(M)
connected with indiscernible set

b which we call stable types. We then note
the connection between unstable types, unstable (x, y; c), and formulas
(x, y; c) with the independence property, i.e. on singletons.]
2 Characteristics of types, p.24-28
[Each indiscernible sequence

b =
b
t
: t I), has for each = ( x, y) a
characteristic number n = n
b,
, the maximal number of intervals to which
an instance ( x, c) can divide

b. We wonder what we can say about it.]
3 Shrinking indiscernbles, p.29-34
[For an indiscernible sequence
b
t
: t I) over a set A, if we increase the set
a little, i.e. if A
= A B then not much indiscernability is lost. An easy

case is: if I has conality > [B[ +[T[ then for some end segment J of I the
sequence
b
t
: t J) is an indiscernible sequence over A
.]
4 Perpendicular endless indiscernible sequences, p.35-52
[We dene perpendicularity and investigate its basic properties; any two
mutually indiscernible sequences are perpendicular. E.g., (for theories with
the dependence property) one indiscernible sequence can be perpendicular
to at most [T[
+
pairwise perpendicular indiscernible sequences. We then
deal with F
sp
|T|
+
-constructions.]
5 Indiscernible sequences perpendicular to cuts, p.53-61
[Using constructions as above we show that we can build models controlling
quite tightly the dual conality of such sequences where dual-cf(
b, M) =
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

DEPENDENT THEORIES 3
Min[B[ : B M and the average of

b over B

b is not realized in M.
That is, for any pairwise perpendicular
: <
) we can nd a model
M including them with dual-cf(
, M) being any (somewhat large) pregiven

regular cardinal).
6 Concluding Remarks, p.62-65
[We speculate on a parallel to DOP and to deepness. Also on the existence
of indiscernibles (starting with any set).]
Notation: 1) T is a rst order theory in a vocabulary
T
, C a monster model of T,
L is rst order logic so L
(T)
the rst order language with vocabulary , i.e., the
set of the rst order formulas in that vocabulary. Let L((T)) = L
(T)
. We may
write a A( C) for a
(g( a))
A.
2) Let , , be rst order formulas, = ( x) mean that x is a sequence of variables
with no repetitions including all free variables of (usually x = x
: < g( x))).
Let ( x, y) mean that we have a sequence of variables x and parameters y where
x y is with no repetitions; so = ( x, y) is not exactly equality.
3) I, J denote linear orders (used to index indiscernible sequences or sets). We shall
use

b =
b
t
: t I) with I (or J) a linear order, and
b
t
: t I
m
C for some
m < . We use I, J as subsets of
m
C for some m (not constant).
4) t denotes truth values,
t
is if t = truth or one and is if t = false or 0. So
if(statement)
is , if the statement is true, false resp., and
if(i)
means
if(i=1)
.
5) dcl(A) is the denable closure of A, acl(A) is the algebraic closure of A (inside
C or, here more interesting, C
eq
).

Our main interest here is in (rst order complete) theories with the dependence
property, but in the beginning we do not always assume this.
This work is continued in [Sh 783].
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

4 SAHARON SHELAH
1 Indiscernible sequences and averages
We try to continue [Sh:c, Ch.II,4.13], but we do not rely on it so there is some
repetition. In [Sh:c], the notion stable (complete rst order) theories, the notions
of indiscernible set and its average (and local versions of them) play an important
role. In an unstable theory, indiscernible sequences are not necessarily indiscernible
sets. Still for an indiscernible set I if T has the dependence property, the basic claim
guaranteeing the existence of averages (any formula ( x, a) divide I into a nite
and a co-nite set) holds. Moreover, any ( x, a) divides any indiscernible sequence
into the union of < n
convex sets. For any T, we can still look at the rst

order formulas ( x, y) which behaves well, i.e. any ( x,
b) divide any indiscernible

sequence

b to a nite/conite set.
In 1.3 - 1.9 + 1.9(c) + (d) we dene the relevant notions: average type for an
ultralter, Av(J, D) or Av
(J, D) or any average type for an indiscernible sequence

Av(J,
b
t
: t I)) and majority (maj) for nite sequences (saying how the majority
behave). We dene the set of stable formulas for an indiscernible

b (stfor(
b); also
dpfor(
b) which is L
(T)
for T with the dependence property), and state some basic
properties.
We dene a notion of distance (< ) between indiscernible sequences (as in
[Sh:93]). Being of nite distance is an equivalence relation and this is related to
canonical bases (of types, of indiscernible sets) which play important role for stable
theories, hence we try to dene parallels in 1.9, see 1.13(2).
Then we note a dichotomy for the types p S
m
(M). Such a type p may be
stable (see Denition 1.19, Claim 1.16 - 1.31); not only is then the type denable,
but for every ultralter D on
m
M with Av(M, D) = p, any indiscernible sequence
constructed from D is an indiscernible set, and the denition of p comes from an
appropriate nite large enough (, k)-indiscernible set. If p S
m
(M) is non-stable,
that is not stable, then there is a partial order with innite chains closely related to
it. We note that if T is unstable with the dependence property, then some (x, y, c)
dene a quasi order with innite chains and also that if T is unstable some (x, y; c)
has the order property (though not necessarily the property (E) of Eherenfeucht,
see [Eh57], some (x
1
, . . . , x
n
) is an irreexive relation on some innite set in some
model of T). It may be hard now to see how he could have not dened stable, but
without C
eq
, and no dichotomical theorem, looking for a reasonable property for
which the proof works, it had been quite natural.
On the subject see [Sh:c, II].
By [Sh 10] the possible function f
T,( x, y)
() = Sup[S
m
(A)[ : [A[ are charac-

terizing (if ( x; y) stable is: , if unstable without independence then Ded()
and if with the independence property 2
). Also ( x, y) is unstable in T i it has the

order property i ( x; y) has the strict order property or the independence prop-
erty. Hence T is unstable i it has the order property or the strict order property
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

even by a formula
(x, y). But it follows there that even if T has both the inde-
pendence property and the strict order property, if ( x, y) has the independence
property some
(x, y) has the independence property. Then Lachlan proves that

if ( x, y) has the strict order property some
(x, y) has the strict order property.

More on ( x, y) with the independence property, see Laskowski [Lw92].
The above settles f
T,
, now for f
T
() = sup[S(A)[ : [A[ this gives only
(f
T,
())
|T|
, Keisler [Ke76] show that if a countable T is unstable without the
independence property but if Ded() does not bound for some 2
|T|
then
f
T
() = Ded()
0
for 2
|T|
. In [Sh:c, III] it is proved (for T not nec-
essarily countable, unstable without the independence property) that for some
=
ord
(T) [T[
+
, f
T
() = Ded()
<
Ord
(T)
for 2
|T|
.
In [Sh 10], [Sh:a, II,4], contains some claims on averages when T (or ) has
the dependence property continued here. On other possible dividing lines among
unstable theories see [Sh 500, 2], (SOP
n
, n 3), [DjSh 692], (NOP
n
, n = 1, 2),
[DjSh 710] (the oak property), [Sh 702].
We may look at Th(M, P
1
, . . . , P
n
) with P
a unary relation, this is connected

to investigating a model of T in logics with extra quantiers; see Baldwin Shelah
[BlSh 156], Shelah [Sh 205], [Sh 284b], [Sh 284c].
Baldwin Benedikt [BlBn00] deal with the following. Let N be a model of T, T
with the dependence property, we expand N by adding a unary predicate P inter-
preted as an indiscernible sequence, I. We extend (N, I) to an [I[
+
-saturated model
and ask is (M, P) benign (see, e.g., [BBSh 815]).
Grossberg Lessman [GrLe00, 4] deals with the context: let T be complete rst
order theory, C its monster model, p say a 1-type, now we deal with the order
property, independence property and strict order property when we restrict the
parameters and variables, i.e., satisfaction, to p(C), mainly generalizing parallels of
results in [Sh:c], in particular the average (x, y) has the order property i it has
the strict order property or the independence property
1
. Compare with [Sh:c, II].
We thank John Baldwin and an anonymous reader for some comments and Eyal
Firstenberg for a very careful reading, corrections and comments.
1.1 Context. T a complete rst order theory, its monster model being C = C
T
as
usual in [Sh:c] the monster, C
eq
is when we add elements to designate equivalence
classes.
1.2 Denition. 1) T has the dependence property or is dependent means it does
not have the independence property whose denition is repeated below.
1
so though the names are similar this is not specially related to 1, e.g., the notion of stable
types are not related
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

6 SAHARON SHELAH
2) T has the independence property if some formula ( x, y) has the independence
property (in T), which means that for every n
C [= ( y
0
, . . . , y
n1
)
n
2
( x)(
<n
( x, y
)
if(())
).
3) The formula ( x; y) has the dependence property (in T) if it does not have the
independence property.
We can use below just J just of the form
>
B.
1.3 Denition. Consider a set of I of nite sequences of length m from the monster
model C where m < , and an ultralter D over I.
1) Let Dom(D) = I.
2) For J a set of nite sequences from C we let Av(J, D) be ( x, a) : x = x
: <
m), a J and
b I :[= (
b, a) D. It will be called the D-average over J or

the average type over J by D. If J =
>
B abusing our notation we may write B
instead of J (or M if B = [M[). (Av stands for average).
3) Av
(A, D) where = ( x, y) is the set of formulas of the form ( x, a) or the

form ( x, a) belonging to Av(A, D) and Av
(A, D) is the union of Av
(A, D) for
. Similarly Av
(J, D), Av
(J, D).
4) Let D be an ultralter on
n
B and B A and I is an innite linear order.
We say

b =
b
t
: t I) is based on D over A or

b is an (A, D)-indiscernible
sequence or is a D-indiscernible sequence over A if for each t I the type tp(
b
t
, A
b
s
: s <
I
t) is the average by D over A
b
s
: s <
I
t. If A = c : c Dom(D),
i.e. A is the domain of D we may write
b is a D-indiscernible sequence or

b is
based on D. This makes sense for I nite but we shall mention it in such a case.
5) Let p S
m
(B) and B A. We say

b =
b
t
: t I) is based on p over A if for
some ultralter D on
m
B the sequence

b is based on D over A and p = Av(B, D).
1.4 Comment: 1) If D is a principal ultralter on J, say
D then Av(B, D) =
tp(
, B).
2) If

b =
b
t
: t I) is based on D over A then D is a principal ultralter i
(s, t I)(
b
s
=

b
t
).
1.5 Claim. 1) For D an ultralter on I
m
C and A C the set Av(A, D) is a
complete m-type over A, i.e., S
m
(A) (and it does not split over I, see Denition
4.23).
2) Let p S
m
(A). Then p is nitely satisable in B i p = Av(A, D) for some
ultralter D on
m
A.
3) If p S
m
(M) then p is nitely satisable in M.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

Proof. Easy.
1.6 Claim. Let D be an ultralter on
m
B and B A.
0) Assume that

b
1
=
b
1
t
: t I
1
) is based on D over A and

b
2
=
b
2
t
: t I
2
) and
for any n < , t
1
<
I
1
. . . <
I
1
t
n
and s
1
<
I
2
. . . <
I
2
s
n
we have tp(
b
1
t
1
. . .
b
1
t
n
, A) =
tp(
b
2
s
1
. . .
b
2
s
n
, A). Then

b
2
is based on D over A, (see part (5)); (if I
1
is nite
we should add [I
2
[ [I
1
[).
1) For any linear order I there is

b =
b
t
: t I) based on D over A.
2) If I is a linear order and

b =
b
t
: t I) is based on D over A, then

b is an
indiscernible sequence over A.
3) If

b =
b
t
: t I) is based on D over A and J I, then

b J =
b
t
: t J) is
based on D over A; here we allow that J is nite.
4) If

b
t
: t I) is based on D over A for = 1, 2 then

b
1
,

b
2
realizes the
same type over A (in C); again we allow I to be nite.
5) If

b
t
: t I
) is based on D over A and n < and I
[= t
1
< . . . < t
n
for
= 1, 2, then the sequences

b
1
t
1
1
. . . b
1
t
1
n
and

b
2
t
2
1
. . .
b
2
t
2
n
realize the same type over
A.
6) p = Av(A, D) does not split over B, which means that: if

b, c
m
A, M < and
tp(
b, B) = tp( c, B) and = ( x, y) a formula, then ( x,
b) p ( x, c) p.
Proof. Easy, E.g.
1) If I is well ordered, choose

b
t
by induction on t. By compactness this holds for
any I.
1.6
1.7 Denition. 1) For an innite linear order I such that for some
2
innite linear
orders I
1
, I
2
we have I = J
1
+ I
2
and an indiscernible sequence

b =
b
t
: t I),
having g(
b
t
) = m for t I, we dene:
(a) stfor
pa
(
b) = ( x, y, c) : g( x) = m, and for every a

g( y)
C, the set
t I : C [= (
b
t
; a, c) is nite or the set t I : C [= (
b
t
; a, c) is
nite
(stfor stands for stable formulas, pa stands for parameters)
(b) stfor(
b) = ( x; y) : ( x, y) stfor
pa
(
b), i.e., no parameters

(c) dpfor(
b) = ( x, y) : g( y) = m and for every a

g( y)
C the set
t I : C [= [ a,
b
t
] is a nite union of convex subsets of I.
Let dpfor
pa
(
b) be dened similarly allowing parameters, obviously dpfor

stands for formula with the (relevant) dependence property
2
or demand in each of the denitions below that if J is a linear order extending I and

b
t
for
t J\I such that
b
t
: t J is an indiscernible sequence the condition on ( x, y) holds
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

8 SAHARON SHELAH
(d) dpfor
n
(
b) is dened similarly when the union is of n convex sets; similarly

is the other cases
(e) writing stfor
pa
(
b) = L() we mean the set of relevant formulas.

2) For a sequence

b =
b
t
: t < k) with
b
t

m
C, and formula = ( y, z), g( y) = m,
we dene
maj
(A,
: < k))) = ( y, c)
t
:t true,false,
c
g( z)
A, and [ :[= (
, c)
t
[ > k/2.
Clearly ( y, c) maj
(C,
: < k)) is a rst order property of c (with the

parameters

b
0
. . .
b
k1
).
If not said otherwise we use k odd so that we have completeness.
(Note that maj stands for majority; this is not necessarily a type, just a set of
formulas).
3) E = E
k
( y, z)
, where ( y, z) is a formula in L
(T)
, written z
1
E z
2
with g( z
1
) =
g( z
2
) = (g( x)) k (written ( x
1
, . . . , x
n
) instead of x
1
. . . x
n
, abusing notation;
normally k is odd) is dened as follows: ( x
0
, . . . , x
k1
)E( x
0
, . . . , x
k1
) =:
( z)(
uk,|u|>k/2
u
( x
, z)
uk,|u|>k/2
u
( x
, z)).
Of course, it is an equivalence relation.
1.8 Claim. 1) If

b =
b
t
: t I) is an innite indiscernible sequence, g(
b
t
) = m
and ( y; z) stfor(
b) so g( y) = m, then for every k large enough we have:

(a) for any c of length g( z), for some truth value t the set t I :[= (
b
t
, c)
t
has < k/2 members

(b) if t
0
, . . . , t
k1
are distinct members of I then maj
(A,
b
t
: < k)) S
m
(A)
for every A C, in fact for every nonprincipal ultralter D over
b
t
: t I
and set A we have maj
(A,
b
t
: < k)) is a subset of Av(A, D), in fact is

equal to Av
(A, D)
(c) if t
0
, . . . , t
k1
I with no repetitions and s
0
, . . . , s
k1
I with no repetition
then (
b
t
0
, . . . ,
b
t
k1
) E
k
( x, y)
(
b
s
0
, . . . ,
b
s
k
1
)
(d) for some nite we have: if I
, I J where J is a linear order,

b
t
:
t J),
I =

b and

b
is -indiscernible sequence, [I
[ k, then
() (b),(c) holds for

b
and
() ( y, z) stfor(
).
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

2) Let

b =
b
t
: t I) be an indiscernible sequence ( y; z) dpfor(
b) so g( y) = m
then for some k = k
,
b
:
(a) for any c of length g( z) and t the set t I : [
b
t
, c]
t
is the union of k
intervals.
3) The k above depends only on p
n
: n < ) where p
n
is tp(
b
t
1
. . .
b
t
n
, ) for any
t
1
<
I
. . . <
I
t
n
.
Proof. 1)
(a) By compactness
(b) just think of the denitions
(c) follows from clause (b)
(d) by compactness.
2), 3) Similarly.
1.8
1.9 Denition. Let

b =
b
t
: t I) be an innite indiscernible sequence.
1) We dene (Cb stands for canonical bases, working in C
eq
):
(a) for ( y; z) stfor(
b) let Cb
( y; z)
(
b) be (
b
t
0
, . . . ,
b
t
k1
)/E
k
( y, z)
C
eq
,
with k = k
( y, z)
(
b) minimal as in 1.8(1)(a) and any pairwise distinct

t
0
, . . . , t
k1
I
(b) Cb(
b) = dclCb
( y, z)
(
b) : ( y, z) stfor(
b) C
eq
.
2) If I has no last element then
Av
(A,

b) = ( x, a)
t
: for every large enough t I we have
C [= (
b
t
, a)
t
where t is a truth value
Av
(A,

b) = Av
(A,

b) :
Av(A,

b) = ( x, a) :( x, y) L
(T)
, a
>
A
and [= (
b
t
, a) for every large enough t I.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

10 SAHARON SHELAH
3) Let Av
stfor
(A,

b) be Av
b, A) for = stfor(
b), similarly for replacements to

stfor.
4) If the order on I = Dom(
b) does not matter then we can replace

b by
b
t
: t I.
1.10 Claim. [T is dependent]
Assume I is a linear order with no last element and

b =
b
t
: t I) an indis-
cernible sequence, g(
b
t
) = m.
1) dpfor(
b) = ( x, y): any and y but x = x
: < m).
2) Av
(A,
b) S
m
(A), see Denition 1.9(2), i.e., for every ( x, c), for some t for
every large enough t we have [= (
b
t
, c)
t
.
3) Av(A,
b) S
m
(A), see Denition 1.9(2).
4) Av
stfor
(A,

b) does not depend on the order of I (so by 1.7 we can use I a set or
innite linear order which is not necessarily endless).
5) If

b is an (innite) indiscernible set, then stfor(
b) = dpfor(
b) = L
(T)
(and
stfor
pa
(
b) = dpfor
pa
(
b) = ( x, y, c) : c C, (and of course g( x) = m).

Proof. Left to the reader or see [Sh:c, II.4.13] (for part (5) see (A) (B) in 1.28).
To formalize clause (d) of 1.8(1) let
1.11 Denition. 1) For a set of formulas and k we say that
b
1
t
: t
I
1
),
b
2
t
: t I
2
) are immediate (, k)-nb-s (or the rst is an immediate (, k)-nb
of the second over A) if:
(a) both are (, k)-indiscernible sequences of length k
(b) for some (, k)-indiscernible sequence
b
t
: t I) and order preserving
functions h
1
, h
2
from I
1
, I
2
into I respectively we have t I
(t)
=

b
t
for = 1, 2
(nb stands for neighbors).
2) The relation being (, k)-nb-s is the closure of being an immediate (, k)-nb
to an equivalence relation. We say of distance n if there is a chain of immediate
(, k)-nb-s of length n starting with one ending in the other. We may write
instead of (, ) and if = L
(T)
we may omit .
3) We can replace k by < k (and even < ).
4) We write (A, , k) if the indiscernibility is over A (also in part (2)).
5) If

b
1
,

b
2
are innite indiscernible sequences, we say they are essentially nb-s
(over A) if for every nite L
(T)
and k < they are (, k)-nb-s (they are
(A, , k)-nb-s).
6) If

b is an innite indiscernible sequence over A we let C
A
(
b) =
b: for some

b
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

an essentially nb of

b over A,
b appears in

b
.
7) If

b is an innite indisernible sequence over A we let C
A
(
b) =
b :

b appears in
some

b
, an innite indiscernible sequence over A which is an A-nb of

b.
1.12 Remark. C
A
(
b) was dened in [Sh:93, Def.5.1](4).

1.13 Claim. 1) If
b
t
: t I) is an innite indiscernible sequence and ( y, z)
dpfor
n
(
b
t
: t I)), then for some nite and k, for any (, k)-nb sequence
t
: t I
) of
b
t
: t I) we have ( y, z) dpfor
n
(
t
: t I
)).
2) The result in (1) holds also for stfor
n
(
b
t
: t I)). If

b =
b
t
: t I) is an
innite indiscernible sequence and ( y, z) stfor(
b), then for some nite and k

for any (, k)-nb

b
of

b we have Cb
( y, z)
(
) = Cb
(y, z)
(
b) and Av
( y, z)
(C,

b
) =
Av
( y, z)
(C,

b).
3) If

b
1
,

b
2
are (, k)-nb-s of distance n for every nite and k < , then they
are L
(T)
-nb-s of distance n.
4) Like part (3) with a xed k, i.e., if

b
1
,

b
2
are (, k)-nb-s of distance n for
every nite , then they are (L
(T)
; k)-nb-s of distance n.
5) If

b is an innite indiscernible sequence, then stfor(
b) = stfor
n
(
b) : n < .
Proof. Easy. (Use compactness for (3) and (4).)

1.14 Denition. Let p S
m
(B).
1) We say p is denable where = ( x, y) and g( x) = m, if some ( y, c) dene
it with c B, where
2) We say ( y, c) dened p where = ( x, y) if:
() for every a
g( y)
B we have
( x, a) p C [= [ a, c].
3) We say p is -denable if it is denable by some ( y, c) with ( y, z) .
4) We say p S
m
(B) is denable if every p is denable.
1.15 Claim. The number of denable p S
m
(B) is at most ([B[ + 2)
|T|
.
Proof. For every denable p S
m
(B) choose a sequence
,p
( y, c
,p
) : =
(x
0
, . . . , x
m1
; y) L
(T)
) such that
,p
( y, c
,p
) dene p and c
,p
B. Now
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

12 SAHARON SHELAH
(a) the number of such sequences is ([B[ + 2)
|T|
(b) if p
1
, p
2
S
m
(B) and
,p
1
( y, c
,p
1
) : (x
0
, . . . , x
m1
; y) L
(T)
) is equal
to
,p
2
( y, c
,p
2
) : (x
0
, . . . , x
m1
; y) L
(T)
), then p
1
= p
2
.
Together we are done.
1.15
1.16 Claim. 1) Assume that B C, p S
m
(B) and D is an ultralter on
m
B
and

b =
b
t
: t I) is an innite D-indiscernible sequence over B (see ?(4)) such
> scitenp.1 undened
that Av(B, D) = p. Then there is a function F from (A, ) : A B is nite and
L
(T)
is nite into D, but we may write F
(A) instead of F((A, )) and

F
b
0
, . . . ,
b
1
, A) instead of F
b
0
. . .
b
1
A), such that:
() if and for each < we have

b

m
B,
b
0
, . . . ,
b
1
, A)( D)
then the sequence
b
t
: t I)
) (where the superscript in
means invert the order) is a -indiscernible sequence over A.

2) Let p S
m
(B), D,

b =
b
t
: t I), A = be as in part (1). For any ( y, z)
there are a nite
L
(T)
and k
< [also large enough as in 1.8] such that:

if

b

m
B for < k where k k
are as in part (1) for =
, then we have:
if ( y, z) stfor(
b) then
(i) Av
(C,

b) = maj
(C,
: < k)) = Av
(C, D)
(ii) p ( y, z) and even p
+
= Av
(C, D) is denable by a rst order

formula with parameters from B (see Denition 1.7(2)).
3) Assume that B
, p
, D
,

b
= b
t
: t I
) are as in part (1) for = 1, 2 and

B
1
= B = B
2
, p
1
= p = p
2
, I
1
= I = I
2
. Then Av
( y, z)
(C, D
) does not depend

on provided that ( y, z) stfor(
b
1
) stfor(
b
2
) and B = M C; recall that
Av
( y, z)
(C,

b
) = Av
( y, z)
(C, D).
Proof of 1.16. 1) There is no harm with increasing I, so without loss of generality,
(as we can supply appropriate

b
t
s) I has no last element. For simplicity without
loss of generality is closed under permuting and identifying the variables. Next
we shall choose F
(A).
Let n < be above the number of free variables in any formula in and let
t
0
< . . . < t
n1
< t be in I. Now tp
b
t
, A
b
t
: < n, C) is a nite set

of formulas and let ( x, c) be its conjunction. So ( x, c) tp(
b
t
, B
b
s
: s <
t, C) = Av(B
b
s
: s < t, D) hence J =
b
m
B : C [= (
b, c) D.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

Choose F
(A) as any such J; why is F as required? So suppose that A B is

nite and let

b
b
0
. . .
b
1
A) for < () or for < . Clearly
()
1

b
b
0
. . .
b
1
A) implies that [t
n1
< t I
b
t
,
realizes the
same -type over A
b
t
0
, . . . ,
b
t
n1
b
1
, . . . ,
b
0
].
But as
b
s
: s I) is an indiscernible sequence over B, and

b
b
1
, . . . ,
b
0
B, we
can replace t
0
< . . . < t
n1
< t, by any t
0
< . . . < t
n1
< t
from I, so
()
2
if

b
b
0
. . .
b
1
A) and t
0
< . . . < t
n1
< t
in I then

b
b
t
realizes the same -type over A
b
t
0
, . . . ,
b
t
n1

b
0
, . . . ,
b
1
.
But by the choice of n,
()
3
for any

b F
b
0
. . .

b
1
A) we have tp
b, A
b
s
: s I
b
1
, . . . ,
b
0
) = tp
b, A
b
s
0
, . . . ,
b
s
n1
b
1
, . . . ,
b
0
) : s
0
< . . . <
s
n1
are in I.
By induction on < we prove that
()
4
for k < for any m
i
0
< . . . < m
i
1
< and t
i
0
<
I
. . . <
I
t
i
k1
for i = 1, 2 the
-type which
b
m
1
0
. . .
b
m
1
1
b
t
1
k1
. . .
b
t
1
0
and
b
m
2
0
. . .
b
m
2
1
b
t
2
k1
. . .
b
t
2
0
realizes over A are equal.
So together we are done.
2) In Clause (i) the rst equality holds by Claim 1.8, Clause (a); that is such
exists by it. For the second equality, if ( x, c)
t
Av(C, D), then we can nd
t
: t I) based on (B c, D), by Claim 1.6(1). Apply the rst equality with
b
standing for

b. Note that tp(
, B) = tp(
b, B) hence there is an automorphism

of C over B mapping

b to

b
hence we can in clause (i) replace

b by

b
(not changing
: < 2k).
Clause (ii) follows from clause (i) as
for c
(g( z))
C we have: (x, c) p
+
i ( x, c) Av(B c,

b) i
( y, c) Av
(C,
: < k)) i C [= [ c, b
0
, . . . , b
k1
] where
(y, b
0
, . . . ,
b
k1
) =
uk,|u|k/r
u
(y,
).
[Why? The second i holds as Av
(C,
b) restricted to B c is p
+
, by
an assumption of our claim and similarly the rst. The third i holds by
clause (i) which we have proved, the fourth i holds by Denition 1.7(2).]
3) Obvious from clause (i) of part (2).
1.16
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

14 SAHARON SHELAH
The notions p S
m
(B) is nitely satisable in B, is denable, has uniqueness
and is stable are closely related and important here; we shall dene now the two
later ones and investigate their relationships.
1.17 Denition. We say p S
m
(B) has unique indiscernibles (or has uniqueness)
when it is nitely satisable in B and
if ()(a) D
is an ultralter on
m
B for = 1, 2
(b) p = Av(B, D
) S
m
(B) for = 1, 2
(c)

b
t
: t I) is a (B, D
)-indiscernible sequence,
see Denition 1.3(4), i.e.,

b
t
realizes
Av(B
s
: s <
I
t, D) for t I where of course
I is an innite linear order
then

b
1
,

b
2
realizes the same type over B.
1.18 Claim. If p S
m
(B) is denable and nitely satisable in B (the second
follows from the rst if B = M), then p has unique indiscernibles.
Proof. For each formula = ( x, y), g( y) = m let
( y, c
) be such that:
() c
B
() for every a
g( y)
B we have
( x, a) p C [=
( a, c
).
Let D
1
, D
2
be ultralters on
m
B such that Av(B, D
1
) = p = Av(B, D
2
) and let
b
t
: t I) be as in Denition 1.17.
Now we prove that
()
n
if t
n
<
I
. . . <
I
t
1
and = ( x
n
, . . . , x
1
, z), g( x
) = m,

d
g( z)
B then
C [= [
b
1
t
n
, . . . ,
b
1
t
1
,

d] [
b
2
t
n
, . . . ,
b
2
t
1
,

d].
Let ( x
n
, . . . , x
1
, y) and

d
g( y)
B be given.
We dene the formula
k
( x
k
, x
k1
, . . . , x
1
, z
k
) and

d
k

g( z
k
)
B by downward in-
duction on k n.
Case 1: k = n.
n
( x
n
, . . . , x
1
,

d
0
) = ( x
n
, . . . , x
1
,

d) so

d
0
=

d.
Case 2: k < n.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

We have
k+1
( x
k+1
, . . . , x
1
,

d
k+1
), now as p is denable (and xing some of the
parameters causes no harm) there is
k
( x
k
, . . . , x
1
,

d
k
) with

d
k
B of course, such
that:
for every a
k
, . . . , a
1

m
B we have:
k+1
( x, a
k
, . . . , a
1
,

d
k+1
) p C [=
k
( a
k
, . . . , a
1
,

d
k
).
So we have carried the induction.
Now let 1, 2 and let A =
d
k
: k n, it is a nite subset of B, and let
be any nite subset of L
(T)
which includes
k
: k n and let F
be as in
1.16 with D
, B, A, here standing for D, M, A, there and let a
i
F
( a
j
: j <
i) A)
n
B for i < and 1, 2.
So a
i
: i < ) is -indiscernible over A, and clearly by 1.16(1)
1
if i
1
< . . . < i
n
< , t
n
<
I
. . . <
I
t
1
then
tp
( a
i
1
. . . a
i
n
, A) = tp
t
1
. . .
t
n
, A).
Easily
2
for 1, 2, k n and i
1
< . . . < i
k+1
we have
[=
k+1
[ a
i
k+1
, . . . , a
i
1
,

d
k+1
]
k
[ a
i
k
, . . . , a
i
1
,

d
k
]
[Why? It suces to note that [=
k+1
[ a
i
k+1
, a
i
k
, . . . , a
i
1
,

d
k+1
], i a
m
B :
k+1
[ a, a
i
k+1
, . . . , a
1
,

d
k+1
] D, i
k+1
( x, a
i
k
, . . . , a
i
1
,

d
k+1
) p, i
[=
k
(a
i
k+1
, . . . , a
i
1
,

d
k
).
Why? The rst i holds by the choice of F
( a
0
, . . . , a
i
k+1
1
)), the second
i as p = Av(B, D
) and the last i by the choice of

k
,

d
k
.]
hence by the transitivity of
3
[=
n
[ a
i
n
, . . . , a
i
1
,

d
n
]
0
[
d
0
]
but
n
= and

d
n
= c so
4
[= [ a
i
n
, . . . , a
i
1
, c]
0
[
d
0
].
So the truth value of [ a
i
n
, . . . , a
i
1
, c] does not depend on 1, 2 so together
with 1.16(1) we get ()
n
so we are done.
1.17
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

16 SAHARON SHELAH
1.19 Denition. 1) We call p S
m
(B) a stable type if:
(a) p is nitely satisable in B
(b) for some ultralter D on
m
B satisfying p = Av(B, D) and sequence

b =
b
t
: t I) based on D (see Denition 1.3(4) so I is innite)

b is an
indiscernible set over B.
2) We call p S(B) a non-stable type if it satises (a) above but not (b).
3) An innite indiscernible sequence

b is stable/nonstable if it is an indiscernible
set/it is not an indiscernible set over the empty set (see 1.28).
1.20 Claim. Assume T is dependent.
1) If p S
m
(B) is stable then p is denable and has uniqueness.
2) If p S
m
(B) is stable, then every innite

b based on p is an indiscernible set,
that is for every D,

b =
b
t
: t I) as in Denition 1.19 the sequence

b is an
indiscernible set over B.
3) The number of stable p S
m
(B) is ([B[ + 2)
|T|
.
Proof. 1) The type p is denable by 1.16(2)(ii), 1.10(5) and so has uniqueness by
1.18.
2) Let

b be based on D over B and let D
1
= D and

b
1
t
=

b
t
. As p is a stable
type there are D
t
: t I
) as in Denition 1.19. Let D

2
= D
and we can nd
b
2
t
: t I) such that:

b
2
t
realizes Av(B
s
: s <
I
t, D
) (by 1.6). Now clearly I [=

t
1
< . . . < t
n
, I
[= s
1
< . . . < s
n
implies tp(
t
n
. . .
t
1
, B) = tp(
b
2
s
n
. . .
b
2
s
1
, B)
for every n (by 1.6(5)) hence also
b
2
t
: t I) is an indiscernible set over B. By
part (1) the sequences
b
1
t
: t I),
b
2
t
: t I) realizes the same type over B hence
also
b
1
t
: t I) is an indiscernible set over B.
3) By part (1) and 1.15.
1.20
1.21 Claim. 1) Assume p S
m
(B) has uniqueness. If for = 1, 2, D
is an
ultralter on
m
B satisfying p = Av(B, D
) and

b
t
: t I) is based on D
,
then

b
1
,

b
2
has distance 2.
2) In part (1), instead uniqueness it suces to assume that n < , t
0
<
I
. . . <
I
t
n1
the sequences

b
1
t
0
. . .
b
1
t
n1
and

b
2
t
0
. . .
b
2
t
n1
realizes the same type.
Proof. Should be easy.
1) By the denition of uniqueness (see Denition 1.17) this follows by part (2).
2) By 1.13 it is enough to prove, for any nite L
(T)
and k < that

b
1
,

b
2
are (, k)-nb-s of distance 2. This in turn follows by 1.16.
1.21
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

The denability of p proved in 1.16(2) for stfor(
b) say more than stated.

The dening formulas are canonical (i.e., the formula depends on and T, the
parameters depend also on p).
1.22 Denition. Let ( x, y) be a formula L
(T)
which has the dependence
property (holds if T has it).
1) Let k
1
[( x, y)] = k
T
[( x, y)] be the minimal k < such that there is no sequence
a
: < k) with a

g( y)
C for < k such that
C [=
n
2
( x)
<k
( x, a
)
()
.
2) We say (k, ) is suitable to

b, if:
(a)

b is a -indiscernible sequence
(b) the formulas in have arity < g(
b)
(c) if

b
is a -indiscernible sequence of the same -type as

b, g(
) > 2k
then for any a
g( y)
C for some truth value t the set t Dom(
) : C [=
[
t
, a]
t
has < k members.
3) We say (k, , n) is strongly suitable to

b, if n = n
: k), (a), (b), (c) as

above hold and
(d) if

b
t
: t I) is a -indiscernible sequence of the same -type as

b and
[I[ >
i<k
n
i
then for no a
g( y)
C are there t
,m
I (for k, m < n
) with
no repetitions satisfying [t
1
,m
1
< t
2
,m
2
(
1
<
2
(
1
=
2
m
1
< m
2
))]
such that for any m
1
, m
2
< n
we have [
1
,m
1
, a] [
2
,m
2
, a] i
1
2
is even.
4) ( y, z) =
k,( x, y)
( y, z) is the canonical denition if it is as in the proof of
1.16(2).
5) Similarly for ( x, y, c).
Trivially
1.23 Claim. In 1.16(2), if ( x, y) stfor(
b), then p is dened by

k,( x, y)
( y, c)
for some c B (of length g( z
( x, y)
)).

Now it should be clear that
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

18 SAHARON SHELAH
1.24 Claim. If p S
m
(B) is stable and

b is an innite indiscernible set based
on p, then for any automorphism F of C
eq
which is the identity over Cb(p), the
sequences

b, f(
b) are nb-s of distance 2. [See Denition 1.11.]

We shall show
1.25 Claim. [T is dependent] If

b =
b
t
: t I) is an innite indiscernible set
over and an indiscernible sequence over A, then

b is an indiscernible set over A.
Proof. By 1.28 below.
But we rst prove some local claim.
1.26 Claim. Assume
()
= ( x
1
, . . . , x
n
, y), g( x
) = m and for any permutation of 1, . . . , n

we let
( x
1
, . . . , x
m
, y) = ( x
(1)
, . . . , x
(n)
, y).
1) If

b =
b
t
: t I) is a ( x
1
, . . . , x
n
, c)-indiscernible sequence but not set and
I = I
0
+I
1
+I
2
, [I
0
[ n2, [I
2
[ n2, then for some permutation of 1, . . . , n
and t
3
, t
4
, . . . , t
n
I
0
I
2
, we have:
( x
1
, x
2
,

d) =: ( x
1
, x
2
,
b
t
3
, . . . ,
b
t
1
, c) linearly
ordered
b
t
: t I
1
) that is, for t ,= s I we have [= [ a
t
, a
s
,

d] t <
I
s.
2) In part (1), with I is innite of course,
( x
1
; x
2
. . . x
n
, z) / stfor(
b I);
moreover
( x
1
, x
2
,

d) / stfor(
b I
1
).
Proof. See e.g. [Sh:c, II] and history there; the point being that the permuations
exchanging k, k + 1 for k = 1, . . . , n 1 generate the group of permutations of
1, . . . , n.
1.26
More information concerning 1.26 is
1.27 Claim. [T is dependent.]
1) Assume
(a)

b =
b
t
: t I) is an indiscernible sequence of m-tuples over

d and I =
J
0
+ J
1
+ J
2
(b) ( x, y,

d) / stfor(
b) and let
( y, x,

d) = ( x, y,

d)
(c) [J
0
[, [J
2
[ are nite large enough, in fact just k
( x; y, z)
from 1.8(2).
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

Then we can nd ( x, e) and truth value t such that
() for t J
1
, [= ( y)(( y, e)
( y, a
t
,

d)
t
)
() for t ,= s J
1
we have
t <
J
1
s i
( y, e),
( y, a
t
,

d)
t

( y, a
s
,

d)
t
() for some n k
( x
1
, x
2
, z)
and t
0
<
J
. . . <
J
t
n1
from J
0
J
2
and
n
2 we
have ( y, e) =
<n
( y,
b
t
,

d)
()
.
2) In part (1) we can deduce that
() ( x
1
, x
2
, e) is a partial order and

b J
1
is ( x
1
, x
2
, e)-increasing where
( x
1
, x
2
, e) = ( y)[( y, e)
( y, x
1
,

d)
( y, x
2
,

d)].
3) Assume in part (1) that J
2
= and [J
0
[ large enough. Then still we can nd
( x, e), t and n() such that:
() if t
1
<
J
1
. . . <
J
1
t
n()
then
( y)[( y, e) &
n()
i=1

( y, a
t
i
,

d)
t
() if s <
J
1
t
1
<
J
1
. . . <
J
1
t
n()
<
J
n
s then
( y, a
s
,

d)
t
, ( y, e)
( y, a
t
1
,

d)
t
. . .
( y, a
t
n()
,

d)
t
() as in part (1).
Proof. 1) Without loss of generality I is dense without rst and without last
element. As ( x, y,

d) / stfor
pa
(
b) for some sequence a, for both truth value

t the set t I : C [= (
b
t
, a,

d)
t
is innite. As we can replace by
without loss of generality t I : C [= (
b
t
, a,

d) is unbounded from above (in
I). As

b is an indiscernible sequence over

d easily
() for every s I the type (
b
t
, y,

d)
if(s<t)
: t I is consistent.
Let n < be minimal such that for some
n
2 and t
0

= (
b
t
, y,

d)
()
: < n (
b
t
, y,

d) : t <
I
t
0
(
b
t
, y,

d) :
t
n1
<
I
t is inconsistent. We let
n
2 be as above such that for this n, is
minimal under the lexicographic order. Such n, exists as ( x, y;

d) dpfor
pa
(
b).
By () above clearly n 1. Also by the minimality on n clearly (0) ,= false,
(n1) ,= true, so n 2, (0) = true and (n1) = false; so for some k < n1
we have (0) = (1) = . . . = (k) ,= (k + 1).
As p
<t
0
,...,t
n1
>
is inconsistent we can nd nite sets J
s : s <
I
t
0
, J
+
s : t
n1
<
I
s such that the set (
b
s
, y,

d) : s J
b
t
, y,

d)
()
: <
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

20 SAHARON SHELAH
n (
b
s
, y,

d) : s J
+
is inconsistent.
Let ( x, d
) be the conjunction of (
b
s
, y,

d) : s J
b
t
, y,

d)
()
: <
n, / k, k + 1 (
b
s
, y,

d) : s J
+
. So ( y,

d
), (
b
t
k
, y,

d), (
b
t
k+1
, y,

d)
is inconsistent, so
C [= ( y)[( y,

d
) (
b
t
k
, y,

d) (
b
t
k+1
, y,

d)].
However ( y,

d
), (
b
t
k
, y,

d), (
b
t
k1
, y,

d) is consistent (otherwise
n
2 such
that (
() = ()) / k, k + 1 and t
: < n) contradict the minimality of

by the lexicographic order). So by the indiscernibility
for s
1
, s
2
[t
k
, t
k+1
]
I
we have (y,

d
), (
b
s
1
, y,

d) (
b
s
2
, y,

d) i s
1
< s
2
.
By indiscernibility this clearly nishes the proof of part (1), except the bounds on
[J
0
[ +[J
2
[, which we can get by redening k
( x, y, z)
or repeating the proof.
(2), (3) Follows.
1.27
1.28 Claim. [T is the dependent.]
For an innite indiscernible sequence

b =
b
t
: t I) (over ) the following condi-
tions are equivalent:
(A)

b is an (innite) indiscernible set
(B) stfor(
b) = L
(T)
(C) for every set A, if

b is an indiscernible sequence over A, then

b is an
indiscernible set over A.
Proof. We shall prove three implications completing a circle.
(C) (A):
Holds as (A) is a special case of (C), i.e. choosing A = .
(A) (B):
Assume ( x, y) / stfor(
b) then for every n for some c

n

g( y)
C the sets I
true
n
=:
t I :[= [
b
t
, c
n
] and I
false
n
= t I :[= [
b
t
, c
n
] has n members. Let
h : I true,false, and dene p
h
= (
b
t
, y)
h(t)
: t I, and we shall show below
that p
h
is consistent, so T has the independence property, contradiction, so there
is no ( x, y) / stfor(
b), hence stfor(
b) = L
(T)
, i.e., condition (B) holds.
So let q be a nite subset of p
h
, so for some nite J I we have q = (
b
t
, y)
h(t)
:
t J. Let n = [J[ and so there is a permutation of the set I such that
t J (t) I
h(t)
n
, possible as [I
t
n
[ n. As

b is an indiscernible set there
is an automorphism F of C such that t I F(
b
t
) =

b
(t)
. Clearly F(q) =
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

(
b
(t)
, y)
h(t)
: t J tp( c
n
, b
t
: t I), hence F
1
( c
n
) realizes q, as
promised.
(B) (C):
So assume that

b is an indiscernible sequence over A. Let I
+
= I
0
+I +I
2
with
I
0
, I
2
any innite linear orders. Clearly there are

b
t

m
C for t I
0
I
2
such that
b
+
=
b
t
: t I
+
) is an indiscernible sequence over A.
If

b
+
is not an indiscernible set over A, then for some c A and = ( x
1
, . . . , x
n
, c)
the sequence

b
+
is -indiscernible sequence but not a -indiscernible set, so by
1.26 some
( x
1
, x
2
,

d) =
( x
1
, x
2
,

d) linearly ordered
b
t
: t I where is
a permutation of 1, . . . , n, t
3
, . . . , t
n
I
0
I
2
and

d =

b
t
3
b
t
4
. . .
b
t
n
c c
b
s
:
s I
0
I
2
; hence
( x
1
, x
2
, z) is not a stable formula for

b. So stfor(
b) ,= L
(T)
,
contradiction. So

b
+
hence

b is an indiscernible set over A as required.
1.28

We now may think on ( x, y, c) which are stable for

b which we get in the ap-
proximation of order in 1.25(4). We may wonder can we not by expanding p (with
more variables, over the same B preserving nite satisability in B) get clearer
picture. This may help in getting indiscernible sequences. (See more in concluding
remarks).
1.29 Denition. If p S
m
(B) is nitely satisable in B let
(a) stfor(p) = stfor(
b): for some ultralter D on

m
B satisfying p = Av(B, D),
the sequence

b =
: < ) an indiscernible sequence obeying D (this

does not matter if we take one or all such

b by 1.6(4),(5))
(b) Cb(p) = Cb(b): for some ultralter D on
m
B satisfying Av(B, D) = p
the sequence b is a (D, B)-indiscernible sequence (in C
eq
, of course).
Of course
1.30 Observation. For any M C and p S
m
(M), or just p S
m
(B) is nitely
satisable in
m
B we have stfor(p), Cb(p) are well dened as there are ultralters D
on
m
M such that Av(M, D) = p.
1.31 Observation [T is dependent.]
If

b is a ( x
, . . . , x
1
; c)-indiscernible sequence over A but is a
-indiscernible
set over and has > k
members, then

b is a ( x
0
, . . . , x
k1
; c)-indiscernible
set over A where
= x
<n
( x, y
)
if(())
:
n
2 and is a permutation of
1, . . . , and n = n
is such that
n
( x, y)
from 1.2 fail.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

22 SAHARON SHELAH
Proof. By the proof of (B) (C) inside the proof of 1.28 it is enough to prove
for every permutation of 1, . . . , that
( x
1
, y,

d) =
( x
1
; x
2
, . . . , x
,

d)
stfor
pa
(
b).
By the proof of (A) (B) inside the proof of 1.28 this follows from
b is
-
indiscernible over .
1.31
1.32 Remark. Note that p S(M) may be denable but not stable, e.g. M N are
models of the theory of (R, <), and a NM is above all b N, then tp(a, M, N)
is denable but not stable.
1.33 Conclusion: [T is dependent and is unstable.]
1) There are M C and non-stable p S
1
(M) [in C, not just C
eq
!].
2) There is an indiscernible sequence of elements which is not an indiscernible set
of elements (over !)
Proof. 1) As T is unstable, for some M C we have [S
1
(M)[ > |M|
|T|
hence by
1.20(3) some type p S
1
(M) is non-stable.
2) By part 1) and Denition 1.19.
1.33
Now 1.27(3) applies to 1.33(2) (where x
i
is x
i
) gives
1.34 Conclusion. If T has the dependence property but is unstable, then some
formula (x, y; c) dene on C a quasi order and even partial order with innite
chains, (so x, y singletons).
Proof. By 1.33(2) and 1.27.
1.34
1.35 Remark. So if T satises some version of -stable (see [Sh 300, Ch.II] or [Sh
702]) then T is stable or T has the independence property.
So we may wonder
1.36 Question: 1) Does the has the dependence property case in 1.35 is needed?
2) If T has the independence property does some (x, y, c) have the independence
property?
3) Let p S
m
(B) be non-stable, letting D D
: D
an ultralter on
m
B satisfying
Av(B, D
) = p, how many

b
D
can we nd in 1.20 which are pairwise not nb-s,
(for T with the dependence property)?
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

Note that
1.37 Observation. If T is unstable, then some formula (x, y, c) has the order
property (equivalently is unstable, hence some (x, y, c)) dene a partial order
with innite chains or has the independence property.
Proof. We know that some (x, y) is unstable so choose a formula (x, y, c) with
the order property, such that g( y) is minimal. So there is an indiscernible sequence
a
i
: i < 4) such that C [= [b
i
, a
j
, c] i j < i. Clearly b
i
: i < 4) is
an indiscernible sequence over c, if it is not an indiscernible set, say not (, k)-
indiscernible set, = (x
0
, . . . , x
k1
, c), then possibly permuting the variables of
the formula (x, y,
b
0
, . . . , b
m1
, b
n2
, b
2+1
, . . . , b
2+k,m3
, c) linear orders a
+i
:
i < ), hence has the order property. So assume b
i
: i < ) is an indiscernible set
over c, and let a
i
be the rst element of the sequence a
i
. If b
2i+1
: i < 4) is not
an indiscernible sequence over c a
2
then, by the indiscernibility of a
i
:
i < ) over c, we can nd a formula (x, y, c
), c
c a
i
: i < or 3 i such
that [= [b
2
, a
+2i+1
, c
] for i < but [= [b

2
, a
2+2i+1
, c
] for i < and we are

done. So assume b
2i+1
: i < 4) is an indiscernible sequence over c a
2
, hence
all a
2j
: j < 4 realizes the same type over b
2i+1
: i < 4 c hence for j < 2
we can nd a
2j
realizing tp( a
2j
, b
2i+1
: i < 4 c, C) and the rst element of a
2j
is a
0
. This contradicts the choice of (x, y, c) as having the order property with
g( y) minimal as we can move a
0
to c.
1.37

1.38 Remark. Note that for indiscernible sets, the theorems on dimension in [Sh:c,
III] holds for theories T with the dependence property, see 3.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

24 SAHARON SHELAH
2 Characteristics of types
We continue to speak on canonical bases and we deal with the characteristics of
types and of indiscernible sets. More elaborately, for any indiscernible sequence

b =
b
t
: t I), I an innite linear order, we have a measure Ch(
b) = Ch
( x, y
)
(
b) :
( x, y
) L
(T)
) with x = x
i
: i < m), m = g(
b
t
) for t J, where Ch
( x, y
)
(
b)
measure how badly ( x, y
) fail to be in stfor(
b) (see Denition 2.5), we can nd

such

bs with maximal such Ch(
b) and wonder what can we say about them. The

case of an indiscernible set is covered by 1.
2.1 Hypothesis. T has the dependence property.
2.2 Denition/Claim. Let

b =
b
t
: t I) be an innite indiscernible sequence,
k < . Then
(a) (Claim) if t
i
I and i < j t
i
<
I
t
j
for i < j < and
b
k
=
b
t
ki
b
t
ki+1
. . .
b
t
ki+k1
: i < ) then
() Cb(
b) = Cb(
b
1
) Cb(
b
k
),
() if
( x
1
, . . . , x
k
; y) = ( x
, y) then:
( x
1
, . . . , x
k
; y) stfor(b
k
) i ( x
; y) stfor(
b)
() if

b
k,1
,

b
k,2
are related like

b
k
above to our

b then Cb(
b
k,1
) = Cb(
b
k,2
)
(even for the local version this is true)
() if ( x, y) dpfor(
b),
( x
1
, . . . , x
k
, y) = ( x
, y) & ( x
m
, y)
or
( x
1
, . . . , x
k
; y) = (( x
, y) ( x
m
, y)) then
stfor(b
k
)
(b) (Denition) let Cb
k
(
b) = Cb(
b
k
) and Av
k
(
b, C) = Av
stfor
(
b
k
, C)
for any

b
k
as above
(c) (Denition) Cb
b) = Cb
k
(
b) : k <
(d) (Fact) if I
1
, I
2
are innite subsets of J and

b =
b
t
: t J) is an
indiscernible sequence (recall J linear order) then
Cb
b I
1
) = Cb
b I
2
)
(e) (Fact) if the innite indiscernible sequences

b
1
,

b
2
are nb-s, then
Cb
b
1
) = Cb
b
2
) for
(f) (Denition) if p S
m
(B) is nitely satisable in
m
B and
then let Cb
(p) = Cb
b): for some ultralter D on

m
B
satisfying Av(B, D) = p, the sequence

b is a
(D, B)-indiscernible sequence.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

Proof. Easy.
Recall: non-forking is quite worthwhile for stable theory; and have several equiv-
alent denitions. There is worthwhile generalization for simple theories; but of
course, not all denitions for stable theories give it for simple T. We may look for a
generalization of non-forking for dependent theories (see alternatives in [Sh 783]).
2.3 Denition. For . We say p S
m
(A) does not -fork over B A, if for
every model M A for some q S
m
(M) extending p we have Cb
(q) acl
C
eq(B).
Similarly we say that C/B does not -fork over A B if c C tp(C, B) does
not -fork over it.
2.4 Remark. Assume that T is a simple theory,

b =
b
t
: t I) is an innite
indiscernible sequence. Then we cannot nd a
n
: n < ) indiscernible sequence,
( x, a
n
) : n < ) pairwise contradictory (or just m-contradiction for some m) and
n<
(
t I)((
b
t
, a
n
)).
Proof. Assume toward contradiction that , a
n
: n < ) form a counterexample.
By thinning and compactness without loss of generality the set I
n
= t I : C [=
[b
t
, a
n
] are pairwise disjoint, and each is a convex subset of I.
Now we can repeat and get the tree property. More fully, for any cardinals >
we consider J =

as a linearly ordered set, ordered lexicographically and for

>
let J
= J : ; without loss of generality let I

0
I has order
type and let h : I J be order preserving. We can nd c
C for J such
that c
: J) is an indiscernible sequence satisfying t I c

h(t)
=

b
t
. By
compactness, for each < we can nd a
:

) such that:
() ( x, a
) :

) are pairwise contradictory (or just any m of them)
() J
,

C [= [ c
, a
].
Now ( x, a
) :
>
) exemplied the tree property.
2.4
We have looked at indiscernible sequences which are stable. We now look after
indiscernible sequences which are in the other extreme.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

26 SAHARON SHELAH
2.5 Denition. 1) For

b =
b
t
: t I) an indiscernible sequence, we dene its
character
Ch(
b) = Ch
( y, z)
(
b) : ( y, z) L
(T)
)
where
Ch
( y, z)
(
b) = Maxn :for some c the sequence TV((
b
t
, c)) : t I)
change sign n times (i.e. I is divided to n + 1 intervals)
(so it is 2k
( y, z)
).
2) For p S
m
(A), let
(a) CH(p) = Ch(
b) :

b is an innite indiscernible sequence such that every

b
t
realizes p
(b) for a formula = ( x
0
, . . . , x
k1
) let CH(p, ( x
0
, . . . , x
k
)) = Ch(
b) :

b =
b
t
: t I) is an innite indiscernible sequence such that t
0
<
I
t
1
<
I
. . . <
I
t
k1
C [= [
b
t
0
, . . . ,
b
t
k1
] and each

b
t
realizes p
(c) CH
max
(p) = n CH(p) : there is no bigger such n
CH(p), when n
is bigger than n mean ()(n
) & ()(n
< n
)
(d) CH
min
(p, ( x
0
, . . . , x
k1
)) = n CH(p, ( x
0
, . . . , x
k1
)): there is no
smaller n
CH(p, ( x
0
, . . . , x
k1
)).
Note: for the trivial , CH(p, ) = CH(p) hence CH
max
(p, ) = CH
max
(p).
2.6 Remark. 1) Instead of counting the number of interchanges of signs we can look
at
(a) ch
( y, z)
(
b) = Maxn: for some sequence c, if J is a conite subset of I

then the sequence TV((
b
t
, c)) : t J) change signs n times and then
dene cH(p), cH
max
(p), etc.
2) Alternatively use
(b)
( y, z)
(
b) =
>
2 : C [= ( y)
<g()
(
b
t
, y)
()
for t
0
< . . . < t
g()1
and/or other variants.

2) Clearly 2.5(2) is a try to get maximal, most general extensions of p (as
non-forking is for stable T)
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

2.7 Claim. Let p S
m
(A) be non-algebraic, x = x
: < m).
1) If n = n
: = ( x, y)) CH(p), then there is n
CH
max
(p) such that
n n
.
2) CH
max
(p) is non-empty.
3) If CH(p, ) ,= then CH
min
(p, ) ,= and CH
max
(p, ) ,= .
Proof. Let R, < be an n-place and 2n-place respectively predicate not in
T
and let
p
= Th(C
T
, c)
cA
( x)[R( x) ( x, c)] : ( x, c) p
( x
0
, . . . , x
k1
)(
<k
R( x
) &
1
<
2
x
1
,= x
2
) : k <
x < y R( x) R( y)
linearly ordered x : R( x)
( x
1
), . . . , ( x
k
)( y
1
) . . . ( y
k
)( x
1
< x
2
< . . . < x
k
& y
1
< . . . < y
k
( x
1
, . . . , x
k
, c) ( y
1
, . . . , y
k
, c) :
c A and L
(T)
and k <
(with x
i
= x
i,
: < m)). For n = n
( x, y)
: ( x, y) L(
T
)) and =
i
( x, y
i
) :
i < [T[) listing the formulas for L
(T)
let
n,
=
n
i
,
i
: i < [T[ where
n,( x, y)
= ( y)( x
0
, . . . , x
n
)[ x
0
< x
1
< . . . < x
n
&
<n
(( x
, y)) ( x
+1
, y)].
Now easily
(a)
p
is a consistent type (using p being non algebraic and Ramsey theorem)
(b) if n n
then
n,

n
,
(c)
p

n,
is consistent i ( n
)[ n n
CH(p)]
(d) if J is a directed (partial) order, n
t
= n
t,( x, y
)
: ( x, y
) L
(T)
)
CH(p, ) increases with t J and n
= n
( x, y
)
: (x, y
)) and
3
n
( x, y
)
=
maxn
t,( x, y
)
: t J, then n
CH( p, )
(e) like (d) inverting the order.
3
well dened as T is dependent
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

28 SAHARON SHELAH
Together we can deduce the desired conclusions.
2.7
2.8 Question: For p S(A) (or just p S(M)), does indiscernible sequences

b of
elements realizing p such that Ch(
b) CH
max
(p), CH(
b) CH
min
(p, ) play a
special role?
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

3 Shrinking indiscernibles
For stable theories we know that if
b
t
: t I) is an indiscernible set over A and
c
>
C then for some J I we have [J[ [T[ and
b
t
: t IJ) is an indiscernible
set over A c
b
t
: t J. We look more closely at the generalization for theories
with the dependence property (continuing [Sh:c, II]).
The case of indiscernible sets is easier so we delay it.
3.1 Notation. For a linear order I, let comp(I) be its completion.
3.2 Claim. If

b =
b
t
: t I) is an indiscernible sequence over A and c
>
C
(so nite), then
(a) there is J
comp(I), [J
[ [T[ such that

() if n < and s,
t
n
I, s
J

t (i.e., s,
t realize the same quantier

free type over J
in the linear order comp(I)) then a

s
= a
s
: <
n), a
t
= a
t
: < n) realize in C the same type over A c

(b) if we x n and deal with -types we can demand [J
[ < k
,n
<
(c) if in addition

b is an indiscernible set, then in () of clause (a) we can
weaken s
J

t to (, k)[(s
= s
k
t
= t
k
) & s
=
t
]
(d) if we replace c by C C in (a) we just use [J
[ [C[ +[T[.
Proof.
(a) by (b)
(b) follows by Claim 3.4 below
(c) similarly
(d) follows.
3.2

The reader may restrict himself in 3.3, 3.4 to the case n = 2 so a is just an
indiscernible sequence; this suces for 3.2.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

30 SAHARON SHELAH
3.3 Denition. 1) For L
(T)
, a linear order I, m
, n ,
an ordinal
for < n, a model M and a set A M, we say that a = a
u,,
: < n, u
[I]
, <
=
|u|
) is (, m
)-indiscernible over A of the
: < n)-kind if the

following holds:
() if m < 1 + m
, I [= t
0
< < t
m1
, I [= s
0
< . . . < s
m1
and for every
v m we let u
v
= t
: v, w
v
= s
: v then a
u
v
,,
: <
n, v [m]
, <
) and a
w
v
,,
: n, v [m]
, <
) realizes the same

-type over A in M.
2) If we omit we mean all rst order formulas, if we omit m
we mean . Also
in a
u,,
we may omit (because it is [u[). Of course nothing changed if we allow
a
u,,
to be a nite sequence (with length depending on (, ) only) but we can
instead increase
.
3) We add above J where J I (or J is included in the completion of I) if in ()
we demand (r J)
(r < t
r < s
& r = t
r = s
& t
< r s
< r).
We say almost above J if we add J t
: < n = .
3.4 Claim. 1) Assume
(a) is a nite set of formulas
(b) M a model of T and A M
(c) a = a
u,k,
: < n, k < k
, u [I]
) is indiscernible over A
(d) c
>
M.
Then there is a nite subset J of I or of the completion comp(I) of I such that
a
u,k,
: < n, k < k
, u [I]
) is -indiscernible over A

d above J.
2) Moreover, there is a bound on [J[ which depend just on , k
: < n) (and T),

and so it is enough that a is
1
-indiscernible for appropriate nite
1
.
3) So for every C C there is J comp(I) of cardinality [C[ +[T[ such that a
is indiscernible above J over A C.
Proof. 1) Toward contradiction assume that the conclusion fails. Let m
be the
maximal number of free variables in members of times n. Without loss of gen-
erality I is a complete linear order with a rst and a last element. For every nite
J I we choose t
J
: < m
J
), s
J
: < m
J
) such that:
(i) m
J
m
(ii) t
J
0
< t
J
1
< . . . < t
J
m
J
1
and s
J
0
< s
J
1
< . . . < s
J
m
J
1
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

(iii) for at least one m < m
J
we have t
J
m
, s
J
m
/ J (actually follows from the rest)
(iv) t
J
: < m
J
), s
J
: < m
J
) exemplify that J is not as required
(v) m
J
is minimal.
Let

b
0
J
= a
J
u,k,
: < n, k < k
, u [t
J
0
, . . . , t
J
m
J
1
]
) and

b
1
J
= a
J
u,k,
: < n, k <
k
, u [s
J
0
, . . . , s
J
m
J
1
]
).
So clearly
()
1
the -types of c
b
0
J
, c
b
1
J
over A are dierent
[why? by their choice].
For J [I]
<
0
let t
J
= s
J
m
J
for = m
J
, m
J
+ 1, . . . , 2m
J
1. Let D
be an
ultralter on [I]
<
0
such that t I J : t J [I]
<
0
D
.
As m
J
m
<
0
, and D
is an ultralter, clearly for some m() m
we have
Y
0
= J [I]
<
0
: m
J
= m() D
. For < 2m(), let

I
1
= t I : J Y
0
: t
J
<
I
t D
I
0
= t I : J Y
0
: t
J
= t D
I
1
= t I : J Y
0
: t <
I
t
J
.
Clearly I
1
, I
0
, I
1
) is a partition of I, I
0
is a singleton or empty, I
1
is an initial
segment of I and I
1
is an end segment of I. As I is complete there is t
I such
that I
0
,= I
0
= t
and t I : t
I
t
I
1
and t I : t

I
t I
1
. So
there are functions g, h : 0, . . . , 2m() 1 0, . . . , 2m() 1 1, 0, 1 such
that for each , k < 2m() we have
J Y
0
: t
J
I
h(,k)
k
D
and
g(, k) = 1 J Y
0
: t
J
<
I
t
J
k
D
g(, k) = 0 J Y
0
: t
J
= t
J
k
D
g(, k) = 1 g(k, ) = 1.
For any , k < 2(m()) if there is t satisfying t
<
I
t <
I
t
k
then choose such t
,k
.
Now let
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

32 SAHARON SHELAH
Y

= J Y
0
:for every , k < 2m() we have t
J
I
h(,k)
k
,
and t
J

I
if I
I
1
. If none of them hold, then I
= t
, <
= and
clearly
0
if <
= then in (I
h(,i)
, <
) there is no last element and t I

h(,i)
J :
t <
t
J
I
h(,i)
1
e = (, k) : , k < 2m(), I
k
,= is an equivalence relation and
ek <
=<
k
& t
= t
2
let w = : I
= t
then ek w k w & t
= t
3
for each J
0
Y

, the set J Y

: if ek, / w then t
J
0
k
<
t
J
belongs to
D
.
We now choose J
n
Y

by induction on n such that ek & / w t
J
n
k
<
t
J
n+1
.
Now
()
2
if i() < ,
i()
2 then the types of

b
0
J
0
b
0
J
1
. . .
b
0
J
i()1
and of
b
(0)
J
0
b
(2)
J
1
. . .
b
(i()1)
J
i()1)
over A are equal.
[Why? By the indiscernibility, see Denition 3.3.]
Now by clauses (iv) + (v) in the choice of t
J
, s
J
(for < m
J
= m()) for each i

there is
i
and

d
i
A such that C [=
i
[ c,
b
0
J
i
,

d
i
]
i
[ c,
b
1
J
i
,

d
i
].
Now
()
3
the sequence
i
( y,
b
0
J
i
,

d
i
) : i < ) of formulas is independent.
[Why? For each i() < and
i()
2 we need to prove that C [=
( y)[
i<i()
i
( y,
b
0
J
i
,

d
i
)
(i)
]. Now by ()
2
it is enough to prove that C [=
( y)[
i<i()
i
( y,
b
(i)
J
i
,

d
1
i
)].
But c exemplies the satisfaction of this formula so ()
3
holds.]
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

As is nite, one appears as
n
for innitely many ns (though not necessarily
with the same

d
n
as we allow A to be innite), so we get contradiction to T has
the dependence property.
2) Similar.
3) Follows.
3.4
3.5 Claim. Assume a
= a
t
: t I
) is an indiscernible sequence (with the

linear order) I
of conality > [T[ for = 1, 2. Then we can nd s
for
= 1, 2, < such that a
1
s
1
a
2
s
2
: < ) is an indiscernible sequence with

s
: < ) being <

I
-increasing unbounded in I
for = 1, 2.
Proof. Easy by 3.3, just choose s
1
i
, s
2
i
by induction on i.
3.5
See more in 4.11.
As 3.5 deal with = L
(T)
, we can derive the parallel result for nite L
(T)
.
3.6 Conclusion 1) Assume
()
b
t
: t I) is an indiscernible sequence over A, comp(I) the completion of
I.
For every C C there are n
( x, y)
: ( x, y) L
(T)
), a sequence of nite numbers,
J comp(I) of cardinality [C[ +[T[ and J
( x, y, c)
: L
(T)
), J
( x, y, c)
a nite
subset of J such that:
()
0
if J
is an initial segment of J including J

( x, y, c)
: L
(T)
, c C
then

b (JJ
) is an indiscernible sequence over A C
b
t
: t J
()
1
for every a
g( y)
A and = ( x, y, c), c C there are n n
( x, y)
and
t
1
< . . . < t
n
from J
such that if r, s It
1
, . . . , t
n
and m [1, n]
(s <
I
t
m
) (r <
I
t
m
) then [= [
b
s
, a] [
b
r
, a]
()
2
for every k < , a
g( y)
A, c
g( z)
A and = ( x
1
, . . . , x
k
, z, y) there
are n n
and t
1
< . . . < t
n
from J
such that if s
1
<
I
. . . <
I
s
k
and r
1
<
I
. . . <
I
r
k
are from J and m [1, n] & [1, k] (s
<
I
t
m
r
<
I
t
m
) & (t
m
<
I
s
t
m
<
I
r
) then [= [
b
s
1
, . . . ,
b
s
k
, c, a]
[
b
r
1
, . . . ,
b
r
k
, c, a].
2) Assume
()
3

b
u,,
: < n, u [I]
, <
) is indiscernible over A and
< for < n

(and n < ).
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

34 SAHARON SHELAH
For every c there are J I, [J[ [T[ and nite J
J for L
(T)
such that the
parallel of ()
1
, ()
2
hold.
Proof. 1) This restates 3.2, 3.4.
2) Similar.
3.7 Question: If < (= (x, y, c)) is a partial order with innite increasing sequences,
we may consider -directed subsets, = cf() > [T[, they dene a Dedekind cut.
What about orthogonality of those?
3.8 Conclusion. 1) Assume
b
t
: t I) is an indiscernible set over A. For every B
there is J I such that [J[ [T[ + [B[ and
b
t
: t IJ) is an indiscernible set
over A B.
Proof. Easy.
3.9 Claim/Denition. Assume p S
m
(B) is a stable type and B M. Then
dim(p, M) = [I[ +[T[ for any I a maximal indiscernible set
m
M base on p is well
dened.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

4 Perpendicular endless indiscernible sequences
Dimension and orthogonality play important role in [Sh:c], see in particular
Ch.V. Now, as our prototype is the theory Th(Q, <), it is natural to look at co-
nality, this is dual-cf(
b, A) dened below (4.5(3)), measuring the conality of

approaching

b from above (here

b is always indiscernible sequences with no last
member). So a relative of orthogonality which we call perpencidularity suggest
itself as relevant. It is dened in 4.5, as well as equivalence and dual-cf. Now
perpendicularity is closely related to mutual indiscernibility (see 4.7(1), 4.11(2)),
hence if T is unstable, then there are lots of pairwise perpendicular indiscernible
sequences: if a
: < ) is an indiscernible sequence, not an indiscernible set and
= a
+n
: n < ) for < then
: < are pairwise perpendicular.

In this section we present basic properties of perpendicularity. In particular, it
is preserved by equivalence (4.11(5)). For perpendicular sequences, we can more
easily restrict them to get mutually indiscernible sets than in 3. In particular we
show that if cf(Dom(
b
1
)) ,= cf(Dom(
b
2
)) then

b
1
,

b
2
are perpendicular.
But for indiscernible sets perpendicularity does not become orthogonality, in fact
it is trivial (see 4.15).
The case of looking at more than two indiscernible sequences reduced to looking
at all pairs (4.14(2), 4.17(2)). Also, as in [Sh:c, V], if

b is not perpendicular to a
for <
and the a
-s are pairwise perpendicular then
< [T[
+
(see 4.19).
Lastly, we recall (from [Sh:c]) the density of types not splitting over small sets
(for theories with the non independence property), hence the existence of a quite
constructible model over any A.
We think
4.1 Thesis: First order T with the dependence property is somewhat like the theory
of the rational order (or real closed elds).
If M is a model of (Q, <) and (I
, I
+
) : <
) is a sequence of pairwise
distinct Dedekind cuts of M, and N
is a dense linear order for <
and N is
M when in the cut (I
, I
+
) we insert N
, then M N; so we have total freedom

of what we put in the cuts.
In the next section we shall prove that if a
: <
) is a sequence of pairwise
perpendicluar endless indiscernible sequences that we have quite a total freedom
in choosing dual-cf( a
, M) : <
), this is a parallel for the above property of

Th(Q, <).
4.2 Thesis: If

b
1
,

b
2
are endless indiscernible sequences, which are not perpendicu-
lar, then there is in a sense an inside denable function showing that I
= c M : c
realizes Av(
,

b
) has the same conality.

(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

36 SAHARON SHELAH
4.3 Hypothesis. T has the dependence property.
4.4 Denition. 1) We say the innite sequences

b
1
,

b
2
are mutually indiscernible
over A if

b
is an indiscernible sequence over
b
3
t
: t Dom(
b
3
) A for
= 1, 2. If we omit over A we mean A = .
2) We say that the family
: <
of sequences is mutually indiscernible over

A, if for <
the sequence

b
t
: ,= , <
, t Dom(
) A.
3) We say b
1
,

b
2
are mutually -indiscernible over A if b
is a -indiscernible
sequence over
b
3
t
: t Dom(
b
3
) A for = 1, 2. Similarly in part (2).
4.5 Denition. Let a
= a
t
: t I
) be an indiscernible sequence which is endless

(i.e., I
having no last element) for = 1, 2.

1) We say that a
1
, a
2
are perpendicular when:
() if

b
n
realizes Av(
b
k
m
: we have m < n & k 1, 2 or we have m = n &
k < a
1
a
2
, a
) for = 1, 2 then

b
1
,

b
2
are mutually indiscernible (see
4.4 above) where

b
n
: n < ) for = 1, 2.
We dene -perpendicular in the obvious way.
2) We say a
1
, a
2
are equivalent and write if for every A C we have Av(A, a
1
) =
Av(A, a
2
).
3) If a
1
A we let dual-cf( a
1
, A) = Min[B[ : B A and no c
>
A realizes
Av(B, a
1
); we usually apply this when A = M.
4.6 Example: M a model of Th(Q, <).
= b
n
: n <
) is an increasing sequence in M.
Then

b
1
,

b
2
are not perpendicular i they dene the same cut of M.
4.7 Claim. 1) If a
1
, a
2
are endless mutually indiscernible sequences, then they are
perpendicular.
2) Mutually indiscernible and perpendicular are symmetric relations.
3) On the family of endless indiscernible sequences, being equivalent is an equiva-
lence relation.
4) In Denition 4.5(1) in () there, to say for every such
n
: n < , = 1, 2)
and to say for some
n
: n < , = 1, 2) are equivalent.
5) If a
1
, a
2
are endless indiscernible sequences and -mutually indiscernible se-
quence then they are -perpendicular.
6) If a
1
, a
2
are endless indiscernible sequence, a
2
is indiscernible over a
1
then a
1
, a
2
are perpendicular.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

4.8 Remark. By 4.12 below, in () of Denition 4.5(1), for any set A we can
add:

b
1
,

b
2
are mutually indiscernible over A; that is for any A, if

b
n
realizes
Av(
b
k
m
: m < n & k 1, 2 or m = n & k < a
1
a
2
A, a
) then

b
1
,

b
2
are mutually indiscernible over A where

b
n
: n < ) for = 1, 2.
Proof. 1) Let

b
n
for 1, 2, n < be as in Denition 4.5(1).
Now we prove by induction on k < that
()
k
the sequences a
1,k
= a
1
b
1
k1
, . . . ,
b
1
0
), a
2,k
= a
2
b
2
k1
, . . . ,
b
2
0
) are mutu-
ally indiscernible.
For k = 0 this is assumed. For k = m + 1, by the choise of

b
1
k
as realizing
Av( a
1,k
a
2,k
, a
1
) clearly a
1,k+1
, a
2,k
are mutually indiscernible. Similarly by the
choice of

b
2
k
, clearly a
1,k+1
, a
2,k+1
are mutually indiscernible.
Now the statement
b
1
n
: n < ),
b
2
n
: n < ) are mutually indiscernible is a
local condition, i.e., it is enough to check it for
b
1
n
: n < k),
b
2
n
: n < k) for each
k < , but this holds by ()
k
above.
2) Read the denition and rename.
3) Let a
1
, a
2
, a
3
be endless indiscernible sequences. Clearly Av(A, a
1
) = Av(A, a
1
)
so a
1
a
1
by Denition 4.5, so is reexive. Also Av(A, a
1
) = Av(A, a
2
)
Av(A, a
2
) = Av(A, a
1
) so is symmetric. Lastly, if a
1
a
2
and a
2
a
3
then for
any A we have Av(A, a
1
) = Av(A, a
2
) Av(A, a
2
) = Av(A, a
3
) hence Av(A, a
1
) =
Av(A, a
3
), as this holds for any A we can deduce that a
1
a
3
, i.e. is transitive.
So is really an equivalence relation.
4) Suppose that for i 1, 2 we have
b
i,
n
: n < , 1, 2) such that

b
i,
n
realizes Av(
b
i,k
m
: m < n & k 1, 2 or m = n & k < a
1
a
2
, a
). We can
choose an increasing sequence of elementary mapping f
n
(n < , < 2) such that
n
1
< n
2
(n
1
= n
1
1
<
2
) f
1
n
1
f
2
n
2
, f
0
0
is the identity on a
1
a
2
, Dom(f
1
n
) =
Dom(f
0
n
)
b
1,1
n
, Dom(f
0
n+1
) = Dom(f
1
n
)
b
1,2
n
, f
1
n
(
b
1,1
n
) =

b
2,1
n
, f
0
n+1
(
b
1,2
n
) =

b
2,2
n
. No
problem to carry the induction and f
= f
0
n
: n < can be extended to an
automorphism of C thus proving the claim.
5) Similar to (1).
6) Left to the reader (see 4.16).
4.7
4.9 Claim. Assume that for = 1, 2 we have:
()
(a) I
, J
are endless linear orders

(b)

b
t
: t I
) is an indiscernible sequence
(c) a
= a
t
: t J
) is an indiscernible sequence
(d) for s J
we have a
s
=

b
t(,s,1)
. . .
t(,s,n
)
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

38 SAHARON SHELAH
(e) t(, s, 1) <
I
t(, s, 2) <
I
. . . <
I
t(, s, n
)
(f) if s
1
<
J
s
2
then t(, s
1
, n
) <
I
t(, s
2
, 1).
0) If

b
is an indiscernible sequence, then so is a
.
1) If

b
1
,

b
2
are mutually indiscernible, then a
1
, a
2
2) Assume that t(, s, 1) : s J
is an unbounded subset of I
for = 1, 2. If
b
1
,

b
2
are perpendicular, then a
1
, a
2
are perpendicular.
3) Like part (2) for equivalent.
Proof. Just think.
(Concerning (2) see 4.11(4) below).
4.9
4.10 Claim. Assume that

b
t
: t I
) is an endless indiscernible sequence

of m
-tuples, so

b
t
= b
t,m
: m < m
). Assume that u
0, . . . , m
1 and
a
t
=

b
t
u
for t I
, = 1, 2.
1) If
b
1
t
: t I
1
),
b
2
t
: t I
2
) are mutually indiscernible, then a
1
t
: t I
1
), a
2
t
:
t I
2
) are mutually indiscernible.
2) If
b
1
t
: t I
1
),
b
2
t
: t I
2
) are perpendicular, then a
1
t
: t I
1
), a
2
t
: t I
2
) are
perpendicular.
3) Of course, permuting, duplicating or renaming the indiscernible in b
t,m
: m <
m
), etc., also is O.K.

Proof. Easy.
4.11 Claim. 1) If a
= a
t
: t I
) is an indiscernible sequence for = 1, 2 and

[T[ < cf(I
1
), [I
1
[ < cf(I
2
), then for some end segments J
1
, J
2
of I
1
, I
2
respectively,
a
1
J
1
, a
2
J
2
are mutually indiscernible; similarly with over A.
1A) If is nite to deduce just -mutually indiscernible, we can omit [T[ < cf(I
1
).
2) If a
= a
t
: t I
) is an indiscernible sequence for = 1, 2 and cf(I

1
), cf(I
2
)
are innite and distinct then a
1
, a
2
are perpendicular.
3) Assume that a
= a
t
: t I
) is an endless indiscernible sequence for = 1, 2,

is limit ordinal and

b
realizes Av(
b
k
: < & k 1, 2 or = &

k < a
1
a
2
, a
) and

b
: < ) for = 1, 2. Then the following are

equivalent:
() a
1
, a
2
are perpendicular
()

b
1
,

b
2
are perpendicular.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

4) If a
= a
t
: t I
) is an endless indiscernible sequence and J
is unbounded
for = 1, 2, then a
1
, a
2
are perpendicular i a
1
J
1
, a
2
J
2
are perpendicular.
5) If a
= a
t
: t I
) are an endless indiscernible sequence for = 1, 2, 3, 4

and a
1
, a
3
are equivalent and a
2
, a
4
are equivalent, then a
1
, a
2
are perpendicular i
a
3
, a
4
are perpendicular; so perpendicularity of a
1
, a
2
depend just on a
1
/ , a
2
/ .
Proof. 1) By 3.6(1) applied to A = a
1
t
: t I
1
and a
2
, there is an end segment
J
2
of I
2
such that a
2
J
2
is an indiscernible sequence over A. Let J
2
be a countable
subset of J
2
and apply 3.6(1) to A
= a
2
t
: t J
2
and a
1
, so there is an end
segment J
1
of I
1
such that a
1
J
1
is an indiscernible sequence over A
. Reecting
on the meaning clearly a
1
J
1
, a
2
J
2
1A) Similar to (A); without loss of generality is closed under permuting and iden-
tifying the variables, and we use the relevant variant of 3.6(1) or just 3.4.
2) Without loss of generality cf(I
1
) < cf(I
2
). It is enough for every formula
= ( x
1
, . . . , x
m
, y
1
, . . . , y
k
) with g( x
) = g( a
1
t
) and g( y
) = g( a
2
t
) to show
that for some t
1
I
1
, t
2
I
2
:
for some truth value t, for all t
1
<
I
1
t
1
1
<
I
1
. . . <
I
1
t
1
m
, t
2
<
I
2
t
2
1
<
I
2
. . . <
I
2
t
2
k
we have [= [
b
1
t
1
1
, . . . ,
b
1
t
1
m
,
b
2
t
2
1
, . . . ,
b
2
t
2
k
]
t
.
By (1A) and 4.7(5) this is easy (as for each nite we can use suitable end
segments).
3) If a
1
, a
2
are perpendicular then by denition 4.5(1) we have

b
1
,

b
2
are mutually
indiscernible and by 4.7(1) this implies that

b
1
,

b
2
are perpendicular. The other
direction is even easier, though we have to use 4.7(4).
4) Let

b
1
n
,
b
2
n
(let n < be as in () of Denition 4.5(1)) for a
1
, a
2
. Now for every set
A, the types Av(A, a
), Av(A, a
) are equal (see 1.10(2)). As ( a

1
J
1
) ( a
2
J
2
) is included in ( a
1
a
2
) clearly
b
1
n
: n < ),
b
2
: n < ) are as required in ()
for a
1
J
1
, a
2
J
2
.
By 4.7(4) we have a
1
, a
2
are perpendicular i
b
1
n
: n < ), b
2
n
: n < ) are mutually
indiscernible i a
1
J
1
, a
2
J
2
are perpendicular; so we are done.
5) We choose
n
by induction on 2n+ for n < , 1, 2 as any sequence realizing
p
n
= Av( a
1
a
2
a
3
a
4

b
k
m
: m < n & k 1, 2 or m = n & k < , a
).
So a
1
, a
2
are perpendicular i
b
1
n
: n < ),
b
2
n
: n < ) are mutually indiscernible
(by 4.7(4)).
Now by the assumption (on the equivalence) the type p
n
is also equal to Av( a
1
a
2
a
3
a
4

b
k
m
: m < n or m = n & k < , a
2+
).
Using again 4.7(4) we have: a
3
, a
4
are perpendicular i
b
1
n
: n < ),
b
2
n
: n < )
Together we get the desired conclusion.
4.11
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

40 SAHARON SHELAH
4.12 Claim. Let a
1
, a
2
be endless indiscernible sequences. The following are equiv-
alent:
(A) a
1
, a
2
are perpendicular
(B) there are A,
b
1
n
,
b
2
n
(for n < ) such that
(a) a
1
a
2
A and

b
n
realizes Av(A a
1
a
2
b
k
m
: n < n & k 1, 2
or m = n & k < , a
) and
(b) the sequences

b
1
=
b
1
n
: n < ),

b
2
=
b
2
n
: n < ) are mutually
indiscernible
(C) for every A a
1
, a
2
and

b
1
realizing Av(A, a
1
) and

b
2
realizing Av(A, a
2
)
the sequence

b
1
realizes Av(A
b
2
, a
1
)
(D) if A a
1
, a
2
and

b
1
n
,
b
2
n
are as in clause (B)(a) then
b
1
n
: n < ),
b
2
n
: n <
) are mutually indiscernible over A.
Proof.
(D) (B):
We can nd A,
n
for n < , 1, 2 which are as in clause (B) except possibly
the mutual indiscernibility in the end, i.e., as in () of Denition 4.5(1) but with
A a
1
a
2
. By clause (D) this suces.
(C) (D):
Now we can prove by induction on i < that:
if m
1
< m
2
< n
1
< . . . < n
i
< and
1
, . . . ,
i
1, 2 then
b
1
m
1
b
2
m
2
1
n
1
. . . b
i
n
i
and

b
1
m
2
b
2
m
1
1
n
1
. . .
i
n
i
realized the same type over A
n
: n < m
1
,
1, 2.
[How? For i = 0 as we are assuming clause (C), for i +1, because the type
tp(
i+1
n
i+1
, A b
n
: n < n
i+1
, 1, 2) does not split over a
i+1
by the
denition of Av.]
By transitivity of equality of type from we can prove that
b
1
2n+1
: n < ),
b
2
2n+2
:
n < ) are mutually indiscernible over A; as in 4.7(4) this suces.
(B) (A):
If A,
n
for n < , 1, 2 are as in clause (B), then as [B
1
B
2

Av(B
1
, a
) Av(B
2
, a
)] they are as in () of Denition 4.5. So by 4.7(4) the

sequences a
1
, a
1
are perpendicular, i.e., clause (A) holds.
(A) (C):
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

Assume that a
1
, a
2
are perpendicular but clause (C) fails for the
4
set A. So
a
1
, a
2
A, let
0
.
Choose

b
for <
+
, 1, 2 by induction on such that

b
realizes Av(A
b
k
: < ( = & k < , a
). Easily by the choice of A for some c A

and ( x, y, c) we have C [= [
b
1
b
2
, c] i . By the mutual indiscernibility

of
b
1
: < ),
b
2
: < ) (which holds as a

1
, a
2
are perpendicular) the set
(
b
1
b
2
, z) : < of formulas is independent, contradiction.

4.12
4.13 Claim. Assume a
= a
t
: t I
) are endless indiscernible sequences for

= 1, 2.
1) If a
1
is an indiscernible sequence over A, then: a
1
is an indiscernible set over
A i a
1
is an indiscernible set over .
2) Assume that a
1
is an innite indiscernible sequence over A, then: a
1
is non-
stable in C i a
1
is non-stable in (C, c)
cA
.
3) If a
1
, a
2
are equivalent, then a
1
is non-stable i a
2
is non-stable.
4) Assume that for = 1, 2, J
is convex and innite and a

1
, a
2
are mutually
indiscernible then a
1
J
1
, a
2
J
2
are mutually indiscernible over
2
=1
( a
(I
)).
5) Let k 1, 2. In part (4) we can omit J
k
a convex (subset of I
k
) if a
k
is an
indiscernible set.
Proof. 1) By 1.28.
2) Follows.
3), 4), 5) Check directly.
4.13
4.14 Claim. 1) Assume a
1
, a
2
are endless indiscernible sequences. If a
1
, a
2
has
conality > [T[ and are mutually indiscernible and

b
>
C, then for some end-
segments J
1
, J
2
of Dom( a
1
), Dom( a
2
) respectively a
1
J
1
, a
2
J
2
are mutually
indiscernible over

b.
1A) Like (1) for mutual -indiscernibility, when
2
is nite,
1
nite large enough,
a
1
, a
2
just endless.
2) Assume

b
1
,

b
2
,

b
3
are endless indiscernible sequences A and I is an innite
linear order and a
t
realizes Av( a
k
s
: s <
I
t & k 1, 2, 3 or s = t & k <
A,

b
) for 1, 2, 3 and t I and let a
= a
t
: t I), then:
(a) a
1
t
a
2
t
a
3
t
: t I) is an indiscernible sequence over A;
4
the choice of

b
1
,
b
2
of course is immaterial as: if

b
realizes Av(A, a
1
) and

b
2
realizes
Av(A +
, a
2
) then

b
2
b
2
realizes the same type over A

(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

42 SAHARON SHELAH
(b) if a
1
is an indiscernible set then a
1
, a
2
are mutually indiscernible over A
(c) if any two of a
1
, a
2
, a
3
are mutually indiscernible and I
1
, I
2
, I
3
are disjoint
5
unbounded subsets of I, then a
1
I
1
, a
3
I
3
A ( a
2
I
2
)
(d) if a
1
, a
2
are mutually indiscernible then they are mutually indiscernible over
A.
Proof. 1) No new point so left to the reader.
2) Without loss of generality I is dense with no rst, no last elements, and I is not
a complete even restricted to an interval and every interval has cardinality > [T[.
Now
Clause (a):
Easy as in 1.6(2).
Clause (b):
For any s
1
<
I
< . . . <
I
s
n1
, stipulating s
0
= , s
n
= + and letting I
=
t I : s
 [T[, by 3.8 this implies that a
1
is an indiscernible set
(see 4.13(1)) over a
2
s
1
. . . a
2
s
m1
; as this holds for any n < and s
1
<
I
. . . <
I
s
n1
,
we get that a
1
is an indiscernible set over a
2
. But for any t I the sequence
a
2
s
: s I & t
I
s) is an indiscernible sequence over a
1
s
a
2
s
: s <
I
t. So clearly
if s : s <
I
t is innite then by the last two sentences, a
2
s I : t
I
s, a
1
are mutually indiscernible (even over a
2
s
: s <
I
t). By the assumption on I in the
beginning of the proof we are done.
Clause (c):
Note that by the assumption on a
t
, it is enough for any pairwise disjoint I
I
for = 1, 2, 3, each as we assume in the beginning of the proof of part (2), to prove
that a
1
I
1
, a
2
I
2
, a
3
I
3
By transitivity of equality it is enough to prove:
() if (1) ,= (2) 1, 2, 3, t
1
< t
2
in I, then the sequences a
(1)
t
1
a
(2)
t
2
and
a
(1)
t
2
a
(2)
t
1
realizes the same type over A a
t
: 1, 2, 3 and (t
1

I
t
I
t
2
).
To prove () it suces for any n and s
1
<
I
. . . <
I
s
n
with t
1
<
I
t
2
<
I
s
1
and
k 0, 1, 2, 3 to prove that a
(1)
t
1
a
(2)
t
2
and a
(1)
t
2
a
(2)
t
1
realize the same type over
5
the disjointness can be omitted
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

A a
s
m
: m 1, . . . , n 1 & 1, 2, 3 or m = n & k a
s
: s <
t
1
, 1, 2, 3. We do it by induction on 4n + k.
For n = 0: The sequences a
(1)
, a
(2)
are perpendicular by 4.7(1) hence the required
conclusion holds by (A) (C) of 4.12.
For n + 1: If k = 0 this is known (being equivalent to the case (n
, k
) = (n, 3)),
otherwise this follows by the denition of average more exactly as tp( a
k
s
n+1
, A
a
s
m
: m 1, . . . , n & 1, 2, 3 or m = n + 1 & k a
s
: s < t
1
,
1, 2, 3) does not split over a
k
by 1.6(6).
Clause (d):
By 4.12.
4.14
4.15 Conclusion. If

b
1
,

b
2
are endless indiscernible sequences and

b
1
is an indis-
cernible set, then

b
1
,

b
2
are perpendicular.
Proof. By 4.14(2), clause (b).
4.16 Claim. 1) Assume a,

b are endless indiscernible sequences. Then a,

b are
perpendicular sequences, i for any ( x, y, c) for some truth values t we have:
(a) for every large enough s Dom( a), for every large enough t Dom(
b) we
have C [= [ a
s
,
b
t
, c]
t
(b) for every large enough t Dom(
b) for every large enough s Dom( a) we

have C [= [ a
s
,
b
t
, c]
t
.
2) For any = ( x, y, c), there is a truth value t = t
( x, y, c)
for which clause (a)
holds and there is a truth value t = t
( x, y, c)
such that clause (b) holds.
Proof. By 4.12(C) (and 1.10(2)).
Part (2) is easy too.
4.16
4.17 Claim. 1) The parallel of the relevant earlier claims holds for several indis-
cernible sequences, that is, assuming a
= a
t
: t I
) is an endless indiscernible
sequence for <
(A) If the intervals [cf(I
), [I
[] are pairwise disjoint, cf(I
) > [T[ +
, then
for some end segment J
of I
for <
, we have a
: <
)
is mutually indiscernible, which means: each a
is indiscernible over
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

44 SAHARON SHELAH
a
: <
& ,= (in fact we can get indiscernibility over

a
: <
& [I
[ < cf(I
) a
: (,
))
(B) Assume a
: <
) are mutually indiscernible,

b
>
C and I
= Dom( a
)
and cf(Dom( a
)) > [T[+
. Then there are end segments J
of I
for <
such that a
: <
) is mutually indiscernible over

b.
(C) If A is a set, J
is an innite linear order disjoint to I
: <
and a
t
realizes p
t
= Av(A a
s
: <
& s I
or = & s J
& s <
J
t
or s J
& < A, a
) for any <
, t J
then a
s
: s J
) : <
are mutually indiscernible over a
s
: <
, s I
A.
(D) If a
: <
) are pairwise perpendicular and J
= J for <
then in
clause (C), a
t
realizes q
t
= Av( a
s
: <
& s I or <
& s
J & s <
J
t or s = t & < A, a
).
(E) if J
= J is an innite linear order (disjoint to Dom( a
) : <
)
and a
t
for <
, t J
realizes the type q
t
from part (D), and for any
< <
, a
s
: s J
), a
s
: s J
) are mutually indiscernible or just

perpendicular, then a
s
: s J
) : <
) are mutually indiscernible,

moreover even over A.
2) If we weaken in the conclusion of clause (A) of part (1) the mutually indiscernible
by mutually -indiscernible, then we can weaken cf(I
) > [T[ +[
[ to cf(I
) > [
[.
Proof. Similar to earlier proofs (4.11).
4.17
4.18 Claim. 1) If a = a
t
: t I) is an indiscernible sequence,

b
>
C then
we can divide I to 2
|T|
convex subsets I
: <
) such that a I
: <
, I
innite) is mutually indiscernible over

b.
2) Similarly in 4.17: if a
is an endless indiscernible sequence over A for <
,
and they are mutually indiscernible and

b
>
C then we can nd w
, [w[ [T[
and for w a partition of I
to 2
|T|
convex sets I
,
: <
) such that the

family a
w a
I
,
: w, <
is mutually indiscernible over

A
b (the partition of I
is induced by some subset I
of comp(I
) of cardinality
[T[).
Proof. 1) By 3.6.
2) Similarly.
4.18
Generalizing another claim for stable theories:
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

4.19 Claim. Assume that
(a)

b, a
are endless indiscernible sequences for <
(b) a
, a
are perpendicular for ,=

(c)

b, a
are not perpendicular.

Then
< [T[
+
.
Proof. Assume toward contradiction that
[T[
+
. We let A =

b
and by
induction on n < , we choose a
,
n
: <
) and

b
n
and for a x n < we choose
a
,
n
by induction on <
and then we choose

b
n
such that:
(a) a
,
n
realized the average of a
over A
m
: m < n a
,
m
: m < n &
<
or m = n & <
(b)

b
n
realizes the average of

b over A
m
: m < n a
,
m
: m n, <
.
For each , as

b, a
are not perpendicular, we can nd n
< , u
[]
n
for
= 0, 1, 2 such that
n
: n u
0
) a
,
n
: n u
1
) and
n
: n u
0
) a
,
n
: n u
2
)
does not realize the same type; say one satises
( x, y) the second not. As we can

replace a
: < [T[
+
) by any subsequence of length [T[
+
, without loss of generality
< [T[
+
n
= n
, u
= u
= . Now for every U [T[

+
let f
U
be the
elementary mapping with domain a
,
n
: n u
1
, < [T[
+
, mapping a
,
n
1
to a
,
n
2
i U , n
1
= n
2
or [T[
+
U , n
1
u
1
, n
2
u
2
, [n
1
u
1
[ = [n
2
u
2
[. Let g
U
be
an automorphism of C extending f
1
U
. We have gotten the independence property
for ( x, y) as g
U
(
n
: n u
0
0
)) realizes ( x
n
: n u
0
), a
,
n
: n u
1
))
if(U)
:
< [T[
+
, contradiction.
4.19

We can deal with perpenducularity of ultralters instead of indiscernible sequences.
4.20 Denition. Let D
be an ultralter on
m()
(B
) for = 1, 2.
We say that D
1
, D
1
are perpendicular if:
() if

b
n
realizes Av(
m
: m < n or m = n k < B
1
B
2
, D
) for
n < , 1, 2, then
b
1
n
: n < ),
b
2
n
: n < ) are mutually indiscernible.
Parallel claims hold, e.g.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

46 SAHARON SHELAH
4.21 Claim. Let D
, B
, m
( = 1, 2) be as in the denition 4.20 above.

1) D
1
, D
1
are perpendicular i
() if A B
1
B
2
,
b
1
realizes Av(A, D
1
), and

b
2
realizes Av(A
b
1
, D
2
) then
b
1
realizes Av(A
b
2
, D
1
).
2) Let

b
t
: t I
) be endless indiscernible sequences, and let D
be an
ultralter on
t
: t I
containing
t
: t J for all the co-bounded subsets J
of I
, for = 1, 2. Then D
1
, D
2
are perpendicular i

b
1
,

b
2
are perpendicular.
3) In Denition 4.20 we can replace mutually indiscernible by perpendicular.
Proof. No new point.
We can translate:
4.22 Claim. 1) Assume we are given set B( C) and D is an ultralter on
m
B
and I is an endless linear order. Then for some ultralter D
on the cardinal
= [T[ + [B[, in C
/D
we can nd an indiscernible sequence

b = b
t
: t I) in
B
/D such that:
() if A
1
C, a
m
C then:
a realizes Av(A
1
, D) i a realizes Av(A
1
,

b) (in C
/D
) i every

b
t
realizes Av(A
1
, D) = Av(A
1
,

b) = tp( a, A, C).
2) If

b =
b
t
: t I) is indiscernible, I endless,

b
t

m
B for m < and B A,
then there is an ultralter D on
m
B such that () of part (1) holds.
Proof. Straightforward.

As background for the following note that for T a totally transcendental (=
0
-
stable), for every A C the set of isolated types in S(A) is dense (i.e. if C [=
(x)(x, a), a A then (x, a) belongs to some q S(A) which is isolated, i.e.
such that for some ( x, a
) q we have (x, a) q). This gives that we can extend

A to a model such that few types over A
are realized in it, so in some sense

M is understood over A ([Mo65] or see [Sh:c]). This enables us to preserve much
(e.g. respecting, see the next section), this is ne but the assumption makes it
irrelevant here.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

For stable T we can replace isolated by [T[
+
-isolated (see [Sh:c, IV]). Also we can
replace isolated by does not fork over some nite subset of A; this looks like an
opposite to being isolated (as a non-forking is an opposite to an isolated one) but
is still managable (and helpful, called F
f
0
-isolated). But all these seem under too
strong assumptions so irrelevant here. We use a substitute: does not split over a
small set, a precursor of non-forking (from [Sh 3]), still of interest when non-forking
is not available.
Recall ([Sh:c, Ch.III,7,IV])
4.23 Denition. 1) p F
sp
(B) if for some set A we have p S

<
(A), B
A, [B[ < and p does not split over B, see part (4) below. Let p F
sp
mean that
for some set B we have p F
sp
(B).
2) A = (A,
b
i
, B
i
: i < i
)) is an F
sp
-construction (or
b
i
, B
i
: i < i
) is an F
sp
-
construction over A) if tp(
b
i
, A
b
j
: j < i) F
sp
(B
i
), so B
i
A
A
i
=: A
b
j
:
j < i) for every i < i
.
3) Omitting B
i
means for some B
i
; let i
= g(A).
4) Recall that p S
m
(A) splits over B A if for some formula ( x, y) and
sequences

b, c from
g( y)
A realizing the same type over B we have ( x,
b), ( x, c)
p.
We give a proof of 4.24 for self-containment.
4.24 Claim. 1) If B A and p is an m-type over B, then there are q S
m
(A)
extending p and B
1
A, [B
1
[ [T[ such that q does not split over B B
1
.
2) For any A and > [T[ there is a model M and F
sp
-construction A = (A,
b
i
, B
i
:
i < i
)) such that:
(a) M = A
A
i
and |M| = [A[
<
+
<
2
|T|+
(b) M is -saturated,
(c) cf(i
) = or singular, cf(i
) =
+
.
3) If A is an F
sp
-construction, = cf(),

b A
A
g(A)
has length < , then tp(
b, A)
does not split over some B A, [B[ < .
4) In part (2) we can add (d) if we replace (a) by (a)
where
(d) if p S
m
(M) does not split over B M, [B[ < then i
= supi : B A
A
i
and

b
i
realizes the type p A
A
i

(a)
M = A
A
i
and |M| = [A[
<
+ 2
2
+|T|
.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

48 SAHARON SHELAH
Proof. 1) For any set C A let us dene
p
C
= p( x) ( x,
b) ( x, c) :( x, y) L
(T)
and

b, c
g( y)
A realizes the same type over B C.
Now if there is C A of cardinality [T[ such that p
C
is nitely satisable (in C),
then choosing B
1
= C and q S
m
(A) any extension of p
C
we are done. So assume
toward contradiction
() if C A has cardinality [T[, then p
C
is not nitely satisable.
Now we choose by induction on < [T[
+
a set C
and then a sequence

,n
,
b
,n
, c
,n
:
n < n
) such that
()
1

,n
=
,n
( x, y
,n
) L
(T)
and

b
,n
, c
,n
are sequences from A of length
g( y
,n
)
as follows. In stage we let C
b
,n
c
,n
: < and n < n
.
So C
A and [C
[ <
0
+[[
+
< [T[
+
; of course, C
0
= . Now by () we know that
p
C
is not nitely satisable, hence we can nd n
< and
,n
( x, y
,n
),
b
,n
, c
,n
as in ()
1
such that
()
2
b
,n
, c
,n
realizes the same type over B C
for n < n
()
3
p
,n
( x,
b
,n
)
,n
( x, c
,n
) : n < n
is not nitely satisable, that is

p( x)
n<n
(
,n
( x,
b
,n
)
,n
( x, c
,n
)).
Having carried the denition note that the number of possible sequences
,n
( x, y
,n
) :
n < n
) is [T[ hence for some unbounded U [T[

+
we have U n
=
n
&
n<n
,n
( x, y
,n
) =
n
( x, y
n
).
Now note
()
4
if p q S
m
(C
) then for some n < n
there are q
0
, q
1
S
m
(C
+1
)
extending q such that
,n
( x,
b
,n
) q
0
,
,n
( x,
b
,n
) q
1
.
[Why? Let q q
S
m
(C
+1
), now by ()
3
, as p q q
clearly for some n < n
we have [
,n
( x,
b
,n
)
,n
( x, c
,n
)] q
, and as

b
,n
, c
,n
C
+1
for some truth
value t we have
,n
( x,
b
,n
)
t
q
hence
,n
( x, c
,n
)
t
q
.
So q
,n
( x, c
,n
)
t
is nitely satisable, but by ()
2
the sequences

b
,n
, c
,n
realizes the same type over C
hence also q
,n
( x,
b
,n
)
t
is nitely satisable
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

hence can be extended to some q
S
m
(C
+1
). So q
, q
can serve as q
0
, q
1
(in
some order); so ()
4
holds.]
Hence
()
5
for any nite set u [T[
+
, the following set has at least 2
|u|
members
: is a function from (, n) : u, n < n
to the truth values such that

p
= p
,n
( x,
b
,n
)
(,n)
: u, n < n
is nitely satisable.
[Why? By induction on [u[ (or on sup(u)) using ()
4
.]
Now let =
n
( x; y
n
) : n < n
, so for every nite u U by ()

5
we have
S
m
b
,n
: u, n < n
) has at least 2
|u|
members
whereas
b
,n
: u, n < n
has at most u m
members where we let m
n<n
g( y
n
).
By [Sh:c, II,4], T has the independence property.
2) Let
be if is regular and
+
if is singular. We shall choose by induction
on
the tuple A
= (A, (
i
, B
i
) : i < i
)) such that:
(a) A
is an F
sp
-construction
(b) i
0
= 0
(c) if < then i
< i
and (
i
, B
i
) = (
i
, B
i
) for every i < i
, so we call
them (b
i
, B
i
)
(d) if is a limit ordinal then i
= i
: < and so A
is determined by
clause (c)
(e) [i
[ =: [A[
<
+
<
2
+|T|
.
If A
is chosen, let A
= A
b
i
: i < i
and let P
= p: for some m < , p is

an m-type over some set B A
of cardinality < hence of cardinality < . We

know that [P
[ and let i
+1
= i
+ let p
i
: i [i
, i
+1
)) list P
(possibly
with repetition). We choose (A
i
, B
i
, q
+
i
,
b
i
) by induction on i [i
, i
+1
) as follows.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

50 SAHARON SHELAH
Let A
i
= A
b
j
: j [i
, i); and let B

i
, q
+
i
be such that q
+
i
S
m
i
(A
i
) be an
extension of p
i
which does not split over B
i
, where Dom(p
i
) B
i
A
i
& [B
i
[ <
where p
i
is an m
i
-type. Why can we nd such B
i
, q
+
i
? by part (1) applied to A
i
, p
i
,
Dom(p
i
). Lastly, let

b
i

m
i
C be any sequence realizing q
+
i
.
So we have carried the induction on i [i
, i
+1
) hence A
+1
is dened. As the
case of limit and = 0 were done we have nished the induction on , so A
is
dened also for =
and A
is as required.
3) Let A = (A,
b
i
, B
i
: i 
(B
). For each i 
(A
b
i
: i u
0
), and
dened u
n
i
of cardinality < for n < by (u
0
as above and) u
n+1
= u
n
u
i
:
i u
n
. Let u
= u
n
: n < and B = A [B
i
: i u

b], so B [A]
<
.
Now we can prove by induction on i u
that tp
b
j
: j u
i, A) does
not split over B. From the case i = i
we can deduce the desired conclusion.

4) Like the proof of part (2) let P
= (p, B) : p S
<
(A
), p does not split over

some set B A
of cardinality < and

6
let (p
i
, B
i
) : i [i
, i
+1
)) listing P
i
j
.
But choosing

b
i
, B
i
for i [i
, i
+ ) we now have two cases.

Case 1: i = i
+ 2j.
As in the proof of part (2) using p
i
+j
where p
i
+j
is as in the proof of part (2).
Case 2: i = i
+ 2j + 1.
If there is q S
<
(A
i
: [i
, i) extending p
+j
not splitting over the
set B
+j
which has cardinality < , choose q
+
i
as some such q. If not, act as in
case 1.
4.24
Similarly, but if we like not to assume > [T[, we need to assume more on T.
4.25 Denition. 1) p F
esp
(B) if for some A, m we have p S

m
(A), B A, and
for every = ( x, y) for some L
(T)
and B
B both of cardinality < the

type p does not (, )-split over B
, see part (4) below.

1A) Let F
esp
(B) = F
esp
0
(B).
2) A = A, (
b
i
, B
i
) : i < i
)) is an F
esp
-construction or (b
i
, B
i
) : i < i
)) is an
F
esp
-construction over A if tp(
b
i
, A
b
j
: j < i) F
esp
(B
i
) so B
i
A
i
=:
A
b
j
: j |T|, {p S
m
(A) : p does not split over some subset of A
of cardinality < } is |A|
<
+
X
<
2
2
; similarly for strong splitting

(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

g( x) = m and
b, c
g( y)
B we have tp
2
(
b, A) = tp
2
( c, A) but ( x,
b), ( x, c)
p.
4.26 Claim. 1) Assume [T[ +
1
.
If B A, p is an m-type over B of cardinality < and [B[ < , then there are
B
[A]
<
extending B and q S
m
(A) from F
esp
(B
) extending p.
2) For any A and = cf() [T[ +
1
and is a model M and F
esp
-construction
A = (A, (
b
i
, A
i
) : i 
(A
A
g(A)
) and
( x, y) L
(T)
for some B A and L
(T)
of cardinality < the type
tp(
b, A) does not (, )-split over A.

4) The parallel of 4.24(4) holds.
The proof of 4.26 is similar to the proof of 4.24.
Proof. 1) Fix m and a m-type p over B such that B A, [B[ < . Without loss of
generality = [T[.
Let x = x
: < m) and let

i
( x, y
i
) : i < [T[ be a list of all such formulas.
For any set C A and L
(T)
we dene
q
i
,C
=
i
( x,
b)
i
( x, c) :
b, c
g( y
i
)
A
and tp
b, C) = tp
( c, C).
We now dene by induction on < [T[, a pair (C
) such that:
1
(a) C
A is increasing continuous
(b)
L
(T)
is increasing continuous
(c)
, C
are of cardinality
0
+[B[ +[[
(d) C
0
= B,
0
=
(e) p q
+1
,C
+1
: < is nitely satisable.
If we succeed, clearly we are done, and there is no problem for = 0 and limit.
So assume = +1. We now try to choose
of the right cardinality

(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

52 SAHARON SHELAH
close enough, e.g. of the form
= L
for some vocabulary

T
. We now try
to choose by induction on i <
1
, C
,i
, n
i
,
b
,i,
, c
,i,
for < n
i
such that:
2
() C
,i
is increasing continuous
() C
,0
= C
() n
i
<
()

b
,i,
, c
,i
,

g( y
i
)
A realizes the same type over C
,i
()

b
,i,
, c
,i,
c
,i
and C
,i+1
C
,i
is nite
() p p
,c
c
,i+1
: <
,i,
( x,
b
,i,
)
,i,
( x, c
,i,
) : < n
i
is
inconsistent.
For i = 0, i limit no problem. For i successor, if the choice C
+1
= C
,i
(and
+1
chosen above) is as required in
1
we are done choosing (C
) thus nishing
the proof. Otherwise p q
+1
,C
+1
: q
+1
,C
,i
is inconsistent hence
has a nite inconsistent subset p
,i
and let C
,i+1
= C
,i
Dom(p
,i
), let n
,i
=
[p
,i
q
+1
,C
,i
[ and
( x,
b
,i,
)
( x, c
,i,
) : i < n
list p
,i
q
+1
,C
,i
.
So for some n() the set U = i <
1
: n
,i
n() is innite. Now we prove
for every i() <
1
and u
1
i(), the following set has at least 2
|u|
members : is a function from (j, n) : j u and n < n
to the truth
values such that pq
+1
,c
+1
C
,i()
: <
( x,
b
,j,n
)
(j,n)
: j u
is nitely satisable.
We do this by induction on [u[; this gives that T has the independence property,
contradiction.
We would have liked to look at all =
0
, but we would get by the proof above
less; say for a pregiven k
0
< , say for = 0, we get every subset of p p
0
,C
0
of
cardinality < k
0
is satisable.
2) Similar to the proof of 4.24(2).
3) Like 4.24(3), only we have to take care of the , too.
4) Like 4.24(4).
4.26
4.27 Claim. If a
1
, a
2
are perpendicular indiscernible sequences each of conality
> [T[, then we can nd J
1
Dom( a
1
), J
2
Dom( a
2
) unbounded such that a
1
J
1
, a
2
J
2
Proof. If Dom( a
1
), Dom( a
2
) has dierent connalities we can apply 4.11 to a
1
I
1
, a
2
I
2
where I
Dom( a
) is unbounded and conal in Dom( a
) and has
cardinality cf(Dom( a
)).
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

Otherwise, let be the common conality, choose t
: < Dom( a
)
increasing unbounded. As in 3.5 we can chooose (i, ) < by induction on 2+,
increasing such that = (i, ) implies that a
realizes Av( a
k
t
k
: < & k
1, 2 or = k = 1 < = 2, a
). So a
1
t
1
a
2
t
2
: < ) is an indiscernible
sequence. By the perpendicularity easily J
= t
: < for = 1, 2 are as

required.
4.27
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

54 SAHARON SHELAH
5 Indiscernible sequence perpendicular to cuts
Our aim is to show that for a set of
: <
of pairwise perpendicu-
lar endless indiscernible sets, we can nd a model M
: <
with
dual-cf(
, M) : <
) essentially as we like, and other

b
in M has such dual

conality i this essentially follows. In fact we can demand M M
0
for any given
M
0
. Toward this we dene and investigate when an endless indiscernible sequence
c is perpendicular to a (Dedekind) cut (I
1
, I
2
) in an indiscernible sequence a.
We use
the Downward L.S. (on M N) can replace [T[ < [P
N
[ < [Q
N
[ by [P
M
[ =
[Q
M
[ but in general cannot invert the inequality, however for conality it
can.
For our purpose respecting dened in 5.2 is a central notion.
5.1 Discussion: 1) We can reformalize the aim as:
given

D = D
: <
), D
an ultralter and Dom(D
) : <
M
and given a sequence
: <
) of regular cardinals ( = cf() > [T[)

and, for simplicity, Dom(D
) in C an indiscernible sequence over Dom(D
) :
,= , then there is a -saturated model M A such that <
= dual-cf(D
, M) dened naturally. This property is meaningful also

for (complete rst order theories) T with the independence property (and
sequence

D). However, at least for some of them, e.g., for number theory
(a) assume
1
,=
2
<
and F is a one to one function from Dom(D
1
)
onto Dom(D
2
) maping D
1
to D
2
and is included in a function de-
nable in C with parameters from A.
Then A M C dual-cf(D
1
, M) = dual-cf(D
2
, M).
In other words if D
1
, D
2
are isomorphic as ultralters then D
1
, D
2
are not perpendicular in C for T = Th(N) because for every such F
there is a denable function extending it.
For dependent theory this gives just that denably isomorphic not
perpendicular
(b) we can weaken the demands on D
1
, D
2
.
5.2 Denition. 1) We say

I = (I
1
, I
2
) is a Dedekind cut (or just a cut) of the
linear order I, if I is the disjoint union of I
1
, I
2
and s I
1
& t I
2
s <
I
t
and we write I = I
1
+ I
2
, and the conality of

I is (cf(I
1
), cf(I
2
)) where I
2
is I
2
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

inverted. If I
1
+ I
2
is a convex subset of J and I
1
,= , = I
2
we may abuse our
notation saying (I
1
, I
2
) is a Dedekind cut of J. We say (I
1
, I
2
) is a Dedekind cut
of a if it is a Dedekind cut of Dom( a). If not say otherwise, I
1
,= , = I
2
, and the
cut is non-trivial if both its conalities are innite.
2) (J
1
, J
2
) (I
1
, I
2
) if J
1
is an end segment of I
1
and J
2
is an initial segment of I
2
.
3) We say that the set A respects the Dedekind cut (I
1
, I
2
) of a if (I
1
, I
2
) is a
Dedekind cut of a and for every

b
>
A for some (J
1
, J
2
) (I
1
, I
2
) the sequence
a (J
1
+ J
2
) is indiscernible over

b.
4) For indiscernible sequences a,

b such that

b is endless and a Dedekind cut (I
1
, I
2
)
of a we say that

b is perpendicular to the cut over A
0
when:
(a) the set A
0

b a respects the cut (I
1
, I
2
) of a
(b) for any set A A
0
respecting the Dedekind cut (I
1
, I
2
) of a and c realizing
Av(A,

b) also the set A c respects the Dedekind cut (I
1
, I
2
) of a.
We also say
b is perpendicular to ( a I
1
, a I
2
) over A
0
. If we omit A
0
we mean
A
0
=

b a.
4A) For indiscernible sequences a,

b such that

b is endless and a Dedekind cut
(I
1
, I
2
) of a we say that

b is truely perpendicular to the cut (I
1
, I
2
) of a when for
any set A, if A a respects the cut (I
1
, I
2
) of a and c realizes Av(A a,

b) then
A a c respects the cut (I
1
, I
2
) of a.
5) For endless indiscernible sequence a and A a we say an endless indiscernible
sequence

b =
b
t
: t I) over A is based on a or

b is based on (A, a) if each

b
t
realizes Av(A
b
s
: s <
I
t, a).
6) We say that the set A weakly respects the Dedekind cut (I
1
, I
2
) of a if (I
1
, I
2
) is
a Dedekind cut of a and for every formula ( x,
b) with

b A for some (J
1
, J
2
)
(I
1
, I
2
) and truth value t we have s J
1
+ J
2
C [= [ a
s
,
b]
t
.
5.3 Claim. 1) If (I
1
, I
2
) is a nontrivial cut of the indiscernible sequence a and
A = a
t
: t I
1
I
2
then A respects the cut (I
1
, I
2
) of a.
2) For every A and indiscernible a A and endless I there is
b
t
: t I) based on
(A, a).
3) Assume that (I
1
, I
2
) is a nontrivial cut of the indiscernible sequence a and c is
an endless indiscernible sequence. Then (a) (b) (c) and if the conalities of
the cut are >
0
then (a) (b) (c) where
(a) c is perpendicular to the cut (I
1
, I
2
) of a
(b) if

b is an indiscernible sequence based on ( c a, c) then the set

b c a
respects the cut (I
1
, I
2
) of a
(c) there is

b an indiscernible sequence based on ( c a, c) such that the set
b c a respects the cut (I

1
, I
2
) of a.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

56 SAHARON SHELAH
4) If A
i
(i < ) is increasing, A
0
A and each A
i
[truely, weakly] respects the cut
(I
1
, I
2
) of a then also
i<
A
i
does.
5) If

b A
0
A
1
and the conalities (of I
1
, I
2
and Dom(
b)) are > [T[ then:

b is
perpendicular to the cut (I
1
, I
2
) of a over A
0
then this holds over A
1
.
Proof. 1) Let

b
>
A; hence for some n < and t
0
, . . . , t
n1
I
1
I
2
we have
b a
t
: < n and dene J

1
= t I
1
: ( < n & t
I
1
t
< t, J
2
=
t I
2
: < n & t
I
2
t < t
. Clearly (J
1
, J
2
) (I
1
, I
2
) and a (J
1
+ J
2
)
is indiscernible over a
t
: t (I
1
I
2
)(J
1
J
2
) hence over a
t
: < n hence
over

b.
2) Also trivial.
3) (a) (c):
Let
b
n
: n < ) be an indiscernible sequence over a
c based on c and for

let A
= a
t
: t I
1
I
2
c
b
n
: n < . We can prove by induction
on that A
respects the cut (I

1
, I
2
) of a; for = 0 by part (1), for = n + 1
by clause (a), see Denition 5.2(4) for limit by part (4) and is straightforward.
Hence by the rst phrase of part (2),
b
n
: n < ) exemplies (a), as required.
(b) (c): Trivial.
(c) (b): Easy.
(b) (a): So we are assuming that the conalities are > [T[.
Let a set A c a respecting the cut (I
1
, I
2
) be given and let

b
n
realize Av( c
A
b
m
: m < n, c) for n < . By clause (b) we know that the set B =
b
n
:
n <
a respect the cut (I

1
, I
2
) of a. Let

d be a nite sequence from A, then
we shall prove
()
d
for some (J
1
, J
2
) (I
1
, I
2
) the sequence a (J
1
J
2
b
n
: n <

d.
Clearly this suces. Note that b
n
: n < ) is indiscernible over A a

d.
As the set
b a respects the cut (I

1
, I
2
) of a, for each n there is (J
n
1
, J
n
2
)
(I
1
, I
2
) such that a (J
n
1
J
n
2

b
0
. . .
b
n
. As the conalities of
(I
1
, I
2
) are >
0
, also (J
1
, J
2
) (I
1
, I
2
) where J
=:
n<
J
n
, and clearly a (J
1
J
2
)

b. By the last two sentences a (J
1
J
2
),
b
n
: n < ) are
mutually indiscernible. Also possibly replacing (J
1
, J
2
) by some (J
1
, J
2
) (J
1
, J
2
),
the sequence a (J
1
J
2

d and similarly (as the conalities
are large), a J
b
n

d a J
3
for = 1, 2. Together if
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

()
d
fails, we get a contradiction to the assumption on the

b
n
s.
4),5) Check.
5.3
5.4 Claim. Assume
(a) (I
1
, I
2
) is a non-trivial Dedekind cut of the indiscernible sequence a
(b) J
1
is an unbounded subset of I
1
(c) J
2
is a subset of I
2
unbounded from below
7
1) If A a then: A respects the Dedekind cut (I
1
, I
2
) of a i A respects the
Dedekind cut (J
1
, J
2
) of a (J
1
J
2
).
2) If

b is an endless indiscernible sequence and

b a A
0
, then

b is perpendicular
to the Dedekind cut (I
1
, I
2
) of a over A
0
i

b is perpendicular to the Dedekind cut
(J
1
, J
2
) of a (J
1
J
2
) over A
0
.
3) If J is an unbounded subset of Dom(
b), in (2) we can replace the last

b by

b J.
Proof. 1), 2), 3) As T has the dependence property.
5.4
5.5 Claim. 1) Assume a = a
1
a
2
is an indiscernible sequence and A a
1
a
2
respects the non-trivial cut ( a
1
, a
2
) both conalities of which are > [T[ and c A
is an endless indiscernible sequence perpendicular to a
1
and to the inverse of a
2
.
Then c is perpendicular to ( a
1
, a
2
) over A.
2) Let (I
1
, I
2
) be a Dedekind cut of the indiscernible sequence a with conalities
> [T[ and a A. The set A respects the cut (I
1
, I
2
) of a i the set A weakly
respected the cut (I
1
, I
2
) of a.
3) Let (I
1
, I
2
) be a Dedekind cut of the innite indiscernible sequence a with in-
nite conalities,

b an endless indiscernible sequence such that a,

b are mutually
indiscernible. Then

b is perpendicular to the cut (I
1
, I
2
) of a.
Proof. 1) It is enough to show that if c realizes Av(A, c) then A c respects the
cut (I
1
, I
2
) of a. By the second part (proved below) it is enough to show that
A c weakly respects ( a
1
, a
2
). Let

d A c, and consider ( x,

d). Without
loss of generality for some

d
A,

d = c
. By the choice of c and as c, a

1
are
perpendicular for some truth value t we have:
(a) for every large enough t Dom( a
1
) we have [= [ a
1
t
, c,

d
]
t
7
we can omit this here (and many other places) if in Denition 5.2(3),(4), we say for every
nite . But this does not help much because of > |T| in 4.23. We could replace |T|
+
by a
kind of
r
ind
(T).
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

58 SAHARON SHELAH
(b) for every large enough t Dom( a
1
) for every large enough s Dom( c) we
have [= [ a
1
t
, c
s
,

d
]
t
(c) for every large enough s Dom( c) for every large enough t Dom( a
1
) we
have [= [ a
1
t
, c
s
,

d
]
t
.
As

d
, c
s
A and A respects ( a
1
, a
2
) by clause (c) clearly
(d) for every large enough s Dom( c) for every small enough t Dom( a
2
)
we have [= [ a
2
t
, c
s
,

d
]
t
.
As c and the inverse of a
2
are perpendicular by clause (d) we have
(e) for every small enough t Dom( a
2
) for every large enough s Dom( c)
we have [= [ a
2
t
, c
s
,

d
]
t
.
By the choice of c and clause (e)
(f) for every small enough t Dom( a
2
) we have [= [ a
2
t
, c,

d
]
t
.
Together we are done.
2) Trivially respect implies weakly respect. So assume

b A. By 4.14(1) we can
nd (J
1
, J
2
) (I
1
, I
2
) such that a J
1
, a J
2
b. Toward contradiction assume that a (J

1
J
2
) is not indiscernible over

b, so
there is ( x
1
, . . . , x
n
,
b) witnessing it. We prove by induction on m n the natural

statement: for some truth value t if t
1
< . . . < t
n
are in J
1
J
2
and t
m+1
, . . . , t
n
J
2
then C [= [ a
t
1
, . . . , a
t
n
,
b]
t
.
For m = 0 this holds as a J
2

b so we can choose
the appropriate t.
For m+1, if t
m
J
2
uses the induction hypothesis, so assume t
m
J
1
, let t
be t
if
,= m and any member t
of J
2
satisfying t
m
< t
< t
m+1
if = m & m+1 n and
t
any member of J
2
if = m = n. Let c = a
t
: ,= m) and ( x, y
: = 1, . . . , n
and ,= m), z) = ( y
1
, . . . , y
m1
, x, y
m+1
, . . . , y
n
, z). So [= [ a
t
m
, c,
b]
t
by the
induction hypothesis and [= [ a
t
n
, c,
b]
t
by the assumption toward contradiction.
So by mutual indiscernibility ( a
t
s
, c,
b)
t
holds for every large enough s J
1
and
fails for every small enough s J
2
. But c,
b A and this shows that weakly

respect fails.
3) Easy, e.g., by part (1) and 4.7(1), clause (b) of Denition 5.2(4) holds and
similarly to the proof of 5.3, clause (a) of Denition 5.2(4) holds.
5.5
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

5.6 Denition. We say an endless indiscernible sequence a A has true dual
conality inside A, tr-d-cf( a, A) = if there is

b A such that a
b is an
indiscernible sequence with the cut ( a,

b) being respected by A and Dom(
b) has
downward conality .
5.7 Observation. 1) tr-d-cf( a, A) is well dened, i.e., has at most one value.
2) If tr-d-cf( a, A) is well dened then it is equal to dual-cf( a, A), see Denition
4.5(3).
Proof. Easy, as is unique.
5.8 Claim. 1) If is a limit ordinal, A
i
: i < ) is increasing, a A
0
is an
endless indiscernible sequence, a
i
A
i+1
realizes Av(A
i
, a), a
= a
i
: i < ) and
a
is the inverse of a
then
(a) a a
is an indiscernible sequence
(b) the set
i<
A
i
respects the cut (Dom( a), Dom( a
)) of a a
.
2) If a is a non-stable indiscernible sequence, a A, the set A respects the non-
trivial cut (I
1
, I
2
) of a and the conalities of the cut are > [T[ then dual-cf( a
I
1
, M) = cf(I
2
) = tr-d-cf( a I
1
, M).
3) [Not used] If a is an indiscernible sequence with Dedekind cut (I
1
, I
2
) of conality
(
1
,
2
),
0
<
1
,
2
and c an endless indiscernible sequence perpendicular to the
cut (I
1
, I
2
) of a, then: for every formula ( x, y, z) and sequence

b such that a
bc
respects the cut (I
1
, I
2
) of a for some truth value t we have:
() for some (J
1
, J
2
) (I
1
, I
2
) and for every t J
1
J
2
, for every large enough
s Dom( c) we have C [= [ a
t
,
b, c
s
]
t
.
4) If in part (3), [T[ <
1
,
2
, then for some (J
1
, J
2
) (I
1
, I
2
) and end segment J
of Dom( c) we have: a (J
1
+ J
2
), c J are mutually indiscernible.
Proof. 1), 2) Straightforward.
3) Let = [T[
+
and let

d
realize Av( a

b c
: < , a), for <

so by the denition of respect the Dedekind cut and as
1
,
2
>
0
there is
(J
1
, J
2
) (I
1
, I
2
) such that a (J
1
J
2
d
n
: n <
b hence
a (J
1
+J
2
) and

d =
: < ) are -mutually indiscernible over

b. So we have
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

60 SAHARON SHELAH
truth value t such that t J
& < C [= [ a
t
,
b,

d
]
t
. Recall that (I
1
, I
2
) have
conality (
1
,
2
) and for our purpose without loss of generality
1
,
2
> [T[. Now
clearly the three indiscernible sequences a J
1
, the inverse of a J
2
and
: < )
are mutually indiscernible, hence by 4.17, clause (B) without loss of generality they

b (i.e., omitting an initial segment of each and
renaming). By the choice of the

d
s, for every t J
1
+ J
2
for every large enough
s Dom( c) we have [= [ a
t
,
b, c
s
]
t
.
4) Should be clear (compare with 1.30, 5.5, 5.8).
5.9 Claim. Assume
(a) I = I
1
+ I
2
and the Dedekind cut (I
1
, I
2
) has conality (
1
,
2
)
(b) a = a
t
: t I) is an indiscernible sequence
(c) a A
(d) the set A respects the cut (I
1
, I
2
) of a
(e) [T[ <
1
,
2
(not used in part (1)).
1) If tp(
d, A) F
sp
and
1
,
2
, then the set A

d respects the cut (I
1
, I
2
) of a.
2) If A
+
= A a
i
: i < i
and for each i, tp(a

i
, A a
j
: j < i) belongs to
F
sp
min{
1
,
2
}
or is Av(A a
j
: j < i,

b) where

b A a
j
: j < i is an endless
indiscernible sequence perpendicular to the cut (I
1
, I
2
) of a over A a
j
: j < i,
then A
+
respects the cut (I
1
, I
2
) of a.
Proof. 1) As p = tp(
d, A) belongs to F
sp
, there is a subset B of A of cardinality

< such that p does not split over B.
Let

d
>
(A

d) and we should nd (J
1
, J
2
) (I
1
, I
2
) such that a (J
1
+J
2
)

d
. As we can increase

d
(this just makes our task harder)

without loss of generality

d
=

d e with e A. Now for every e
B e there is
(J
1
e
, J
2
e
) (I
1
, I
2
) such that a (J
1
e
+J
2
e
) is an indiscernible sequence over e
. Let
J
be J
e
: e
B e for = 1, 2 if >
0
and let J
= J
e
, e
listing B e
if =
0
. As
1
,
2
and cf(I
1
, I
2
) = (
1
,
2
) clearly (J
1
, J
2
) (I
1
, I
2
) and
a (J
1
+ J
2
) is indiscernible over B e.
Now for any formula and

b
>
B and s J
1
, t J
2
we have
()
1
tp( a
1
s
e, B) = tp( a
2
t
e, B).
[Why? By a (J
1
+ J
2
) indiscernible over B e.]
()
2
C [= [
d, a
1
s
, e,
b] i (x, a
1
s
, e,
b) p.
[Why? By the assumption on p as tp(
d, A).]
()
3
( x, a
1
s
, e,
b) p i ( x, a
2
t
, e,
b) p.
[Why? By ()
1
as p does not split over B.]
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

()
4
( x, a
2
t
, e,
b) p i C [= (
d, a
2
t
, e,
b).
[Why? By the denition of p as tp(
d, A).]
So by ()
2
+ ()
3
+ ()
4
as they hold for every and

b
>
B
()
5

d a
1
s
e and

d a
2
t
e realize the same type over B
hence
()
6
a
1
s
, a
2
t
realized the same type over e
d.
By 5.5(2) this suces as
1
,
2
> [T[ (or we could have repeated the proof for any
increasing sequence s
1
, . . . , s
n
), t
1
, . . . , t
n
) from J
1
+J
2
so
1
,
1
> [T[ will not be
used).
2) Let A
i
= Aa
j
: j < i, and we prove by induction on i that the cut (I
1
, I
2
) of
a is respected by the set A
i
. For i = 0 we use assumption (d). For i limit, we use
5.3(4). For i = j + 1 we use part (1) if tp(
b
j
, A
j
) F
sp
and we use the denition

5.2(4) if not. For i = i
we have gotten the desired conclusion.
5.9
5.10 Claim. Assume
(a) (I
1
, I
2
) is a non-trivial cut of the indiscernible sequence a
(b)

b is an endless indiscernible sequence
(c) a I
1
,

b are perpendicular
(d) for each t I
2
the sequence a
1
t
realizes Av( a
1
s
: s I
1
t [T[.
Then

b is perpendicular to the cut (I
1
, I
2
) of a.
Proof. Now by assumption (c) here and 4.27, for some unbounded I
1
I
1
and I
Dom(
b), the sequences a I
1
,

b I
are mutually indiscernible. By 5.4(1)+(2) +

5.3(5), without loss of generality I
1
= I
1
, I
= Dom(
b), so without loss of generality a

I
1
,

b are mutually indiscernible. Easily also

b and a are mutually indiscernible (by
clauses (c) and (d)). Hence by 5.5(3) clearly a,

b are perpendicular.
5.10
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

62 SAHARON SHELAH
5.11 Theorem. Assume
(a) =
<
2
+ 2
2
<
1
(b) [T[ <
1
= cf(
1
)
2

2
= cf(
2
),
1

1
= cf(
1
) for <
(c) [A[
(d) a
A is endless, non-stable indiscernible for <
and

(e) the a
for <
are pairwise perpendicular

(f) cf(Dom( a
)) is
1
.
Then we can nd a model M such that
() A M
() dual-cf( a
, M) =
1
+cf(Dom( a
)) for every <
, moreover tr-d-cf( a
, M) =
() if a M is a non-stable endless indiscernible sequence of cardinality (hence

conality) <
2
perpendicular to every a
then dual-cf( a, M) =
2
() M is
1
-saturated.
Proof. Let
0
= cf(Dom( a
)). Note that without loss of generality

() there is

b
such that a
is an indiscernible sequence, with the cut

(Dom( a
), Dom(
)) having conality (
0
,
1
) and this Dedekind cut be-

ing respected over the set A.
[Why? By using 5.10 and 5.3(4), of course.]
We can nd a
i
for i <
=:
2
such that letting A
i
= A a
j
: j < i we have
(i) for each i <
we have tp( a
i
, A
i
) F
sp
1
or tp( a
i
, A
i
) = Av(A
i
, a) for
some non-stable endless indiscernible sequence a A
i
of cardinality <
2
perpendicular to a
for every <
(ii) if p S
<
(A
(+1)
), p F
sp
1
then for ordinals j < , p is realized by
b
+j
(iii) if a is as in (i), a A
then for ordinal j < ,
b
+j
realizes
Av(A
+j
, a).
By bookkeeping, as in 5.9 this is straightforward and clauses () and () obviously
hold and in particular there is M with universe A
. By 5.9, M respects ( a
,

b
)
for each <
hence by 5.8(2), clause () holds.

(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

As for clause (), let a = a
t
: t I) M be a non-stable endless indiscernible
sequence [I[ <
2
, a perpendicular to every a
; it is also perpendicular to the inverse

of

b
as their conalities are dierent. Hence it is also perpendicular to the cut

( a
, b
) by 5.5(1).
As
2

2
= cf(
2
) = cf(
) it follows that for some <
, a A
and
so u = i : i (,
) and a
i
realizes Av(A
i
, a) is unbounded in
, let J be
(u, >) and so aa
t
: t J) is an indiscernible sequence. By 5.8(1) the set A
respect the Dedekind cut (I, J) of aa

t
: t J) that is M respects it hence
dual-cf( a, M) = cf(J
) = cf(
) =
2
as required.
5.11
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

64 SAHARON SHELAH
6 Concluding Remarks
We continue to deal with dependent T.
6.1 Conjecture: If p S(M) then there are at most 2
|T|
innite sequences of
indiscernible, pairwise non nb-s based on some ultralter D on M with Av(M, D) =
p.
A major lack of this work is the absence of test questions.
A candidate is the problem of classifying rst order theories by the existence of
indiscernibles (raised by Grossberg and the author, see [Sh 702, 2]), e.g.:
6.2 Question: If A C
T
, = [A[ +[T[ (or e.g. =
7
([A[ + [T[) and = (2
)
+
(or larger, but no large cardinals) and a
i
C
T
for i < then for some w []
+
,
the sequence a
i
: i w) is an indiscernible sequence over A(in C
T
).
Now though this property cannot characterize the dependence property, it is quite
natural in this context. Consider T
n
, the model completion of the empty theory in
the vocabulary R
n
, R
n
an n-place relation. So if =
n
()
+
, we get a positive
answer, but for n 2, T
n
is independent. We may consider replacing indiscernible
sequences a by a = a
t
: t I) as an index structure with n(I) = n(R) + 1 : R
an atomic relation of I < , a is I-indiscernible, i.e., k < , s, t
k
I, s
I
a
t
, a
s
realizes the same type. See also later.
Another direction is generalizing DOP, which in spite of its name is a non rst
order independence property.
On classication by Karp complexity see Laskowski and Shelah [LwSh 560], [LwSh
687] (let the -Karp complexity
(M) of M be the least such that every

L
,
(
M
)-formula is equivalent in M to such a formula of quantier depth < ,
and the (, )-Karp complexity of T is
(M)+1 : M a model of T of cardinality

.
For elementary classes which are unstable but dependent the following parallel to
DOT may help.
6.3 Denition. T has the dual-cf- -dimensional independence property when:
= (
0
,
1
,
2
),
1
,=
0
<
1
,
0
<
2
and for every and symmetric relation
R we can nd M
R
,

b
, c

0
(M
R
) and an indiscernible sequences I
,
=
a
,,i
: i <
0
) M
R
for (, ) R, < such that:
(a) the type of

b
I
,
is the same for all pairs (, ) R
(b) dual-cf(I
,
, M
R
) =
1
for (, ) R
(c) if < and R
and the sequence I
,
= a
,,i
: i <
0
) M
R
is such
that for every (
1
,
1
) R there is an automorphism h of C taking

b
1
to
, c
1
to c
and a
1
,
1
,i
to a
,,i
then dual-cf(I
,
, M) =
2
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

(d) M
R
is
+
0
-saturated.
Note that a sucient condition for this is
(c)
+
if I
,
= a
,,i
: i <
0
) realizes the relevant type over a
and
(, ) / R, < < and
1
<
1
< , (
1
,
1
) R then I
,
, I
1
,
1
are perpendicular.
Recall
6.4 Denition. M is -resplended when: if c
>
M, M N, N the
M
-reduct
of N
+
, [
N
+
N
[ < , then (M, c) can be expanded to a model of Th(N
+
, c).
6.5 Problem: Characterize the -Karp complexities for -resplended models of a
complete rst order theory T of cardinality > [T[ say for > > [T[ regular?
For T stable it is 0 (see [Sh:e, Ch.V]), as all such models are saturated. For T
with the independence property we should consider combining [LwSh 687] (which
constructs models of T
1
T with large -Karp complexity of the
T
-reduct) and
[Sh:e, Ch.V]. But here our concern is for unstable T with the dependence property.
If we look at an indiscernible sequence
b
t
: t I) inside a model M, we know
that distinct Dedekind cuts with at least one side having large conality are quite
unconnected. We shall show in subsequent work [Sh 783] that the large conality
demand is not incidental.
6.6 Question: Investigate the graph (
b
t
:

b
t
an endless indiscernible sequence,
perpendicularly).
As in 5 we can show that many variants are equivalent (using +, to
absorb). We can similarly discuss a parallel to deepness (see [Sh:f, X], recall that
deepness is related to orthogonality).
6.7 Discussion: 1) It is known that e.g. (rst theory of) the p-adics has the depen-
dence property (and are unstable). Does this work tell us anything on them? Well,
the construction in 5 gives somewhat more than what unstability gives: compli-
cated models with more specic freedom. Note that instead dual-cf(I, M) we can
use more complicated invariants (see [Sh:e, Ch.III,3] or earlier works).
We can, of course, (for the p-adic) characterize directly when indiscernible se-
quences are perpendicular.
2) We may like to dene super dependence properties (and
nip
(T)) (parallel
of superstable, i.e., (T) =
0
or super-simple
cdt
(T) =
0
). There are some
possibilities, one dened in [Sh:c, III], another in [Sh 783]. We may try the denition
w(I) <
0
, i.e., weight for every endless indiscernible sequence where
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

66 SAHARON SHELAH
6.8 Denition. For an endless indiscernible sequence I let w(I) = sup : there
is a sequence of length of pairwise perpendicular endless indiscernible sequences
each non perpendicular to I.
But w(I) is not weight for superstable theory just of a variant of it hence exactly
like dimension in the sense of algebraic manifolds.
6.9 Question: Assume I
= a
t
: t I
) for = 1, 2 are endless indiscernible

sequences and they are non perpendicular.
(a) Find a denable equivalence relation E such that a
2
t
/E : t I
2
) is non-
trivial and a
2
t
/E acl(I
1
a
2
s
: s <
I
2
t) for any large enough t (i.e.,
non-perpendicularty implies non-orthogonality for trivial reasons).
(b) If (I
1
, I
2
) is (1, < )-mutual indiscernible (parallel to Hrushovskis theorem),
can we dene a derived group? More generally, it seems persuasive that
groups appear naturally, particularly ordered groups.
(c) Does the fact that putting elements together, make strong splitting implies
dividing helps?
[recall:
(i) p strongly splits over A if there is a sequence a
t
: i < ) indiscernible
over A such that ( x, a
0
) & (x, a
1
)
(ii) p divides over A if there is an indiscernible sequence
b
i
: i < ) over
A and (x,
b
0
) p such that ( x,
b
i
) : i < is contradictory.
So having (i), letting

b
i
= a
2i
a
2i+1
, (x,
b
i
) = (x, a
2i
) & (x, a
2i+1
)
we have (ii).]
(d) Can the canonical bases of 1 help?
(e) What can we say on a perpendicular to a set (or model) A?
6.10 Discussion Cherlin wonders on the place of parallel algebraic geometric dimen-
sion and the place of 0-minimal theory. In my perception probably if we succeed in
6.9(a), we may have a minimality notion which may then be characterized as some
cases, but maybe it does not t.
6.11 Question: If B C C, p S
m
(B), then p has an extension q in S
m
(C)
which does not split strongly over B (and if p does not fork over A, then q does not
fork over A).
6.12 Question: Given two non perpendicular types in S(A) (or ultralter on A)
which are weakly perpendicular can we nd naturally dened groups?
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

6.13 Concluding Remark. We can dene when an endless indiscernible sequence is
orthogonal to a set and the dimensional independence property and prove natural
properties, we hope to pursue this.
6.14 Question: For an innite indiscernible sequence
b
t
: t I),
b
t

C can we
nd A C, [A[ [[ + [T[ and

b
+
t
=

b
t
a
t
, g( a
t
) = ([[ + [T[) for t J such that
b
+
t
: t J) is an indiscernible sequence and letting p( x, y) = tp(
b
+
s

b
+
t
, A) for
s <
J
t we have p( x, y) p( y, z) p( x, z)?
6.15 Question: 1) If M N, a N, (M, N, a) is -saturated > 2
|T|
. Can we nd
for every A M, [A[ < a set B M, [B[ 2
|T|
such that p B p A?
2) Fix nite set = (x, a) : a
m
C. Look at S(), is it true that every p S()
is determined uniquely by q p, [q[ n
1
and a rank < n

2
?
Or another way to use having few types.
6.16 Problem: Let M be -saturated, p S(M) and for = (x; y) let I
(x; y)
(p) =
( y, c) : c M and
b M :[= (
b, c), (x,
b) p is denable. Investigate this;

we can prove that it is not too small.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

68 SAHARON SHELAH
REFERENCES.
[BBSh 815] Bektur Baizhanov, John Baldwin, and Saharon Shelah. Subsets of
superstable structures are weakly benign. Journal of Symbolic Logic,
70:142150, 2005. math.LO/0303324.
[BlBn00] John T. Baldwin and Michael Benedikt. Stability theory, permutations
of indiscernibles, and embedded nite model theory. Transactions of
the American Mathematical Society, 352:49374969, 2000.
[BlSh 156] John T. Baldwin and Saharon Shelah. Second-order quantiers and the
complexity of theories. Notre Dame Journal of Formal Logic, 26:229
303, 1985. Proceedings of the 1980/1 Jerusalem Model Theory year.
[DjSh 692] Mirna Dzamonja and Saharon Shelah. On <
-maximality. Annals of
Pure and Applied Logic, 125:119158, 2004. math.LO/0009087.
[DjSh 710] Mirna Dzamonja and Saharon Shelah. On properties of theories which
preclude the existence of universal models. Annals of Pure and Applied
Logic, 139:280302, 2006. math.LO/0009078.
[Eh57] Andrzej Ehrenfeucht. On theories categorical in power. Fundamenta
Mathematicae, 44:241248, 1957.
[GrLe00] Rami Grossberg and Olivier Lessmann. Local order property in nonele-
mentary classes. Archive for Mathematical Logic, 39:439457, 2000.
[Ke76] H. Jerome Keisler. Six classes of theories. Journal of Australian Math-
ematical Society, 21:257275, 1976.
[Lw92] Michael C. Laskowski. Vapnik-Chervonenkis classes of denable sets.
Journal of the London Mathematical Society, 45:377384, 1992.
[LwSh 560] Michael C. Laskowski and Saharon Shelah. The Karp complexity
of unstable classes. Archive for Mathematical Logic, 40:6988, 2001.
math.LO/0011167.
[LwSh 687] Michael C. Laskowski and Saharon Shelah. Karp complexity and
classes with the independence property. Annals of Pure and Applied
Logic, 120:263283, 2003. math.LO/0303345.
[Mo65] Michael Morley. Categoricity in power. Transaction of the American
Mathematical Society, 114:514538, 1965.
[Sh:e] Saharon Shelah. Nonstructure theory, accepted. Oxford University
Press.
[Sh 3] Saharon Shelah. Finite diagrams stable in power. Annals of Mathe-
matical Logic, 2:69118, 1970.
(
7
1
5
)

r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
9
-
1
6

m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
9
-
1
8

[Sh 10] Saharon Shelah. Stability, the f.c.p., and superstability; model theoretic
properties of formulas in rst order theory. Annals of Mathematical
Logic, 3:271362, 1971.
[Sh:a] Saharon Shelah. Classication theory and the number of nonisomor-
phic models, volume 92 of Studies in Logic and the Foundations of
Mathematics. North-Holland Publishing Co., Amsterdam-New York,
xvi+544 pp, $62.25, 1978.
[Sh:93] Saharon Shelah. Simple unstable theories. Annals of Mathematical
Logic, 19:177203, 1980.
[Sh 205] Saharon Shelah. Monadic logic and Lowenheim numbers. Annals of
Pure and Applied Logic, 28:203216, 1985.
[Sh 300] Saharon Shelah. Universal classes. In Classication theory (Chicago,
IL, 1985), volume 1292 of Lecture Notes in Mathematics, pages 264
418. Springer, Berlin, 1987. Proceedings of the USAIsrael Conference
on Classication Theory, Chicago, December 1985; ed. Baldwin, J.T.
[Sh:c] Saharon Shelah. Classication theory and the number of nonisomor-
phic models, volume 92 of Studies in Logic and the Foundations of
Mathematics. North-Holland Publishing Co., Amsterdam, xxxiv+705
pp, 1990.
[Sh 284c] Saharon Shelah. More on monadic logic. Part C. Monadically inter-
preting in stable unsuperstable T and the monadic theory of

. Israel
Journal of Mathematics, 70:353364, 1990.
[Sh 284b] Saharon Shelah. Notes on monadic logic. B. Complexity of linear orders
in ZFC. Israel Journal of Mathematics, 69:94116, 1990.
[Sh 500] Saharon Shelah. Toward classifying unstable theories. Annals of Pure
and Applied Logic, 80:229255, 1996. math.LO/9508205.
[Sh:f] Saharon Shelah. Proper and improper forcing. Perspectives in Mathe-
matical Logic. Springer, 1998.
[Sh 702] Saharon Shelah. On what I do not understand (and have some-
thing to say), model theory. Mathematica Japonica, 51:329377, 2000.
math.LO/9910158.
[Sh 783] Saharon Shelah. Dependent rst order theories, continued. Israel Jour-
nal of Mathematic, 173:160, 2009. math.LO/0406440.

Saharon Shelah - Classification Theory For Elementary Classes With The Dependence Property - A Modest Beginning

Caricato da

Informazioni sul documento

Descrizione originale:

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Saharon Shelah - Classification Theory For Elementary Classes With The Dependence Property - A Modest Beginning

Caricato da

Copyright:

Formati disponibili

(

b) of formulas ( x, y) such that every instance ( x, c)

= A B then not much indiscernability is lost. An easy

, M) being any (somewhat large) pregiven

convex sets. For any T, we can still look at the rst

b) divide any indiscernible

(J, D) or any average type for an indiscernible sequence

(A)[ : [A[ are charac-

). Also ( x, y) is unstable in T i it has the

(x, y) has the independence property. Then Lachlan proves that

(x, y) has the strict order property.

a unary relation, this is connected

b, a) D. It will be called the D-average over J or

(A, D) where = ( x, y) is the set of formulas of the form ( x, a) or the

(A, D) is the union of Av

) is based on D over A and n < and I

b, B) = tp( c, B) and = ( x, y) a formula, then ( x,

b) = ( x, y, c) : g( x) = m, and for every a

b), i.e., no parameters

b) = ( x, y) : g( y) = m and for every a

b) be dened similarly allowing parameters, obviously dpfor

b) is dened similarly when the union is of n convex sets; similarly

b) = L() we mean the set of relevant formulas.

: < k)) is a rst order property of c (with the

b) so g( y) = m, then for every k large enough we have:

has < k/2 members

: < k)) is a subset of Av(A, D), in fact is

, I J where J is a linear order,

b) minimal as in 1.8(1)(a) and any pairwise distinct

b), similarly for replacements to

b) does not matter then we can replace

b) = ( x, y): any and y but x = x

b) = ( x, y, c) : c C, (and of course g( x) = m).

, an innite indiscernible sequence over A which is an A-nb of

b) was dened in [Sh:93, Def.5.1](4).

b), then for some nite and k

(A) instead of F((A, )) and

) (where the superscript in

means invert the order) is a -indiscernible sequence over A.

< [also large enough as in 1.8] such that:

are as in part (1) for =

(C, D) is denable by a rst order

) are as in part (1) for = 1, 2 and

) does not depend

: < n, C) is a nite set

(A) as any such J; why is F as required? So suppose that A B is

realizes the same -type over A

b, B) hence there is an automorphism

hence we can in clause (i) replace

, B, A, here standing for D, M, A, there and let a

) and the last i by the choice of

) as in Denition 1.19. Let D

) (by 1.6). Now clearly I [=

b) say more than stated.

is a -indiscernible sequence of the same -type as

: k), (a), (b), (c) as

b), then p is dened by

b) are nb-s of distance 2. [See Denition 1.11.]

) = m and for any permutation of 1, . . . , n

b) for some sequence a, for both truth value

: < n) contradict the minimality of

b) then for every n for some c