3

7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


1
On the number of automorphisms
of uncountable models
Saharon Shelah, Heikki Tuuri and Jouko V a an anen *
Abstract
Let (/) denote the number of automorphisms of a model / of power

1
. We derive a necessary and sucient condition in terms of trees for
the existence of an / with
1
< (/) < 2

1
. We study the suciency
of some conditions for (/) = 2

1
. These conditions are analogous to
conditions studied by D.Kueker in connection with countable models.
The starting point of this paper was an attempt to generalize some results of
D.Kueker [8] to models of power
1
. For example, Kueker shows that for countable
/ the number (/) of automorphisms of / is either or 2

. In Corollary 13
we prove the analogue of this result under the set-theoretical assumption I(): if
I() holds and the cardinality of / is
1
, then (/)
1
or (/) = 2

1
. In
Theorem 16 we show that the consistency strength of this statement + 2

1
>
2
is that of an inaccessible cardinal. We use [[/[[ to denote the universe of a model
/ and [/[ to denote the cardinality of [[/[[ . Kueker proves also that if [/[ ,
[B[ > and / B (in L

), then (/) = 2

. Theorem 1 below generalizes this

to power
1
. If / and B are countable, / ,= B and / B (in L

), then we
know that (/) = 2

. Theorem 7 shows that the natural analogue of this result

fails for models of power
1
. Theorem 14 links the existence of a model / such
that [/[ =
1
,
1
< (/) < 2

1
, to the existence of a tree T which is of power

1
, of height
1
and has (/) uncountable branches.
We use /

1
B to denote that has a winning strategy in the Ehrenfeucht-
Frasse game G(/, B) of length
1
between / and B. During this game two
players and extend a countable partial isomorphism between / and B. At
the start of the game is empty. Player begins the game by choosing an element
a in either / or B. Then has to pick an element b in either / or B so that a
and b are in dierent models. Suppose that a /. If the relation (a, b) is
not a partial isomorphism, then loses immediately, else the game continues in the
same manner and the new value of is the mapping (a, b). The case a B
is treated similarly, but we consider the relation (b, a). The length of our
game is
1
moves. Player wins, if he can move
1
times without losing. The
only dierence between this game and the ordinary game characterizing partial
* The rst author would like to thank the United StatesIsrael Binational
Science Foundation for support of this research (Publication # 377). The second
and third author were supported by Academy of Finland grant 1011040.
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


2
isomorphism is its length. M.Karttunen and T.Hyttinen have proved ([3,4,7])
that /

1
B is equivalent to elementary equivalence relative to the innitely
deep language M

1
. It may also be observed that /

1
B is equivalent to
isomorphism in a forcing extension, where the set of forcing conditions is countably
closed [9]. For the denition of M

1
and other information of

1
the reader is
referred to [3,4,7,9,10,11]. Our treatment is selfcontained, however. The denition
of the language M

1
is not needed in this paper.
One of the basic consequences of /

1
B is that if / and B both have power

1
, then /

= B [7]. The proof of this is similar to the proof of the corresponding
result for countable models.
We note in passing that there is a canonical innitary game sentence
A
(see
[3], [4] or [7]), a kind of generalized Scott sentence, with the property that B [=
A
i /

1
B for any B. So, if /

1
B happens to imply that B has power
1
,
then
A
characterizes / up to isomorphism.
The authors are indebted to Wilfrid Hodges for his help in the early stages of
this work and to Alistair Lachlan and Alan Mekler for suggesting improvements.
Theorem 1. If a model of power
1
is

1
-equivalent to a model of power
>
1
, then it has 2

1
automorphisms.
For the proof of this theorem we dene the following game G(/) where / is
a model of power
1
: There are
1
moves and two players and . During the
game a countable partial isomorphism is extended. At each move rst plays
a point, to which then tries to extend . can tell whether the point is to be
on the image side or in the domain side. Moreover, has to come up with two
contradictory extensions of , from which chooses the one the game goes on
with. wins, if he can play all
1
moves.
A model / is called perfect, if has a winning strategy in G(/).
Proposition 2. If /

1
B for some B of power >
1
, then / is perfect.
Proof. Let S be a winning strategy of in the Ehrenfeucht-Frasse-game.
An S-mapping is a partial isomorphism between / and B arising from S. We
describe a winning strategy of in G(/). During the game constructs S-
mappings : / B and : B / simultaneously with the required . The
idea is to keep = .
Suppose now plays x and asks to extend the domain of to x. If
x / dom() (= dom()), uses S to extend to x. Likewise, if (x) / dom(),
uses S to extend so that (x) dom(). Let (x) = ((x)). This completes
the rst part of the move of .
For the second part, has to come up with

and

, which are contra-

dictory extensions of . For any b B S gives some s(b) /. If b / ran(),
then s(b) / dom(). As [Bran()[ > [/[ , there are b ,= b

Bran() with
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


3
Figure 1.
s(b) = s(b

). We extend using S rst to get an element a so that (b) = a

and after that we extend further to get (b

) = a

. Now, a ,= a

, since b ,= b

(Figure 1). Now we can dene

and

. In the rst case we extend so that

(s(b)) = b and we let

= . (Note here, that we do not extend to b

.
It is not necessary to keep ran() = dom().) In the second case we extend so
that (s(b)) = b

and we dene

= . Because

(s(b)) ,=

(s(b)), the two

extensions are contradictory.
Proposition 3. If / is perfect, then (/) = 2

1
.
Proof. Suppose S is a winning strategy of in G(/). Let us consider all games
in which enumerates all of /. Each such play determines an automorphism of
/. Since has a chance of splitting the game at each move, there are 2

1
dierent
automorphisms.
This ends the proof of Theorem 1.
Now we dene a game that characterizes the elementary submodel relation
for the language M

1
. Suppose / B. We describe the game G

(/, B). The

game resembles very much the ordinary Ehrenfeucht-Frasse-game between / and
B. The dierence is that at the start of the game can pick a countable set C of
elements of / and set as the initial partial isomorphism = (a, a) [ a C. Then
and continue the game like the usual Ehrenfeucht-Frasse-game extending .
We write / _

1
B, if has a winning strategy in the game G

(/, B). If
/ _

1
B and / ,= B, then we write /

1
B. It can be proved that the relation
/ _

1
B holds if and only if / is an elementary submodel of B relative to the
language M

1
. In this denition the formulas of M

1
may contain only a
countable number of free variables. The proof is very similar to the proof of the
fact that /

1
B is equivalent to elementary equivalence of / and B ([7], [3],
[4]).
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


4
We describe the game G

(/, B), which is more dicult for to win than

G

(/, B). The length of the game is

1
and it resembles the Ehrenfeucht-Frasse
game. During it must extend a countable partial isomorphism : / B and
at each move the rules are the following:
(i) if a /, a / dom() and a / ran(), then can move a / and demand
to extend to (a, a);
(ii) if a / (a B) then can move a / (a B) and demand to extend
so that a dom() (a ran()).
We write /

1
B, if / B and has a winning strategy in the game
G

(/, B). If /

1
B and / ,= B, then we write / <

1
B.
Our aim is next to prove that if / <

1
B for some B, then there are 2

1
automorphisms of /.
Lemma 4. Let (/

)
<
( limit) be uncountable models such that:
(i) /

if < ;
(ii) /

<
/

if is a limit;
(iii) /

1
/
+1
if < .
Let /

<
/

. Then /
0
_

1
/

. (The arity of relations and functions must

be nite.)
Proof. For simplicity of notation, we assume that in the games G

(/, B) and
G

(/, B) at each round , extends the partial isomorphism by just a single

ordered pair (a

, b

), where a

/ and b

B.
For each < , let

be s xed winning strategy in G

(/

, /
+1
).
We describe a winning strategy for in G

(/
0
, /

). We modify the game

G

(/
0
, /

) so that and only move at innite limit ordinal rounds, which is

clearly equivalent to the original game. At each round <
1
, also constructs a
sequence s

of length +1, such that s

() /

for all . At limit rounds

, rst constructs s

and then extends the partial isomorphism in the game

G

(/
0
, /

) by (a, b), where a = s

(0) and b = s

().
Before round , we assume that the following conditions are true:
(1) For all < , the sequence ((s

(), s

(+1)))
<
is a play in G

(/

, /
+1
)
according to s winning strategy

.
(2) For all < , s

is continuous, that is, if is a limit ordinal, and s

() = a,
there is < , such that for all < , s

() = a.
(3) Suppose a is in the range of some sequence s

, < , and is the least

ordinal such that a A

. Then there is an ordinal such that [, ] = [

for some < , s

() = a. If is a successor, then is a successor ordinal

or . If is a limit, then = .
starts the game G

(/
0
, /

) by choosing the countable set C of elements

of /
0
. chooses as the rst sequences s
n
, n < , constant sequences whose
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


5
values enumerate C. Let us consider round in the game, where is an innite
limit. In general there are two cases.
First the case where picks a /
0
as his th move. If there is some s

such that s

(0) = a, then responds by s

() /

and denes s

= s

. Else,
by (3), can move a /

and choose the appropriate constant sequence as s

.
The inductive hypotheses are met and we can let s
+n
= s

, for n < .
Suppose then picks b /

as his th move. Again, if for some < ,

s

() = b, we are done. Else, let us construct the required sequence s

. Let
0
be the least ordinal such that b /

0
and s

(
0
) ,= b for all < . Note that
by hypothesis (3) and condition (ii) of the lemma,
0
=
0
+ 1, for some
0
(or

0
= 0). We dene s

() = b for all >

0
. Let c be the response of according
to

0
if continues G

(/

0
, /

0
) by moving b /

0
. Let s

(
0
) = c. Then we
continue the construction of s

by downward induction. then moves s

(0) /
0
in the game G

(/
0
, /

). Similarly, by a closing procedure, can construct s

+n
,
n < , so that clause (3) is satised at +.
Proposition 5. If / is of cardinality
1
and / <

1
B for some B, then
/

1
B for some B of power
2
, whence / is perfect.
Proof. We may assume / and B have both power
1
. Thus, by remarks
preceding Theorem 1, /

= B. We construct a sequence (/

)
<
2
of models
so that each is isomorphic to /, /

, if < , and /

<

1
/
+1
for
all <
2
. We handle the successor step by identifying /

with / via the

isomorphism. Then from B we get /
+1
. At limits we take the union of models.
Lemma 4 makes sure that the union is isomorphic to /, if it is not of power
2
.
So, if / fullls the condition of Proposition 5, then it has 2

1
automorphisms.
The proof of the following result shows that /

1
B is a much stricter condition
than / _

1
B.
Proposition 6.
/

1
B / _

1
B
but
/ _

1
B , /

1
B.
Proof. The rst claim is trivial. For the second consider the following models.
There is one equivalence relation R in the vocabulary. The model / contains
simply
1
equivalence classes of size
1
. The model B / contains one additional
equivalence class of size
1
. Then it is very easy to see that wins G

(/, B).
But can win G

(/, B) in two moves. First chooses some b B, b / /. Let

be s response. Let a /, / [= R(a,
1
(b)), a / ran() dom(). Then
demands to map a identically.
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


6
If / and B are countable, / ,= B and / B (relative to L

), then
(/) = 2

. This would suggest the analogous conjecture for uncountable models:

if [/[ = [B[ =
1
and /

1
B, then (/) = 2

1
. But this conjecture is false, as
the following counterexample constructed by S.Shelah shows.
Theorem 7. Let > be regular. There are models /
1
/
2
, /
1
,= /
2
,
[/
1
[ = [/
2
[ = , such that
(i) for every A [[/
1
[[ , [A[ < , there is an isomorphism from /
2
onto /
1
which is the identity on A;
(ii) (/
1
) .
Remark. Hence /
1

/
2
but there is no /
3
such that /
1

/
3
and
[/
3
[ > , as then (/
1
) = 2

.
Proof. We rst dene such /
1
and /
2
with the vocabulary L = R

[ 0 <
< , limit, where R

has places and [R

M
1

[ = [R
M
2

[ = . We can then
replace these models (in Proposition 8) by models with a vocabulary consisting of
just one binary relation.
We dene A, A

, f

and

, < , such that:

(1)

< for all < and

[ < ) is increasing and continuous;

(2)
0
= , if > 0 is a limit, then

=

<

, and if = + 1, then

;
(3) A

= i <

[ i even, A = i < [ i even;

(4) f

is a 11 function from onto A mapping

+1
onto A
+1
;
(5) f

maps the interval [

,
+1
) onto [

,
+1
) A for > ;
(6) f

[ A

is the identity function on A

;
(7) f

, < , are dened using free groups (see the construction of f

below).
The denition of

and A

is clear from (1)(3). We now describe the

construction of f

, < . If < , let T

at
= s

[ and T

nat
= (s

)
1
[
be sets of arbitrary symbols. Let T

be the set of all such sequences

=
1
. . .
n
that:
(T1) 0 n < ;
(T2)
k
T

at
T

nat
for all 1 k n;
(T3) if n > 0 then
n
= s

;
(T4)
k
T

nat

k+1
T

at
for all 1 k < n;
(T5) (k, )(
k
,
k+1
= s

, (s

)
1
).
Thus we see that T

is a subset of the normal forms of the free group generated

by s

[ . If =
1
. . .
n
T

and s

at
, then we dene the operation
s

in the following way:

(a) if
1
,= (s

)
1
or = , then s

= s

1
. . .
n
(i.e. just concatenate);
(b) if
1
= (s

)
1
, then s

=
2
. . .
n
.
It is easy to check that s

. Thus is dened like the multiplicative

operation for the free group.
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


7
Lemma A. Let ,

and . If ,=

, then s

,= s

.
Proof. Straightforward. Lemma A.
For each < let
(

, j

) [

<
+1

list the set

P

= (, j) [ T

, ,= , j <

, j / A

without repetitions in such a way that

is even if and only if

1
T

at
,
where we denote

1
. . .

.
If (, j) P

for some < , let (, j) be the unique such that (, j) =

(

, j

). Now we dene f

, < (see Figure 3). For < let

f

() =

if <

and A

,
(s

, ) if <

and / A

,
(s

, j) if

<
+1
and = (, j),
(s

, j) if

<
+1
, > and = (, j).
Figure 3.
We have to check that f

is well-dened, that is, (s

, j) and (s

, j)
must be dened above in appropriate conditions and their values must be even.
We check only (s

, j), the other case is similar. Suppose

<
+1
and
= (, j). Then T

, ,= . Let =
1
. . .
n
. If
1
,= (s

)
1
, then
s

= s

1
. . .
n
,= . Thus (s

, j) is dened and it is even, since s

at
.
Suppose
1
= (s

)
1
. Then s

=
2
. . .
n
. Now n 2 by (T3) and
2
T

at
by (T4). Thus
2
. . .
n
,= and (s

, j) is dened and even.

Lemma B. Conditions (4), (5) and (6) above are met.
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


8
Proof. From the denition of f

we see easily that f

maps
+1
to A
+1
and
[

,
+1
) to [

,
+1
) A, if > . We show rst that f

is a 11 function
A. Suppose
1
,=
2
. We prove f

(
1
) ,= f

(
2
). There are several cases,
of which we treat the two most interesting. The proof in other cases is similar or
trivial.
(a) Suppose
1
<

,
1
/ A

and
2
[

,
+1
). Let
2
= (, j). Since
,= , by Lemma A s

,= s

. Thus f

(
1
) = (s

,
1
) ,= (s

, j) = f

(
2
).
(b) Suppose
1
,
2
[

,
+1
). Let
1
= (
1
, j
1
) and
2
= (
2
, j
2
). If
j
1
,= j
2
, then the claim is clear. If j
1
= j
2
, then
1
,=
2
and by Lemma A
s

1
,= s

2
and again the claim holds.
Next we prove that f

is onto. Let A. We try to nd < , for which

= f

(). If A

, then we set = . Suppose then [

,
+1
)A. Denote
= (, j), where =
1
. . .
n
, ,= . We know
1
T

at
, since is even.
(a) If n = 1, then = s

by (T3) and we set = j .

(b) If n > 1 and
1
= s

, then we set = (
2
. . .
n
, j).
(c) If n > 1 and
1
,= s

, then = ((s

)
1

1
. . .
n
, j). Here is dened and
(T4) fullled because
1
T

at
.
Suppose then [

,
+1
) A, > .
(a) If
1
= s

, then n > 1 by (T3) and = (

2
. . .
n
, j).
(b) If
1
,= s

, then = ((s

)
1

1
. . .
n
, j).
Thus we have proved that f

: A is 11 and onto. Now (4), (5) and (6)

are clear. Lemma B.
If < , let () denote the unique for which

<
+1
. Let G
1
be the group of permutations of A generated by f

(f

)
1
[ , < . Let G
2
be the group of permutations of generated by (f

)
1
f

[ , < .
We are ready to dene the models. We dene /
1
and /
2
as follows:
(i) [[/
1
[[ = A;
(ii) [[/
2
[[ = ;
(iii) R
M
k

= i
0
i
2
. . . i

. . .)
<, even
[ g G
k
(

< even
g(i

) = ), k = 1, 2,
0 < < , limit.
Lemma C. /
1
/
2
.
Proof. Suppose i

[ < even) R
M
1

. Thus there are k < ,

r
,
r
< ,
for 1 r k such that (using (f

(f

)
1
)
1
= f

(f

)
1
)

< even
f

1
(f

1
)
1
f

2
(f

2
)
1
. . . f

k
(f

k
)
1
(i

) = .
If < is chosen large enough, then by (6) f

(i

) = i

and f

() = for all
< , even, and thus

< even
((f

)
1
f

1
)((f

1
)
1
f

2
) . . . ((f

k1
)
1
f

k
)((f

k
)
1
f

)(i

) = .
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


9
But this means i

[ < even) R
M
2

. The other direction is similar. Lemma

C.
Lemma D. For each , f

is an isomorphism from /
2
onto /
1
which is the
identity on A

. (Hence G
k
is a group of automorphisms of /
k
.)
Proof. Suppose i

[ < even) R
M
2

. Then

< even
(f

1
)
1
f

1
. . . (f

k
)
1
f

k
(i

) = .
If is chosen large enough, then

< even
f

(f

1
)
1
f

1
. . . (f

k
)
1
f

k
(f

)
1
(f

(i

)) = ,
which means f

(i

) [ < even) R
M
1

. The other direction is similar.

Lemma D.
Since is regular, Lemma D proves part (i) of the theorem. To show (ii) it
is enough to prove the following lemma, because [G
1
[ .
Lemma E. G
1
is the group of all automorphisms of /
1
.
Proof. Let g

AUT(/
1
), g

/ G
1
. Let G

1
be the group generated by
f

(f

)
1
[ , < . As is regular, by taking successive closures we can nd
a limit ordinal < such that:
(1) g

maps A

onto A

;
(2) for every g G

1
, g

[ A

,= g [ A

.
(In fact the set of such is a closed unbounded subset of .)
Let i

= g

() for < , =

, even. As g

AUT(/
1
) and [
< even) R
M
1

, there is some g
1
G
1
with

< even
g
1
(i

) = . Let
g = g
1
1
G
1
. Then

< even
g() = i

.
Thus g

[ A

= g [ A

. By (2) g [ A

/ h [ A

[ h G

1
and by (1) g maps
A

onto itself. To get a contradiction it is enough to prove:

() If g G
1
and g [ A

/ h [ A

[ h G

1
, then g does not map A

onto
itself.
Proof of (). So let
g = f

k
(f

k
)
1
. . . f

1
(f

1
)
1
be a counterexample with k minimal. Clearly
i
,=
i
and
i+1
,=
i
by the
minimality of k.
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


10
As g / G

1
, for some 1 r k holds
r
or
r
. If
r
, then
we can consider g
1
= f

1
(f

1
)
1
. . . f

k
(f

k
)
1
, which is also a counterexample
with k minimal. Thus we may assume without loss of generality that
r
for
some r. Let
= max(
r
[ r 1, . . . , k,
r
<
r
[ r 1, . . . , k,
r
< ) + 1.
Let
0
A

be arbitrary. We denote

1
= (f

1
)
1
(
0
),

1
= f

1
(
1
),
.
.
.

k
= (f

k
)
1
(
k1
),

k
= f

k
(
k
).
Thus
k
= g(
0
). For i = 0, . . . , k let
b
i
= max,
1
, . . . ,
i
.
Lemma F. Suppose
0
A

. Then (
i
) < maxb
i
+ 1, for i = 0, . . . , k.
Proof. By induction. First, (
0
) < . Suppose (
i
) < maxb
i
+1, . From
the denition of f

we see ((f

)
1
()) () for all , . Thus (
i+1
) (
i
).
We see also that if (f

()) > (), then (f

()) = . Thus (
i+1
) (
i+1
)
or (
i+1
) =
i+1
. In both cases (
i+1
) < maxb
i+1
+ 1, . Lemma F.
Lemma G. For all 1 i k either
i
b
i1
or
i
b
i1
.
Proof. Suppose
i
> b
i1
and
i
> b
i1
. Since b
i1
, this implies

i
,
i
. Thus
i
,
i
maxb
i1
+ 1, . Suppose
0
A

is arbitrary. By
Lemma F (
i1
) < maxb
i1
+ 1, and by (6) f

i
(f

i
)
1
(
i1
) =
i1
. But
now we see
f

k
(f

k
)
1
. . . f

1
(f

1
)
1
[ A

= f

k
(f

k
)
1
. . . f

i+1
(f

i+1
)
1
f

i1
(f

i1
)
1
. . . f

1
(f

1
)
1
[ A

,
a contradiction with the minimality of k. Lemma G.
The following lemma shows that g maps (s

, 1) outside A

, which contra-
dicts our assumption and proves ().
Lemma H. Let
0
= (s

, 1). Then for all 1 i k

i
is of the form
(s
b
i

i
2
. . .
i
n
i
, j
i
), where s
b
i

i
2
. . .
i
n
i
T
b
i
and n
i
1. Hence (
i
) = b
i
.
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


11
Proof. Suppose rst the claim holds for
i
, i 1. We prove it holds for
i+1
.
(a) Suppose
i+1
> b
i
= (
i
). Then
i+1
= (f

i+1
)
1
(
i
) =
i
. By Lemma
G
i+1
b
i
. Now

i+1
= f

i+1
(
i
) = (s

i+1
s

i
2
. . .
i
n
i
, j
i
).
Hence the claim holds for i + 1.
(b) Suppose
i+1
b
i
= (
i
). Then

i+1
= (f

i+1
)
1
(
i
) = ((s

i+1
)
1
s

i
2
. . .
i
n
i
, j
i
),
where
i+1
,=
i
by the minimality of k. Note that
i+1
is odd. If
i+1
> b
i
,
then

i+1
= f

i+1
(
i+1
) = (s

i+1

i+1
,
i+1
)
and the claim holds for i + 1. If
i+1
b
i
, then

i+1
= f

i+1
(
i+1
) = (s

i+1
(s

i+1
)
1
s

i
2
. . .
i
n
i
, j
i
),
where
i+1
,=
i+1
by the minimality of k and the claim holds.
Next we prove that the claim is true for i = 1.
(a) Suppose
1
> b
0
= . Then
1
=
0
= (s

, 1) and
1
b
0
= . As
above we get
1
= (s

1
s

, 1).
(b) Suppose
1
b
0
= . Then
1
= ((s

1
)
1
s

, 1), where
1
,= by
the denition of . If
1
> , then
1
= (s

1
,
1
). If
1
, then
2
=
(s

1
(s

1
)
1
s

, 1). Lemma H.
Let
0
= (s

, 1). By Lemma H (
k
) = b
k
, since
i
for some i.
Thus
k
/ A

, which proves (). This ends the proof of Lemma E and the whole
theorem.
Proposition 8. We can nd models /
1
and /
2
which satisfy Theorem 7 and
have a vocabulary of one binary relation.
Proof. Suppose / is a model of the vocabulary R

[ 0 < < , limit, such

that [/[ = , [R
M

[ and R

has places. We dene a model / = F(/) of

one binary relation R. Let
[[/[[ = [[/[[

((a

)
<
, ) [ / [= R

(a
0
, . . . , a
<
, . . .), < .
The relation R holds in / exactly in the following two cases:
(i) if b
1
, b
2
[[/[[ , b
1
= ((a

)
<
,
1
) and b
2
= ((a

)
<
,
2
), where
1
<
2
,
then / [= R(b
1
, b
2
);
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


12
(ii) if b [[/[[ and b = ((a

)
<
, ), then / [= R(a

, b).
In other words, for each tuple (a

)
<
, such that / [= R

(a
0
, . . . , a
<
, . . .) we
add new elements to [[/[[ . The new elements are wellordered by R and for
all < a

is in relation R with the th added element.

Obviously [F(/)[ = . It is a routine task to check that there is a 11 cor-
respondence between AUT(/) and AUT(F(/)). (Note that / [= xR(x, a)
i a [[/[[ .) Thus (/) = (F(/)). It also easy to see that if /

,
then F(/)

F(/

). Let / and /

be the models constructed in Theorem

7. Let /
1
= F(/) and /
2
= F(/

).
We say that a chain of models (/

)
<
is continuous, if /

<
/

for
a limit. A chain is an elementary chain, if /

1
/

for all < . If the

relation _

1
were preserved under unions of continuous chains of models, then
we could replace <

1
by

1
in Proposition 5, as is easy to see. This raised
the question, whether _

1
is preserved under unions of continuous chains. Since
Theorem 7 shows that <

1
cannot be replaced by

1
, it also proves that _

1
is not always preserved. Below we present also two other counterexamples. They
are continuous elementary chains of length and
1
. The problem, whether _

1
is preserved under unions of continuous chains of length
2
or greater, is open to
the authors.
We dene the linear order , which we shall use in the proofs below. The linear
order consists of functions f :
1
, for which the set n [ f(n) ,= 0
is nite. If f, g , then f < g i f(n) < g(n), where n is the least number,
where f and g dier. By
<
we mean the restriction of to those functions f
for which f(0) < . Similarly we dene

.
Let and be arbitrary linear orders. By we mean a linear order where
we have a copy of for every x . The order between the copies is determined
by . By + we mean a linear order, where is on top of . If is an ordinal,
then

denotes in a reversed order.

We rst prove a lemma about .
Lemma 9.
(i)

= for all ,
(ii) n

= for all n ,
(iii)

= for all <
1
.
Proof. (i) Let f

. Simply map f to g , where g(0) = f(0) and

g(n) = f(n), if n ,= 0.
(ii) We prove the claim by induction on n. Suppose n

= . Clearly

<1

= , thus n

=
<1
. By (i)

=
1
. So (n + 1)

=
<1
+
1

= .
(iii) We prove this by induction on . The successor step is easy, because
+

= . Suppose then that is a limit ordinal. Let (
n
)
n<
be an increasing
sequence conal in . Then =

n<

n+1

n
. All the dierences in the sum
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


13
are < , so we can use our induction assumption and we get

. Thus
the limit case is reduced to showing that

= . We describe the isomorphism.
First we map the topmost copy of in

to f [ f(0) > 0. This mapping

goes as in (i). Then we map the next copy of to f [ f(0) = 0, f(1) > 0,
and continuing this way we get an isomorphism.
Proposition 10. There exists an elementary chain (/
n
)
n<
of models of car-
dinality
1
such that
/
n
,_

n<
/
n
for all n.
Proof. We let /
n
= n. Then the union of the chain is / = . We can
choose an increasing sequence of points in / so that the length of the sequence
is and the sequence has no upper bound in /. It is not possible to nd such
a sequence in any /
n
. Thus it is clear that no /
n
is an elementary submodel of
/.
It remains to prove that our chain is really an elementary chain. We start to
play the game G

(/
n
, /
m
), m > n. First chooses a countable set C in /
n
,
which is mapped identically to /
m
. Some of the points of C are in the topmost
copy of in /
n
. Let <
1
be so big that none of these points f has f(0) .
We form an isomorphism between /
n
and /
m
so that it maps the points in C
identically. We map the part (n 1) +
<
in /
n
identically to /
m
. The
remaining part of /
n
is

and thus isomorphic to . The remaining part of

/
m
is isomorphic to + (m n) and thus isomorphic to . So we get an
isomorphism between the remaining parts. Now can win the game simply by
playing according to our isomorphism.
Proposition 11. There exists an elementary chain (/

)
<
1
of models of
cardinality
1
such that
/

,_

<
1
/

for all . In this chain /

<
/

, if is a limit ordinal.
Proof. We let /

= +

. Then there is a descending

1
-sequence in
/ =

<
1
/

, but no descending
1
-sequence in any /

. This shows that

/

,_

1
/.
We have to prove that our chain is elementary. We start to play the game
G

(/

, /

), where < . First chooses a countable set C of points in /

.
Let <
1
be so big that for no f C f(0) . We form an isomorphism
between our models so that it maps the points in C identically. First we map the
part

in /

identically to /

. We map the part

<
in the bottom copy of
in /

again identically to /

. Now it remains to map

to

, where
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


14
= . But, according to Lemma 7 (i) and (ii), these both are isomorphic to
, so we get the isomorphism between /

and /

. Then wins the game by

playing according to this isomorphism.
We shall now consider a totally dierent kind of condition which also guaran-
tees perfectness. Let I() denote the assumption (taken from [2]) that
there is an ideal I on
2
which is
2
-complete, normal, contains all
singletons , <
2
, and
I
+
= X
2
[ X / I
has a dense subset K such that every descending chain of length <
1
of elements of K has a lower bound in K.
Remark. I() implies that I is precipitous and hence that
2
is measurable
in an inner model. On the other hand, if a measurable cardinal is Levy-collapsed
to
2
, I() becomes true [1].
We prove that I() implies CH. Suppose 2

2
. Let T be a full binary
tree of height + 1. Let A t T [ height(t) = , [A[ =
2
. Let I be the
ideal on A given by I(). Now it is very easy to construct t
0
< < t
n
<
and X
0
X
n
, n < , such that height(t
n
) = n, X
n
K, and
for all a X
n
holds a > t
n
. Now

n<
X
n
contains at most one element, a
contradiction.
Theorem 12. Assume I(). If a model / of power
1
satises (/) >
1
,
then / is perfect.
Proof. (Inspired by [2].) Let I satisfy I(). We may assume I is an ideal on
a set AUT of automorphisms of power
2
. We describe a winning strategy of
in G(/). Let X AUT and f X. We say that f is an I -point of X, if for all
countable f , it holds that [] X I
+
, where [] = the set of all extensions
of .
Claim: Every X I
+
has an I -point.
Otherwise every f X has a
f
f with X[
f
] I . Because CH holds, there
are only
1
countable . This implies X

fX
X [
f
] I , a contradiction.
The idea of is to construct a descending sequence (X

)
<
1
of elements
of K. We denote by

the countable partial isomorphism at stage . The

descending sequence is chosen so that for all f X

holds

f .
Suppose the players have played moves. Then demands to extend

to a point x and give two contradictory extensions. For example, demands x

to be on the domain side. Because functions f can have only
1
dierent values
at x and I is
2
-closed, we can nd Y I
+
, Y X

, such that all the functions

3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


15
in Y agree at x. Now let f be an I -point of Y and let f

be an I -point of
Y f. Because f and f

are two dierent mappings, we can choose countable

f and

so that and

are contradictory extensions of

and they
are dened at x. Now we can choose X K and X

K, (X, X

Y ), so that
for all g X g and for all g X

g . The extensions and

are
the demanded contradictory extensions. For example, if picks , then we set
X
+1
= X and
+1
= .
Limit steps in the game do not cause trouble, because countable descending
chains in K have a lower bound in K.
Corollary 13. Assume I(). Then the following condition () holds:
() If / is a model of power
1
, then the conditions
(i) (/) >
1
,
(ii) (/) = 2

1
,
(iii) / is perfect,
are equivalent.
Remark. T. Jech has proved [5] it consistent that 2

=
1
, 2

1
>
2
and there
is a tree of power
1
with
2
automorphisms. Hence () cannot hold without
some set-theoretical assumption. We shall later show that the consistency strength
of () is that of an inaccessible cardinal. Note that () implies CH.
The following result of S.Shelah shows a dependence between trees and the
number of automorphisms of an uncountable model.
Theorem 14. Suppose that there exists a tree T of height
1
such that:
(i) T has uncountable branches, where
1
< < 2

1
;
(ii) each level in the tree has
1
nodes.
Then we can build a structure / of cardinality
1
with exactly automorphisms.
Proof. Let T

= t T [ height(t) = and
G

= X T

[ [X[ <
for each <
1
. If X, Y G

, we dene
X +Y = (XY ) (Y X),
i.e. X + Y is the symmetric dierence of X and Y . Clearly, + makes G

into
an Abelian group. Actually, G

is a linear vector space over the eld Z

2
= 0, 1,
but below we need only to know that G

is Abelian.
Let G be the Abelian group, which consists of all functions (
1
-sequences)
s :
1


<
1
G

, where s() G

, and addition is dened coordinatewise:

3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


16
(s
1
+ s
2
)() = s
1
() + s
2
(). If B = (t

)
<
1
is an
1
-branch in T , then B
determines naturally a sequence b G, where b() = t

. Let G

G be
the Abelian group generated by all sequences b corresponding to
1
-branches.
(Equivalently, G

is the vector subspace spanned by such sequences.)

Suppose s G

is arbitrary. Then s = b
1
+ + b
n
for some
1
-branches
b
1
, . . . , b
n
. Clearly, if t T

, then t s() i an odd number of branches

b
1
, . . . , b
n
passes through t. From this we see that if < and t T

, then
() t s() i t has an odd number of successors in s().
Let /

be a model of vocabulary R
s
[ s G

such that
(i) [[/

[[ = s [ s G

;
(ii) /

[= R
s
(s
1
, s
2
) i s
2
= s
1
+s.
The model /

is like an ane space, where the set of points is [[/

[[ and the
space of dierences G

is kept rigid. Obviously, [/

[ = and AUT(/

) consists
of all mappings

s
, s [[/

[[ , where

s
(x) = x + s. Thus /

has exactly
automorphisms.
Let / be a model such that:
(i) [[/[[ = s [ [ s [[/

[[, <
1
;
(ii) the vocabulary of /

is F R
s
[ s [[/[[;
(iii) / [= R
s
(s
1
, s
2
) i the domains of s, s
1
, s
2
are equal and s
2
= s
1
+s (where
the sum is dened coordinatewise);
(iv) / [= F(s
1
, s
2
) i s
1
is an initial segment of s
2
.
Since [T[ =
1
, there are only
1
countable initial segments of
1
-branches, and
[/[ =
1
. We show that there is a 11 correspondence between AUT(/

) and
AUT(/). Let s [[/

[[ be arbitrary. Then

s
AUT(/

). We dene from

s
an automorphism
s
of /: if r [[/[[ and dom(r) = , then
s
(r) = r +s [ .
Obviously, if s ,= s

, then
s
,=
s
.
Suppose then is an automorphism of /. We denote by s

a function,
such that dom(s

) = and s

() = for all < . We dene s G in the

following way: s [ = (s

) for all <

1
. We show that s [[/

[[ . By ()
[s()[ [s()[ if . Since [s()[ is nite for all , there must be n and
such that [s()[ = n for all . Thus from () we see that from up s
determines some
1
-branches b
1
, . . . , b
n
, such that s [ (
1
) = b [ (
1
), where
b = b
1
+ + b
n
. It remains to show that s [ ( + 1) = b [ ( + 1). We know
s [ ( + 1) = (s
+1

) = s

[ ( + 1) for some s

[[/

[[ . Since s

() = b(), ()
implies that s

[ ( + 1) = b [ ( + 1), and thus s = b [[/

[[ .
Now it is very easy to show that =
s
. Thus there is a 11 correspondence
and / has exactly automorphisms.
Remark. If the tree T above is a Kurepa tree, then the resulting model / is
clearly not perfect.
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


17
We can modify the preceding proof to get a suitable model with a nite
vocabulary. We add to the model / the set a
s
[ s [[/[[ of new elements
and wellorder them with a new relation <. Then we can use these new elements
to code the relations R
s
into a single relation and we get a nite vocabulary. This
modication does not aect the number of automorphisms.
Theorem 14 is of use only, if the conditions in it are consistent with ZFC. We
show that this is indeed the case.
A tree T is a Kurepa tree if:
(i) height(T) =
1
;
(ii) each level of T is at most countable;
(iii) T has at least
2
uncountable branches.
It is well-known (see e.g. [6]) that Kurepa trees exist in the constructible
universe. Let / be a countable standard model of ZFC + V = L. Let T be a
Kurepa tree in /. Let be the number of uncountable branches in T . Now we
use forcing to get a model where 2

1
> . We utilize Lemma 19.7 of [6]. In /
the equation 2
<
1
=
1
holds. Let > be such that

1
= . Let P be the
set of all functions p such that:
(i) dom(p)
1
and [dom(p)[ <
1
,
(ii) ran(p) 0, 1,
and let p be stronger than q i p q . The generic extension /[G] has the
same cardinals as / and /[G] [= 2

1
= . P is a countably closed notion
of forcing. Hence Lemma 24.5 of [6] says that the Kurepa tree T contains in
/[G] just those branches that are in the ground model. Thus there are exactly
uncountable branches in T also in the extended model /[G] . CH is true in
L, therefore /[G] [= 2

=
1
by the countable closure of P . We have obtained
a model /[G] of ZFC + CH with a tree T , which has the properties (i)(ii) of
Theorem 14.
From Theorem 14 and the above remarks we obtain a new proof of Jechs
result [5]:
If ZF is consistent, then ZFC + 2

=
1
+ there exists a model of
cardinality
1
with automorphisms,
1
< < 2

1
is consistent.
If we assume CH, we can prove the other direction in Theorem 14.
Proposition 15. Assume CH. Suppose that we have a model / of cardinality

1
and / has automorphisms,
1
< < 2

1
. Then there exists a tree T of
height
1
such that the conditions (i)(ii) in Theorem 14 hold.
Proof. To avoid some complications, we assume that / has a relational vo-
cabulary. If not, we can transform the vocabulary to relational and that does not
aect the number of automorphisms. The tree T will consist of partial automor-
phisms of /. Let (a

)
<
1
enumerate /. Let /

= /[`a

[ < . We
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


18
let T = f [ f is an automorphism of some /

. If f, g T , then f g i g
extends f .
Suppose f is an automorphism of /. Let <
1
be arbitrary. It may be
that the restriction of f to /

is not a bijection from /

to /

, but by taking
successively closures we nd > , for which f gives an automorphism of /

.
Thus f determines an uncountable branch in T .
For the other direction, if we have an uncountable branch in T , it is clear
that it determines an automorphism of /. Thus T has uncountable branches.
The tree T may contain at most
1

nodes. Since we assumed CH, this

is equal to
1
. So, each level of T contains
1
nodes.
Theorem 16. CH + () is equiconsistent with the existence of an inaccessible
cardinal. Also CH + 2

1
>
2
+ for all / of power
1
, (/) >
1
implies
(/) = 2

1
is equiconsistent with the existence of an inaccessible cardinal.
Proof. Let be a strongly inaccessible cardinal and so that =

1
.
Let P = QR, where Q is the Levy collapse of to
2
(see [6], p. 191) and R is
the set of Cohen conditions for adding subsets to
1
. We show that V
P
[= ().
Suppose p [= (/) >
1
. We may assume, without loss of generality, that / V .
Hence there is a P -name

f and p P so that p [=

f is an automorphism of
/ and

f / V . For any extension q of p let
f
q
= (, ) [ q [=

f() = .
Now for each extension q of p and for all countable sets A, B
1
there are
extensions q
0
and q
1
of q in P and an element a of
1
so that
(i) A a dom(f
q
0
) dom(f
q
1
),
(ii) B ran(f
q
0
) ran(f
q
1
),
(iii) f
q
0
(a) ,= f
q
1
(a).
Using this fact it is easy to see that p [= wins G(/). This ends the proof of
one half of the claims.
For the other half of the rst claim we assume that CH + () holds. If
2
is
not inaccessible in L, then there is a Kurepa tree with
2
branches, and hence
by the remark after Theorem 14, a non-perfect model of cardinality
1
with >
1
automorphisms.
For the other half of the second claim we show that under our assumption

2
has to be inaccessible in L. For this end, suppose
2
is not inaccessible in L.
Then there is A
1
so that
L[A]
2
=
2
,
L[A]
1
=
1
and GCH holds in L[A]
(see, e.g., Jech [6], p.252). We shall construct a tree with
1
nodes and exactly

2
branches. Let C be the set of with
1
< <
2
, and L

[A] [= ZFC- +
there is cardinal
1
and there are no cardinals >
1
. Note that C L[A] .
If < , we denote by (L

[B], ) a model of vocabulary (, U

1
, U
2
), where
U
1
and U
2
are unary relations, the interpretation of U
1
is B and the interpretation
of U
2
is the single element L

[B] .
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


19
We form the Skolem hulls in this proof by choosing as a witness the element
which is the smallest possible in the canonical well-ordering of the corresponding
model.
Fact A. An easy argument shows that if C and < , then there cannot
be any gaps between ordinals which are included in the Skolem hull of
1

(or
1
, as is denable in the model) in (L

[A], ).
Let B be the class of pairs (, (L

[B], )) L[A] , where L

[B] [= ZFC-
+ there is cardinal
1
and there are no cardinals >
1
, B = A
L

[B]
1
,
<
L

[B]
1
, < and >
L

[B]
1
.
We dene a partial ordering of these pairs as follows:
(, (L

[B], )) < (

, (L

[B

],

))
if <

and (L

[B], ) is the transitive collapse of the Skolem hull of

in (L

[B

],

). We dene a tree T as follows: Nodes of the tree are

pairs (, (L

[B], )) B with < <

1
. The ordering of T is the same as that
of B. The cardinality of T is
1
.
If G = (

, (L

[B

],

)), <
1
, is an uncountable branch in T , then the
direct limit of (L

[B

],

), <
1
, is isomorphic to some (L

[A], ), where
C. If we denote by H

the transitive collapse of the Skolem hull of

, <
1
, in (L

[A], ), then (, H

), <
1
, is a branch H in T . A
straightforward argument shows that G and H coincide. So the original branch
G is in fact in L[A] . Since T has at most
2
uncountable branches in L[A] , it
has at most
2
uncountable branches altogether. On the other hand, by Fact A
above, T clearly has at least
2
uncountable branches. We have shown that T
has
1
nodes and exactly
2
uncountable branches.
In this paper we have considered models of cardinality
1
and games of length

1
. When we generalize the model theory of countable models to uncountable
cardinalities, many problems arise. We chose to concentrate our attention on

1
, because it oers the simplest example of an uncountable cardinal, and even
this simple case seems to present enough problems. Naturally, the results in this
paper can be generalized to many other cardinalities , i.e. we can consider
models of power and games of length . The results 16 above are valid for
any uncountable cardinal . Proposition 10 can be generalized for any regular
uncountable cardinal , thus we get an elementary chain of length , for which _

is not preserved under the union. From the ideas of Proposition 11 we obtain the
following result: if is a regular uncountable cardinal, is a successor cardinal
and , then there is an elementary chain of length , for which _

is
not preserved under the union. Theorem 14, which shows a dependence between
trees and automorphisms, holds for any uncountable . Proposition 15 has a
counterpart for any regular uncountable .
3
7
7


r
e
v
i
s
i
o
n
:
1
9
9
5
-
1
1
-
2
5







m
o
d
i
f
i
e
d
:
1
9
9
5
-
1
1
-
2
6


20
References
[1] F.Galvin, T.Jech, M.Magidor, An ideal game, J. Symbolic Logic vol. 43 no.
2 (1978) 284292.
[2] W.Hodges and S.Shelah, Innite games and reduced products, Ann. Math.
Logic 20 (1981) 77108.
[3] T.Hyttinen, Games and innitary languages, Ann. Acad. Sci. Fenn. Ser. A
I Math. Diss. vol 64 (1987).
[4] T.Hyttinen, Model theory for innite quantier languages, Fund. Math. 134
(1990) 125142.
[5] T.Jech, Automorphisms of
1
-trees, Trans. A.M.S. 173 (1972) 5770.
[6] T.Jech, Set theory (Academic Press, 1978).
[7] M.Karttunen, Model theory for innitely deep languages, Ann. Acad. Sci.
Fenn. Ser. A I Math. Diss. vol 50 (1984).
[8] D.Kueker, Denability, automorphisms and innitary languages, in: Barwise,
ed., The syntax and semantics of innitary languages (Springer, 1968).
[9] M.Nadel and J.Stavi, L

-equivalence, isomorphism and potential isomor-

phism, Trans. A.M.S. 236 (1978) 5174.
[10] J.Oikkonen, How to obtain interpolation for L

, in: Drake, Truss, eds.,

Logic Colloquium 86 (North-Holland, 1988).
[11] J.V a an anen, Games and trees in innitary logic: A survey, in Quantiers (eds.
M.Krynicki, M.Mostowski and L.Szczerba), to appear.
Saharon Shelah
Department of Mathematics
Hebrew University of Jerusalem
Jerusalem
Israel
Heikki Tuuri
Department of Mathematics
University of Helsinki
Hallituskatu 15
00100 Helsinki
Finland
Jouko V a an anen
Department of Mathematics
University of Helsinki
Hallituskatu 15
00100 Helsinki
Finland