Sie sind auf Seite 1von 35

(

7
7
1
)


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







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


POLISH ALGEBRAS,
SHY FROM FREEDOM
SH771
Saharon Shelah
The Hebrew University of Jerusalem
Einstein Institute of Mathematics
Edmond J. Safra Campus, Givat Ram
Jerusalem 91904, Israel
Department of Mathematics
Hill Center-Busch Campus
Rutgers, The State University of New Jersey
110 Frelinghuysen Road
Piscataway, NJ 08854-8019 USA
Abstract. Every Polish group is not free whereas some F

group may be free. Also


every automorphism group of a structure of cardinality, e.g.

is not free.
This research was partially supported by the Israel Science Foundation founded by the Israel
Academy of Sciences and Humanities.
Publication 771
I would like to thank Alice Leonhardt for the beautiful typing.
Latest Revision - 09/June/26
Typeset by A
M
S-T
E
X
1
(
7
7
1
)


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







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


2 SAHARON SHELAH
0 Introduction
Our rst motivation was the question: can a countable structure have an auto-
morphism group, which a free uncountable group? This is answered negatively in
[Sh 744].
This was a well known problem in group theory at least in England (David Evans
in a meeting in Durham 1987) and we thank Simon Thomas for telling us about
it. Independently in descriptive set theory, Howard and Kechris [BeKe96] ask if
there is an uncountable free Polish group, i.e. which is on a complete separable
metric space. A related result (before [Sh 744]) was gotten by Solecki [So98], [So99]
who proved that the group of automorphisms of a countable structure cannot be an
uncountable free abelian group. Having the problem arise independently supported
the feeling that it is a natural problem.
The idea of the proof in [Sh 744] was to prove that such a group has some
strong algebraic completeness or compactness, more specically for any sequence
d
n
: n < ) of elements of the group converging to the identity many countable
sets of equations are solvable. This is parallel in some sense to Hensel lemma for
the p-adics, and seem to me interesting in its own right.
Lecturing in a conference in Rutgers, February 2001, I was asked whether I am
really speaking on Polish groups. We then prove the parallel result (i.e. 0.1 below)
using a more restrictive condition on the set of equations. Parallel theorems hold,
for semi groups and for metric algebras, e.g. with non isolated unit (e is a unit
means e is a subalgebra).
More specically we prove (see Conclusion 4.3, 4.1(1)).
0.1 Theorem. 1) There is no Polish group which as a group is free and uncountable.
2) Slightly more generally, assume
(a) G is a metric space
(b) G is a group with continuous xy, x
1
(c) G is complete
(d) the density of G is < [G[.
Then G is not free.
0.2 Thesis: If G is a Polish algebra satisfying one of the compactness conditions
dened below, then it is in fact large in the sense of lots of sets of equations has a
solution.
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 3
A reader interested just in this theorem 0.1 can read just 3.
0.3 Question: What are the restrictions on Aut(A) for uncountable structures A?
We also prove that if A is a structure of cardinality , is strong limit of
conality
0
(e.g.

) and the automorphism group of A is of cardinality > , then


it is far from being free; this does not follow directly from 0.1(2) as the natural
metric considered here does not satisfy all the conditions.
In [Sh 744] this is proved in the special case where A =

n
P
A
n
satisfying [P
A
n
[ < .

Note: An arbitrary subgroup e.g. of the symmetric group of size
1
can consistently
be made into an automorphism group by Just, Shelah, Thomas [JShT 654]. So

is more interesting from this point of view.


(
7
7
1
)


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







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


4 SAHARON SHELAH
0.4 Question: Is there a model theory of Polish spaces?
Naturally we would like to develop a parallel to classication theory (see [Sh:c]).
A natural test problem is to generalize Morley theorem = Los conjecture. But
we only have one model so what is the meaning?
Well, we may change the universe. If we deal with abelian groups (or any variety)
it is probably more natural to ask when is such (Borel) algebra free.
0.5 Example: If P is adding (2

0
)
+
-Cohen subsets of then
(C)
V
and (C)
V[G]
are both algebraically closed elds of characteristic 0 which are not isomorphic (as
they have dierent cardinalities).
So we restrict ourselves to forcing P
1
P
2
such that
(2

0
)
V[P
1
]
= (2

0
)
V[P
2
]
and compare the Polish models in V
P
1
, V
P
2
. We may restrict our forcing notions
to c.c.c. or whatever...
0.6 Example: Under any such interpretation
(a) C = the eld of complex numbers is categorical
(b) R = the eld of the reals is not
(by adding 2

0
many Cohen reals).
(Why? Trivially: R
V[P
2
]
is complete in V[P
2
] while R
V[P
1
]
in V
[P
2
]
is not complete,
but there are less trivial reasons).
0.7 Conjecture: We have a dychotomy, i.e. either the model is similar to categorical
theories, or there are many complicated models.
0.8 Thesis: Classication theory for such models resemble more the case of L

1
,
than the rst order.
See [Sh:h]; as support for this thesis we prove:
0.9 Theorem. There is an F

abelian group (i.e. a F

-denition, in fact an
explicit denition) such that V [= G is a free abelian group i V [= 2

0
<
736
.
Comments: In the context of the previous theorem we cannot do better than F

,
but we may hope for some other example which is not a group or categoricity is
not because of freeness.
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 5
0.10 Conclusion: Freeness (of an F

-abelian group) can stop at


n
(any n).
A connection with the model theory is that by Hart-Shelah [HaSh 323] such things
can also occur in L

1
,
whereas (by [Sh 87a], [Sh 87b] Theorem) if

n
(2

n
< 2

n+1
)
and L

1
,
, categorical in every
n
, then is categorical in every . See more
in [ShVi 648].
The parallels here are still open.
This casts some light on the thesis that non-rst order logics are more distant
from the so-called mainstream mathematics dealing with automorphisms.
Lecturing on [Sh 744], in MAMLS meeting (Rutgers, Feb.2001) some asked if
this apply to Polish group, (not just automoprhisms of countable structures) and
showed it. Subsequently (during Hattingham, Aug. 2001 Euresco conference),
telling Sabbagh about the results for Polish groups he wondered whether we get
non
0
-freeness, but the results give non-strongly-
1
-freeness. Subsequently his
ex-student Anatole Khelif who continues this work and proves non-
0
-freeness.
Lately Blass asks on denable abelian subgroups of Z

, answers are derived for


this from [Sh 402] and 5. The works originally have a section on stability theory in
this context, proving a generalization of the existence of large indiscernibility set.
Very lately (Feb 2009) it was moved to [Sh 849], also the parts on semi-groups,
the semi-group of endo-morphism and for metric algebras were removed.
We may be more humble than in 0.4.
0.11 Question: Is there model theory for equational theories, stressing free algebras?
The material in 1 - 5 (except some generalizations) was presented in a course
in Rutgers, Sept. - Oct 2001 and I thank the audience for their comments. We
thank the referee for doing much to improve the presentation in particular by less
is better.
0.12 Notation. 1) Let denote the set of natural numbers, and let x < mean x
is a natural number.
2) Let a, b, c, d denote members of G (a group).
3) Let

d denote a nite sequence d
n
: n < n

), and similarly in other cases.


4) Let k, , m, n, i, j, r, s, t denote natural numbers.
5) Let , , denote reals > 0.
0.13 Denition. 1) A group word is a sequence x
n
1
1
x
n
2
2
, . . . , x
n
k
k
where the x

are
variables or elements of a group and n

Z for = 1, . . . , k.
2) The word is reduced if n

,= 0, x

,= x
+1
.
3) The length of a word w = x
n
1
1
x
n
2
2
. . . x
n
k
k
is
k

=1
[n

[.
(
7
7
1
)


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







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


6 SAHARON SHELAH
4) A group term w(x
1
, . . . , x
n
) is a word of the form x
n
1

1
x
n
2

2
. . . x
n
k

k
with
i

1, . . . , n, n

Z (actually r

1, . . . , 1 suce).
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 7
1 metric groups and metric models
We rst dene metric [complete] group, and give a natural major example: au-
tomorphism groups. The natural example of metric group is the group of automor-
phism of a countable structure.
1.1 Denition. a = (G, d) is called a metric group if:
(a) G is a group
(b) G is a metric space for the metric d
(c) the functions xy, x
1
are continuous.
2) Saying (G, d) is complete, means complete as a metric space.
1.2 Notation: For a metric group a, the metric is denoted by d
a
and the unit is
denoted by e
a
and the group by G
a
. When no confusion arises G is a metric
group means (G, d
G
) is a metric group.
Now we dene cases closer to automorphism groups, in those cases the proof is very
similar to the one in [Sh 744].
1.3 Denition. 1) G is a specially metric group when:
() G is a metric group
() for every , R
+
there is R
+
such that: if x
1
, x
2
, y
1
, y
2
a :
d
G
(a, e
G
) < then d
G
(x
1
, x
2
) < d(y
1
, y
2
) < implies d
G
(x
1
y
1
, x
2
y
2
) <
d(x
1
1
, x
1
2
) < ; this is a kind of uniform continuity (inside the -
neighborhood of e
G
; this is harder if we increase and/or decrease )
() for arbitrarily small R
+
the set a G : d(a, e
G
) < is a subgroup of
G.
2) We say

=
n
: n < ) is strongly O.K. for G if:
(a)
n
R
+
n
(b)
n
satises clause () of part (1), i.e. a G : d(a, e
G
) <
n
is a subgroup
of G
(c)
n+1

n
and 0 = inf
n
: n <
(d) if x
1
, x
2
, y
1
, y
2
a G : d
G
(a, e
G
) <
0
and d(x
1
, x
2
) <
n+1
, d
G
(y
1
, y
2
) <

n+1
and r(1), r(2) 1, 1 then d
G
(x
r(1)
1
y
r(2)
1
, x
r(1)
2
y
r(2)
2
) <
n
.
(
7
7
1
)


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







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


8 SAHARON SHELAH
3) We say G is specially
+
metric group if in part (1) we have (), () and
()
+
for every R
+
the set a G : d(a, e
G
) < is a subgroup of G.
1.4 Observation. 1) If G is a special metric group then there is a sequence

which
is strongly O.K. for G.
2) We can in clause (d) of 1.3(2) above omit r(1), r(2) and conclude only d(x
1
y
1
, x
2
y
2
) <

n
and d(x
1
1
, x
1
2
) <
n
. This causes just slight changes in the computations of
length in the proof or replacing

by a suitable subsequence.
3) Every specially
+
metric group is a specially metric group.
Proof. Easy.
1.5 Denition. Assume A is a countable structure with automorphism group
G = Aut(A) and for notational simplicity its set of elements is , the set of
natural numbers (and, of course, it is innite, otherwise trivial).
We dene a metric d = d
A
= d
aut
A
on G by
d(f, g) = sup2
n
: f(n) ,= g(n) or f
1
(n) ,= g
1
(n).
Let Aut
A
= (Aut(A), d
aut
A
), but we may write G
A
or G
aut
A
.
1.6 Claim. : For A as above:
(Aut
A
, d
aut
A
) is a complete separable specially
+
metric group.
Proof. Easy.
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 9
2 Semi metric groups: on automorphism
groups of uncountable structures
It seems natural investigating the automorphism groups of a model A say of
cardinality, e.g.

, intending to put in a framework where we shall be able to


prove it is not a free group of cardinality >

. Now choose

A = A
n
: n < )
such that, e.g. the universe of A is

and A
n
=
n
. For such

A there is a
natural metric on Aut(A) under which it is a complete metric group, but usually
of too big density; there is another natural metric on Aut(A) under which it is a
complete metric space with density

but multiplication is not continuous and


Cauchi sequences may not converge to any point. To get the desired results we
use indirectly complete-metric dened in 2.1, which combine the two metrics;
in other words we weaken the completeness demand; hopefully this will have other
applications as well, e.g. also for Borel groups. We also look at some generalizations:
replacing Aut(A) by other derived structures.
The reader may skip this section if not interested in the results concerning

. One
way to present what we are doing is
2.1 Denition. We say G is an indirectly complete metric group if:
(a) G is a group
(b) G is a metric space under d
G
(c) if c = c
n
: n < ) satises d
G
(c
n
, c
n+1
) < 1/2
n
, then letting d
n
= c
1
2n
c
2n+1
and

d = d
n
: n < ) we have
() for some metric d

= d
G,

d
, under d

, G is a metric group, and



d con-
verges to some c under d

, moreover (G, d

) is complete.
2.2 Denition. Assume A is a structure.
1)

A is an -representation of A if

A = A
n
: n < ), A
n
A
n+1
for n < and
A
n
: n < is the universe of A.
2) For every -representation

A of A let Aut
A
(A) = f Aut(A): for every n <
for some m < we have (x A
n
)(f(x) A
m
& f
1
(x) A
m
).
3) If

A is an -representation of A and G = Aut(A) then we dene d = d
A,

A
= d
aut
A,

A
,
a metric
1
on G by
1
this is proved in 2.3
(
7
7
1
)


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







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


10 SAHARON SHELAH
d(f, g) = sup2
n
: n = 0 or there is a A
n
such that
for some (f

, g

) (f, g), (f
1
, g
1
)
one of the following possibilities holds
(a) for some m < we have f

(a) A
m
g

(a) / A
m
,
(b) f

(a) ,= g

(a) are in A
n
.
4) If

A is an -representation of A and G = Aut(A) then we dene d

= d

A,

A
a
metric on G by d

(f, g) = sup2
n
: n = 0 or f A
n
= g A
n
and f
1
A
n
=
g
1
A
n
.
2.3 Claim. Assume

A is an -representation of an innite structure A.
1) (Aut(A), d
A,

A
) is an indirectly complete metric group with density 2

0
+
2
|A
n
|
: n < ; in fact it is a complete metric space (but in general not a
metric group).
2) (Aut(A), d

A,

A
) is a complete metric group of density |A|
|A
n
|
: n < ; so
if each A
n
is nite the density is
0
.
3) If the universe of A is and A
n
= n = 0, . . . , n 1 then

A = A
n
: n < )
is an -representation of A with each A
n
nite and d

A,

A
= d
aut
A
from Denition
1.5(1) and under it Aut(A) is a complete separable specially metric group.
Proof. 1) Let d = d
A,

A
. First we show that
()
1
d is a metric (even ultrametric).
()
2
d(f
1
, g
1
) = d(f, g)
()
3
G as a metric space under d
G
has density 2
|A
n
|
: n < .
We may wonder whether (Aut(A), d) is complete, i.e. whether every d-Cauchi
sequence f
n
: n < ) in G, d-converge to some f G.
Before answering we prove a weaker substitute (in fact, it is the one we shall use
for proving the indirect completeness below).
Now we note that:
()
4
if f
n
: n < ) is a d-Cauchi sequence, then for every a A, f
n
(a) : n < )
is eventually constant and f
1
n
(a) : n < ) is eventually constant so the
limit f is a well dened permutation of A, moreover belongs to G = Aut(A).
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 11
Very nice, but multiplication
2
is not continuous in general for this metric, d. We
also have
()
5
(Aut(A), d) is a complete metric space.
[Why? Let

f = f
n
: n < ) be a d-Cauchi sequence, by ()
4
there is
f Aut(A) to which

f pointwise converges. Now let n() < be given
so for some n(1) < we have: n n(1) d(f
n
, f
n(1)
) < 2
n()
and
we shall prove n n(1) = d(f, f
n
) < 2
n()
. So for each c A
n()
we
shall check clauses (a),(b) in the denition of d in Denition 2.2(3). First
for every m < we have n n(1) f
n
(c) A
m
f
n(1)
(c) A
m
, but
f
n
(c) : n < ) is eventually constantly f(c) hence f(c) A
m
f
n(1)
A
m
.
This takes care of clause (a).
As for clause (b), similarly n n(1) [f
n
(a), f
n(1)
(a) A
n
f
n
(a) =
f
n(1)
(a)) hence n n(1) [f(a), f
n(1)
(a) A
n
f(a) = f
n(1)
(a)). The
same holds for f
1
n
: n < ), f
1
so we are done.]
But we have to prove that (Aut(A), d) is an indirectly complete metric group. The
only clause left is (c) of Denition 2.1. So assume g
n
G, d(g
n
, g
n+1
) < 1/2
n
, hence
by ()
4
+ ()
5
the sequence g
n
: n < ) converge to some g G by the metric
and also hence pointwise converge to g. Let f
n
= g
1
2n
g
2n+1
, easily f
n
: n < )
pointwise converge to e
G
= id
A
, let f = e
G
; so it suces to nd a metric d

such
that (G, d

) is a complete metric group in which f


n
: n < ) converge to f. We
prove this assuming just


f = f
n
: n < ) is an -sequence of members of G which pointwise converge
to f G.
Let B
n
= B
n

f
= a A : a A
n
and for every m (n, ) we have f
m
(a) =
f
n
(a) & f
1
m
(a) = f
1
n
(a). Clearly

B = B
n

f
: n < ) is an increasing -sequence
of subsets of A with union (the universe of) A. Recall that d

=: d

A,

B
was dened
by
d

(f, g) = sup2
n
: n = 0 or f B
n
,= g B
n
or f
1
B
n
,= g
1
B
n
.
Now by parts (2),(3)
2
E.g. let A be a trivial structure (i.e. with the empty vocabulary so G is the group of
permutations of A) and b
n
= c
n
A
n+1
\A
n
and a
n
A
0
be pairwise distinct (for n < ). Let
h
k
exchange a
n
, c
n
if n < k and is the identity otherwise. Let f interchange a
n
, b
n
if n < and
is the identity otherwise. Let g
k
interchange b
n
, c
n
if n < k and is the identity otherwise. Now
g
k
: k < is a Cauchi sequence and so has a limit g, but h
k
= fg
k
f and h
k
: k < is not a
Cauchi sequence as d(h
k
1
, h
k
1
) = 1 if k
1
< k
2
as witnessed by a = a
k
1
.
(
7
7
1
)


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







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


12 SAHARON SHELAH
()
6
the group G with d

is a complete metric group


and obviously
()
7
d(f
n
, f
m
) < Max2
n
, 2
m
and d(f
n
, f) 2
n
hence
()
8
f
n
: n < ) converge to f by d

.
So by ()
6
+ ()
8
we are done.
2),3) Left to the reader.
2.3
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 13
3 Compactness of metric algebras
Note that below if u
n
= t
n
= n we may write x
n
instead of x
n
, n instead t u
n
and d
n
instead of

d
n
. Also 3.1 - 3.6 works for Polish algebras as well.
3.1 The completeness Lemma. Assume a is a Polish group G = G
a
such that
(a) u
n
: n < ) is a sequence of pairwise disjoint non-empty nite sets
(b) x
n
= x
t
: t u
n
)
(c)
n
( x
n+1
) =
n,t
( x
n+1
) : t u
n
) is a sequence of
M
-terms, possibly with
parameters (from G
a
) so
n
(

d) =
n,t
(

d) : t u
n
) for any

d = d
s
: s
u
n+1
), d
s
M
(d) =
n
: n < ) is a sequence of positive reals converging to 0
(e)

d
n+1
= d
n+1,t
: t u
n+1
) with each d
n+1,t
an element of M such that if

n+1
= d

n+1,t
: t u
n+1
) is of distance <
n+1
from

d
n+1
, (that is d

n+1,t

Ball(d
n+1,t
,
n+1
) for each t u
n+1
), then
n
(

n+1
) Ball(

d
n
,
n
) which
means: t u
n

n,t
(. . . , d

n+1,s
, . . . )
su
n+1
Ball(d
n,t
,
n
)
(f) for every n < and a positive real there is m > n such that
()

if d

m,t
Ball(d
m,t
,
n
) for every t u
m
then the distance between

n
(
n1
(. . . ,
m1
(

m
), . . . ),
n
(
n+1
(. . . ,
m1
(

d
m
)) . . . ) is < .
Then there are d

n,t
M for n < , t u
n
which solves the set of equations
d

n,t
=
n+1
(d

n+1,s
)
su
n+1
and satises
d

n+1,t
Ball
G
(d
n+1
,
n
).
3.2 Remark. 1) In special versions we have

d
n
=
n
(

d
n+1
) (and in [Sh 744] we
have d
n
=
n
(d
n+1
)) but here there is no the true solution which we perturb.
2) Condition (e) in Lemma 3.1 says that if in large m we perturb

d
m
with error
<
m
and compute down by the s we still get a reasonable

d
n
but not necessarily
a very good one.
Proof. For every k we shall dene c
k
n,t
: t u
n
, n < ), a sequence of elements of
the algebra.
(
7
7
1
)


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







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


14 SAHARON SHELAH
First, if n k let c
k
n,t
= d
n,t
. Second, we dene c
k
n
= c
k
n,t
: t u
n
) by downward
induction on n k.
n = k by the rst case.
n < k let c
k
n
=
n
( c
k
n+1
).
By clauses (e) and (f) we have, respectively
()
1
c
k
n,t
Ball(d
n,t
,
n
).
()
2
for every positive real > 0 and n < , there is m > n such that if k m
then t u
n
c
k
n,t
Ball(c
m
n,t
, ).
So, for each n and t u
n
the sequence c
k
n,t
: k < ) is a Cauchy sequence by ()
2
;
hence it converges to some c
n,t
M. Now
the sequence c
n,t
: n < , t u
n
) forms a solution: for every n < and
t u
n
the equation c
k
n,t
=
n,t
(. . . , c
k
n+1,s
, . . . )
su
n+1
is satised whenever
n > k hence in the limit c
n,t
=
n,t
(. . . , c
n+1,s
, . . . )
su
n+1
.
3.1
Recall about groups
3.3 Fact: A free group is torsion free and the group is not divisible, in fact, every
element c has at most one n-th root for each n = 1, 2, . . . and has no root for every
large enough n except when c is the unit.
Recall
3.4 Denition. We say N is a retract of the group (or the algebra) M when
N M and there is a homomorphism form M onto N which is the identity on N.
3.5 Fact: Every countable subgroup of a free group G is contained in a countable
subgroup which is a retract of G.
We now give a criterion to show non-freeness. We could use x instead of x, of
course.
3.6 Claim. 1)
(a) a is a complete metric group, the group is G
a
and e

M
a
(b) B M
a
is countable with e

belonging to the closure of Be

(c) is a set of terms of the form (x, y)


(d) if (x, y) and

b B and c M
a
then x M
a
: c = (x,

b) is nite
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 15
(e) for every nite A G
a
and a positive real there are a sequence

b from B
and term (x, y) such that:
() (e

b) Ball(e

, )
() (c,

b) / A for every c M
a
.
Then no countable sub-group of G
a
containing B is a retract (in the algebraic sense)
of M

. Hence G
a
is not free.
2) We can omit e

, i.e. omit clause (b) and the last phrase of clause (a) and change
clause (e) to
(e)

for any nite A M


a
and real > 0 and d G
a
there is a term (x, y)
and sequence

b
g( y)
B and element d

G
a
such that
()
M
a
(d

b) Ball
a
(d, )
() for no c M
a
do we have
M
a
(c,

b) A.
3.7 Remark. We can similarly phrase sucient conditions for G
a
is unstable in

0
[for quantier free formulas, see [Sh 849]].
Proof. 1) Like the proof of part (2) below except that we add to ():
() e
n
= e

.
2) We rely on 3.1.
Assume toward contradiction that G is a countable reduct of G
a
which includes
B, so we can choose h

, a homomorphism from M
a
onto G which extends id
M
a
.
Let a
n
: n < ) list G. Let u
n
= n. We choose

b
n
and
n
(x, y
n
) and
n
by
induction on n such that
()()
n
(x, y
n
)
()

b
n
a sequence from M of length g( y
n
)
()
n
a positive real,
n+1
<
n
/2
() e
n
M
a
()
n
(e
n+1
,

b
n
) Ball
a
(e
n
,
n
), moreover if c Ball
a
(e
n+1
,
n+1
) then
n
(c,

b
n
)
Ball
a
(e
n
,
n
)
() if k < n and c, c

Ball
a
(e
n+1
,
n+1
) and we dene the terms
n+1,
(x)
for n + 1, with parameters by downward induction on as follows

n+1,n+1
(x) = x,
n+1,
(x) =

(
n+1,+1
(x),

) then the d-distance be-


tween
n+1,k
(c) and
n+1,k
(c

) is <
n
.
(
7
7
1
)


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







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


16 SAHARON SHELAH
Let us carry the induction, in stage n we choose e
n
,
n
and
n1
(x,

b
n1
) if n > 0.
Case 1: n = 0.
This is straightforward.
Case 2: n = k + 1.
Let D be the set of c = c
m
: m k) which satises

k
(i) m < k c
m
=
m
(c
m+1
, h

b
m
))
(ii) c
0
= a
k
.
We can prove by induction on m k that c
m
: c D is nite, and let A =
c
k
: c D. By clause (e)

of the assumptions (see 3.6(2)) applied with (A,


k
, e
k
)
here standing for (A, , d) there. There are r < , = (x, y
0
, . . . , y
r1
) and

b
r
(G
a
) and d

as there. We let

b
k
=

b,
k
= , y
k
= y
0
, . . . , y
r1
), e
n
= d

.
Lastly, we should choose
n
R
+
. There are several demands but each holds
for every small enough > 0, more exactly one for clause () and for each m < n,
one for clause ().
Having carried the induction, clearly 3.1 apply hence there is a solution d

n
:
n < ), that is G
a
[= d

m
=
m
(d

m+1
,

b
m
) for m < . But h

is a homomorphism
from G
a
into G so h

(d

n
) : n < ) satises all the equations in
k
hence by our
choice in stage n = k + 1, h

(d

0
) ,= a
k
. As this holds for every k and a
k
: k <
list the elements of M
a
we are done.
3.6
3.8 Remark. 1) If we phrase algebraic compactness, it is preserved by taking reducts.
2) In a reasonable variant we can replace M countable by |M| < cov(meagre);
we hope to return to this elsewhere.
3) We can change the demand on : at most one solution in clause (e), A a singleton
in clause (f). This suces for groups.
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 17
4 Conclusions
4.1 Conclusion 1) If (G, d) is a complete metric group of density < [G[, then:
(a) G is not free,
(b) if G is
1
-free (see 4.2) then for some countable A G, there is no countable
reduct H of G including A, recalling 3.4.
2) It suces that G is an indirectly complete metric group and as a metric space
it is of density < [G[.
3) Instead density < [G[ it is enough to assume that the topology induces by
the metric is not discrete.
Recall (see [EM02])
4.2 Denition. A group G is -free when every subgroup of G generated by <
elements is a gree group.
Proof. 1), 2) Clause (a) follows from clause (b) by Fact 3.5. Let = density(G)
and y
i
G be pairwise distinct for i <
+
. Without loss of generality y
i
/ y
j
:
j < i)
G
. So for some increasing sequence i
n
: n < ) the sequence y
i
n
: n < )
is a Cauchi sequence.
For part (1), by completeness it converges say to y

, the convergence is for d.


Now b
n
: n < ) =: (y

)
1
y
i
2n+1
: n < ) converges to e
G
, the members are
pairwise distinct so without loss of generality ,= e
G
.
However for part (2) we know that some (G, d

), i.e. equal to (G, d) as a group but


with a dierent metric; is a complete metric group with an -sequence b
n
: n < )
of members of Ge
G
converging to e
G
.
Let = x
m
y : m < and B = b
n
: n < and e

= e
G
. Now we shall apply
3.6(1). In the assumptions, clauses (a)-(c) are obvious. As for clause (d) we are
using: equations of the form x
m
a

= a

has at most one solution in G, see 3.3. We


are left with clause (e), so we are given a real > 0 and a nite set A G. We can
choose b ((BA)e
G
) of distance < from e
G
. Let (x, y) = x
n
y and

b = b)
where n < is the minimal n > 1 such that [a A ab
1
has no n-th root]. This
is possible, see Fact 3.3 so (e
G
,

b) = b Ball
d
(e
G
, ) as required in subclause ()
of clause (e) and (e
G
,

b) = b / A as required in subclause () of clause (e) so by


3.6(1) we are done.
3) Choose y
n
: n < ) converging to some y

such that y
n
: n < )y

) is with
no repetitions, possible on (G, d) is not discrete. Now continue as above.
4.1
(
7
7
1
)


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







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


18 SAHARON SHELAH
In particular
4.3 Conclusion: There is no free uncountable Polish group.
4.4 Conclusion: 1) Assume A is a countable structure. Then Aut(A), the group
of automorphisms of A, is not a free uncountable group, in fact it satises the
conclusions of 4.1.
2) Assume A is a structure of cardinality and = =

or more generally
=
n
: n < , 2

n
< 2

n+1
, = 2

n
: n < < 2

. Then Aut(A) cannot


be free of cardinality > , in fact, it satises the conclusions of 4.1.
Proof. 1) By 1.6, Aut(A) is a Polish group and apply 4.3.
2) Without loss of generality the universe of A is , using

A =
n
: n < )
we know by 2.3(1) that (Aut(A), d
A,

A
) is an indirectly complete metric group and
apply 4.1(2).
4.4
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 19
5 Quite free but not free abelian groups
If uncountable Polish groups are not free, we may look at wider classes: F

,
Borel analytic, projective, L[R].
5.1 Question: 1) Is the freeness of a reasonably denable abelian group absolute?
2) For which cardinals does -freeness imply freeness (or
+
-freeness) for nicely
denable abelian groups, in particular for =

?
3) Similarly for other varieties (or any case when free is denable, say for any
universal Horn theory).
This is connected also to [Sh 402] whose original aim was a question of Marker
are there non-free Whitehead Borel Abelian groups. But already in [Sh 402] it
seems to me the basic question is to clarify freeness in such groups; that is, question
5.1 above.
Blass asked about denable subgroups of Z

(see question 5.10): by [Sh 402] and


the construction here we quite resolve this.
Recall that [Sh 402] analyze
1
-free abelian groups which are
1
1
or so. A natural
dividing line was suggested; the complicated half was proved to be not Whitehead,
and at least for me is an analog to not
0
-stable. The low half is
2
-free. So under
CH we were done, but what if 2

0
>
1
? Are they also free? This was left open by
[Sh 402].
We shed some light by giving an example (an F

one) showing that the non-CH


case in [Sh 402] is a real problem. This resolves the original problem:
(a) it is consistent that there are non-free Whitehead groups
(b) it is consitent that there are no non-free Whitehead groups.
[Why? Clause (a) is derived in 5.13 whereas by [Sh 402] we have clause (b), in fact
derive it from CH.]
But what about the further question, e.g. 5.1(2)? The examples seem to indicate
(at least to me) that the picture in [Sh 87a], [Sh 87b] is the right one here, connecting
theories of L

1
,
with
1
1
-models. Also related are [EM2], [MkSh 366] on almost
freeness for varieties, and see [EM] on abelian groups. In particular we conjecture
every

-free Borel group is free.


We shall use freely the well known theorem saying
a subgroup of a free abelian group is a free abelian group.
(
7
7
1
)


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







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


20 SAHARON SHELAH
5.2 Denition. For k() < we dene an abelian group G = G
k()
, it is generated
by x
m, ,
: m k() and
>
2 and =

: k(), ,= m) where

2 y
,n
: n < and =

: k()) where



2 freely except the
equations:

,n
(n!)y
,n+1
= y
,n
+

x
m,
m
,
: m k() and
m
= m

k() : m

,= m and
=
m
n.
(Note that if m
1
< m
2
k() then
m
1
,=
m
2
having dierent index sets).
Explanation. A canonical example of a non-free group is (Q, +). Other examples are
related to it after we divide by something. The ys here play that role of providing
(hidden) copies of Q. What about xs? For each
k()+1
(

2) we use m k()
to give y
,n
: n < ), k() + 1 chances, opportunities to avoid having (Q, +)
as a quotient, one for each cardinal
k()
. More specically, if H G is the
subgroup which is generated by X = x
m, ,
: m ,= m() and is a function from
k() : ,= m to and
>
2, still in G/H the y
,n
: n < does not
generate a copy of Q, as witnessed by x
m(),
m()
,
m()
n
: n < .
5.3 Claim. The abelian group G
k()
is a Borel group, even an F

-one, that is the


set of elements and the graphs of + and the function x x (i.e. (x, y, z) :
G
k()
[= x +y = z hence also (x, x) : x G
k()
) are F

-sets; hence Borel.


Proof. Let cd be a one-to-one function from the set of nite sequences of natural
numbers onto the set of natural numbers and we dene:

1
(a) cd(x
m, ,
) = cd(m, cd(), . . . ,

(i), . . . )
k(),=m
: i < ) so it


and let X = cd(x
n, ,
) : (n, , ) as in Denition 5.2
(b) cd(y
,n
) = cd(n, . . . ,

(i) . . . )
k()
: i < ) and
Y = cd(y
,n
) : ( , n) as in Denition 5.2
(c) for a sequence a = a

: < n) of integers let


abs( a) = [a

[ : < n) and sign( a) = sign(a

):
< n), where sign(a

) is 0,1,2 if a

is negative, zero,
positive respectively.
We say represents x G
k()
as witnessed by (z

, a

, m) : < n) when:

2
(a) G [= x =

<n
a

,
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 21
(b) z

x
n, ,
: (, , ) as in Denition 5.2 y
,m
: ( , m

) as
in Denition 5.2
(c) z

: < n) is without repetitions


(d) cd(z

) m : < n) are pairwise distinct


(e) if (z

, a

, m

) : < n

) satises clauses (a)-(d), then m m

(f) if n = 0 then m = 0
(g) cd(z
0
) <
lex
cd(z
1
) <
lex
. . . where <
lex
is the lexicographic order
on

(h) = cd(n) sign( a) abs( a)cd(z

)(i) : k())) : i < ) where


a = a

: < n), of course.


Now for n < , a = a

: < n)
n
(Z0), i < and =

: < n)
n
(
i
) is
<
lex
-increasing hence without repetitions (and if n = 0 we let i = 0) we let
Z
a,
= : represent some x G
k()
as witnessed by
(z

, a

, i) : < n) and cd(z

) m =

for < n.
Let Y be the set of such pairs ( a, )
()
1
Z
a,
: ( a, ) Y ) is a sequence of pairwise disjoint closed subsets of

()
2
every member of G is represented by one and only one member of Z :=
Z
a,
: ( a, ) Y .
[Why? For any i < n clearly x
m, ,
: (m, , ) as in Denition 5.2 y
,i
:
k()+1
(

Z) generates freely a subgroup G

k(),i
of G
k()
.
Why? E.g. let G

be a vector space over the eld Q with basis x

m, ,
: (m, , )
as in Denition 5.2 y

,i
:
k()+1
(

2). We dene a mapping h into G

from
the set of generators x
m, ,
: (m, , ) as in Denition 5.2 y
,j
:
k()+1
(

2)
and j < as follows:
(a) h(x
m, ,
) = x

m, ,
(b) h(y
,i
) = y

,i
(c) we choose h(y
,j
) for j (i, ) by induction on j such that h maps
,j1
to an equation satised in G

(d) we choose h(y


,j
) for j [0, i) by downward induction on j such that h
maps
,j+1
to an equation satised in G

.
(
7
7
1
)


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







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


22 SAHARON SHELAH
So as h maps the set of generators of G
k()
into G

preserving the equations in 5.2


we can extend h to a homomorphism from G
k()
into G

.
Now the set which generates G

k(),i
is mapped in a one-to-one way, onto a basis
of G

, hence it generates G

k(),i
freely, as stated above.
Inspecting 5.2 clearly the quotient G
k()
/G

k(),i
is torsion. The rest should be
clear, too.]
()
3
U = (
1
,
2
,
3
) :

represent x

G
k()
for = 1, 2, 3 and G
k()
[=
x
1
+x
2
= x
3
is the graph of a two-place function
()
4
for any ( a

) Y for = 1, 2, 3 the set (


1
,
2
,
3
) U :

Z
a

for
= 1, 2, 3 is a closed set.
Clearly we are done.
5.3
As a warm up we note:
5.4 Claim. G
k()
is an
1
-free abelian group.
Proof. Let U

2 be countable (and innite) and dene G

U
like G restricting
ourselves to

U; by the Lowenheim-Skolem-Tarski argument it suces to prove


that G

U
is a free abelian group. List
k()+1
U without repetitions as
t
: t < ),
and choose s
t
< such that [r < t &
r
k() =
t
k() =
t,k()
:
[s
t
, )
r,k()
: [s
r
, )].
Let
Y
1
= x
m, ,
: m < k(),
k()+1\{m}
U and
>
2
Y
2
=

x
m, ,
:m = k(),
k()
U and for no t < do we have
=
t
k() &
t,k()
: s
t
<

Y
3
= y

t
,n
: t < and n [s
t
, ).
Now
()
1
Y
1
Y
2
Y
3
generates G

U
.
[Why? Let G

be the subgroup of G

U
which Y
1
Y
2
Y
3
generates. First we prove
by induction on n < that for
k()
U and
n
2 we have x
k(), ,
G

. If
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 23
x
k(), ,
Y
2
this is clear; otherwise, by the denition of Y
2
for some < and
t < such that s
t
we have =
t
k(), =
t,k()
.
Now
(a) y

t,+1
, y

t
,
are in Y
3
G

.
Hence by the equation
,n
in Denition 5.2, clearly x
k(), ,
G

. So as Y
1

G

U
, all the generators of the form x
m, ,
with each

U are in G

. Also we
have
(b) x
m,
t
{ik(),i=m},
belong to Y
1
G

if m < k().
Now for each t < we prove that all the generators y

t
,n
are in G

. If n s
t
then clearly y

t
,n
Y
3
G

. So it suces to prove this for n s


t
by downward
induction on n; for n = s
t
by an earlier sentence, for n < s
t
by
,n
. The other
generators are in this subgroup so we are done.]
()
2
Y
1
Y
2
Y
3
generates G

U
freely.
[Why? Translate the equations.
Alternatively, let z = z

: < ()) list the set x


m, ,
: (m, , ) as in
Denition 5.2 but
k()+1
U y
,n
:
k()+1
U and n < which generates
G

U
and z lists it without repetitions such that for some increasing continuous
sequence
i
: i +) we have
0
= 0,
+
= () and
(a) z

: <
1
= Y
1
Y
2
Y
3
(b) z

: [
1+n
,
1+n+1
) = x
k(), ,
G

U
: x
k(), ,
/ Y
2
and g() = n
(c) z

: [
+r
,
+r+1
) = y

t
,n
: t < , n < s
t
and r = s
t
n.
Now the proof above shows that:
there is a one-to-one function from the set of equations dening G

U
onto
[
1
,
+
) such that:
if the equation is mapped to the ordinal then: if z

appears in the
equation then and z

appears in the equation and its coecient is 1


or 1.
This clearly suces.]
5.4
Now systematically
(
7
7
1
)


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







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


24 SAHARON SHELAH
5.5 Denition. 1) For U

2 let G
U
be the subgroup of G generated by
Y
U
= y
,n
:
k()+1
(U) and n < x
m, ,
: m k() and
(k()+1)\{m}
(U)
and
>
2. Let G
+
U
be the divisible hull of G
U
and G
+
= G
+
(

2)
.
2) For U

2 and nite u

2 let G
U,u
be the subgroup
3
of G generated by
G
U(u\{})
: u; and for
k()+1
U let G
U,
be the subgroup of G
generated by G
U{
k
:k<g( ) and k=}
: < g( ).
3) For U

2 let
U
= the equation
,n
:
k()+1
U and n < . Let

U,u
=
U(u\{})
: u.
5.6 Claim. 0) If U
1
U
2


2 then G
+
U
1
G
+
U
2
G
+
.
1) For any n() < , the abelian group G
+
U
(which is a vector space over Q), has the
basis Y
n()
U
i
:= y
,n()
:
k()+1
(U) x
m, ,
: m k(),
(k()+1)\{m}
(U)
and
>
2.
2) For U

2 the abelian group G
U
is generated by Y
U
freely (as an abelian group)
except the set
U
of equations.
3) If U
m


2 for m < m() then the subgroup G
U
0
+. . .+G
U
m()1
of G is generated
by Y
U
0
Y
U
1
. . . Y
U
m()1
freely (as an abelian group) except the equations in

U
0

U
1
. . .
U
m()1
. Also G
U
0
+ . . . + G
U
m()1
is a direct summand of G
when
if
0
, . . . ,
k()
U
m
: m < m() such that
( k())(m < m())[
0
, . . . ,
k()

U
m
)
then for some m < m() we have
0
, . . . ,
k()
U
m
.
4) If U

= UU

for < m() k() and U

: < m()) are pairwise disjoint then


holds.
5) G
U,u
G
Uu
if U

2 and u

2U; moreover G
U,u

pr
G
Uu

pr
G.
6) If U

: < ()) is -increasing continuous then also G


U

: < ()) is
-increasing continuous.
7) If U
1
U
2
U

2 and u

2U is nite, [u[ < k() and U
2
U
1
= and
v = u then (G
U,u
+G
U
2
u
)/(G
U,u
+G
U
1
u
) is isomorphic to G
U
1
v
/G
U
1
,v
.
8) If U

2 and u

2U has k() members then (G
U,u
+ G
u
)/G
U,u
is
isomorphic to G
u
/G
,u
.
Proof. 0) Obvious.
1) See the proof of ()
2
inside the proof of 5.3.
2),3),4) Follows.
3
note that if u = {} then G
U,u
= G
U
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 25
5) First, G
U,u
G
Uu
follows by the denition. Second, we deal with proving
G
U,u

pr
G
Uu
. So let [u[ = m() + 1 and

: m()) list u, necessarily


with no repetitions and let U

= U (u

) (so G
U,u
= G
U
0
+. . . +G
U
m()
) and
assume z G
Uu
, a Z0 and az belongs to G
U
0
+ . . . + G
U
m()
so it has the
form b
i
x
m
i
,
i
,
i
: i < i() + c
j
y

j
,n
j
: j < j() with b
i
, c
j
Z and
i
,
j
are (nite) sequences of members of U
(i)
, U
k(j)
respectively and are as required in
Denition 5.2 where (i), k(j) m().
Now similarly as z G
Uu
, we can nd z = b

i
x
m

i
,

i
,

i
: i < i

() +
c

j
y

j
,n

j
: j < j

().
By the equations in Denition 5.2 without loss of generality for some n() we
have: j < j() n
j
= n() and j < j

() n

j
= n(). Also without loss of generality each
of the sequences (m
i
,
i
,
i
) : i < i()),
j
: j < j()) is with no repetitions,
and also in (m

i
,

i
,

i
) : i < i

()),

j
: j < j

()) there is no repetition (for

j
: j < j()) and

j
: j < j

()) we use n
j
= n(), n

j
= n()). Together
b
i
x
m
i
,
i
,
i
: i < i() + c
j
y

j
,n()
: j < j() = ab

i
x
m

i
,

i
,
i
: i <
i

() + ac

j
y

j
,n()
: j < j

().
Now this equation holds in G
Uu
hence is G and even in G
+
. By part (1) and the
no repetitions, after possible permuting we get i() = i

(), j() = j

(), (m
i
,
i
,
i
) =
(m

i
,

i
,

i
), b
i
= ab

i
for i < i(),
j
=

j
for j < j(), c
j
= ac

j
for j < j(). But
this proves that x
m

i
,

i
,
i
: i < i

()y

j
,n()
: j < j

() G
U,u
hence z G
U,u
as required.
Third, the proof of G
Uu

pr
G is similar.
6) Easy.
7) Clearly U
1
v = U
2
u hence G
U
1
u
G
U
1
v
= G
U
2
u
hence G
U,u
+G
U
1
u
is
a subgroup of G
U,u
+G
U
2
u
, so the rst quotient makes sense.
Hence by an isomorphism theorem (G
U,u
+G
U
2
u
)/(G
U,u
+G
U
1
,u
) is isomorphic
to G
U
2
u
/(G
U
2
u
(G
U,u
+ G
U
1
,u
)). Now G
U
1
,v
G
U
1
v
= G
U
2
u
and G
U
1
,v
=
G
U
1
(v\{})
: v = G
U
1
(v\{})
: u + G
U
1
,(v\{})
G
U,u
+ G
U
1
,u
.
Together G
U
1
,v
is included in their intersection, i.e. G
U
2
u
(G
U,u
+G
U
1
,u
) include
G
U
1
,v
and using part (1) both has the same divisible hull inside G
+
. But G
U
1
,v
is a
pure subgroup of G by part (5) hence of G
U
1
v
. Hence necessarily G
U
1
u
(G
U,u
+
G
U
1
,u
) = G
U
1
,v
, so as G
U
2
u
= G
U
1
v
we are done.
8) The proof is similar to the proof of part (7). Note that G
U,u
G
U,u
+G
u
hence
the rst quotient makes sense. So by an isomorphism theorem (G
U,u
+ G
u
)/G
U,u
is isomorphic to G
u
/(G
U,u
G
u
). Now G
U,u
G
u
includes G
,u
and using part (1)
both has the same divisible hull inside G
+
. But G
,u
is a pure subgroup of G
u
by
part (5). So necessarily G
U,u
G
u
= G
,u
, so G
u
/(G
U,u
G
u
) = G
u
/G
,u
, so we
are done.
5.6
(
7
7
1
)


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







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


26 SAHARON SHELAH
Discussion: For the readers convenience we write what the group G
k()
is for the
case k() = 0. So, omitting constant indexes and replacing sequences of length one
by the unique entry we get that it is generated by y
,n
(for

2, n < ) and x

(for

>
2) freely as an abelian group except the equations (n!)y
,n+1
= y
,n
+x
n
.
Note that if K is the countable subgroup generated by x

:
>
2 then G/K
is a divisible group of cardinality continuum hence G is not free. So G is
1
-free
but not free.
Now we have the main proof
5.7 Main Claim. 1) The abelian group G
Uu
/G
U,u
is free if U

2, u

2U
and 1 [u[ k k() and [U[
k()k
.
2) If U

2 and [U[
k()
, then G
U
is free.
Proof. 1) We prove this by induction on [U[; without loss of generality [u[ = k as
also k

= [u[ satises the requirements.


Case 1: U is countable.
So let

: < k list u be with no repetitions, now if k = 0, i.e. u = then


G
Uu
= G
U
= G
U,u
so the conclusion is trivial. Hence we assume u ,= , and let
u

:= u

for < k.
Let
t
: t < t

) list with no repetitions the set


U,u
:=
k()+1
(U u):
for no < k does
k()+1
(U u

). Now comes a crucial point: let t < t

, for
each < k for some r
t,
k() we have
t,r
t,
=

by the denition of
U,u
, so
[r
t,
: < k[ = k < k() +1 hence for some m
t
k() we have < k r
t,
,= m
t
so for each < k the sequence
t
(k() + 1m
t
) is not from
s
: s k() and
s ,= m
t
) :
s
(U u

) for every s k() such that s ,= m


t
.
For each t < t

we dene S(t) = m k() :


t,s
: s k() & s ,= m
is included in U u

for no < k. So m
t
S(t) 0, . . . , k() and m
S(t)
t
j k() : j ,= m / (U u

) for every < k. For m k() let


t,m
:=
t
j k() : j ,= m and

t
:=

t,m
t
. Now we can choose s
t
< by
induction on t such that
() if t
1
< t, m k() and

t
1
,m
=

t,m
, then

t,m
s
t
/
t
1
,m
: < .
Let Y

= x
m, ,
G
Uu
: x
m, ,
/ G
Uu

for < k y
,n
G
Uu
: y
,n
/
G
Uu

for < k.
Let
Y
1
= x
m, ,
Y

: for no t < t

do we have m = m
t
& =

t
.
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 27
Y
2
= x
m, ,
Y

: x
m, ,
/ Y
1
but for no
t < t

do we have m = m
t
& =

t
&

t,m
t
s
t

t,m
t

Y
3
= y
,n
: y
,n
Y

and n [s
t
, ) for the t < t

such that =
t
.
Now the desired conclusion follows from
()
1
y +G
U,u
: y Y
1
Y
2
Y
3
generates G
Uu
/G
U,u
()
2
y +G
U,u
: y Y
1
Y
2
Y
3
generates G
Uu
/G
U,u
freely.
Proof of ()
1
. It suces to check that all the generators of G
Uu
belong to G

Uu
=:
Y
1
Y
2
Y
3
G
U,u
)
G
.
First consider x = x
m, ,
where
k()+1
(U u), m < k() and
n
2 for some
n < . If x / Y

then x G
U,u

for some < k but G


Uu

G
U,u
G

Uu
so
we are done, hence assume x Y

. If x Y
1
Y
2
Y
3
we are done so assume
x / Y
1
Y
2
Y
3
. As x / Y
1
for some t < t

we have m = m
t
& =

t
. As x / Y
2
,
clearly for some t as above we have
t,m
t
s
t

t,m
t
. Hence by Denition
5.2 the equation

t
,n
from Denition 5.2 holds, now y

t
,n
, y

t
,n+1
G

Uu
. So
in order to deduce from the equation that x = x
m, ,
= x
m
t
,

t
,
t,m
t
n
belongs to
G
Uu
, it suces to show that x
j,

t,j
,
t,j
n
G

Uu
for each j k(), j ,= m
t
. But
each such x
j,

t,j
,
t,j
n
belong to G

Uu
as it belongs to Y
1
Y
2
.
[Why? Otherwise necessarily for some r < t

we have j = m
r
,

t,j
=

r,m
r
and

r,m
r
s
r

t,j
n
r,m
r
so n s
r
and as said above n s
t
. Clearly r ,= t as
m
r
= j ,= m
t
, now as

t,m
r
=

r,m
r
and
t
,=
r
(as t ,= r) clearly
t,m
r
,=
r,m
r
.
Also (r < t) by () above applied with r, t here standing for t
1
, t there as
r,m
r

s
r

t,j
n
r,m
r
. Lastly for if t < r, again () applied with t, r here standing
for t
1
, t there as n m
t
gives contradiction.]
So indeed x G

Uu
.
Second consider y = y
,n
G
Uu
, if y / Y

then y G
U,u
G

Uu
, so assume
y Y

. If y Y
3
we are done, so assume y / Y
3
, so for some t, =
t
and
n < s
t
. We prove by downward induction on s s
t
that y
,s
G

Uu
, this clearly
suces. For s = s
t
we have y
,s
Y
3
G

Uu
; and if y
,s+1
G

Uu
use the
equation

t
,s
from 5.2, in the equation y
,s+1
G

Uu
and the xs appearing in
the equation belong to G

Uu
by the earlier part of the proof (of ()
1
) so necessarily
y
,s
G

Uu
, so we are done.
Proof of ()
2
. We rewrite the equations in the new variables recalling that G
Uu
is generated by the relevant variables freely except the equations of
,n
from
(
7
7
1
)


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







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


28 SAHARON SHELAH
Denition 5.2. That is, let z

: < ()) list Y

without repetitions such that for


some -increasing continuous sequence
i
: i +) of ordinals we have:
(a)
0
= 0,
+
= ()
(b) z

: <
1
= Y
1
Y
2
Y
3
(c) z

: [
1+n
,
1+n+1
) = x
m, ,
Y

: x
m, ,
/ Y
1
Y
2
and is of
lenth n
(d) z

: [
+n
,
+r+1
) = y
,n
Y

: y
,n
/ Y
3
and r = s
t
n.
Clearly
there is a one-to-one function from the set
Uu

Uu\{}
: u into
[
1
,
+
) such that:
if the equation is mapped to the ordinal then if z

appears in the
equation then < and z

appears in the equation with coecient one or


minus one.
This clearly suces.
Case 2: U is uncountable.
As
1
[U[
k()k
, necessarily k < k().
Let U =

: < where = [U[, list U with no repetitions. Now for each


[U[ let U

:=

: < , u

= u

. Now

1
(G
U,u
+ G
U

u
)/G
U,u
: < [U[) is an increasing continuous sequence of
subgroups of G/G
U,u
[Why? By 5.6(6).]

2
G
U,u
+G
U
0
u
/G
U,u
is free.
[Why? This is (G
U,u
+G
u
)/G
U,u
= (G
U,u
+G
u
)/G
U,u
which by 5.6(8) is
isomorphic to G
u
/G
,u
which is free by Case 1.]
Hence it suces to prove that for each < [U[ the group (G
U,u
+G
U
+1
u
)/(G
U,u
+
G
U

u
) is free. But easily

3
this group is isomorphic to G
U

/G
U

,u

.
[Why? By 5.6(7) with U

, U
+1
, U,

, u here standing for U


1
, U
2
, U, , u
there.]

4
G
U

/G
U

,u

is free.
[Why? By the induction hypothesis, as
0
+ [U

[ < [U[
k()(k+1)
and
[u

[ = k + 1 k().]
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 29
2) If k() = 0 just use 5.4, so assume k() 1. Now the proof is similar to (but
easier than) the proof of case (2) inside the proof of part (1) above.

5.7
5.8 Claim. If U

2 and [U[
k()+1
then G
U
is not free.
Proof. Assume toward contradiction that G
U
is free and let be large enough; for
notational simplicity assume [U[ =
k()+1
. O.K. as a subgroup of a free abelian
group is a free abelian group. We choose N

by downward induction on k()


such that
(a) N

is an elementary submodel
4
of (H (), , <

)
(b) |N

| = [N


k()
[ =

and

+ 1 N

(c) G
U
N

and N
+1
, . . . , N
k()
N

.
Let G

= G
U
N

, a subgroup of G
U
. Now
()
0
G
U
/(G

: k()) is a free (abelian) group.


[Easy. Let Z G
U
be the set of free generators. Without loss of generality
Z N
1
. Then Z

= Z N

is the set of free generators of G


U
N

= G

is the set of free generators of G

. So G
U
/G

is free.]
Now
()
1
letting U
1

be U for = k() + 1 and


k()

m=
(N
m
U) for k(); we have:
U
1

has cardinality

for k() + 1
[Why? By downward induction on . For = k() + 1 this holds by an
assumption. For = k() this holds by clause (b). For < k() this holds
by the choice of N

as the set
k()

m=+1
(N
m
U) has cardinality
+1

and belong to N

and clause (b) above.]


()
2
U
2

=: U
1
+1
(N

U) has cardinality
+1
for k()
[Why? As [U
1
+1
[ =
+1
>

= |N

| [N

U[.]
4
H () is {x: the transitive closure of x has cardinality < } and <

is a well ordering of
H ()
(
7
7
1
)


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







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


30 SAHARON SHELAH
()
3
for m < k() the set U
3
m,
=: U
2

r=m
N
r
has cardinality
m
[Why? By downward induction on m. For m = 1 as U
2

N
m
and
[U
2

[ =
+1
and clause (b). For m < similarly.]
Now for = 0 choose

U
2

, possible by ()
2
above. Then for > 0, k()
choose

U
3
0,
. This is possible by ()
3
. So clearly
()
4

U and

N
m
U ,= m for , m k().
[Why? If = 0, then by its choice,

U
2

, hence by the denition of U


2

in
()
2
we have

/ N

, and

U
1
+1
hence

N
+1
. . . N
k()
by ()
1
so ()
4
holds for = 0. If > 0 then by its choice,

U
3
0,
but U
3
m,
U
2

by ()
3
so

U
2

hence as before

N
+1
. . . N
k()
and

/ N

.
Also by ()
3
we have

r=0
N

so ()
4
really holds.]
Let

: k()) and let G

be the subgroup of G
U
generated by x
m, ,
:
m k() and
k()+1\{m}
U and
>
2 y
,n
:
k()+1
U but ,=

and
n < . Easily G

recalling G

= N

G
U
hence G

: k() G

, but
y

,0
/ G

hence
()
5
y

,0
/

: k().
But for every n
()
6
n!y

,n+1
y

,n
= x
m,

(k())+1\{m}),

m
n
: m k() G

:
k().
[Why? x
m,

(k())+1\{m}),

m
n
G
m
as

(k()) + 1m) N
m
by
()
4
.]
We can conclude that in G
U
/

: k(), the element y


,0
+

:
k() is not zero (by ()
5
) but is divisible by every natural number by ()
6
.
This contradicts ()
0
so we are done.
5.8
5.9 Conclusion. G
k()
is a Borel and even F

abelian group which is


k()+1
-free
but if 2

0

k()+1
is not free and even not
k()+2
-free.
Proof. G
k()
is Borel and F

by 5.3, it is
k()+1
-free by 5.7 and if 2

0

k()+1
it
is not
k()+2
-free by 5.8.
5.9
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 31
Blass asks
5.10 Question: Suppose (a) + (b) below, does it follow that forcing with Q add
reals?
(a) G is a Borel denition of an uncountable abelian subgroup of

Z (the
Specker group) which is not free
(b) the forcing Q satises
Q
G
V
is free.
Now
5.11 Fact: For just Borel abelian group G: if CH, then the answer to 5.10 is yes, if
not CH then the answer is not for Q = Levy(
1
, 2

0
).
Proof. First, assume CH holds and G is as in (a) of 5.10; (or just dened absolutely
enough such that G
V
is a subgroup of G
V
Q
for any forcing notion Q and is still not
free). Then by [Sh 402] the group G
V
is non-free in some strong way such that no
forcing not collapsing 2

0
to
0
can make it free (that is, for some countable G
0

G
V
, G
V
/G
0
contains the direct sum of 2

0
nite rank non-free abelian groups).
This is a strong yes answer.
On the other hand, if 2

0
>
1
we can nd such group: for k() 1, our G
k()
if

1
<
k()+1
2

0
, see below, is a strong negative answer. So together this gives
answers to a question of Blass.
5.11
5.12 Corollary. 1) The group G
k()
is embeddable into

Z, even purely.
2) Hence forcing which does not add bounded subsets to
k()
can make it free (i.e.
Levy(

, 2

0
) if k() while if our universe satises 2

0
>
k()
it is not free
there).
Proof. 1) For every n < we dene a function f
n
from Y to G
k()
where Y is the
set of generators of G
k()
, i.e.
Y = y
,n+1
:n < ,
k()+1
(

2) x
m,
m
,
: m k(),

m

{:k(),=m}
(

2) and
>
2.
First dene a function h
n
: for

2, h
n
() is a sequence of length g() and
(h
n
())() =

() if < n & < g()


0 if n & < g()
.
(
7
7
1
)


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







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


32 SAHARON SHELAH
For =

: u)
u
(

2) we let h
n
( ) = h
n
(

) : u).
Lastly, let
f
n
(y
,n+1
) = y
h
n
( ),n+1
f
n
(x
m,
m
,
) = x
m,h
n
(
m
),h
n
()
.
Does f
n
induce a homomorphism from G
k()
into G
k()
? For this it is enough to
check that for every one of the relations from Denition 5.2, its f
n
-image is satised
in G
k()
, but this is obvious as it is mapped to another one of the equations in the
denition of G
k()
: the equation in
,m
is mapped to the equation in
h
n
( ),m
.
So f
n
extends to an endomorphism

f
n
of G
k()
. Easily
if L G
k()
is a nite rank subgroup (so free) then for n large enough

f
n
L is one to one.
Now the range of

f
n
is clearly countable hence free, and is innite; let it be

<
Zz
n,
.
Hence for some homomorphisms g
n,
from Rang(f
n
) to Z for < we have
z Rang(

f
n
) z = g
n,
(z)z
n,
: <
where g
n,
(z) = 0 for every large enough
Let f
n,
= g
n,


f
n
Hom(G
k()
, Z). Those homomorphisms give, by renaming
the f
n,
s, an embedding of G
k()
into

Z. Looking at the construction, it is a pure
one.
2) By 5.7.
5.12
5.13 Claim. Assume MA +2

0
>
2
.
If k() > 2 and 2

0

k()+1
then G = G
k()
is a non-free Whitehead Borel
(abelian) group.
Recall
5.14 Denition. An abelian group G is called a Whitehead group when: if H is
an abelian group and g is a homomorphism from H onto G with kernal isomorphic
to Z, then the kernel is a direct summand of H (equivalently, g is invertible).
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 33
Proof. By 5.3 we know that G
k()
is a Borel group and by 5.8 it is not free. Let

: < 2

0
) list

2 with no repetitions and U

= : < .
So U

: < 2

0
) be -increasing continuous with union

2 such that U
0
=
, [U

[ [[; and let H

:= G
U

, see Denition 5.5(1). So H

: < 2

0
) is a
-increasing continuous sequence of subgroups of G with union G. For < 2

0
,
letting u

= recalling Denition 5.5 we have G


U

= G
U
+1
= H
+1
and
G
U

,u

= G
U

= H

, hence H
+1
/H

= G
U

/G
U

,u

and by 5.7(1) the latter


group is
2
-free so H
+1
/H

is
2
-free. As MA holds and [H
+1
/H

[ < 2

0
and
H
+1
/H

is
2
-free we know that it is a Whitehead group.
As H

is -increasing continuous, H
0
= 0 and each H
+1
/H

is a Whitehead
group, it follows that H

: < 2

0
is a Whitehead group, which means that G
is as required.
5.13
(
7
7
1
)


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







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


34 SAHARON SHELAH
REFERENCES.
[BeKe96] Howard Becker and Alexander S. Kechris. The descriptive set theory
of Polish group actions, volume 232 of London Mathematical Society
Lecture Notes Series. Cambridge University Press, Cambridge, 1996.
[EM2] Paul C. Eklof and Alan Mekler. Categoricity results for L
,
-free
algebras. Annals of Pure and Applied Logic, 37:8199.
[EM] Paul C. Eklof and Alan Mekler. Almost free modules: Set theoretic
methods, volume 46 of NorthHolland Mathematical Library. North
Holland Publishing Co., Amsterdam, 1990.
[EM02] Paul C. Eklof and Alan Mekler. Almost free modules: Set theoretic
methods, volume 65 of NorthHolland Mathematical Library. North
Holland Publishing Co., Amsterdam, 2002. Revised Edition.
[HaSh 323] Bradd Hart and Saharon Shelah. Categoricity over P for rst order T or
categoricity for L

can stop at
k
while holding for
0
, ,
k1
.
Israel Journal of Mathematics, 70:219235, 1990. math.LO/9201240.
[JShT 654] Winfried Just, Saharon Shelah, and Simon Thomas. The automor-
phism tower problem revisited. Advances in Mathematics, 148:243
265, 1999. math.LO/0003120.
[MkSh 366] Alan H. Mekler and Saharon Shelah. Almost free algebras . Israel
Journal of Mathematics, 89:237259, 1995. math.LO/9408213.
[Sh 849] Saharon Shelah. Beginning of stability theory for Polish Spaces. Israel
Journal of Mathematics, submitted.
[Sh 87a] Saharon Shelah. Classication theory for nonelementary classes, I. The
number of uncountable models of L

1
,
. Part A. Israel Journal of
Mathematics, 46:212240, 1983.
[Sh 87b] Saharon Shelah. Classication theory for nonelementary classes, I. The
number of uncountable models of L

1
,
. Part B. Israel Journal of
Mathematics, 46:241273, 1983.
[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 402] Saharon Shelah. Borel Whitehead groups. Mathematica Japonica,
50:121130, 1999. math.LO/9809198.
(
7
7
1
)


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







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


POLISH ALGEBRAS, SHY FROM FREEDOM SH771 35
[Sh 744] Saharon Shelah. A countable structure does not have a free uncount-
able automorphism group. Bulletin of the London Mathematical Soci-
ety, 35:17, 2003. math.LO/0010305.
[Sh:h] Saharon Shelah. Classication Theory for Abstract Elementary
Classes, volume 18 of Studies in Logic: Mathematical logic and foun-
dations. College Publications, 2009.
[ShVi 648] Saharon Shelah and Andres Villaveces. Categoricity may fail late.
Journal of Symbolic Logic, submitted. math.LO/0404258.
[So98] Slawomir Solecki. Analytic ideals and their applications. Annals of
Pure and Applied Logic, 99:5172, 1999.
[So99] Slawomir Solecki. Polish group topologies. In Sets and proofs (Leeds,
1997), volume 258 of London Math. Soc. Lecture Note Ser., pages 339
364, Cambridge, 1999. Cambridge Univ. Press.

Das könnte Ihnen auch gefallen