a

r
X
i
v
:
m
a
t
h
/
9
9
1
0
1
6
1
v
1


[
m
a
t
h
.
L
O
]


2
8

O
c
t

1
9
9
9
ALMOST FREE SPLITTERS
R udiger Gobel and Saharon Shelah
Abstract
Let R be a subring of the rationals. We want to investigate self splitting
R-modules G that is Ext
R
(G, G) = 0 holds. For simplicity we will call such
modules splitters, see [16]. Also other names like stones are used, see a dictio-
nary in Ringels paper [14]. Our investigation continues [10]. In [10] we answered
an open problem by constructing a large class of splitters. Classical splitters are
free modules and torsion-free, algebraically compact ones. In [10] we concen-
trated on splitters which are larger then the continuum and such that countable
submodules are not necessarily free. The opposite case of
1
-free splitters of
cardinality less or equal to
1
was singled out because of basically dierent tech-
niques. This is the target of the present paper. If the splitter is countable, then
it must be free over some subring of the rationals by Hausen [12]. Contrary to
the results in [10] and in accordance to [12] we can show that all
1
-free splitters
of cardinality
1
are free indeed.
1 Introduction
Throughout this paper R will denote a subring of the rationals Q and we will consider
R-modules in order to nd out when they are splitters. Splitters were introduced in
Schultz [16]. They also come up under dierent names as mentioned in the abstract.
Denition 1.1 An R-module G is a splitter if and only if Ext
R
(G, G) = 0 or equiva-
lently if Ext
Z
(G, G) = 0 which is the case if and only if any R-module sequence
0 G

X

G 0
This work is supported by the project No. G-0294-081.06/93 of the German-Israeli Foundation
for Scientic Research & Development
AMS subject classication:
primary 13C05, 18E40, 18G05, 20K20, 20K35, 20K40
secondary: 13D30, 18G25, 20K25, 20K30, 13C10
Key words and phrases: self-splitting modules, criteria for freeness of modules
GbSh 682 in Shelahs list of publications
1
splits.
A short exact sequence
0 B

C

A 0
represents 0 in Ext (A, B) if and only if there is a splitting map : A C such that
= id
A
. Here maps are acting on the right.
Recall an easy basic observation, see [5]:
If Ext (A, B) = 0, A

A and B

B, then Ext (A

, B/B

) = 0 as well.
The rst result showing freeness of splitters is much older then the notion of splitters
and is due to Hausen [12]. It says that any countable, torsion-free abelian group is a
splitter if and only if it is free over its nucleus. The nucleus is the largest subring R of
Q which makes the abelian group canonically into an R-module. More precisely
Denition 1.2 The nucleus R of a torsion-free abelian group G ,= 0 is the subring R
of Q generated by all
1
p
(p any prime) for which G is p-divisible, i.e. pG = G.
The xed ring R mentioned at the beginning will be the nucleus R = nuc G of the
associated abelian group G.
The following result reduces the study of splitters among abelian groups to those which
are torsion-free and reduced modules over their nuclei.
Theorem 1.3 ([16]) Let G be any abelian group and G = D C a decomposition
of G into the maximal divisible subgroup D and a reduced complement C. Then the
following conditions are equivalent.
(i) G is a splitter.
(ii)
_

_
(a) D is torsion (possibly 0), C is a torsion-free (reduced) splitter with
pC = C for all p primary components D
p
,= 0 of D.
(b) D is not torsion and C is cotorsion.
Many splitters are constructed in [10], in fact we are also able to prescribe their
endomorphism rings. This shows that uncountable splitters are not classiable in any
reasonable way, a result very much in contrast to classical well-known (uncountable)
splitters which are the torsion-free algebraically compact (or cotorsion) groups.
2
The classical splitters come up naturally among many others when considering
Salces work [15] on cotorsion theories: A cotorsion theory is a pair of classes of R-
modules (F, C) which are maximal, closed under extensions such that the torsion-free
class F is closed under subgroups and the cotorsion class C is closed under epimorphic
images and Ext (F, C) = 0 for all F F and C C. The elements in F C are
splitters and in case of Harrisons classical cotorsion theory these are the torsion-free,
algebraically compact groups. For the trivial cotorsion theory these are free R-modules.
Hausens [12] theorem mentioned above can be slightly extended without much
eort, see [10].
Theorem 1.4 If R = nuc G is the nucleus of the torsion-free group G and G is a
splitter of cardinality < 2

0
, then G is an
1
-free R-module.
Recall that G is an
1
-free R-module if any countably generated R-submodule is free.
The algebraic key tool of this paper can be found in Section 2. We consider torsion-free
R-modules M of nite rank which are minimal in rank and non-free. They are (by
denition) n-free-by-1 R-modules if rk M = n + 1; the name is self explaining: They
are pure extensions of a free R-module of rank n by an R-module of rank 1. Similar
to simply presented groups, n-free-by-1 groups are easy represented by free generators
and relations. Using these minimal R-modules we will show the following
Main Theorem 1.5 Any
1
-free splitter of cardinality
1
is free over its nucleus.
The proof will depend on the existence of particular chains of
1
-free R-module of
cardinality
1
which we use to divide
1
-free R-modules of cardinality
1
into three
types (I,II,III). This may be interesting independently and we would like to draw
attention to Section 3. In Sections 4-7 we use our knowledge about these chains to
show freeness of splitter. The proof is divided into two main cases depending on the
continuum hypothesis CH (Section 5) and its negation (Section 4). In the appendix
Section 8 we present a proof of the main result of Section 5 under the weaker set
theoretic assumption WCH 2

0
< 2

1
, a weak form of CH which will be interesting
(only) for splitters of cardinality >
1
. The results in Section 6 and 7 on splitters of
type II and III do not use the case distinction by additional axioms of set theory.
2 Solving Linear Equations
Let R be a subring of Q. Then R-modules of minimal nite rank which are not free
will lead to particular innite systems of linear equations. Consider the Baer-Specker
R-module R

of all R-valued functions f : R on , also denoted by f = (f

m
)
m
.
3
Lemma 2.1 Let p = (p
m
)
m
, k
i
= (k
im
)
m
R

(i < n) where each p

m
is not
a unit of R. Then we can nd a sequence s = (s
m
)
m
R

such that the following

system of equations ( s) has no solution x = (x
0
, . . . , x
n
) R
n+1
, y
m
R with
y
0
= x
n
, y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ s
m
(m ). ( s)
Proof: We will use Cantors argument which shows that there are more real numbers
than rationals. First we enumerate all elements in R
n+1
as
W = W = x
m
= (x
m
0
, . . . , x
m
n
) : m
and construct s R

inductively.
It is interesting to note that the set of bad elements
B = s R

: x R
n+1
, (y
m
)
m
R

solving ( s) R

is a submodule of R

but [B[ is uncountable in many cases. Hence enumerating B

would not help.
Suppose s
0
, . . . , s
m1
R are chosen and we must nd s
m
. We calculate y
0
, . . . , y
m
from s
0
, . . . , s
m1
and y
0
= x
m
n
, x
m
1
, . . . , x
m
n1
and equation ( s) up to m1. The values
are uniquely dened by torsion-freeness and in particular
z = y
m
+

i<n
x
m
i
k
im
R (2.1)
is uniquely dened. Recall that p
m
is not a unit and either p
m
does not divide z, then
we set s
m
= 0 or we can choose some s
m
R 0 and p
m
does not divide z + s
m
. In
any case
y p
m
= z + s
m
has no solution in R (2.2)
and s
m
is dened.
Suppose that ( s) has a solution x R
n+1
, then x = x
m
for some m by our enumer-
ation. We calculate y
m+1
from ( s) substituting x, hence
y
m+1
p
m
= y
m
+

i<m
x
m
i
k
im
+ s
m
= z + s
m
is solvable by (2.1), which contradicts (2.2).
4
If G

G is a pure R-submodule of some R-module G which is of nite rank, not

a free R-module such that all pure R-submodules of G

of smaller rank are free, then

we will say that G

is minimal non-free. Such minimal non-free modules are simply

presented in the sense that there are x
i
, y
m
G

(i < n, m ) such that

G

= B, y
m
R : m ) with B =

i<n
x
i
R (2.3)
and the only relations
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
(m ) (2.4)
and coecients p
m
, k
im
R. The submodule B is pure in G

. If G is not
1
-free, then
the existence of minimal non-free submodules is immediate by Pontryagins theorem.
Non-freeness of G

implies that the Baer type of

G

/B = T Q
is strictly greater than the type of R, see Fuchs [5, Vol 2, pp 107-112].
In more details we have that B is a pure submodule of G

, hence G

/B is torsion-
free of rank 1 and since G

is not a free R-module, G

/B = T cannot be isomorphic
to R. If : G

Q is the canonical homomorphism taking B to 0 and y

0
to 1 Q,
then Im Q represents the type of y
0
+ B)

= G

/B. There are p

m
N, not units
in R such that T =

m
q
1
m
Z Q and q
m
=

i<m
p
i
. In order to derive the crucial
equations as in the above denition we choose preimages y
m
G

of q
1
m
such that
y
0
= q
0
= 1 and y
m
= q
1
m
(m ).
Using q
m+1
= q
m
p
m
we nd elements k
im
R (i < n), g
m
G

such that (2.3) and

(2.4) holds. We will constantly use the representations (2.3) and (2.4) which are basic
for the following
Proposition 2.2 Let G

G
+1
be a countable free resolution of G

as in (2.3) and
let the relations (2.4) be expressed in G
+1
by
y

m+1
p
m
= y

m
+

i<n
x

i
k
im
+ g
m
for some g
m
G

; let z
m
(m ) be non-trivial elements of an
1
-free R-module H
0
of cardinality
1
and
0 H
0
H

h
G

0
5
be a short exact sequence. Then we can nd an R-module
H

= H

, y

m
: m )
with B

i<n
x

i
R, g
m
h = g
m
and the only relations in H

m+1
p
m
= y

m
+

i<n
x

i
k
im
+ z
m
+ g
m
. (m )
The map h extends to h

by x

i
h

= x

i
, y

m
h

= y

m
such that the new diagram with
vertical maps inclusions commutes:
0 H
0
H

h
G

0

0 H
0
H

G
+1
0
Proof: Let
F
+1
= H

i<n
x
i
R

m
y
m
R
and dene
N
+1
= ( y
m+1
p
m
y
m

i<n
x
i
k
im
z
m
g
m
)R : m ).
Hence H

= F
+1
/N
+1
and let
x

i
= x
i
+ N
+1
, y

m
= y
m
+ N
+1
and x

= x + N
+1
.
First we see that
(a)
_
x x

(x H

) denes an embedding H

and then we identify H

with its image in H

.
It remains to show that H

N
+1
= 0 viewed in F
+1
. If x H

N
+1
, then
there are k
m
R for m l and some l such that
l

m=0
( y
m+1
p
m
y
m

i<n
x
i
k
im
z
m
g
m
)k
m
= x H

.
6
We get x =
l

m=0
(z
m
+ g
m
)k
m
and
l

m=0
( y
m+1
p
m
y
m

i<n
x
i
k
im
)k
m
= 0. The coecient
of y
l+1
is p
l
k
l
= 0, hence k
l
= 0 and going down we get k
m
= 0 for all m l, hence
x = 0 and (a) holds. Due to N
+1
we have the useful system of equations in H

.
(b) y

m+1
p
m
= y

m
+

i<n
x

i
k
im
+ z
m
+ g
m
with g
m
H

.
In view of (a) we also have
(c) H

= H

, y

m
: m ) with B

i<n
x

i
R H

.
Next we claim that
(d)
_
If h

= h, x

i
h

= x
i
and y

m
h

= y
m
(i < n, m ), then
h

: H

G
+1
is a well-dened homomorphism with
(e) ker h

= H
0
and Imh

= G
+1
.
As h

is dened on non-free generators, we must check that the relations between

them are preserved, when passing to the proposed image. The relations are given by
N
+1
or equivalently by (b). Using the denition (d) we see that the relations (b) are
mapped summand-wise under h

as follows.
y

m+1
p
m
+ y

m
+

i<n
x

i
k
im
+ z
m
+ g
m

y
m+1
p
m
? y
m
+

i<n
x
i
k
im
+ 0 + g
m
and inspection of (2.4) and the relations in G
+1
shows that ? is an equality sign
=. Hence h

is well-dened. Notice that H

= H

h = G

, therefore h

induces a
homomorphism
H

/H

G
+1
/G

and the last argument and g

m
G

show that this is an isomorphism. Hence passing

from h to the extended map h

the kernel cannot grow, we have H

0
= ker h

= ker h
and Imh

= G
+1
is obvious, so (d) and (e) and the proposition are shown.
3 The Main Reduction Lemma - Types I, II and III
The Chase radical G of a torsion-free R-module G is the characteristic submodule
G =

U G, G/U is
1
-free .
7
Since G/G is also
1
-free, the Chase radical is the smallest submodule with
1
-free
quotient. If U is a submodule of G we write

U
G = G

for the Chase radical of G over U which is dened by (G/U) = G

/U.
Given any
1
-free R-module G of cardinality [G[ =
1
, we x an
1
-ltration
G =
_
<
1
G
0

which is an ascending, continuous chain of countable, free and pure R-submodule G

0

of G with G
0
0
= 0.
We want to nd a new ascending, continuous chain of pure R-submodules G

(not
necessarily countable) such that G =

<
1
G

. However we do require that

G/G

is
1
-free if is not a limit ordinal. (3.1)
We will use the new chain to divide
1
-free R-module of cardinality
1
into three types.
This distinction helps to show that
1
-free splitters of cardinality
1
are free.
Suppose G

G is constructed for all < . Next we want to dene G

. If is a
limit ordinal, then
G

=
_
<
G

.
Hence we may assume that = + 1 and we must dene G

= G
+1
. In order to
ensure G =

1
G

we let
G
0
= (G

+ G
0

G (3.2)
the pure R-submodule generated by G

+G
0

. In any case we want to ensure that (3.1)

holds, hence
G
0
G G

. Therefore we construct an ascending, continuous chain of

pure R-submodules
G
j
: j <
1
with G
j+1
/G
j
= 0 or minimal non-free for each 0 < j <
1
(3.3)
such that G

j
1
G
j
. Suppose that G
i
is dened for all i < j <
1
. If j is a limit
ordinal we take G
j
=

i<j
G
i
and if j = i + 1 we distinguish two cases:
If G/G
i
is
1
-free, then G
j
= G
i
hence G

= G
i
and [G

/G
1
[ =
0
. (3.4)
8
Otherwise G/G
i
is not
1
-free, and by Pontryagins theorem we can nd a nite
rank minimal non-free pure R-submodule M/G
i
of G/G
i
. Since G =

i
1
G
0
i
and
, = M G
i
G, there is also a least ordinal = (M) = (M/G
i
) <
1
, such that
(M G
i
) (G
0
+1
G
0

) ,= . (3.5)
Among the candidates M we choose one with the smallest (M) and take it for
M = G
i+1
. This completes the construction of the G
i

s. Notice that either the

construction of G

stops as in case (3.4) or we arrive at the second possibility:

G
i+1
/G
i
is minimal non-free for each i <
1
and [G

/G
1
[ =
1
. (3.6)
It remains to show that in case (3.6) the following holds.

G
0
G = G

or equivalently G/G

is
1
-free. (3.7)
Suppose that G/G

is not
1
-free and let X be a non-free submodule of minimal nite
rank in G/G

which exists by Pontryagins theorem. Representing X in G we have

G

= x
i
, y
m
, G

: i < n, m )

with G

/G

= X,
see also Gobel, Shelah [10]. There are elements g
m
G

(m ) such that
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ g
m
for some p
m
, k
im
R (p
m
not units of R). We take
G

= x
i
, y
m
, g
m
: i < n, m )

G,
hence X = G

+ G

/G

was our starting point. Since G

is obviously countable, there

is a

1
with G

G
0

. If G
j
: j <
1
is the chain constructed above, we
also nd i
1
with g
m
G
i
for all m . If i j
1
, then G

+ G
j
/G
j
is
an epimorphic image of X, hence minimal non-free or 0. The second case leads to the
immediate contradiction:
G

G
j
G

but X ,= 0.
Hence G

+ G
j
/G
j
,= 0 was a candidate for constructing G
j+1
for any i j
1
.
Has it been used? We must compare the -invariant (G

+G
j
/G
j
) with the various
(G
j+1
/G
j
). From G

G
0

we see that there is

j
<

such that
(G

+ G
j
G
j
) (G
0

j
+1
G
0

j ,= .
9
By minimality of
j
=: (G
j+1
/G
j
) we must have
j

j
<

and
(G
j+1
G
j
) (G
0

j
+1
G
0

j
) ,=
and (G
j
G
0

) (j
1
) is a strictly increasing chain of length
1
of the countable
module G
0

, which is impossible. Hence G/G

is
1
-free and (3.7) is shown. We have
a useful additional property of the constructed chain which reects (3.7).
Corollary 3.1 If 0 ,=
1
is not a limit ordinal, then G

=
G
0

G.
Proof: We concentrate on the case (3.6) and only note that the case (3.4) is similar.
Recall from (3.7), that G/G

is
1
-free, hence the claim of the corollary is equivalent
to say that any submodule X of G

must be G

if only G
0

X with G

/X
1
-free.
Let > 0 and suppose G
0

X G

and 0 ,= G

/X is
1
-free. First we claim
that
G

X for all < .

If this is not the case, then let < be minimal with G

, X. Recall that can not

be a limit ordinal and we can write = +1 for some < . We have G

j
1
G
j
,
hence
i

= min j
1
: G
j
, X
1
exists. If i

= 0, then G
0

G
0

X from > and G

0
, X. We get G
0
=
G

, G
0

, X and G
0

X requires G

, X, contradicting minimality of . Hence

i

> 0 and i

= j + 1. We have G
i

, X and G
j
X from j < i

and minimality
of i

. However G
j+1
/G
j
is minimal non-free, and 0 ,= G
j+1
+ X/X G/X is an
epimorphic image, hence non-free as well. Therefore G/X is not
1
-free, a contradiction
showing our rst claim.
From the rst claim we derive

<
G

X. Now there must be a minimal

i

= min j : G
j
, X
1
,
which cannot be a limit ordinal, and again i

> 0, hence i

= j + 1. We nd
G
j
X, G
j+1
, X and G

/X cannot be
1
-free, a nal contradiction.
We now distinguish cases for G depending on the existence of particular ltrations.
Let G =

<
1
G

be the ltration constructed from the

1
-ltration G =

<
1
G
0

.
If there is an ordinal <
1
(which we assume to be minimal) such that G = G

,
then let C = G
0

which is a countable, free and pure R-submodule of G. From Corollary

3.1 we see that
C
G = G. Hence, beginning with C we get a new
1
-ltration (we use
10
the same notation) G

:
1
of countable, pure and free R-submodules of G such
that G
0
= C and each G
+1
/G

( > 0) is minimal non-free. In this case we say that

G and the ltration are of type I.
In the opposite case the chain only terminates at the limit ordinal
1
i.e. G

,= G
for all <
1
. We have a proper ltration G =

<
1
G

such that Corollary 3.1

holds. If for each
1
for some i <
1
case (3.4) occurs, then the constructed
chain G

:
1
is an
1
-ltration of countable, pure and free R-submodules with
the properties of Corollary 3.1 and (3.1). We say that the chain and G are of type II
respectively.
If G is not of type I or of type II we say that G is of type III. In this case, there is
a rst
1
such that G
+1
/G

is uncountable. We may assume that = 0. With

the new enumeration we see that the following holds for type III:
(III)
_

_
G =

<
1
G

, G
0
= 0, [G
1
[ =
1
and (3.1) holds,
G
1
=

j
1
G
0j
is an
1
-ltration of pure submodules of G
1
with each G
0j+1
/G
0j
minimal non-free.
We have a
Reduction-Lemma 3.2 Any
1
-free module G of cardinality
1
is either of type I, II
or III.
4 Splitters Of Cardinality
1
< 2

0
Are Free
In this section we do not need the classication of
1
-free R-modules of cardinality

1
given in Lemma 3.2. Moreover, we note that
1
-freeness of splitter of cardinality

1
< 2

0
follows by Theorem 1.4. In fact we will present a uniform proof showing
freeness of splitters up to cardinality
1
< 2

0
which extends Hausens result [12]
concerning countable splitters. We begin with a trivial observation
Proposition 4.1 Let G =

1
G

be an
1
-ltration of pure and free R-submodules
G

of G. Then nuc (G

) = R for all
1
.
Proof: Choose any basic element b G

for some
1
. If r Q divides b in G,
then r divides b in G

by purity, hence r R from bR C = G

and nuc G = R.
11
Corollary 4.2 (
1
< 2

0
). If G is a splitter of cardinality
1
and nuc G = R,
then there is an
1
-ltration G =

1
G

of pure and free R-submodules G

such that
nuc G

= R for all
1
.
Proof: From
1
< 2

0
and Gobel and Shelah [10], see Theorem 1.4, follows that G
is an
1
-free R-module and G has an
1
-ltration as in the hypothesis of Proposition
4.1. Hence Corollary 4.2 follows by the Proposition 4.1.
Denition 4.3 Let G be a torsion-free abelian group with nuc G = R and X an R-
submodule of G. Then X is contra-Whitehead in G if the following holds.
There are z
m
G, p
m
, k
im
R (i < n, m ) such that the system of equations
Y
m+1
p
m
Y
m
+

i<n
X
i
k
im
+ z
m
mod X (m )
has no solutions y
m
, a
i
G (for Y
m
, X
i
respectively) with

i<n
(a
i
+ X)R free of rank n
and pure in G/X. Otherwise we call X pro-Whitehead in G.
For X G as in the denition let W be the set of all nite sequences a =
(a
0
, a
1
. . . , a
n
) such that
(i) a
i
G (i n)
(ii)

i<n
(a
i
+ X) R is pure in G/X.
(iii) (a
i
+ X)R : i n)

is not a free R-module in G/X.

In particular G

a
=

i<n
a
i
R X is a pure submodule of G
a
= X, a
i
R : i n)

and
of G, the module G
a
/X is an n-free-by-1 R-module. From (2.4) we nd p
am
N not
units in R and elements k
aim
R (i < n), g
am
G
a
such that
y
am+1
p
am
= y
am
+

i<n
a
i
k
aim
+ g
am
(m ) (4.1)
The equations (4.1) are the basic systems of equations which decide about G to be
a splitter. We will also consider an inhomogeneous counter part of (4.1) and choose
a sequence z = (z
m
: m ) of elements z
m
G. The z-inhomogeneous counter part
of (4.1) is the system of equations
Y
m+1
p
am
Y
m
+

i<n
X
i
k
aim
+ z
m
mod X (m ) (4.2)
12
According to the above denition we also say that a W is contra-Whitehead if (4.2)
has no solutions y
m
(m ) in G (hence in G
a
) for some z and X
i
= a
i
. Otherwise
we say that a is pro-Whitehead. If G =

1
G

is an
1
-ltration of G, then we dene
W

for X = G

and let S =
1
: there exists a W

contra-Whitehead .
Proposition 4.4 If G =

1
G

and S as above is stationary in

1
, then G is not a
splitter.
Before proving this proposition we simplify our notation. If S we choose
z

m
G, a = (a

0
, a

1
, . . . , a

n
), p
am
= p
m
, g
am
= g

m
, k
aim
= k
im
, y
am
= y

m
so that equations (4.1) and (4.2) become for X = G

m+1
p
m
= y

m
+

i<n
a

i
k
im
+ g

m
(m ) (4.3)
with z

-inhomogeneous counter part

Y
m+1
p
m
Y
m
+

i<n
X
i
k
im
+ z

m
mod G

(m ). (4.4)
Hence (4.3) is a system of equations with solutions y

m
, a

i
, g

m
in G
+1
, while (4.4)
with variables Y
m
, X
i
(m , i < n) has no solutions in G, as discussed in Denition
4.3 for X = G

. The set of limit ordinals is a cub, hence we may restrict S to this cub
and assume that S consists of limit ordinals only. If S we also may assume that
G
+1
= G

, a

i
R : i < n)

= G

, a

i
R, y

m
R : m
1
, i < n).
We begin the
Proof of the Proposition 4.4:
We will use the last remarks for constructing h : H G such that
() 0 H
0
H
h
G 0
does not split, hence Ext (G, H
0
) ,= 0. We will have H
0

= G, hence Ext (G, G) ,= 0
and G is not a splitter.
Choose an isomorphism : G H
0
which carries the
1
-ltration G

:
1
to
H
0
=

1
H

and z

m
to z

m
. Inductively we want to dene short exact sequences
() 0 H
0
id
H

0 ( < )
13
which are increasing continuously. Let
(0) 0 H
0
id
H
0
h
0
0 0
be dened for H
0
= H
0
with h
0
the zero-map and suppose () is dened for all
< <
1
with a limit ordinal. We take unions and () is dened. If
1
S,
we extend () trivially to get ( + 1) and if S we must work for ( + 1):
We apply Proposition 2.2 to nd H

H
+1
with
H
+1
= H

, e
m
, x
m
: m , i < n).
and relations
e
m+1
p
m
= e
m
+

i<n
x
i
k
im
+ y
m
+ z

m
(m ) (4.5)
with y
m
h

= g

m
G

. We want to extend the homomorphism h

: H

to
h
+1
: H
+1
G
+1
, and set e
m
h
+1
= y

m
and x
i
h
+1
= a

i
. By Proposition 2.2
the map h
+1
is a well dened homomorphism. It is clearly surjective with kernel H
0
.
Hence ( + 1) is well-dened for all
1
and h =

1
h

shows ().
Finally we must show that () does not split and suppose that : G H is a
splitting map for (), hence h = id
G
and H = H
0
Im and g

m
= y
m
h

= y
m
h, so
(y
m
g

m
) h = 0 implies y
m
g

m
H
0
for all S. The set
C =
1
: a limit ordinal y
m
g

e
H

- by a back-and-forth argument - is a cub and hence S C is stationary in

1
. We can
nd C S and consider the attached equations. In G holds (4.3)
y

m+1
p
m
= y

m
+

i<n
a

i
k
im
+ g

m
and moves these equations to H:
(y

m+1
) p
m
= (y

m
) +

i<n
(a

i
) k
im
+ (g

m
)
which we subtract from (4.5). Hence
(e

m+1
y

m+1
) p
m
= (e

m
y

m
) +

i<n
(x
i
a

i
) k
im
+ (y
m
g

m
) + z

m
14
Put
f
m
= e

m
g

m
, v
i
= x
i
a

i
, w
m
= y
m
g

m

and note that
f
m
h = e
m
h g

m
h = g

m
g

m
= 0
hence f
m
ker h = H
0
. Similarly w
m
, v
m
H
0
. The last equation turns into
f
m+1
p
m
= f
m
+

i<n
v
im
k
im
+ w
m
+ z

m
(m )
which, as just seen, is a system of equations in H
0
. From C we have w
m
H

.
The isomorphism
1
moves the last equation back into G and w
m

1
G

. Using
f

m
= f
m

1
, w

m
= w
m

1
, v

i
= v
i

1
we derive
f

m+1
p
m
= f

m
+

i<n
v

i
k
im
+ w

m
+ z
m

(m )
with w

m
G

and z
m

as in (4.4), which is impossible in the case S which is

contra-Whitehead, where we have chosen z
m

accordingly.
Theorem 4.5 Let G be a splitter of cardinality < 2

0
with nuc G = R. If X is a pure,
countable R-submodule of G which is pro-Whitehead in G, then G/X is an
1
-free
R-module.
Proof: First we assume that nuc (G/X) = R and suppose for contradiction that G/X
is not an
1
-free R-module. By Pontryagins theorem we can nd an R-submodule
Y G/X of nite rank which is not free. We may assume that Y is of minimal rank.
Hence
Y = B, y
m
R : m ), B =

i<n
x
i
R
with the only relations
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
(m )
as in Section 2 such that each p
m
R is not a unit of R for m . Choose x
i
G
such that x
i
+ X = x
i
for each i < n. We can also choose a sequence of elements
z
m
G such that z
m
+ X is not divisible by p
m1
from nuc(G/X) = R (m ).
15
If

2, then let
z

= (e)z
e
: e ) = (z

e
).
Recall that X is pro-Whitehead in G, hence the systems of equations
y

m+1
p
m
y

m
+

i<n
x
k
i
k
im
+ z

m
mod X (m ) ()
has solutions x

i
, y

m
G for each

2. Note that
[x

i
: i < n)

0
) :

2[ [G[ < 2

0
.
We can nd ,=

2 such that x

i
= x

i
for all i < n and y

0
= y

0
. From ,= we
nd a branching point j such that
(j) ,= (j) but j = j.
We may assume
(j) = 1 and (j) = 0
and put w
m
= y

m
y

m
. Subtracting the equations () from () we get from x

i
x

i
= 0
that
w
m+1
p
m
= w
m
+ (z

m
z

m
) mod X
and w
0
= y

0
y

0
= 0 as well. For m j we have z

m
z

m
= 0 and z

j
z

j
= z
j
, hence
w
m
= 0 for m < j by torsion-freeness and
w
j
p
j1
= z
j
mod X
which contradicts our choice of z
m
s and p
m
s.
If nuc (G/X) = Q, then G/X is divisible, hence X is dense and pure in G, we have
X

X, where

X is the Z-adic completion of X, and X is a free R-modules of
countable rank. Hence G/X

X/X

=

0
Q and there are 2

0
independent elements
in

X/X. Using these independent elements, we nd systems of equations expressing
them as solutions Z-adic limits, which must be solvable by pro-Whitehead. Hence
[G[ = [

X[ = 2

0
>
1
, which is a contradition.
So we nd p
m
R and z
m
G such that p
m1
does not devide z
m
+X in G for all
m . The above argument applies again for n = 0 and leads to a contradiction.
Corollary 4.6 Any splitter of cardinality at most
1
< 2

0
is free over its nucleus.
16
Proof: Let
G =
_

1
G

be an
1
-ltration of the splitter G. By Corollary 4.2 we may assume that each G

is
a pure and free R-submodule of G with R = nuc G. If S denotes the set

1
: G

is contra-Whitehead in G,
then S is not stationary in
1
by the last Proposition 4.4. We may assume that all G

are pro-Whitehead in G and each G

+1
/G

is countable, hence free by Theorem 4.5.

We see that G must be free as well.
5 Splitters Of Type I Under CH
In view of Section 4 we may assume CH to derive a theorem in ZFC showing freeness
for
1
-free splitters of cardinality
1
of type I. The advantage of the set theoretical
assumption is - compared with the proof based on the weak continuum hypothesis
WCH in Section 8 - that the proof given here by no means is technical. Recall that G
is of type I if G =

1
G

for some
1
-ltration G

:
1
of pure submodules G

such that each G

+1
/G

( > 0) is a minimal non-free R-module. In this section we

want to show the following
Proposition 5.1 (ZFC + CH) Modules of type I are not splitters.
Combining Proposition 5.1 and Corollary 4.6 we have can remove CH and have the
immediate consequence which holds in ZFC.
Corollary 5.2 Any
1
-free splitter of type I (and cardinality
1
) is free over its nu-
cleus.
The proof of Proposition 5.1 is based on an observation strongly related to type
I concerning splitting maps. Then we want to prove a step lemma for applications
of CH. Finally we use CH to show Ext (G, G) ,= 0 in Theorem 5.1. In Section 1 we
noticed that if
0 B

C

A 0
is a short exact sequence, hence representing an element in Ext (A, B), then this element
is 0 if and only if there is a splitting map : A C such that = id
A
. This simple
fact is the key for the next two results.
17
Observation 5.3 Let G =

1
G

be a ltration of type I. For

1
, let
0 H
0
H

0 ()
be a continuous, increasing chain of short exact sequences with union
0 H
0
H G 0 (
1
)
and let H
0
= G be
1
-free. Then any splitting map of (1) has at most one extension
to a splitting map of (
1
).
Proof: We may assume that the splitting map : G
1
H
1
of (1) has two extensions
,

: G H which split. Since

1
is a limit ordinal, there is some <
1
minimal
with (

) G

,= 0. Clearly is not a limit ordinal and

induces a non-trivial
map : G

/G
1
H

. The domain of this map is minimal non-free, while its range

is
1
-free, hence must be 0, a contradiction.
Step-Lemma 5.4 Let G =

1
G

be a ltration of type I and let

0 H
0
H

h
G

0
be a short exact sequence with H
0
= G. If : G

is a splitting map, then there

is an extension of this sequence such that does not extend to a splitting map

of the
new short exact sequence:
0 H
0
H

0

0 H
0
H

G
+1
0
Moreover, the vertical maps in the diagram are inclusions and if G
+1
/G

is n-free-
by-1, then B

+1
is a free R-module of rank n and
H

= H

+1
, y

m
: m )
and B

+1
is mapped under h

mod G

onto a free maximal pure R-submodule of

G
+1
/G

.
18
Proof of the Step-Lemma 5.4: We will use special elements s
m
R (m ) to
kill extensions. It will help the reader to pose precise conditions on the choice of the
s
m
s only when needed, which will be at the end of the proof. Readers familiar with
such proofs will know that we are working to produce a p-adic catastrophe.
First we use the fact that G

= G
+1
/G

is minimal non-free, say n-free-by-1. By

(2.3) and (2.4) we have
G

i<n
x
i
R, y
m
R, m )
with the only relations
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
(m )
and coecients
p
m
, k
im
R.
By the last equations we can nd g
m
G

and x
i
, y
m
G
+1
such that
()
_
_
_
G
+1
= G

, x
i
R, y
m
R : i < n, m )
with the relations y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ g
m
(m ).
The action of is known to us on G

, hence we can choose a pure element 0 ,= z H

0
and let z
m
= zs
m
, hence H
0
/zR is
1
-free by purity of z in an
1
-free R-module. We
also choose preimages g
m
= g
m
H

, hence g
m
h = g
m
. We are now in the
position to apply Proposition 2.2. Let
H

= H

+1
, y

m
: m ) with B

+1
=

i<n
x

i
R H

be the extension given by the proposition with the useful relations

y

m+1
p
m
= y

m
+

i<n
x

i
k
im
+ zs
m
+ g
m
(m ) (5.1)
and an extended homomorphism h

: H

G
+1
with
h

= h, x

i
h

= x
i
and y

m
h

= y
m
(i < n, m )
such that
ker h

= H
0
and Imh

= G
+1
.
It remains to show the non-splitting property of the Lemma.
19
Suppose that

: G
+1
H

is an extensions of : G

H such that

= id
G
+1
.
Now we want to derive a contradiction when choosing the s
m
s accordingly (inde-
pendent of

!) We apply

to () and get the equations in H

:
(

) y
m+1

p
m
= y
m

i<n
x
i

k
im
+ g
m

.
If d
m
= y

m
y
m

, and e
i
= x

i
x
i

then d
m
H
0
from
d
m
h

= (y

m
y
m

)h

= y

m
h

y
m

= y
m
y
m
= 0 and ker h

= H
0
.
Similarly we argue with e
i
and get
d
m
, e
i
H
0
.
Subtracting (

) from (5.1) leads now to a system of equations in H

0
.
d
m+1
p
m
= d
m
+

i<n
e
i
k
im
+ zs
m
.
We consider the submodule
W = d
m
+ zR, e
i
+ zR : i < n, m )
R
H
0
/zR.
The last displayed equations tell us that W is an epimorphic image of a minimal non-
free R-module, hence 0 or non-free of nite rank. On the other hand H
0
/zR is
1
-free
as noted above, hence W = 0 or equivalently
d
m
, e
i
: m , i < n)
R
zR

= R
The original equations
d
m+1
p
m
= d
m
+

i<n
e
i
k
im
+ s
m
(5.2)
still hold, but this time require solutions d
m
, e
i
R. We get to an end: just choose
rational numbers s
m
R such that (5.2) has no solutions. The existence of these s
m
s
follows from Lemma 2.1. Finally note that dealing with (5.2) is independent of the
20
particular choices of the extensions of as required in the Lemma.
Proof of Proposition 5.1: Let
H
0

= G =
_

1
G

be the module of type I. We must show that Ext (G, H

0
) ,= 0 and need a non-splitting
short exact sequence
0 H
0
H
h
G 0 (5.3)
which we construct inductively as an ascending, continuous chain of short exact se-
quences
0 H
0
H

h
G

0
with union (5.3). Let
0 H
0
H
1
h
1
G
1
0
be the rst step with G
1
a free R-module of countable rank. By Observation 5.3 and
CH we can enumerate all possible splitting maps : G H of extensions h as in
(5.3) of all h
1
s by
1
, and let

: G H,
1
be such a list. Using the
Step-Lemma 5.4 and the uniqueness in Observation 5.3 we can discard any

at stage
when constructing
0 H
0
H
+1
h
+1
G
+1
0.
The resulting extension (5.3) can not split.
6 Splitters Of Type II
An R-module G is of type II if G has an
1
-ltration G =

1
G

of pure submodules
G

such that G/G

is
1
-free for all non-limit ordinals
1
, see Section 3. In this
section we want to show our second main
Theorem 6.1 If G is of type II, then G is a splitter if and only if G is free over its
nucleus R.
21
Remark Theorem 6.1 includes that strongly
1
-free R-modules are never splitters,
except if trivially the module is free. This was very surprising to us.
Proof: If S =
1
: G/G

is not
1
-free, then S is a set of limit ordinals by
(3.1), and if S we also may assume that G
+1
/G

is minimal non-free, compare

3.
We get a - invariant (G) dened by S modulo the ideal of thin sets, see e.g. [4]. If
(G) = 0, then we nd a cub C
1
, with CS = and G =

C
G

. Let
0
= min C.
Then G = G

0
F for some free R-module F, and G

0
is a countable submodule of G
which must be free over R by Hausens [12] result, see also [10]. Hence G is free. Note
that the hypothesis of G being
1
-free is not used in this case! If (G) ,= 0 we want to
show that Ext (G, G) ,= 0. Theorem 6.1 can be rephrased as
If G is of type II, then G is a splitter if and only if (G) = 0. (6.1)
Now assume that S is stationary in
1
. We want to construct some H
h
G 0
with kernel ker h = H
0
, G isomorphic to H
0
by , which does not split. If G

=
H

(
1
), then
H
0
=
_

1
H

is a (canonical)
1
-ltration of H
0
copied from G. First we pick elements z

H
0
such that z

R

= R and H
0
/z

R is
1
-free, e.g. take any basis element from a layer
H

+2
H

+1
of the ltration of H
0
. Then we dene inductively a continuous chain of
short exact sequences (
1
).
()
_
0 H
0
H

0
countable, free submodules H

, and ordinals <

1
subject to various conditions. At the end we want in particular H =

1
H

1
H

.
If = 0, then G
0
= 0 and we take the zero map h
0
: H
0
G
0
0 with kernel
H
0
.
Suppose () is constructed for all < . If is a limit, we take unions h

<
h

,
H

<
H

and H

<
H

, assuming that at inductive steps sequences extend

(naturally) by inclusions. Then visibly () holds.
We may assume that () is known, and we want to construct ( + 1).
If , S, then we extend () trivially:
22
Put H
+1
= H

with F

a free R-module of the same rank as the free R-module

G
+1
/G

. As G
+1
= F

, we may choose an isomorphism h

: F

and extend h

to h
+1
by h
+1
= h

. Clearly ker h
+1
= ker h

= H
0
and
Imh
+1
= G
+1
.
If S, then we must work. We have G
+1
/G

= B

+1
, y

m
: m ) from (2.3)
and (2.4). Hence
G
+1
= G

, B
+1
, y
m
R : m ), B
+1
=

i<n
x
i
R (6.2)
with relations
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ g
m
(m ), (6.3)
where g
m
G

. Let

B
+1
=

i<n
x
i
R be a copy of B
+1
. Then we pose the following
additional conditions on ( + 1).
(a) H
+1
h
+1
= H
+1
h
+1
= G
+1
(b)

B
+1
H
+1
(c) H
+1
/H

= G
+1
/G

(d) H
+1
/H

+ H

is free for all , , S and

1
(e) H
+1
/H

+ H

is
1
-free for all ,
1
(f) H

+1
H
+1
H
0
= H

(+1)

.
We choose preimages g
m
H

such that g
m
h

= g
m
and apply Proposition 2.2 to
dene the extension
H

H
+1
= H

B
+1
, y
m
R : m )
with the relations
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ z
m
s
m
+ g
m
(6.4)
where

B
+1

=

i<n
x

i
R
23
as required in (b). Similarly, by Proposition 2.2 the map h

extends to an epimorphism
h
+1
: H
+1
G
+1
. It is now easy to check that (c) holds and it is also easy to see
that ker h
+1
= H
0
. Next we extend H

H
+1
carefully such that ( + 1), (a), (b),
(d), (e) and (f) hold.
Imh
+1
= G
+1
is a countable module of the
1
- free R-module G, hence free and h
+1
must split. There is a splitting map
: G
+1
H
+1
such that h
+1
= id
G
+1
,
hence H
+1
= H
0
G
+1
. Let

0
: H
+1
H
0
,
1
: H
+1
G
+1

be the canonical projections with

0
+
1
= id
H
+1. Recall that

B
+1
H
+1
. Choose

1
large enough such that , S,

, + 1 and (

B
+1
+ H

)
0
H

. This
is easy because
1
S is unbounded and (

B
+1
+ H

)
0
is countable. Put H
+1
=
H

(G
+1
) and = ( + 1)

. Note that
G
+1
= H
+1
h
+1
H
+1
h
+1
G
+1
h
+1
= G
+1
and (a) follows.
If (+1)

, then H
+1
+H

= H
+1
and if (+1)

, then H
+1
+H

= H

G
+1

and (d) follows. We see immediately H

1
G
+1
and H

0
H

(+1)

, hence
H

0
+ H

1
H
+1
and (+1) holds; similarly

B
+1
H
+1
for (c). From H

(+1)

H
0
and the modular
law we have H
+1
H
0
= H

(+1)

(G
+1
H
0
) = H

(+1)

and (f) holds.

Finally we choose H =

1
H

, h =

1
h

and
0 H
0
H
h
G 0 (6.5)
is established and it remains to show that (6.5) does not split. Suppose for contradiction
that : G H is a splitting map for h. We have H =

1
H

and G =

1
G

1
-ltration. Using the above properties of the H

s, it follows by a back and forth

argument that
E =
1
: H

H
0
= H

, G

24
is a cub. On the other hand S is stationary in
1
and we nd
S E.
From (6.3) and (6.4) we have
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ g
m

and
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ z

s
m
+ g
m
with g
m
, g
m
H

and G

.
Put d
m
= y
m
y
m
, f
m
= g
m
g
m
, e
i
= x
i
x
i
and notice that
d
m
h = e
i
h = f
m
h = 0, hence d
m
, e
i
, f
m
H
0
.
Subtracting the last displayed equations we get
(j) d
m+1
p
m
= d
m
+

i<n
e
i
k
im
+ f
m
+ z

s
m
in H
0
.
Recall that f
m
H

H
0
H

by (f) and mod T = H

+ z

R the equations
(j) say that W = d
m
, e
i
: i < n, m ) + T/T is either minimal non-free or 0. On
the other hand H
0
/T is
1
-free, hence W = 0 and d
m
, e
i
, z

) H

+ z

R. Recall
from (e) that H
0
/H

is
1
-free. Hence (j) turns into
d
m+1
p
m
d
m
+

i<n
e
i
k
im
+ z

s
m
mod H

.
Using
1
-freeness of H
0
/z

R these equations tell us that we must have solutions

d
m
, e
i
R for
(k) d
m+1
p
m
= d
m
+

i<n
e
i
k
im
+ s
m
.
In Lemma 2.1 we selected particular s
m
s in R such that (k) has no solution in R. Now
we are ready to make this choice which we should have done right at the beginning of
the proof and hence derive a contradiction; we conclude Ext (G, G) ,= 0.
25
From Theorem 6.1 we see that non-free but strongly
1
-free abelian groups are never
splitters. We nd this very surprising. Particular groups like the Grith-group G
below which is a Whitehead group (Ext (G, Z) = 0) under Martins axiom and CH is
not a splitter. Recall a nice and easy construction of G which is sometimes Whitehead
but always fails to be a splitter in general.
Let P = Z

1
=

1
Z the cartesian product of Z. If
1
is a limit ordinal choose
an order preserving map

: with sup(

) = . Then, along this ladder system

we dene branch-elements
c
n
=

in
(i

)
i!
n!
which are a divisibility chain of c

0
modulo

1
Z, hence
G =

1
Z, c
n
:
1
, a limit ordinal, n )
is a pure subgroup of P. We see that [G[ =
1
and G is
1
- free by
1
-freeness of P; see
[5] (Vol 1, p 94, Theorem 19.2). Moreover
G
,= 0 because G

= G

<
Z (
1
)
is an
1
-ltration of G with G
+1
/G

divisible for all limit ordinals . Hence G is not

free. It is easy to check that G is
1
-separable, hence strongly
1
-free; see also [4], p.
183, Theorem 1.3.
7 Splitters Of Type III
If G is of type III then we recall from Section 3 that G =

1
G

, G
0
= 0 with (3.4)
(3.6) and G
1
=

j
1
G
0j
and G
0j
: j
1
is an
1
-ltration of pure submodules G
0j
such that each G
0j+1
/G
0j
is minimal non-free. Here we will show:
Theorem 7.1 Modules of type III are not splitters.
Proof: Let G

1
G

be an isomorphic copy of G taking G

to G

, and choose a
sequence of elements z

+2
(
1
) such that
G

+1
z

R = 0 and G

+2
/G

+1
z

R is
1
-free.
This is possible by (III).
26
By a basic observation from Section 1 it is enough to show that Ext (G
1
, G) ,= 0.
Inductively we will construct a non-trivial element in Ext (G
1
, G). We consider the
following diagram
(0) 0 G

H
0
h
0
G
00
= 0 0
|
() 0 G

G
0
0
|
(
1
) 0 G

H
h
G
1
0
The rst row is the trivial extension with G

= H
0
and h
0
= 0. Vertical maps and
maps between G

and H

s are inclusions. The sequences () are increasing continuous

and suppose () is constructed for all < . Then h

<
h

and
0 G

_
<
H

h
G
0
0
if is a limit. Next we want to construct ( + 1) from () and recall that G

=
G
0+1
/G
0
is minimal non-free generated as in (2.3), (2.4). We can write
G
0+1
= G
0
, B
+1
, y
m
R : m ), B
+1
=

i<n
x
i
R (7.1)
with relations
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ g
m
(m ), g
m
G
0
. (7.2)
Then we dene
h
+1
: H
+1
G
0+1
0.
by Proposition 2.2. Hence
H
+1
= H

B
+1
, y
m
R : m )
has the relations
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ z

s
m
+ g
n
(m ), (7.3)
27
where the s
m
R will be specied later on, and g
m
H

.
Suppose that (
1
) splits and consequently : G
1
H is a splitting map for h.
Then let
d
m
= y
m
y
m
, e
i
= x
i
x
i
and f
m
= g
m
g
m
.
From splitting we get again
d
m
, e
i
, f
m
G

(
1
, m , i < n).
Using G

1
G

for
1
we nd
1
such that
d
m
, e
i
, f
m
G

for all i < n, m .

Consider a map :
1

1
taking any
1
to
() = min
1
; a limit ordinal, d
m
, e
i
, f
m
G

, , m , i < n
and note that C =
1
: () = is a cub in
1
and a subset of
E = : a limit ordinal, d
m
, e
i
, f
m
G

for all < , i < n, m .

Hence E is a cub in
1
. Next we apply to (7.2) and subtract (7.3). Hence we get a
system of equations in H
+1
.
d
m+1
p
m
= d
m
+

i<n
e
i
k
im
+ f
m
+ z

s
m
(m ) (7.4)
If E, then modulo H

the equations (7.4) turn into

d
m+1
p
m
d
m
+

i<n
e
i
k
im
+ z

s
m
(m )
and modulo z

R an earlier argument and

1
-freeness of G

/G

R show that the

last equation requires solutions d
m
, e
i
R for
d
m+1
p
m
= d
m
+

i<n
e
i
k
im
+ s
m
(m )
By a special choice of s
m
s in Lemma 2.1 this is now excluded, a contradiction. Hence
(
1
) has no splitting map and Theorem 7.1 follows.
28
8 Appendix: Splitters Of Type I Under 2

0
< 2

1
In Section 5 we have seen a proof that CH implies modules of type I are never splitters.
A slight variation but some what technical modication of the proof, shows that this
result can be extended to WCH that is 2

0
< 2

1
. Due to Section 5 this is not needed
for the main result of this paper dealing with modules of cardinality
1
but it will be
interesting when passing to cardinals >
1
. We outline the main steps, their proofs are
suggested by the proofs in Section 5.
Theorem 8.1 ( ZFC + 2

0
< 2

1
) Modules of type I are not splitters.
Step-Lemma 8.2 Let G =

1
G

be a ltration of type I and let

0 K H

h
G

0
be a short exact sequence with some z K such that zR

= R and K, K/zR are
1
-free.
Then there are two commuting diagrams ( = 0, 1)
0 K H

h
G

0

0 K

h
G
+1
0
with vertical maps inclusions such that any third row with H

1
-free,
0 K

0
and any splitting map of h cannot have two splitting extensions

of h

:
0 K H

0

0 K

h
G
+1
0

0 K

0
Moreover H

= H

+1
, y

m
: m ) and B

+1
is mapped under h

mod G

onto a free maximal R-submodule of G

+1
/G

cf. (2.3).
29
Denition 8.3 If such an extension

as in (8.2) exists for some 0, 1 we say

that splits over (H

, h

).
Proof of Lemma 8.2: Compare the proof of the Step-Lemma 5.4 but note that at
the end you must take once more dierences of the elements d
m
, e
i
for = 1 and
= 0 respectively. Then we are able to apply Lemma 2.1 to get a contradiction from
splitting.
We then apply the Step-Lemma and weak diamond

1
to construct a short exact
sequence
0 H
0
H
h
G 0.
Let : G H
0
be a xed isomorphism. Later we will use consequences of

1
to show that h does not split.
Proof of Theorem 8.1: If H

= G

, then H
0
=

1
H

is an
1
-ltration if
G =

1
G

is the given ltration of type I.

Let T =

1
>
2 be the tree of all branches : 2 for some
1
. We call = l()
the length of . Branches are ordered as usually, hence <

if

Dom = . The
empty set is the bottom element of the tree. If T, then we construct triples
(H

, H

, h

)
of R-modules H

with H

free of countable rank and a homomorphism

h

: H

G
subject to various natural conditions.
(i)
_
H

= H
0
, H

= 0 and h

= 0, hence
0 H
0
H

G
0
= 0 0 is short exact.
(ii)
_
If <

, then (H

, H

, h

) (H

, H

, h

), i.e.
H

, H

and h

.
(iii) H
0
= ker h

and Imh

= G
l()
= H

.
30
(iv)
_
_
_
If
1
is a limit and

2, then we take unions
(H

, H

, h

) =

<
(H

, H

, h

) = (

<
H

<
H

<
h

).
If l() = we put further restrictions on those triples. In this case G

= G
+1
/G

is minimal non-free, and G

can be represented by (2.3), (2.4). There are elements

g
m
G

, x
i
, y
m
G
+1
with
G
+1
= G

, B
+1
, y
m
R : m ), B
+1
=

i<n
x
i
R
and relations
y
m+1
p
m
= y
m
+

i<n
x
i
k
im
+ g
m
.
We choose an isomorphic copy B
+1
=

i<n
x

R of B
+1
and now continue dening the
tree with triples.
If 0, 1, then we require more from (H
<>
, H
<>
, h
<>
).
(S i) H

B
+1
H
<>
(S ii) H

+ H

H
<>
for all

2, and
1
.
(S iii) H

H
0
H

for some

[,
1
)
Note that H

H
0
= ker (h

).
(S iv) The crucial condition:
Suppose : G

is a homomorphism extending to

: G
+1
H
<>
, then
not both of them can be splitting maps over (H
<>
, h
<>
) for = 0, 1.
Before we begin with the inductive construction, we observe from (iii) that G
+1
/G

=
H
<>
/H

= H
<>
/H

for Dom = .
If

1
2 and H = H() =

<
1
H

, then H

<
1
H

H from (ii). (S iii)

ensures H
0
H

and from (iii) we get H/H

0
= H

/H
0
, hence H

= H. This will show

that
(1) H() =
_
<
1
H

=
_
<
1
H

(

1
2)
31
Similarly h() =

<
1
h

is a well-dened homomorphism h() : H() G by (ii),

it is onto with kernel H
0
by (iii), hence
(2) 0 H
0
H()
h()
G 0 (

1
2).
Condition (1) provides an
1
- ltration used to apply weak diamond for showing
that (2) does not split for some .
Next we will show that the tree with triples exists. This will follow by induction
along the length branches

2. The case = 0 is (i) and already established.
Suppose the construction is completed for all < and <
1
is a limit ordinal. For


2 we dene (H

, H

, h

) as in (iv) and it is easy to verify that all conditions hold,

notably (S iii), because we take only countable unions. We come to the inductive step
constructing (H
<>
, H
<>
, h
<>
) from (H

, H

, h

) for = Dom.
First we adopt the Step-Lemma for K = H

H
0
= ker h

, H

= H

, H

= H
<>
, h =
h

, h

= h
<>
and note that the needed element z exists because H

H
0
H

is
free. We must still dene H
<>
H

carefully satisfying (ii), (S i) - (S iii) and the

last equality in (iii): Write again h

for h
<>
and H

for H
<>
. We know that
Imh

= G
+1
is a countable submodule of the
1
-free module G, hence free and h

must split. There is a splitting map

: G
+1
H

such that

= id
G+1
, hence
H

= H
0
(G
+1

)
from the rst part of (iii). Let

0
: H

H
0
and

1
: H

H
0
be the canonical
projections, hence

0
+

1
= id
H
.
Choose = ( + 1)

<
1
large enough such that B
+1

0
H

0
H

, where B
+1
is taken from the Step-Lemma. We can choose because B
+1
and H

are countable.
Put
H
<>
= H

(G
+1

),
hence by the known half of (iii)
G
+1
= H

H
<>
h

G
+1

= G
+1
and the other half of (iii) follows.
If , then H
<>
+H

= H
<>
and if , then H
<>
+H

= H

(G
+1

)
with quotient H
<>
/H

(G
+1

)

= H
0
/H

which shows (S ii). Trivially H

1

(G
+1

) and H

0
H

by the choice of , hence

H

= H

id
H
H

0
+ H

1
H
<>
32
and (ii) holds. Similarly B
+1
H
<>
and (S i) is shown. From H

H
0
and the
modular law we have
H
<>
H
0
= H

(G
+1

H
0
) = H

and (S iii) holds.

The construction of the tree with triples is complete. We are ready to use the weak
diamond

1
(S) to show that G is not a splitter.
We will use

1
(S) as stated in Eklof, Mekler [4, p. 143, Lemma 1.7] and note that
G =

<
1
G

, H() =

<
1
H

are
1
-ltrations.

1
(S) must tell us which

1
2 we
should pick. We dene a partition P so that for S a homomorphism : G

(

2) has value P

() = 0 if and only if does not split over (H

0
, h
0
). By the
Step-Lemma build into the construction, we observe that
(a) if P() = 1, then cannot split over (H
1
, h
1
)
The prediction principle nds us a branch

1
2 with the -property
(b)
_
If : G H is any map, then
S

= S : P

( G

) = () is stationary in
1
.
We pick that branch and build H = H() and h = h() accordingly, hence
0 H
0
H
h
G 0 is short exact. After the branch is xed we let
H

= H

, h H

= h

and H

= H

. Now we claim that the last sequence does

not split. Suppose to the contrary that : G H is a splitting map, hence h = id
G
.
Notice that the set
C = <
1
, H

H
0
= H

, G

is a cub. Since S

1
is stationary, we nd an S

C and also let G

= ,
hence
(c)
_
: G

, P

() = () and
is a splitting map of h

: H

.
We also nd some < C. The diculty is that G
+1
H
+1
does not follow, as
in the case S is not costationary. Hence we need the stronger Step-Lemma (as usual).
33
If () = P() = 0, then (H
0
, h
0
) is part of the construction of
0 H
0
H
+1
G
+1
0
and does not split over (H
0
, h
0
), but is a global splitting map, hence splits at
over (H
0
, h
0
), a contradiction.
Necessarily () = P() = 1 and by (a) does not split over (H
1
, h
1
), but this
time (H
1
, h
1
) was used in the construction of H
h
G and a contradiction follows.
This shows that is no splitting map, and G is not a splitter.
References
[1] T. Becker, L. Fuchs, S. Shelah, Whitehead modules over domains, Forum
Mathematicum 1 (1989), 5368.
[2] A.L.S. Corner, R. Gobel, Prescribing endomorphism algebras - A unied treat-
ment, Proceed. London Math. Soc. (3) 50 (1985), 471 483.
[3] P. Eklof Set Theoretic Methods in Homological Algebra and Abelian Groups, Les
Presses de l Universit`e de Montreal 1980.
[4] P. Eklof, A. Mekler Almost free modules, Set-theoretic methods, North-
Holland, Amsterdam 1990.
[5] L. Fuchs, Innite abelian groups - Volume 1, 2 Academic Press, New York (1970,
1973).
[6] P. A. Grith, A solution of the splitting mixed problem of Baer, Trans. Amer-
ican Math. Soc. 139 (1969), 261269.
[7] R. Gobel, R. Prelle, Solution of two problems on cotorsion abelian groups,
Archiv der Math. 31 (1978), 423431.
[8] R. Gobel, Cotorsion-free abelian groups with only small cotorsion images, Aus-
tral. Math. Soc. (Ser. A) 50 (1991), 243247.
[9] R. Gobel, New aspects for two classical theorems on torsion splitting, Comm.
Algebra 15 (1987), 24732495.
34
[10] R. Gobel, S. Shelah, Cotorsion theories and splitters, to appear in Trans. Amer.
Math. Soc. (1999)
[11] R. Gobel, J. Trlifaj, Cotilting and a hierarchy of almost cotorsion groups, to
appear in Journal of Algebra (1999)
[12] J. Hausen, Automorphismen gesattigte Klassen abzahlbaren abelscher Gruppen,
Studies on Abelian Groups, Springer, Berlin (1968), 147181.
[13] C. M. Ringel, The braid group action on the set of exceptional sequences of a
hereditary artin algebra, pp. 339 352 in Abelian group theory and related topics,
Contemporary Math. 171 American Math. Soc., Providence, R.I. 1994.
[14] C. M. Ringel, Bricks in hereditary length categories, Resultate der Mathematik
6 (1983), 64 70.
[15] L. Salce, Cotorsion theories for abelian groups, Symposia Math. 23 (1979), 1132.
[16] P. Schultz, Self-splitting groups, Preprint series of the University of Western
Australia at Perth (1980)
[17] S. Shelah Innite abelian groups, Whitehead problem and some constructions,
Israel Journal Math. 18 (1974), 243 256.
[18] S. Shelah On uncountable abelian groups, Israel Journal Math. 32 (1979), 311
330.
[19] S. Shelah, A combinatorial theorem and endomorphism rings of abelian groups
II, pp. 37 86, in Abelian groups and modules, CISM Courses and Lectures, 287,
Springer, Wien 1984.
R udiger Gobel
Fachbereich 6, Mathematik und Informatik
Universitat Essen, 45117 Essen, Germany
email: R.Goebel@Uni-Essen.De
and
Saharon Shelah
Department of Mathematics
Hebrew University, Jerusalem, Israel
and Rutgers University, Newbrunswick, NJ, U.S.A
e-mail: Shelah@math.huji.ae.il
35