Hereditarily Separable Groups and Monochromatic Uniformization

P. C. Eklof University of California, Irvine A. H. Mekler Simon Fraser University

S. Shelah Hebrew University and Rutgers University

Abstract We give a combinatorial equivalent to the existence of a non-free hereditarily separable group of cardinality 1 . This can be used, together with a known combinatorial equivalent of the existence of a non-free Whitehead group, to prove that it is consistent that every Whitehead group is free but not every hereditarily separable group is free. We also show that the fact that Z is a p.i.d. with innitely many primes is essential for this result.

modified:2000-10-30

Introduction
An abelian group G is said to be separable if every nite rank pure subgroup is a free direct summand of G; G is hereditarily separable if every subgroup of G is separable. It is well-known that a Whitehead group is hereditarily separable. In fact, we have the following implications: free W-group hereditarily separable

revision:2000-10-27

The authors thank Rutgers University for its support. Research partially supported by NSERC grant #9848 Research partially supported by the BSF. Publication #442

442

2 Whiteheads Problem asks if the rst arrow is reversible. For each of the two arrows, it has been proved by the third author that it is independent of ordinary set theory whether the arrow reverses. (See [7], [9] and [10], or the account in [3].) Now if we consider the two arrows together, there are four possible cases, three of which have already been shown to be consistent: 1. Both arrows reverse. That is, every hereditarily separable group is free. This is true in a model of V = L. (See [3, VII.4.9].) 2. Neither arrow reverses. That is, there are Whitehead groups which are not free, and hereditarily separable groups which are not Whitehead. This is true in a model of MA + CH. (See [3, VII.4.5, VII.4.6 and XII.1.11].) 3. The second arrow reverses but not the rst. That is, every hereditarily separable group is Whitehead and there are Whitehead groups which are not free. This is true in a model of Ax(S) + (1 \ S) plus (E) for every regular > 1 and every stationary subset E of . (See [3, Exer. XII.16].)
modified:2000-10-30

4. The rst arrow reverses but not the second. That, is every Whitehead group is free, but there are non-free hereditarily separable groups. It is an application of the main theorem of this paper that this case is consistent. (See section 3.) We also give additional information about the circumstances under which Cases 2 and 3 can occur. (See section 4.) Finally, we show that Case 4 is impossible for modules over a p.i.d. with only nitely many (but at least two) primes. (See section 5.) Our methods involve the use of notions of uniformization, which have played an important role in this subject since [8]. (See, for example, [3] and the recent [4].) It has been proved that there exists a non-free Whitehead group of cardinality 1 if and only if there is a ladder system on a stationary subset of 1 which satises 2-uniformization. (Denitions are given in detail in the next section.) Our main theorem here is the following:

442

revision:2000-10-27

3 Theorem 1 A necessary and sucient condition for the existence of a nonfree hereditarily separable group of cardinality 1 is the existence of a ladder system on a stationary subset of 1 which satises monochromatic uniformization for colours. A ladder system = { : S} on S is said to satisfy monochromatic uniformization for colours if for every function c: S , there is a function f : 1 such that for every S, f ( (n)) = c() for all but nitely many n . We believe the main theorem is of independent interest aside from its use in Case 4. We will prove suciency in section 1 and necessity in section 2. We will then derive the consistency of Case 4 by standard forcing techniques like those used in [10]. (Actually, we need only the suciency part of the main theorem for this.) A knowledge of forcing is required only for sections 3 and 4. We would like to thank Bill Wickless for his help in answering a question about nite rank torsion-free groups.

Preliminaries
We will always be dealing with abelian groups or Z-modules; we shall simply say group. A group G is said to be 1 -free if every countable subgroup of G is free, or equivalently, every nite rank subgroup is free. (See [3, IV.2.3]; throughout the paper we will usually cite [3] for results we need, rather than the original source.) An 1 -free group G is separable if and only if every pure subgroup H of nite rank is a direct summand of G, i.e., there is a projection h: G H (a homomorphism such that h H is the identity on H). The following are two useful facts (cf. [3, IV.2.7 and VII.4.2]): Lemma 2 (i) An 1 -free group G is separable if every pure cyclic subgroup of G is a direct summand of G. (ii) An 1 -free group G is hereditarily separable if B is separable whenever B is a subgroup of G such that G/B is isomorphic to a subgroup of Q/Z and there is a nite set P of primes such that the order of every element of G/B is divisible only by primes in P .
modified:2000-10-30

442

revision:2000-10-27

4 A group G is said to be a Whitehead group if Ext(G, Z) = 0. Every Whitehead group is separable ([3, XII.1.3]). Since a subgroup of a Whitehead group is also a Whitehead group, every Whitehead group is hereditarily separable. A group is said to be a Shelah group if it has cardinality 1 , is 1 -free, and for every countable subgroup B there is a countable subgroup B B such that for every countable subgroup C satisfying C B = B, C/B is free. In that case, we say that B has the Shelah property over B. In [7] and [9] it is proved a consequence of Martins Axiom plus CH that the Whitehead groups of cardinality 1 are precisely the Shelah groups. Notions of uniformization (in our sense) were rst dened in [2] and [8]. Let S be a subset of lim(1 ). If S, a ladder on is a function : which is strictly increasing and has range conal in . A ladder system on S is an indexed family = { : S} such that each is a ladder on . The ladder system is tree-like if whenever (n) = (m), then n = m and (k) = (k) for all k < n. For a cardinal 2, a -coloring of a ladder system on S is an indexed family c = {c : S} such that c : . A uniformization of a coloring c of a ladder system on S is a pair g, g where g: 1 , g : S and for all S and all n g (), g( (n)) = c (n). If such a pair exists, we say that c can be uniformized. We say that (, )-uniformization holds or that satises -uniformization if every -coloring of can be uniformized. A monochromatic colouring c of a ladder system is one such that for each S, c is a constant function. We shall, from now on, consider a monochromatic colouring with colours to be a function c: S (which gives the constant value, c(), of the colouring of ). Then a uniformization of a monochromatic colouring c is a pair f, f where for all S and all n f (), f ( (n)) = c(). If every monochromatic -colouring of can be uniformized we say satises monochromatic uniformization for colours. Dene a ladder system based on a countable set to be an indexed family = { : S} such that each is a function from to a xed countable set I. We can dene notions of colouring and uniformization analogous to those above. (See [3, pp. 367369].) The following two results, though stated and proved for ladder systems on a stationary subset of 1 , are true also for ladder systems based on a countable set. For ease of reference, we include the following result (compare [3, XII.3.3]).

442

revision:2000-10-27

modified:2000-10-30

5 Lemma 3 If there is a ladder system on a stationary subset of 1 which satises -uniformization (resp. monochromatic uniformization for colours), then there is a tree-like ladder system on a stationary subset of 1 which satises -uniformization (resp. monochromatic uniformization for colours). Proof. Suppose { : S} satises -uniformization (resp. monochromatic uniformization for colours). Choose a one-one map from < 1 to 1 , such that () ( ) if is a sequence extending and such that for any < 1 , ( ) (n) for all n dom( ). Let C be a closed unbounded subset of 1 consisting of limit ordinals such that for every C, [< ] . Let S = S C. For S , dene (n) = ( (0), . . . , (n) ). Then {: S } is tree-like and satises -uniformization (resp. monochromatic uniformization for colours). Remark. With a little more care we can prove that if there is a ladder system on S which satises -uniformization, then there is a tree-like ladder system on the same set S which satises -uniformization. (Compare [3, Exer. XII.17].) The third author has proved that there is a non-free Whitehead group of cardinality 1 if and only if there is a ladder system on a stationary subset of 1 which satises 2-uniformization. (See [3, XII.3].) The main theorem of this paper is an analogous necessary and sucient condition for the existence of a non-free hereditarily separable group. Since every Whitehead group is hereditarily separable, we can conclude that if there is a ladder system on a stationary subset of 1 which satises 2-uniformization, then there is a ladder system on a stationary subset of 1 which satises monochromatic uniformization for colours. It is perhaps reassuring to know that there is a simple direct proof of this consequence: Proposition 4 If there is a ladder system on a stationary subset of 1 which satises 2-uniformization, then there is a ladder system on a stationary subset of 1 which satises monochromatic uniformization for colours. Proof. By Lemma 3, we can assume that = { : S} is tree-like. Fix a monochromatic colouring c of with colours. Let c be a 2-colouring of such that for S, and k such that 2k > c(), c (2k + 1), c (2k + 2), . . . , c (2k+1 )

442

revision:2000-10-27

modified:2000-10-30

6 is the sequence 0c() , 1, 1, 1, . . . , 1 i.e., c() zeroes followed by 2k c() ones. Let g, g be a uniformization of c . To dene f, f , let f () be the least n so that n = 2k+1 where 2k > max{c(), g ()}. We dene f so that for all S and m f (), f ( (m)) = c(). To see that f is well-dened, consider the case when (m) = (m) and m f ( ), f (). Since is tree-like, (j) = (j) for all j m. By denition of f , there is a k such that 2k+1 m, and 2k > max{c(), c( ), g (), g ( )}. But then the values of g( (j)) for j = 2k + 1, . . . , 2k+1 code c() and also c( ), so c() = c( ). Remark. The proof actually shows that if is tree-like and satises 2uniformization, then satises monochromatic uniformization for colours.

Suciency

Theorem 5 If there is a ladder system on a stationary subset of 1 which satises monochromatic uniformization for colours, then there is a non-free group of cardinality 1 which is hereditarily separable. Proof. By hypothesis there is a stationary subset S of 1 and a ladder system = { : S} such that every monochromatic colouring with colours can be uniformized. By Lemma 3, without loss of generality we can assume that is tree-like. We begin by dening the group. Let pn (n < ) be an enumeration of the primes. The group G will be generated by {x : < 1 }{y,n: S, n < }, subject to the relations pn y,n+1 = y,n + x (n)
modified:2000-10-30

442

revision:2000-10-27

For any , we let G denote the subgroup of G generated by {x , yn : < , S }. It is standard that G is 1 -free (in fact 1 -separable) but not free. The rest of the proof will be devoted to proving that G is hereditarily separable. Assume B is a subgroup of G such that G/B is isomorphic to a subgroup of Q/Z and there is a nite set P of primes such that the order of every element of G/B is divisible only by primes in P ; we need to prove that if

7 Z is a rank 1 pure subgroup of B, then there is a projection of B onto Z. (See Lemma 2.) Since G/B is countable there is so that G + B = G and Z G . Fix such an and call it . Next, choose in G a system of representatives for G/B and let g: G G be the function which assigns to an element of G its coset representative. Finally choose n so that for all n n , G/B is uniquely pn -divisible. Let h1 be a projection of B G onto Z. Such a projection exists since B G is free. We will extend h1 to a projection h from B to Z by dening h on {x g(x ), y,n g(y,n): , n , S, }. Such a denition suces (provided it works), since B is generated by this set together with B G . Dene the colouring c: (S \ ) G so that c() = g(y,n ). (Since the values are taken in G , which is a countable set, this is an allowable colouring.) Let the pair f, f uniformize c. We can assume that f () n and that (f ()) for all S \ . We dene the function h in three stages. First for each and n f (), h(y,n g(y,n)) = 0 and h(x (n) g(x (n) )) = h1 (pn g(y,n+1) g(y,n) g(x (n) )). There are two potential problems with the second denition: namely, why is the right-hand side of the equation dened; and why is the denition independent of ? (Note that x (n) may equal x (n) for some ). For the rst problem, note pn g(y,n+1) g(y,n) g(x (n) ) pn y,n+1 y,n x (n) 0 (mod B), hence pn g(y,n+1) g(y,n) g(x (n) ) B G . The independence of the denition from is a consequence of the following lemma, after noting that n f () implies g(y,n ) = g(y,n ) when (n) = (n). Lemma 6 Let n n . Suppose that , S, (n) = (n) and g(y,n ) = g(y,n ). Then g(y,n+1) = g(y,n+1).

modified:2000-10-30

revision:2000-10-27

Proof. The proof is by induction on n n . Since the ladder system is tree-like, we can assume by induction that g(y,n) = g(y,n). Now pn g(y,n+1) g(y,n)+g(x (n) ) = g(y,n)+g(x (n) ) pn g(y,n+1) (mod B).

Since G/B is uniquely pn -divisible (by choice of n ), g(y,n+1) g(y,n+1) (mod B) and hence g(y,n+1) = g(y,n+1) by denition of g.

442

8 To complete the denition, the second step is to dene h(x g(x )) arbitrarily (say 0) for any not covered in the rst step (i.e., and = (n) for any and any n f ()). Finally, for all and n < f () dene h(y,n g(y,n)) as required by the equation (y,n g(y,n))+ (x (n) g(x (n) )) pn (y,n+1 g(y,n+1))+ g(y,n) + g(x (n) ) pn g(y,n+1) = 0. (Do this by downward induction.) It remains to see that h induces a homomorphism. Consider the free group F = L (B G ) where L is the group freely generated by {u , w,n: S, n , > , (n) > }. There is a surjective map : F B which is the identity on B G and such that (u ) = x g(x ) and (w,n) = y,n g(y,n). The kernel K of is generated by elements of the form (w,n + u (n) pn w,n+1)+(g(y,n)+g(x (n) )pn g(y,n+1)). Let h: F Z be dened so that h B G = h1 , h(u ) = h(x g(x )) and h(w,n) = h(y,n g(y,n)). Since h is constantly 0 on K, it induces a homomorphism from B to Z which agrees with h on the generators of B. Remark. The same proof works with any tree-like ladder system based on a countable set. (The assumption that the ladder system is tree-like is necessary, as witnessed by Hausdor gaps). In particular, if there is a set of 1 branches through the binary tree of height which satises monochromatic uniformization for colours, then the group built from these branches is hereditarily separable. This group is just the group constructed in [3, VII.4.3]; there it is shown that MA + CH implies this group is hereditarily separable. Given these comments, one might expect that it is possible to show that MA + CH implies that any system of 1 branches through the binary tree satises monochromatic uniformization for colours. Indeed, this is the case: given a set of 1 branches and a monochromatic colouring c by colours, let the poset, P, consist of pairs (s, B) where s is a function from n 2 and B is a nite subset of the branches such that for all b B, s(b n) = c(b). If (t, c) P and dom(t) = m 2, we dene (t, C) (s, B) i s t, B C and for all b B and n k m, t(b k) = s(b n) = c(b). The proof that for each n, {(s, B): n 2 dom(s)} is dense uses the fact that the colouring is monochromatic. On the other hand the poset is c.c.c., since any two conditions with the same rst element are compatible.

442

revision:2000-10-27

modified:2000-10-30

Necessity

The following lemma can be derived as a consequence of the fact that the Richman type of a nite rank torsion free group is well-dened (see [6] or [5]); but for the convenience of the reader we give a self-contained proof. Lemma 7 Suppose A is a torsion free group of rank r + 1 and every rank r subgroup is free. If B and C are pure rank r subgroups, then the type of A/B is the same as the type of A/C. Proof. The proof is by induction on r. We can assume that r 1 and B = C. Consider rst the case r = 1; then B C = 0. If b B and c C are generating elements, then A Qb Qc and it is enough to prove that m divides b (mod C) if and only if m divides c (mod B). Now if m divides b (mod C), then b = ma + nc for some a A and n Z. Since B = b is pure in A, m and n must be relatively prime. Hence there exist s, t Z such that ns + mt = 1. But then c = m(tc sa) + sb, so m divides c (mod B). Now suppose r > 1. Consider B C; since r + 1 = rk(B + C) = rk(B) + rk(C) rk(B C) and 2r > r + 1, we have that rk(B C) 1. Since B C is a pure free subgroup of A we can nd x B C which is a pure subgroup of A. Note that A/ x has the property that every subgroup of rank r 1 is free. Now apply the induction hypothesis to A/ x , B/ x and C/ x . Theorem 8 If there is a non-free hereditarily separable group G of cardinality 1 , then there is a ladder system on a stationary subset of 1 which satises monochromatic uniformization for colours. Proof. We can write G = <1 G , a union of a continuous chain of countable free pure subgroups where, without loss of generality, we can assume that there is a stationary subset S of 1 , consisting of limit ordinals, and an integer r 0 such that for all S, G+1 /G is non-free of rank r + 1 and every subgroup of G+1 /G of rank r is free. (We use the fact that if a stationary subset of 1 is partitioned into countably many pieces, then one of the pieces must be stationary: cf. [3, II.4.5].) Thus for each S there is def a pure free subgroup F /G of G+1 /G of rank r such that M = G+1 /F is rank 1 and non-free. Moreover either M is divisible or there is a prime

442

revision:2000-10-27

modified:2000-10-30

10 p and an element y + F of M which is not divisible by p . Without loss of generality (again using [3, II.4.5]), we can assume that there is a prime p such that p = p for all S such that M is not divisible. For each S, let {y : 0 r} G+1 be such that {y + G : 0 r 1} is a basis of F /G , yr F and yr + F is not divisible by p in M if / M is not divisible. Then G+1 is generated by G {y : r} {zn : n }, where the zn satisfy equations
rn zn = r where gn G and rn , s,n Z. Dene functions on for each S by: (n) = gm , s,m , rm : r, m n . s,ny + gn

revision:2000-10-27

Notice that (n) determines the isomorphism type of the nitely generated subgroup of G+1 /G generated by (the cosets of) {y : r} {zm : m n}. As in [3, XII.3] see especially Theorem XII.3.3 and the beginning of the proof of XII.3.1 (p. 381) if we show that = { : S} satises monochromatic uniformization for colours, then there is a ladder system on a stationary subset of 1 with the same property. So x a monochromatic colouring c: S of . We are going to use c to dene a subgroup B of G with a pure cyclic subgroup Z. By the hypothesis on G, there will be a projection h: B Z. Because of the way we dene B we will be able to use h to dene f : { (n): S, n } such that for each S, f ( (n)) = c() for all but nitely many n . We will dene a continuous chain of subgroups B of G by induction on and let B = 1 B . To begin, let {xn : n } be a basis of G0 , and let B0 be the subgroup of G0 generated by {px0 } {pxn+1 xn : n }. Thus def G0 /B0 Z(p ) and Z = Zpx0 is a pure subgroup of B0 . = Let A = {tn : n } G0 be a complete set of representatives of G0 /B0 such that t0 = 0. For each pair (d, a) where d > 0 and a A, x an element [d, a] such that dt[d,a] + B0 = a + B0 . We will dene the B so that for all 1. B + G0 = G and

442

modified:2000-10-30

11 2. for all < , B G = B . Notice then that G /B G0 /B0 , and Z is pure in each B . = The crucial case is when we have dened B already and S. We will dene B,m by induction on m and then let B+1 = m B,m . Let
B,0 = B {y : < r} {yr tc() } .

Then B,0 G = B since {y : r} is independent mod G . Suppose B,m has been dened so that B,m G = B . Thus (B,m + G )/B,m G0 /B0 . = Let dm > 0 be minimal such that dm zm B,m + G . If dm zm am A (mod B,m ), let B,m+1 = B,m + Z(zm t[dm ,am ] ). Then we will have B,m+1 G = B . So, in the end, B+1 G = B . Moreover, B+1 + G0 = G+1 , because, by construction, every generator of G+1 belongs to B+1 + G0 . If S, the construction of B+1 is essentially the same, except that the / colouring c plays no role; we begin with a set Y G+1 which is maximal independent mod G and let B,0 = B Y ; then dene B,m by induction as before (using a well-ordering of type of a set of generators of G+1 mod G ). This completes the description of the construction of B. Now x a projection h: B Z and x a well-ordering, , of Z r+1 of order type . We are going to dene the uniformizing function f . We must dene f () for each of the form (n). (Note that there may be many such that = (n).) Suppose = gm , sm , rm : r, m n . Let = () be minimal such that gm G for all m n. For each k (k) we can construct a group B just as in the construction of B+1 , which is generated by B {y : < r} {yr tk } together with elements of the form zm t[dm ,am ] (m n) where the zm satisfy the relations () rm zm = sm y + g m .

revision:2000-10-27

modified:2000-10-30

This is an abstract group, which can be regarded as a subgroup of the free group on G {y : r} {zm : m n} modulo the relations in G and

442

12 the relations given by (i.e., the equations ()). If is such that (n) = (k) and c() = k, then there is an embedding of B into B,n+1 which xes B (and is an isomorphism if = ). As before, Z = px0 is a pure subgroup (k) (k) of B . Since B is isomorphic to a subgroup of G, it is separable. Since h exists, there is a -least tuple w : r k in Z r+1 for (k) which there is a projection h : B Z with h B = h B , h (y ) = w (k) for < r and h (yr tk ) = wr . Note that this tuple determines h on B . Dene f () = k. We have to show that this denition works, that is, for each S, f ( (n)) equals c() for suciently large n . Fix S. With respect to the wellordering , there are only nitely many wrong guesses which come before the right answer h(y ): < r h(yr tc() ) c() . So we just have to show that no wrong guess can work for all n if it involves a k = c(). If there were a wrong guess that worked for all n for some k = c(), then there would be a projection h onto Z whose domain, B , contains {y : < r} {yr tk }, elements of the form zn a,n for all n (with a,n A), and B where is minimal such that gn G for all n . Let G denote
G0 + B = G + {y : r} {zn : n } .

Notice that for each g G0 , there is a j such that pj g B0 B , which is a subset of dom(h) and dom(h ). Hence we can extend h and h uniquely to homomorphisms from G into Q(p) Z. (Here Q(p) is the group of rationals whose denominators are powers of p.) Denote the extension by h (resp. h ). =h. We claim that h Assume for the moment that this is true. Then h(yr ) = h (yr ). Now h(yr tc() ) Z and h (yr tk ) Z. So h(tc() tk ) = h(yr tc() ) h (yr tk ) Z.

modified:2000-10-30

442

revision:2000-10-27

Since k = c(), there is an s Z so that s(tc() tk ) x0 (mod B0 ), so h(x0 ) Z. But p(h(x0 )) = h(px0 ) = px0 ; this contradicts the fact that px0 generates Z. It remains to prove the claim. Let H = G/G , which is isomorphic to G+1 /G . Now h h induces a homomorphism from H into Q(p) since h and h agree on B , hence on G0 (since G0 /B0 Z(p )) and so on G . So =

13 it suces to prove that Hom(H, Q(p) ) = 0. Assume, to the contrary, that there is a non-zero : H Q(p) . Let K = ker(); then the rank of K is r. Now H/K is isomorphic to a subgroup of Q(p) and hence is not divisible. By Lemma 7, H/K is isomorphic to M = G+1 /F . So by the choice of S and p, H/K is not p-divisible; since H/K is isomorphic to a subgroup of Q(p) , this implies H/K is free. But this is impossible, since H is not free and K is a subgroup of rank r, and hence free. Corollary 9 If there is an hereditarily separable group of cardinality 1 which is not free, then there exist 21 dierent 1 -separable groups of cardinality 1 which are hereditarily separable. Proof. By the theorem, the given hypothesis implies that there is a ladder system on a stationary subset of 1 which satises monochromatic uniformization for colours. Using this ladder system, we can construct an 1 -separable group which is hereditarily separable as in the proof of Theorem 5. By a standard trick we can, in fact, construct such groups with 21 dierent -invariants. (Compare [3, VII.1.5].) Similarly to the proof of Theorem 8 we can prove the following:
modified:2000-10-30

Theorem 10 If there is an hereditarily separable group G of cardinality 1 which is not a Shelah group, then there is a ladder system based on a countable set which satises monochromatic uniformization for colours.

Consistency of Case 4

The consistency of Case 4 in the Introduction will now follow from Theorem 5 and the following set-theoretic result.

442

revision:2000-10-27

Theorem 11 It is consistent with ZFC + GCH that the following all hold: (i) there is a ladder system on a stationary subset of 1 which satises monochromatic uniformization for colours; (ii) there is no ladder system on a stationary subset of 1 which satises 2-uniformization; (iii) (E) holds for every stationary subset, E, of every regular cardinal > 1 .

14 Proof. We assume familiarity with the methods of [10]. For simplicity let our ground model be L; x a stationary, co-stationary subset S of 1 and a ladder system on S. Our forcing P will be an iterated forcing with countable support using two types of posets: R, which adds a Cohen subset of 1 , and Q(c) which is the poset uniformizing a monochromatic colouring c: S of with countable conditions, i.e., Q(c) = {f : f : for some successor < 1 and for all S , f ( (n)) = c() for almost all n }. In the iteration P we force with R at successors of even ordinal stages and c) at the successors of odd ordinal stages, where, as usual, the force with Q( names c are chosen so that all possibilities occur. The posets R and Q(c) are proper, so stationary sets are preserved by P. Also, P is (1 \ S)-closed and of cardinality 2 , so GCH holds in the generic extension as well as (E) for every stationary subset of every regular cardinal > 1 . It remains to show that in the generic extension 2-uniformization fails for every stationary subset E of 1 and every ladder system = { : E}. By doing an initial segment of the forcing we can assume that E and are both in the ground model. Let X be the generic set for the rst copy of R in the iteration of P. Consider the 2-colouring {c : E} of dened as follows: c (n) = 0 if and only if + n X. The proof that this colouring is not uniformized now follows along the same lines as that in [10]. Corollary 12 It is consistent with ZFC + GCH that there is an hereditarily separable group of cardinality 1 which is not free, and every Whitehead group (of arbitrary cardinality) is free. Proof. We use the model of ZFC + GCH constructed in Theorem 11. Clause (i) in Theorem 11 together with Theorem 5 imply that there is a nonfree hereditarily separable group of cardinality 1 . Clause (ii) implies that there is no non-free Whitehead group of cardinality 1 . (See [3, XII.3.1(i)].) Finally clause (iii) enables one to do an inductive proof that there is no nonfree Whitehead group of any cardinality (as, for example, in [3, XII.1.6]).

442

revision:2000-10-27

modified:2000-10-30

15

Cases 2 and 3

In Cases 2 and 3 of the Introduction we are in the situation where there is a Whitehead group which is not free; here we shall consider two hypotheses which are stronger than this hypothesis: rst, that there is a Whitehead group which is not a Shelah group; and, second, that every Shelah group is a Whitehead group. The following theorem says that the hypothesis that there is a Whitehead group which is not a Shelah group is not consistent with Case 3. It also gives another consistency proof for Case 2 since it is known that it is consistent that there are Whitehead groups which are not Shelah groups (see [3, XII.3.11]). Theorem 13 If there is a Whitehead group of cardinality 1 which is not a Shelah group, then there is an hereditarily separable group which is not a Whitehead group. Proof. By [3, XII.3.19] there is a ladder system = { : 1 } based on a countable set I which satises 2-uniformization. Without loss of generality we can assume that I = and each : is strictly increasing. Moreover, as in the proof of Lemma 3, we can assume that is tree-like, and hence, as in the proof of Lemma 4, satises monochromatic uniformization for colours. For each 1 and n let k,n = ( (n) + 1)! and k,n = (n)! Let G be the group generated by {xn : n < } {y,n: 1 , n < }, subject to the relations k,n+1y,n+1 = y,n + x (n) . (1)

modified:2000-10-30

442

revision:2000-10-27

As in the proof of Theorem 5, G is hereditarily separable. It remains to show that G is not a Whitehead group. For this we shall dene an epimorphism : H G with kernel Z which does not split. Let H be the group generated by {xn : n < } {y,n: 1 , n < } {z}, subject to the relations k,n+1y,n+1 = y,n + x (n) + k,n+1z. (2)

16 There is an epimorphism taking y,n to y,n, xm to xm , and z to 0; the kernel of is the pure subgroup of H generated by z. Aiming for a contradiction, assume there is a splitting of , i.e., a homomorphism : G H such that = 1G . So (y,n) y,n ker() for all < 1 , n . Since a countable union of countable sets is countable, there exists = such that (0) = (0), (1) = (1) and (y,0) y,0 = (y,0) y,0 . Let m ( 2) be minimal such that (m) = (m). We claim that (y,n) y,n = (y,n) y,n if n < m. The proof is by induction on n < m; the initial case n = 0 is by choice of and . So supposing the result is true for n < m1, we will prove it for n+1. Applying the homomorphism to equation (1) for as well as and subtracting we get that (in H) k,n+1(y,n+1) k,n+1(y,n+1) = (y,n) (y,n) since x (n) = x (n) because n < m. But then by induction k,n+1(y,n+1) k,n+1(y,n+1 ) = y,n y,n . (4) (3)

Now by equation (2), since x (n) = x (n) and k,n+1 = k,n+1 (the latter because n < m 1), we have
modified:2000-10-30

k,n+1y,n+1 k,n+1y,n+1 = y,n y,n. so by equations (4) and (5) we have k,n+1((y,n+1) y,n+1) = k,n+1 ((y,n+1) y,n+1 ).

(5)

(6)

Since n < m 1, k,n+1 = k,n+1, so cancelling k,n+1 from equation (6), we obtain the desired result, and the claim is proved. Now equation (4) holds for n = m 1 so k,m (y,m) k,m (y,m ) = y,m1 y,m1 . (7)

revision:2000-10-27

In this case, instead of (5) we have k,m y,m k,m y,m (k,m k,m )z = y,m1 y,m1 . so combining (7) and (8) we have k,m (y,m) k,m (y,m ) = k,my,m k,m y,m (k,m k,m )z. (9) (8)

442

17 Say (m) < (m). Then k,m , k,m and k,m are all divisible by k,m = ( (m) + 1)! so equation (9) implies that ( (m) + 1)! divides k,m z = (m)!z in H which is a contradiction, since z generates a pure subgroup of H. Now we consider the hypothesis that every Shelah group is a Whitehead group. This is true in a model of Martins Axiom, in which case there are hereditarily separable groups which are not Whitehead groups, i.e., Case 2 holds. Here we show that it is consistent that every Shelah group is a Whitehead group but every hereditarily separable group is a Whitehead group, i.e., there is a model for Case 3 in which every Shelah group is a Whitehead group. For this purpose we use the notion of stable forcing. A poset, P, is stable if for every countable subset P0 there is a countable subset P1 so that for every p P there is an extension p of p and an element p P1 so that p and p are compatible with exactly the same elements of P0 . In [1] the basic facts about c.c.c. stable forcings are proved. There are a few basic facts that we will use: Proposition 14 1. [1] Any iteration of c.c.c. stable forcings with nite support is c.c.c. and stable. 2. The forcing adding any number of Cohen reals is stable.
modified:2000-10-30

3. If A is a Shelah group and 0 ZB A 0 is a short exact sequence, then the nite forcing, Q(), constructing the splitting of is (c.c.c. and) stable. Proof. We will prove only the last of the statements. Write A as < A (an 1 -ltration) where each A is pure in A and A+1 has the Shelah property over A . The forcing Q() is the set of partial splittings of whose domains are nite rank pure subgroups of A. (This forcing is c.c.c. see, e.g., [3, XII.1.11].) Given P0 , choose so that every element of P0 has domain contained in A . Let P1 be the set of elements of Q() whose domains are contained in A+ . Given p Q(), let G be the pure subgroup of A generated by A dom(p). There exists n such that GA+ = GA+n (since G has nite rank over A ). Then G = (G A+n ) y0 , . . . , ym

442

revision:2000-10-27

18 for some y0 , . . . , ym since G A+n+1 = G A+n and A+n+1 has the Shelah property over A+n . Extend p to p Q() such that dom(p ) = M y0 , . . . , ym where M G A+n is a nite rank pure subgroup of G A+n such that dom(p) M y0 , . . . , ym . Let p = p M P1 . It suces to prove that if q P0 is compatible with p , then q is compatible with p . So suppose that r Q() such that r q, p . Without loss of generality dom(r) G A+n . Dene r with dom(r ) = dom(r) y0 , . . . , ym by: r dom(r) = r and r y0 , . . . , ym = p y0 , . . . , ym . Clearly r q, p . Moreover, dom(r ) is pure in A since G is pure in A and dom(r) is pure in G A+n ; so r Q(). Theorem 15 It is consistent that every Shelah group is a Whitehead group and every hereditarily separable group is a Whitehead group. Proof. We do our forcing over L by iteratively adding subsets of 1 by nite conditions and adding splittings for Shelah groups. More precisely, our forcing P will be an iterated forcing with nite support and of length 2 using two types of posets: R, the nite functions from 1 to 2, and Q() which is the nite forcing splitting as in Proposition 14(3). If we choose the iterants correctly, then in the generic extension every Shelah group of cardinality 1 will be a Whitehead group and (E) will hold for every stationary subset of every regular cardinal greater than 1 . It will suce then to show that every hereditarily separable group of cardinality 1 is a Shelah group (because we have all instances of diamond above 1 : cf. [3, Exer. XII.16(ii)].) By Theorem 10 it is enough to show that, in the generic extension, if = { : < 1 } is a ladder system based on , then does not satisfy monochrome uniformization for colours. In fact, we will show that does not satisfy monochrome uniformization for 2 colours. By absorbing an initial segment of the forcing into the ground model we can assume that is in the ground model and the forcing P is rst R, the nite functions from 1 to 2, followed by a name T for a c.c.c. stable forcing. We dene the colouring c: 1 2 to be the generic set for R; let c be a name for c.

442

revision:2000-10-27

modified:2000-10-30

19 In order to obtain a contradiction, assume that this colouring can be uni formized. Then there is a pair f , f of names for functions and there is a p P such that p f , f uniformizes c. Now let P0 be a countable subset of P containing p as well as for every n < , a maximal antichain which determines the value of f (n). Let P1 be as given by the denition of a stable poset for this P0 . For each 1 choose p p so that p determines the values of f () and c() and there exists p P1 so that p and p are compatible with exactly the same elements of P0 . Say p f () = m c() = e .

By the pigeon-hole principle, there exists an uncountable set E 1 and p P1 so that for all E, p = p . Since p is compatible with p , there exists q1 p , p . By the denition of R, there exists 0 E and q2 P such that q1 q2 and q2 c(0 ) = e0 . So q2 k > m0 s.t. f ( (k)) = e0 . ( (k0 )) = e . Thus there exists q3 q2 and k0 > m0 such that q3 f 0 But there is a maximal antichain in P0 of conditions forcing the value of f ( (k0 )). Hence there exists r P0 and q4 such that r q4 , q3 q4 and r f ( (k0 )) = 1 e0 . Then r is compatible with p = p 0 and hence with p0 . But this is a contradiction since p0 c() = e0 f ( (k0 )) = c() since k0 > m0 .
modified:2000-10-30

Finitely many primes

442

revision:2000-10-27

The proof of Theorem 5 uses innitely many primes. Otherwise said, the type of the (torsion-free rank one) non-free quotients G+1 /G in that construction is (1, 1, 1, . . .). We may ask what happens if we are allowed only nitely many primes. For example, we may consider modules over Z(P ) (where P is a set of primes and Z(P ) denotes the rationals whose denominators in reduced form are not divisible by an element of P ) and ask whether the main theorem, Theorem 1, holds. If P is innite, i.e., Z(P ) has innitely many primes, then our proofs apply and there is a non-free hereditarily separable Z(P ) module of cardinality 1 if and only if there is a ladder system on 1 which satises monochrome uniformization for colours. On the other hand if the cardinality of P is nite but at least two, we can show that Theorem 1 does not hold, and Case 4 in the Introduction is impossible. In fact, this section is devoted to proving the following result:

20 Theorem 16 Suppose R is a countable p.i.d. with only nitely many but at least 2 primes. If there is an hereditarily separable R-module of cardinality 1 which is not free, then there is a Whitehead R-module of cardinality 1 which is not free. Proof. The method of proof is to show that if there is an hereditarily separable R-module of cardinality 1 which is not free, then there is a ladder system on a stationary subset of 1 which satises 2-uniformization. We rst prove that there is a tree-like ladder system = { : S} on a stationary subset S of lim(1 ) such that for every 2-colouring c = {c : S} of , there is a function f : 1 2 such that for all S there exists m such that f ( (n), m ) = c (n) for all n .

()

Let N be an hereditarily separable R-module of cardinality 1 . As in the proof of Theorem 8, we write N = <1 N as a union of a continuous chain of countable free pure submodules where there is a stationary subset S of 1 , consisting of limit ordinals, and an integer r 0 such that for all S, N+1 /N is non-free of rank r + 1 and every subgroup of N+1 /N of rank r is free. There is a pure free subgroup F /N of N+1 /N of rank r such that N+1 /F is rank 1 and non-free. It follows from the fact that there are only nitely many primes that the type of N+1 /F is (t1 , t2 , . . . , tn ) where each ti is either 0 or and at least one ti = . Thus without loss of generality we may assume that there is a xed prime p R such that for all S there exists {y : 0 r} N+1 / such that {y + N : 0 r 1} is a basis of F /N , yr F and yr + F is p-divisible in N+1 /F . Then N+1 contains elements zn (n ), where z0 = yr and the zn (n 1) satisfy equations (
n) pzn = r s,n y + gn + ,n j<n rj zj

revision:2000-10-27

modified:2000-10-30

where gn N , r ,n, s,n R, and no element of N+1 / F {zn : n } has order p. Dene functions on for each S by: ,m (n) = gm , s,m , rj : r, m n, j < m .

442

21 Let c be a 2-colouring of = { : S}. Following the pattern of the proof of Theorem 8, we will use c to dene a subgroup B of N . We begin by letting {xn : n } be a basis of N0 , and letting B0 be the subgroup of N0 generated by {px0 } {pxn+1 xn : n }. Also, let A = {tn : n } N0 be a complete set of representatives of N0 /B0 such that t0 = 0 and for each a A, x an element [p, a] such that pt[p,a] + B0 = a + B0 . Assume we have dened B so that B + N0 = N and for all < , B N = B . We now dene B,m inductively so that zm B,m + N0 . Let B,0 be generated by B {y0 , . . . , yr }. If B,m1 has been dened, we have pzm B,m1 + N0 , so pzm a (mod B,m1 ) for some a A. Let m m
B,m = B,m1 + R(zm t[p,a ] c (m)x0 ). m

Having dened B,m for all m, we can extend m B,m to B+1 such that B+1 + N0 = N+1 and B+1 N = B . Finally, let B = <1 B and x a projection h: B Rpx0 and a wellordering, , of Rr+1 of order type . Extend h to a homomorphism, also denoted h, from N into Qpx0 , where Q is the quotient eld of R. Given of the form (n) and m , we are going to dene f (, m). Let w m : r be the mth tuple in Rr+1 according to . We shall suppose that
modified:2000-10-30

(#m )

h(y ) = w m px0

for r for some such that = (n), and show that under this supposition (and with the information given by ) we can compute c (n); we will then dene this value of c (n) to be f (, m). Since one of our suppositions (#m ) about the values of h(y ) must be right, () will be proved. The proof is by induction on k n that we can compute h(zk ), a , and k c (k). In fact, for 0 < k n we have an equation (
k) pzk = r s,k y + gk + ,k j<k rj zj

revision:2000-10-27

satised by zk . Since by induction and our supposition we know the value of h for all the elements on the right-hand side, we can compute h(zk ). Since by induction we also know c k, we know B,k1 , so we can calculate a ( pzk k (mod B,k1 )). Finally, we know that h(zk t[p,a ] c (k)x0 ) = h(zk t[p,a ] ) c (k)x0 k k

442

22
belongs to Rpx0 , and we know h(zk t[p,a ] ) by induction (because we know k h N0 ). Now h(zk t[p,a ] ) x0 and h(zk t[p,a ] ) cannot both belong to Rpx0 . k k If h(zk t[p,a ] ) belongs to Rpx0 , c (k) must equal 0; otherwise let c (k) = 1. k (If the latter value does not make h(zk t[p,a ] c (k)x0 ) belong to Rpx0 , then k our supposition must have been wrong, and we can let f (, m) be arbitrary.) This completes the proof of (). At this point we use the assumption that there are at least two primes. Then the proof of necessity, i.e. of Theorem 8, is still valid. (Referring to the last paragraph of that proof, we use the fact that there are two primes when we assert that H/K is not divisible.) Moreover, there is a single ladder system which satises the property of () as well as monochromatic uniformization for colours. (Indeed, by reducing to a smaller set we can assume that the same set S is used in both the proof of () and the proof of Theorem 8; then we can let be a ladder system derived from functions which give combined information about the equations used in the proof of () and the equations used in the proof of Theorem 8.) Given a 2-colouring c of , let f be as in (). Dene a monochromatic colouring c of by: c () = m where m is such that f ( (n), m ) = c (n) for all n . There is a uniformization g, g of c . Dene h: 1 2 by: h() = f (, g()). Then for all S for suciently large n,

modified:2000-10-30

h( (n)) = f ( (n), g( (n))) = f ( (n), m ) = c (n).

The third author, in [9, Thm. 3.6], claimed to prove that if the non-freeness of G involves only nitely many primes, then G is hereditarily separable if and only if G is Whitehead. However, the proof given seems to be irredeemably defective. We do not know if the result claimed is true. Thus we still have the following open questions: If R is a countable p.i.d. with exactly one prime, does Theorem 16 hold? If R has nitely many primes, is every hereditarily separable R-module of cardinality 1 a Whitehead module? If not, nd a combinatorial equivalent, analogous to Theorem 1, to the existence of a hereritarily-separable R-module which is not a Whitehead module.

442

revision:2000-10-27

23

References
[1] U. Avraham and S. Shelah, Forcing with stable posets, J. Symbolic Logic 47, 3742 (1982). [2] K. Devlin and S. Shelah, A weak version of which follows from 20 < 21 , Israel J. Math. 29, 239247 (1978). [3] P. C. Eklof and A. H. Mekler, Almost Free Modules, NorthHolland (1990). [4] P. C. Eklof, A. H. Mekler and S. Shelah, Uniformization and the diversity of Whitehead groups, Israel J. Math. 80, 301321 (1992). [5] H. P. Goeters and W. J. Wickless, Hyper- groups, Comm. in Algebra, 17, 12751290 (1989). [6] F. Richman, A class of rank-2 torsion free groups, in Studies on Abelian Groups (B. Charles, ed.), Springer-Verlag, 327334, (1968). [7] S. Shelah, Innite abelian groups, Whitehead problem and some constructions, Israel J. Math 18, 243256 (1974).
modified:2000-10-30

[8] S. Shelah, Whitehead groups may not be free even assuming CH, I, Israel J. Math. 28 193203 (1977). [9] S. Shelah, On uncountable abelian groups, Israel J. Math. 32, 311330 (1979). [10] S. Shelah, Whitehead groups may not be free even assuming CH, II, Israel J. Math. 35, 257285 (1980).

442

revision:2000-10-27