Introduction
Factorization into irreducibles is unique in a factorial domain, but uniqueness of the number of factors is a substantially weaker condition. A domain is half-factorial if every nonzero nonunit factors into irreducibles and all such factorizations have the same length. Carlitz’s theorem characterizes half-factorial rings of integers of algebraic number fields by class number at most two [1 ] . The corresponding question for arbitrary Dedekind domains is different because the distribution of prime ideals among the ideal classes is no longer fixed by the number-field setting.
Claborn proved that every abelian group is the ideal class group of a Dedekind domain [3 ] . Grams and subsequent realization theorems permit prescribed classes of prime divisors under the appropriate generating condition [11 , 2 ] . Thus the abstract class group does not by itself determine the factorization lengths of a particular Dedekind domain. Nevertheless, it is natural to ask whether every abelian group admits some prime-divisor distribution for which the resulting domain is half-factorial. Equivalently, one asks:
Is every abelian group isomorphic to the ideal class group of a half-factorial Dedekind domain?
This is Problem 15 in the survey of Geroldinger, Kim, and Loper [10 §12.1] .
The group-theoretic formulation uses zero-sum sequences. For S ⊆ G S\subseteq G S ⊆ G , let B ( S ) \mathcal{B}(S) B ( S ) be the monoid of finite sequences over S S S whose sum is zero. The classical realization criterion says that G G G occurs as the class group of a half-factorial Dedekind domain precisely when it has a monoid-generating subset S S S for which B ( S ) \mathcal{B}(S) B ( S ) is half-factorial [2 Theorem 3.6] . Group generation and monoid generation agree for torsion groups. For such a group, the Skula–Śliwa cross-number criterion characterizes half-factoriality by k ( U ) = 1 k(U)=1 k ( U ) = 1 for every minimal nonempty zero-sum sequence U U U , where k ( U ) = ∑ s v s ( U ) / ord ( s ) k(U)=\sum_s v_s(U)/\operatorname{ord}(s) k ( U ) = ∑ s v s ( U ) / ord ( s ) [7 Lemma 3.2] .
Michel and Steffan obtained early positive realization results [12 ] . Geroldinger and Göbel proved that every Warfield group admits half-factorial generators and, in particular, that every countable abelian group does so [8 Theorems 5.1 and 5.3] . They also reduced the universal problem to reduced uncountable unbounded primary groups and singled out the torsion-completion of ⨁ j ≥ 1 C p j \bigoplus_{j\geq1} C_{p^j} ⨁ j ≥ 1 C p j as an unresolved example [8 pp. 855–856] . The positive lifting and reduction statements are established results; we use them with explicit attribution rather than treating them as contributions of this paper.
Our main contribution is a restriction on arbitrary half-factorial generating sets. It does not require generators of the socle, compatible generators of bounded layers, or a prescribed character. Say that an abelian p p p -group is p p p -adically separated if ⋂ r ≥ 1 p r P = 0 \bigcap_{r\geq1}p^rP=0 ⋂ r ≥ 1 p r P = 0 . The principal conclusions are as follows.
Main Theorem. Let p p p be a prime.
(1) If a p p p -adically separated abelian p p p -group P P P has a half-factorial generating subset, then
∣ P ∣ ≤ max { ℵ 0 , ∣ P / p P ∣ } . |P| \leq\max\{\aleph_0, |P/pP|\}. ∣ P ∣ ≤ max { ℵ 0 , ∣ P / pP ∣ } . (2) The group P p = t ( ∏ j ≥ 1 Z / p j Z ) P_p=t\left(\prod_{j\geq1}\mathbb{Z}/p^j\mathbb{Z}\right) P p = t ( ∏ j ≥ 1 Z / p j Z ) , where t t t denotes the torsion subgroup, has no half-factorial generating subset. Consequently it is not the ideal class group of a half-factorial Dedekind domain, nor the divisor class group of a half-factorial Krull monoid.
(3) For every uncountable cardinal κ ≤ 2 ℵ 0 \kappa\leq2^{\aleph_0} κ ≤ 2 ℵ 0 , there is a reduced abelian p p p -group of cardinality κ \kappa κ with no half-factorial generating subset. The least cardinality of an abelian group not realizable as the class group of a half-factorial Dedekind domain is therefore ℵ 1 \aleph_1 ℵ 1 .
The proof of the cardinality bound is finite at its decisive step. Given a finite abelian p p p -group G G G , a generating set S S S , and r ≥ 1 r\geq1 r ≥ 1 , we select A ⊆ S A\subseteq S A ⊆ S such that K = ⟨ A ⟩ K=\langle A\rangle K = ⟨ A ⟩ is a direct summand with p r G ⊆ K p^rG\subseteq K p r G ⊆ K and K [ p ] ⊆ p r G K[p]\subseteq p^rG K [ p ] ⊆ p r G . Two distinct generators of the same order that become equal modulo K K K without losing order yield a minimal zero-sum sequence of cross number greater than one. It follows that, for a half-factorial generating set of any abelian p p p -group,
ord ( s ) = ord ( t ) = p n , s − t ∈ p r P , p n − 1 s ∉ p r P ⟹ s = t . \operatorname{ord}(s)=\operatorname{ord}(t)=p^n,\quad s-t\in p^rP,\quad p^{n-1}s\notin p^rP\quad\Longrightarrow\quad s=t. ord ( s ) = ord ( t ) = p n , s − t ∈ p r P , p n − 1 s ∈ / p r P ⟹ s = t .
Stratifying the generators by n n n and r r r gives the cardinality bound. This separates the present argument from obstructions to particular extension procedures: the necessary subgroup is selected from the very generating set being tested.
The complementary results address how half-factorial generating sets behave under algebraic constructions. We extend the Geroldinger–Göbel free-kernel decomposition to quotients that may contain elements of infinite order. For torsion kernels, we prove that half-factoriality of an order-preserving lifted generating set is equivalent to splitting along those representatives. We also construct weakly half-factorial generators for every torsion abelian group, impose a prescribed character in bounded primary groups, and give explicit obstructions to thinning, adjoining roots, and repeated multiplication by a prime. The weak character criterion and its finite-group antecedents belong to the established theory [7 , 13 , 14 ] ; the additional constructions and proofs are presented separately from the main nonrealization argument.
Section 2 fixes conventions and recalls the realization and cross-number criteria. Section 3 proves the Main Theorem and its cardinal refinements. Section 4 treats lifting, the classical reduction, and pure extensions. Section 5 studies characters and finite obstructions. Section 6 concludes with socle compatibility, contraction, and the limitations of the cardinality criterion.
Zero-sum sequences and realization
This section establishes the conventions used throughout the paper, records the exact connection with Dedekind class groups, and gives the elementary positive examples needed later. We distinguish the ambient group of labels from the intrinsic divisor class group of an unlabelled block monoid. Standard references for the group and factorization theory are [5 , 6 , 9 ] .
Conventions and factorization lengths
We write N = { 1 , 2 , … } \mathbb{N} = \{1, 2, \ldots\} N = { 1 , 2 , … } and N 0 = N ∪ { 0 } \mathbb{N}_0 = \mathbb{N} \cup\{0\} N 0 = N ∪ { 0 } . All groups are abelian and written additively. All monoids are commutative and cancellative and have an identity. A monoid is reduced if its only unit is its identity; an atom is a nonunit that is not a product of two nonunits. A monoid is atomic if every nonunit is a finite product of atoms. Write A ( M ) \mathcal{A}(M) A ( M ) for the set of atoms of a monoid M M M and q ( M ) \mathrm{q}(M) q ( M ) for its quotient group. For an atomic monoid M M M and a ∈ M a \in M a ∈ M , let L M ( a ) \mathsf{L}_M(a) L M ( a ) be the set of lengths of its factorizations into atoms, and set L M ( a ) = { 0 } \mathsf{L}_M(a) = \{0\} L M ( a ) = { 0 } for a unit. The monoid is half-factorial if every such set is a singleton. A domain has the corresponding property when its multiplicative monoid of nonzero elements does. All rings are commutative with 1 ≠ 0 1 \ne0 1 = 0 , and ring homomorphisms preserve the identity.
For a group G G G and S ⊆ G S \subseteq G S ⊆ G , write ⟨ S ⟩ \langle S\rangle ⟨ S ⟩ for the generated subgroup and [ S ] [S] [ S ] for the generated submonoid, including the empty sum. Thus [ S ] = G [S] = G [ S ] = G requires nonnegative coefficients. If G G G is torsion, then [ S ] = ⟨ S ⟩ [S] = \langle S\rangle [ S ] = ⟨ S ⟩ . Write C n = Z / n Z C_n = \mathbb{Z}/n\mathbb{Z} C n = Z / n Z and identify C p ∞ C_{p^\infty} C p ∞ with the p p p -primary subgroup of Q / Z \mathbb{Q}/\mathbb{Z} Q / Z . An abelian p p p -group is a group all of whose elements have finite p p p -power order. For m ≥ 1 m \ge1 m ≥ 1 , put G [ m ] = { g ∈ G : m g = 0 } G[m] = \{g \in G : mg = 0\} G [ m ] = { g ∈ G : m g = 0 } . The subgroup P [ p ] P[p] P [ p ] of a p p p -group is its socle , an F p \mathbb{F}_p F p -vector space. A group is divisible if multiplication by every positive integer is surjective, and reduced if it contains no nonzero divisible subgroup. A subgroup H ⊆ G H \subseteq G H ⊆ G is pure if H ∩ m G = m H H \cap mG = mH H ∩ m G = m H for every m ≥ 1 m \ge1 m ≥ 1 . A p p p -group is bounded if p N P = 0 p^N P = 0 p N P = 0 for some N N N , and unbounded otherwise. The axiom of choice is used for bases and cardinal comparisons.
Let F ( S ) \mathcal{F}(S) F ( S ) be the free commutative monoid on S S S . Its members are finite sequences, written B = ∏ s ∈ S s v s ( B ) B = \prod_{s \in S} s^{v_s(B)} B = ∏ s ∈ S s v s ( B ) with finite support. The sequence length and sum are ∣ B ∣ = ∑ s ∈ S v s ( B ) |B| = \sum_{s \in S} v_s(B) ∣ B ∣ = ∑ s ∈ S v s ( B ) and σ ( B ) = ∑ s ∈ S v s ( B ) s \sigma(B) = \sum_{s \in S} v_s(B)s σ ( B ) = ∑ s ∈ S v s ( B ) s . A subsequence means a divisor in F ( S ) \mathcal{F}(S) F ( S ) , including multiplicities. A zero-sum sequence over an additive abelian group G G G is a finite sequence of elements of G G G , with repetitions allowed and order disregarded, whose sum is 0 0 0 . It is minimal if it is nonempty and no proper nonempty subsequence has sum 0 0 0 . The block monoid is
B ( S ) = { B ∈ F ( S ) : σ ( B ) = 0 } . \mathcal{B}(S) = \{B \in\mathcal{F}(S) : \sigma(B) = 0\}. B ( S ) = { B ∈ F ( S ) : σ ( B ) = 0 } .
Its identity is the empty sequence. Its atoms are exactly the minimal nonempty zero-sum sequences; induction on ∣ B ∣ |B| ∣ B ∣ proves atomicity. A subset S S S is half-factorial if B ( S ) \mathcal{B}(S) B ( S ) is half-factorial. We write H \mathcal{H} H for the class of groups possessing a half-factorial subset that generates them as a monoid. The empty set is allowed for the trivial group.
Adding or deleting the letter zero does not affect generation or half-factoriality: if 0 ∉ S 0 \notin S 0 ∈ / S , then B ( S ∪ { 0 } ) ≅ B ( S ) × N 0 \mathcal{B}(S \cup\{0\}) \cong\mathcal{B}(S) \times\mathbb{N}_0 B ( S ∪ { 0 }) ≅ B ( S ) × N 0 . Unless stated otherwise, generating sets omit zero. We use the usual convention ord ( 0 ) = 1 \operatorname{ord}(0) = 1 ord ( 0 ) = 1 when zero occurs in a sequence. Half-factoriality has finite character: a set is half-factorial precisely when each of its finite subsets is half-factorial [7 Lemma 3.1] . In particular, subsets and unions of nested chains of half-factorial sets are half-factorial. Indeed, any failure is witnessed by two factorizations using only finitely many letters. This observation concerns chains of sets, not arbitrary increasing unions of the groups they generate.
The realization and cross-number criteria
A Dedekind domain is a Noetherian integrally closed domain of Krull dimension at most one. Its ideal class group is the group of nonzero fractional ideals modulo nonzero principal fractional ideals. Let H H H be a commutative cancellative monoid and let F = F ( P ) F = \mathcal{F}(P) F = F ( P ) be the free commutative monoid on a set P P P . A monoid homomorphism φ : H → F \varphi: H \to F φ : H → F is a divisor homomorphism if, for all a , b ∈ H a, b \in H a , b ∈ H , φ ( a ) ∣ φ ( b ) \varphi(a) \mid\varphi(b) φ ( a ) ∣ φ ( b ) implies a ∣ H b a \mid_H b a ∣ H b . It is a divisor theory if, in addition, for every f ∈ F f \in F f ∈ F there exist m ≥ 1 m \ge1 m ≥ 1 and a 1 , … , a m ∈ H a_1, \ldots, a_m \in H a 1 , … , a m ∈ H such that
f = gcd F ( φ ( a 1 ) , … , φ ( a m ) ) . f = \gcd_F(\varphi(a_1), \ldots, \varphi(a_m)). f = F g cd ( φ ( a 1 ) , … , φ ( a m )) .
Here greatest common divisors in F F F are computed by taking the minimum exponent of each generator in P P P . A commutative cancellative monoid is a Krull monoid if it admits a divisor theory. For a divisor theory φ : H → F \varphi: H \to F φ : H → F , its divisor class group is
C ( H ) = q ( F ) / q ( φ ( H ) ) . C(H)=\mathsf{q}(F)/\mathsf{q}(\varphi(H)). C ( H ) = q ( F ) / q ( φ ( H )) .
This group is independent, up to isomorphism, of the chosen divisor theory; see [9 ] . We recall the realization theorem in the form needed here. The existence statement is classical [11 , 2 ] ; its formulation in terms of generating subsets is also recorded in [7 p. 598] and [8 p. 842] .
Theorem 2.1 (Classical realization criterion). Let G G G be an abelian group. The following are equivalent:
G G G has a half-factorial subset S S S with [ S ] = G [S]=G [ S ] = G ;
G G G is isomorphic to the ideal class group of a half-factorial Dedekind domain;
G G G is isomorphic to the divisor class group of a half-factorial Krull monoid.
More precisely, for every nonempty S ⊆ G S\subseteq G S ⊆ G with [ S ] = G [S]=G [ S ] = G , there is a Dedekind domain whose class group is identified with G G G and whose classes containing nonzero prime ideals are exactly S S S . Its factorization lengths are transferred to B ( S ) \mathcal{B}(S) B ( S ) , so it is half-factorial precisely when S S S is half-factorial.
Explanation of the equivalence. The prescribed prime-class realization theorem supplies the domain. If ( a ) (a) ( a ) factors into prime ideals, replace each prime ideal by its class to obtain a block over S S S . A zero-sum subsequence corresponds to a principal subproduct; its generator gives the corresponding element factor. Partitioning the prime-ideal occurrences lifts factorizations, including their lengths. Conversely, irreducible elements give minimal zero-sum subsequences under this transfer. These are the standard block-transfer properties [2 , 9 ] .
The same argument applies to a divisor theory of a Krull monoid. The classes of its prime divisors generate the class group as a monoid, and the block transfer preserves sets of lengths. Thus (3) implies (1), while (1) implies (2) by realization and (2) implies (3) by the divisor theory of a Dedekind domain. The trivial group may be realized by k [ X ] k[X] k [ X ] for any field k k k . □ \square □
For a subset T T T of a torsion group and B ∈ F ( T ) B\in\mathcal{F}(T) B ∈ F ( T ) , define its cross number by
k ( B ) = ∑ t ∈ T v t ( B ) ord ( t ) . \mathsf{k}(B)=\sum_{t\in T}\frac{\mathsf{v}_t(B)}{\operatorname{ord}(t)}. k ( B ) = t ∈ T ∑ ord ( t ) v t ( B ) .
The following criterion is due to Skula and Śliwa; see [7 Lemma 3.2] and [14 Theorem 4.1] . We include its short proof because it is the only numerical factorization criterion used in the nonrealization argument.
Proposition 2.2 (Cross-number criterion). Let G G G be a torsion abelian group and let T ⊆ G T\subseteq G T ⊆ G , with ord ( 0 ) = 1 \operatorname{ord}(0)=1 ord ( 0 ) = 1 if 0 ∈ T 0\in T 0 ∈ T . Then T T T is half-factorial if and only if k ( U ) = 1 \mathsf{k}(U)=1 k ( U ) = 1 for every U ∈ A ( B ( T ) ) U\in\mathcal{A}(\mathcal{B}(T)) U ∈ A ( B ( T )) .
Proof. If every atom has cross number one, additivity of k \mathsf{k} k shows that all factorizations of a fixed block have the same length. Conversely, assume half-factoriality and let U U U be an atom. Choose a positive integer m m m divisible by the orders of all elements in its support. The block U m U^m U m has a factorization of length m m m using U U U , and a factorization using the pure atoms t ord ( t ) t^{\operatorname{ord}(t)} t ord ( t ) of length ∑ t m v t ( U ) / ord ( t ) = m k ( U ) \sum_t m\mathsf{v}_t(U)/\operatorname{ord}(t)=m\mathsf{k}(U) ∑ t m v t ( U ) / ord ( t ) = m k ( U ) . Equality of these lengths gives k ( U ) = 1 \mathsf{k}(U)=1 k ( U ) = 1 . □ \square □
Labels and the intrinsic class group
The quotient of the ambient free group by the group of blocks need not be the intrinsic divisor class group of an unlabelled block monoid. The distinction is relevant when applying a realization theorem; compare the divisor-theoretic discussion in [15 §2] .
Example 2.3. Let G = C n = ⟨ g ⟩ G=C_n=\langle g\rangle G = C n = ⟨ g ⟩ , with n ≥ 2 n\ge2 n ≥ 2 , and let S = { g } S=\{g\} S = { g } . Then [ S ] = G [S]=G [ S ] = G and B ( S ) = { g n k : k ∈ N 0 } \mathcal{B}(S)=\{g^{nk}:k\in\mathbb{N}_0\} B ( S ) = { g nk : k ∈ N 0 } is a free monoid on one generator. Its intrinsic divisor class group is therefore zero. On the other hand, q ( F ( S ) ) / q ( B ( S ) ) ≅ Z / n Z \mathsf{q}(\mathcal{F}(S))/\mathsf{q}(\mathcal{B}(S))\cong\mathbb{Z}/n\mathbb{Z} q ( F ( S )) / q ( B ( S )) ≅ Z / n Z . The inclusion B ( S ) ⊂ F ( S ) \mathcal{B}(S)\subset\mathcal{F}(S) B ( S ) ⊂ F ( S ) is saturated, but is not a divisor theory. Indeed, g g g is not a greatest common divisor of finitely many members of B ( S ) \mathcal{B}(S) B ( S ) .
An inclusion H ⊆ F ( P ) H \subseteq\mathcal{F}(P) H ⊆ F ( P ) is saturated if H = q ( H ) ∩ F ( P ) H = \mathsf{q}(H) \cap\mathcal{F}(P) H = q ( H ) ∩ F ( P ) inside q ( F ( P ) ) \mathsf{q}(\mathcal{F}(P)) q ( F ( P )) ; equivalently, the inclusion is a divisor homomorphism. For such an inclusion to be a divisor theory, it suffices that every p ∈ P p \in P p ∈ P be the greatest common divisor of finitely many elements of H H H . For such an inclusion the divisor class group is q ( F ( P ) ) / q ( H ) \mathsf{q}(\mathcal{F}(P))/\mathsf{q}(H) q ( F ( P )) / q ( H ) .
Proposition 2.4. Let G G G be an abelian group and let S ⊆ G S \subseteq G S ⊆ G satisfy [ S ] = G [S] = G [ S ] = G . Set P = S × N P = S \times\mathbb{N} P = S × N , let β : F ( P ) → G \beta: \mathcal{F}(P) \to G β : F ( P ) → G send ( s , j ) (s,j) ( s , j ) to s s s , and put H = β − 1 ( 0 ) H = \beta^{-1}(0) H = β − 1 ( 0 ) . Then H ⊆ F ( P ) H \subseteq\mathcal{F}(P) H ⊆ F ( P ) is a divisor theory with class group G G G . Forgetting the second coordinate gives a surjective homomorphism H → B ( S ) H \to\mathcal{B}(S) H → B ( S ) that preserves all sets of factorization lengths. In particular, H H H is half-factorial if and only if S S S is half-factorial.
Proof. The inclusion is saturated because the quotient of two blocks, when it is a sequence, still has sum zero. Fix a label p = ( s , j ) p = (s,j) p = ( s , j ) . Since [ S ] = G [S] = G [ S ] = G , choose a sequence with sum − s -s − s . Give its letters labels in two disjoint sets, both avoiding p p p , to obtain W 1 , W 2 W_1, W_2 W 1 , W 2 . Then p W 1 , p W 2 ∈ H pW_1, pW_2 \in H p W 1 , p W 2 ∈ H and their greatest common divisor in F ( P ) \mathcal{F}(P) F ( P ) is p p p . This also works for s = 0 s = 0 s = 0 , using the empty sequence. Thus the inclusion is a divisor theory.
Extend β \beta β to q ( F ( P ) ) \mathsf{q}(\mathcal{F}(P)) q ( F ( P )) . To verify its kernel explicitly, let z ∈ ker β z \in\ker\beta z ∈ ker β . Choose a sequence M M M clearing all negative coordinates of z z z , so that z M ∈ F ( P ) zM \in\mathcal{F}(P) z M ∈ F ( P ) . Choose a further sequence C C C with β ( C ) = − β ( M ) \beta(C) = -\beta(M) β ( C ) = − β ( M ) . Then M C , z M C ∈ H MC, zMC \in H M C , z M C ∈ H and z = ( z M C ) / ( M C ) ∈ q ( H ) z = (zMC)/(MC) \in\mathsf{q}(H) z = ( z M C ) / ( M C ) ∈ q ( H ) . Since β \beta β is surjective, the quotient group is G G G .
Every block over S S S can be labelled, so the forgetting map is surjective. A factorization of its image can be lifted by partitioning the occurrences of each class among the factors, retaining their original labels. Conversely, forgetting labels sends atoms to atoms: a proper zero-sum subsequence of the image would lift to a proper zero-sum subsequence of the original block. The map therefore preserves exactly the factorization lengths. □
The labelled construction identifies the divisor class group and the length-preserving map explicitly. Its Dedekind-domain interpretation uses Theorem 2.1 ; no identification of an arbitrary Krull monoid with a domain’s reduced multiplicative monoid is needed.
Elementary positive constructions
Direct sums provide the elementary positive examples used later. The block decomposition is standard [8 Lemma 3.5] ; it is a restricted product, so each sequence involves only finitely many components.
Lemma 2.5. Let ( G j ) j ∈ J (G_j)_{j \in J} ( G j ) j ∈ J be abelian groups and suppose that T j ⊆ G j ∖ { 0 } T_j \subseteq G_j \setminus\{0\} T j ⊆ G j ∖ { 0 } is half-factorial with [ T j ] = G j [T_j] = G_j [ T j ] = G j . Embed the T j T_j T j in G = ⨁ j ∈ J G j G = \bigoplus_{j \in J} G_j G = ⨁ j ∈ J G j and put T = ⋃ j ∈ J T j T = \bigcup_{j \in J} T_j T = ⋃ j ∈ J T j . Then [ T ] = G [T] = G [ T ] = G and B ( T ) ≅ ⨁ j ∈ J B ( T j ) \mathcal{B}(T) \cong\bigoplus_{j \in J} \mathcal{B}(T_j) B ( T ) ≅ ⨁ j ∈ J B ( T j ) , where the right side is the restricted direct product of monoids. In particular, H \mathcal{H} H is closed under arbitrary direct sums.
Proof. Both group elements and sequences have finite support. A sequence on T T T has sum zero precisely when its component in each T j T_j T j has sum zero. Consequently each atom belongs to a single component, and every factorization length is the sum of the uniquely determined component lengths. □
The elementary cyclic examples already give all direct sums of cyclic groups. For a finite cyclic group generated by c c c of order n ≥ 2 n \ge2 n ≥ 2 , the set { c } \{c\} { c } has the single block atom c n c^n c n . For the infinite cyclic group generated by e e e , the set { e , − e } \{e,-e\} { e , − e } has the single block atom e ( − e ) e(-e) e ( − e ) . These familiar examples and Lemma 2.5 show that all direct sums of cyclic groups lie in H \mathcal{H} H ; see also [7 p. 598] .
The divisible primary example is equally explicit. The existence assertion below is classical; compare [7 Corollary 5.4] . We include the short argument because it supplies the divisible part in the reduction theorem.
Proposition 2.6. For every prime p p p , the Prüfer group C p ∞ C_{p^\infty} C p ∞ belongs to H \mathcal{H} H . More precisely, identifying it with the p p p -primary subgroup of Q / Z \mathbb{Q}/\mathbb{Z} Q / Z , the set T = { p − n + Z : n ≥ 1 } T = \{p^{-n} + \mathbb{Z} : n \ge1\} T = { p − n + Z : n ≥ 1 } is monoid-generating and half-factorial.
Proof. Put c n = p − n + Z c_n=p^{-n}+\mathbb{Z} c n = p − n + Z . The c n c_n c n generate the group; because it is torsion, they also generate it as a monoid. Give c n c_n c n weight p − n p^{-n} p − n . A sequence is a block exactly when its total weight is an integer.
A finite collection of these weights having positive integer sum m m m can be partitioned into m m m collections of weight one. Indeed, let N N N be the largest occurring denominator exponent. For N ≥ 1 N \ge1 N ≥ 1 , the number of occurrences of p − N p^{-N} p − N is divisible by p p p : multiply the total weight by p N p^N p N and reduce modulo p p p . Bundle these occurrences into packets of p p p and regard each packet as a weight p − ( N − 1 ) p^{-(N-1)} p − ( N − 1 ) . Repeat. This process ends with m m m packets of weight one, each recording a disjoint subsequence of the original sequence.
It follows that every block of weight greater than one is reducible. Conversely, a block of weight one has no proper nonempty block subsequence, because any such subsequence has positive integral weight. Thus the atoms have weight exactly one, and the total weight of a block is the length of each of its factorizations. □
We will also use the established countable case: every countable abelian group belongs to H \mathcal{H} H [8 Theorem 5.3, with $F = 0] . Every bounded torsion abelian group belongs to H \mathcal{H} H as well, because it is a direct sum of finite cyclic groups [5 ] . These positive cases concern existence of a suitable generating set, not the extendibility of an arbitrarily prescribed one.
Separation and a cardinality obstruction
We first construct a finite zero-sum obstruction from coincident quotient images. We then select a suitable direct summand from an arbitrary finite generating set and derive the separation theorem. The resulting cardinality bound excludes the torsion-completion and its uncountable pure subgroups. The proof uses neither the later lifting results nor a hypothesis about generators of the socle.
Collisions and a finite separation subgroup
The first observation turns two order-preserving representatives of one quotient element into an atom of cross number greater than one. No half-factoriality hypothesis is needed to construct this atom.
Lemma 3.1. Let G G G be a torsion abelian group, let S ⊆ G ∖ { 0 } S \subseteq G \setminus\{0\} S ⊆ G ∖ { 0 } , and let A ⊆ S A \subseteq S A ⊆ S . Put K = ⟨ A ⟩ K = \langle A\rangle K = ⟨ A ⟩ . Suppose that distinct s , t ∈ S ∖ K s,t \in S \setminus K s , t ∈ S ∖ K satisfy
s + K = t + K , ord G ( s ) = ord G ( t ) = ord G / K ( s + K ) = q . (1) s + K = t + K,\qquad\operatorname{ord}_{G}(s) = \operatorname{ord}_{G}(t) = \operatorname{ord}_{G/K}(s + K) = q. \tag*{(1)} s + K = t + K , ord G ( s ) = ord G ( t ) = ord G / K ( s + K ) = q . (1) Then there is a nonempty sequence C ∈ F ( A ) C \in\mathcal{F}(A) C ∈ F ( A ) such that U = s t q − 1 C U = st^{q-1}C U = s t q − 1 C is an atom of B ( S ) \mathcal{B}(S) B ( S ) . In particular, k ( U ) = 1 + k ( C ) > 1 \mathsf{k}(U) = 1 + \mathsf{k}(C) > 1 k ( U ) = 1 + k ( C ) > 1 , so S S S is not half-factorial.
Proof. Since G G G is torsion, A A A generates K K K as a monoid. Choose a sequence C C C over A A A of least possible sequence length subject to σ ( C ) = t − s \sigma(C) = t-s σ ( C ) = t − s . It is nonempty because s ≠ t s \ne t s = t . The sequence U = s t q − 1 C U = st^{q-1}C U = s t q − 1 C has sum zero.
Let V V V be a nonempty zero-sum subsequence of U U U . The letters from C C C vanish in G / K G/K G / K , while all q q q occurrences in s t q − 1 st^{q-1} s t q − 1 have the same image of order q q q . Thus V V V contains either none or all of these q q q occurrences. In the first case V V V is a nonempty block subsequence of C C C ; deleting it gives a shorter sequence with sum t − s t-s t − s , contrary to the choice of C C C . In the second case V = s t q − 1 C ′ V = st^{q-1}C' V = s t q − 1 C ′ for a subsequence C ′ C' C ′ of C C C , and σ ( C ′ ) = t − s \sigma(C') = t-s σ ( C ′ ) = t − s . Minimality forces C ′ = C C' = C C ′ = C . Therefore U U U is an atom. Its first q q q letters have cross number one by (3.1), and C C C has positive cross number. Apply Proposition 2.2. □
The next lemma is purely group-theoretic. Its generating subset is selected from the given set, a point essential to the subsequent application of Lemma 3.1.
Lemma 3.2. Let p p p be a prime, let G G G be a finite abelian p p p -group, let S ⊆ G S \subseteq G S ⊆ G generate G G G , and let r ≥ 1 r \ge1 r ≥ 1 . There is a subset A ⊆ S A \subseteq S A ⊆ S such that, for K = ⟨ A ⟩ K = \langle A\rangle K = ⟨ A ⟩ ,
p r G ⊆ K , K [ p ] ⊆ p r G . (2) p^rG \subseteq K,\qquad K[p] \subseteq p^rG. \tag*{(2)} p r G ⊆ K , K [ p ] ⊆ p r G . (2)
In fact, K K K can be chosen as a direct summand isomorphic to the sum of the cyclic factors of G G G whose orders exceed p r p^r p r . Here K [ p ] = { x ∈ K : p x = 0 } K[p] = \{x \in K : px = 0\} K [ p ] = { x ∈ K : p x = 0 } .
Proof. Step 1: separate the cyclic factors by exponent. By the finite abelian-group structure theorem [5 ] , write
G = L ⊕ H , p r L = 0 , H = ⨁ j = 1 d ⟨ h j 0 ⟩ , ord ( h j 0 ) = p a j , a j ≥ r + 1. G = L \oplus H,\qquad p^rL = 0,\qquad H = \bigoplus_{j=1}^{d}\langle h_j^0\rangle,\quad\operatorname{ord}(h_j^0) = p^{a_j},\quad a_j \ge r+1. G = L ⊕ H , p r L = 0 , H = j = 1 ⨁ d ⟨ h j 0 ⟩ , ord ( h j 0 ) = p a j , a j ≥ r + 1.
If d = 0 d = 0 d = 0 , take A = ∅ A = \varnothing A = ∅ and K = 0 K = 0 K = 0 . Suppose d > 0 d > 0 d > 0 . The projections of S S S generate H H H , so choose distinct u 1 , … , u d ∈ S u_1,\ldots,u_d \in S u 1 , … , u d ∈ S whose projections form a basis of H / p H H/pH H / p H . Write u i = ℓ i + h i u_i = \ell_i + h_i u i = ℓ i + h i with ℓ i ∈ L \ell_i \in L ℓ i ∈ L and h i ∈ H h_i \in H h i ∈ H .
Step 2: show that the selected subgroup is a graph. The elements h i h_i h i generate H H H . Indeed, for H 0 = ⟨ h 1 , … , h d ⟩ H_0 = \langle h_1,\ldots,h_d\rangle H 0 = ⟨ h 1 , … , h d ⟩ one has H = H 0 + p H H = H_0 + pH H = H 0 + p H ; iteration and the finite exponent of H H H give H = H 0 H = H_0 H = H 0 . Moreover,
H / p r + 1 H ≅ ( Z / p r + 1 Z ) d . H/p^{r+1}H \cong(\mathbb{Z}/p^{r+1}\mathbb{Z})^d. H / p r + 1 H ≅ ( Z / p r + 1 Z ) d .
The images of the h i h_i h i constitute a basis of this module: they generate it, and the resulting surjection from ( Z / p r + 1 Z ) d (\mathbb{Z}/p^{r+1}\mathbb{Z})^d ( Z / p r + 1 Z ) d is an isomorphism, since source and codomain have the same finite cardinality. Consequently, every integer relation ∑ i c i h i = 0 \sum_i c_i h_i = 0 ∑ i c i h i = 0 satisfies p r + 1 ∣ c i p^{r+1} \mid c_i p r + 1 ∣ c i for every i i i . Since p r L = 0 p^rL = 0 p r L = 0 , the same coefficients give ∑ i c i ℓ i = 0 \sum_i c_i\ell_i = 0 ∑ i c i ℓ i = 0 .
It follows that the assignments h i ↦ ℓ i h_i \mapsto\ell_i h i ↦ ℓ i define a homomorphism ψ : H → L \psi: H \to L ψ : H → L . The subgroup K = ⟨ u 1 , … , u d ⟩ K = \langle u_1,\ldots,u_d\rangle K = ⟨ u 1 , … , u d ⟩ is its graph:
K = { ψ ( h ) + h : h ∈ H } , G = L ⊕ K . K = \{\psi(h) + h : h \in H\},\qquad G = L \oplus K. K = { ψ ( h ) + h : h ∈ H } , G = L ⊕ K .
Step 3: verify the two power conditions. For every h ∈ H h \in H h ∈ H , p r ψ ( h ) = 0 p^r\psi(h) = 0 p r ψ ( h ) = 0 , so p r K = p r H = p r G p^rK = p^rH = p^rG p r K = p r H = p r G , proving the first inclusion in (2) . Also H [ p ] ⊆ p r H H[p] \subseteq p^rH H [ p ] ⊆ p r H , because all cyclic factors of H H H have order at least p r + 1 p^{r+1} p r + 1 . The homomorphism ψ \psi ψ vanishes on p r H p^rH p r H , and hence on H [ p ] H[p] H [ p ] . Thus K [ p ] = H [ p ] ⊆ p r G K[p] = H[p] \subseteq p^rG K [ p ] = H [ p ] ⊆ p r G , with H H H regarded as the chosen summand of G G G . Take A = { u 1 , … , u d } A = \{u_1,\ldots,u_d\} A = { u 1 , … , u d } . □
Separation of arbitrary half-factorial generators
Only finitely many generators are needed to witness a relation s − t ∈ p r P s-t \in p^rP s − t ∈ p r P . This permits the finite lemma to be used in groups of arbitrary cardinality, without a pure-subgroup assumption.
Theorem 3.3. Let p p p be a prime, let P P P be an abelian p p p -group, and let S ⊆ P ∖ { 0 } S \subseteq P \setminus\{0\} S ⊆ P ∖ { 0 } be a half-factorial generating set. Suppose that s , t ∈ S s,t \in S s , t ∈ S and n , r ≥ 1 n,r \ge1 n , r ≥ 1 satisfy
ord ( s ) = ord ( t ) = p n , s − t ∈ p r P , p n − 1 s ∉ p r P . (3) \operatorname{ord}(s) = \operatorname{ord}(t) = p^n,\qquad s-t \in p^rP,\qquad p^{n-1}s \notin p^rP. \tag*{(3)} ord ( s ) = ord ( t ) = p n , s − t ∈ p r P , p n − 1 s ∈ / p r P . (3) Then s = t s = t s = t .
Proof. Step 1: choose a finite witness for divisibility. Suppose that s ≠ t s \ne t s = t . Choose y ∈ P y \in P y ∈ P such that s − t = p r y s-t = p^r y s − t = p r y . Since S S S generates P P P , there is a finite subset S 0 ⊆ S S_0 \subseteq S S 0 ⊆ S containing s , t s,t s , t with y ∈ G 0 = ⟨ S 0 ⟩ y \in G_0 = \langle S_0\rangle y ∈ G 0 = ⟨ S 0 ⟩ . The group G 0 G_0 G 0 is a finite abelian p p p -group. Orders of its elements are the same as in P P P , and
s − t ∈ p r G 0 , p n − 1 s ∉ p r G 0 . s-t \in p^rG_0,\qquad p^{n-1}s \notin p^rG_0. s − t ∈ p r G 0 , p n − 1 s ∈ / p r G 0 . Step 2: preserve the orders in a quotient. Apply Lemma 3.2 to G 0 G_0 G 0 , S 0 S_0 S 0 , r r r , obtaining A ⊆ S 0 A \subseteq S_0 A ⊆ S 0 and K = ⟨ A ⟩ K = \langle A\rangle K = ⟨ A ⟩ with p r G 0 ⊆ K p^rG_0 \subseteq K p r G 0 ⊆ K and K [ p ] ⊆ p r G 0 K[p] \subseteq p^rG_0 K [ p ] ⊆ p r G 0 .
The element p n − 1 s p^{n-1}s p n − 1 s has order p p p and does not belong to K K K ; otherwise it would lie in K [ p ] ⊆ p r G 0 K[p] \subseteq p^rG_0 K [ p ] ⊆ p r G 0 . Every nonzero subgroup of the cyclic group ⟨ s ⟩ \langle s\rangle ⟨ s ⟩ contains p n − 1 s p^{n-1}s p n − 1 s . Therefore ⟨ s ⟩ ∩ K = 0 \langle s\rangle\cap K = 0 ⟨ s ⟩ ∩ K = 0 , and s + K s+K s + K has order p n p^n p n in G 0 / K G_0/K G 0 / K . Since s − t ∈ p r G 0 ⊆ K s-t \in p^rG_0 \subseteq K s − t ∈ p r G 0 ⊆ K , the elements s , t s,t s , t have the same nonzero image in this quotient.
Step 3: produce a large-cross-number atom. Lemma 3.1 gives an atom over S 0 S_0 S 0 of cross number greater than one. It remains an atom over S S S , because its subsequences have the same sums in G 0 G_0 G 0 and P P P . This contradicts Proposition 2.2. □
For an abelian p p p -group P P P , the p p p -adic topology is the group topology having { p r P : r ≥ 1 } \{p^rP:r\ge1\} { p r P : r ≥ 1 } as a neighbourhood basis at zero; see [4 Example 4.1(a)] . We call P P P p p p -adically separated if ⋂ r ≥ 1 p r P = 0 \bigcap_{r\ge1}p^rP=0 ⋂ r ≥ 1 p r P = 0 , or equivalently, if its p p p -adic topology is Hausdorff. The separation theorem bounds the number of generators at each order and divisibility level. Here and below, cardinal inequalities are taken in the usual set-theoretic sense.
Theorem 3.4. Let p p p be a prime and let P P P be a p p p -adically separated abelian p p p -group. If P P P has a half-factorial generating subset, then
∣ P ∣ ≤ max { ℵ 0 , ∣ P / p P ∣ } . (4) |P| \le\max\{\aleph_0, |P/pP|\}. \tag*{(4)} ∣ P ∣ ≤ max { ℵ 0 , ∣ P / pP ∣ } . (4) In particular, a p p p -adically separated abelian p p p -group with countable P / p P P/pP P / pP can admit a half-factorial generating subset only if P P P is countable.
Proof. Let S ⊆ P ∖ { 0 } S\subseteq P\setminus\{0\} S ⊆ P ∖ { 0 } be a half-factorial generating set. For positive integers n , r n,r n , r , put
S n , r = { s ∈ S : ord ( s ) = p n and p n − 1 s ∉ p r P } . S_{n,r}=\{s\in S:\operatorname{ord}(s)=p^n\text{ and }p^{n-1}s\notin p^rP\}. S n , r = { s ∈ S : ord ( s ) = p n and p n − 1 s ∈ / p r P } .
The natural map S n , r → P / p r P S_{n,r}\to P/p^rP S n , r → P / p r P is injective by Theorem 3.3 . Every s ∈ S s\in S s ∈ S belongs to some S n , r S_{n,r} S n , r : if ord ( s ) = p n \operatorname{ord}(s)=p^n ord ( s ) = p n , its nonzero multiple p n − 1 s p^{n-1}s p n − 1 s cannot belong to all p r P p^rP p r P . Thus
S = ⋃ n , r ≥ 1 S n , r , ∣ S n , r ∣ ≤ ∣ P / p r P ∣ . (5) S=\bigcup_{n,r\ge1}S_{n,r},\qquad|S_{n,r}|\le|P/p^rP|. \tag*{(5)} S = n , r ≥ 1 ⋃ S n , r , ∣ S n , r ∣ ≤ ∣ P / p r P ∣. (5)
Set κ = max { ℵ 0 , ∣ P / p P ∣ } \kappa=\max\{\aleph_0,|P/pP|\} κ = max { ℵ 0 , ∣ P / pP ∣ } . Multiplication by p j p^j p j induces a surjection P / p P → p j P / p j + 1 P P/pP\to p^jP/p^{j+1}P P / pP → p j P / p j + 1 P for every j ≥ 0 j\ge0 j ≥ 0 . The finite filtration of P / p r P P/p^rP P / p r P with these successive quotients therefore gives ∣ P / p r P ∣ ≤ κ |P/p^rP|\le\kappa ∣ P / p r P ∣ ≤ κ . (5) yields ∣ S ∣ ≤ κ |S|\le\kappa ∣ S ∣ ≤ κ , and the group generated by at most κ \kappa κ elements has cardinality at most κ \kappa κ . This proves (3.4). □
A countable dense subset meets every open coset, so the same obstruction has a topological formulation.
Corollary 3.5. Let p p p be a prime and let P P P be an abelian p p p -group whose p p p -adic topology is Hausdorff and has a countable dense subset. If P P P admits a half-factorial generating subset, then P P P is countable.
Proof. Every coset of p P pP pP is open and must meet a countable dense subset. Hence P / p P P/pP P / pP is countable. Apply Theorem 3.4 . □
The torsion-completion counterexample
The following group is the explicit test case proposed in [8 p. 855] . We verify directly all group properties needed for the obstruction. Its identification as the torsion-completion of ⨁ j ≥ 1 C p j \bigoplus_{j\ge1}C_{p^j} ⨁ j ≥ 1 C p j is also discussed in Proposition 4.10 below.
Theorem 3.6. For every prime p p p , let
P p = t ( ∏ j ≥ 1 Z / p j Z ) = { ( x j ) j : p m x j = 0 for all j and some m ≥ 0 } , (6) P_p=t\left(\prod_{j\ge1}\mathbb{Z}/p^j\mathbb{Z}\right)=\{(x_j)_j:p^m x_j=0\text{ for all }j\text{ and some }m\ge0\}, \tag*{(6)} P p = t ( j ≥ 1 ∏ Z / p j Z ) = {( x j ) j : p m x j = 0 for all j and some m ≥ 0 } , (6) where t ( − ) t(-) t ( − ) denotes the torsion subgroup. Then P p P_p P p is a reduced, unbounded abelian p p p -group of cardinality 2 ℵ 0 2^{\aleph_0} 2 ℵ 0 and
⋂ r ≥ 1 p r P p = 0 , P p / p r P p ≅ ⨁ j ≥ 1 Z / p min ( j , r ) Z ( r ≥ 1 ) . (7) \bigcap_{r\ge1}p^rP_p=0,\qquad P_p/p^rP_p\cong\bigoplus_{j\ge1}\mathbb{Z}/p^{\min(j,r)}\mathbb{Z}\quad(r\ge1). \tag*{(7)} r ≥ 1 ⋂ p r P p = 0 , P p / p r P p ≅ j ≥ 1 ⨁ Z / p m i n ( j , r ) Z ( r ≥ 1 ) . (7) It has no half-factorial generating subset. Consequently P p P_p P p is not isomorphic to the ideal class group of any half-factorial Dedekind domain, so the universal half-factorial Dedekind realization problem has a negative answer. The group P p P_p P p is not the divisor class group of a half-factorial Krull monoid either.
Proof. Step 1: cardinality and separation. Write C j = Z / p j Z C_j = \mathbb{Z}/p^j\mathbb{Z} C j = Z / p j Z . The description in (6) shows that P p P_p P p is a p p p -group. It contains ⨁ j ≥ 1 C j \bigoplus_{j\geq1} C_j ⨁ j ≥ 1 C j , so its exponent is unbounded. Its order-p p p subgroup is
P p [ p ] = ∏ j ≥ 1 C j [ p ] ≅ ∏ j ≥ 1 Z / p Z . P_p[p] = \prod_{j\geq1} C_j[p] \cong\prod_{j\geq1} \mathbb{Z}/p\mathbb{Z}. P p [ p ] = j ≥ 1 ∏ C j [ p ] ≅ j ≥ 1 ∏ Z / p Z .
This subgroup has cardinality 2 ℵ 0 2^{\aleph_0} 2 ℵ 0 , and the ambient countable product of finite groups has the same cardinality. Thus ∣ P p ∣ = 2 ℵ 0 \lvert P_p\rvert= 2^{\aleph_0} ∣ P p ∣ = 2 ℵ 0 . If x ∈ ⋂ r ≥ 1 p r P p x \in\bigcap_{r\geq1} p^rP_p x ∈ ⋂ r ≥ 1 p r P p , then for each j j j the j j j th coordinate of x x x lies in p j C j = 0 p^jC_j = 0 p j C j = 0 . Hence x = 0 x = 0 x = 0 . This also proves reducedness: every divisible subgroup is contained in every p r P p p^rP_p p r P p .
Step 2: compute all finite power quotients. Fix r ≥ 1 r \geq1 r ≥ 1 . Coordinate reduction induces a homomorphism
ρ r : P p ⟶ ∏ j ≥ 1 C j / p r C j . \rho_r : P_p \longrightarrow\prod_{j\geq1} C_j/p^rC_j. ρ r : P p ⟶ j ≥ 1 ∏ C j / p r C j .
To see that its values have finite support, choose m m m with p m x = 0 p^m x = 0 p m x = 0 . For j ≥ m + r j \geq m+r j ≥ m + r , the coordinate x j x_j x j belongs to p j − m C j ⊆ p r C j p^{j-m}C_j \subseteq p^rC_j p j − m C j ⊆ p r C j . The map is surjective because finitely supported representatives belong to P p P_p P p . Clearly p r P p ⊆ ker ρ r p^rP_p \subseteq\ker\rho_r p r P p ⊆ ker ρ r . Conversely, suppose that x ∈ ker ρ r x \in\ker\rho_r x ∈ ker ρ r and p m x = 0 p^m x = 0 p m x = 0 . Choose y j ∈ C j y_j \in C_j y j ∈ C j with p r y j = x j p^r y_j = x_j p r y j = x j for each j j j . Then p m + r y j = 0 p^{m+r}y_j = 0 p m + r y j = 0 for all j j j , so y = ( y j ) j ∈ P p y = (y_j)_j \in P_p y = ( y j ) j ∈ P p and x = p r y x = p^r y x = p r y . Thus ker ρ r = p r P p \ker\rho_r = p^rP_p ker ρ r = p r P p . Since C j / p r C j ≅ Z / p min ( j , r ) Z C_j/p^rC_j \cong\mathbb{Z}/p^{\min(j,r)}\mathbb{Z} C j / p r C j ≅ Z / p m i n ( j , r ) Z , this proves (7) .
Step 3: apply the obstruction and realization criterion. In particular, P p / p P p P_p/pP_p P p / p P p is countable, while P p P_p P p is uncountable and p p p -adically separated. Theorem 3.4 excludes a half-factorial generating subset. The final assertion follows from the classical realization criterion, Theorem 2.1 . □
The argument can be strengthened from nonexistence to the production of a specific kind of finite obstruction inside every generating set.
Corollary 3.7. For every prime p p p and every generating subset S ⊆ P p ∖ { 0 } S \subseteq P_p \setminus\{0\} S ⊆ P p ∖ { 0 } , there is an atom of B ( S ) \mathcal{B}(S) B ( S ) whose cross number is greater than one. More precisely, there are distinct s , t ∈ S s,t \in S s , t ∈ S of a common order q q q and a nonempty sequence C C C over S ∖ { s , t } S \setminus\{s,t\} S ∖ { s , t } such that s t q − 1 C st^{q-1}C s t q − 1 C is an atom.
Proof. A generating set of the uncountable group P p P_p P p is uncountable. Partition it using the sets S n , r S_{n,r} S n , r in (5) , without assuming half-factoriality. Since every quotient P p / p r P p P_p/p^rP_p P p / p r P p is countable, some S n , r S_{n,r} S n , r contains distinct s , t s,t s , t with s − t ∈ p r P p s-t \in p^rP_p s − t ∈ p r P p . Choose a finite S 0 ⊆ S S_0 \subseteq S S 0 ⊆ S containing these elements and generating a p r p^r p r -root of s − t s-t s − t . Apply Lemma 3.2 to ⟨ S 0 ⟩ \langle S_0\rangle ⟨ S 0 ⟩ . Exactly as in the proof of Theorem 3.3 , the resulting subgroup K K K gives the hypotheses of Lemma 3.1 , with q = p n q = p^n q = p n . The correction sequence there lies in F ( A ) \mathcal{F}(A) F ( A ) with A ⊆ S 0 ∩ K A \subseteq S_0 \cap K A ⊆ S 0 ∩ K , whereas s , t ∉ K s,t \notin K s , t ∈ / K . The resulting atom is also an atom over S S S . □
Sharpness and smaller counterexamples
The known countable case gives a converse to the cardinal obstruction when the quotient modulo p p p is countable.
Corollary 3.8. Let p p p be a prime and let P P P be a p p p -adically separated abelian p p p -group with countable P / p P P/pP P / pP . Then P P P has a half-factorial generating subset if and only if P P P is countable.
Proof. Necessity follows from Theorem 3.4 . Sufficiency is the countable-group theorem of Geroldinger–Göbel [8 Theorem 5.3] . □
Both the size of the quotient and the separatedness hypothesis matter. The following examples make those boundaries explicit.
Example 3.9. Let p p p be a prime and κ \kappa κ an infinite cardinal. The elementary abelian group E = ⨁ κ C p E = \bigoplus_{\kappa} C_p E = ⨁ κ C p is p p p -adically separated and has a half-factorial generating set consisting of its cyclic basis elements. It satisfies ∣ E ∣ = ∣ E / p E ∣ = κ \lvert E\rvert= \lvert E/pE\rvert= \kappa ∣ E ∣ = ∣ E / pE ∣ = κ , so the bound in Theorem 3.4 is attained at every infinite cardinal.
In contrast, for uncountable κ \kappa κ , the divisible group D = ⨁ κ C p ∞ D = \bigoplus_{\kappa} C_{p^\infty} D = ⨁ κ C p ∞ belongs to H \mathcal{H} H by Lemma 2.5 and Proposition 2.6, but p D = D pD = D p D = D and ∣ D ∣ = κ > ℵ 0 |D| = \kappa> \aleph_0 ∣ D ∣ = κ > ℵ 0 . Thus the separatedness assumption cannot be omitted. The theorem does not claim that its cardinal bound is sufficient in general.
The continuum-sized example contains pure counterexamples of every smaller uncountable cardinality. We include the subgroup construction rather than assuming that half-factorial realizability is inherited by subgroups.
Corollary 3.10. Let p p p be a prime and let ℵ 1 ≤ κ ≤ 2 ℵ 0 \aleph_1 \le\kappa\le2^{\aleph_0} ℵ 1 ≤ κ ≤ 2 ℵ 0 . There is a pure subgroup H κ H_\kappa H κ of P p = t ( ∏ j ≥ 1 Z / p j Z ) P_p = t\left(\prod_{j\ge1} \mathbb{Z}/p^j\mathbb{Z}\right) P p = t ( ∏ j ≥ 1 Z / p j Z ) such that
∣ H κ ∣ = κ , H κ / p r H κ ≅ ⨁ j ≥ 1 Z / p min ( j , r ) Z ( r ≥ 1 ) . |H_\kappa| = \kappa,\qquad H_\kappa/p^r H_\kappa\cong\bigoplus_{j\ge1}\mathbb{Z}/p^{\min(j,r)}\mathbb{Z}\quad(r\ge1). ∣ H κ ∣ = κ , H κ / p r H κ ≅ j ≥ 1 ⨁ Z / p m i n ( j , r ) Z ( r ≥ 1 ) . The group H κ H_\kappa H κ is reduced, unbounded, and p p p -adically separated, and it has no half-factorial generating subset. Consequently the least cardinality of a group not realizable as the ideal class group of a half-factorial Dedekind domain is ℵ 1 \aleph_1 ℵ 1 .
Proof. Put B = ⨁ j ≥ 1 Z / p j Z ⊆ P p B = \bigoplus_{j\ge1}\mathbb{Z}/p^j\mathbb{Z} \subseteq P_p B = ⨁ j ≥ 1 Z / p j Z ⊆ P p and choose X ⊆ P p X \subseteq P_p X ⊆ P p of cardinality κ \kappa κ . Let H 0 = ⟨ B ∪ X ⟩ H_0 = \langle B \cup X\rangle H 0 = ⟨ B ∪ X ⟩ . Having constructed H i H_i H i , choose, for each h ∈ H i ∩ p r P p h \in H_i \cap p^rP_p h ∈ H i ∩ p r P p and r ≥ 1 r \ge1 r ≥ 1 , one element y h , r ∈ P p y_{h,r} \in P_p y h , r ∈ P p with p r y h , r = h p^r y_{h,r} = h p r y h , r = h , and let H i + 1 H_{i+1} H i + 1 be the subgroup generated by H i H_i H i and all these roots. Each H i H_i H i has cardinality κ \kappa κ . Hence H κ = ⋃ i ≥ 0 H i H_\kappa= \bigcup_{i\ge0} H_i H κ = ⋃ i ≥ 0 H i has cardinality κ \kappa κ and satisfies H κ ∩ p r P p = p r H κ H_\kappa\cap p^rP_p = p^rH_\kappa H κ ∩ p r P p = p r H κ for every r r r . Multiplication by integers prime to p p p is an automorphism of every p p p -group, so H κ H_\kappa H κ is pure in P p P_p P p .
Purity gives an injection H κ / p r H κ → P p / p r P p H_\kappa/p^rH_\kappa\to P_p/p^rP_p H κ / p r H κ → P p / p r P p . It is surjective because B ⊆ H κ B \subseteq H_\kappa B ⊆ H κ and the coordinate calculation in Theorem 3.6 shows that B B B maps onto P p / p r P p P_p/p^rP_p P p / p r P p . This proves the displayed quotient formula. Separatedness is inherited from P p P_p P p , and it implies reducedness; unboundedness follows from B ⊆ H κ B \subseteq H_\kappa B ⊆ H κ . Theorem 3.4 now excludes half-factorial generators. Finally, every countable abelian group belongs to H \mathcal{H} H , whereas H ℵ 1 H_{\aleph_1} H ℵ 1 does not. Apply Theorem 2.1. □ \square □
The nonrealization proof is now complete. It depends only on the classical criteria in Section 2, finite abelian-group structure, and the arguments in this section. The following sections study complementary phenomena and are not needed for this proof.
Lifting and extensions
We compare constructions that preserve half-factoriality with extensions that cannot do so. Free kernels admit an explicit block-monoid decomposition, while order-preserving lifts across torsion kernels impose a splitting condition. We also recall the classical primary reduction and show that the class H \mathcal{H} H is not closed under pure extensions.
An arbitrary free kernel
The essential observation is that lifts can be moved into a common nonnegative coordinate cone along a free kernel. The resulting correction terms are then additive and nonnegative. When the free subgroup has torsion quotient, the block-monoid decomposition below is [8 Proposition 3.7] . The proof given here allows a quotient with elements of infinite order as well.
Theorem 4.1. Let G G G be an abelian group and let F = ⨁ i ∈ I Z e i F = \bigoplus_{i\in I}\mathbb{Z}e_i F = ⨁ i ∈ I Z e i be a free subgroup. Write Q = G / F Q = G/F Q = G / F , and suppose that T ⊆ Q ∖ { 0 } T \subseteq Q\setminus\{0\} T ⊆ Q ∖ { 0 } satisfies [ T ] = Q [T] = Q [ T ] = Q . There is a subset S ⊆ G ∖ { 0 } S \subseteq G\setminus\{0\} S ⊆ G ∖ { 0 } with [ S ] = G [S] = G [ S ] = G and a monoid isomorphism
B ( T ) × N 0 ( I ) ≅ B ( S ) . (8) \mathcal{B}(T) \times\mathbb{N}_0^{(I)} \cong\mathcal{B}(S). \tag*{(8)} B ( T ) × N 0 ( I ) ≅ B ( S ) . (8) Here N 0 ( I ) \mathbb{N}_0^{(I)} N 0 ( I ) denotes the monoid of finitely supported nonnegative integer vectors, with coordinatewise addition. Consequently, G / F ∈ H G/F \in\mathcal{H} G / F ∈ H implies G ∈ H G \in\mathcal{H} G ∈ H . No purity or splitting hypothesis on F ⊆ G F \subseteq G F ⊆ G is required.
Proof. Step 1: choose nonnegative representatives. Since tensoring with Q \mathbb{Q} Q is exact, Q ⊗ Z F \mathbb{Q}\otimes_{\mathbb{Z}} F Q ⊗ Z F embeds in Q ⊗ Z G \mathbb{Q}\otimes_{\mathbb{Z}} G Q ⊗ Z G . Choose a Q \mathbb{Q} Q -linear projection onto this subspace. Composing the natural map from G G G with this projection gives a group homomorphism
λ : G ⟶ Q ( I ) , λ ( e i ) = ε i , \lambda:G\longrightarrow\mathbb{Q}^{(I)},\qquad\lambda(e_i)=\varepsilon_i, λ : G ⟶ Q ( I ) , λ ( e i ) = ε i ,
where ε i \varepsilon_i ε i is the i i i th coordinate vector. In particular, λ \lambda λ is injective on F F F and identifies F F F with Z ( I ) \mathbb{Z}^{(I)} Z ( I ) .
For each t ∈ T t\in T t ∈ T , choose a representative y t ∈ G y_t\in G y t ∈ G . All vectors λ ( y t ) \lambda(y_t) λ ( y t ) have finite support. Adding an appropriate integer combination of the e i e_i e i therefore produces a representative x t ∈ y t + F x_t\in y_t+F x t ∈ y t + F such that λ i ( x t ) ≥ 0 \lambda_i(x_t)\geq0 λ i ( x t ) ≥ 0 for every i i i . Define
S = { x t : t ∈ T } ∪ { e i , − e i : i ∈ I } . (9) S=\{x_t:t\in T\}\cup\{e_i,-e_i:i\in I\}. \tag*{(9)} S = { x t : t ∈ T } ∪ { e i , − e i : i ∈ I } . (9)
The three parts of this union are disjoint: the x t x_t x t lie in nonzero, pairwise distinct cosets of F F F , and the basis elements have infinite order. Projecting an arbitrary g ∈ G g\in G g ∈ G to Q Q Q expresses its coset as a finite sum of members of T T T . Subtracting the corresponding x t x_t x t leaves an element of F F F , which is a nonnegative sum of the e i e_i e i and − e i -e_i − e i . Thus [ S ] = G [S]=G [ S ] = G .
Step 2: define additive correction terms. For B ∈ B ( T ) B\in\mathcal{B}(T) B ∈ B ( T ) , the element ∑ t ∈ T v t ( B ) x t \sum_{t\in T}v_t(B)x_t ∑ t ∈ T v t ( B ) x t lies in F F F . Write
∑ t ∈ T v t ( B ) x t = ∑ i ∈ I a i ( B ) e i . (10) \sum_{t\in T}v_t(B)x_t=\sum_{i\in I}a_i(B)e_i. \tag*{(10)} t ∈ T ∑ v t ( B ) x t = i ∈ I ∑ a i ( B ) e i . (10)
The coefficients a i ( B ) a_i(B) a i ( B ) are integers with finite support. Applying λ \lambda λ shows that a i ( B ) = ∑ t v t ( B ) λ i ( x t ) ≥ 0 a_i(B)=\sum_{t}v_t(B)\lambda_i(x_t)\geq0 a i ( B ) = ∑ t v t ( B ) λ i ( x t ) ≥ 0 . The map B ↦ ( a i ( B ) ) i B\mapsto(a_i(B))_i B ↦ ( a i ( B ) ) i is additive under concatenation. Hence
B ^ = ∏ t ∈ T x t v t ( B ) ∏ i ∈ I ( − e i ) a i ( B ) (11) \widehat{B}=\prod_{t\in T}x_t^{v_t(B)}\prod_{i\in I}(-e_i)^{a_i(B)} \tag*{(11)} B = t ∈ T ∏ x t v t ( B ) i ∈ I ∏ ( − e i ) a i ( B ) (11)
belongs to B ( S ) \mathcal{B}(S) B ( S ) .
Define
Φ ( B , ( n i ) i ) = B ^ ∏ i ∈ I ( e i ( − e i ) ) n i . (12) \Phi\bigl(B,(n_i)_i\bigr)=\widehat{B}\prod_{i\in I}\bigl(e_i(-e_i)\bigr)^{n_i}. \tag*{(12)} Φ ( B , ( n i ) i ) = B i ∈ I ∏ ( e i ( − e i ) ) n i . (12)
Additivity of the correction vector shows that Φ \Phi Φ is a monoid homomorphism.
Step 3: recover a block from its unique components. To prove bijectivity, let C ∈ B ( S ) C\in\mathcal{B}(S) C ∈ B ( S ) . Its sequence expression is uniquely of the form
C = ∏ t ∈ T x t v t ∏ i ∈ I e i u i ( − e i ) w i . C=\prod_{t\in T}x_t^{v_t}\prod_{i\in I}e_i^{u_i}(-e_i)^{w_i}. C = t ∈ T ∏ x t v t i ∈ I ∏ e i u i ( − e i ) w i .
Projection to Q Q Q shows that B = ∏ t t v t B=\prod_t t^{v_t} B = ∏ t t v t belongs to B ( T ) \mathcal{B}(T) B ( T ) . The zero-sum condition in G G G , together with (4.3) and independence of the e i e_i e i , gives w i = a i ( B ) + u i w_i=a_i(B)+u_i w i = a i ( B ) + u i for every i i i . Therefore C = Φ ( B , ( u i ) i ) C=\Phi(B,(u_i)_i) C = Φ ( B , ( u i ) i ) . The lift multiplicities determine B B B , and the e i e_i e i -multiplicities determine ( u i ) i (u_i)_i ( u i ) i , so this expression is unique. This proves (4.1). A product with a free commutative monoid is half-factorial whenever the other factor is half-factorial, proving the last assertion. □
The same isomorphism identifies the entire factorization-length structure, not only half-factoriality.
Corollary 4.2. Let G G G be an abelian group, let F = ⨁ i ∈ I Z e i F=\bigoplus_{i\in I}\mathbb{Z}e_i F = ⨁ i ∈ I Z e i be a free subgroup, and let T ⊆ ( G / F ) ∖ { 0 } T\subseteq(G/F)\setminus\{0\} T ⊆ ( G / F ) ∖ { 0 } satisfy [ T ] = G / F [T]=G/F [ T ] = G / F . Choose the representatives x t x_t x t , the set S S S , and the isomorphism Φ \Phi Φ by the construction in the proof of Theorem 4.1. Then the atoms of B ( S ) \mathcal{B}(S) B ( S ) are precisely the sequences U ^ \widehat{U} U , with U ∈ A ( B ( T ) ) U\in\mathcal{A}(\mathcal{B}(T)) U ∈ A ( B ( T )) , and the pairs e i ( − e i ) e_i(-e_i) e i ( − e i ) , with i ∈ I i\in I i ∈ I . For every B ∈ B ( T ) B\in\mathcal{B}(T) B ∈ B ( T ) and ( n i ) i ∈ N ( I ) (n_i)_i\in\mathbb{N}^{(I)} ( n i ) i ∈ N ( I ) ,
L B ( S ) ( Φ ( B , ( n i ) i ) ) = L B ( T ) ( B ) + ∑ i ∈ I n i . (13) \mathsf{L}_{\mathcal{B}(S)}\bigl(\Phi(B,(n_i)_i)\bigr)=\mathsf{L}_{\mathcal{B}(T)}(B)+\sum_{i\in I}n_i. \tag*{(13)} L B ( S ) ( Φ ( B , ( n i ) i ) ) = L B ( T ) ( B ) + i ∈ I ∑ n i . (13) In particular, if T T T is half-factorial and its factorization-length function is ℓ T \ell_T ℓ T , then the factorization-length function on B ( S ) \mathcal{B}(S) B ( S ) is given by ℓ S ( Φ ( B , ( n i ) i ) ) = ℓ T ( B ) + ∑ i n i \ell_S(\Phi(B,(n_i)_i))=\ell_T(B)+\sum_i n_i ℓ S ( Φ ( B , ( n i ) i )) = ℓ T ( B ) + ∑ i n i .
Proof. An atom of a direct product of reduced atomic monoids has exactly one nonidentity component, and that component is an atom. In the factor N 0 ( I ) \mathbb{N}_0^{(I)} N 0 ( I ) , the atoms are the coordinate vectors and the factorization length of ( n i ) i (n_i)_i ( n i ) i is ∑ i n i \sum_i n_i ∑ i n i . Transport these statements through Φ \Phi Φ . □ \square □
The construction applies to nonsplit extensions in a concrete way.
Example 4.3. Let p p p be a prime, G = Z [ 1 / p ] G = \mathbb{Z}[1/p] G = Z [ 1/ p ] , and F = Z F = \mathbb{Z} F = Z . Then G / F ≅ C p ∞ G/F \cong C_{p^\infty} G / F ≅ C p ∞ . For c n = p − n + Z c_n = p^{-n} + \mathbb{Z} c n = p − n + Z , take T = { c n : n ≥ 1 } T = \{c_n : n \ge1\} T = { c n : n ≥ 1 } and the representatives x c n = p − n x_{c_n} = p^{-n} x c n = p − n . Theorem 4.1 gives
S = { − 1 , 1 } ∪ { p − n : n ≥ 1 } , B ( S ) ≅ B ( T ) × N 0 . S = \{-1,1\} \cup\{p^{-n} : n \ge1\}, \qquad\mathcal{B}(S) \cong\mathcal{B}(T) \times\mathbb{N}_0. S = { − 1 , 1 } ∪ { p − n : n ≥ 1 } , B ( S ) ≅ B ( T ) × N 0 .
The half-factoriality of T T T is verified in Proposition 2.6 . A block B ∈ B ( T ) B \in\mathcal{B}(T) B ∈ B ( T ) has integral weight m = ∑ n v c n ( B ) p − n m = \sum_n v_{c_n}(B)p^{-n} m = ∑ n v c n ( B ) p − n , and its lift is B ^ = ∏ n ( p − n ) v c n ( B ) ( − 1 ) m \widehat{B} = \prod_n (p^{-n})^{v_{c_n}(B)}(-1)^m B = ∏ n ( p − n ) v c n ( B ) ( − 1 ) m . The exact sequence does not split: a nonzero torsion quotient C p ∞ C_{p^\infty} C p ∞ cannot be a subgroup of the torsion-free group Z [ 1 / p ] \mathbb{Z}[1/p] Z [ 1/ p ] . Also, F F F is not pure in G G G , since 1 ∈ p G ∩ F 1 \in pG \cap F 1 ∈ pG ∩ F but 1 ∉ p F 1 \notin pF 1 ∈ / pF .
The classical primary reduction and positive consequences
The reduction from arbitrary groups to reduced uncountable unbounded primary groups is due to Geroldinger–Göbel [8 pp. 855–856] . We give the argument because it explains the role of free kernels and distinguishes the known reduction from the separation theorem of § 3 .
Theorem 4.4 (Geroldinger–Göbel reduction). The following assertions are equivalent.
Every abelian group belongs to H \mathcal{H} H .
Every torsion abelian group belongs to H \mathcal{H} H .
For every prime p p p , every abelian p p p -group belongs to H \mathcal{H} H .
For every prime p p p , every reduced abelian p p p -group belongs to H \mathcal{H} H .
For every prime p p p , every reduced uncountable unbounded abelian p p p -group belongs to H \mathcal{H} H .
Consequently each assertion is equivalent to an affirmative answer to the universal half-factorial Dedekind realization problem.
Proof. The implications (1) ⇒ \Rightarrow ⇒ (2) ⇒ \Rightarrow ⇒ (3) ⇒ \Rightarrow ⇒ (4) ⇒ \Rightarrow ⇒ (5) are immediate. Assertion (5) implies (4) by the countable and bounded cases recalled in § 2 . We prove the other converse steps.
Assume (4), and let P P P be a p p p -group. Its maximal divisible subgroup D D D is a direct summand, so P = D ⊕ R P = D \oplus R P = D ⊕ R with R R R reduced. Moreover, D D D is a direct sum of copies of C p ∞ C_{p^\infty} C p ∞ . These are standard structure results for abelian groups [5 ] . By Proposition 2.6 , Lemma 2.5 , and (4), both summands and hence P P P belong to H \mathcal{H} H . This proves (3).
Assume (3). Every torsion abelian group is the direct sum of its p p p -primary components [5 ] . Apply Lemma 2.5 to obtain (2).
Finally, assume (2), and let G G G be any abelian group. Choose a maximal Z \mathbb{Z} Z -linearly independent subset { e i : i ∈ I } \{e_i : i \in I\} { e i : i ∈ I } of G G G and let F = ⨁ i Z e i F = \bigoplus_i \mathbb{Z}e_i F = ⨁ i Z e i . For every g ∈ G g \in G g ∈ G , either g ∈ F g \in F g ∈ F or maximality gives a relation n g + ∑ i n i e i = 0 ng + \sum_i n_i e_i = 0 n g + ∑ i n i e i = 0 with n ≠ 0 n \ne0 n = 0 : a relation with n = 0 n = 0 n = 0 would contradict independence of the e i e_i e i . Thus a nonzero multiple of every coset in G / F G/F G / F is zero, and G / F G/F G / F is torsion. By (2), G / F ∈ H G/F \in\mathcal{H} G / F ∈ H . Theorem 4.1 gives G ∈ H G \in\mathcal{H} G ∈ H , proving (1). The final assertion follows from Theorem 2.1 . □ \square □
The reduction is a logical equivalence between universal assertions; Theorem 3.6 shows that all of them are false. For an individual quotient, however, the lifting theorem remains a useful positive construction.
Corollary 4.5. Let G G G be an abelian group and let F F F be a free subgroup. If G / F G/F G / F is countable, divisible, or a direct sum of cyclic groups and Prüfer groups, then G ∈ H G \in\mathcal{H} G ∈ H . In particular, every torsion-free abelian group of finite rank is the ideal class group of a half-factorial Dedekind domain.
Proof. The countable case is [8 Theorem 5.3] . A divisible group is a direct sum of copies of Q \mathbb{Q} Q and of Prüfer groups [5 ] ; these summands belong to H \mathcal{H} H by the countable case. The cyclic and Prüfer case follows from Lemma 2.5 and Proposition 2.6. Apply Theorem 4.1. A torsion-free group of finite rank embeds in Q r \mathbb{Q}^{r} Q r for some finite r r r , and is therefore countable. Its realization follows from the countable case and Theorem 2.1.
More generally, the quotient may be any group already known to belong to H \mathcal{H} H . In particular, the Warfield-group theorem [8 Theorem 5.1] applies; the corresponding full-rank quotient consequence is explicitly [8 Corollary 5.2] .
Root adjunction and torsion kernels
A single root adjunction admits an exact criterion. The special case in which the lower multiple is already among the generators is [8 Proposition 3.3] . The formulation below also covers a lower multiple not originally in the generating set.
Proposition 4.6. Let G G G be a torsion abelian group, let H ⊆ G H \subseteq G H ⊆ G be a subgroup, and let T ⊆ H ∖ { 0 } T \subseteq H \setminus\{0\} T ⊆ H ∖ { 0 } generate H H H . For x ∈ G ∖ H x \in G \setminus H x ∈ G ∖ H , put m = ord ( x + H ) m = \operatorname{ord}(x+H) m = ord ( x + H ) and h = m x ∈ H h=mx\in H h = m x ∈ H . There is a surjective homomorphism
θ : B ( T ∪ { x } ) ⟶ B ( T ∪ { h } ) , A x m k ⟼ A h k , \theta: \mathcal{B}(T \cup\{x\}) \longrightarrow\mathcal{B}(T \cup\{h\}), \qquad Ax^{mk} \longmapsto Ah^k, θ : B ( T ∪ { x }) ⟶ B ( T ∪ { h }) , A x mk ⟼ A h k , which preserves all sets of factorization lengths. Consequently, T ∪ { x } T \cup\{x\} T ∪ { x } is half-factorial if and only if T ∪ { h } T \cup\{h\} T ∪ { h } is half-factorial. The statement includes h = 0 h=0 h = 0 .
Proof. The multiplicity of x x x in a block is divisible by m m m , by projection to G / H G/H G / H , and substitution of x m x^m x m by h h h preserves its sum. If h ∉ T h \notin T h ∈ / T , the displayed map is an isomorphism: replace every occurrence of h h h by x m x^m x m to obtain its inverse.
If h ∈ T h \in T h ∈ T , distinguish temporarily the occurrences of h h h coming from x m x^m x m from the original occurrences of h h h in A A A . A factorization of θ ( A x m k ) \theta(Ax^{mk}) θ ( A x mk ) can be lifted by distributing these marked occurrences among its factors, replacing each marked occurrence by x m x^m x m . The resulting factors are nonempty blocks and multiply to A x m k Ax^{mk} A x mk . Conversely, the image of a nonempty block is nonempty; if the image of an atom factored, this lifting would factor the atom. Surjectivity follows by taking k = 0 k=0 k = 0 . Therefore the map preserves exactly the sets of lengths.
For a sequence over a subgroup or a quotient we compute its cross number using the orders in that group, indicating the group by a subscript when necessary.
Lemma 4.7. Let
0 ⟶ K ⟶ G → π Q ⟶ 0 0 \longrightarrow K \longrightarrow G \xrightarrow{\pi} Q \longrightarrow0 0 ⟶ K ⟶ G π Q ⟶ 0 be an exact sequence of torsion abelian groups. Let A ⊆ K ∖ { 0 } A \subseteq K \setminus\{0\} A ⊆ K ∖ { 0 } and T ⊆ Q ∖ { 0 } T \subseteq Q \setminus\{0\} T ⊆ Q ∖ { 0 } be half-factorial generating sets. Choose x t ∈ G x_t \in G x t ∈ G with π ( x t ) = t \pi(x_t)=t π ( x t ) = t and put S = A ∪ { x t : t ∈ T } S=A\cup\{x_t:t\in T\} S = A ∪ { x t : t ∈ T } . For U ∈ A ( B ( T ) ) U\in\mathcal{A}(\mathcal{B}(T)) U ∈ A ( B ( T )) , define
V T = ∏ t ∈ T x t v t ( U ) , d ( U ) = 1 − ∑ t ∈ T v t ( U ) ord G ( x t ) . V_T=\prod_{t\in T}x_t^{v_t(U)}, \qquad d(U)=1-\sum_{t\in T}\frac{v_t(U)}{\operatorname{ord}_G(x_t)}. V T = t ∈ T ∏ x t v t ( U ) , d ( U ) = 1 − t ∈ T ∑ ord G ( x t ) v t ( U ) . If S S S is half-factorial, then there is a sequence C ∈ F ( A ) C\in\mathcal{F}(A) C ∈ F ( A ) such that V U C V_UC V U C is an atom of B ( S ) \mathcal{B}(S) B ( S ) and
σ ( C ) = − σ ( V U ) , k K ( C ) = d ( U ) . (14) \sigma(C)=-\sigma(V_U), \qquad\mathsf{k}_K(C)=d(U). \tag*{(14)} σ ( C ) = − σ ( V U ) , k K ( C ) = d ( U ) . (14) If σ ( V U ) ≠ 0 \sigma(V_U)\ne0 σ ( V U ) = 0 , then d ( U ) > 0 d(U)>0 d ( U ) > 0 and at least one letter t t t in U U U satisfies ord G ( x t ) > ord Q ( t ) \operatorname{ord}_G(x_t)>\operatorname{ord}_Q(t) ord G ( x t ) > ord Q ( t ) .
Proof. Since σ ( V U ) ∈ K \sigma(V_U)\in K σ ( V U ) ∈ K and A A A generates K K K as a monoid, choose D ∈ F ( A ) D\in\mathcal{F}(A) D ∈ F ( A ) with σ ( D ) = − σ ( V U ) \sigma(D)=-\sigma(V_U) σ ( D ) = − σ ( V U ) . Factor the block V U D V_UD V U D into atoms over S S S . Projection to Q Q Q , omitting the letters in A A A , sends each factor to a block dividing U U U . Because U U U is an atom, exactly one factor contains all the letters of V U V_U V U ; all other factors contain only letters of A A A . Write that distinguished factor as V U C V_UC V U C , with C ∣ D C\mid D C ∣ D . Half-factoriality and Proposition 2.2 give k G ( V U ) + k K ( C ) = 1 \mathsf{k}_G(V_U)+\mathsf{k}_K(C)=1 k G ( V U ) + k K ( C ) = 1 .
If σ ( V U ) ≠ 0 \sigma(V_U)\ne0 σ ( V U ) = 0 , then C C C is nonempty and its cross number is positive. Also ord Q ( t ) \operatorname{ord}_Q(t) ord Q ( t ) divides ord G ( x t ) \operatorname{ord}_G(x_t) ord G ( x t ) , while k Q ( U ) = 1 \mathsf{k}_Q(U)=1 k Q ( U ) = 1 . Thus d ( U ) > 0 d(U)>0 d ( U ) > 0 requires strict increase of the order of at least one lifted letter.
In particular, there is no room for a nonempty correction when all lifted orders equal the corresponding quotient orders. This yields an equivalence, rather than merely a necessary numerical test.
Theorem 4.8. Let $0 \to K \to G \xrightarrow{\pi} Q \to 0$ be an exact sequence of torsion abelian groups. Let A ⊆ K ∖ { 0 } A \subseteq K \setminus\{0\} A ⊆ K ∖ { 0 } and T ⊆ Q ∖ { 0 } T \subseteq Q \setminus\{0\} T ⊆ Q ∖ { 0 } be half-factorial generating sets. Choose representatives x t ∈ G x_t \in G x t ∈ G such that
π ( x t ) = t , ord G ( x t ) = ord Q ( t ) ( t ∈ T ) , \pi(x_t)=t,\qquad\operatorname{ord}_G(x_t)=\operatorname{ord}_Q(t)\quad(t\in T), π ( x t ) = t , ord G ( x t ) = ord Q ( t ) ( t ∈ T ) , and put S = A ∪ { x t : t ∈ T } S=A\cup\{x_t:t\in T\} S = A ∪ { x t : t ∈ T } . The following are equivalent.
S S S is half-factorial.
For every U ∈ A ( B ( T ) ) U\in\mathcal{A}(B(T)) U ∈ A ( B ( T )) , one has ∑ t v t ( U ) x t = 0 \sum_t v_t(U)x_t=0 ∑ t v t ( U ) x t = 0 in G G G .
The assignment t ↦ x t t\mapsto x_t t ↦ x t extends to a homomorphism s : Q → G s:Q\to G s : Q → G with π s = id Q \pi s=\operatorname{id}_Q π s = id Q .
When these conditions hold, G = K ⊕ s ( Q ) G=K\oplus s(Q) G = K ⊕ s ( Q ) and B ( S ) ≅ B ( A ) × B ( T ) B(S)\cong B(A)\times B(T) B ( S ) ≅ B ( A ) × B ( T ) . In particular, if the exact sequence does not split, no choice of order-preserving representatives yields a half-factorial set of the displayed form.
Proof. Under the order hypothesis, d ( U ) = 1 − k Q ( U ) = 0 d(U)=1-k_Q(U)=0 d ( U ) = 1 − k Q ( U ) = 0 for every quotient atom U U U . Lemma 4.7 proves (1) ⇒ \Rightarrow ⇒ (2).
Assume (2). Since every block over T T T factors into atoms, the lifted sum of every such block is zero. Consider a finite integer relation ∑ t n t t = 0 \sum_t n_t t=0 ∑ t n t t = 0 in Q Q Q . For each occurring t t t , choose m t ≥ 0 m_t\geq0 m t ≥ 0 with m t ≡ n t ( m o d ord Q ( t ) ) m_t\equiv n_t\pmod{\operatorname{ord}_Q(t)} m t ≡ n t ( mod ord Q ( t )) . The sequence ∏ t t m t \prod_t t^{m_t} ∏ t t m t is a block, so ∑ t m t x t = 0 \sum_t m_tx_t=0 ∑ t m t x t = 0 by the preceding observation. Since ord G ( x t ) = ord Q ( t ) \operatorname{ord}_G(x_t)=\operatorname{ord}_Q(t) ord G ( x t ) = ord Q ( t ) , replacing m t m_t m t by n t n_t n t does not change its lifted sum. Thus ∑ t n t x t = 0 \sum_t n_tx_t=0 ∑ t n t x t = 0 .
All defining group relations among T T T are therefore respected. The assignment extends to a homomorphism s : Q → G s:Q\to G s : Q → G , and π s = id Q \pi s=\operatorname{id}_Q π s = id Q because this equality holds on the generating set T T T . This proves (2) ⇒ \Rightarrow ⇒ (3).
Finally, a section gives G = K ⊕ s ( Q ) G=K\oplus s(Q) G = K ⊕ s ( Q ) . The two parts of S S S lie in the two summands, so the block monoid is the displayed product. Both factors are half-factorial. This proves (3) ⇒ \Rightarrow ⇒ (1) and the remaining assertions. □
The hypotheses concern the chosen generating sets. They do not say that a nonsplit torsion extension fails to belong to H H H . The following countable example makes that limitation explicit.
Example 4.9. Let p p p be a prime and let E = ⨁ n ≥ 1 ⟨ e n ⟩ E=\bigoplus_{n\geq1}\langle e_n\rangle E = ⨁ n ≥ 1 ⟨ e n ⟩ with ord ( e n ) = p n \operatorname{ord}(e_n)=p^n ord ( e n ) = p n , and define an epimorphism ρ : E → C p ∞ \rho:E\to C_{p^\infty} ρ : E → C p ∞ by ρ ( e n ) = c n = 1 / p n + Z \rho(e_n)=c_n=1/p^n+\mathbb{Z} ρ ( e n ) = c n = 1/ p n + Z . Put k n = e n − p e n + 1 k_n=e_n-pe_{n+1} k n = e n − p e n + 1 . Then
K = ker ρ = ⨁ n ≥ 1 ⟨ k n ⟩ , ord ( k n ) = p n . K=\ker\rho=\bigoplus_{n\geq1}\langle k_n\rangle,\qquad\operatorname{ord}(k_n)=p^n. K = ker ρ = n ≥ 1 ⨁ ⟨ k n ⟩ , ord ( k n ) = p n . Indeed, a relation among finitely many k n k_n k n is trivial by successively inspecting its highest e n + 1 e_{n+1} e n + 1 -coordinate. To identify the kernel, replace e n e_n e n successively by k n + p e n + 1 k_n+pe_{n+1} k n + p e n + 1 in a finitely supported element of ker ρ \ker\rho ker ρ ; the remaining last coefficient is divisible by the order of its basis element and hence contributes zero.
Both A = { k n : n ≥ 1 } A=\{k_n:n\geq1\} A = { k n : n ≥ 1 } and T = { c n : n ≥ 1 } T=\{c_n:n\geq1\} T = { c n : n ≥ 1 } are half-factorial generating sets. The representatives x c n = e n x_{c_n}=e_n x c n = e n preserve orders, but the exact sequence does not split, since E E E is reduced. Thus A ∪ { e n : n ≥ 1 } A\cup\{e_n:n\geq1\} A ∪ { e n : n ≥ 1 } is not half-factorial. Directly, for every n ≥ 1 n\geq1 n ≥ 1 the sequence e n + 1 p n k n e n p n − 1 e_{n+1}^{p^n}k_ne_n^{p^n-1} e n + 1 p n k n e n p n − 1 is an atom with cross number 1 + 1 / p n 1+1/p^n 1 + 1/ p n : a zero-sum subsequence must either omit k n k_n k n and be empty, or contain k n k_n k n and then have all the indicated multiplicities. Nevertheless, { e n : n ≥ 1 } \{e_n:n\geq1\} { e n : n ≥ 1 } alone is a half-factorial generating set of E E E . Thus the obstruction comes from insisting on adjoining the prescribed kernel generators.
The pure extension associated to the torsion-completion
The group from Theorem 3.6 also exhibits a failure of extension closure. We identify its pure direct-sum subgroup and divisible quotient explicitly. By the torsion-completion of a p p p -group B B B we mean the torsion subgroup of its p p p -adic completion B ^ = l i m ← r B / p r B \widehat{B}=\varprojlim_r B/p^rB B = lim r B / p r B ; see [6 ] .
Proposition 4.10. Fix a prime p p p . Set C n = Z / p n Z C_n = \mathbb{Z}/p^n\mathbb{Z} C n = Z / p n Z , B = ⨁ n ≥ 1 C n B = \bigoplus_{n\ge1} C_n B = ⨁ n ≥ 1 C n , and
P = { ( a n ) n ∈ ∏ n ≥ 1 C n : p r a n = 0 for every n and some r ≥ 0 } . (15) P = \left\{(a_n)_{n \in\prod_{n\ge1} C_n} : p^r a_n = 0\text{ for every }n\text{ and some }r \ge0\right\}. \tag*{(15)} P = { ( a n ) n ∈ ∏ n ≥ 1 C n : p r a n = 0 for every n and some r ≥ 0 } . (15) Then P P P is the torsion-completion of B B B , is reduced, uncountable, and unbounded, and B B B is pure in P P P . The quotient D = P / B D = P/B D = P / B is a nonzero divisible p p p -group.
Let A ⊆ B ∖ { 0 } A \subseteq B \setminus\{0\} A ⊆ B ∖ { 0 } and T ⊆ D ∖ { 0 } T \subseteq D \setminus\{0\} T ⊆ D ∖ { 0 } be any half-factorial generating sets. Order-preserving representatives x t ∈ P x_t \in P x t ∈ P of all t ∈ T t \in T t ∈ T exist, but for every such choice the set A ∪ { x t : t ∈ T } A \cup\{x_t:t \in T\} A ∪ { x t : t ∈ T } is not half-factorial.
Proof. The p p p -adic completion of B B B identifies with the subgroup of ∏ n C n \prod_{n} C_n ∏ n C n consisting of sequences whose coordinates belong to p k C n p^kC_n p k C n for all but finitely many n n n , for every k k k . This follows by viewing B / p k B B/p^kB B / p k B as the direct sum of the coordinate quotients C n / p k C n C_n/p^kC_n C n / p k C n and taking the inverse limit. If a sequence is annihilated by p r p^r p r , its n n n th coordinate for n > r n>r n > r belongs to p n − r C n p^{n-r}C_n p n − r C n . Thus every sequence in (15) lies in the completion. Multiplication by an integer prime to p p p is invertible in every coordinate. Thus every torsion element of the completion is killed by a power of p p p , and its torsion subgroup is exactly P P P .
For each fixed coordinate n n n , every element of p k P p^kP p k P has zero n n n th coordinate when k ≥ n k \ge n k ≥ n . Hence ⋂ k ≥ 0 p k P = 0 \bigcap_{k\ge0}p^kP=0 ⋂ k ≥ 0 p k P = 0 , which implies that P P P is reduced. Moreover, P [ p ] = ∏ n ≥ 1 C n [ p ] P[p] = \prod_{n\ge1} C_n[p] P [ p ] = ∏ n ≥ 1 C n [ p ] is a product of countably many nontrivial groups of order p p p , so P P P is uncountable. It is unbounded because it contains B B B .
If b ∈ B ∩ p k P b \in B \cap p^kP b ∈ B ∩ p k P , choose y ∈ P y \in P y ∈ P with p k y = b p^ky=b p k y = b and truncate y y y to the finite support of b b b . The truncation lies in B B B and still maps to b b b , proving B ∩ p k P = p k B B \cap p^kP = p^kB B ∩ p k P = p k B . Multiplication by any integer prime to p p p is an automorphism on both groups, so B B B is pure in P P P .
To prove divisibility of D D D , let x = ( x n ) n ∈ P x=(x_n)_n \in P x = ( x n ) n ∈ P be annihilated by p r p^r p r , with r ≥ 1 r \ge1 r ≥ 1 . Truncate the first r r r coordinates to obtain b ∈ B b \in B b ∈ B . For n > r n>r n > r , the coordinate x n x_n x n is divisible by p p p in C n C_n C n . Choose roots coordinatewise to obtain y y y with p y = x − b py=x-b p y = x − b . The roots are annihilated by p r + 1 p^{r+1} p r + 1 , so y ∈ P y \in P y ∈ P . It follows that multiplication by p p p on D D D is surjective, and D D D is divisible since it is a p p p -group. The element ( p n − 1 + p n Z ) n ≥ 1 (p^{n-1}+p^n\mathbb{Z})_{n\ge1} ( p n − 1 + p n Z ) n ≥ 1 belongs to P ∖ B P\setminus B P ∖ B , so D ≠ 0 D\ne0 D = 0 .
The group B B B has a half-factorial generating set by its cyclic decomposition; D D D has one by its decomposition into Prüfer groups and Proposition 2.6 . For any element t ∈ D t \in D t ∈ D of order p r p^r p r , choose a representative y ∈ P y \in P y ∈ P . Purity gives p r y ∈ B ∩ p r P = p r B p^ry \in B \cap p^rP = p^rB p r y ∈ B ∩ p r P = p r B . Choose b ∈ B b \in B b ∈ B with p r b = p r y p^rb=p^ry p r b = p r y . Then x t = y − b x_t=y-b x t = y − b represents t t t and has order exactly p r p^r p r .
Finally, $0 \longrightarrow B \longrightarrow P \longrightarrow D \longrightarrow 0$ cannot split: a section would embed the nonzero divisible group D D D in the reduced group P P P . The last assertion therefore follows from Theorem 4.8 . □ \square □
Remark 4.11. Proposition 4.10, considered on its own, excludes only a specified class of lifting constructions. A putative half-factorial generating set of P P P need not contain generators of this particular subgroup B B B , have a half-factorial image in D D D , or preserve orders under projection. These possibilities are all excluded separately by Theorem 3.6 , whose proof makes none of those assumptions. Example 4.9 shows that the first limitation matters even in a countable group already known to belong to H \mathcal{H} H . Theorem 5.2 also applies to P P P , so a weakly half-factorial generating set does exist in this test case.
The pure extension just computed yields a closure failure for existence of half-factorial generators, not merely for a selected choice of representatives.
Corollary 4.12. The class H \mathcal{H} H is not closed under extensions, even under pure extensions of abelian p p p -groups. For every prime p p p , the exact sequence
0 ⟶ B ⟶ P p ⟶ P p / B ⟶ 0 , B = ⨁ j ≥ 1 C p j , 0 \longrightarrow B \longrightarrow P_p \longrightarrow P_p/B \longrightarrow0,\qquad B=\bigoplus_{j\ge1} C_{p^j}, 0 ⟶ B ⟶ P p ⟶ P p / B ⟶ 0 , B = j ≥ 1 ⨁ C p j , has pure kernel, B ∈ H B\in\mathcal{H} B ∈ H , and P p / B ∈ H P_p/B\in\mathcal{H} P p / B ∈ H , but P p ∉ H P_p\notin\mathcal{H} P p ∈ / H .
Proof. Proposition 4.10 proves purity and divisibility of P p / B P_p/B P p / B . Direct sums of cyclic groups and divisible groups belong to H \mathcal{H} H , whereas Theorem 3.6 excludes P p P_p P p . □ \square □
Characters and compatibility of generating sets
Integrality of cross numbers is weaker than half-factoriality. We first show that every torsion abelian group admits generators satisfying this weaker condition. We then impose a prescribed character on each bounded primary group and examine the obstructions to extending or thinning these generators. The completion examples show that these positive constructions cannot be made globally half-factorial in general.
Weak half-factoriality and characters
A subset T T T of a torsion group is weakly half-factorial if every atom of B ( T ) \mathcal{B}(T) B ( T ) has integral cross number. Since every block factors into atoms, this is equivalent to integrality of the cross number of every block. The terminology and finite-group character criterion are standard [7 Lemma 4.1] ; see also [13 , 14 ] . The next proof works without a boundedness or finiteness assumption.
Lemma 5.1. Let G G G be a torsion abelian group and let T ⊆ G ∖ { 0 } T \subseteq G \setminus\{0\} T ⊆ G ∖ { 0 } generate G G G . The following are equivalent: every B ∈ B ( T ) B \in\mathcal{B}(T) B ∈ B ( T ) has integral cross number; and there is a homomorphism χ : G → Q / Z \chi:G \to\mathbb{Q}/\mathbb{Z} χ : G → Q / Z satisfying χ ( t ) = 1 / ord ( t ) + Z \chi(t)=1/\operatorname{ord}(t)+\mathbb{Z} χ ( t ) = 1/ ord ( t ) + Z for every t ∈ T t \in T t ∈ T .
Proof. A character with the stated values gives k ( B ) + Z = χ ( σ ( B ) ) = 0 k(B)+\mathbb{Z}=\chi(\sigma(B))=0 k ( B ) + Z = χ ( σ ( B )) = 0 for each block B B B . Conversely, assign t t t the value 1 / ord ( t ) + Z 1/\operatorname{ord}(t)+\mathbb{Z} 1/ ord ( t ) + Z . Given a finite integer relation ∑ t n t t = 0 \sum_t n_t t=0 ∑ t n t t = 0 , replace each n t n_t n t by a nonnegative integer m t m_t m t congruent to n t n_t n t modulo ord ( t ) \operatorname{ord}(t) ord ( t ) . Then ∏ t t m t \prod_t t^{m_t} ∏ t t m t is a block, so ∑ t m t / ord ( t ) \sum_t m_t/\operatorname{ord}(t) ∑ t m t / ord ( t ) is integral. The same is true of ∑ t n t / ord ( t ) \sum_t n_t/\operatorname{ord}(t) ∑ t n t / ord ( t ) . Thus the assignment respects every relation and defines the required homomorphism. □
The character condition is always achievable when the generating set is allowed to vary. The following proof uses only a linear functional on the socle and the injectivity of divisible abelian groups. The finite-group character theory is developed in [13 ] ; the following construction treats arbitrary cardinality and unbounded exponent.
Theorem 5.2. Every torsion abelian group G G G has a generating subset T ⊆ G ∖ { 0 } T \subseteq G \setminus\{0\} T ⊆ G ∖ { 0 } and a homomorphism χ : G → Q / Z \chi:G \to\mathbb{Q}/\mathbb{Z} χ : G → Q / Z such that
χ ( t ) = 1 ord ( t ) + Z ( t ∈ T ) . (16) \chi(t)=\frac{1}{\operatorname{ord}(t)}+\mathbb{Z}\qquad(t\in T). \tag*{(16)} χ ( t ) = ord ( t ) 1 + Z ( t ∈ T ) . (16) Consequently every zero-sum sequence over T T T has integral cross number. In particular, every reduced abelian p p p -group has a weakly half-factorial generating subset.
Proof. We first treat a nonzero abelian p p p -group P P P . Write P [ p ] = { x ∈ P : p x = 0 } P[p]=\{x\in P:px=0\} P [ p ] = { x ∈ P : p x = 0 } and identify the subgroup of order p p p in C p ∞ ⊆ Q / Z C_{p^\infty}\subseteq\mathbb{Q}/\mathbb{Z} C p ∞ ⊆ Q / Z with the additive group of F p \mathbb{F}_p F p , sending 1 1 1 to 1 / p + Z 1/p+\mathbb{Z} 1/ p + Z .
Step 1: bounded primary groups. Suppose first that P P P has exponent p N p^N p N . Choose z ∈ P z\in P z ∈ P of order p N p^N p N and extend the map z ↦ 1 / p N + Z z\mapsto1/p^N+\mathbb{Z} z ↦ 1/ p N + Z on ⟨ z ⟩ \langle z\rangle ⟨ z ⟩ to a homomorphism χ : P → C p ∞ \chi:P\to C_{p^\infty} χ : P → C p ∞ . This extension exists because divisible abelian groups are injective [5 ] . Put
T = { t ∈ P ∖ { 0 } : χ ( t ) = 1 / ord ( t ) + Z } . T=\{t\in P\setminus\{0\}:\chi(t)=1/\operatorname{ord}(t)+\mathbb{Z}\}. T = { t ∈ P ∖ { 0 } : χ ( t ) = 1/ ord ( t ) + Z } .
Then z ∈ T z\in T z ∈ T . For g ∈ P g\in P g ∈ P , write χ ( g ) = a / p N + Z \chi(g)=a/p^N+\mathbb{Z} χ ( g ) = a / p N + Z with 0 ≤ a < p N 0\leq a<p^N 0 ≤ a < p N . If p ∤ a p\nmid a p ∤ a , then g g g has order p N p^N p N and multiplication by an inverse of a a a modulo p N p^N p N produces a member of T T T generating the same cyclic subgroup as g g g . If p ∣ a p\mid a p ∣ a , apply the same argument to g + z g+z g + z , whose character value has numerator a + 1 a+1 a + 1 prime to p p p . Then g = ( g + z ) − z ∈ ⟨ T ⟩ g=(g+z)-z\in\langle T\rangle g = ( g + z ) − z ∈ ⟨ T ⟩ . Thus T T T generates P P P .
Step 2: a functional on the socle in the unbounded case. Now suppose that P P P is unbounded. For n ≥ 0 n\geq0 n ≥ 0 let V n = ( p n P ) [ p ] ⊆ P [ p ] V_n=(p^nP)[p]\subseteq P[p] V n = ( p n P ) [ p ] ⊆ P [ p ] . These are nonzero descending F p \mathbb{F}_p F p -subspaces. There is a linear functional f : P [ p ] → F p f:P[p]\to\mathbb{F}_p f : P [ p ] → F p that is nonzero on every V n V_n V n . Indeed, if ⋂ n V n ≠ 0 \bigcap_n V_n\neq0 ⋂ n V n = 0 , prescribe f ( v ) = 1 f(v)=1 f ( v ) = 1 on a nonzero vector in the intersection and extend to a basis. If the intersection is zero, choose increasing integers n j n_j n j with V n j ≠ V n j + 1 V_{n_j}\neq V_{n_{j+1}} V n j = V n j + 1 and vectors v j ∈ V n j ∖ V n j + 1 v_j\in V_{n_j}\setminus V_{n_{j+1}} v j ∈ V n j ∖ V n j + 1 . The n j n_j n j are unbounded, and the v j v_j v j are linearly independent: in a nontrivial finite relation, the term with the least index cannot lie in the subspace containing all later terms. Prescribe f ( v j ) = 1 f(v_j)=1 f ( v j ) = 1 and extend to a basis. This again gives the required functional. Step 3: extend the functional and normalize generators. Regard f f f as a homomorphism to the subgroup of order p p p in C p ∞ C_{p^\infty} C p ∞ and extend it to χ : P → C p ∞ \chi: P \to C_{p^\infty} χ : P → C p ∞ . For every n ≥ 1 n \ge1 n ≥ 1 , choose v ∈ V n − 1 v \in V_{n-1} v ∈ V n − 1 with f ( v ) ≠ 0 f(v) \ne0 f ( v ) = 0 and write v = p n − 1 y v = p^{n-1}y v = p n − 1 y . Then y y y and χ ( y ) \chi(y) χ ( y ) both have order p n p^n p n . Multiplying y y y by a suitable integer prime to p p p gives an element z n z_n z n of order p n p^n p n with χ ( z n ) = 1 / p n + Z \chi(z_n) = 1/p^n + \mathbb{Z} χ ( z n ) = 1/ p n + Z .
Define T T T by the same displayed formula as in the bounded case. If g ∈ P g \in P g ∈ P is annihilated by p m p^m p m , then χ ( g + z m + 1 ) = a / p m + 1 + Z \chi(g + z_{m+1}) = a/p^{m+1} + \mathbb{Z} χ ( g + z m + 1 ) = a / p m + 1 + Z for an integer a ≡ 1 ( m o d p ) a \equiv1 \pmod p a ≡ 1 ( mod p ) . Hence g + z m + 1 g + z_{m+1} g + z m + 1 has order p m + 1 p^{m+1} p m + 1 . For an inverse u u u of a a a modulo p m + 1 p^{m+1} p m + 1 , the element t = u ( g + z m + 1 ) t = u(g + z_{m+1}) t = u ( g + z m + 1 ) belongs to T T T , and g = a t − z m + 1 ∈ ⟨ T ⟩ g = at - z_{m+1} \in\langle T\rangle g = a t − z m + 1 ∈ ⟨ T ⟩ . Thus T T T generates P P P . Since P P P is torsion, group generation is monoid generation.
Step 4: combine the primary components. For a general torsion group use its primary decomposition G = ⨁ p G p G = \bigoplus_p G_p G = ⨁ p G p . Apply the construction to each nonzero G p G_p G p , take the union of the generating subsets, and define χ ( ( g p ) p ) = ∑ p χ p ( g p ) \chi((g_p)_p) = \sum_p \chi_p(g_p) χ (( g p ) p ) = ∑ p χ p ( g p ) . This sum is finite for every g ∈ G g \in G g ∈ G . Equation (16) follows, and Lemma 5.1 gives integrality. For the trivial group use T = ∅ T = \varnothing T = ∅ and the zero homomorphism. □
For a homomorphism χ : P → Q / Z \chi: P \to\mathbb{Q}/\mathbb{Z} χ : P → Q / Z on an abelian p p p -group, put
W χ ( P ) = { x ∈ P ∖ { 0 } : χ ( x ) = 1 / ord ( x ) + Z } . (17) W_\chi(P) = \{x \in P \setminus\{0\} : \chi(x) = 1/\operatorname{ord}(x) + \mathbb{Z}\}. \tag*{(17)} W χ ( P ) = { x ∈ P ∖ { 0 } : χ ( x ) = 1/ ord ( x ) + Z } . (17)
Every block over this set has integral cross number, whether or not it generates P P P . We refer to W χ ( P ) W_\chi(P) W χ ( P ) as the full character set; the next results contrast its arithmetic with that of suitable subsets.
A prescribed character on every bounded layer
The bounded-group existence assertion is classical. The next construction additionally prescribes the character and gives a uniform sequence length for all block atoms. We use the standard cyclic decomposition of a bounded abelian p p p -group [5 ] .
Theorem 5.3. Let p p p be a prime, let N ≥ 1 N \ge1 N ≥ 1 , and let P P P be an abelian group of exponent p N p^N p N . Suppose that χ : P → Q / Z \chi: P \to\mathbb{Q}/\mathbb{Z} χ : P → Q / Z has image the subgroup of order p N p^N p N . Then there is a half-factorial generating subset T ⊆ W χ ( P ) T \subseteq W_\chi(P) T ⊆ W χ ( P ) such that every t ∈ T t \in T t ∈ T has order p N p^N p N . Every atom of B ( T ) \mathcal{B}(T) B ( T ) has sequence length p N p^N p N , and
L B ( T ) ( B ) = { ∣ B ∣ / p N } ( B ∈ B ( T ) ) . L_{\mathcal{B}(T)}(B) = \{|B|/p^N\} \qquad(B \in\mathcal{B}(T)). L B ( T ) ( B ) = { ∣ B ∣/ p N } ( B ∈ B ( T )) . More explicitly, choose z ∈ P z \in P z ∈ P with χ ( z ) = 1 / p N + Z \chi(z) = 1/p^N + \mathbb{Z} χ ( z ) = 1/ p N + Z and write ker χ = ⨁ i ∈ I ⟨ e i ⟩ \ker\chi= \bigoplus_{i \in I}\langle e_i\rangle ker χ = ⨁ i ∈ I ⟨ e i ⟩ , where ord ( e i ) = p r i \operatorname{ord}(e_i) = p^{r_i} ord ( e i ) = p r i and 1 ≤ r i ≤ N 1 \le r_i \le N 1 ≤ r i ≤ N . One may take
T = { z } ∪ { z + e i : i ∈ I } . (18) T = \{z\} \cup\{z + e_i : i \in I\}. \tag*{(18)} T = { z } ∪ { z + e i : i ∈ I } . (18) Proof. The element z z z has order p N p^N p N , and χ \chi χ restricts to an isomorphism from ⟨ z ⟩ \langle z\rangle ⟨ z ⟩ onto its image. Hence P = ⟨ z ⟩ ⊕ ker χ P = \langle z\rangle\oplus\ker\chi P = ⟨ z ⟩ ⊕ ker χ . The displayed set generates P P P , since e i = ( z + e i ) − z e_i = (z + e_i) - z e i = ( z + e i ) − z . Group generation is monoid generation in a torsion group. Its elements are distinct, have character value 1 / p N + Z 1/p^N + \mathbb{Z} 1/ p N + Z , and therefore have order p N p^N p N .
A sequence B = z a 0 ∏ i ( z + e i ) a i B = z^{a_0}\prod_i(z + e_i)^{a_i} B = z a 0 ∏ i ( z + e i ) a i is a block precisely when p r i ∣ a i p^{r_i} \mid a_i p r i ∣ a i for every i i i and p N ∣ a 0 + ∑ i a i p^N \mid a_0 + \sum_i a_i p N ∣ a 0 + ∑ i a i . Partition the occurrences of z + e i z + e_i z + e i into packets of p r i p^{r_i} p r i identical letters, and regard each occurrence of z z z as a packet of size one. The packet sizes are powers of p p p not exceeding p N p^N p N . Whenever their total size is m p N mp^N m p N , they can be grouped into m m m collections of size p N p^N p N . Indeed, the number of packets of the smallest size p j < p N p^j < p^N p j < p N is divisible by p p p , by reducing the total size modulo p j + 1 p^{j+1} p j + 1 . Bundle them in groups of p p p and repeat, retaining the constituent packets. This ends with m m m packets of size p N p^N p N .
Each resulting collection is a zero-sum subsequence: its ker χ \ker\chi ker χ -coordinates vanish packet by packet, and its z z z -coordinate has total coefficient p N p^N p N . Conversely, every nonempty block has sequence length a positive multiple of p N p^N p N , so a block of length p N p^N p N is an atom. Thus all atoms have that length, proving the formula and half-factoriality. □
The same character can be used on all bounded layers of an unbounded group. The assertion concerns the individual layers; it does not assert that the resulting sets are nested.
Corollary 5.4. Let p p p be a prime and let P P P be an unbounded abelian p p p -group. There is a homomorphism χ : P → C p ∞ ⊆ Q / Z \chi:P\to C_{p^\infty}\subseteq\mathbb{Q}/\mathbb{Z} χ : P → C p ∞ ⊆ Q / Z such that, for every n ≥ 1 n\geq1 n ≥ 1 , the subgroup P [ p n ] = { x ∈ P : p n x = 0 } P[p^n]=\{x\in P:p^nx=0\} P [ p n ] = { x ∈ P : p n x = 0 } has a half-factorial generating set T n ⊆ W χ ( P ) T_n\subseteq W_\chi(P) T n ⊆ W χ ( P ) consisting entirely of elements of order p n p^n p n . Consequently, ⋃ n ≥ 1 T n \bigcup_{n\geq1}T_n ⋃ n ≥ 1 T n is a weakly half-factorial generating set of P P P .
Proof. The unbounded-case construction in Theorem 5.2 gives, for every n n n , an element z n ∈ P [ p n ] z_n\in P[p^n] z n ∈ P [ p n ] with χ ( z n ) = 1 / p n + Z \chi(z_n)=1/p^n+\mathbb{Z} χ ( z n ) = 1/ p n + Z . Hence the restriction of χ \chi χ to P [ p n ] P[p^n] P [ p n ] has image of order p n p^n p n . Apply Theorem 5.3 to each restriction. The union generates P = ⋃ n P [ p n ] P=\bigcup_nP[p^n] P = ⋃ n P [ p n ] and is contained in W χ ( P ) W_\chi(P) W χ ( P ) , so its blocks have integral cross number. □
The completion example converts this lack of an asserted compatibility into an actual obstruction, for every prime.
Corollary 5.5. For every prime p p p , the group P p = t ( ∏ j ≥ 1 C p j ) P_p=t\left(\prod_{j\geq1}C_{p^j}\right) P p = t ( ∏ j ≥ 1 C p j ) is a countable increasing union of subgroups belonging to H \mathcal{H} H , but P p ∉ H P_p\notin\mathcal{H} P p ∈ / H . Moreover, there do not exist nested half-factorial sets T n ⊆ P p [ p n ] ∖ { 0 } T_n\subseteq P_p[p^n]\setminus\{0\} T n ⊆ P p [ p n ] ∖ { 0 } generating P p [ p n ] P_p[p^n] P p [ p n ] for all n ≥ 1 n\geq1 n ≥ 1 . Thus H \mathcal{H} H is not closed under countable increasing unions of subgroups.
Proof. Each P p [ p n ] P_p[p^n] P p [ p n ] is bounded, hence belongs to H \mathcal{H} H , and their union is P p P_p P p . Nonmembership follows from Theorem 3.6 . If the sets T n T_n T n were nested, their union would be half-factorial by finite character and would generate P p P_p P p , a contradiction. □
Compatibility does not follow merely from using the same character, even for two successive bounded layers.
Example 5.6. Let P = ⟨ e ⟩ ⊕ ⟨ a ⟩ ⊕ ⟨ b ⟩ P=\langle e\rangle\oplus\langle a\rangle\oplus\langle b\rangle P = ⟨ e ⟩ ⊕ ⟨ a ⟩ ⊕ ⟨ b ⟩ , with ord ( e ) = 4 \operatorname{ord}(e)=4 ord ( e ) = 4 and ord ( a ) = ord ( b ) = 2 \operatorname{ord}(a)=\operatorname{ord}(b)=2 ord ( a ) = ord ( b ) = 2 . Define χ ( e ) = 1 / 4 + Z \chi(e)=1/4+\mathbb{Z} χ ( e ) = 1/4 + Z and χ ( a ) = χ ( b ) = 1 / 2 + Z \chi(a)=\chi(b)=1/2+\mathbb{Z} χ ( a ) = χ ( b ) = 1/2 + Z . Then
T 1 = { 2 e , a , b } , T 2 = { e , a − e , b − e } T_1=\{2e,a,b\},\qquad T_2=\{e,a-e,b-e\} T 1 = { 2 e , a , b } , T 2 = { e , a − e , b − e }
are half-factorial generating sets of P [ 2 ] P[2] P [ 2 ] and P [ 4 ] = P P[4]=P P [ 4 ] = P , respectively, and both lie in W χ ( P ) W_\chi(P) W χ ( P ) . The first assertion follows from independence of 2 e , a , b 2e,a,b 2 e , a , b . For the second, observe that ker χ = ⟨ a − 2 e ⟩ ⊕ ⟨ b − 2 e ⟩ \ker\chi=\langle a-2e\rangle\oplus\langle b-2e\rangle ker χ = ⟨ a − 2 e ⟩ ⊕ ⟨ b − 2 e ⟩ and apply Theorem 5.3 with z = e z=e z = e .
Nevertheless, their union is not half-factorial. The sequence
U = ( 2 e ) a b ( a − e ) ( b − e ) (19) U=(2e)\,a\,b\,(a-e)\,(b-e) \tag*{(19)} U = ( 2 e ) a b ( a − e ) ( b − e ) (19)
is an atom with cross number two. To check minimality, write the multiplicities of 2 e , a , b , a − e , b − e 2e,a,b,a-e,b-e 2 e , a , b , a − e , b − e in a subsequence as c , α , β , u , v ∈ { 0 , 1 } c,\alpha,\beta,u,v\in\{0,1\} c , α , β , u , v ∈ { 0 , 1 } . Its sum vanishes exactly when α = u \alpha=u α = u , β = v \beta=v β = v , and 2 c − u − v ≡ 0 ( m o d 4 ) 2c-u-v\equiv0\pmod{4} 2 c − u − v ≡ 0 ( mod 4 ) . These conditions force all five multiplicities to be zero or all to be one. This proves the claim. Thus individually valid character-compatible constructions do not establish the half-factoriality of their union.
A different proposed compatibility condition would be to require p z n + 1 = z n pz_{n+1}=z_n p z n + 1 = z n for pivots of order p n p^n p n . In a reduced group such an infinite sequence cannot exist: its generated subgroup would be a nonzero copy of C p ∞ C_{p^\infty} C p ∞ . This observation excludes that particular choice of compatible pivots, not all possible half-factorial generating sets.
Finite obstructions to extension and thinning
The following examples distinguish failure of a chosen generating set from nonexistence of all half-factorial generating sets. In particular, a character controls the integrality of cross numbers, not their value on atoms.
Example 5.7. Let G = C 2 3 G=C_2^3 G = C 2 3 have basis a , b , c a,b,c a , b , c , put d = a + b + c d=a+b+c d = a + b + c , and set T = { a , b , c , d } T=\{a,b,c,d\} T = { a , b , c , d } . The character that sends each basis element to 1 / 2 + Z 1/2+\mathbb{Z} 1/2 + Z sends every member of T T T to 1 / 2 + Z 1/2+\mathbb{Z} 1/2 + Z . Lemma 5.1 therefore applies. Nevertheless, U = a b c d U=abcd U = ab c d is an atom with κ ( U ) = 2 \kappa(U)=2 κ ( U ) = 2 . Indeed, a zero-sum subsequence not containing d d d must be empty, while a zero-sum subsequence containing d d d must contain all of a , b , c a,b,c a , b , c . The two factorizations
U ⋅ U = ( a 2 ) ( b 2 ) ( c 2 ) ( d 2 ) U\cdot U=(a^2)(b^2)(c^2)(d^2) U ⋅ U = ( a 2 ) ( b 2 ) ( c 2 ) ( d 2 )
гրբեջ have lengths two and four. Thus T T T is not half-factorial. This is not a counterexample to the universal realization problem: the basis subset { a , b , c } \{a,b,c\} { a , b , c } is a half-factorial generating set of C 2 3 C_2^3 C 2 3 .
Half-factoriality of a chosen set cannot be transported through an arbitrary quotient or multiplication map. The next example is completely explicit.
Proposition 5.8. There is a finite abelian 2-group G G G with a half-factorial generating subset S S S such that 2 S 2S 2 S is a non-half-factorial generating subset of 2 G 2G 2 G . Consequently, neither homomorphic images of half-factorial subsets nor adjoining all their 2-multiples necessarily preserves half-factoriality.
Proof. Take G = C 4 3 G=C_4^3 G = C 4 3 with basis e 1 , e 2 , e 3 e_1,e_2,e_3 e 1 , e 2 , e 3 , put h = − e 1 − e 2 − e 3 h=-e_1-e_2-e_3 h = − e 1 − e 2 − e 3 , and let S = { e 1 , e 2 , e 3 , h } S=\{e_1,e_2,e_3,h\} S = { e 1 , e 2 , e 3 , h } . A sequence e 1 a 1 e 2 a 2 e 3 a 3 h b e_1^{a_1}e_2^{a_2}e_3^{a_3}h^b e 1 a 1 e 2 a 2 e 3 a 3 h b is a block precisely when a i ≡ b ( m o d 4 ) a_i\equiv b\pmod{4} a i ≡ b ( mod 4 ) for each i i i . An atom with any exponent at least four is one of the pure blocks e 1 4 , e 2 4 , e 3 4 , h 4 e_1^4,e_2^4,e_3^4,h^4 e 1 4 , e 2 4 , e 3 4 , h 4 . If all exponents are less than four, they must be equal; the only atom of this kind is e 1 e 2 e 3 h e_1e_2e_3h e 1 e 2 e 3 h . All atoms therefore have sequence length four, and S S S is half-factorial.
In 2 G ≅ C 2 3 2G\cong C_2^3 2 G ≅ C 2 3 , the elements 2 e 1 , 2 e 2 , 2 e 3 2e_1,2e_2,2e_3 2 e 1 , 2 e 2 , 2 e 3 are a basis and 2 h = 2 e 1 + 2 e 2 + 2 e 3 2h=2e_1+2e_2+2e_3 2 h = 2 e 1 + 2 e 2 + 2 e 3 . Thus 2 S 2S 2 S is exactly the configuration of Example 5.7. It is not half-factorial. Since any subset of a half-factorial set is half-factorial, the set S ∪ 2 S S\cup2S S ∪ 2 S cannot be half-factorial either. □
There is a separate obstruction to extending prescribed generators of a subgroup. The example for p = 2 p=2 p = 2 below appears in [8 p. 851, Remark 2] . The same construction works for every prime, with an explicit atom witnessing failure.
Proposition 5.9. Let p p p be a prime and let G = C p ⊕ C p 2 = ⟨ a ⟩ ⊕ ⟨ e ⟩ G=C_p\oplus C_{p^2}=\langle a\rangle\oplus\langle e\rangle G = C p ⊕ C p 2 = ⟨ a ⟩ ⊕ ⟨ e ⟩ , where ord ( a ) = p \operatorname{ord}(a)=p ord ( a ) = p and ord ( e ) = p 2 \operatorname{ord}(e)=p^2 ord ( e ) = p 2 . Put b = a + p e b=a+pe b = a + p e , H = G [ p ] H=G[p] H = G [ p ] , and T = { a , b } T=\{a,b\} T = { a , b } . Then T T T is a half-factorial generating set of the proper subgroup H H H , but no half-factorial subset of G G G containing T T T generates G G G . More precisely, for every x ∈ G ∖ H x\in G\setminus H x ∈ G ∖ H there is an atom of B ( T ∪ { x } ) \mathcal{B}(T\cup\{x\}) B ( T ∪ { x }) with cross number 1 + 1 / p 1+1/p 1 + 1/ p .
Proof. The elements a , b a,b a , b are independent of order p p p and generate H = ⟨ a , p e ⟩ H=\langle a,pe\rangle H = ⟨ a , p e ⟩ , so T T T is half-factorial. Let x ∈ G ∖ H x\in G\setminus H x ∈ G ∖ H . Then ord ( x ) = p 2 \operatorname{ord}(x)=p^2 ord ( x ) = p 2 and p x = j p e = j ( b − a ) px=jpe=j(b-a) p x = j p e = j ( b − a ) for some j ∈ { 1 , … , p − 1 } j\in\{1,\ldots,p-1\} j ∈ { 1 , … , p − 1 } . The sequence
W = x p a j b p − j (20) W=x^p a^j b^{p-j} \tag*{(20)} W = x p a j b p − j (20)
is a block. In a zero-sum subsequence the multiplicity of x x x is divisible by p p p , by projection to G / H G/H G / H . If it is zero, independence of a , b a,b a , b forces the subsequence to be empty. If it is p p p , the coefficients of a , b a,b a , b must be j , p − j j,p-j j , p − j , respectively. Thus W W W is an atom. Its cross number is p / p 2 + j / p + ( p − j ) / p = 1 + 1 / p p/p^2+j/p+(p-j)/p=1+1/p p / p 2 + j / p + ( p − j ) / p = 1 + 1/ p , so Proposition 2.2 excludes half-factoriality as soon as any x ∉ H x\notin H x ∈ / H is adjoined. □
By contrast, the union of a chain of nested half-factorial subsets is half-factorial: any two factorizations witnessing failure would involve only finitely many letters, all belonging to one member of the chain. This finite-character observation is standard [7 Lemma 3.1] . It does not provide the nested sets. Proposition 5.9 shows why compatibility at successive stages requires its own argument.
The next example strengthens the warning following Theorem 5.2 . Failure is not confined to taking too many generators: the weakly half-factorial set can be irredundant. A generating set is irredundant if none of its proper subsets generates the group.
Theorem 5.10. Let p p p be a prime, let m ≥ 2 m\ge2 m ≥ 2 , and put r = m p r=mp r = m p . Let
E = ⨁ i = 1 r ⟨ e i ⟩ , ord ( e i ) = p 2 , G = E / p ( e 1 + ⋯ + e r ) . E=\bigoplus_{i=1}^{r}\langle e_i\rangle,\qquad\operatorname{ord}(e_i)=p^2,\qquad G=E/p(e_1+\cdots+e_r). E = i = 1 ⨁ r ⟨ e i ⟩ , ord ( e i ) = p 2 , G = E / p ( e 1 + ⋯ + e r ) . Write t i t_i t i for the image of e i e_i e i and let T = { t 1 , … , t r } T=\{t_1,\ldots,t_r\} T = { t 1 , … , t r } . Then T T T is an irredundant weakly half-factorial generating set of G G G , but contains no half-factorial generating subset. More precisely,
every t i t_i t i has order p 2 p^2 p 2 , the atoms over T T T are exactly
t 1 p 2 , … , t r p 2 , U = ∏ i = 1 r t i p , (21) t_1^{p^2},\ldots,t_r^{p^2},\qquad U=\prod_{i=1}^{r}t_i^p, \tag*{(21)} t 1 p 2 , … , t r p 2 , U = i = 1 ∏ r t i p , (21)
and k ( U ) = m \mathsf{k}(U)=m k ( U ) = m . For every integer q ≥ 1 q\ge1 q ≥ 1 ,
L B ( T ) ( U p q ) = { p q + ( r − p ) j : 0 ≤ j ≤ q } . (22) \mathsf{L}_{\mathcal{B}(T)}(U^{pq})=\{pq+(r-p)j:0\le j\le q\}. \tag*{(22)} L B ( T ) ( U pq ) = { pq + ( r − p ) j : 0 ≤ j ≤ q } . (22)
Nevertheless, G ≅ C p ⊕ C p 2 r − 1 G\cong C_p\oplus C_{p^2}^{r-1} G ≅ C p ⊕ C p 2 r − 1 , so G G G itself has a half-factorial generating set.
Proof. The homomorphism on E E E sending every e i e_i e i to 1 / p 2 + Z 1/p^2+\mathbb{Z} 1/ p 2 + Z annihilates p ( e 1 + ⋯ + e r ) p(e_1+\cdots+e_r) p ( e 1 + ⋯ + e r ) , whose image is r / p + Z = m + Z r/p+\mathbb{Z}=m+\mathbb{Z} r / p + Z = m + Z . It therefore descends to a character χ \chi χ of G G G . In particular, ord ( t i ) = p 2 \operatorname{ord}(t_i)=p^2 ord ( t i ) = p 2 and χ ( t i ) = 1 / ord ( t i ) + Z \chi(t_i)=1/\operatorname{ord}(t_i)+\mathbb{Z} χ ( t i ) = 1/ ord ( t i ) + Z . The t i t_i t i are distinct: a relation e i − e j e_i-e_j e i − e j cannot be a multiple of p ( e 1 + ⋯ + e r ) p(e_1+\cdots+e_r) p ( e 1 + ⋯ + e r ) , as is already clear after reduction modulo p E pE pE . Lemma 5.1 proves weak half-factoriality.
The defining relation lies in p E pE pE , so G / p G ≅ E / p E ≅ C p r G/pG\cong E/pE\cong C_p^r G / pG ≅ E / pE ≅ C p r , with the images of the t i t_i t i forming a basis. Consequently no proper subset of T T T generates G G G .
A sequence ∏ i t i v i \prod_i t_i^{v_i} ∏ i t i v i is a block precisely when, for one j ∈ { 0 , … , p − 1 } j\in\{0,\ldots,p-1\} j ∈ { 0 , … , p − 1 } , all its multiplicities satisfy v i ≡ j p ( m o d p 2 ) v_i\equiv jp\pmod{p^2} v i ≡ j p ( mod p 2 ) . If any multiplicity is at least p 2 p^2 p 2 , an atom must be the corresponding pure block. Otherwise all multiplicities equal j p jp j p . A nonempty such block is U j U^j U j , and it is an atom exactly when j = 1 j=1 j = 1 . This proves (5.6). In particular, k ( U ) = r p / p 2 = m ≠ 1 \mathsf{k}(U)=rp/p^2=m\ne1 k ( U ) = r p / p 2 = m = 1 , so T T T is not half-factorial. Together with irredundancy, this proves the assertion about all its subsets.
A factorization of U p q U^{pq} U pq uses some number a a a of copies of U U U and numbers b i b_i b i of the pure atoms. Comparing multiplicities gives p a + p 2 b i = p 2 q pa+p^2b_i=p^2q p a + p 2 b i = p 2 q for every i i i . Thus a = p c a=pc a = p c and b i = q − c b_i=q-c b i = q − c with 0 ≤ c ≤ q 0\le c\le q 0 ≤ c ≤ q . The factorization length is p c + r ( q − c ) pc+r(q-c) p c + r ( q − c ) , giving (5.7). For example, U p U^p U p has lengths p p p and r = m p r=mp r = m p .
Finally, e 1 + ⋯ + e r , e 2 , … , e r e_1+\cdots+e_r,e_2,\ldots,e_r e 1 + ⋯ + e r , e 2 , … , e r is a basis of the free Z / p 2 Z \mathbb{Z}/p^2\mathbb{Z} Z / p 2 Z -module E E E . In this basis the only additional relation is that p p p times its first member is zero. This gives the asserted cyclic decomposition, whose standard generators are half-factorial. □
This phenomenon persists among reduced uncountable unbounded p p p -groups. For any uncountable cardinal κ \kappa κ , take the direct sum of κ \kappa κ copies of the group in Theorem 5.10, together with ⨁ n ≥ 1 C p n \bigoplus_{n\ge1}C_{p^n} ⨁ n ≥ 1 C p n . The union of the displayed weak generating sets and the cyclic generators is irredundant and weakly half-factorial, but has no half-factorial generating subset. The group is reduced, uncountable, and unbounded. It is nonetheless a direct sum of cyclic groups and belongs to H \mathcal{H} H . Thus the obstruction to thinning is not a counterexample to the universal primary-group realization problem.
The full character set in the torsion-completion
One might instead keep all character-compatible elements, so that no useful generator is discarded. The following elementary construction shows why this also fails in groups with sufficiently large socle. Its conclusion concerns the full set in (5.2), not every generating subset of that set.
Proposition 5.11. Let p p p be a prime, let P P P be an abelian p p p -group, and let χ : P → Q / Z \chi:P\to\mathbb{Q}/\mathbb{Z} χ : P → Q / Z be a homomorphism with χ ( P [ p ] ) ≠ 0 \chi(P[p])\ne0 χ ( P [ p ]) = 0 . If dim F p ( P [ p ] ∩ ker χ ) \dim_{\mathbb{F}_p}(P[p]\cap\ker\chi) dim F p ( P [ p ] ∩ ker χ ) is infinite, then, for every integer m ≥ 1 m\ge1 m ≥ 1 , the block monoid B ( W χ ( P ) ) \mathcal{B}(W_\chi(P)) B ( W χ ( P )) has an atom of cross number m m m . For each m ≥ 2 m\ge2 m ≥ 2 it has an element whose set of lengths contains both p p p and m p mp m p . In particular, the ratios of maximal to minimal factorization lengths in this block monoid are unbounded.
Proof. Choose c ∈ P [ p ] c\in P[p] c ∈ P [ p ] with χ ( c ) = 1 / p + Z \chi(c)=1/p+\mathbb{Z} χ ( c ) = 1/ p + Z . The block c p c^p c p is an atom of cross number one. For m ≥ 2 m\ge2 m ≥ 2 , put r = m p − 1 r=mp-1 r = m p − 1 and choose independent elements a 1 , … , a r ∈ P [ p ] ∩ ker χ a_1,\ldots,a_r\in P[p]\cap\ker\chi a 1 , … , a r ∈ P [ p ] ∩ ker χ . Define
s 0 = c − ∑ i = 1 r a i , s i = c + a i ( 1 ≤ i ≤ r ) . s_0=c-\sum_{i=1}^{r}a_i,\qquad s_i=c+a_i\quad(1\le i\le r). s 0 = c − i = 1 ∑ r a i , s i = c + a i ( 1 ≤ i ≤ r ) .
The elements c , a 1 , … , a r c,a_1,\ldots,a_r c , a 1 , … , a r are independent over F p \mathbb{F}_p F p . Hence the s i s_i s i are distinct elements of order p p p with character value 1 / p + Z 1/p+\mathbb{Z} 1/ p + Z , and all belong to W χ ( P ) W_\chi(P) W χ ( P ) . The sequence U = s 0 s 1 ⋯ s r U=s_0s_1\cdots s_r U = s 0 s 1 ⋯ s r has sum ( r + 1 ) c = m p c = 0 (r+1)c=mpc=0 ( r + 1 ) c = m p c = 0 .
If a subsequence of U U U has multiplicities ε i ∈ { 0 , 1 } \varepsilon_i\in\{0,1\} ε i ∈ { 0 , 1 } and sum zero, its a i a_i a i -coordinate forces ε i = ε 0 \varepsilon_i=\varepsilon_0 ε i = ε 0 for each i i i . Therefore that subsequence is empty or is all of U U U . Thus U U U is an atom with k ( U ) = ( r + 1 ) / p = m k(U)=(r+1)/p=m k ( U ) = ( r + 1 ) / p = m . Finally,
U p = ∏ i = 0 r s i p U^p=\prod_{i=0}^{r}s_i^p U p = i = 0 ∏ r s i p
has a factorization of length p p p into copies of U U U and a factorization of length r + 1 = m p r+1=mp r + 1 = m p into pure atoms. Their length ratio is m m m , which proves the last assertion. □ \square □
For the torsion-completion P P P in Proposition 4.10 , this applies to every character for which W χ ( P ) W_\chi(P) W χ ( P ) generates P P P . Indeed, such a character is nonzero on P [ p ] P[p] P [ p ] : if t ∈ W χ ( P ) t\in W_\chi(P) t ∈ W χ ( P ) has order p n p^n p n , then χ ( p n − 1 t ) = 1 / p + Z \chi(p^{n-1}t)=1/p+\mathbb{Z} χ ( p n − 1 t ) = 1/ p + Z . Also, P [ p ] = ∏ n ≥ 1 C n [ p ] P[p]=\prod_{n\geq1}C_n[p] P [ p ] = ∏ n ≥ 1 C n [ p ] is infinite-dimensional, so the kernel of its nonzero functional to the subgroup of order p p p has infinite dimension. Thus the full character set always has atoms of arbitrarily large integral cross number in this test case. This proves the failure of the full-set construction; nonexistence of every half-factorial generating subset follows instead from Theorem 3.6 .
The logical implications can now be summarized without confusing a property of a specified set with existence of some suitable set.
Figure 1. Implications for an abelian p p p -group P P P . The equivalence is classical; the left downward implication requires the displayed separatedness hypothesis. Weakly half-factorial generators always exist, but they do not imply half-factorial realizability. No converse to the cardinality implication is asserted.
Socle compatibility and repeated contraction
We finish with two structural restrictions on chosen generating sets. At the prime two, containing a socle basis forces compatibility with the entire finite-height filtration. Under independence modulo p p p , one contraction preserves every set of lengths, but this property need not persist under iteration. These results provide finer local information than the global cardinality obstruction.
A socle-filtration obstruction at the prime two
For an abelian 2-group P P P , its socle P [ 2 ] P[2] P [ 2 ] is a vector space over F 2 \mathbb{F}_2 F 2 . A subset of P [ 2 ] ∖ { 0 } P[2]\setminus\{0\} P [ 2 ] ∖ { 0 } is half-factorial precisely when it is linearly independent. Indeed, independence leaves only the pure atoms e 2 e^2 e 2 . A dependent subset has a finite minimally dependent subset of at least three distinct elements, whose product is an atom of cross number at least 3 / 2 3/2 3/2 .
The next estimate bounds the socle support of a sequence by its cross number. It applies to arbitrary sequences, not only to blocks.
Lemma 6.1. Let P P P be an abelian 2 2 2 -group, let S ⊆ P ∖ { 0 } S \subseteq P \setminus\{0\} S ⊆ P ∖ { 0 } be half-factorial, and suppose that S S S contains a vector-space basis E E E of P [ 2 ] P[2] P [ 2 ] . For v ∈ P [ 2 ] v \in P[2] v ∈ P [ 2 ] , let d E ( v ) d_E(v) d E ( v ) be the number of nonzero coordinates of v v v in the basis E E E . If A ∈ F ( S ) A \in\mathcal{F}(S) A ∈ F ( S ) and σ ( A ) ∈ P [ 2 ] \sigma(A) \in P[2] σ ( A ) ∈ P [ 2 ] , then
d E ( σ ( A ) ) ≤ 2 k ( A ) . (23) d_E(\sigma(A)) \leq2k(A). \tag*{(23)} d E ( σ ( A )) ≤ 2 k ( A ) . (23) In particular, if k ( A ) = 1 / 2 k(A) = 1/2 k ( A ) = 1/2 and σ ( A ) ∈ P [ 2 ] ∖ { 0 } \sigma(A) \in P[2] \setminus\{0\} σ ( A ) ∈ P [ 2 ] ∖ { 0 } , then σ ( A ) ∈ E \sigma(A) \in E σ ( A ) ∈ E .
Proof. Write σ ( A ) = e 1 + ⋯ + e d \sigma(A) = e_1 + \cdots+ e_d σ ( A ) = e 1 + ⋯ + e d with distinct e i ∈ E e_i \in E e i ∈ E , and put C = e 1 ⋯ e d C = e_1 \cdots e_d C = e 1 ⋯ e d . Since each e i e_i e i has order two, A C AC A C is a block. Mark the d d d occurrences belonging to C C C and factor A C AC A C into atoms. No atom contains two marked occurrences: they are distinct letters of cross number 1 / 2 1/2 1/2 each, so two of them already have cross number one and do not sum to zero. Any additional letter would give cross number greater than one, contrary to Proposition 2.2.
Consequently the factorization has at least d d d atoms. Its length is k ( A C ) = k ( A ) + d / 2 k(AC) = k(A) + d/2 k ( A C ) = k ( A ) + d /2 , again by Proposition 2.2. Thus k ( A ) + d / 2 ≥ d k(A) + d/2 \geq d k ( A ) + d /2 ≥ d , which proves (6.1). The final assertion follows because a nonzero socle element has positive support size. □
For example, if s ∈ S s \in S s ∈ S has order 2 r 2^r 2 r , applying the lemma to s 2 r − 1 s^{2^{r-1}} s 2 r − 1 shows that 2 r − 1 s ∈ E 2^{r-1}s \in E 2 r − 1 s ∈ E . The same observation applies to a sequence A A A with k ( A ) = 2 − r k(A) = 2^{-r} k ( A ) = 2 − r whenever σ ( A ) \sigma(A) σ ( A ) has order 2 r 2^r 2 r . This allows a carrying procedure that respects the socle basis.
Theorem 6.2. Let P P P be an abelian 2 2 2 -group and let S ⊆ P ∖ { 0 } S \subseteq P \setminus\{0\} S ⊆ P ∖ { 0 } be a half-factorial generating set. Suppose that E ⊆ S E \subseteq S E ⊆ S is a vector-space basis of P [ 2 ] P[2] P [ 2 ] . Put V n = ( 2 n P ) ∩ P [ 2 ] V_n = (2^nP) \cap P[2] V n = ( 2 n P ) ∩ P [ 2 ] for n ≥ 0 n \geq0 n ≥ 0 . Then
V n = span F 2 ( E ∩ V n ) ( n ≥ 0 ) . (24) V_n = \operatorname{span}_{\mathbb{F}_2}(E \cap V_n) \qquad(n \geq0). \tag*{(24)} V n = span F 2 ( E ∩ V n ) ( n ≥ 0 ) . (24) Thus the same basis E E E must be adapted to every finite-height subspace of the socle.
Proof. Step 1: establish the invariants for a carrying procedure. We first record an elementary consequence of half-factoriality. If a nonempty sequence A A A has cross number 2 − r 2^{-r} 2 − r , with r ≥ 1 r \geq1 r ≥ 1 , and 2 r σ ( A ) = 0 2^r\sigma(A) = 0 2 r σ ( A ) = 0 , then σ ( A ) \sigma(A) σ ( A ) has order exactly 2 r 2^r 2 r . Otherwise, for its order d < 2 r d < 2^r d < 2 r , the nonempty block A d A^d A d would have cross number d / 2 r < 1 d/2^r < 1 d / 2 r < 1 , which is impossible. When its order is 2 r 2^r 2 r , Lemma 6.1, applied to A 2 r − 1 A^{2^{r-1}} A 2 r − 1 , gives 2 r − 1 σ ( A ) ∈ E 2^{r-1}\sigma(A) \in E 2 r − 1 σ ( A ) ∈ E .
Fix n ≥ 0 n \geq0 n ≥ 0 and v ∈ V n ∖ { 0 } v \in V_n \setminus\{0\} v ∈ V n ∖ { 0 } . Choose z ∈ P z \in P z ∈ P with 2 n z = v 2^n z = v 2 n z = v ; then 2 n + 1 z = 0 2^{n+1}z = 0 2 n + 1 z = 0 . Since S S S generates P P P as a monoid, choose a sequence W ∈ F ( S ) W \in\mathcal{F}(S) W ∈ F ( S ) with σ ( W ) = z \sigma(W) = z σ ( W ) = z . Initially partition W W W into its individual letters. Each piece A A A of this partition satisfies, for some r ≥ 1 r \geq1 r ≥ 1 ,
k ( A ) = 2 − r , ord ( σ ( A ) ) = 2 r , 2 r − 1 σ ( A ) ∈ E . (25) k(A) = 2^{-r}, \qquad\operatorname{ord}(\sigma(A)) = 2^r, \qquad2^{r-1}\sigma(A) \in E. \tag*{(25)} k ( A ) = 2 − r , ord ( σ ( A )) = 2 r , 2 r − 1 σ ( A ) ∈ E . (25)
The last condition follows from Lemma 6.1.
Step 2: lower the largest order by pairing. Suppose that the largest exponent r r r among the pieces exceeds n + 1 n+1 n + 1 . Multiplying the sum of all pieces by 2 r − 1 2^{r-1} 2 r − 1 gives zero. Pieces of smaller order disappear, while each piece of order 2 r 2^r 2 r contributes its indicated basis element of E E E . Independence of E E E therefore shows that, for each e ∈ E e \in E e ∈ E , an even number of the order-2 r 2^r 2 r pieces have top multiple e e e . Pair those pieces within each such group.
If A A A and A ′ A' A ′ are paired, then A A ′ AA' A A ′ has cross number 2 − ( r − 1 ) 2^{-(r-1)} 2 − ( r − 1 ) and its sum is annihilated by 2 r − 1 2^{r-1} 2 r − 1 , since the two top multiples are equal. As r − 1 ≥ 1 r-1 \geq1 r − 1 ≥ 1 , the observation at the start of the proof shows that its sum has order exactly 2 r − 1 2^{r-1} 2 r − 1 ; the same observation and the socle support lemma establish the last condition of (6.3) for the new piece. Replace every pair by its concatenation. The largest exponent decreases. Repetition therefore yields a finite partition of W W W into pieces satisfying (6.3) with r ≤ n + 1 r \leq n+1 r ≤ n + 1 .
Step 3: recover the required socle expansion. Finally multiply their sums by 2 n 2^n 2 n . The pieces with r < n + 1 r < n+1 r < n + 1 vanish. Every remaining piece contributes a member of E E E that belongs to 2 n P 2^nP 2 n P , hence to E ∩ V n E \cap V_n E ∩ V n . Their sum is v v v . This proves the nontrivial inclusion in (6.2); the reverse inclusion is immediate. □
The theorem gives an obstruction when the socle filtration has small successive quotients but a large underlying vector space. Indeed, if ⋂ n ≥ 0 V n = 0 \bigcap_{n \geq0} V_n = 0 ⋂ n ≥ 0 V n = 0 , then every e ∈ E e \in E e ∈ E lies in exactly one set E ∩ ( V n ∖ V n + 1 ) E \cap(V_n \setminus V_{n+1}) E ∩ ( V n ∖ V n + 1 ) . The images of this set form a basis of V n / V n + 1 V_n/V_{n+1} V n / V n + 1 by (24) . In particular, countable-dimensional successive quotients would force E E E to be countable.
Corollary 6.3. Let
P = t ( ∏ j ≥ 1 Z / 2 j Z ) , P=t\left(\prod_{j\geq1}\mathbb{Z}/2^j\mathbb{Z}\right), P = t ( j ≥ 1 ∏ Z / 2 j Z ) , where t ( − ) t(-) t ( − ) denotes the torsion subgroup. There is no half-factorial generating subset S ⊆ P ∖ { 0 } S\subseteq P\setminus\{0\} S ⊆ P ∖ { 0 } for which S ∩ P [ 2 ] S\cap P[2] S ∩ P [ 2 ] generates P [ 2 ] P[2] P [ 2 ] . Consequently, there are no subsets T n ⊆ P [ 2 n ] ∖ { 0 } T_n\subseteq P[2^n]\setminus\{0\} T n ⊆ P [ 2 n ] ∖ { 0 } , n ≥ 1 n\geq1 n ≥ 1 , such that all T n T_n T n are half-factorial, all T n T_n T n generate P [ 2 n ] P[2^n] P [ 2 n ] , and T n ⊆ T n + 1 T_n\subseteq T_{n+1} T n ⊆ T n + 1 for every n n n .
Proof. Identify P [ 2 ] P[2] P [ 2 ] with ∏ j ≥ 1 F 2 \prod_{j\geq1}\mathbb{F}_2 ∏ j ≥ 1 F 2 by writing its j j j th coordinate as ε j 2 j − 1 + 2 j Z \varepsilon_j2^{j-1}+2^j\mathbb{Z} ε j 2 j − 1 + 2 j Z . For V n = ( 2 n P ) ∩ P [ 2 ] V_n=(2^nP)\cap P[2] V n = ( 2 n P ) ∩ P [ 2 ] one has
V n = { ( ε j ) j : ε 1 = ⋯ = ε n = 0 } . (26) V_n=\{(\varepsilon_j)_j:\varepsilon_1=\cdots=\varepsilon_n=0\}. \tag*{(26)} V n = {( ε j ) j : ε 1 = ⋯ = ε n = 0 } . (26) The forward inclusion is immediate. For the reverse inclusion, choose the j j j th coordinate of a 2 n 2^n 2 n th root to be ε j 2 j − n − 1 + 2 j Z \varepsilon_j2^{j-n-1}+2^j\mathbb{Z} ε j 2 j − n − 1 + 2 j Z when j > n j>n j > n , and zero when j ≤ n j\leq n j ≤ n . This root is annihilated by 2 n + 1 2^{n+1} 2 n + 1 and therefore belongs to P P P .
It follows that ⋂ n V n = 0 \bigcap_nV_n=0 ⋂ n V n = 0 and dim F 2 ( V n / V n + 1 ) = 1 \dim_{\mathbb{F}_2}(V_n/V_{n+1})=1 dim F 2 ( V n / V n + 1 ) = 1 for every n ≥ 0 n\geq0 n ≥ 0 . On the other hand, P [ 2 ] = ∏ j ≥ 1 F 2 P[2]=\prod_{j\geq1}\mathbb{F}_2 P [ 2 ] = ∏ j ≥ 1 F 2 is uncountable and consequently has no countable vector-space basis. Suppose that S S S as in the first assertion existed. The subset E = S ∩ P [ 2 ] E=S\cap P[2] E = S ∩ P [ 2 ] would be half-factorial and would generate the socle. By the observation at the start of this subsection it would be a basis of P [ 2 ] P[2] P [ 2 ] . Theorem 6.2 and (26) imply that E ∩ ( V n ∖ V n + 1 ) E\cap(V_n\setminus V_{n+1}) E ∩ ( V n ∖ V n + 1 ) has exactly one member for each n n n . Their union is E E E , since ⋂ n V n = 0 \bigcap_nV_n=0 ⋂ n V n = 0 . This makes E E E countable, a contradiction.
For the last assertion, the union S = ⋃ n T n S=\bigcup_nT_n S = ⋃ n T n would be half-factorial by finite character and would generate P = ⋃ n P [ 2 n ] P=\bigcup_nP[2^n] P = ⋃ n P [ 2 n ] . Its intersection with P [ 2 ] P[2] P [ 2 ] contains T 1 T_1 T 1 and hence generates the socle, contradicting the first assertion. □ \square □
The conclusion is stronger than Example 5.6: that example exhibits unsuccessful choices, whereas Corollary 6.3 excludes every nested choice in the indicated group. The socle argument by itself proves only a conditional obstruction; the unconditional nonrealization result is Theorem 3.6 . A global generating set need not generate the socle using its order-two members. The next elementary example makes the distinction explicit.
Example 6.4. Let G = C 2 ⊕ C 4 = ⟨ a ⟩ ⊕ ⟨ z ⟩ G=C_2\oplus C_4=\langle a\rangle\oplus\langle z\rangle G = C 2 ⊕ C 4 = ⟨ a ⟩ ⊕ ⟨ z ⟩ , where ord ( a ) = 2 \operatorname{ord}(a)=2 ord ( a ) = 2 and ord ( z ) = 4 \operatorname{ord}(z)=4 ord ( z ) = 4 , and put S = { z , z + a } S=\{z,z+a\} S = { z , z + a } . This is a generating set and both its elements have order four. Its block atoms are exactly
z 4 , ( z + a ) 4 , z 2 ( z + a ) 2 . z^4,\qquad(z+a)^4,\qquad z^2(z+a)^2. z 4 , ( z + a ) 4 , z 2 ( z + a ) 2 . Indeed, a sequence z u ( z + a ) v z^u(z+a)^v z u ( z + a ) v is a block precisely when v v v is even and u + v u+v u + v is divisible by four, from which the list follows. Every listed atom has cross number one. Thus S S S is half-factorial, although S ∩ G [ 2 ] = ∅ S\cap G[2]=\varnothing S ∩ G [ 2 ] = ∅ and G [ 2 ] ≠ 0 G[2]\neq0 G [ 2 ] = 0 . Therefore the socle-generation hypothesis in Corollary 6.3 cannot be inferred merely from half-factoriality and generation of the whole group.
Contraction under independence modulo a prime
A different possible approach is to contract generators repeatedly by multiplication by p p p . The following criterion explains when one contraction preserves all factorization lengths. It does not assert preservation for an arbitrary generating set.
Proposition 6.5. Let p p p be a prime, let P P P be an abelian p p p -group, and let S ⊆ P ∖ { 0 } S\subseteq P\setminus\{0\} S ⊆ P ∖ { 0 } generate P P P . Suppose that the images of the members of S S S are linearly independent in P / p P P/pP P / pP . Every B ∈ B ( S ) B\in\mathcal{B}(S) B ∈ B ( S ) then has all its multiplicities divisible by p p p , and the map
θ : B ( S ) ⟶ B ( p S ) , ∏ s ∈ S s p a s ⟼ ∏ s ∈ S ( p s ) a s \theta:\mathcal{B}(S)\longrightarrow\mathcal{B}(pS),\qquad\prod_{s\in S}s^{p a_s}\longmapsto\prod_{s\in S}(ps)^{a_s} θ : B ( S ) ⟶ B ( pS ) , s ∈ S ∏ s p a s ⟼ s ∈ S ∏ ( p s ) a s
is a surjective homomorphism preserving every set of factorization lengths. Here coincident letters on the right are combined, and 0 is allowed in p S pS pS . In particular, S S S is half-factorial if and only if p S pS pS is half-factorial.
Proof. Reducing a zero-sum relation modulo p P pP pP and using the stated independence shows that each multiplicity is divisible by p p p . The displayed substitution consequently defines a homomorphism into B ( p S ) \mathcal{B}(pS) B ( pS ) . Given a block over p S pS pS , choose a preimage in S S S for each occurrence and replace that occurrence by p p p copies of its preimage. The resulting sequence is a block over S S S , proving surjectivity.
For a given B ∈ B ( S ) B \in\mathcal{B}(S) B ∈ B ( S ) , group its occurrences of each s s s into packets of p p p identical letters, retaining their labels in θ ( B ) \theta(B) θ ( B ) . Any factorization of θ ( B ) \theta(B) θ ( B ) can be lifted by distributing those labelled packets among the factors. Conversely, if the image of an atom factored, this lifting procedure would factor the original atom. Thus atoms map to atoms, and every factorization of the image lifts to a factorization with the same length. This proves the asserted equality of length sets. Adding or deleting the letter zero does not affect half-factoriality. □
The independence needed for the next contraction can disappear. In fact, arbitrarily many successful contractions can precede the first failure, even when the original generating set is irredundant and the group is finite.
Theorem 6.6. Let p p p be a prime and r ≥ 1 r \ge1 r ≥ 1 . Put k = p 2 k=p^2 k = p 2 , N = r + 2 N=r+2 N = r + 2 , and
G = ( ⨁ i = 1 k C p N e i ) / ⟨ p r ( e 1 + ⋯ + e k ) ⟩ . G=\left(\bigoplus_{i=1}^{k} C_{p^N}e_i\right)\big/\left\langle p^r(e_1+\cdots+e_k)\right\rangle. G = ( i = 1 ⨁ k C p N e i ) / ⟨ p r ( e 1 + ⋯ + e k ) ⟩ . Let t i t_i t i be the image of e i e_i e i and set S = { t 1 , … , t k } S=\{t_1,\ldots,t_k\} S = { t 1 , … , t k } . Then S S S is an irredundant half-factorial generating set of G G G , and p j S p^jS p j S is half-factorial for every 0 ≤ j ≤ r 0\le j\le r 0 ≤ j ≤ r , but p r + 1 S p^{r+1}S p r + 1 S is not half-factorial. More precisely, ∏ i = 1 k ( p r + 1 t i ) \prod_{i=1}^{k}(p^{r+1}t_i) ∏ i = 1 k ( p r + 1 t i ) is an atom of cross number p p p . Moreover, G ≅ C p r ⊕ C p r + 2 p 2 − 1 G\cong C_{p^r}\oplus C_{p^{r+2}}^{p^2-1} G ≅ C p r ⊕ C p r + 2 p 2 − 1 .
Proof. Each t i t_i t i has order p N p^N p N . Indeed, a multiple of e i e_i e i can belong to the displayed relation subgroup only if all the other coordinates of that relation vanish, in which case its i i i th coordinate also vanishes. The relation vector belongs to p p p times the original direct sum, so the images of the t i t_i t i form a basis of G / p G G/pG G / pG . In particular, S S S is irredundant.
For 0 ≤ j ≤ r + 1 0\le j\le r+1 0 ≤ j ≤ r + 1 put x i = p j t i x_i=p^jt_i x i = p j t i , q = p N − j q=p^{N-j} q = p N − j , and d = p max { r − j , 0 } d=p^{\max\{r-j,0\}} d = p m a x { r − j , 0 } . The x i x_i x i are nonzero and distinct, and all have order q q q . Direct inspection of the relation subgroup shows that ∏ i x i a i \prod_i x_i^{a_i} ∏ i x i a i is a block precisely when all a i a_i a i have the same residue modulo q q q and that common residue is divisible by d d d . Therefore its atoms are exactly the pure blocks x i q x_i^q x i q and the diagonal block
D j = ( x 1 ⋯ x k ) d . D_j=(x_1\cdots x_k)^d. D j = ( x 1 ⋯ x k ) d .
To verify completeness of this list, an atom with an exponent at least q q q is a pure block. Otherwise all exponents are the same integer ℓ d \ell d ℓ d in [ 0 , q − 1 ] [0,q-1] [ 0 , q − 1 ] , and a nonempty such block is a power of D j D_j D j .
If j ≤ r j\le r j ≤ r , then k ( D j ) = k d / q = 1 k(D_j)=kd/q=1 k ( D j ) = k d / q = 1 , so the cross-number criterion proves half-factoriality. If j = r + 1 j=r+1 j = r + 1 , then q = p q=p q = p , d = 1 d=1 d = 1 , and k ( D j ) = k / p = p > 1 k(D_j)=k/p=p>1 k ( D j ) = k / p = p > 1 . For an explicit factorization witness,
D r + 1 p = ∏ i = 1 p 2 x i p . D_{r+1}^{p}=\prod_{i=1}^{p^2}x_i^p. D r + 1 p = i = 1 ∏ p 2 x i p .
has lengths p p p and p 2 p^2 p 2 . Finally, replace e k e_k e k in the standard basis of ⨁ i C p N e i \bigoplus_i C_{p^N}e_i ⨁ i C p N e i by e 1 + ⋯ + e k e_1+\cdots+e_k e 1 + ⋯ + e k . The relation then affects only that cyclic summand, reducing its order to p r p^r p r and giving the stated group decomposition. □
Thus one-step preservation under the hypotheses of Proposition 6.5 does not justify an unrestricted induction. In particular, successful testing through any prescribed finite number of contractions supplies no uniform guarantee for the next contraction. This obstruction, like the socle obstruction above, delimits a construction method rather than producing a group with no half-factorial generating set.
Concluding remarks
The separation theorem answers the universal Dedekind realization problem negatively by controlling arbitrary global generators. For each prime p p p , the torsion-completion P p P_p P p has countable finite power quotients but is uncountable. Half-factorial generators would be a countable union of sets injecting into these quotients, which is impossible. The construction of pure subgroups shows that the obstruction already occurs at the least uncountable cardinal.
Several positive facts remain compatible with this negative result. Every torsion abelian group has weakly half-factorial generators; every bounded subgroup has half-factorial generators; and the completion is a pure extension with both kernel and quotient in H \mathcal{H} H . None of these statements ensures a single global half-factorial generating set. The failure of closure under pure extensions and countable increasing unions makes the distinction an algebraic one, not merely a failure of a particular proof method.
The cardinal bound is a necessary condition for p p p -adically separated groups. Outside that hypothesis the proof does not apply, since some nonzero top multiples may belong to every p r P p^rP p r P . Even within the separated class, the bound alone is not proved sufficient. A structural characterization of the groups belonging to H \mathcal{H} H therefore remains a separate problem. The splitting, character, socle, and contraction results offer restrictions and constructions for that classification without being needed for the negative answer established in § 3 .
Competing interests. The authors declare that they have no competing interests.
Availability of data and material: No datasets were generated or analysed during the current study.
Use of large language models, AI, and machine learning tools. During the preparation of this work the authors used ChatGPT, an AI assistant developed by OpenAI, to explore various mathematical concepts related to the topics of this paper and to verify the English grammar. After using this tool/service, the author reviewed and edited the content as needed and take full responsibility for the content of the published article.
Funding. H. Kim was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Education (grant no. RS-2026-25589739). Department of Computer Engineering, Hoseo University, Republic of Korea
Email address: hkkim@hoseo.edu