Structure

Finitely generated groups are not closed under ℵ1\aleph_1-cofiltered limits

We construct a diagram Δ:ω1op→Grp\Delta : \omega_1^{\op} \to \Grp in which every group is isomorphic to a fixed finitely generated group GG, but whose limit is uncountable, and hence not finitely generated. Here, ω1\omega_1 denotes the poset of countable ordinals.

Throughout, we fix a non-trivial finitely generated group GG with G≅G×GG \cong G \times G. Such a group has been constructed by Tyrer Jones (1974), Theorem B.

Definition. A good pair (p,s):A⇄B(p,s) : A \rightleftarrows B consists of two homomorphisms p:A→Bp : A \to B and s:B→As : B \to A such that

  • BB and ker⁡(p)\ker(p) are isomorphic to GG,
  • p∘s=idBp \circ s = \id_B,
  • the subgroups s(B)s(B) and ker⁡(p)\ker(p) of AA commute elementwise.

In this case, AA is the internal direct product of s(B)s(B) and ker⁡(p)\ker(p), which we indicate by writing A=s(B)×ker⁡(p)A = s(B) \times \ker(p). In particular, A≅B×ker⁡(p)≅G×G≅GA \cong B \times \ker(p) \cong G \times G \cong G.

Lemma 1.

Let BB be a group with B≅GB \cong G. Let p:B×G→Bp : B \times G \to B be the projection defined by p(b,g)≔bp(b,g) \coloneqq b and s:B→B×Gs : B \to B \times G be the inclusion defined by s(b)≔(b,1)s(b) \coloneqq (b,1). Then (p,s):B×G⇄B(p,s) : B \times G \rightleftarrows B is a good pair.

Proof. We have ker⁡(p)=1×G≅G\ker(p) = 1 \times G \cong G. Clearly, p∘s=idBp \circ s = \id_B, and s(B)=B×1s(B) = B \times 1 commutes with 1×G1 \times G. □\square

Lemma 2.

Good pairs compose: If (p,s):A⇄B(p,s) : A \rightleftarrows B and (q,t):B⇄C(q,t) : B \rightleftarrows C are good pairs, then (q∘p,s∘t):A⇄C(q \circ p, s \circ t) : A \rightleftarrows C is a good pair.

Proof. We have q∘p∘s∘t=q∘t=idCq \circ p \circ s \circ t = q \circ t = \id_C.

Next, we compute the kernel of q∘pq \circ p. By the internal direct product decomposition A=s(B)×ker⁡(p)A = s(B) \times \ker(p), every a∈Aa \in A can be written uniquely as a=s(b)ka = s(b) k with b∈Bb \in B and k∈ker⁡(p)k \in \ker(p), namely b=p(a)b = p(a). Then q(p(a))=q(b)q(p(a)) = q(b), so a∈ker⁡(q∘p)a \in \ker(q \circ p) if and only if b∈ker⁡(q)b \in \ker(q). This shows ker⁡(q∘p)=s(ker⁡(q))⋅ker⁡(p)\ker(q \circ p) = s(\ker(q)) \cdot \ker(p). Since s(ker⁡(q))⊆s(B)s(\ker(q)) \subseteq s(B), the two factors commute elementwise and intersect trivially. Therefore, ker⁡(q∘p)≅s(ker⁡(q))×ker⁡(p)≅ker⁡(q)×ker⁡(p)≅G×G≅G.\ker(q \circ p) \cong s(\ker(q)) \times \ker(p) \cong \ker(q) \times \ker(p) \cong G \times G \cong G.

Finally, we show that s(t(C))s(t(C)) commutes elementwise with ker⁡(q∘p)=s(ker⁡(q))⋅ker⁡(p)\ker(q \circ p) = s(\ker(q)) \cdot \ker(p). Since t(C)t(C) commutes elementwise with ker⁡(q)\ker(q), applying ss shows that s(t(C))s(t(C)) commutes elementwise with s(ker⁡(q))s(\ker(q)). And since s(t(C))⊆s(B)s(t(C)) \subseteq s(B), it commutes elementwise with ker⁡(p)\ker(p). □\square

Lemma 3.

Good pairs lift: Let (π,ι):X⇄B(\pi,\iota) : X \rightleftarrows B and (p,s):A⇄B(p,s) : A \rightleftarrows B be good pairs. Then there is a good pair (π′,ι′):X⇄A(\pi',\iota') : X \rightleftarrows A with p∘π′=π,ι′∘s=ι.p \circ \pi' = \pi, \qquad \iota' \circ s = \iota.

Proof. Let C≔ker⁡(π)C \coloneqq \ker(\pi) and K≔ker⁡(p)K \coloneqq \ker(p). Since A≅GA \cong G and C≅GC \cong G, we may choose an isomorphism c:A→Cc : A \to C. We have the internal direct product decompositions A=s(B)×KA = s(B) \times K and X=ι(B)×CX = \iota(B) \times C. Applying cc to the former, we get C=c(K)×c(s(B))C = c(K) \times c(s(B)), and therefore X=ι(B)×c(K)×c(s(B))X = \iota(B) \times c(K) \times c(s(B)) as an internal direct product of three subgroups. That is, every element of XX can be written uniquely as ι(b) c(k) c(s(b′))\iota(b) \, c(k) \, c(s(b')) with b,b′∈Bb,b' \in B and k∈Kk \in K. We define the homomorphisms ι′:A→X\iota' : A \to X and π′:X→A\pi' : X \to A by ι′(s(b)k)≔ι(b) c(k),π′(ι(b) c(k) c(s(b′)))≔s(b)k.\iota'(s(b) k) \coloneqq \iota(b) \, c(k), \qquad \pi'\bigl(\iota(b) \, c(k) \, c(s(b'))\bigr) \coloneqq s(b) k. They are well-defined homomorphisms, since they are defined factor by factor on internal direct products. We now verify the required properties.

  • We have π′∘ι′=idA\pi' \circ \iota' = \id_A by construction.
  • We have p(π′(ι(b) c(k) c(s(b′))))=p(s(b)k)=bp\bigl(\pi'(\iota(b) \, c(k) \, c(s(b')))\bigr) = p(s(b) k) = b. This equals π(ι(b) c(k) c(s(b′)))\pi\bigl(\iota(b) \, c(k) \, c(s(b'))\bigr), since π∘ι=idB\pi \circ \iota = \id_B and c(A)=C=ker⁡(π)c(A) = C = \ker(\pi). Hence, p∘π′=πp \circ \pi' = \pi.
  • We have ι′(s(b))=ι(b)\iota'(s(b)) = \iota(b), i.e. ι′∘s=ι\iota' \circ s = \iota.
  • The kernel of π′\pi' is c(s(B))≅B≅Gc(s(B)) \cong B \cong G.
  • The image ι′(A)=ι(B)⋅c(K)\iota'(A) = \iota(B) \cdot c(K) commutes elementwise with ker⁡(π′)=c(s(B))\ker(\pi') = c(s(B)). Indeed, ι(B)\iota(B) commutes elementwise with CC, and c(K)c(K) commutes elementwise with c(s(B))c(s(B)) because KK commutes elementwise with s(B)s(B). □\square

Next, we construct the diagram Δ:ω1op→Grp\Delta : \omega_1^{\op} \to \Grp.

Proposition 4.

There are groups Δβ≅G\Delta_\beta \cong G for β<ω1\beta < \omega_1 and homomorphisms παβ:Δβ→Δα\pi_{\alpha\beta} : \Delta_\beta \to \Delta_\alpha and ιαβ:Δα→Δβ\iota_{\alpha\beta} : \Delta_\alpha \to \Delta_\beta for α≤β<ω1\alpha \leq \beta < \omega_1 with the following properties:

  1. We have παα=ιαα=id\pi_{\alpha\alpha} = \iota_{\alpha\alpha} = \id, and for α≤β≤γ\alpha \leq \beta \leq \gamma we have παβ∘πβγ=παγ\pi_{\alpha\beta} \circ \pi_{\beta\gamma} = \pi_{\alpha\gamma} and ιβγ∘ιαβ=ιαγ\iota_{\beta\gamma} \circ \iota_{\alpha\beta} = \iota_{\alpha\gamma}.
  2. For α<β\alpha < \beta, the pair (παβ,ιαβ):Δβ⇄Δα(\pi_{\alpha\beta},\iota_{\alpha\beta}) : \Delta_\beta \rightleftarrows \Delta_\alpha is good.

Proof. We construct Δβ\Delta_\beta and the homomorphisms παβ,ιαβ\pi_{\alpha\beta},\iota_{\alpha\beta} for α≤β\alpha \leq \beta by transfinite recursion on β\beta.

We start with Δ0≔G\Delta_0 \coloneqq G and π00≔ι00≔idG\pi_{00} \coloneqq \iota_{00} \coloneqq \id_G.

For the successor step, let Δβ+1≔Δβ×G\Delta_{\beta+1} \coloneqq \Delta_\beta \times G, and let (πβ,β+1,ιβ,β+1)(\pi_{\beta,\beta+1},\iota_{\beta,\beta+1}) be the good pair from Lemma 1. For α<β\alpha < \beta we define πα,β+1≔παβ∘πβ,β+1,ια,β+1≔ιβ,β+1∘ιαβ.\pi_{\alpha,\beta+1} \coloneqq \pi_{\alpha\beta} \circ \pi_{\beta,\beta+1}, \qquad \iota_{\alpha,\beta+1} \coloneqq \iota_{\beta,\beta+1} \circ \iota_{\alpha\beta}. Then (1) clearly holds, and (2) follows from Lemma 2.

For the limit step, let λ<ω1\lambda < \omega_1 be a limit ordinal. Since λ\lambda is countable, we may choose a cofinal sequence α0<α1<⋯\alpha_0 < \alpha_1 < \cdots in λ\lambda. First, we define good pairs (παnλ,ιαnλ):Δλ⇄Δαn(\pi_{\alpha_n\lambda},\iota_{\alpha_n\lambda}) : \Delta_\lambda \rightleftarrows \Delta_{\alpha_n} by recursion on nn. Let Δλ≔Δα0×G\Delta_\lambda \coloneqq \Delta_{\alpha_0} \times G, and let (πα0λ,ια0λ)(\pi_{\alpha_0\lambda},\iota_{\alpha_0\lambda}) be the good pair from Lemma 1. If (παnλ,ιαnλ)(\pi_{\alpha_n\lambda},\iota_{\alpha_n\lambda}) has been defined, we apply Lemma 3 to this good pair and to the good pair (παnαn+1,ιαnαn+1):Δαn+1⇄Δαn(\pi_{\alpha_n\alpha_{n+1}},\iota_{\alpha_n\alpha_{n+1}}) : \Delta_{\alpha_{n+1}} \rightleftarrows \Delta_{\alpha_n}. This gives a good pair (παn+1λ,ιαn+1λ):Δλ⇄Δαn+1(\pi_{\alpha_{n+1}\lambda},\iota_{\alpha_{n+1}\lambda}) : \Delta_\lambda \rightleftarrows \Delta_{\alpha_{n+1}} with παnαn+1∘παn+1λ=παnλ,ιαn+1λ∘ιαnαn+1=ιαnλ.\pi_{\alpha_n\alpha_{n+1}} \circ \pi_{\alpha_{n+1}\lambda} = \pi_{\alpha_n\lambda}, \qquad \iota_{\alpha_{n+1}\lambda} \circ \iota_{\alpha_n\alpha_{n+1}} = \iota_{\alpha_n\lambda}. By induction, using (1) below λ\lambda, we get παnαm∘παmλ=παnλ,ιαmλ∘ιαnαm=ιαnλ\pi_{\alpha_n\alpha_m} \circ \pi_{\alpha_m\lambda} = \pi_{\alpha_n\lambda}, \qquad \iota_{\alpha_m\lambda} \circ \iota_{\alpha_n\alpha_m} = \iota_{\alpha_n\lambda} for all n≤mn \leq m.

Now, for an arbitrary α<λ\alpha < \lambda, choose some nn with α<αn\alpha < \alpha_n and define παλ≔πααn∘παnλ,ιαλ≔ιαnλ∘ιααn.\pi_{\alpha\lambda} \coloneqq \pi_{\alpha\alpha_n} \circ \pi_{\alpha_n\lambda}, \qquad \iota_{\alpha\lambda} \coloneqq \iota_{\alpha_n\lambda} \circ \iota_{\alpha\alpha_n}. Also, let πλλ≔ιλλ≔id\pi_{\lambda\lambda} \coloneqq \iota_{\lambda\lambda} \coloneqq \id. By the previous equations and (1) below λ\lambda, this definition does not depend on the choice of nn, and it agrees with the previous definition when α=αm\alpha = \alpha_m. For (1), let α≤β<λ\alpha \leq \beta < \lambda and choose nn with β<αn\beta < \alpha_n. Then παβ∘πβλ=παβ∘πβαn∘παnλ=πααn∘παnλ=παλ,\pi_{\alpha\beta} \circ \pi_{\beta\lambda} = \pi_{\alpha\beta} \circ \pi_{\beta\alpha_n} \circ \pi_{\alpha_n\lambda} = \pi_{\alpha\alpha_n} \circ \pi_{\alpha_n\lambda} = \pi_{\alpha\lambda}, and similarly ιβλ∘ιαβ=ιαλ\iota_{\beta\lambda} \circ \iota_{\alpha\beta} = \iota_{\alpha\lambda}. For (2), the pair (παλ,ιαλ)(\pi_{\alpha\lambda},\iota_{\alpha\lambda}) is the composite of the good pairs (παnλ,ιαnλ)(\pi_{\alpha_n\lambda},\iota_{\alpha_n\lambda}) and (πααn,ιααn)(\pi_{\alpha\alpha_n},\iota_{\alpha\alpha_n}) in the sense of Lemma 2, hence it is good. □\square

By (1), we obtain a diagram Δ:ω1op→Grp\Delta : \omega_1^{\op} \to \Grp with α↦Δα\alpha \mapsto \Delta_\alpha and (α≤β)↦παβ(\alpha \leq \beta) \mapsto \pi_{\alpha\beta}. The homomorphisms ιαβ\iota_{\alpha\beta} are not part of the diagram; they are only used to produce elements of its limit.

Proposition 5.

The limit of Δ\Delta in Grp\Grp is uncountable.

Proof. The limit is the subgroup L≔{x∈∏β<ω1Δβ:πβγ(xγ)=xβ for all β≤γ}.\textstyle L \coloneqq \bigl\{x \in \prod_{\beta < \omega_1} \Delta_\beta : \pi_{\beta\gamma}(x_\gamma) = x_\beta \text{ for all } \beta \leq \gamma\bigr\}. For α<ω1\alpha < \omega_1 we define a homomorphism jα:Δα→∏β<ω1Δβj_\alpha : \Delta_\alpha \to \prod_{\beta < \omega_1} \Delta_\beta by jα(x)β≔{ιαβ(x)β≥α,πβα(x)β≤α.j_\alpha(x)_\beta \coloneqq \begin{cases} \iota_{\alpha\beta}(x) & \beta \geq \alpha, \\ \pi_{\beta\alpha}(x) & \beta \leq \alpha. \end{cases} For β=α\beta = \alpha, both cases give xx, so this is well-defined. We claim that jαj_\alpha maps into LL, i.e. that for all x∈Δαx \in \Delta_\alpha and β≤γ\beta \leq \gamma we have πβγ(jα(x)γ)=jα(x)β.\pi_{\beta\gamma}(j_\alpha(x)_\gamma) = j_\alpha(x)_\beta. This follows from (1) and παβ∘ιαβ=id\pi_{\alpha\beta} \circ \iota_{\alpha\beta} = \id, distinguishing three cases:

πβγ(jα(x)γ)=πβγ(ιαγ(x))=πβγ(ιβγ(ιαβ(x)))=ιαβ(x)=jα(x)βif α≤β,πβγ(jα(x)γ)=πβγ(ιαγ(x))=πβα(παγ(ιαγ(x)))=πβα(x)=jα(x)βif β≤α≤γ,πβγ(jα(x)γ)=πβγ(πγα(x))=πβα(x)=jα(x)βif γ≤α. \begin{align*} \pi_{\beta\gamma}(j_\alpha(x)_\gamma) & = \pi_{\beta\gamma}(\iota_{\alpha\gamma}(x)) = \pi_{\beta\gamma}(\iota_{\beta\gamma}(\iota_{\alpha\beta}(x))) = \iota_{\alpha\beta}(x) = j_\alpha(x)_\beta && \text{if } \alpha \leq \beta, \\ \pi_{\beta\gamma}(j_\alpha(x)_\gamma) & = \pi_{\beta\gamma}(\iota_{\alpha\gamma}(x)) = \pi_{\beta\alpha}(\pi_{\alpha\gamma}(\iota_{\alpha\gamma}(x))) = \pi_{\beta\alpha}(x) = j_\alpha(x)_\beta && \text{if } \beta \leq \alpha \leq \gamma, \\ \pi_{\beta\gamma}(j_\alpha(x)_\gamma) & = \pi_{\beta\gamma}(\pi_{\gamma\alpha}(x)) = \pi_{\beta\alpha}(x) = j_\alpha(x)_\beta && \text{if } \gamma \leq \alpha. \end{align*}

So we get a homomorphism jα:Δα→Lj_\alpha : \Delta_\alpha \to L. A coordinatewise check of the same kind shows jβ∘ιαβ=jαj_\beta \circ \iota_{\alpha\beta} = j_\alpha for α≤β\alpha \leq \beta. Hence, the subgroups Lα≔jα(Δα)L_\alpha \coloneqq j_\alpha(\Delta_\alpha) of LL satisfy Lα⊆LβL_\alpha \subseteq L_\beta for α≤β\alpha \leq \beta. Next, we show that Lα≠Lα+1L_\alpha \neq L_{\alpha+1}. Since GG is non-trivial and ker⁡(πα,α+1)≅G\ker(\pi_{\alpha,\alpha+1}) \cong G by (2), there is some element 1≠k∈ker⁡(πα,α+1)1 \neq k \in \ker(\pi_{\alpha,\alpha+1}). Assume that jα+1(k)=jα(x)j_{\alpha+1}(k) = j_\alpha(x) for some x∈Δαx \in \Delta_\alpha. Comparing the (α+1)(\alpha+1)-coordinates gives k=ια,α+1(x)k = \iota_{\alpha,\alpha+1}(x). Applying πα,α+1\pi_{\alpha,\alpha+1} gives 1=x1 = x, and hence k=1k = 1, a contradiction. Thus, jα+1(k)∈Lα+1∖Lαj_{\alpha+1}(k) \in L_{\alpha+1} \setminus L_\alpha.

Now choose an element yα∈Lα+1∖Lαy_\alpha \in L_{\alpha+1} \setminus L_\alpha for every α<ω1\alpha < \omega_1. These elements are pairwise distinct: if α<β\alpha < \beta, then yα∈Lα+1⊆Lβy_\alpha \in L_{\alpha+1} \subseteq L_\beta, but yβ∉Lβy_\beta \notin L_\beta. Therefore, LL has at least ℵ1\aleph_1 elements. □\square

Corollary 6.

The full subcategory of finitely generated groups Grpfg⊆Grp\Grp_\fg \subseteq \Grp is not closed under ℵ1\aleph_1-cofiltered limits.

Proof. The diagram Δ:ω1op→Grp\Delta : \omega_1^{\op} \to \Grp is ℵ1\aleph_1-cofiltered, and each Δα≅G\Delta_\alpha \cong G is finitely generated. But its limit is uncountable by Proposition 5, whereas every finitely generated group is countable. □\square

Context

This page is referenced by the following categories.