Degrees of compression and inertia for free-abelian times free groups
Mallika Roy
Departament de Matemàtiques, Universitat Politècnica de Catalunya, Catalonia
[email protected]
and
Enric Ventura
Departament de Matemàtiques, Universitat Politècnica de Catalunya, Catalonia
[email protected]
(Date: March 12, 2024)
Abstract.
We introduce the concepts of degree of inertia, diG(H), and degree of compression, dcG(H), of a finitely generated subgroup H of a given group G. For the case of direct products of free-abelian and free groups, we compute the degree of compression and give an upper bound for the degree of inertia. Imposing some technical assumptions to the supremum involved in the definition of degree of inertia, we introduce the notion called restricted degree of inertia, diG′(H), and, again for the case Zm×Fn, we provide an explicit formula relating it to the restricted degree of inertia of its projection to the free part, diFn′(Hπ).
Key words and phrases:
free groups, subgroups, degree of compression, degree of inertia
2010 Mathematics Subject Classification:
20E05, 20E07, 20K25
1. Introduction
For a group G, we write r(G) to denote the rank of G, i.e., the minimum cardinal of a generating set for G. To work with group morphisms, we use the notational convention of writing arguments on the left, i.e., ϕ:G1→G2, g↦gϕ; and so, compositions as written: gϕψ=(gϕ)ψ. Accordingly, we write conjugations on the right, Hg=g−1Hg, and commutators in the form [a,b]=a−1b−1ab. For a subgroup H⩽G, we shall use the notation H⩽fgG to emphasize that H is finitely generated, and H⩽fiG (resp., H⩽∞G, or H⩽lG) to emphasize it is of finite index (resp., infinite index, or index l) in G.
In the commutative realm, the rank function is increasing in the sense that H⩽K⩽G implies r(H)⩽r(K). This is far from true in general, and the main expression of this phenomena can be found in the context of free groups Fn, where the free group of countably infinite rank easily embeds into the free group of rank 2, Fℵ0⩽F2. However, when restricting ourselves to certain families of groups and subgroups, the rank function tends to behave less wildly and somehow closer to the commutative behaviour. An example of this situation is again in finitely generated free groups, but restricting our attention to subgroups fixed by automorphisms or endomorphisms: the story began in [7], where Dyer–Scott showed that Fix(φ) is a free factor of Fn for every finite order automorphism φ∈Aut(Fn), and conjectured that r(Fix(φ))⩽n, in general. This was proved later by Bestvina–Handel [2], and extended several times in subsequent papers, all of them pointing to the direction that the rank function, when restricted to subgroups fixed by endomorphisms, tends to behave similarly to the abelian case. In this spirit, the following concepts were first introduced by Dicks–Ventura in [6] and turned out to be quite relevant in the subsequent literature:
Definition 1.1**.**
Let G be a group. A finitely generated subgroup H⩽fgG is said to be compressed in G if r(H)⩽r(K), for every H⩽K⩽G. And H is said to be inert in G if r(H∩K)⩽r(K), for every K⩽G. (Note that, equivalently, in both definitions one can restrict the attention to those subgroups K’s being finitely generated.)
Observe that, directly by induction from the definition, inert subgroups are closed under finite intersections. Also, inert subgroups are compressed, while the other implication is not known to be true or false in free groups (this is the so-called inertia conjecture, see Zhang–Ventura–Wu [21] for partial results), and it is not true in general:
Example 1.2**.**
Consider the direct product of the Klein bottle group with the group of integers, say G=⟨a,b∣bab−1a⟩×⟨c∣⟩, and its subgroup H=⟨a,b2,c⟩≃Z3. By Corollary 4.3 and Proposition 4.4 from [21], H is compressed but not inert in G.
Several known results involving these concepts include the following:
Theorem 1.3**.**
(Dicks–Ventura, **[6]**): Arbitrary intersections of fixed subgroups of injective endomorphisms of Fn are inert in Fn;
(Martino–Ventura, **[12]**): arbitrary intersections of fixed subgroups of endomorphisms of Fn are compressed in Fn;
(Wu–Zhang, **[20]**): arbitrary intersections of fixed subgroups of automorphisms of closed surface groups G with negative Euler characteristic are inert in G;
(Wu–Ventura–Zhang, **[19]**): arbitrary intersections of fixed subgroups of endomorphisms of surface groups G are compressed in G.
Also, in [19] and [21], Zhang–Ventura–Wu studied similar questions within the family of finite direct products of free and surface groups, where more interesting phenomena show up.
In the present paper we introduce a quantification for these two concepts and study it within the families of free groups, and free-abelian times free groups. For technical reasons it is better to work with the so-called reduced rank of a group G, defined as r~(G)=max{0,r(G)−1}, i.e., one unit less than the rank except for the trivial group for which we take zero (note that then, r~(1)=r~(Z)=0 while 0=r(1)=r(Z)=1). Observe that H⩽G is compressed in G if and only if r~(H)/r~(K)⩽1 for every H⩽K⩽fgG; and that H⩽G is inert in G if and only if r~(H∩K)/r~(K)⩽1 for every K⩽fgG (understanding always 0/0=1). This motivates the following quantitative definitions:
Definition 1.4**.**
Let G be a group and H⩽fgG. The degree of compression of H in G is dcG(H)=supK{r~(H)/r~(K)}, where the supremum is taken over all subgroups H⩽K⩽fgG. Similarly, the degree of inertia of H in G is diG(H)=supK{r~(H∩K)/r~(K)}, where the supremum is taken over all K⩽fgG satisfying H∩K⩽fgG and, in both cases, 0/0 is understood to be 1.
Note that, taking K=H, we get dcG(H)⩾1 and diG(H)⩾1. So, the possibility of K being cyclic (which leads in both cases to 0/0=1) is irrelevant in both definitions and we can restrict the two supremums to non-cyclic K’s without changing their final values.
Note also that the supremum in the definition of degree of compression is always a maximum, since the numerator has a fixed value and the denominator takes only natural values. Although we do not have any particular example, the supremum in the definition of degree of inertia could, in principle, not be attained at any particular subgroup K. In this sense, the following is an intriguing question for which, at the time of writing, we have no idea how to answer:
Question 1.5**.**
Is there a (finitely generated) group G and a subgroup H⩽fgG such that diG(H) is a transcendental number? or an irrational number? Or such that the supremum in diG(H) is not a maximum?
Observe that in the definition of degree of inertia, we take the supremum only over those subgroups K⩽fgG whose intersection with H is again finitely generated. In groups G with the Howson property (the intersection of any two finitely generated subgroups is again finitely generated), like free groups or surface groups, this is no restriction at all and that supremum is over all finitely generated K’s. Otherwise, if G is not Howson we are eliminating, on purpose, those possible finitely generated K’s having non-finitely generated intersection with H (which would force diG(H) to be automatically infinite). However observe that, even with the present limited definition, diG(H) may be infinite as well; explicit examples will be shown later.
We adapt the definition of inertia to the non-Howson environments by saying that a subgroup H⩽G is finitary inert in G if r(H∩K)⩽r(K) for every K⩽fgG such that H∩K⩽fgG. The following observation then follows directly from the definitions and presents the values of dcG(H) and diG(H) as a quantification of how far is the subgroup H⩽fgG from being compressed and from being finitary inert in G, respectively:
Observation 1.6**.**
Let G be a group and H⩽fgG. Then,
1⩽dcG(H)⩽diG(H);
dcG(H)=1* if and only if H is compressed in G;*
diG(H)=1* if and only if H is finitary inert in G.*
The following intriguing question is open, as far as we know:
Question 1.7**.**
Is there a (finitely generated) group G with a subgroup H⩽fgG being finitary inert but not inert? (i.e., satisfying r~(H∩K)⩽r~(K) for every K⩽fgG with H∩K⩽fgG, but simultaneously admitting some K0⩽fgG with r~(H∩K0)=∞?).
In the present paper we study these notions for the case of the free group and obtain the following result in Section 2:
Theorem 1.8**.**
For any finitely generated free group G=Fn, the function dcFn(⋅) is computable; more precisely, there is an algorithm which, on input h1,…,hr∈Fn, it computes the value of dcFn(⟨h1,…,hr⟩) and outputs a free basis of a subgroup K⩽fgFn where it is attained.
The question whether diFn(⋅) is computable in free groups (related to the question whether the corresponding supremum is a maximum or not) seems to be much more delicate. In Section 2 we refer to a quite similar question, which was successfully solved recently by S. Ivanov in [9]. However, at the time of writing, we do not know how to use this result to eventually compute diFn(⋅).
Then, we concentrate in free-abelian times free groups, G=Zm×Fn, where the situation is richer and trickier because, for m⩾1, n⩾2, G is known to be non-Howson (the easy example G=Z×F2=⟨t⟩×⟨a,b⟩, H=⟨a,b⟩, K=⟨ta,b⟩, and H∩K=⟨w(a,b)∣∣w∣a=0⟩=⟨a−nban,n∈Z⟩ already appears in [3] attributed to Moldavanskii). Denoting by π:G↠Fn the natural projection, in Section 3 we study the degree of compression and prove the following result:
Theorem 1.9**.**
For any given H⩽fgG=Zm×Fn, and using the notation from Section 3, we have
[TABLE]
Moreover, dcG(H) is algorithmically computable; more precisely, there is an algorithm which, on input h1,…,hr∈G, it computes the value of dcG(⟨h1,…,hr⟩) and outputs a basis of a subgroup K⩽fgG where that supremum is attained.
In Section 4, we study the degree of inertia, also for groups of the form G=Zm×Fn, and get the following result:
Theorem 1.10**.**
Let H⩽fgG=Zm×Fn, and let LH=H∩Zm.
If r(Hπ)⩽1 then diG(H)=1;
if r(Hπ)⩾2 and [Zm:LH]=∞ then diG(H)=∞;
if r(Hπ)⩾2 and [Zm:LH]=l<∞ then diG(H)⩽ldiFn(Hπ).
We conjecture that the inequality in Theorem 1.10 (iii) is, in fact, an equality, i.e.,
Conjecture 1.11**.**
For G=Zm×Fn and H⩽fgG with r(Hπ)⩾2, we have the equality diG(H)=[Zm:LH]⋅diFn(Hπ), where LH=H∩Zm.
Unfortunately, we cannot complete a general proof for this equality. Instead, we get an approximation to it by introducing a technical modification to the definition of diG(⋅), the so-called restricted degree of inertia:
Definition 1.12**.**
Let G be a group and π:G↠G/Z(G) the projection modulo its center Z(G)⊴G. Let H⩽fgG be such that Hπ is not virtually cyclic and Hπ[Gπ,Gπ]. The restricted degree of inertia of H in G is diG′(H)=supK{r~(H∩K)/r~(K)}, where the supremum is taken over all K⩽fgG satisfying H∩K⩽fgG, [Hπ:Hπ∩Kπ]=∞, and Hπ∩Kπ[Gπ,Gπ] (again, understanding 0/0=1).
Remark 1.13**.**
The conditions on the projection of the subgroup Hπ not to be virtually cyclic and not to be contained in the commutator [Gπ,Gπ] are just to make sure the supremum is not over the empty set: assuming these two conditions, let h∈H be such that hπ∈/[Gπ,Gπ] and take K=⟨h⟩⩽G; clearly, K=H∩K=⟨h⟩ is cyclic and so finitely generated, Hπ∩Kπ=⟨hπ⟩⩽∞Hπ because Hπ is not virtually cyclic, and hπ∈Hπ[Gπ,Gπ]. Moreover, r~(H∩K)/r~(K)=0/0=1 and so, diG′(H)⩾1.
Note that, the two cases of our interest are the following: (i) G=Fn, n⩾2, we have Z(G)=1, G/Z(G)=Fn, and the definition applies to any non-cyclic subgroup H⩽G such that H⩽[Fn,Fn]; and (ii) G=Zm×Fn, n⩾2, we have Z(G)=Zm, G/Z(G)=Fn, and the definition applies to any subgroup H⩽G such that Hπ is not cyclic and Hπ⩽[Fn,Fn].
Observe that the definition of restricted degree of inertia coincides with that of degree of inertia, except for the extra technical conditions required to the subgroups K over which the supremum runs. Hence, 1⩽diG′(H)⩽diG(H) and we conjecture that, at least in the free and free-abelian times free cases, they do coincide.
Conjecture 1.14**.**
For G=Fn and G=Zm×Fn, diG′(⋅)=diG(⋅).
Our main result in Section 5 is the desired equality from Theorem 1.10 (iii), but expressed with the restricted degrees of inertia:
Theorem 1.15**.**
Let H⩽fgG=Zm×Fn be such that Hπ is not cyclic and Hπ[Fn,Fn], and let LH=H∩Zm.
- (i)
If [Zm:LH]=∞ then diG′(H)=∞;
2. (ii)
if [Zm:LH]=l then diG′(H)=ldiFn′(Hπ).
Theorem 1.15 is the most involved and technical result in the paper. We hope in the future some new ideas come up allowing to avoid the technical working conditions (namely, Hπ∩Kπ having infinite index in Hπ, and not being contained in [Fn,Fn]) and to recycle the proof of Theorem 1.15 into a proof for Conjecture 1.11 (or, better, for Conjecture 1.14).
We conclude the present section with a straightforward result which will be used later.
Lemma 1.16**.**
Let ϕ:G1→G2 be an isomorphism of groups. For every H⩽fgG1,
dcG2(Hϕ)=dcG1(H);
diG2(Hϕ)=diG1(H);
with the extra assumptions that Hπ1 is not virtually cyclic and Hπ1[G1π1,G1π1], where πi:Gi→Gi/Z(Gi) is the natural projection modulo the center Z(Gi)⊴Gi, i=1,2, we have diG2′(Hϕ)=diG1′(H).
Proof.
For every K⩽fgG1 with H⩽K, we have Kϕ⩽fgG2 and Hϕ⩽Kϕ so, r~(H)=r~(Hϕ)⩽dcG2(Hϕ)⋅r~(Kϕ)=dcG2(Hϕ)⋅r~(K). Therefore, dcG1(H)⩽dcG2(Hϕ). By symmetry, we get (i).
Similarly, for every K⩽fgG1 with H∩K⩽fgG1, we have Kϕ⩽fgG2 and Hϕ∩Kϕ=(H∩K)ϕ⩽fgG2 so, r~(H∩K)=r~((H∩K)ϕ)=r~(Hϕ∩Kϕ)⩽diG2(Hϕ)⋅r~(Kϕ)=diG2(Hϕ)⋅r~(K). Therefore, diG1(H)⩽diG2(Hϕ). By symmetry, we deduce (ii).
The argument in the previous paragraph also shows (iii), provided we see that the technical conditions in the supremum of the definition of restricted degree of inertia get preserved under ϕ. And in fact they do: suppose K⩽fgG1 is such that H∩K⩽fgG1, Hπ1∩Kπ1⩽∞Hπ1 and Hπ1∩Kπ1[G1π1,G1π1]; since, ϕ is an isomorphism, Kϕ⩽fgG2 and Hϕ∩Kϕ=(H∩K)ϕ⩽fgG2; also, since ϕ maps Z(G1) onto Z(G2), there exists an isomorphism ϕˉ:G1/Z(G1)→G2/Z(G2) such that π1ϕˉ=ϕπ2 and, hence, Hπ1∩Kπ1⩽∞Hπ1 implies Hϕπ2∩Kϕπ2=Hπ1ϕˉ∩Kπ1ϕˉ=(Hπ1∩Kπ1)ϕˉ⩽∞Hπ1ϕˉ=Hϕπ2; finally, Hϕπ2∩Kϕπ2=Hπ1ϕˉ∩Kπ1ϕˉ=(Hπ1∩Kπ1)ϕˉ[G1π1,G1π1]ϕˉ=[G1ϕπ2,G1ϕπ2]=[G2π2,G2π2]. This completes the proof of the lemma.
∎
Corollary 1.17**.**
Let G be a group. For every H⩽fgG and every g∈G, dcG(Hg)=dcG(H), diG(Hg)=diG(H), and (with the extra assumptions on H and so, on Hg) diG′(Hg)=diG′(H). ∎
2. The free case
For all the paper, we fix an alphabet of n letters, X={x1,…,xn}, and consider the free group on it, F(X), also denoted by Fn. In the present section we study the degrees of compression and inertia in the context of the free group, i.e., the functions dcFn(⋅) and diFn(⋅).
Hanna Neumann proved in [15] that r~(H∩K)⩽2r~(H)r~(K), for every H,K⩽fgFn. And the same assertion removing the factor “2” became soon known as the Hanna Neumann conjecture. This has been a major problem in Geometric Group Theory, with lots of partial results and improvements appearing in the literature since then. An interesting one was done by W. Neumann in [16], who proved the stronger fact ∑x∈Sr~(H∩Ks)⩽2r~(H)r~(K) (known as the strengthened Hanna Neumann inequality), where S is any set of double coset representatives of Fn modulo H on the left and K on the right (i.e., S⊆Fn contains one and only one element in each double coset H\Fn/K); in particular, this implies that, for all H,K⩽fgFn, all except finitely many of the intersections H∩Kx, x∈Fn, are trivial or cyclic. Few years ago the Hanna Neumann conjecture, even in its strengthened version, has been completely resolved in the positive, independently by J. Friedman [8] and by I. Mineyev [14] (see also W. Dicks [5]). This can be interpreted as the following upper bound for dcFn(H) and diFn(H) in terms of the subgroup H⩽fgFn:
Observation 2.1**.**
For H⩽fgFn, we have 1⩽dcFn(H)⩽diFn(H)⩽r~(H).
Friedman–Mineyev’s inequality is easily seen to be tight (consider, for example, the subgroups H=⟨a,b−1ab⟩ and K=⟨b,a2,aba⟩ of F2, and its intersection H∩K=⟨a2,b−1a2b,b−1aba⟩); therefore, it can be interpreted in the following way: “the smallest possible multiplicative constant α∈R satisfying r~(H∩K)⩽αr~(H)r~(K), for every H,K⩽fgFn, is α=1”. Now fix the subgroup H: by definition, the smallest possible constant α∈R satisfying r~(H∩K)⩽αr~(H)r~(K), for every K⩽fgFn, is α=diFn(H)/r~(H).
S. Ivanov [9] already considered and studied the strengthened version of what we call here the degree of inertia. He defined the Walter Neumann coefficient of H⩽fgFn as σ(H):=supK⩽fgFnr~(H,K)/r~(H)r~(K), where r~(H,K)=∑s∈H\Fn/Kr~(H∩Ks) (understanding 0/0=1). In other words, σ(H) is the smallest possible constant α∈R such that r~(H,K)⩽αr~(H)r~(K), for every K⩽fgFn. Using linear programming techniques, Ivanov was able to prove the following remarkable result:
Theorem 2.2** (Ivanov, [9]).**
For any finitely generated free group Fn, the function σ is computable and the supremum is a maximum; more precisely, there is an algorithm which, on input h1,…,hr∈Fn, it computes the value of σ(⟨h1,…,hr⟩) and outputs a free basis of a subgroup K⩽fgFn where that supremum is attained.
Ivanov’s proof is involved and technical. Although it looks quite similar, we have been unable to adapt Ivanov’s arguments to answer any of the following questions which, as far as we know, remain open:
Question 2.3**.**
Is the function diFn(⋅) computable? Is that supremum always a maximum? More precisely, is there an algorithm which, on input h1,…,hr∈Fn, computes the value of diFn(⟨h1,…,hr⟩)? Or even more, it outputs a free basis of a subgroup K⩽fgFn where it is attained?
The corresponding questions for the degree of compression are much easier and can be established with the use of Stallings’ graphs (we assume the reader is familiar with these techniques), algebraic extensions, and Takahasi’s Theorem.
Definition 2.4**.**
Let H⩽fgK⩽fgFn. If H is a free factor of K we write H⩽ffK. On the other extreme, the extension H⩽K is said to be algebraic, denoted by H⩽algK, if H is not contained in any proper free factor of K, i.e., if H⩽A⩽K=A∗B implies B=1; we denote by AEFn(H) the set of algebraic extensions of H in Fn.
Theorem 2.5** (Takahasi, [17]; see also [13]).**
Every H⩽fgFn has finitely many algebraic extensions, say AEFn(H)={H=H0,H1,…,Hr} (r depending on H), each Hi is finitely generated, and free bases for all of them are algorithmically computable from a given set of generators for H. Furthermore, for every extension H⩽K⩽Fn, there exists a unique (and computable) 0⩽i⩽r such that H⩽algHi⩽ffK; this Hi is called the K-algebraic closure of H.
Sketch of the proof.
The original proof by M. Takahasi [17] was combinatorial, playing with words and cancellation in the free group. We sketch the modern proof given in [13] following ideas of Ventura [18], Kapovich–Miasnikov [10] and Margolis–Sapir–Weil [11].
We have the alphabet X fixed as a free basis for the ambient free group, Fn=F(X). Now, given generators for H⩽fgF(X), one can compute the Stallings’ graph Γ(H) for H (denote the basepoint by ⊙). Attaching the necessary infinite hanging trees so that it becomes a complete graph (i.e., with all vertices having an incoming and an outgoing edge labelled xi for every xi∈X), we obtain the Schreier graph χ(Fn,H,X) (which is finite if and only if H is of finite index in Fn). Of course, χ(Fn,H,X) is a covering, χ(Fn,H,X)↠R(X), of the bouquet R(X), the graph with a single vertex and one loop labelled xi for every xi∈X; more precisely, it is the covering of R(X) corresponding to the subgroup H⩽fgπ(R(X))=Fn. By standard covering theory, K↔χ(Fn,K,X) is a bijection between intermediate subgroups H⩽K⩽Fn and intermediate coverings, χ(Fn,H,X)↠χ(Fn,K,X)↠R(X) (mapping finitely generated subgroups to graphs with finite core, and viceversa).
Fix H⩽fgK⩽fgFn, and consider their Stallings’ graphs Γ(H)=core(χ(Fn,H,X)) and Γ(K)=core(χ(Fn,K,X)), both being finite graphs. The above bijection means that χ(Fn,K,X) is a quotient of χ(Fn,H,X), i.e., the result of χ(Fn,H,X) after identifying vertices and edges in a compatible way (i.e., modulo a congruence, an equivalence relation on the set of vertices satisfying that if p∼q and e1 and e2 are edges with the same label and ιe1=ιe2=p, then e1∼e2). There are two cases: if no pair of vertices in Γ(H)⩽χ(Fn,H,X) become identified then Γ(H) is a subgraph of Γ(K)=core(χ(Fn,K,X)) and so, H⩽ffK; otherwise, we loose H from the picture, but we can still say that some compatible quotient of Γ(H) will be visible as a subgraph of Γ(K). Since Γ(H) is finite, it has finitely many compatible quotients and, therefore, computing all of them and computing free bases for their fundamental groups, we obtain a finite list of finitely generated subgroups OFn(H)={H=H0,H1,…,Hr} (r depending on H), called the fringe of H in [13], all of them containing H and satisfying the following property: for every H⩽fgK⩽fgFn there exists (a non necessarily unique) i=0,…,r such that H⩽Hi⩽ffK.
It only remains to clean this list by checking, for each pair of indices i,j, whether Hi⩽ffHj and, in this case, delete Hj from the list. It is not difficult to see that the resulting reduced list is precisely AEFn(H)⊆OFn(H). Uniqueness of the K-algebraic closure follows directly from the definition of algebraic extension.
∎
As an easy corollary, we obtain the following result which immediately proves Theorem 1.8:
Corollary 2.6**.**
For any subgroup H⩽fgFn, we have dcFn(H)=supH⩽K⩽fgFn{r~(H)/r~(K)}=maxK∈AEFn(H){r~(H)/r~(K)}; furthermore, we can effectively compute dcFn(H) and a free basis of a subgroup K where the maximum is attained.
Proof.
By Theorem 2.5, every H⩽K⩽fgFn uniquely determines the K-algebraic closure of H, i.e., an H′∈AEFn(H) such that H⩽algH′⩽ffK. Since r~(H′)⩽r~(K), we can restrict the supremum in the definition of dcFn(H) to those subgroups in AEFn(H). And, since ∣AEFn(H)∣ is finite and computable, this supremum is a maximum and we can effectively compute both dcFn(H) and a free basis of a subgroup K where the maximum is attained.
∎
3. Degree of compression in free-abelian times free groups
For the rest of the paper we work in free-abelian times free groups G=Zm×Fn, i.e., direct products of a free-abelian group Zm and a free group Fn, investigating here the degrees of compression and inertia of subgroups.
Taking a free-abelian basis {t1,…,tm} for Zm (with multiplicative notation), and a free basis {x1,…,xn} for Fn, we have
[TABLE]
where i,j=1,2,…,m and k=1,2,…,n. A normal form for elements in G is
[TABLE]
where a=(a1,…,am)∈Zm is a row integral vector, and u=u(x1,…,xn) is a reduced word in Fn. Note that the symbol t by itself has no real meaning; it just allows us to convert the ambient notation for the abelian group Zm from multiplicative into additive (since tatb=ta+b, for a,b∈Zm).
At a first glance, solving problems in Zm×Fn seems to be reducible to the corresponding problems in Zm and Fn. However, this is not always the case and many naive looking questions are much more complicated to answer in Zm×Fn, rather than in Zm and Fn. This is the case, for example, with the Howson property: both Zm and Fn are Howson but, as we saw above, G=Zm×Fn is not (as soon as m⩾1 and n⩾2).
Let π:G=Zm×Fn↠Fn, tau↦u, be the natural projection. For a subgroup H⩽fgG, a basis of H is a set of generators for H of the form {ta1u1,ta2u2,…,tarur,tb1,tb2,…,tbs}, where {u1,…,ur} is a free basis of Hπ, ai∈Zm for i=1,2,…,r, and {b1,…,bs} is a free-abelian basis for LH=H∩Zm (to avoid confusions, we will maintain the full names, free-abelian basis, free basis, and just basis, to refer to Zm, Fn, and G=Zm×Fn, respectively). According to [4, Prop. 1.9], every subgroup H⩽fgG admits a basis, computable from any given set of generators. Observe also that a subgroup H⩽G is finitely generated if and only if Hπ⩽Fn is so.
In this section we study the degree of compression of a given subgroup H⩽fgG. The first lemma says that it is enough to consider those overgroups K such that Hπ⩽algKπ.
Lemma 3.1**.**
Let H⩽fgG=Zm×Fn. Then,
[TABLE]
Proof.
We already observed above that the supremum defining the degree of compression is always a maximum. The inequality ⩾ is clear.
Fix a basis for H, say {ta1u1,…,tarur,tb1,…,tbs}. To see the other inequality, take a subgroup H⩽K⩽fgG and we shall construct H⩽K′⩽fgG such that Hπ⩽algK′π and r~(H)/r~(K)⩽r~(H)/r~(K′).
We have LH=H∩Zm=⟨tb1,…,tbs⟩⩽K∩Zm=LK and Hπ⩽Kπ so, r(LH)⩽r(LK) and Hπ⩽algJ⩽ffKπ, for some J∈AEFn(Hπ). Take a free basis {v1,…,vp} for J and extend it to a free basis {v1,…,vp,vp+1,…,vq} for Kπ, p⩽q. Now, consider a basis for K of the form {tc1v1,…,tcpvp,tcp+1vp+1,…,tcqvq,td1,…,tdℓ}, where ci∈Zm, i=1,…,q, are certain vectors, and {td1,…,tdℓ} is a free-abelian basis for LK.
Let, K′=⟨tc1v1,…,tcpvp,td1,…,tdℓ⟩⩽fgK⩽G and we claim that H⩽K′. In fact, we already know that tbi∈LH⩽LK=LK′=⟨td1,…,tdℓ⟩⩽K′ for i=1,…,s. Now, for i=1,…,r we see that taiui∈K′: write ui as a word ui=wi(v1,…,vp) (unique up to reduction) and compute wi(tc1v1,…,tcpvp)=teiwi(v1,…,vp)=teiui∈K′⩽K, where ei=∣wi∣v1c1+⋯+∣wi∣vpcp. But taiui∈H⩽K so, tei−ai∈LK=LK′⩽K′ and hence, taiui=(tei−ai)−1(teiui)∈K′.
So, for every H⩽K⩽fgG we have found a finitely generated subgroup in between, H⩽K′⩽K, such that Hπ⩽algJ=K′π and
[TABLE]
therefore, r~(H)/r~(K′)⩾r~(H)/r~(K) and the proof is completed.
∎
Fix H⩽fgG together with a basis for it {ta1u1,…,tarur,tb1,…,tbs}, and consider the matrices
[TABLE]
For every J∈AEFn(Hπ) given with a free basis, say J=⟨v1,…,vp⟩, we can consider the (unique reduced) word expressing each ui in terms of v1,…,vp, say ui=wi(v1,…,vp), abelianize, and get the vector (∣wi∣v1,…,∣wi∣vp)∈Zp, i=1,…,r; collecting all of them into the rows of a matrix,
[TABLE]
According to Lemma 3.1, to compute dcG(H) it is enough to consider the subgroups of the form K=⟨tc1v1,…,tcpvp,LK⟩⩽fgG (where LK=K∩Zm, assume the given set of generators to be a basis for K) such that H⩽K⩽G, Hπ=⟨u1,…,ur⟩⩽algKπ=⟨v1,…,vp⟩, compute r~(H)/r~(K), and take the maximum of these values. (Observe that, although ∣AEFn(Hπ)∣<∞, there are, possibly, infinitely many such K’s; however, r~(K)=p−1+r(LK) takes only finitely many values.)
So, fix such a K and consider the matrix
[TABLE]
Observe that CK satisfies row(A−UKπCK)⩽LK: in fact, for every i=1,…,r, we have
[TABLE]
where (UKπ)i is the i-th row of UKπ; therefore, H⩽K implies that ai−(UKπ)iCK∈LK, for i=1,…,r. This motivates the following definition, which allows us to obtain the main result in this section.
Definition 3.2**.**
For given matrices A∈Mr×m(Z), B∈Ms×m(Z), and U∈Mr×p(Z), define d(A,B,U)=minL⩽Zm{r(L)∣∃C∈Mp×m(Z) such that row(A−UC)⩽L, and row(B)⩽L}.
Proof of Theorem 1.9.
By Lemma 3.1, we know that the supremum in dcG(H) is attained at a certain H⩽K⩽fgG such that Kπ∈AEFn(Hπ). And, for every such K, r~(K)=r~(Kπ)+r(LK) so,
[TABLE]
[TABLE]
since, by the argument above, every K with Kπ=J∈AEFn(Hπ) satisfies r(LK)⩾d(A,B,UJ), one of them with equality.
In order to compute the value of dcG(H) we can do the following: first compute AEFn(Hπ); for each member J=⟨v1,…,vp⟩, write each ui in the free basis of Hπ in terms of the free basis {v1,…,vp} of J, and obtain the matrix UJ; then compute d(A,B,UJ)+r~(J) (which is effectively doable by the following Proposition 3.3). When this procedure is done for each of the finitely many J∈AEFn(Hπ), take the minimum of the values d(A,B,UJ)+r~(J) and, by (1), we are done. Moreover, the elements of the free basis for the subgroup J attaining this minimum, together with the rows of the matrix C just computed and realizing the minimum in d(A,B,UJ), are the ingredients to build a basis for a subgroup K⩽fgG attaining the minimum in dcG(H).
∎
Proposition 3.3**.**
For any given matrices A∈Mr×m(Z), B∈Ms×m(Z), and U∈Mr×p(Z), the value of d(A,B,U) is algorithmically computable, together with a free-abelian basis of an L⩽Zm attaining the minimum, and the corresponding matrix C∈Mp×m(Z).
Proof.
Recall that d(A,B,U) is the minimum rank of those subgroups L⩽Zm satisfying row(B)⩽L, and row(A−UC)⩽L for some C∈Mp×m(Z). Observe first that, replacing B by B′ with row(B)⩽firow(B′)⩽⊕Zm, we have d(A,B′,U)=d(A,B,U); in fact, d(A,B′,U)⩾d(A,B,U) is clear from the definition, and for every L⩽Zm containing row(B) and row(A−UC) for some C∈Mp×m(Z), we have the subgroup L+row(B′)⩽Zm which contains row(B′) and row(A−UC) for the same matrix C, and has the same rank, r(L+row(B′))=r(L), since L⩽fiL+row(B′); this proves the equality.
Let us do a few reductions to the problem. Compute matrices P∈GLr(Z), Q∈GLp(Z), and positive integers d1,…,dℓ∈N, ℓ⩽min{r,p}, satisfying 1⩽d1∣d2∣⋯∣dℓ=0, such that PUQ=U′, where U′=diag(d1,…,dℓ)∈Mr×p(Z) (understanding the last r−ℓ⩾0 rows and the last p−ℓ⩾0 columns full of zeros); this is the Smith normal form of U, see [1] for details. Writing A′=PA, B′=B, and doing the change of variable C=QC′, we have row(A−UC)=row(PA−PUQC′)=row(A′−U′C′). So, d(A,B,U)=d(A′,B′,U′).
To compute d(A′,B′,U′), we have to find a subgroup L⩽Zm of the minimum possible rank, and vectors c1′,…,cp′∈Zm, such that row(B′)⩽L,
[TABLE]
and
[TABLE]
Note that the last p−ℓ⩾0 columns of U′ are full of zeroes so, no condition concerns the vectors cℓ+1′,…,cp′ and we can take them to be arbitrary (say zero, for example). That is, taking cℓ+1′=⋯=cp′=0, denoting A′′=A′∈Mr×m(Z), B′′=B′∈Ms×m(Z), U′′∈Mr×ℓ(Z) the matrix U′ after deleting the last p−ℓ⩾0 columns (and C′′∈Mℓ×m(Z) the matrix C′ after deleting the last p−ℓ⩾0 rows), we have d(A′,B′,U′)=d(A′′,B′′,U′′).
Now, we can ignore conditions (3) by adding the vectors aℓ+1′′,…,ar′′ as extra rows at the bottom of B: let A′′′∈Mℓ×m(Z) be A′′ after deleting the last r−ℓ⩾0 rows, B′′′∈M(s+r−ℓ)×m(Z) be B′′ enlarged with r−ℓ extra rows with the vectors aℓ+1′′,…,ar′′, U′′′∈Mℓ×ℓ(Z) be the matrix U′′ after deleting the last r−ℓ⩾0 rows (and C′′′=C′′), and we have that d(A′′,B′′,U′′)=d(A′′′,B′′′,U′′′). Note that now, U′′′=diag(d1,…,dℓ) is a square matrix.
Finally, if d1=1 we can take c1′=a1′ and the first condition in (2) becomes trivial; so, deleting the possible ones at the beginning of the list d1∣d2∣⋯∣dℓ (and their rows and columns from U′′′), and deleting also the corresponding first rows of A and C, we can assume d1=1.
Altogether, and resetting the notation to the original one, we are reduced to compute d(A,B,U) in the special situation where A∈Mr×m(Z), B∈Ms×m(Z), and U=diag(d1,…,dr)∈Mr×r(Z), with 1=d1∣d2∣⋯∣dr=0, and further, by the argument in the first paragraph of the present proof, with row(B) being a direct summand of Zm. That is, we have to compute a subgroup L⩽Zm of the minimum possible rank, and vectors c1,…,cp∈Zm satisfying row(B)⩽L and
[TABLE]
where ai is the i-th row of A. Let us think the conditions in (4) as saying that ai∈L modulo diZm, i=1,…,r. To solve this, let us start with L0=row(B)⩽⊕Zm and let us increase it the minimum possible in order to fulfill conditions (4).
Since d1∣d2∣⋯∣dr, the natural projections πi:Zm↠(Z/diZ)m factorize through the chain of morphisms Zm↠(Z/drZ)m↠(Z/dr−1Z)m↠⋯↠(Z/d1Z)m. Starting with L⩾L0 and collecting the last condition in (4), we deduce that L must further satisfy Lπr⩾L0πr+⟨vr0πr⟩, where vr0=ar∈Zm. Now the second condition from below in (4) adds the requirement Lπr−1∋ar−1πr−1. But ar−1πr−1∈(Z/dr−1Z)m has finitely many (more precisely, (dr/dr−1)m) pre-images in (Z/drZ)m; compute them all, take pre-images vr−1 up in Zm, and we get that L must further satisfy Lπr⩾L0πr+⟨vr0πr,vr−1πr⟩, where vr−1πr is one of these (dr/dr−1)m pre-images. Repeat this same argument with all the conditions in (4), working from bottom to top: we deduce that L must further satisfy Lπr⩾L0πr+⟨vr0πr,vr−1πr,…,v1πr⟩, where vi∈Zm is a vector such that viπr is one of the computed (dr/di)m pre-images of aiπi∈(Z/diZ)m up in (Z/drZ)m, i=r−1,…,1, i.e., vi≡aimoddi. This makes a total of (dr/dr−1)m⋯(dr/d1)m possible lower bounds for Lπr: compute them all, find one with minimal possible rank, say Lπr⩾L0πr+⟨vr0πr,vr−10πr,…,v10πr⟩, and we deduce that d(A,U,B)⩾r(L1πr), where L1=L0+⟨vr0,vr−10,…,v10⟩⩽Zm.
We claim that this lower bound is tight, i.e., d(A,B,U)=r(L1πr). To see this, we have to construct a subgroup L⩽Zm of rank exactly r(L1πr), containing L0 and satisfying (4) for some vectors c1,…,cr∈Zm (which must also be computed). Since L0 is a direct summand of Zm, say with free-abelian basis {w1,…,wk}, we deduce that L0πr is a direct summand of (Z/drZ)m with abelian basis {w1πr,…,wkπr}. So, L0πr is also a direct summand of L1πr⩽(Z/drZ)m; compute a complement and get vectors v1′,…,vl′∈Zm, l⩽r, such that {w1πr,…,wkπr,v1′πr,…,vl′πr} is an abelian basis of L1πr=L0πr⊕V; in particular, r(L1πr)=k+l.
Finally, take L=⟨w1,…,wk,v1′,…,vl′⟩⩽Zm. This subgroup has the desired rank r(L)=k+l=r(L1πr) (since the given generators are linearly independent because their πr-projections are so), and satisfies the required conditions: on one hand, L0=⟨w1,…,wk⟩⩽L; on the other hand, for every i=1,…,r, vi0πr∈L1πr=⟨w1πr,…,wkπr⟩⊕⟨v1′πr,…,vl′πr⟩ so,
[TABLE]
for some integers λ1,…,λk,μ1,…,μl∈Z; thus, L contains the vector ci=λ1w1+⋯+λkwk+μ1v1′+⋯+μlvl′ which satisfies ci≡vi0moddr and so, ci≡vi0moddi too; since vi0≡aimoddi, we deduce ci≡aimoddi and we are done.
∎
It is natural to ask whether the minimum minJ∈AEFn(Hπ){r~(J)+d(A,B,UJ)} in Theorem 1.9 is attained at an algebraic extension J∈AEFn(Hπ) of minimal rank. Unfortunately, this is not always the case, as shown in the following example. In order to compute dcG(H), this fact forces us to run over all algebraic extensions J of Hπ, and compute d(A,B,UJ) following the algorithm given in Proposition 3.3, for each one. We do not see any shortcut to this procedure, for the general case.
Example 3.4**.**
We exhibit an explicit example of a subgroup H⩽fgG having two algebraic extensions J,J′∈AEFn(Hπ) with r~(J)<r~(J′) but r~(J)+d(A,B,UJ)>r~(J′)+d(A,B,UJ′).
Let H=⟨t(−1,0)b2,t(1,0)ac−1ac−1,t(0,1)bac−1⟩⩽fgG=Z2×F3. Projecting, we have Hπ=⟨b2,ac−1ac−1,bac−1⟩, and Fig. 1 represents the Stallings’ graph ΓA(Hπ) for Hπ as a subgroup of F3 with respect to the ambient free basis A={a,b,c}. Successively identifying pairs of vertices of ΓA(Hπ) and reducing the resulting A-labeled graph in all possible ways, one concludes that ΓA(Hπ) has nine congruences, whose corresponding quotient graphs are depicted in Figs. 1 and 2; this is the fringe of Hπ, OFn(Hπ); see the proof of Theorem 2.5 above.
Now following the cleaning process, we get the set of algebraic extensions for Hπ, namely AE(Hπ)={Hπ,J}, where J=⟨b,ac−1⟩}. (To this goal, the following fact helps: suppose N is obtained from M by a single identification of a pair of vertices followed by foldings; if r(N)=r(M)+1 then M is a free factor of N, otherwise, M⩽algN.)
Following the notation above, we have
[TABLE]
According to Theorem 1.9,
[TABLE]
Since H⩽H, d(A,B,UHπ)=r(LH)=0 and the first term on the minimum in (5) is r~(Hπ)+d(A,B,UHπ)=(3−1)+0=2.
Following the algorithm given in Proposition 3.3, let us compute now d(A,B,UJ), where J=⟨b,ac−1⟩; we have r=3, m=2, s=0, and p=2. Computing the Smith normal form for UJ, we get
[TABLE]
with d1=1, d2=2, and ℓ=min{r,p}=2. Diagonalyzing the problem, we obtain
[TABLE]
and d(A,B,UJ)=d(A′,B′,U′) (under the change of variable C=QC′). Since p=ℓ=2 the next reduction is empty and A′′=A′, B′′=B′, and U′′=U′. Applying the following reduction to delete the last r−ℓ=3−2=1 zero rows in U′′, we get
[TABLE]
Finally, in order to delete d1=1 from the list of divisors, we take c1′′′=(0,1) and get
[TABLE]
Going up by finite index, we replace the matrix B′′′′ to (0,1), and are reduced to compute d(A′′′′,(0,1),U′′′′); this is the smallest rank of a subgroup L⩽Z2 such that ⟨(0,1)⟩⩽L and (1,0)−2c2′′′′∈L for some c2′′′′∈Z2. Clearly, d(A′′′′,(0,1),U′′′′)=2, and one (non unique) solution is given by L=Z2 and c2′′′′=(1,0). Collecting the c1 computed before, and undoing the change of variable, we get
[TABLE]
We conclude that d(A,B,UJ)=2 and one of the subgroups K with the smallest possible rank satisfying Kπ=J and H⩽K⩽Z2×F3 is K=⟨t(−1,1)b,t(1,0)ac−1,t(1,0),t(0,1)⟩. So, the second term on the minimum in (5) is r~(J)+d(A,B,UJ)=(2−1)+2=3. Therefore,
[TABLE]
In particular, H is compressed in G.
As seen in this example, the algebraic extension J looks better than the other one Hπ because it contributes to the free rank in 2 units instead of 3. However, in order to match the free-abelian part, J forces us to take two more units of rank, while Hπ requires zero units. Note that in this example, d(A,B,UJ) is as big as it could be since, in general, d(A,B,UJ)⩽m=2. The example can easily be extended to an arbitrary m.
4. Degree of inertia in free-abelian times free groups
In this section, we study the degree of inertia for subgroups H of G=Zm×Fn and relate it to the corresponding degree of inertia of Hπ in Fn; it turns out that the index of H∩Zm in Zm (whether finite or infinite) is closely related to the degree of inertia of H. Unfortunately, the situation here is more complicated and we can only prove an upper bound for diG(⋅) in terms of diFn(⋅) and the mentioned index; the computability of this function remains open, as in the free case.
Lemma 4.1**.**
For positive real numbers a,b,c,d>0,
[TABLE]
Proof of Theorem 1.10.
(i). The hypothesis r(Hπ)⩽1 implies that H=⟨tau,LH⟩, for some a∈Zm and u∈Fn (possibly trivial). Then, for every K⩽fgG, we have (H∩K)π⩽Hπ∩Kπ⩽⟨u⟩ so, (H∩K)π=⟨ur⟩ for some r∈Z. Hence, H∩K=⟨tbur,LH∩LK⟩ for some b∈Zm and we get r(H∩K)⩽r(K). Therefore, r~(H∩K)/r~(K)⩽1, which is valid for every K⩽fgG. Thus, diG(H)=1 (i.e., H is inert in G).
(ii). Consider the (unique) subgroup L~H satisfying LH⩽fiL~H⩽⊕Zm, and take a free-abelian basis {b1,…,bs} of L~H, such that {λ1b1,…,λsbs} is a free-abelian basis of LH for appropriate integers λ1,…,λs∈Z (there is always a basis like this by standard linear algebra arguments). By hypothesis, s=r(LH)<m and, completing to a free-abelian basis {b1,…,bs,bs+1,…,bm} of the ambient Zm, we get at least one extra vector bs+1 (which, of course, is primitive in Zm and so has relatively prime coordinates).
Now fix a basis for H of the form {ta1u1,…,tan1un1,tλ1b1,…,tλsbs}, where a1,…,an1∈Zm, and {u1,…,un1} is a free basis for Hπ; in particular, we have r(Hπ)=n1⩾2, r(LH)=s<m, and r(H)=n1+s.
For proving diG(H)=∞, we shall construct a family of subgroups KN⩽fgZm×Fn, indexed by N∈N, all of them with constant rank 3 (i.e., r~(KN)=2), with all the intersections H∩KN being finitely generated, but with r~(H∩KN) tending to ∞, as N→∞.
Let KN=⟨ta1′u1,ta2′u2,LKN⟩⩽Zm×Fn, where the vectors a1′,a2′∈Zm and the subgroup LKN⩽Zm are to be determined; note that for all choices r(KNπ)=2, and here we are already using the hypothesis n1⩾2.
Let us understand the intersection H∩KN following the procedure (and notation) given in [4, Thm. 4.5]: we have n2=r(KNπ)=2, Hπ∩KNπ=⟨u1,u2⟩ and so n3=r(Hπ∩KNπ)=2, and we consider the matrices
[TABLE]
Let ρ1:Hπ↠Zn1, ρ2:KNπ↠Z2, and ρ3:Hπ∩KNπ↠Z2 be the corresponding abelianization maps (not to be confused with the restrictions of the global abelianization map Fn↠Zn to the corresponding domains). Clearly, the inclusion maps ιH:Hπ∩KNπ↪Hπ and ιK:Hπ∩KNπ↪KNπ abelianize, respectively, to the morphisms Z2→Zn1 and Z2→Z2 given by the matrices
[TABLE]
Moreover, let
[TABLE]
and let us put all these ingredients into the following diagram:
[TABLE]
According to the argument in [4, Thm. 4.5], the subgroup (H∩KN)π⩽Hπ∩KNπ is, precisely, the full preimage by R and ρ3 of LH+LKN⩽Zm.
Let us choose now the vectors a1′=a1−bs+1 and a2′=a2, and the subgroup LKN=⟨Nbs+1⟩, N∈N; the matrix R becomes
[TABLE]
We have LH+LKN=⟨λ1b1,…,λsbs,Nbs+1⟩ and then,
[TABLE]
(the last equality being true because bs+1 has relatively prime coordinates). As ρ3 is onto, taking ρ3-preimages preserves the index and we have
[TABLE]
Thus, by the Schreier index formula, r~((H∩KN)π)=Nr~(Hπ∩KNπ)=N and we deduce that r~(H∩KN)=N+r(LH∩LKN)=N+0=N tends to ∞, as N→∞. This completes the proof that diG(H)=∞.
(iii). Fix a basis for H, say {ta1u1,…,tan1un1,tb1,…,tbm}, where a1,…,an1∈Zm, {u1,…,un1} is a free basis for Hπ, and {b1,…,bm} is a free-abelian basis for LH⩽lZm; in particular, r(Hπ)=n1⩾2, r(LH)=m, and r(H)=n1+m.
In order to show the inequality diG(H)⩽ldiFn(Hπ), let us fix an arbitrary subgroup K⩽fgG, assume that H∩K is finitely generated, and let us prove that r~(H∩K)/r~(K)⩽ldiFn(Hπ). Fix a basis for K, say K=⟨ta1′v1,…,tan2′vn2,LK⟩ and we have
[TABLE]
As in the proof of part (ii), we consider the intersection diagram to understand H∩K:
[TABLE]
where ρ1:Hπ↠Zn1, ρ2:Kπ↠Zn2, and ρ3:Hπ∩Kπ↠Zn3 are the corresponding abelianization maps (here, n3=r(Hπ∩Kπ)<∞), where ι and ι′ are the natural inclusions, where P∈Mn3×n1(Z) and P′∈Mn3×n2(Z) are the matrices of their respective abelianizations (note that ι and ι′ being injective do not imply P and P′ necessarily being so; in particular, n3 may very well be bigger than n1 or n2), where A∈Mn1×m(Z) and A′∈Mn2×m(Z) are the matrices with rows {a1,…,an1} and {a1′,…,an2′} respectively, and where R=PA−P′A′∈Mn3×m(Z). According to the argument in [4, Thm. 4.5], the crucial property of diagram (8) is the fact that (H∩K)π=(LH+LK)R−1ρ3−1.
From the hypothesis, LH⩽lZm and so, LH+LK⩽l′Zm, where 1⩽l′⩽l. As in general R is not necessarily onto, (LH+LK)R−1⩽l′′Zn3 with 1⩽l′′⩽l′. And, since ρ3 is onto, (H∩K)π=(LH+LK)R−1ρ3−1⩽l′′Hπ∩Kπ. Therefore, by the Schreier index formula,
[TABLE]
Now, using (7), we have
[TABLE]
where the second inequality is an equality if LK={0}, and follows from applying Lemma 4.1 to r(LK)r(LH∩LK)⩽1⩽l′′diFn(Hπ) otherwise. Therefore
[TABLE]
as we wanted.
∎
5. Restricted degree of inertia for free-abelian times free groups
To improve the inequality from Theorem 1.10 (iii) into an equality, we need to add a couple of technical restrictions on the subgroups K over which the supremum in the definition of degree of inertia runs. This gives rise to the notion of restricted degree of inertia given in Definition 1.12; in the particular case of interest, G=Zm×Fn, it is the following:
[TABLE]
applied to subgroups H⩽fgG such that r(Hπ)⩾2 and Hπ[Fn,Fn]. The main result in the present section is Theorem 1.15. The proofs for part (i) and for the inequality ⩽ from (ii) work almost exactly equal as the corresponding parts from Theorem 1.10. The inequality diG(H)⩾ldiFn(Hπ) from (ii) is more involved and will require the previous development of several lemmas about intersections of subgroups of Fn, and a strong use of the well-known tool of pull-backs of graphs for working with intersections of finitely generated subgroups of Fn; we concentrate these technicalities into Claim 5.1 and postpone its proof until having the lemmas available.
Proof of Theorem 1.15.
(i). Follow the same arguments as in Theorem 1.10 (ii) with the following detail in mind: by the assumption Hπ∈[Fn,Fn] we can assume, from the very beginning and without loss of generality, that u1∈/[Fn,Fn], i.e., the first element in the chosen free-basis for Hπ is outside the commutator [Fn,Fn]. Now the goal is to construct a family of subgroups KN⩽fgZm×Fn, indexed by N∈N, all of them having rank 3, with all the intersections H∩KN being finitely generated, and further satisfying [Hπ:Hπ∩KNπ]=∞ and Hπ∩KNπ[Fn,Fn], such that r~(H∩KN) tends to infinity, as N→∞.
The construction of these KN’s will be similar to that in Theorem 1.10(ii), but with slight technical modifications in order to get the extra conditions. Take KN=⟨ta1′u12,ta2′u22,LKN⟩⩽Zm×Fn, where the vectors a1′,a2′∈Zm and the subgroup LKN⩽Zm are to be determined. Note that Hπ∩KNπ=⟨u12,u22⟩⩽∞Hπ, and also Hπ∩KNπ[Fn,Fn] as u12∈/[Fn,Fn] (since u1∈/[Fn,Fn] and Fn/[Fn,Fn]=Zn is torsion-free).
The rest of the argument works in a parallel way, just realizing that now P=(200200⋯⋯00)∈M2×n1(Z) and so,
[TABLE]
Choosing the vectors a1′=2a1−bs+1 and a2′=2a2, and the subgroup LKN=⟨Nbs+1⟩⩽Zm, the rest of the proof proceeds verbatim.
(ii)-⩽. In order to show the inequality diG′(H)⩽ldiFn′(Hπ), let us fix an arbitrary subgroup K⩽fgG, assume H∩K⩽fgG and also [Hπ:Hπ∩Kπ]=∞ and Hπ∩Kπ[Fn:Fn], and we have to prove that r~(H∩K)/r~(K)⩽ldiFn′(Hπ). Exactly the same arguments as in Theorem 1.10 (iii) work here, with the caution that the inequality in Equation (9) is still true with diFn(Hπ) replaced by the (eventually smaller) value diFn′(Hπ), since the involved subgroup Kπ further satisfies Hπ∩Kπ⩽∞Hπ and Hπ∩Kπ[Fn,Fn], by construction.
(ii)-⩾. By hypothesis, r(Hπ)⩾2 and so, the Stallings’ graph Γ(Hπ) has at least one vertex p of degree bigger than 2. Without loss of generality, we can assume that it is the basepoint ⊙ who has degree at least 3: in fact, let w∈Fn be the label of any path from ⊙ to p and, replacing H by Hw (and so, Hπ by Hwπ=(Hπ)w), the inequality to prove does not change; see Lemma 1.17.
Let {ta1u1,…,tan1un1,tb1,…,tbm} be a basis for the subgroup H⩽fgG, where a1,…,an1∈Zm, {u1,…,un1} is a free basis for Hπ, u1∈[Fn,Fn], and {b1,…,bm} is a free-abelian basis for LH⩽lZm; in particular, r(Hπ)=n1⩾2, r(LH)=m, and r(H)=n1+m.
In order to prove the inequality, diG′(H)⩾ldiFn′(Hπ), we fix ϵ>0 and will construct a subgroup Kϵ⩽fgG satisfying H∩Kϵ⩽fgG, [Hπ:Hπ∩Kϵπ]=∞, Hπ∩Kϵπ[Fn,Fn] and, furthermore, r~(H∩Kϵ)/r~(Kϵ)>ldiFn′(Hπ)−ϵ. For any candidate K⩽G, equations (9), (10), and (11) above (with diFn′ instead of diFn) contain all possible reasons for which the quotient r~(H∩K)/r~(K) may be less than ldiFn′(Hπ), namely:
- (I)
r~(Hπ∩Kπ)/r~(Kπ)⩽diFn′(Hπ);
2. (II)
r~(Kπ)+r(LK)l′′diFn′(Hπ)r~(Kπ)+r(LH∩LK)⩽r~(Kπ)l′′diFn′(Hπ)r~(Kπ);
3. (III)
l′⩽l;
4. (IV)
l′′⩽l′.
Choosing Kπ so that r~(Hπ∩Kπ)/r~(Kπ)>diFn′(Hπ)−ϵ we can make the inequality in (I) arbitrarily tight; choosing LK=0 inequalities (II) and (III) become equalities; and, finally, if the linear map R:Zn3→Zm from diagram (8) is onto then inequality (IV) becomes an equality. In view of these, we claim that
Claim 5.1**.**
Given ϵ>0, there exists M⩽fgFn (with a free basis {v1,…,vn2}), and there exist vectors a1′,…,an2′∈Zm such that:
[Hπ:Hπ∩M]=∞;
Hπ∩M[Fn,Fn];
r~(Hπ∩M)/r~(M)>diFn′(Hπ)−ϵ;
R=PA−P′A′:Zn3→Zm* is onto, where P,A,P′ are the matrices appearing in diagram (8) and A′ is the matrix with rows a1′,…,an2′∈Zm.*
Observe that the existence of M satisfying (i), (ii), and (iii) is immediate from the definition of diFn′(Hπ). Assuming, on top of these, condition (iv) is more tricky: the choice of M determines the ranks n2=r(M) and n3=r(Hπ∩M), and it could very well happen that n3<m, making then impossible to choose the vectors a1′,…,an2′∈Zm in such a way that R is onto. This situation forces us to manipulate M and make sure to get n3 big enough, to have enough freedom, to choose A′, so that R is onto; and all this without loosing the ϵ inequality (iii) (neither (i) nor (ii)). Here is where the extra technical conditions (i) and (ii) added to the definition of restricted degree of inertia are going to play a crucial role. Let us postpone the proof of the claim and continue with the main argument.
Given ϵ>0, apply Claim 5.1 to ϵ/l: we get M=⟨v1,…,vn2⟩⩽fgFn, and vectors a1′,…,an2′∈Zm satisfying (i), (ii), r~(Hπ∩M)/r~(M)>diFn′(Hπ)−ϵ/l, and (iv). The subgroup Kϵ=⟨ta1′v1,…,tan2′vn2⟩⩽fgG satisfies Kϵπ=M and LKϵ=0 hence,
H∩Kϵ⩽fgG, since (H∩Kϵ)π=(LH+LKϵ)R−1ρ3−1⩽l′′Hπ∩Kϵπ,
[Hπ:Hπ∩Kϵπ]=[Hπ:Hπ∩M]=∞,
Hπ∩Kϵπ=Hπ∩M[Fn,Fn],
and also
[TABLE]
[TABLE]
Therefore, diG′(H)⩾ldiFn′(Hπ) as we wanted to prove.
∎
Before proving Claim 5.1, we need to develop several lemmas about intersections of subgroups of Fn. A well-known tool for understanding these intersections is the pull-back of graphs.
Definition 5.2**.**
Let N,M⩽fgFn and consider its Stallings’ graphs Γ(N),Γ(M), respectively. Its direct product, Γ(N)×Γ(M), is defined as the new graph having as set of vertices VΓ(N)×VΓ(M), set of xi-labelled edges ExiΓ(N)×ExiΓ(M) (here, ExiΓ denotes the set of edges in Γ labelled by the letter xi∈X), with the natural incidence functions ι(e,f)=(ιe,ιf) and τ(e,f)=(τe,τf), and with basepoint being the pair of basepoints (⊙,⊙).
Clearly, Γ(N)×Γ(M) is folded, but neither connected nor free of degree one vertices, in general. The pull-back of Γ(N) and Γ(M), denoted Γ(N)∧Γ(M) is the result of trimming (i.e., repeatedly deleting vertices of degree one different from the basepoint) the connected component of the basepoint (⊙,⊙).
It is well known (see, for example, [10] for details) that Γ(N)∧Γ(M)≃Γ(N∩M), the Stallings’ graph for N∩M. In particular, if both N,M are finitely generated then so is N∩M; this is the Howson property for Fn.
Definition 5.3**.**
Let Γ(N) be the Stallings’ graph for N⩽fgFn. For every vertex p∈VΓ(N) and every element w∈Fn, we define pw to be the terminal vertex of the unique reduced path γ in Γ(N) starting at p and with label w, in case it exists; otherwise, pw is undefined. Note that w∈N if and only if ⊙w is defined and equals ⊙. Note also that N has finite index in Fn if and only if Γ(N) is complete and if and only if ⊙w is defined in Γ(N), for every w∈Fn.
Lemma 5.4**.**
Let N,M⩽fgFn, with the basepoint from Γ(N) having degree at least 3. Then, N∩M has infinite index in N if and only if there exists w∈N such that ⊙w is undefined in Γ(M).
Proof.
Suppose N∩M has finite index in N; so, there is r⩾1 such that, for every w∈N, wr∈N∩M. This means that, for every cyclically reduced w∈N, (⊙,⊙)wr, and so (⊙,⊙)w, is defined in Γ(N∩M); hence, projecting to Γ(M), ⊙w is defined in Γ(M). For those w∈N not cyclically reduced, write w=u−1⋅w′⋅u, without cancellations, with w′ being cyclically reduced, and with u=1; in this case, we just have that (⊙,⊙)wr=(⊙,⊙)u−1w′ru, and so (⊙,⊙)u−1w′, is defined in Γ(N∩M). But the basepoint in Γ(N) has degree at least 3 so we can take a cyclically reduced 1=v∈N such that the product w⋅v=u−1⋅w′⋅u⋅v∈N has no cancellation and is cyclically reduced again; then (⊙,⊙)wv, and so (⊙,⊙)w, is defined in Γ(N∩M) and, hence, ⊙w is defined in Γ(M).
For the other implication suppose that, for every w∈N, ⊙w is defined in Γ(M), say ⊙w∈{p0=⊙,p1,…,pr}⊆VΓ(M). Choose a maximal tree T in Γ(M) and define wi=ℓ(T[⊙,pi])∈Fn for i=0,…,r (note that w0=1). The hypothesis tells us that N⊆M⊔Mw1⊔⋯⊔Mwr. Intersecting with N, we get N⊆(N∩M)⊔(N∩M)v1⊔⋯⊔(N∩M)vs for some vi∈N and s⩽r (where we have deleted the possibly empty intersections). Since the other inclusion is immediate, we deduce that N∩M has finite index in N.
∎
Proposition 5.5**.**
[p-Expansion] Let N,M⩽fgFn, and suppose that r(N)⩾2, the basepoint ⊙ of Γ(N) has degree at least 3, and N∩M⩽∞N. Then, for every 1⩽p⩽∞, there exist p freely independent elements w1,…,wp∈N such that M⩽ffM′=M∗⟨w1,…,wp⟩ and N∩M⩽ff(N∩M)∗⟨w1,…,wp⟩⩽ffN∩M′⩽∞N.
Proof.
Let ea,eb,ec be three different edges going out from ⊙ in Γ(N), ιea=ιeb=ιec=⊙, with pairwise different labels a,b,c∈X±1, respectively. By Lemma 5.4, there is u0∈N such that ⊙u0 is undefined in Γ(M). Realize u0 as a reduced closed path γ0 at ⊙ in Γ(N) and, without lost of generality, we can assume it finishes with ea−1. For α=a,b,c, take a non-trivial reduced path ηα in the graph Γ(N)∖{eα} and closed at τeα (there always exists such a path because r(N)⩾2, even if eα is a bridge since Γ(N) has no vertices of degree 1 except possibly ⊙); now consider γα=eαηαeα−1 , a reduced closed path at ⊙ in Γ(N), beginning with eα and ending with eα−1 (so, its
label uα=ℓ(γα)∈N is a reduced word on X±1 beginning with α and ending with α−1). Note then that the paths γ0,γ1=γ0γb,γ2=γ0γbγa,γ3=γ0γbγaγb,…, and also the paths γiγcγi−1, i⩾1, are reduced as written; furthermore, all of them are closed paths at ⊙ in Γ(N) so, wi=ℓ(γiγcγi−1)∈N, for all i⩾1.
Now, let us extend the graph Γ(M) by adding the necessary vertices and edges so that we can read all the paths γiγcγi−1 from ⊙, i=1,…,p: since ⊙u0 was undefined in Γ(M), possibly an initial segment of γ0 is readable in Γ(M) but not the entire path, forcing us to append at least a new edge sticking out from Γ(M); behind it, we add the rest of the construction, see Fig. 3 (this means adding infinitely many new vertices and edges in the case p=∞). Since the added paths are all reduced, the resulting graph presents no foldings and so it is a (possibly infinite) Stallings’ graph, having Γ(M) as a subgraph. Hence, M is a free factor of its fundamental group, M⩽ffM′=M∗⟨w1,…,wp⟩.
And let us compare the pull-backs Γ(N)∧Γ(M)=Γ(N∩M) and Γ(N)∧Γ(M′)=Γ(N∩M′). Since wi∈N for all i⩾1, it is clear that Γ(N)∧Γ(M′) contains, as a subgraph, Γ(N)∧Γ(M) with the same additions as in Fig. 3, and, possibly, more edges which we do not control. Therefore, N∩M⩽ff(N∩M)∗⟨w1,…,wp⟩⩽ffN∩M′.
Finally, to see that N∩M′ is still an infinite index subgroup of N, observe that w=ℓ(γ0γc)∈N is such that ⊙w is not defined in Γ(M′) (see Fig. 3) and apply Lemma 5.4.
∎
Observation 5.6**.**
Let K⊴Fn be a normal subgroup of Fn. For any M⩽Fn, KM=⟨K,M⟩.
Proof.
It is obvious that KM={km∣k∈K,m∈M}⊆⟨K,M⟩. For the other inclusion note that, by normality, given m∈M and k∈K, there exists k′∈K such that mk=k′m. Repeated applications of this fact converts an arbitrary expression k1m1⋯krmr∈⟨K,M⟩, with ki∈K, mi∈M, into a single product km, k∈K, m∈M. Therefore, KM=⟨K,M⟩.
∎
Lemma 5.7**.**
Let G be a group and N,M⩽G. Then, [N:N∩M]⩽[G:M], with equality if MN=G. If additionally [N:N∩M] is finite, the equality holds if and only if MN=G.
Proof.
Let G=⊔i∈IMgi be the coset decomposition of G modulo M, where ∣I∣=[G:M]⩽∞. Intersecting with N (and removing the possibly empty terms), we have N=⊔i∈I(N∩Mgi)=⊔i′∈I′(N∩M)ni, for some I′⊆I and ni∈N. So, [N:N∩M]=∣I′∣⩽∣I∣=[G:M].
Furthermore, for g∈G, Mg intersects N non-trivially if and only if g∈MN. So, if G=MN then I′=I and [N:N∩M]=∣I′∣=∣I∣=[G:M] (with the converse being also true whenever ∣I′∣<∞).
∎
Corollary 5.8**.**
Let K⊴dFn, and M⩽Fn, then [M:M∩K]=d if and only if ⟨K,M⟩=Fn. ∎
Let us now fix a letter xj∈X, and an integer d∈Z, and consider the particular normal subgroup Kdj={w∈Fn∣∣w∣xj∈dZ}⊴Fn, where ∣w∣xj denotes the xj-th coordinate of the abelianization of w. The Stallings’ graph Γ(Kdj) is depicted in Fig. 4 (the loops at each vertex representing all the n−1 remaining letters). It is clear that Kdj⊴dFn.
Lemma 5.9**.**
Let M⩽fgFn, and fix a letter xj∈X and an integer d∈Z. Then the following conditions are equivalent:
- (a)
KdjM=⟨Kdj,M⟩=Fn;
2. (b)
M∩Kdj⩽dM;
3. (c)
there exists a word m∈M such that gcd(∣m∣xj,d)=1;
4. (d)
the direct product Γ(M)×Γ(Kdj) is connected.
Proof.
(a)⇔(b). True by Observation 5.6 and Corollary 5.8.
(a)⇒(c). From the hypothesis, xj∈Fn can be written as xj=km, for some k∈Kdj and some m∈M. Thus, 1=∣xj∣xj=∣k∣xj+∣m∣xj=λd+∣m∣xj for some λ∈Z and hence, gcd(∣m∣xj,d)=1.
(c)⇒(d). In the direct product Γ(M)×Γ(Kdj), consider the d full subgraphs Δi, i=0,…,d−1, whose vertices are VΔi={(p,i)∣p∈VΓ(M)}. In order to see that Γ(M)×Γ(Kdj) is connected, we shall prove the existence of a path from (⊙,i) to (⊙,i+1), for every i (indices modulo d), and that each Δi is connected.
In fact, by hypothesis there exists m∈M and α,β∈Z such that gcd(∣m∣xj,d)=α∣m∣xj+βd=1. Let γm denote the path in Γ(M), closed at ⊙, whose label is m, and note that ℓ(γmα)=mα. Furthermore, in the complete graph Γ(Kdj), i⋅mα is defined and equals i+1modd (by Bezout’s equality above). Hence, there is also a path labelled mα in Γ(M)×Γ(Kdj) from (⊙,i) to (⊙,i+1), for every i=0,…,d−1.
On the other hand, let p be any arbitrary vertex in Γ(M). As Γ(M) is connected, there is a path, say γ, from ι(γ)=⊙ to τ(γ)=p. Let w=ℓ(γ)∈Fn, let s=∣w∣xj, and consider the path γm−sαw starting at ⊙, reading m−sαw, and ending at p. Since ∣m−sαw∣xj=−sα∣m∣xj+∣w∣xj=0modd, and Kdj is complete, m−sαw is the label of a closed path in Γ(Kdj) at any vertex. Hence, there is a path in Δi⊆Γ(M)×Γ(Kdj) starting at (⊙,i) and ending at (p,i).
(d)⇒(b). The graph Γ(Kdj) has exactly d vertices, and d edges labelled by each letter; see Fig. 4. As Γ(M)×Γ(Kdj) is connected, we have
[TABLE]
Hence, by Schreier index formula, M∩Kdj⩽dM (and not less).
∎
With these ingredients we can already proof Claim 5.1.
Proof of Claim 5.1.
We are given a subgroup H⩽fgG with r(Hπ)⩾2, Hπ∈[Fn,Fn], with the basepoint in Γ(Hπ) having degree at least 3, with LH=H∩Zm⩽lZm (and so, r(LH)=m), and with basis {ta1u1,…,tan1un1,tb1,…,tbm}, such that u1∈[Fn,Fn]; we are also given ϵ>0, and we have to find a subgroup M⩽fgFn and vectors a1′,…,an2′∈Zm, n2=r(M), satisfying simultaneously conditions (i)-(iv). Let us distinguish two cases.
Case-1: diFn′(Hπ)>1. Making ϵ smaller if necessary, there always exists a subgroup M1⩽fgFn such that [Hπ:Hπ∩M1]=∞, Hπ∩M1[Fn,Fn] and
[TABLE]
Hence, both reduced ranks are r~(Hπ∩M1)⩾1 and r~(M1)⩾1 (recall that, in the definition of diFn′(⋅), 0/0 is understood to be 1). As Hπ∩M1[Fn,Fn], there exists v∈Hπ∩M1 and a letter xj∈X such that ∣v∣xj=λ=0. Write λ=p1α1⋯prαr, where each pi is a prime divisor of λ. Now choose a big enough prime number d≫0, such that gcd(λ,d)=1 and d>2mdiFn′(Hπ)/ϵ. We have \epsilon\operatorname{\tilde{r}}(M_{1})\big{(}d\operatorname{\tilde{r}}(M_{1})+m\big{)}\geqslant\epsilon d\operatorname{\tilde{r}}(M_{1})>2m\operatorname{di}^{\prime}_{F_{n}}(H\pi)\operatorname{\tilde{r}}(M_{1})\geqslant 2m\operatorname{\tilde{r}}(H\pi\cap M_{1}) and so,
[TABLE]
Dividing both sides by 2\big{(}d\operatorname{\tilde{r}}(M_{1})+m\big{)}\operatorname{\tilde{r}}(M_{1})\neq 0 we get
[TABLE]
Now, consider Kdj⊴dFn and M2:=M1∩Kdj. Since the element v∈Hπ∩M1⩽M1 satisfies gcd(∣v∣xj,d)=gcd(λ,d)=1, Lemma 5.9 tells us that M2⩽dM1 and Hπ∩M2=(Hπ∩M1)∩Kdj⩽dHπ∩M1, and also Γ(M1)∧Γ(Kdj)=Γ(M1)×Γ(Kdj) and Γ(Hπ∩M1)∧Γ(Kdj)=Γ(Hπ∩M1)×Γ(Kdj); note further that Γ(Hπ∩M1)×Γ(M2) is not necessarily connected, but its connected component containing the basepoint (after trimming, if necessary) coincides with Γ(Hπ∩M1)×Γ(Kdj), both being the Stallings’ graph of (Hπ∩M1)∩Kdj=(Hπ∩M1)∩M2; see Fig. 5, where every graph in the second and third columns is the pull-back of the one left to it and the one above it. Moreover, from the hypothesis we have Hπ∩M1⩽∞Hπ and so, Hπ∩M2⩽∞Hπ.
Now, let us apply Proposition 5.5, with p=m, to the pull-back of Γ(Hπ) and Γ(M2) (which equals Γ(Hπ∩M2)): we are under the hypothesis (r(Hπ)⩾2, the basepoint ⊙ of Γ(Hπ) has degree at least 3, and Hπ∩M2⩽∞Hπ) so we can perform an m-expansion to M2 and get m freely independent elements w1,…,wm∈Fn such that M2⩽ffM:=M2∗⟨w1,…,wm⟩ and Hπ∩M2⩽ff(Hπ∩M2)∗⟨w1,…,wm⟩⩽ffHπ∩M⩽∞Hπ; in particular, n2=r~(M)=r~(M2)+m and r~(Hπ∩M)⩾r~(Hπ∩M2)+m. This is our candidate subgroup M⩽fgFn: in fact, (i) Hπ∩M⩽∞Hπ; (ii) Hπ∩M1⩽[Fn,Fn] so, Hπ∩M2⩽[Fn,Fn] (since Fn/[Fn,Fn]≃Zn is torsion free) and Hπ∩M⩽[Fn,Fn]; and (iii), by the Schreier index formula and equations (12)-(13),
[TABLE]
It remains to choose appropriate vectors a1′,…,an2′∈Zm satisfying (iv). Take a free basis for M2, {v1,…,vk}, and extend it to a free basis for M, {v1,…,vk,w1,…,wm}. We have that n1=r(Hπ), k=r(M2), and n2=k+m=r(M). Similarly, as a free basis for Hπ∩M, take a free basis for Hπ∩M2 (say, made of q=r(Hπ∩M2) freely independent elements) followed by possibly some more, say p⩾0, and finally followed by {w1,…,wm}; we have n3:=r(Hπ∩M)=q+p+m⩾m. Finally, consider the intersection diagram for Hπ and M,
[TABLE]
where P∈M(q+p+m)×n1(Z) and P′∈M(q+p+m)×(k+m)(Z) are the abelianization of the inclusions Hπ∩M↪Hπ and Hπ∩M↪M, respectively, where A∈Mn1×m(Z) is the matrix with rows a1,…,an1, and where A′∈M(k+m)×m(Z) is the matrix with rows a1′,…,ak+m′ to be determined in such a way that R=PA−P′A′:Zq+p+m→Zm becomes onto.
Note that, by construction, the first q elements in the free basis for Hπ∩M are freely independent from {w1,…,wm}, and that {w1,…,wm} are present in the last positions of the chosen bases for both Hπ∩M and M; therefore, P′ has the form
[TABLE]
Let Q be the lower m×m block in PA∈M(q+p+m)×m(Z), and define
[TABLE]
Separating the rows in the natural blocks, we have that
[TABLE]
is a surjective map from Zq+p+m onto Zm, proving condition (iv). This concludes the proof for this case-1.
Case-2: diFn′(Hπ)=1. Recall that we already fixed a basis {ta1u1,…,tan1un1,tb1,…,tbm} for H, with u1∈[Fn,Fn]. Let M=⟨u1,u2−1u1u2,…,u2−(m−1)u1u2m−1⟩⩽fgHπ, a subgroup with n2=r(M)=m already satisfying the first three required conditions: (i) Hπ∩M=M⩽∞Hπ=⟨u1,…,un1⟩; (ii) u1∈Hπ∩M=M⩽[Fn,Fn]; and (iii) r~(Hπ∩M)/r~(M)=1>diFn′(Hπ)−ϵ, independently from the given ϵ.
Finally, we have to choose appropriate vectors a1′,…,am′∈Zm so that (iv) holds. To do this, look at the intersection diagram for Hπ and M (see Fig. 8): we have n3=n2=m and
[TABLE]
and, therefore,
[TABLE]
will become the identity Im, and so onto, after choosing A′ appropriately. This shows (iv) and concludes the proof.
∎
Acknowledgements. The first author thanks the support and hospitality from the Barcelona Graduate School of Mathematics and the Departament de Matemàtiques of the Universitat Politècnica de Catalunya. Both authors are partially supported by the Spanish Agencia Estatal de Investigación, through grant MTM2017-82740-P (AEI/ FEDER, UE), and also by the “María de Maeztu” Programme for Units of Excellence in R&D (MDM-2014-0445).