Embeddings ∗i∈IGi↪Aut⁡(F∑ini+∣I∣+1)\ast_{i\in I}G_i\hookrightarrow\operatorname{Aut}(F_{\sum_i n_i+|I|+1}) from separating cocycles

18.11

Problem

If A,B≤Aut⁡(Fn)A,B\le\operatorname{Aut}(F_n) with n≥2n\ge2, does A∗BA*B embed in Aut⁡(Fm)\operatorname{Aut}(F_m) for some finite mm?

A,B≤Aut⁡(Fn)⟹?A∗B↪Aut⁡(Fm).A,B\le\operatorname{Aut}(F_n) \quad\stackrel{?}{\Longrightarrow}\quad A*B\hookrightarrow\operatorname{Aut}(F_m).

The construction below gives m=2n+3m=2n+3 and extends to finite indexed families with different initial ranks.

The embedding theorem

Write F(X)F(X) for the free group on a set XX, and put Fn=F({x1,…,xn})F_n=F(\{x_1,\ldots,x_n\}), with F0=1F_0=1. Automorphisms act on the left, so (αβ)(x)=α(β(x))(\alpha\beta)(x)=\alpha(\beta(x)). All free products below are free products of groups.

Theorem 1. Let II be a finite set, let ni≥0n_i\geq0 for i∈Ii\in I, and let Gi≤Aut⁡(Fni)G_i\leq\operatorname{Aut}(F_{n_i}). There is an injective homomorphism

∗i∈IGi⟶Aut⁡(FR),R=∑i∈Ini+∣I∣+1.(1) \tag{1} \mathop{\ast}_{i\in I}G_i \longrightarrow \operatorname{Aut}(F_R), \qquad R=\sum_{i\in I}n_i+|I|+1.

The groups GiG_i need not be finitely generated.

Corollary 2. If G≤Aut⁡(Fn)G\leq\operatorname{Aut}(F_n) and H≤Aut⁡(Fm)H\leq\operatorname{Aut}(F_m), then

G∗H↪Aut⁡(Fn+m+3).(2) \tag{2} G*H\hookrightarrow\operatorname{Aut}(F_{n+m+3}).

In particular, if both groups are subgroups of Aut⁡(Fn)\operatorname{Aut}(F_n), then G∗H↪Aut⁡(F2n+3)G*H\hookrightarrow\operatorname{Aut}(F_{2n+3}).

Proof. Apply Theorem 1 to a two-element index set. ◻

Problem 18.11, proposed by V. G. Bardakov, asks whether the free product of two subgroups of Aut⁡(Fn)\operatorname{Aut}(F_n), for n≥2n\geq2, embeds in Aut⁡(Fm)\operatorname{Aut}(F_m) for some finite mm (2, Problem 18.11). Corollary 2 supplies the explicit bound m=2n+3m=2n+3. The construction also permits different initial ranks.

McCullough and Miller (4) study symmetric automorphisms of free products. A related embedding problem is treated by Marchand (3, Corollary 5.5), who constructs embeddings Out⁡(Q)↪Out⁡(F)\operatorname{Out}(Q)\hookrightarrow\operatorname{Out}(F) for suitable free products QQ of a free group and finite abelian groups, with FF a finite-index free subgroup of QQ. In Theorem 1, the source is a free product of subgroups of free-group automorphism groups, and the target is itself an automorphism group. We obtain the embedding from the cocycle criterion of Theorem 8 below.

Separating cocycles and free products

Definition 3. Let ρ:A→Aut⁡(B)\rho:A\to\operatorname{Aut}(B) be a homomorphism. A normalized nonabelian cocycle for ρ\rho is a map v:A→Bv:A\to B satisfying

v(1)=1,v(gh)=v(g)ρ(g)(v(h))(g,h∈A).(3) \tag{3} v(1)=1,\qquad v(gh)=v(g)\rho(g)(v(h))\quad(g,h\in A).

We call vv separating if v(g)=1v(g)=1 implies g=1g=1.

Let II now be an arbitrary set. For each i∈Ii\in I, suppose that AiA_i and BiB_i are groups, αi:Ai→Aut⁡(Bi)\alpha_i:A_i\to\operatorname{Aut}(B_i) is a homomorphism, and ci:Ai→Bic_i:A_i\to B_i satisfies

ci(1)=1,ci(ab)=ci(a)αi(a)(ci(b)),a,b∈Ai,ci(a)=1 ⟹ a=1.\begin{align*} c_i(1)&=1,\tag{4a}\\ c_i(ab)&=c_i(a)\alpha_i(a)(c_i(b)), &&a,b\in A_i,\tag{4b}\\ c_i(a)=1&\ \Longrightarrow\ a=1. \tag{4c} \end{align*}

Put

A=∗i∈IAi,B=∗i∈IBi,P=∏i∈IAut⁡(Bi).(5) \tag{5} A=\mathop{\ast}_{i\in I}A_i, \qquad B=\mathop{\ast}_{i\in I}B_i, \qquad P=\prod_{i\in I}\operatorname{Aut}(B_i).

We identify the factors with their canonical images in the corresponding free products. For p=(pi)∈Pp=(p_i)\in P, let p‾∈Aut⁡(B)\overline p\in\operatorname{Aut}(B) be the automorphism whose restriction to BiB_i is pip_i. The inverse is p−1‾\overline{p^{-1}}, and pq‾=p‾ q‾\overline{pq}=\overline p\,\overline q. For a∈Aia\in A_i, define α^i(a)∈P\widehat\alpha_i(a)\in P by

α^i(a)j={αi(a),j=i,id⁡Bj,j≠i.(6) \tag{6} \widehat\alpha_i(a)_j= \begin{cases} \alpha_i(a),&j=i,\\ \operatorname{id}_{B_j},&j\ne i. \end{cases}

Lemma 4. There are a homomorphism ρ:A→Aut⁡(B)\rho:A\to\operatorname{Aut}(B) and a normalized cocycle v:A→Bv:A\to B such that

ρ(a)=α^i(a)‾,v(a)=ci(a)(a∈Ai).(7) \tag{7} \rho(a)=\overline{\widehat\alpha_i(a)},\qquad v(a)=c_i(a)\quad(a\in A_i).

Every ρ(g)\rho(g) preserves each free factor BiB_i and restricts to an automorphism of that factor.

Proof. Use the componentwise action to form B⋊PB\rtimes P, with multiplication

(b,p)(d,q)=(b p‾(d),pq).(8) \tag{8} (b,p)(d,q)=\bigl(b\,\overline p(d),pq\bigr).

Equations (4a) and (4b) give homomorphisms

Ai⟶B⋊P,a⟼(ci(a),α^i(a)).A_i\longrightarrow B\rtimes P, \qquad a\longmapsto\bigl(c_i(a),\widehat\alpha_i(a)\bigr).

Their free-product extension is a homomorphism L:A→B⋊PL:A\to B\rtimes P. Write

L(g)=(v(g),σ(g)),ρ(g)=σ(g)‾.L(g)=(v(g),\sigma(g)),\qquad \rho(g)=\overline{\sigma(g)}.

Projection onto PP shows that σ\sigma, and hence ρ\rho, is a homomorphism. The first coordinate of L(gh)=L(g)L(h)L(gh)=L(g)L(h) is exactly (3). The restrictions (7) follow from the definition of LL on each factor. Since σ(g)∈P\sigma(g)\in P, the last assertion follows from the definition of the componentwise action. ◻

Lemma 5. The cocycle vv in Lemma 4 is separating.

Proof. Let g≠1g\ne1 and write its reduced free-product normal form as

g=a1⋯aℓ,ℓ≥1,ar∈Air∖{1},ir≠ir+1(1≤r<ℓ).g=a_1\cdots a_\ell, \qquad \ell\geq1,\quad a_r\in A_{i_r}\setminus\{1\},\quad i_r\ne i_{r+1}\quad(1\leq r<\ell).

Set g0=1g_0=1 and gr=a1⋯arg_r=a_1\cdots a_r. Repeated application of (3) gives

v(g)=b1⋯bℓ,br=ρ(gr−1)(cir(ar)).(12) \tag{12} v(g)=b_1\cdots b_\ell, \qquad b_r=\rho(g_{r-1})\bigl(c_{i_r}(a_r)\bigr).

By (4c), cir(ar)≠1c_{i_r}(a_r)\ne1. Lemma 4 therefore gives br∈Bir∖{1}b_r\in B_{i_r}\setminus\{1\}. The indices i1,…,iℓi_1,\ldots,i_\ell are unchanged, so the right-hand side of (12) is a nonempty reduced word in BB. The free-product normal-form theorem implies v(g)≠1v(g)\ne1. ◻

One additional free generator

Let CC be a group and put C+=C∗⟨z⟩C^+=C*\langle z\rangle, where ⟨z⟩\langle z\rangle is infinite cyclic.

The following added-generator construction is standard. Automorphisms fixing CC and sending zz to zuzu occur, for example, in Bardakov and Mikhailov (1, proof of Theorem 4).

Lemma 6. For α∈Aut⁡(C)\alpha\in\operatorname{Aut}(C) and u∈Cu\in C, the assignments

Tα,u(c)=α(c)(c∈C),Tα,u(z)=zu(13) \tag{13} T_{\alpha,u}(c)=\alpha(c)\quad(c\in C), \qquad T_{\alpha,u}(z)=zu

define an automorphism of C+C^+. These automorphisms satisfy

Tα,uTβ,w=Tαβ, uα(w),Tα,u−1=Tα−1, α−1(u−1).\begin{align*} T_{\alpha,u}T_{\beta,w} &=T_{\alpha\beta,\,u\alpha(w)},\tag{14a}\\ T_{\alpha,u}^{-1} &=T_{\alpha^{-1},\,\alpha^{-1}(u^{-1})}. \tag{14b} \end{align*}

Proof. The universal property of C∗⟨z⟩C*\langle z\rangle first gives a homomorphism Tα,uT_{\alpha,u} with the stated values. On CC, the composite in (14a) restricts to αβ\alpha\beta. On the remaining generator,

(Tα,uTβ,w)(z)=Tα,u(zw)=zuα(w).(T_{\alpha,u}T_{\beta,w})(z) =T_{\alpha,u}(zw)=zu\alpha(w).

The same universal property proves (14a). Moreover, Tid⁡,1=id⁡C+T_{\operatorname{id},1}=\operatorname{id}_{C^+}, and the two choices in (14b) give, in the two orders, the pairs

(id⁡,uα(α−1(u−1)))=(id⁡,1),(id⁡,α−1(u−1)α−1(u))=(id⁡,1).\begin{aligned} \bigl(\operatorname{id},u\alpha(\alpha^{-1}(u^{-1}))\bigr)&=(\operatorname{id},1),\\ \bigl(\operatorname{id},\alpha^{-1}(u^{-1})\alpha^{-1}(u)\bigr)&=(\operatorname{id},1). \end{aligned}

Thus (14b) is a two-sided inverse. ◻

Proposition 7. Let ρ:A→Aut⁡(C)\rho:A\to\operatorname{Aut}(C) be a homomorphism and let v:A→Cv:A\to C be a normalized cocycle. Then

Θ:A⟶Aut⁡(C∗⟨z⟩),Θ(g)=Tρ(g),v(g)(17) \tag{17} \Theta:A\longrightarrow\operatorname{Aut}(C*\langle z\rangle), \qquad \Theta(g)=T_{\rho(g),v(g)}

is a homomorphism. If vv is separating, then Θ\Theta is injective.

Proof. The identity element is preserved because ρ(1)=id⁡\rho(1)=\operatorname{id} and v(1)=1v(1)=1. By (14a) and the cocycle identity,

Θ(g)Θ(h)=Tρ(g)ρ(h), v(g)ρ(g)(v(h))=Tρ(gh),v(gh)=Θ(gh).\Theta(g)\Theta(h) =T_{\rho(g)\rho(h),\,v(g)\rho(g)(v(h))} =T_{\rho(gh),v(gh)}=\Theta(gh).

If Θ(g)=id⁡\Theta(g)=\operatorname{id}, evaluation at zz gives zv(g)=zzv(g)=z. Cancellation, followed by injectivity of the canonical map C→C∗⟨z⟩C\to C*\langle z\rangle, gives v(g)=1v(g)=1. If vv is separating, then g=1g=1, proving injectivity. ◻

Theorem 8. Let II be any set, and suppose that the groups Ai,BiA_i,B_i, homomorphisms αi:Ai→Aut⁡(Bi)\alpha_i:A_i\to\operatorname{Aut}(B_i), and maps ci:Ai→Bic_i:A_i\to B_i satisfy (4a)–(4c). There is an injective homomorphism

∗i∈IAi⟶Aut⁡ ⁣((∗i∈IBi)∗⟨z⟩).(19) \tag{19} \mathop{\ast}_{i\in I}A_i \longrightarrow \operatorname{Aut}\!\left(\left(\mathop{\ast}_{i\in I}B_i\right) *\langle z\rangle\right).

Its restriction to AiA_i sends aa to the automorphism whose restriction to BiB_i is αi(a)\alpha_i(a), whose restriction to every BjB_j with j≠ij\ne i is the identity, and which sends zz to zci(a)zc_i(a).

Proof. Lemmas 4 and 5 give an action ρ\rho on B=∗i∈IBiB=\mathop{\ast}_{i\in I}B_i and a separating cocycle v:A→Bv:A\to B. Apply Proposition 7. Equation (7) gives the stated restrictions. ◻

Markers for free-group automorphisms

The separating condition in Theorem 8 can be obtained from a point with trivial stabilizer.

Lemma 9. Let α:A→Aut⁡(C)\alpha:A\to\operatorname{Aut}(C) be a homomorphism and let w∈Cw\in C. Then

c(a)=w−1α(a)(w)(20) \tag{20} c(a)=w^{-1}\alpha(a)(w)

is a normalized cocycle. It is separating if the stabilizer of ww under α\alpha is trivial.

Proof. We have c(1)=w−1w=1c(1)=w^{-1}w=1, and

c(a)α(a)(c(b))=w−1α(a)(w) α(a)(w−1α(b)(w))=w−1α(ab)(w)=c(ab).\begin{align*} c(a)\alpha(a)(c(b)) &=w^{-1}\alpha(a)(w)\, \alpha(a)\bigl(w^{-1}\alpha(b)(w)\bigr)\\ &=w^{-1}\alpha(ab)(w)=c(ab). \end{align*}

Also c(a)=1c(a)=1 is equivalent to α(a)(w)=w\alpha(a)(w)=w, which gives the last assertion. ◻

For a group CC and a finite list (y1,…,yr)(y_1,\ldots,y_r) of its elements, define

μC(y1,…,yr)=y1t⋯yrt∈C∗⟨t⟩,μC(∅)=1,(22) \tag{22} \mu_C(y_1,\ldots,y_r)=y_1t\cdots y_rt \in C*\langle t\rangle, \qquad \mu_C(\varnothing)=1,

where ⟨t⟩\langle t\rangle is infinite cyclic.

Lemma 10. If every yjy_j and every yj′y'_j is nonidentity, then

μC(y1,…,yr)=μC(y1′,…,ys′)⟹r=s  and  yj=yj′ (1≤j≤r).\mu_C(y_1,\ldots,y_r)=\mu_C(y'_1,\ldots,y'_s) \quad\Longrightarrow\quad r=s\ \text{ and }\ y_j=y'_j\ (1\leq j\leq r).

Proof. For r>0r>0, the expression y1t⋯yrty_1t\cdots y_rt is a reduced free-product word of syllable length 2r2r: its syllables are nonidentity and alternate between CC and ⟨t⟩\langle t\rangle. The same holds on the right when s>0s>0. Uniqueness of reduced normal forms gives equal lengths and equal corresponding syllables. If either list is empty, its product is the identity; the other list must then be empty as well. ◻

Lemma 11. Let X={x1,…,xn}X=\{x_1,\ldots,x_n\} be a free basis, put B=F(X)∗⟨t⟩B=F(X)*\langle t\rangle, and set

w=x1t⋯xnt.(24) \tag{24} w=x_1t\cdots x_nt.

Extend each α∈Aut⁡(F(X))\alpha\in\operatorname{Aut}(F(X)) to α‾∈Aut⁡(B)\overline\alpha\in\operatorname{Aut}(B) by fixing tt. If α‾(w)=w\overline\alpha(w)=w, then α=id⁡\alpha=\operatorname{id}.

Proof. The free-product universal property gives the extension and the identity αβ‾=α‾ β‾\overline{\alpha\beta}=\overline\alpha\,\overline\beta. Each xjx_j is nonidentity, so each α(xj)\alpha(x_j) is nonidentity. The assumed equality reads

α(x1)t⋯α(xn)t=x1t⋯xnt.\alpha(x_1)t\cdots\alpha(x_n)t=x_1t\cdots x_nt.

Lemma 10 gives α(xj)=xj\alpha(x_j)=x_j for every jj. An endomorphism of a free group is determined by its values on the free basis, hence α=id⁡\alpha=\operatorname{id}. The same argument applies when n=0n=0: the basis is empty and the free group has only its identity automorphism. ◻

The finite-family construction and its rank

Proof of Theorem 1. Choose pairwise disjoint bases Xi={xi,1,…,xi,ni}X_i=\{x_{i,1},\ldots,x_{i,n_i}\}, marker letters tit_i, and one further letter zz. Define

Bi=F(Xi)∗⟨ti⟩,wi=xi,1ti⋯xi,niti.(26) \tag{26} B_i=F(X_i)*\langle t_i\rangle, \qquad w_i=x_{i,1}t_i\cdots x_{i,n_i}t_i.

Extend g∈Gig\in G_i to g‾∈Aut⁡(Bi)\overline g\in\operatorname{Aut}(B_i) by fixing tit_i, and put ci(g)=wi−1g‾(wi)c_i(g)=w_i^{-1}\overline g(w_i). By Lemmas 9 and 11,

ci(1)=1,ci(gh)=ci(g)g‾(ci(h)),ci(g)=1⟹g=1.(27) \tag{27} c_i(1)=1,\qquad c_i(gh)=c_i(g)\overline g(c_i(h)),\qquad c_i(g)=1\Longrightarrow g=1.

Theorem 8 therefore gives an embedding into Aut⁡(E)\operatorname{Aut}(E), where

E=(∗i∈IBi)∗⟨z⟩.E=\left(\mathop{\ast}_{i\in I}B_i\right)*\langle z\rangle.

The automorphism assigned to g∈Gig\in G_i is determined by

Θi(g)∣Bi=g‾,Θi(g)∣Bj=id⁡Bj(j≠i),Θi(g)(z)=z wi−1g‾(wi).\begin{align*} \Theta_i(g)|_{B_i}&=\overline g, &\Theta_i(g)|_{B_j}&=\operatorname{id}_{B_j}\quad(j\ne i), \tag{29a}\\ \Theta_i(g)(z)&=z\,w_i^{-1}\overline g(w_i). \tag{29b} \end{align*}

It remains to identify EE. Let

Y=(∐i∈IXi)⨿{ti:i∈I}⨿{z}.Y=\left(\coprod_{i\in I}X_i\right) \amalg\{t_i:i\in I\}\amalg\{z\}.

The inclusions of the named generators give a homomorphism F(Y)→EF(Y)\to E. Conversely, their images in F(Y)F(Y) give homomorphisms from every BiB_i and from ⟨z⟩\langle z\rangle, hence a homomorphism E→F(Y)E\to F(Y). The two composites fix every named generator, so both are identity homomorphisms by the respective universal properties. Thus

E≅F(Y),∣Y∣=∑i∈I(ni+1)+1=∑i∈Ini+∣I∣+1=R.(31) \tag{31} E\cong F(Y),\qquad |Y|=\sum_{i\in I}(n_i+1)+1 =\sum_{i\in I}n_i+|I|+1=R.

Choose an isomorphism e:E→FRe:E\to F_R. The map γ↦eγe−1\gamma\mapsto e\gamma e^{-1} is an isomorphism Aut⁡(E)→Aut⁡(FR)\operatorname{Aut}(E)\to\operatorname{Aut}(F_R); composing it with the constructed embedding proves (1). ◻

The rank calculation applies to an empty family as well: the source is the trivial group and E=⟨z⟩E=\langle z\rangle. Zero ranks cause no change to the proof, since Aut⁡(F0)=1\operatorname{Aut}(F_0)=1 and the corresponding marker is the empty product.

For comparison, let k=∣I∣≥2k=|I|\geq2. Iterate Corollary 2, identifying each preceding free product with its embedded image before adjoining the next factor. The iterated and simultaneous bounds are

Rbin=∑i∈Ini+3(k−1),R=∑i∈Ini+k+1,Rbin−R=2k−4.\begin{align*} R_{\mathrm{bin}}&=\sum_{i\in I}n_i+3(k-1), \tag{32a}\\ R&=\sum_{i\in I}n_i+k+1, \qquad R_{\mathrm{bin}}-R=2k-4. \tag{32b} \end{align*}

Thus finite-family existence already follows by binary iteration; using one common generator zz saves 2k−42k-4 generators in this comparison. No minimality is asserted. For a singleton family the given inclusion G1≤Aut⁡(Fn1)G_1\leq\operatorname{Aut}(F_{n_1}) already uses fewer generators.

References

Preprint · Lean (GitHub)

  1. V. G. Bardakov and R. Mikhailov, On certain questions of the free group automorphisms theory, Comm. Algebra 36 (2008), no. 4, 1489–1499. doi:10.1080/00927870701866929.
  1. E. I. Khukhro and V. D. Mazurov (eds.), Unsolved Problems in Group Theory: The Kourovka Notebook, 21st ed., version 46 (2026), Problem 18.11. arXiv:1401.0300v46.
  1. A. Marchand, Free representations of outer automorphism groups of free products via characteristic abelian coverings, J. Group Theory 26 (2023), no. 2, 399–420. doi:10.1515/jgth-2021-0154.
  1. D. McCullough and A. Miller, Symmetric automorphisms of free products, Mem. Amer. Math. Soc. 122 (1996), no. 582, viii+97 pp. doi:10.1090/memo/0582.