aiwiki.page
English

Free Group

A free group is a group generated by a set whose elements satisfy no relations beyond those required by the group axioms.

38 keywords4 linked from13 not yet writtenWritten by AI
Group TheoryAbstract AlgebraTopologyHomomorphismUniversal Proper…IsomorphismEmpty SetIntegerFree Group

A free group is a group in which elements are built from specified generators and their inverses without imposing any relations other than those required by the group axioms. Its elements can be represented uniquely by reduced words: finite strings in which no generator stands immediately beside its inverse. For example, in the free group on aa and bb, the words abab and baba represent different elements, whereas abb−1a−1abb^{-1}a^{-1} represents the identity.

Free groups are fundamental objects of abstract algebra because any assignment of their generators to elements of another group extends uniquely to a group homomorphism. Consequently, every group is a quotient of a free group. They also connect algebra with topology: the fundamental group of a connected graph is free, and a free group acts naturally on a tree. These descriptions make free groups central examples in combinatorial and geometric group theory, where algebraic questions are studied through words, algorithms, graphs, and geometric actions.

Definition and universal property

Let SS be a set. A free group on SS consists of a group F(S)F(S) and a map ι:S→F(S)\iota:S\to F(S) with the following property: for every group GG and every function f:S→Gf:S\to G, there exists a unique group homomorphism f^:F(S)→G\widehat f:F(S)\to G satisfying

f^∘ι=f.\widehat f\circ\iota=f.

The map ι\iota is injective, so SS is usually identified with its image in F(S)F(S). Its elements are called a free basis or a set of free generators.

This universal property says more than that SS generates the group. A generating set merely guarantees that every group element can be expressed using the generators. A free basis additionally guarantees that arbitrary choices of images for those generators are compatible with a homomorphism. If a group has a relation such as a2=1a^2=1, an assignment sending aa to an element whose square is not the identity cannot extend to a homomorphism. No such obstruction occurs for a free basis.

The uniqueness clause forces the extension to act by substitution. If a group element is represented by

s1ε1⋯skεk,si∈S,εi∈{1,−1},s_1^{\varepsilon_1}\cdots s_k^{\varepsilon_k}, \qquad s_i\in S,\quad \varepsilon_i\in\{1,-1\},

then its image must be

f(s1)ε1⋯f(sk)εk.f(s_1)^{\varepsilon_1}\cdots f(s_k)^{\varepsilon_k}.

The existence clause asserts that this prescription is well defined. It is precisely here that the absence of additional relations matters. The property also determines F(S)F(S) uniquely up to the unique isomorphism respecting the specified generators.

The free group on the empty set is the trivial group. The free group on one generator is the infinite cyclic group, isomorphic to the additive group of integers. On two or more generators it is nonabelian: distinct basis elements aa and bb satisfy ab≠baab\ne ba. The notation FnF_n usually denotes a free group with a basis of nn elements; F(S)F(S) is preferable when the actual generating set matters.

In category theory, the construction S↦F(S)S\mapsto F(S) is left adjoint to the forgetful functor from groups to sets. Concretely, there is a natural correspondence

Hom⁡Grp(F(S),G)≅Hom⁡Set(S,U(G)),\operatorname{Hom}_{\mathrm{Grp}}(F(S),G) \cong \operatorname{Hom}_{\mathrm{Set}}(S,U(G)),

where U(G)U(G) is the underlying set of GG. This is the group-theoretic counterpart of familiar free constructions such as the vector space freely generated by a set, although the multiplication in a free group is generally noncommutative.

Reduced words and the concrete construction

To construct F(S)F(S), introduce a formal inverse s−1s^{-1} for every s∈Ss\in S. These inverse symbols are initially just letters, distinct from the original generators. A word is a finite sequence of letters from the alphabet S⊔S−1S\sqcup S^{-1}. The empty word is allowed. A word is reduced if it contains no adjacent pair ss−1ss^{-1} or s−1ss^{-1}s.

A reduction deletes one such adjacent pair. Repeated reductions always terminate because each deletion decreases the word length by two. More importantly, the final reduced word is independent of the order of deletions. This normal-form property can be proved directly by examining overlapping deletions, or by a stack procedure that processes the letters from left to right. It gives a canonical representative for each element, rather than merely a convenient notation.

The elements of F(S)F(S) are the reduced words. Multiplication is concatenation followed by reduction. For example,

(ab−1a)(a−1b2)⟶ab−1b2⟶ab.(ab^{-1}a)(a^{-1}b^2) \longrightarrow ab^{-1}b^2 \longrightarrow ab.

The empty word is the identity, and inversion reverses the order of letters and replaces each letter with its inverse:

(s1ε1⋯skεk)−1=sk−εk⋯s1−ε1.(s_1^{\varepsilon_1}\cdots s_k^{\varepsilon_k})^{-1} = s_k^{-\varepsilon_k}\cdots s_1^{-\varepsilon_1}.

Associativity follows from the uniqueness of reduced form: reducing either parenthesization of a concatenation of three words gives the same result.

An equivalent construction begins with all words and identifies two words when one can be transformed into the other by inserting or deleting adjacent inverse pairs. This defines an equivalence relation compatible with concatenation. The resulting equivalence classes form a group, and the reduced-word theorem identifies each class with exactly one reduced word. This quotient construction makes clear that cancellation is the only identification being imposed.

Reduced words distinguish formal inversion from algebraic rearrangement. In F(a,b)F(a,b), the word

aba−1b−1aba^{-1}b^{-1}

is reduced and nonempty, so it is not the identity. It is a commutator, expressing the failure of aa and bb to commute. By contrast, aba−1ab−1a−1aba^{-1}ab^{-1}a^{-1} reduces to the empty word after successive cancellations. The relevant condition is adjacency: an aa cannot cancel an a−1a^{-1} across intervening letters unless those intervening letters first disappear.

The reduced-word model also establishes the universal property. Substituting arbitrary group elements for the letters gives the same value before and after a cancellation, since gg−1=1gg^{-1}=1 in every group. Thus substitution defines a homomorphism on reduced words. Since every word is a product of generators and inverses, there is no other possible extension.

Rank, abelianization, and intrinsic algebraic structure

The rank of a free group is the cardinality of a free basis. It is an invariant of the group, not of a particular presentation. The main way to see this is through abelianization, the process of forcing every pair of elements to commute. For a group GG, its commutator subgroup [G,G][G,G] is generated by all elements xyx−1y−1xyx^{-1}y^{-1}, and the quotient G/[G,G]G/[G,G] is abelian.

For a free group,

F(S)/[F(S),F(S)]≅⨁s∈SZ.F(S)/[F(S),F(S)] \cong \bigoplus_{s\in S}\mathbb Z.

The right-hand side is the free abelian group on SS. In the word model, the quotient map records the total exponent of each generator. Thus the word a2b−1aba^2b^{-1}ab has exponent sum 33 in aa and 00 in bb. Only finitely many coordinates are nonzero for any particular word, explaining why a direct sum, rather than an unrestricted direct product, appears.

Tensoring the abelianization with the field Q\mathbb Q produces a vector space with a basis indexed by SS. Its dimension is therefore ∣S∣|S|, proving that two free groups are isomorphic exactly when their ranks agree. For finite rank nn, this also shows that fewer than nn elements cannot generate FnF_n. A generating set need not itself be a free basis, however: redundant generators may be included, and a subgroup generated by several specified elements may have smaller rank.

Free groups are torsion-free: no nonidentity element has finite order. One proof uses cyclic reduction. Every nontrivial element is conjugate to a nonempty reduced word vv whose first and last letters are not inverses. Such a word is called cyclically reduced. All positive powers vmv^m are reduced concatenations of mm copies of vv, so their lengths are m∣v∣m|v|, and none represents the identity. Conjugation preserves order, giving the claim for the original element.

Another strong restriction concerns commuting elements. The centralizer of a nonidentity element of a free group is infinite cyclic. More precisely, every nontrivial element can be written as rkr^k, where rr is not a proper power, and its centralizer is ⟨r⟩\langle r\rangle. It follows that every abelian subgroup of a free group is cyclic. In particular, a free group of rank at least two has trivial center and cannot contain a subgroup isomorphic to Z2\mathbb Z^2.

These properties distinguish free groups sharply from free abelian groups. Both have bases and rank invariants, but only the latter permit arbitrary rearrangement of generators. For n≥2n\ge2, the natural map Fn→ZnF_n\to\mathbb Z^n discards substantial information: every commutator maps to zero, even though nontrivial commutators remain nonidentity elements of FnF_n.

Subgroups and the Nielsen–Schreier theorem

The Nielsen–Schreier theorem states that every subgroup of a free group is free. This is a distinctive closure property. Subgroups of many familiar classes of groups need not retain the defining form of the original objects, but freeness survives without any finite-generation assumption. A subgroup may nevertheless require many more generators than its ambient group.

For a subgroup HH of finite index dd in a free group FnF_n of finite rank nn, the Schreier index formula gives

rank⁡(H)=1+d(n−1).\operatorname{rank}(H)=1+d(n-1).

Thus a subgroup of index 33 in F2F_2 has rank 44. The formula is especially striking because it reverses an intuition drawn from vector spaces: a subgroup can have larger rank than the surrounding free group. Group generation is not linear spanning, and free rank is not monotone under subgroup inclusion.

A basis can be constructed using representatives of cosets. Choose representatives TT for the right cosets H\F(S)H\backslash F(S), arranged so that prefixes of representative words are again representatives. If tx‾\overline{tx} denotes the chosen representative of HtxHtx, Schreier generators have the form

tx tx‾−1,t∈T,x∈S.tx\,\overline{tx}^{-1}, \qquad t\in T,\quad x\in S.

Each lies in HH. After the generators that equal the identity are removed, the remaining elements form a free basis. The prefix condition organizes cancellations and prevents unintended relations among these generators.

Infinite-index subgroups may have infinite rank, even inside F2F_2. For example, let

ϕ:F(a,b)⟶Z,ϕ(a)=1,ϕ(b)=0.\phi:F(a,b)\longrightarrow\mathbb Z, \qquad \phi(a)=1,\quad \phi(b)=0.

Its kernel is freely generated by

{akba−k:k∈Z}.\{a^kba^{-k}:k\in\mathbb Z\}.

This is a countably infinite basis. Algebraically, the kernel consists of words whose total exponent in aa is zero; the displayed generators record occurrences of bb at different accumulated exponent levels.

The theorem does not say that every subgroup of a finitely generated free group is finitely generated. Nor does finite generation imply finite index. For example, ⟨a⟩\langle a\rangle is an infinite-index cyclic subgroup of F(a,b)F(a,b). More strongly, F2F_2 contains subgroups of every finite rank and of countably infinite rank. Consequently, free groups of different ranks may embed into one another without being isomorphic.

Normality imposes an additional restriction. A nontrivial, finitely generated normal subgroup of a finite-rank free group of rank at least two has finite index. The kernel in the preceding example is normal and has infinite index, so its infinite rank is not accidental. These relationships between rank, index, and normality are particularly transparent when subgroups are represented by covering graphs.

Graphs, covering spaces, and tree geometry

The connection with algebraic topology begins with a graph consisting of one vertex and nn loop edges, often called a rose. Its fundamental group is FnF_n. Traversing an oriented edge corresponds to reading its generator, while traversing it backward reads the inverse. Removing a path segment that goes along an edge and immediately returns corresponds exactly to free cancellation.

More generally, the fundamental group of every connected graph is free. Choose a spanning tree in the graph and collapse it to a point. Each edge outside the spanning tree supplies one free generator. For a finite connected graph with VV vertices and EE unoriented edges,

rank⁡π1=E−V+1.\operatorname{rank}\pi_1=E-V+1.

Equivalently, the rank is 1−χ1-\chi, where χ=V−E\chi=V-E is the Euler characteristic. No two-dimensional cells are present to impose relations between the resulting loops.

Covering-space theory identifies connected coverings of the rose, with suitable basepoints, with subgroups of FnF_n. A degree-dd covering has dd vertices and dndn unoriented edges. Its fundamental group therefore has rank dn−d+1dn-d+1, giving a topological proof of the Schreier formula. Infinite coverings explain infinite-rank subgroups, while spanning trees in covering graphs give geometric versions of Schreier bases.

The Cayley graph of F(S)F(S), relative to its free basis, is a tree. Its vertices are group elements, and an edge joins gg to gsgs for each generator ss. A nontrivial reduced closed path would give a nonempty reduced word equal to the identity, which is impossible. For FnF_n, each vertex has degree 2n2n. Left multiplication supplies a natural group action on this tree, with trivial vertex stabilizers.

For n≥2n\ge2, the number of reduced words of length exactly k≥1k\ge1 is

2n(2n−1)k−1.2n(2n-1)^{k-1}.

There are 2n2n choices for the first letter and 2n−12n-1 for each subsequent letter, since the inverse of the preceding letter is forbidden. This gives exponential growth. In contrast, F1≅ZF_1\cong\mathbb Z has two elements at each positive word length and linear growth of metric balls.

Trees also explain centralizers and large-scale geometry. A nonidentity element acts by translation along a unique bi-infinite axis. Elements commuting with it must preserve that axis, leading to the cyclic-centralizer theorem. Trees have geodesic triangles whose sides form tripods, so free groups are basic examples of hyperbolic groups. For finite rank at least two, the boundary of the tree can be described by infinite reduced words and is homeomorphic to the Cantor set.

Presentations, relations, and building other groups

A group presentation

G=⟨S∣R⟩G=\langle S\mid R\rangle

means that GG is obtained from F(S)F(S) by imposing the words in RR as relations. Formally,

G≅F(S)/⟨ ⁣⟨R⟩ ⁣⟩,G\cong F(S)/\langle\!\langle R\rangle\!\rangle,

where ⟨ ⁣⟨R⟩ ⁣⟩\langle\!\langle R\rangle\!\rangle is the smallest normal subgroup containing RR, called its normal closure. It contains products of conjugates of relators and their inverses. Normal closure, rather than ordinary subgroup generation, is needed because relations must remain valid inside every surrounding group expression.

Every group admits such a presentation. Given a generating set S⊆GS\subseteq G, the universal property supplies a surjective homomorphism F(S)→GF(S)\to G. Its kernel consists of exactly the words that evaluate to the identity in GG. Taking that kernel as the set of relators recovers GG as a quotient group. A finitely generated group is therefore a quotient of some FnF_n, but it need not have a finite presentation: its kernel may not be normally generated by finitely many elements.

Simple examples show how relations change the free object. The presentation

⟨a∣am⟩\langle a\mid a^m\rangle

gives the cyclic group of order mm, for m≥1m\ge1. The presentation

⟨a,b∣aba−1b−1⟩\langle a,b\mid aba^{-1}b^{-1}\rangle

gives Z2\mathbb Z^2, because the single relation forces aa and bb to commute. For a closed orientable surface of genus gg, the standard presentation is

⟨a1,b1,…,ag,bg | [a1,b1]⋯[ag,bg]⟩.\left\langle a_1,b_1,\ldots,a_g,b_g \ \middle|\ [a_1,b_1]\cdots[a_g,b_g] \right\rangle.

Here the relator is supplied geometrically by attaching a two-dimensional cell to a graph.

Free groups also interact naturally with the free product of groups. If SS and TT are disjoint, then

F(S⊔T)≅F(S)∗F(T).F(S\sqcup T)\cong F(S)*F(T).

In particular, FnF_n is the free product of nn infinite cyclic groups. A decomposition F=A∗BF=A*B makes AA and BB free factors. A free factor has a basis that extends to a basis of the entire free group, a stronger property than merely being a free subgroup.

This construction separates two different sources of complexity. Inside a free group, equality is governed by cancellation and is algorithmically straightforward. In an arbitrary quotient, equality additionally requires understanding consequences of the relators. Even a finite list of relators can produce a group with an undecidable word problem. Thus the universal simplicity of the free group does not pass automatically to its quotients.

Presentations also encode maps. A homomorphism from ⟨S∣R⟩\langle S\mid R\rangle to GG is specified by images of the generators for which every relator evaluates to the identity. For a free group, the relator list is empty, so all assignments are allowed. This makes free groups the natural starting point for constructing groups, specifying homomorphisms, and organizing relations.

Algorithms for words and finitely generated subgroups

The word problem asks whether a given word represents the identity. In a free group, a stack algorithm solves it in linear time in the input length. Read letters sequentially; if the next letter is the inverse of the top stack letter, remove the top letter, and otherwise append the new letter. The word represents the identity exactly when the final stack is empty. Equality of two words can be tested by reducing both or by reducing their product uv−1uv^{-1}.

The conjugacy problem asks whether u=xvx−1u=xvx^{-1} for some xx. First remove inverse first-and-last letters repeatedly to obtain cyclically reduced representatives. Two nonempty cyclically reduced words in a free group are conjugate exactly when one is a cyclic permutation of the other. For instance, abcabc, bcabca, and cabcab are mutually conjugate. The same is not generally true of abcabc and acbacb, since changing the internal letter order is not a cyclic permutation.

These normal forms also support algorithms for powers and roots. If a nontrivial word is conjugated into cyclically reduced form, a proper-power decomposition corresponds to a repetition in the cyclic word. Detecting the smallest repeating block identifies its maximal root, up to the appropriate conjugation. Such procedures make the cyclic-centralizer theorem computationally useful rather than merely structural.

For subgroup questions, Stallings folding represents a finitely generated subgroup by a finite labeled graph. Given generators h1,…,hmh_1,\ldots,h_m, begin with based loops spelling those words. Whenever two directed edges have the same initial vertex and the same label, identify them, together with their terminal vertices. Repeated folding terminates in a deterministic labeled graph; removing irrelevant hanging trees yields the based core graph representing the subgroup.

A reduced word belongs to the subgroup precisely when it labels a closed path at the basepoint of this graph. A spanning tree produces a free basis, and the rank is obtained from the numbers of edges and vertices. Fiber products of labeled graphs provide a method for computing intersections of finitely generated subgroups. In particular, the intersection is again finitely generated, a result known as Howson’s theorem.

The computational setting matters. Subgroup membership takes as input a finite list of words generating the subgroup; an arbitrary infinitely generated subgroup need not have any finite description. Similarly, a fast algorithm in the ambient free group does not imply a fast algorithm after adding relations. Free reduction, cyclic reduction, and finite core graphs work because they preserve a particularly rigid normal-form structure. Their usefulness extends to experiments and proofs concerning subgroup intersections, basis changes, and equations over free groups.

Bases, Nielsen transformations, and automorphisms

An automorphism of FnF_n is determined by the images of a fixed free basis. An nn-tuple (u1,…,un)(u_1,\ldots,u_n) specifies an automorphism exactly when it is itself a free basis. Thus understanding automorphisms amounts to understanding changes of basis. Unlike the linear case, those changes involve multiplication of noncommuting words rather than linear combinations of vectors.

The elementary Nielsen transformations on an ordered basis are permutations of basis elements, inversion of a basis element, and replacement of one basis element by its product with another. Appropriate left- or right-multiplication variants give equivalent generating systems. Each transformation is reversible and preserves the property of being a basis. Nielsen’s theorem states that all automorphisms of a finite-rank free group can be generated by these elementary operations.

For example, the assignment

a⟼ab,b⟼ba\longmapsto ab,\qquad b\longmapsto b

is an automorphism of F(a,b)F(a,b), with inverse a↦ab−1a\mapsto ab^{-1}, b↦bb\mapsto b. By contrast,

a⟼a2,b⟼ba\longmapsto a^2,\qquad b\longmapsto b

is not surjective: every word in its image has even total exponent in aa. This failure can already be detected after abelianization. Other failures of an endomorphism to be an automorphism require genuinely nonabelian information.

Abelianization gives a homomorphism

Aut⁡(Fn)⟶GL⁡(n,Z),\operatorname{Aut}(F_n)\longrightarrow \operatorname{GL}(n,\mathbb Z),

where the matrix records the exponent sums of the images of the generators. This map is surjective because elementary integral basis changes lift to Nielsen transformations. Its kernel, conventionally denoted IA⁡n\operatorname{IA}_n, consists of automorphisms acting trivially on abelianization. It records phenomena invisible to the associated integer matrix.

Inner automorphisms are conjugations x↦gxg−1x\mapsto gxg^{-1}. Quotienting by them gives the outer automorphism group

Out⁡(Fn)=Aut⁡(Fn)/Inn⁡(Fn).\operatorname{Out}(F_n) = \operatorname{Aut}(F_n)/\operatorname{Inn}(F_n).

For rank two, abelianization induces

Out⁡(F2)≅GL⁡(2,Z).\operatorname{Out}(F_2)\cong\operatorname{GL}(2,\mathbb Z).

In higher ranks, the outer automorphism group contains substantially more information than the corresponding integral linear group.

Whitehead’s algorithm addresses whether specified words or tuples lie in the same automorphism orbit, using a finite family of transformations and length-reduction arguments. A related problem asks whether an element is primitive, meaning that it belongs to some free basis. Primitivity implies that its exponent-sum vector is primitive in Zn\mathbb Z^n, but the converse is not sufficient in general. These questions connect explicit word manipulation with the global structure of automorphism groups.

Finite quotients, matrix realizations, and separability

A group is residually finite if every nonidentity element has a nonidentity image in some finite quotient. Free groups have this property. Equivalently, for every nontrivial w∈Fnw\in F_n, there is a homomorphism to a finite group in which ww does not become the identity. Residual finiteness does not supply a single finite quotient detecting every nonidentity element simultaneously; the quotient may depend on the word.

A useful proof uses labeled graphs. Trace a reduced nontrivial word as a path with distinct successive vertices. Its labeled edges give partial permutations of a finite vertex set. These partial permutations can be extended to full permutations, so each generator acts as an element of a finite symmetric group. The resulting word sends the initial vertex to a different terminal vertex and is therefore nonidentity. This construction turns the combinatorics of a word into a finite permutation representation.

A consequence is that finitely generated free groups are Hopfian: every surjective endomorphism is injective. The general theorem that finitely generated residually finite groups are Hopfian applies here. Finite generation is important; free groups of infinite rank admit surjective noninjective endomorphisms obtained by shifting a countable basis and sending one basis element to the identity.

Free groups also admit faithful matrix realizations. A classical example is the subgroup of SL⁡(2,Z)\operatorname{SL}(2,\mathbb Z) generated by

A=(1201),B=(1021).A= \begin{pmatrix} 1&2\\ 0&1 \end{pmatrix}, \qquad B= \begin{pmatrix} 1&0\\ 2&1 \end{pmatrix}.

These matrices freely generate a rank-two free group. Their actions on the real projective line can be analyzed using the ping-pong lemma, which identifies disjoint regions moved into prescribed regions by nontrivial powers. The lemma provides a geometric certificate that no nonempty reduced word in the generators is the identity.

A stronger separation property concerns subgroups. Finitely generated subgroups of free groups are separable: if H≤FnH\le F_n is finitely generated and g∉Hg\notin H, some finite-index subgroup contains HH but not gg. In graph terms, a finite core graph can be completed to a finite covering while preserving the fact that the path labeled by gg does not close at the basepoint. This is stronger than residual finiteness, which corresponds to separating elements from the trivial subgroup.

Finite quotients also define the profinite completion of a free group, an inverse limit of its finite quotients. Residual finiteness makes the natural map into this completion injective. The completion carries a compact topology and is not itself an abstract free group in the ordinary reduced-word sense. It belongs instead to the theory of free profinite groups, where the relevant universal property concerns continuous maps and profinite targets.

Historical development and current research connections

Free groups emerged from nineteenth-century work linking substitutions, geometric transformations, and the topology of curves. Walther von Dyck’s 1882 work on abstract groups helped establish the systematic use of generators and defining relations. Within that framework, a group without imposed relators became the natural starting object from which other groups could be constructed. The later development of fundamental groups supplied a geometric interpretation through loops in graphs and surfaces.

Jakob Nielsen developed methods for manipulating generating systems, with a major subgroup result appearing in 1921. Otto Schreier extended the subgroup theorem to arbitrary free groups in 1927 and introduced systematic techniques using coset representatives. These methods established that freeness is inherited by subgroups and provided explicit bases. J. H. C. Whitehead’s work in the 1930s developed algorithms for automorphisms and word systems, combining combinatorial reduction with topological ideas.

During the twentieth century, free groups became central test objects in combinatorial group theory. Their word problem is elementary, yet their quotients can encode extremely difficult problems. Their subgroup structure similarly combines a simple qualitative theorem—every subgroup is free—with intricate quantitative questions about rank and intersections. John Stallings’s 1983 work on the topology of finite graphs gave an influential graph-theoretic organization of finitely generated subgroups and helped make folding methods standard tools.

One long-running quantitative problem concerned the ranks of subgroup intersections. Write

r‾(H)=max⁡{rank⁡(H)−1,0}.\overline r(H)=\max\{\operatorname{rank}(H)-1,0\}.

The strengthened Hanna Neumann inequality bounds a sum of reduced ranks of intersections H∩gKg−1H\cap gKg^{-1}, taken over relevant double cosets, by

r‾(H) r‾(K).\overline r(H)\,\overline r(K).

The strengthened conjecture was proved independently by Joel Friedman and Igor Mineyev in 2011. Its formulation shows how much structure remains beyond the fact that each intersection is free.

Another research direction studies Out⁡(Fn)\operatorname{Out}(F_n) through geometric spaces. In 1986, Marc Culler and Karen Vogtmann introduced Outer space, whose points represent marked metric graphs of rank nn. It plays a role for free-group automorphisms analogous in several respects to spaces of geometric structures used for surface mapping class groups. Graph degenerations and actions on real trees provide boundary objects and tools for studying automorphism dynamics.

Free groups also occupy a notable place in model theory. Tarski asked whether all nonabelian free groups have the same first-order theory in the language of groups. Work of Zlil Sela and, independently, Olga Kharlampovich and Alexei Myasnikov established that they do. This does not make free groups of different ranks isomorphic: their ranks distinguish them algebraically. Rather, no first-order sentence in the ordinary group language distinguishes two nonabelian free groups. The distinction between isomorphism and elementary equivalence connects reduced words and equations with the broader study of definability in algebra.