Cyclic JSJ Decompositions

• Cyclic JSJ decompositions are graph-of-groups structures that capture all splittings over infinite cyclic or virtually cyclic subgroups, offering a canonical deformation space.
• They leverage Bass–Serre theory to construct universally elliptic trees via trees of cylinders and peripheral vertex classifications, enabling algorithmic group analysis.
• They underpin advanced quasi-isometry classifications and boundary invariant computations, facilitating solutions to group isomorphism and JSJ algorithm challenges.

Cyclic JSJ decompositions are graph-of-groups decompositions that organize all splittings of a group over infinite cyclic subgroups; in formulations adapted to torsion or hyperbolicity, the edge groups are often taken to be two-ended or virtually cyclic. In the modern Bass–Serre formulation, a JSJ tree is a universally elliptic splitting that is maximal for domination, and the collection of all such trees forms a JSJ deformation space rather than a single preferred decomposition. Canonical representatives arise in many settings from trees of cylinders, from Bowditch’s boundary construction for hyperbolic groups, or from compatibility JSJ constructions (Guirardel et al., 2016, Cashen et al., 2016, Barrett, 2016).

1. General framework and deformation spaces

A one-edge cyclic splitting is expressed in Bass–Serre form as either an amalgamated product $G=A*_{C}B$ or an HNN extension $G=\mathrm{HNN}(A;C,\phi)$, where $C$ is cyclic or, in the broader two-ended formulation, virtually cyclic. The associated Bass–Serre tree records vertex stabilizers as conjugates of vertex groups and edge stabilizers as conjugates of edge groups. In the relative setting one fixes a family $\mathcal H$ of subgroups required to remain elliptic, and studies $(\mathcal A,\mathcal H)$-trees whose edge stabilizers lie in the chosen class $\mathcal A$ (Cashen et al., 2016, Guirardel et al., 2016).

Universal ellipticity is the organizing principle. A subgroup is universally elliptic if it fixes a point in every allowed splitting, and a JSJ tree is one whose edge stabilizers are universally elliptic and which dominates every other universally elliptic tree. For finitely presented groups, JSJ deformation spaces exist for arbitrary edge-group classes, hence in particular for cyclic and virtually cyclic classes; the deformation space is contractible and plays the role of the canonical object when no single preferred splitting exists (Guirardel et al., 2016).

This framework is deliberately broader than the classical one-edge language. It accommodates relative splittings, canonicalization procedures, and the later distinction between rigid vertices, flexible vertices, and canonical refinements such as trees of cylinders and compatibility JSJ trees. In this sense, “cyclic JSJ decomposition” denotes both a concrete graph of groups and the deformation-space structure underlying it (Guirardel et al., 2016).

2. Canonical trees, cylinders, and vertex types

A central canonicalization device is the tree of cylinders. Given a splitting over two-ended groups, edges in the Bass–Serre tree are declared equivalent when their stabilizers are commensurable; each equivalence class is a cylinder, and the dual tree to the covering by cylinders is the tree of cylinders. Guirardel–Levitt show that this construction depends only on the deformation space, so it yields a canonical tree. For one-ended hyperbolic groups with boundary not a circle, Bowditch’s canonical simplicial JSJ tree is equivariantly isomorphic to this tree of cylinders (Cashen et al., 2016, Guirardel et al., 2016).

The vertex-level structure is encoded by peripheral data. For a vertex $v$, the peripheral structure $P_v$ is the collection of coarse equivalence classes of incident edge groups inside the vertex group. In the two-ended JSJ studied by Cashen and Martin, every vertex group is either two-ended, hanging, or rigid. Hanging vertices are those for which $(G_v,P_v)$ is quasi-isometric to the universal cover of a hyperbolic pair of pants with peripheral structure coming from boundary components, while rigid vertices are neither two-ended nor hanging and admit no further splitting over two-ended subgroups relative to their incident edge groups (Cashen et al., 2016).

In the more general JSJ theory over slender edge groups, flexible non-slender vertices are quadratically hanging: they fit into an exact sequence

$1 \to F \to Q \to \pi_1(\Sigma) \to 1,$

where $G=\mathrm{HNN}(A;C,\phi)$0 is a compact hyperbolic $G=\mathrm{HNN}(A;C,\phi)$1-orbifold and incident edge groups map to finite or boundary subgroups. In the torsion-free cyclic case, the fiber is trivial, $G=\mathrm{HNN}(A;C,\phi)$2 has no mirrors, boundary components are used by incident edge groups or prescribed elliptic subgroups, and non-elliptic actions of $G=\mathrm{HNN}(A;C,\phi)$3 are dual to families of essential simple closed geodesics on $G=\mathrm{HNN}(A;C,\phi)$4 (Guirardel et al., 2016).

These constructions separate canonical information from presentation-dependent information. The original JSJ graph of groups may vary inside the deformation space, but the tree of cylinders is deformation-space invariant and, in many classes, automorphism invariant as well (Guirardel et al., 2016, Cashen et al., 2016).

3. Hyperbolic groups, boundaries, and relative boundaries

For one-ended hyperbolic groups, Bowditch constructs a canonical JSJ decomposition over virtually cyclic subgroups from topological features of the boundary, especially cut points and local cut pairs. In this setting the JSJ is automorphism invariant, universally elliptic, maximal for domination, and algorithmically accessible through the geometry of the boundary and of the cusped space. The algorithm of Barrett and Cashen computes Bowditch’s canonical VC-JSJ from large finite balls in the cusped space, detects cut points and cut pairs via thickened cylinders around quasi-geodesics, and then produces the cyclic $G=\mathrm{HNN}(A;C,\phi)$5-JSJ by refining hanging Fuchsian vertices through the mirrors splitting and collapsing all dihedral edges; a further folding step yields the $G=\mathrm{HNN}(A;C,\phi)$6-JSJ (Barrett, 2016).

Boundary methods also govern relative free-group JSJ theory. Given a finite-rank free group together with a finite family of conjugacy classes of maximal cyclic subgroups, one forms the relative boundary $G=\mathrm{HNN}(A;C,\phi)$7 by collapsing the two endpoints of each peripheral axis. Cut points and uncrossed cut pairs in $G=\mathrm{HNN}(A;C,\phi)$8 determine a canonical simplicial tree with cocompact $G=\mathrm{HNN}(A;C,\phi)$9-action, and the quotient graph of groups is the JSJ decomposition of $C$0 over cyclic subgroups relative to the multiclass. In this relative setting, non-cyclic vertex groups are rigid or QH-surface type, and a multiclass is virtually geometric if and only if the relative boundary is planar (Cashen, 2010).

The same boundary viewpoint interacts with coarse geometry. For one-ended hyperbolic groups, a boundary homeomorphism induces an isomorphism between trees of cylinders, preserves vertex types, and restricts to homeomorphisms on vertex boundaries. This makes the boundary not merely a detector of splittings, but a canonical topological package from which the cyclic JSJ may be recovered and compared (Cashen et al., 2016).

4. Explicit realizations in major group classes

In generalized Baumslag–Solitar and quadratic Baumslag–Solitar settings, cyclic JSJ decompositions can often be recognized directly from the defining graph of groups. Forester proved that, except for the elementary cases $C$1, $C$2, and the Klein bottle group, reduced GBS splittings are cyclic JSJ decompositions. Guirardel and Levitt’s deformation-space language clarifies the uniqueness statement, while the QBS extension shows that a reduced QBS graph with all edge labels $C$3 and with each GBS component reduced and non-degenerate is a Rips–Sela cyclic JSJ decomposition; the QH vertices record exactly the surface-type pieces and their boundary subgroups (Alonso, 2011).

For right-angled Artin groups, the combinatorics of the defining graph determines both the existence of cyclic splittings and the JSJ itself. If $C$4 has at least three vertices, then $C$5 does not split over $C$6 if and only if $C$7 is biconnected. For one-ended RAAGs, the cyclic JSJ is built from the block tree of cut vertices and maximal biconnected subgraphs, and the associated tree of cylinders admits an explicit description in terms of cut vertices and certain maximal biconnected induced subgraphs. In the one-ended RAAG case there are no quadratically hanging vertices, so the canonical tree alternates cylindrical and rigid pieces only (Clay, 2014, Margolis, 2018).

For Artin groups, Jones, Mangioni, and Sartori prove that $C$8 splits over an infinite cyclic subgroup if and only if the defining graph has a separating vertex or has exactly two vertices. They construct an explicit cyclic JSJ $C$9 from “big chunks,” separating vertices, toral leaves, and braided even leaves. The resulting graph of groups is universally elliptic and dominates every universally elliptic virtually cyclic splitting; although the chosen JSJ graph depends on the Artin generating set, its JSJ deformation space is canonical and its tree of cylinders is $\mathcal H$0-invariant (Jones et al., 20 Jun 2025).

There are also higher-dimensional analogues. For orientable $\mathcal H$1 pairs, the Scott–Swarup regular-neighborhood construction yields JSJ-type decompositions in which every edge splitting is either canonical or a special canonical torus. In the torsion-free $\mathcal H$2 case, two-ended virtually cyclic subgroups are infinite cyclic, and the decomposition becomes an algebraic annulus–torus theory parallel to the classical $\mathcal H$3-manifold picture (Reeves et al., 2020).

5. Quasi-isometry, boundary classification, and structure invariants

In groups with two-ended splittings, the JSJ tree of cylinders can be decorated and converted into a coarse-geometric invariant. Cashen and Martin assign ornaments encoding vertex type and relative quasi-isometry type, refine them by neighbor refinement, cylinder refinement, and vertex refinement, and package the stable data into a matrix-valued structure invariant $\mathcal H$4. For finitely presented one-ended groups not commensurable with a surface group, this invariant is preserved by quasi-isometries; under the technical hypotheses that cylinder stabilizers are two-ended and every non-elementary vertex is relatively rigid or hanging, equality of the resulting invariants reduces global quasi-isometry classification to relative quasi-isometry classification of the vertex groups together with realizability, orientation, and stretch data (Cashen et al., 2016).

The same decorated-tree technology governs boundary classification for one-ended hyperbolic groups. In that setting the initial decorations use vertex type and relative boundary homeomorphism type rather than stretch factors, and boundary homeomorphism $\mathcal H$5 is characterized by matching structure invariants together with local realizability and cylinder-orientation data. The boundary theorem is formally parallel to the quasi-isometry theorem, but stretch disappears because it is invisible to boundary homeomorphism (Cashen et al., 2016).

The role of stretch is particularly explicit in RAAG quasi-isometry theory. Quasi-isometric one-ended RAAGs have weakly equivalent JSJ trees of cylinders, but the converse fails in general. Relative stretch factors attached to rigid standard geodesics along cyclic edges are quasi-isometry invariants, and for the class $\mathcal H$6 of one-ended RAAGs whose rigid vertex groups lie in the dovetail class, the embellished decoration consisting of vertex types, relative quasi-isometry classes, and relative stretch factors is a complete quasi-isometry invariant (Margolis, 2018).

Concrete examples show the distinction between boundary and quasi-isometry invariants. In the HNN examples

$\mathcal H$7

the relative stretch factors across the rigid edge are $\mathcal H$8, $\mathcal H$9, and $(\mathcal A,\mathcal H)$0 for $(\mathcal A,\mathcal H)$1 respectively; these groups are pairwise non-quasi-isometric, yet their boundaries are homeomorphic because the stretch data is not seen by boundary homeomorphism (Cashen et al., 2016).

6. Algorithms, isomorphism problems, and open directions

The algorithmic theory of cyclic JSJ decompositions is now distributed across several classes. For one-ended hyperbolic groups, Bowditch’s VC-JSJ can be computed from large balls in the cusped space and then converted to the $(\mathcal A,\mathcal H)$2-JSJ and $(\mathcal A,\mathcal H)$3-JSJ by mirrors splitting, collapsing dihedral edges, and maximal-$(\mathcal A,\mathcal H)$4 folding; the method avoids Makanin’s algorithm and uses the geometry of the Cayley graph and the cusped space (Barrett, 2016). For torsion-free one-ended hyperbolic groups that are graphs of free groups with cyclic edge groups, Touikan gives a double-exponential-time algorithm based on immersed cycles in CAT(0) square complexes, with the first explicit time bound for a JSJ algorithm in this setting (S, 2018).

In highly specialized classes the complexity is sharper. For hyperbolic two-generator one-relator groups with abelianization $(\mathcal A,\mathcal H)$5, the non-triviality of the $(\mathcal A,\mathcal H)$6-JSJ is equivalent to the Friedl–Tillmann polytope being a straight line but not a point. This yields an $(\mathcal A,\mathcal H)$7 algorithm for computing the $(\mathcal A,\mathcal H)$8-JSJ and, in the torsion-free BS-free case, an $(\mathcal A,\mathcal H)$9 algorithm for deciding whether the decomposition is non-trivial (Gardam et al., 2021).

Recent work also addresses isomorphism questions at the graph-of-groups level. New “swap” and “connection” moves on cyclic graphs of groups preserve the fundamental group and allow witnesses of isomorphism without expansions, so the number of vertices and edges remains constant along the sequence. For a large family of cyclic JSJ decompositions, the isomorphism problem reduces to generalized Baumslag–Solitar groups, and among GBS groups it further reduces to one-vertex graphs; this yields a solution for a broad class of flexible GBSs and reframes vertex elimination as a controlled move system rather than an expansion–collapse process (Ascari et al., 22 Jul 2025).

Several limitations remain explicit in the literature. The complete quasi-isometry classification of groups with two-ended splittings depends on the open question of whether every one-ended hyperbolic group satisfies the condition that every non-elementary JSJ vertex is relatively rigid or hanging (Cashen et al., 2016). In RAAG theory, weak equivalence of JSJ trees of cylinders is not complete without stretch data, and it remains open whether every RAAG is dovetail (Margolis, 2018). For Artin groups, extensions from cyclic splittings to splittings over higher-rank or spherical Artin subgroups are posed as a further direction (Jones et al., 20 Jun 2025). More generally, the deformation-space viewpoint remains essential: cyclic JSJ theory often supplies canonical deformation spaces and canonical trees of cylinders even when no single canonical JSJ graph of groups exists (Guirardel et al., 2016).