Cactus Configurations: Theory & Applications

• Cactus configurations are combinatorial structures defined by strict sparsity conditions where cycles may overlap only at limited points.
• They integrate techniques from graph theory, matroid theory, and group theory to yield explicit extremal bounds, realization algorithms, and structured decompositions.
• Their applications span theoretical analysis in algebraic geometry and topology as well as practical design in high-performance computing frameworks.

Cactus configurations are a family of objects organized around a cactus-type sparsity principle: cyclic structure is allowed, but overlap is tightly constrained. The expression is used in several distinct settings. In rank-three matroid theory, a cactus configuration is a point-line configuration built from lines and cycles by free gluing; equivalently, every line lies in at most one cycle, or the associated graph $G(M)$ is a cactus graph (Vandebrouck, 19 Aug 2025, Liwski et al., 9 Jun 2025). In group theory and topology, cactus configurations are interval-reversing or circular interval-reversing strand configurations encoded by cactus groups and their affine and virtual analogues (Bellingeri et al., 2022, Chemin, 27 Jan 2025, Ilin et al., 2023). In scientific computing, “Cactus Configurations” denotes the thorn-specific configuration files of the Cactus Framework, written in the Cactus Configuration Language (CCL), which specify variables, parameters, scheduling, and inter-component dependencies (Allen et al., 2010).

1. Graph-theoretic prototype

The graph-theoretic cactus is the structural prototype for most later uses of the term. A cactus graph is a connected graph in which each edge is contained in at most one cycle; equivalently, cactus graphs are exactly the connected graphs whose blocks are either single edges or cycles, so cycles can meet only at cut-vertices and the graph has the form of a tree of cycles (Zhang et al., 2023). The classical extremal bound is

$$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$

and this bound is best possible (Zhang et al., 2023).

The same paper introduces $k$-cactus graphs: a connected graph is a $k$-cactus if each edge is contained in at most $k$ cycles, where $k\ge 1$ (Zhang et al., 2023). For $2\le k\le 4$, the blocks can still be classified explicitly. A graph is a $2$-cactus iff each block is either an edge, a cycle, or a $\theta_3$-graph; a $3$-cactus iff each block is either an edge, a cycle, or a $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$0-graph with $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$1; and a $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$2-cactus iff each block is either an edge, a cycle, a $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$3-graph, or a $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$4-graph with $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$5 (Zhang et al., 2023). For 2-connected $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$6-cactus graphs on $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$7 vertices,

$$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$8

and the bound is tight whenever $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$9 (Zhang et al., 2023).

This prototype also supports distinct realizability theories. A cactus metric is a finite metric realized by an edge-weighted $k$0-cactus; cactus metrics have a unique optimal realization, and there is an $k$1 algorithm that recognizes whether a metric is a cactus metric and, if so, computes its optimal realization (Hayamizu et al., 2019). In geometric graph drawing, every cactus has a planar Lombardi drawing for its natural embedding, but there exist planar embeddings of cacti that do not have planar Lombardi drawings (Eppstein et al., 2021). These results show that the graph-theoretic cactus is not only sparse but also algorithmically and geometrically rigid.

2. Point-line configurations and cactus matroids

In the matroidal setting, the relevant objects are point-line configurations, namely simple matroids of rank at most $k$2. Their elements are points, and maximal dependent subsets of size at least $k$3 are called lines (Vandebrouck, 19 Aug 2025). A connected cactus configuration is built inductively by free gluing together lines and cycles: one starts from either a line or a cycle and repeatedly forms a free gluing with another line or cycle. A general cactus configuration is one whose connected components are connected cactus configurations (Vandebrouck, 19 Aug 2025).

The same class admits two equivalent characterizations. First, every line lies in at most one cycle. Second, if $k$4 denotes the graph whose vertices are the points of degree at least two and in which two vertices are adjacent iff the corresponding points lie on a common line, then

$k$5

(Vandebrouck, 19 Aug 2025). This is the precise sense in which cactus configurations are a geometric or matroidal analogue of cactus graphs.

For simple rank-three matroids, the constructive description can be stated more concretely. A connected cactus configuration is obtained by starting from either a line, identified in the paper with the uniform matroid $k$6, or a cycle, and then repeatedly freely gluing another line or cycle along a chosen point (Liwski et al., 9 Jun 2025). The free gluing operation $k$7 identifies a point $k$8 in $k$9 with a point $k$0 in $k$1, preserving the pre-existing line structure and introducing no new unintended dependencies (Liwski et al., 9 Jun 2025). This inductive definition is the combinatorial basis for the algebraic geometry developed later.

3. Realization spaces, matroid varieties, and circuit varieties

For a point-line configuration $k$2 on $k$3, the realization space $k$4 consists of collections of vectors $k$5 such that a subset is dependent in $k$6 iff the corresponding vectors are linearly dependent (Vandebrouck, 19 Aug 2025, Liwski et al., 9 Jun 2025). Its Zariski closure

$k$7

is the matroid variety, with defining ideal $k$8 (Vandebrouck, 19 Aug 2025). The circuit variety

$k$9

has ideal $k$0, and one always has $k$1 (Liwski et al., 9 Jun 2025).

Two further sources of equations are central. The Grassmann–Cayley ideal $k$2 is generated from circuit polynomials via concurrency relations in the Grassmann–Cayley algebra, and the lifting ideal $k$3 is obtained from liftability matrices built from 3-circuits (Vandebrouck, 19 Aug 2025, Liwski et al., 9 Jun 2025). The algebraic-geometric program of these papers is to determine when the matroid ideal can be described, up to radical, by these explicit polynomial systems.

For cactus configurations, the geometry is unusually controlled. Every cactus configuration is realizable, and its matroid variety $k$4 is irreducible (Vandebrouck, 19 Aug 2025, Liwski et al., 9 Jun 2025). The proofs proceed by showing that cactus configurations are nilpotent, then using the implication nilpotent $k$5 solvable, and solvable $k$6 irreducible (Vandebrouck, 19 Aug 2025). A sharper theorem states that if $k$7 is a cactus configuration and $k$8, the set of points of degree at least three, contains no cycle, then

$k$9

(Vandebrouck, 19 Aug 2025). The no-cycle hypothesis on $k\ge 1$0 is necessary in general: the paper gives an example of a cactus configuration with a cycle among the high-degree points where $k\ge 1$1 (Vandebrouck, 19 Aug 2025).

The circuit variety of a cactus configuration also has a controlled decomposition: $k\ge 1$2 where $k\ge 1$3 is obtained by turning the points in $k\ge 1$4 into loops (Vandebrouck, 19 Aug 2025). There are at most $k\ge 1$5 irreducible components, and the decomposition is stated up to irredundancy (Vandebrouck, 19 Aug 2025). This is one of the cleanest instances in which combinatorial cactus structure yields explicit information about realization spaces and defining equations.

4. Cactus groups and configuration spaces of points on the circle

The cactus group $k\ge 1$6 is generated by symbols

$k\ge 1$7

subject to

$k\ge 1$8

$k\ge 1$9

$2\le k\le 4$0

(Bellingeri et al., 2022). Geometrically, $2\le k\le 4$1 is represented by a cactus braid in which strands $2\le k\le 4$2 meet at one point and reverse order (Bellingeri et al., 2022). The papers emphasize that cactus groups are not braid groups: the braid relation fails in general (Bellingeri et al., 2022). The quantity $2\le k\le 4$3 is the leaf number of the generator (Bellingeri et al., 2022).

There is a natural permutation map $2\le k\le 4$4, and its kernel

$2\le k\le 4$5

is the pure cactus group (Bellingeri et al., 2022). A central structural result is that $2\le k\le 4$6 is the fundamental group of the real locus of the Deligne–Mumford compactification $2\le k\le 4$7 (Bellingeri et al., 2022). Another central tool is Mostovoy’s Gauss diagram group $2\le k\le 4$8, a right-angled Coxeter group, together with an injective group 1-cocycle

$2\le k\le 4$9

(Bellingeri et al., 2022). This yields an embedding

$2$0

and makes available the normal-form machinery of right-angled Coxeter groups (Bellingeri et al., 2022).

Several algebraic consequences follow. The word problem in $2$1 is solvable. The group $2$2 has no odd torsion, and for every $2$3, if $2$4 is large enough then $2$5 contains torsion of order $2$6. The pure cactus group $2$7 is torsion-free. The center of $2$8 is trivial for $2$9, and $\theta_3$0 for $\theta_3$1 (Bellingeri et al., 2022). The subgroup structure is also explicit: the twin group $\theta_3$2 and, more generally, all $\theta_3$3, inject into $\theta_3$4 (Bellingeri et al., 2022).

A parallel topological line of work identifies pure cactus groups with compactified configuration spaces of points on the circle. The survey (Hama et al., 11 May 2025) states the conceptual identification

$\theta_3$5

and develops low-degree models using

$\theta_3$6

and its compactification $\theta_3$7 (Hama et al., 11 May 2025). In degree three, $\theta_3$8, with $\theta_3$9, and the paper gives an explicit equivariant bijection between the universal cover of $3$0 and the Cayley complex of a cactus subgroup (Hama et al., 2024). In degree four, $3$1, and $3$2 is the connected sum of five projective planes; the quotient of the relevant Cayley complex by $3$3 is identified cell-by-cell with $3$4 (Hama et al., 11 May 2025).

5. Affine, virtual, and diagrammatic extensions

Affine cactus groups replace intervals on a line by circular intervals on $3$5. The affine cactus group $3$6 is generated by $3$7, where each generator corresponds to a circular interval $3$8, and the defining relations are the affine analogues of the ordinary cactus relations (Chemin, 27 Jan 2025). The main structural theorem identifies affine cactus groups with generalized cactus groups on a Coxeter group of type $3$9 in the paper’s notation, and there is an embedding

$$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$00

where $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$01 is an affine Gauss diagram group generated by involutions attached to circular sets (Chemin, 27 Jan 2025). From this representation one obtains linearity, solvability of the word problem, residual nilpotence of the pure affine cactus group $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$02, trivial center for $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$03 and $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$04 in the stated ranges, absence of odd-order torsion, and torsion-freeness of $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$05 (Chemin, 27 Jan 2025).

Virtual cactus groups arise from a different compactification theory. The space

$$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$06

of $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$07 distinct points on the line modulo translation admits a compactification $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$08, the cactus flower moduli space, together with a map

$$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$09

to the flower-curve compactification $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$10 (Ilin et al., 2023). The fibers of $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$11 are products of genus $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$12 Deligne–Mumford spaces (Ilin et al., 2023). On real loci, the resulting cube complexes are aspherical, and the equivariant fundamental groups are the virtual symmetric group and the virtual cactus group: $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$13 (Ilin et al., 2023). The same paper constructs a natural homomorphism

$$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$14

from the extended affine cactus group to the virtual cactus group via degeneration of a twisted real form of the Deligne–Mumford space (Ilin et al., 2023).

A diagrammatic extension is provided by cactus doodles. A cactus doodle is an immersed closed curve in $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$15 whose singularities may be multi-tuple intersection points, with all tangent lines distinct at each such point (Mostovoy et al., 2022). The equivalence relation is generated by isotopy together with $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$16-moves, which create or delete pairs of $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$17-tuple points, and $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$18-moves, which pass a $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$19-tuple point through an $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$20-tuple point (Mostovoy et al., 2022). Every cactus doodle is equivalent to the closure of some element of $$|E(G)| \le \left\lfloor \frac{3(n-1)}{2} \right\rfloor,$$21, and every cactus doodle is equivalent to its mirror image (Mostovoy et al., 2022). This places cactus groups in a role explicitly analogous to the relation between braids and knots.

6. Cactus configurations in the Cactus Framework

In computational science, “Cactus Configurations” has a different meaning. It denotes the thorn-specific configuration files, written in the Cactus Configuration Language, that describe how a component fits into the Cactus Framework (Allen et al., 2010). In Cactus terminology, components are thorns and the framework core is the flesh (Allen et al., 2010). The CCL is the declarative metadata layer through which the flesh discovers what a thorn provides, what it requires, how it is built, and when its routines run (Allen et al., 2010).

The paper identifies five thorn configuration files:

• interface.ccl
• param.ccl
• schedule.ccl
• configuration.ccl (optional)
• test.ccl (optional)

The flesh parses these files at build time, generates code for variables, parameters, and functions, checks dependency constraints, and at run time reads the user’s parameter file, activates only the required thorns, assigns parameter values, and constructs the execution schedule (Allen et al., 2010). In this sense the configuration language is operational rather than merely descriptive.

interface.ccl defines the thorn interface, inheritance, variables, and aliased functions (Allen et al., 2010). Variables are grouped into variable groups with homogeneous attributes such as data type, group type, rank, dimensions, and number of time levels. The three group types are Grid functions (GFs), Arrays, and Scalars (Allen et al., 2010). param.ccl defines runtime parameters with type, scope, allowed range, and default value; allowed types are Int, Real, Keyword, Boolean, and String, and parameters may be steerable (Allen et al., 2010). schedule.ccl specifies scheduled execution, including standard time bins such as CCTK_STARTUP, CCTK_PARAMCHECK, CCTK_INITIAL, CCTK_PRESTEP, CCTK_EVOL, CCTK_POSTSTEP, and CCTK_ANALYSIS (Allen et al., 2010).

Two notions organize component interchangeability. Interfaces describe runtime dependencies and allow multiple providers; capabilities describe build-time dependencies and allow only one provider in a given configuration (Allen et al., 2010). The canonical example is the driver interface, provided by both the unigrid driver PUGH and the adaptive mesh refinement driver Carpet (Allen et al., 2010). This is the basis of the plug-and-play design emphasized in the paper. The same paper gives large-scale examples: the Einstein Toolkit uses Cactus infrastructure and includes 135 thorns, of which 78 are needed as starting points to reconstruct the whole toolkit from dependency information (Allen et al., 2010).

The CCL is also presented as incomplete for future needs. The paper identifies missing or underdeveloped areas including support for meshless methods and unstructured meshes, multiphysics with multiple domains, constants in CCL rather than include files, enumerations and user-defined structures, natural handling of vectors and tensors and their symmetries, scientific metadata about thorn meaning, and improved syntax or standardized formats such as RDF or YAML (Allen et al., 2010). Here “cactus configurations” designates not a combinatorial object but a formal specification layer for HPC component composition.

7. Conceptual unity and divergence

Across these domains, cactus configurations are not a single invariant notion. In matroid theory they encode point-line incidence with the condition that every line lies in at most one cycle (Vandebrouck, 19 Aug 2025). In group theory they encode multi-strand or circular interval reversals and lead to pure, affine, and virtual cactus groups (Bellingeri et al., 2022, Chemin, 27 Jan 2025, Ilin et al., 2023). In the Cactus Framework they are declarative component specifications for compilation and scheduling (Allen et al., 2010).

What these meanings share is a controlled replacement of tree-like rigidity by sparse cyclic interaction. In the graph and matroid settings, this yields block decompositions, irreducibility theorems, and explicit extremal or realization results (Zhang et al., 2023, Vandebrouck, 19 Aug 2025). In the group-theoretic setting, it yields Coxeter-type embeddings, solvable word problems, and direct links to configuration spaces of points on the circle (Bellingeri et al., 2022, Hama et al., 11 May 2025). In the software setting, it yields a component architecture in which interfaces, capabilities, and schedules can be composed without hard-coded module-level dependencies (Allen et al., 2010). The recurrence of the term across such different literatures is therefore structural rather than merely terminological: each usage isolates a class that remains tractable because cycles are permitted only under explicit combinatorial control.