Crochet Algorithm: Computational Textile Structures

• Crochet algorithm is a computational framework that uses local stitch operations and explicit combinatorial rules to build globally correct textile structures.
• It encompasses executable DSLs like CrochetPARADE that validate, compile, and render complex stitch programs, ensuring geometric and topological correctness.
• The approach extends to holomorphic dynamics where the algorithm decomposes postcritically finite rational maps into crochet and Sierpiński components, highlighting its cross-disciplinary relevance.

Searching arXiv for the cited crochet-related papers and the holomorphic-dynamics usage of “crochet” to ground the article in current literature. arxiv_search(query="CrochetBench Crochet Algorithm crocheting mathematics AmiGo Bour minimal surfaces postcritically finite rational maps crochet", max_results=10, sort_by="relevance") “Crochet algorithm” is not a single standardized term in the arXiv literature. It designates several algorithmic formalisms that use crochet as a computational medium: executable stitch programs expressed in a domain-specific language, gauge-calibrated local update rules for realizing prescribed curvature, geometry-to-pattern pipelines for surfaces and meshes, and, in holomorphic dynamics, a canonical decomposition algorithm for postcritically finite rational maps into crochet and Sierpiński components (Li et al., 12 Nov 2025, Kekkonen, 14 Aug 2025, Dudko et al., 2022). Across these settings, the common structure is local state update under explicit combinatorial or geometric constraints.

1. Terminological scope and conceptual structure

In computational crochet, the basic algorithmic question is how local stitch operations induce a global object. The literature instantiates this question in several technically distinct ways. In multimodal reasoning, crochet is treated as an executable procedural domain in which natural language or images are translated into a strict intermediate representation with labels, references, and repetition constructs (Li et al., 12 Nov 2025). In differential geometry, crochet is modeled as a discrete dynamical system in which each round updates circumference according to curvature-dependent rules, with gauge parameters linking stitch width and row height to the target metric (Kekkonen, 14 Aug 2025). In computational fabrication, the goal is pattern synthesis from continuous or discrete geometry, including surfaces of revolution, triangulated hyperbolic surfaces, minimal surfaces, and closed triangle meshes (Martinez et al., 2023, Trnkova, 2019, Kekkonen, 2023, Edelstein et al., 2022).

A separate usage occurs in complex dynamics. There, a crochet map is a Böttcher expanding map for which any two Fatou components can be connected by a countable chain of Fatou components with common boundary points, and the Crochet Algorithm computes a canonical invariant multicurve that decomposes the system into crochet and Sierpiński blocks (Dudko et al., 2022). The semantic gap between this dynamical meaning and the textile one is substantial, but both rely on invariant local adjacency rules and quotient constructions.

A common misconception is that crochet algorithms are primarily textual recipe generation. The recent benchmark literature explicitly rejects this reduction: the operative target is executable correctness rather than stylistic similarity, and compilation validity becomes the decisive criterion when crochet is formalized as a stateful program (Li et al., 12 Nov 2025).

2. Executable crochet procedures and the CrochetPARADE formalism

CrochetBench frames crochet algorithms as compilable, stateful programs rather than free-form text. Its intermediate representation is CrochetPARADE, a domain-specific language designed to parse and validate textual stitch procedures, enforce numerical and topological constraints, render the resulting structure in 2D/3D, and support debugging such as over-/under-stretched stitches and impossible attachments (Li et al., 12 Nov 2025).

The core primitives observed in CrochetBench include stitches such as ch, sc, hdc, dc, tr, ss, and modifiers such as scbl; operations such as sk and turn; stitch macros such as sc2inc; labels A, B, … used to mark anchor positions; bracketed repeats [ … ]*n; numeric prefixes such as 24sc; the line or round separator ¶; and the attach or reference operator @ (Li et al., 12 Nov 2025). Patterns are represented as sequences of DSL lines, each a comma-separated list of stitch calls, label definitions, skips, repeats, and joins. The approximate grammar given in the paper has the form

$$\text{Program} := \text{Line}^*, \qquad \text{Line} := \text{"¶"}\ \text{Statement},$$

with stitch calls, repeats, joins, skips, and turns as the principal production types (Li et al., 12 Nov 2025).

The validator enforces several distinct constraint classes. Syntax structure errors include unbalanced brackets or parentheses and malformed multiple references. Stitch definition errors arise from tokens outside the DSL dictionary. Labeling and reference errors include undefined labels and non-adjacent label reuse. Structural and formatting issues include the rule that turning can happen only at the end of a row and naming conflicts with stitch tokens. Runtime parser failures indicate deeper structural inconsistencies (Li et al., 12 Nov 2025). Compilation proceeds through parsing to an AST, static checks, structural analysis of stitch counts and geometrical viability, and rendering to SVG or GLTF.

This executable view changes what counts as correctness. A program is correct only if it compiles without syntax, label, or structural errors and produces a renderable structure. This shifts crochet from descriptive language modeling to symbolic program synthesis with state, scope, and topology. The benchmark’s examples make the distinction concrete: a superficially plausible line can fail because of an unbalanced bracket or an undefined stitch such as trr, even when the generated text resembles a pattern (Li et al., 12 Nov 2025).

3. Curvature-driven crochet as a discrete geometric dynamical system

The geometric literature models crochet worked in rounds with one live loop as a discrete dynamical system. In “Crocheting Mathematics,” the circumference of a geodesic circle in a target geometry determines the stitch count on a round through the gauge relation

$N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$

where $w$ is stitch width, $h$ is row height, and $\gamma$ is stitch density per unit length (Kekkonen, 14 Aug 2025). The curvature enters through the circumference law. For Euclidean geometry,

$C(r)=2\pi r,$

so the per-round increase is approximately constant and, when $h \approx w$, yields the classical flat-circle rule of about six increases per round. For hyperbolic geometry,

$C(r)=\frac{2\pi}{\kappa}\sinh(\kappa r),$

which leads after a few rounds to the multiplicative rule $N_{t+1}\approx N_t e^{\kappa h}$ and the practical instruction “increase one stitch every $m$ stitches” with $N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$0. For the sphere,

$N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$1

so increases grow from the pole to the equator and then turn into symmetric decreases (Kekkonen, 14 Aug 2025).

The Lobachevskian-surface literature makes the same point with a different discretization. There the hyperbolic circumference is taken as $N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$2, with stitch count $N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$3, exact round-to-round growth factor

$N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$4

and large-$N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$5 approximation $N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$6 (Reyes et al., 2024). The paper also gives block patterns such as 3332, interpretable as controlled mixtures of doubles, triples, and quadruples into one base stitch. This yields an exact algebraic realization of the target increase $N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$7 via counts of singles, doubles, triples, and quadruples distributed uniformly around the round (Reyes et al., 2024).

A related triangulated construction appears in “Hyperbolic Flowers,” where curvature is encoded by a $N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$8 tiling with $N_t \approx \gamma\, C(r_t), \qquad r_t = t h,$9. For triangular tilings this means $w$0, and the crocheting rule becomes “add triangles on boundary edges; increase at vertices to reach valence $w$1,” or, in gauge form, place increases approximately every

$w$2

stitches, where $w$3 and $w$4 (Trnkova, 2019).

These formulations establish crochet as a metric discretization device rather than merely a craft notation. The local update rule carries curvature, and gauge calibration controls the realized intrinsic geometry.

4. Pattern synthesis from analytic and discrete geometry

A second major line of work generates crochet patterns directly from geometric input. For surfaces of revolution, “Automating Crochet Patterns for Surfaces of Revolution” uses a generating curve $w$5 on $w$6, places rows by arclength

$w$7

solves $w$8, and converts circumference

$w$9

to stitch counts

$h$0

(Martinez et al., 2023). Increases and decreases are then distributed by a remainder method with optimal cyclic shift. Candidate placements are evaluated with circular-distance statistics $h$1 and $h$2, which maximize separation from the previous round’s increase or decrease positions and reduce stacking artifacts (Martinez et al., 2023).

AmiGo generalizes algorithmic generation to closed manifold triangle meshes. Given a closed manifold triangle mesh $h$3, a seed vertex $h$4, and a stitch width $h$5, the algorithm builds a Crochet Graph $h$6 whose rows are geodesic isolines of the distance function $h$7 and whose column edges are couplings between consecutive rows (Edelstein et al., 2022). The function $h$8 ordering vertices within rows is computed on a cut mesh by minimizing

$h$9

subject to boundary conditions; consecutive rows are coupled by Dynamic Time Warping minimizing

$\gamma$0

and the coupled-row property guarantees valid instructions using only sc, inc(x), and dec(x) (Edelstein et al., 2022). Branching is handled by slicing at saddle isolines of the geodesic distance, topologically sorting the resulting segments, and attaching them by join-as-you-go without sewing (Edelstein et al., 2022).

For Bour’s $\gamma$1 minimal surfaces, the geometric driver is the intrinsic metric

$\gamma$2

from which one obtains

$\gamma$3

and

$\gamma$4

The algorithm advances one round by intrinsic stitch height $\gamma$5, updates $\gamma$6, and sets the round stitch count to $\gamma$7 (Kekkonen, 2023). Enneper’s, Richmond’s, and $\gamma$8 are treated as explicit special cases, including handling of self-intersection structure.

This family of pipelines shows that crochet pattern generation can be driven by arclength inversion, geodesic parameterization, coupling on cyclic sequences, or intrinsic rotational metrics, depending on the source geometry.

5. Benchmarking procedural competence and executable correctness

CrochetBench is currently the most explicit evaluation framework for multimodal crochet algorithms. It contains Task A (Stitch Recognition), Task B (Instruction Selection), Task C (Instruction Generation), and Task D (Instruction-to-DSL Translation), with datasets of 6,009, 6,003, 6,009, and 119 or 100 items depending on the Task D variant (Li et al., 12 Nov 2025). The corpus comprises 6,085 patterns across 55 project categories; 98.77% include product images; instruction length ranges from 20 to $\gamma$9 characters, with mean 3,216 and median 2,453 (Li et al., 12 Nov 2025).

The benchmark explicitly separates surface similarity from executability. Task C uses BLEU, ROUGE-L, and ChrF, whereas Task D uses validator-based metrics: Compilation Success Rate,

$C(r)=2\pi r,$0

and Partial Executable Rate,

$C(r)=2\pi r,$1

The stated key insight is that surface-level generation quality is not predictive of executable correctness (Li et al., 12 Nov 2025).

The reported results quantify that gap. For Task A, the best F1 is Claude Sonnet 4 at approximately 0.6094, while the open-source best is DeepSeek-VL at approximately 0.6060. For Task B, GPT-4o reaches approximately 0.5811 accuracy, while Qwen2-VL is the open-source best at approximately 0.4196. For Task C, Gemini 2.5 Flash-Lite has the best BLEU, ROUGE-L, and ChrF. In Task D$C(r)=2\pi r,$2, the best CSR is Claude Sonnet 4 at approximately 52.1%, and in Task D$C(r)=2\pi r,$3 the best CSR is Qwen2-VL at approximately 21.0%; DeepSeek-VL has CSR approximately 8.1% with PER approximately 37.49%, while closed-source models lag in project-level CSR at 4–5% and PER at approximately 2.76–8.16% (Li et al., 12 Nov 2025).

The dominant failure modes are undefined stitches, bracket or syntax errors, and label or reference errors. The authors therefore recommend hybrid neuro-symbolic parsing, constraint-driven decoding, 3D geometry priors, multimodal verification, and state-tracking scaffolds (Li et al., 12 Nov 2025). This suggests that practical crochet algorithms for multimodal systems must integrate compiler feedback rather than rely on unconstrained sequence generation.

6. The Crochet Algorithm in holomorphic dynamics

In “A canonical decomposition of postcritically finite rational maps and their maximal expanding quotients,” the Crochet Algorithm belongs to an entirely different domain. Let $C(r)=2\pi r,$4 be postcritically finite, or equivalently a Thurston map $C(r)=2\pi r,$5 in the topological model. Under the standing assumption of non-empty Fatou set, the paper defines crochet maps through several equivalent criteria, including the existence of a connected forward-invariant zero-entropy graph $C(r)=2\pi r,$6 with $C(r)=2\pi r,$7, the fact that $C(r)=2\pi r,$8 is a singleton, and the property that any two points in $C(r)=2\pi r,$9 can be connected by a path whose intersection with the Julia set is countable (Dudko et al., 2022).

The algorithmic objective is to compute a unique canonical invariant multicurve $h \approx w$0 such that the small maps in the decomposition along $h \approx w$1 are either Sierpiński maps or crochet maps. The construction starts by computing maximal clusters of touching Fatou components and their boundary multicurve. It then decomposes the map with respect to that multicurve, iterates the procedure on the resulting small maps, and finally glues small crochet maps that correspond to the same point in the quotient by eliminating primitive crochet unicycles (Dudko et al., 2022).

The stated outputs are the canonical invariant multicurve $h \approx w$2, the decomposition into small spheres, and the quotient cactoid $h \approx w$3 with quotient map $h \approx w$4 (Dudko et al., 2022). Theorem 1.6 gives the defining quotient behavior: small Julia sets of Sierpiński small maps project onto spheres, crochet small Julia sets project to points, and distinct crochet small Julia sets project to distinct points. Theorem 1.1 then states maximality: the induced quotient is maximal among totally topologically expanding quotients (Dudko et al., 2022).

This usage makes clear that “crochet algorithm” is not intrinsically textile. In this context, crochet names a topological connectivity class of dynamical pieces and an effective symbolic decomposition procedure operating on a sphere biset. The overlap with textile crochet is etymological rather than substantive, but both usages are organized by adjacency, iteration, and collapse or expansion under local rules.