Unverified 2026

Holonomy-preserving quad-mesh augmentation

Implementation & benchmark of arXiv:2608.01444 — Which Holonomy Signatures Are Realizable? A Complete Answer for Closed Surfaces

Usefulness5/10

Difficulty5/10

Novelty7/10

Architecture Generalization Accuracy Graph nn Data augmentation Geometry Combinatorics Topology Geometry Graph theory

Source paper: Which Holonomy Signatures Are Realizable? A Complete Answer for Closed Surfaces arXiv:2608.01444 ⓘ · analyzed Aug 31, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Use the paper's explicit square insertion surgery to generate new quad-mesh examples with altered local valence patterns but unchanged genus, unchanged non-target vertices, and unchanged rotational-holonomy subgroup. Train a mesh GNN with consistency loss or label-preserving augmentation across the original and surgically modified meshes, forcing predictions to depend on global structure rather than accidental local tessellation.

Formulas

$$\alpha^{\prime}(x)=y_{2},\qquad \alpha^{\prime}(\alpha(x))=y_{0},\qquad \alpha^{\prime}(y_{1})=y_{3},$$

$$\operatorname{valence}(v)=w\ \longmapsto\ \operatorname{valence}(v_1)=\gamma+2,\qquad \operatorname{valence}(v_2)=w-\gamma+2,$$

$$\operatorname{im}\rho^{\prime}=\operatorname{im}\rho,\qquad F\geq 2g-2+n,$$

$$4F=\sum_i(m_i+4)+4r=4(2g-2)+4n+4r.$$

Mathematical statement

A quad mesh is represented by oriented edge darts with an involution $\alpha$ pairing the two sides of each edge. In Lemma 4.1, if $x$ and $\alpha(x)$ are the darts of a chosen edge and $y_0,y_1,y_2,y_3$ are the four darts of a newly inserted square, the even split surgery replaces the pairing by $\alpha'(x)=y_2$, $\alpha'(\alpha(x))=y_0$, $\alpha'(y_1)=y_3$, while leaving all other pairings unchanged. If the original vertex has valence $w$, and the two ends of the slit have $\gamma$ intervening corners on one side, the vertex is replaced by vertices of valences $\gamma+2$ and $w-\gamma+2$. The surgery adds one square, preserves genus, leaves every other vertex untouched, and preserves $\operatorname{im}\rho\leq\mathbb Z_4$, where $\rho:H_1(M\setminus C;\mathbb Z)\to\mathbb Z_4$ is the rotational holonomy and $m_i\equiv\rho(\gamma_i)\pmod 4$ encodes cone order. For a target signature with $n$ prescribed nonregular vertices and genus $g$, every mesh needs $F\geq 2g-2+n$ faces because $4F=\sum_i(m_i+4)+4r=4(2g-2)+4n+4r$, with $r$ regular valence-4 vertices; this provides a minimality check for generated examples.

Implementation notes

Integrate this at the mesh preprocessing and training-augmentation stage of a graph neural network or mesh transformer. Represent each quad mesh with a dart table: each dart stores its face, cyclic next-dart pointer, endpoint vertex, and paired dart $\alpha(d)$. Also store vertex corner cycles so that the two sides of a selected edge can be ordered and the gap $\gamma$ counted.

Pseudocode:

augment(mesh, edge=(x, alpha[x]), side_gap_gamma):
    create new face with darts y0,y1,y2,y3 in cyclic order
    old_partner = alpha[x]
    alpha_new[x] = y2
    alpha_new[old_partner] = y0
    alpha_new[y1] = y3
    copy all other alpha pairings
    rebuild vertex cycles and face adjacency
    assert number_faces(new) == number_faces(mesh) + 1
    assert genus(new) == genus(mesh)
    assert valences are unchanged except target split
    assert holonomy_subgroup(new) == holonomy_subgroup(mesh)
    return new

Apply one to three surgeries per training sample, choosing edges incident to a nonregular vertex and sampling $\gamma$ uniformly among valid gaps. Compute labels such as genus, cone-order multiset, or graph property on both versions and impose $L_{cons}=\|f(G)-f(G')\|^2$ when the task label is surgery-invariant. The implementation derives the pairing and valence identities directly from the formulas; genus can be checked empirically from $V-E+F=2-2g$, while holonomy can initially be approximated by storing the mod-4 turn sum around a cycle basis and verifying that its generated subgroup is unchanged. As a cheap first experiment, use a small message-passing GNN on synthetic genus-0 and genus-1 quad meshes, train on unsurgerized meshes, and test on meshes augmented by held-out local surgeries. Compare ordinary augmentation against holonomy-preserving augmentation at equal sample count. Success is higher accuracy and lower performance drop under changed local valence distributions, with all generated samples passing the Euler-characteristic, face-count, and holonomy checks.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.