Trivial Connections on Discrete Surfaces

Crane, Desbrun, Schröder — Computer Graphics Forum 29(5), 1525–1533 (SGP 2010)
Explained assuming: you know tangent vector fields on meshes, singularities, and Euler characteristic at a working level. No deep learning — this is discrete differential geometry. The key new idea (parallel transport / connections as a first-class discrete object) is explained from scratch below.

One-sentence version: Instead of smoothing a field and hoping singularities land somewhere reasonable, prescribe exactly where you want them (and with what strength) up front, and solve one sparse linear system for the smoothest field on the mesh that has singularities there and nowhere else — an exact, non-iterative alternative to both greedy local smoothing and interactive manual placement.

The core idea: treat "how a vector rotates as it moves" as the variable, not the vector itself

Take a tangent vector and slide it across the surface from one triangle to an adjacent one. On a flat plane, "the same vector" just means the same direction. On a curved surface, there's an unambiguous notion of "as parallel as possible" too (the Levi-Civita connection, the standard one induced by the surface's own curvature) — but it isn't literally "no rotation," because curvature itself forces a small rotation as you go. A discrete connection is exactly this: an assigned rotation angle for each mesh edge, describing how a vector is carried from one triangle's tangent plane to its neighbor's.

The key move this paper makes: rather than treating the vector field as the primary unknown to solve for, as most field-design methods do, treat the connection itself — the per-edge rotation angles — as the unknown, and solve for it directly. Once you have a good connection, recovering an actual smooth field from it (by picking a starting direction somewhere and parallel- transporting it everywhere) is a trivial, cheap final step — hence "trivial connections."

Why walking a loop reveals a singularity

Take any small loop on the surface — say, the triangles surrounding one vertex — and parallel- transport a vector all the way around it back to where it started. On a flat region, it comes back pointing the same way it left. Where it doesn't, that mismatch angle (the holonomy of the loop) is exactly what a singularity is: a point where the field genuinely cannot be smooth, and the size of the mismatch (quantized in units of 360°/N for an N-RoSy field) is the singularity's index.

Hand-drawn illustration of an arrow being carried around a dashed loop on a hill, gradually rotating, ending up pointing a different direction than it started, labeled left-over twist.
Carry a vector all the way around a loop by parallel transport. On a flat region it comes back unchanged; where it doesn't, that leftover rotation — the holonomy — is exactly what a singularity is, and its size is the singularity's index.

Crucially, this reframes "where are the singularities" as a direct, computable property of whatever connection you choose — not something that only emerges after the fact from smoothing. That's what makes the paper's core promise possible: choose the holonomy pattern you want (i.e. choose exactly which loops should have a mismatch, and how much), and solve for the connection that's globally smoothest subject to producing exactly that pattern.

One linear system, hard constraints, a global optimality guarantee

Because a connection is just one number per edge, and both "how smooth is it" (deviation from Levi-Civita) and "what's its holonomy around each loop" (the singularity constraint) are linear in those numbers, the whole problem reduces to a single sparse linear system — no iteration, no local relaxation, no risk of getting stuck in a bad local optimum the way iterative smoothing can. The paper proves the result is globally optimal in a precise sense: among every connection achieving the prescribed set of singularities, it's the one closest to Levi-Civita (i.e. the smoothest possible one consistent with your constraints).

Poincaré-Hopf still applies. You can't prescribe an arbitrary singularity pattern — the indices you choose must sum to a value fixed by the surface's Euler characteristic (the same topological constraint that makes some singularities unavoidable on any curved surface — the same reason you can't comb a hairy ball perfectly flat). But within that one global constraint, placement and per-singularity index are entirely up to you.
Prescribed singularities (location, index) pairs, summing to the surface's Euler characteristic One sparse linear solve per-edge rotation angles, smoothest connection matching those constraints Trivial recovery of the field pick a start direction, parallel-transport it everywhere using the solved connection
Blue = your input, a chosen singularity configuration. Orange = the paper's actual contribution — a linear solve for the connection itself, not the field. Green = the "trivial" last step this method is named for, cheap once the connection is known.

Why this composes well with other design decisions

Because this paper's input is simply a list of desired (location, index) pairs, it composes naturally with any upstream decision about where those should go — including symmetry constraints (make the list symmetric under a chosen mirror map to get a bilaterally symmetric field) or feature-driven placement (put singularities at natural anatomical/mechanical joints rather than mid-panel). The linear solve itself doesn't care how the list was chosen.

Source: Crane, K., Desbrun, M., Schröder, P., "Trivial Connections on Discrete Surfaces," Computer Graphics Forum 29(5), 1525–1533 (Symposium on Geometry Processing 2010).