JOVANA
Explore Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

Simplicial Complexes, Discrete Curvature & Discrete Gauss-Bonnet

How do you give a polyhedral mesh — flat triangles glued edge to edge — a notion of curvature, when every face is perfectly flat? The whole answer hides at the vertices, and when you add it all up it lands, exactly and on the nose, on the number the smooth Gauss-Bonnet theorem already promised.

From smooth surfaces to combinatorial ones

Everything in this rung lives on a simplicial complex, so let us be precise about the object before we put geometry on it. A simplicial complex is a set of simplices — vertices (0-simplices), edges (1-simplices), triangles (2-simplices), tetrahedra (3-simplices) — glued together by two clean rules: every face of a simplex in the complex is also in the complex, and any two simplices meet in a common face or not at all. The second rule is the one beginners drop: two triangles may share a whole edge or a single vertex, but they may not cross through each other's interiors or overlap on half an edge. That rigidity is exactly what makes the complex a faithful combinatorial skeleton of a space rather than a tangle.

When such a complex is glued so that every edge is shared by exactly two triangles and the triangles around every vertex fan out into a single disk, the result is a triangulated surface — a combinatorial model of an honest topological surface. This is the discrete cousin of the smooth manifold from Vol I: it has the same global shape (a sphere, a torus, a Klein bottle) but it is built from finitely many flat pieces a computer can store exactly. We do not approximate the surface and then differentiate; we keep the combinatorics exact and ask which geometric quantities survive. The surprise of this guide is how much does.

Where did the curvature go? Hunting it down to the vertices

Here is the puzzle stated sharply. Take a sphere and triangulate it with thousands of flat triangles. Every face is flat — Gaussian curvature zero. Every edge is a fold, but a fold is developable: you can flatten the two adjacent triangles onto a table without stretching, so a creature walking across an edge feels no intrinsic distortion either. Yet the whole thing is, unmistakably, curved. So where is the curvature? The answer is that ALL of the intrinsic curvature of a triangulated surface is concentrated at the vertices, and the smooth recipe tells us how to find it: curvature is the failure of the surrounding angles to add up to a flat 2 pi.

Stand at a vertex v and look at the triangles fanning around it. Each contributes its corner angle — its interior angle at v. On a perfectly flat sheet those angles sum to exactly 2 pi: you can lay the fan flat without overlap or gap. If they sum to LESS than 2 pi there is a gap; closing it forces a cone, and the surface bulges outward like the tip of a party hat — positive curvature. If they sum to MORE than 2 pi the fan cannot lie flat — it ruffles like the edge of a kale leaf — negative curvature. The angle defect at v, which we write kappa(v), is precisely 2 pi minus the sum of the corner angles around v.

Angle defect at an interior vertex v:

    kappa(v)  =  2*pi  -  sum over triangles T touching v of  theta_T(v)

    theta_T(v) = interior angle of triangle T at the vertex v

  angles sum < 2pi   ->  kappa > 0   (cone, party-hat tip, positive curvature)
  angles sum = 2pi   ->  kappa = 0   (flat, fan lies flat, zero curvature)
  angles sum > 2pi   ->  kappa < 0   (saddle, kale-leaf ruffle, negative curvature)

For a boundary vertex use   pi - (sum of angles)   in place of  2*pi - (...).
The angle defect: 2 pi minus the angles meeting at an interior vertex. Its sign is the sign of curvature, and it is the discrete analogue of Gaussian curvature integrated and concentrated at the vertex.

Why the angle defect IS curvature — not merely an analogy

It is tempting to call the angle defect a clever proxy for curvature, but it is far more honest than that — it is the integral of curvature, made exact. Recall from the curves-and-surfaces rung that the smooth Gaussian curvature K controls precisely the angle excess of a geodesic triangle: a small geodesic triangle of interior angles summing to pi + (integral of K over the triangle). Equivalently, parallel-transporting a vector around a small loop rotates it by the enclosed integral of K — curvature is angle deficit per unit area. The angle defect at a vertex is exactly that integral, with all the curvature crushed onto a single point: kappa(v) is the total Gaussian curvature of any smooth surface you might bend into that cone.

This is why the angle defect is INTRINSIC — it depends only on the corner angles, which depend only on the edge lengths, never on how the surface sits in space. That is the discrete shadow of Gauss's Theorema Egregium: curvature is determined by the metric alone. Two meshes with the same edge lengths but folded differently in R^3 have the same defect at every vertex, just as the smooth theorem promised. Cut a cone along a slit and unroll it flat: the angles do not change, the defect does not change, and a flat sheet rolled into a cylinder genuinely has zero defect everywhere — matching the fact that a cylinder is intrinsically flat even though it looks bent.

Discrete Gauss-Bonnet: a topological invariant falls out exactly

Now add up the angle defect over every vertex of a closed triangulated surface. The smooth Gauss-Bonnet theorem says the total Gaussian curvature of a closed surface equals 2 pi times its Euler characteristic, integral of K = 2 pi chi(M). The discrete Gauss-Bonnet theorem is the same statement with the integral replaced by a finite sum of defects: the sum over all vertices of kappa(v) equals 2 pi times the Euler characteristic chi = V - E + F. And here is the beautiful part — this is not an approximation that improves with refinement. It is an EXACT identity on every triangulated surface, no matter how coarse or how skewed the triangles.

The proof is a few lines of counting, and it is worth carrying it out because it shows the theorem is forced, not lucky. Sum kappa(v) = sum over v of (2 pi - sum of angles at v). The first piece is 2 pi V. The second piece is the sum of ALL corner angles of ALL triangles, merely regrouped by vertex; regroup it instead by triangle, and since the three angles of each triangle sum to pi, you get pi F. So the total defect is 2 pi V - pi F. Now use that the surface is closed and triangulated: each face has 3 edges and each edge is shared by 2 faces, so 3F = 2E. Substitute pi F = pi F and rewrite using 2E = 3F, and the whole thing collapses to 2 pi (V - E + F) = 2 pi chi.

Plug in a concrete shape to feel it. The octahedron has 6 vertices, 12 edges, 8 faces, so chi = 6 - 12 + 8 = 2, the Euler characteristic of the sphere; the theorem predicts a total defect of 4 pi. Check it directly: each vertex is surrounded by four equilateral-ish faces of an octahedron — four right-angle corners, summing to 4 times pi/2 = 2 pi... no wait, an octahedron's faces are equilateral triangles meeting four to a vertex, so each vertex sees 4 times pi/3 = 4 pi/3, giving kappa = 2 pi - 4 pi/3 = 2 pi/3 per vertex, and 6 vertices yield 6 times 2 pi/3 = 4 pi. The torus tells the complementary story: chi = 0, so the defects must cancel to zero total — every bit of positive curvature on the outer rim is paid for by negative curvature in the inner hole, which is exactly the saddle geometry your fingers feel on the inside of a donut.

What this buys you, and the honest limits

Read the discrete theorem the way Vol I taught you to read its smooth parent: as a bridge between local geometry and global topology. The left side is local — a finite, computable number at each vertex, decided by edge lengths alone. The right side is a topological invariant, blind to the metric: chi = V - E + F is the same for a coarse octahedron (chi = 2) and a million-triangle sphere (chi = 2). So you can compute a topological invariant of a shape WITHOUT knowing its topology in advance — just total the angle defects and divide by 2 pi. A spiky sphere mesh and a smooth one both report 2; a coffee-mug mesh reports 0, the Euler characteristic of a torus, telling you it has exactly one handle.

  1. Build the mesh as a simplicial complex: a vertex list and a face list of triangles, oriented consistently so each interior edge is traversed once in each direction by its two faces.
  2. At each interior vertex v, gather the incident triangles, compute each corner angle at v from the edge lengths via the law of cosines, and form kappa(v) = 2 pi minus their sum.
  3. Sum kappa(v) over all vertices and divide by 2 pi: you have just computed chi = V - E + F from the geometry alone, an exact integer.
  4. Cross-check by directly counting V - E + F from the combinatorics; the two numbers must agree exactly. A mismatch means your mesh is not a clean closed surface (a non-manifold edge, a boundary, or a hole).

Be honest about what does NOT carry over, because the slogan 'discrete equals smooth' is only half true. The defect captures INTRINSIC curvature flawlessly, but discrete EXTRINSIC quantities are messier: there is no single canonical discrete mean curvature — the cotangent-Laplacian definition, the volume-gradient definition, and the normal-cycle definition all disagree away from the smooth limit, and choosing among them is a real modeling decision you will meet in Guide 3. Convergence also has teeth: the angle defect converges to integrated curvature only for reasonable meshes, and a pathological 'Schwarz lantern' triangulation of a cylinder can make discrete area diverge even as the mesh refines. The exact theorems are the gift; pointwise convergence is a separate, conditional promise. Keep the two ideas apart and you will not be fooled.