Gaussian Curvature

Why Gaussian curvature matters

Gaussian curvature is the simplest non-trivial invariant of a surface — the first quantity that distinguishes a sphere from a saddle without reference to how either sits inside 3-space. The Theorema Egregium, by showing K is intrinsic, founded a whole subject: Riemannian geometry, where one studies "curved spaces" without needing an ambient flat space to host them. From there, general relativity, gauge theory, optimal transport, and machine learning on manifolds all became possible.

• Foundation of intrinsic geometry. Theorema Egregium gives the framework for studying surfaces (and higher-dimensional manifolds) without embedding. Riemann's habilitation lecture (1854) generalized to n dimensions, leading to Riemannian geometry, the Riemann curvature tensor, and ultimately general relativity, where spacetime is a 4-manifold whose intrinsic curvature is gravity.
• General relativity. Einstein's field equations relate intrinsic curvature (Ricci tensor, scalar curvature) to matter-energy content. The 2-dimensional case — Gaussian curvature — captures the same idea more visibly: matter curves spacetime, and the curvature is intrinsic, observable to anyone living on the manifold.
• Cartography and mapmaking. Maps of the Earth (K = 1/R² ≈ 2.5 × 10⁻¹⁴ m⁻²) onto a plane (K = 0) must distort either angles, areas, or distances. Different projections compromise differently: Mercator preserves angles (conformal), Lambert preserves areas, gnomonic preserves geodesics. No map preserves everything — that is Theorema Egregium in everyday life.
• Geodesic deviation and triangulation. A geodesic triangle's angle sum exceeds π by ∫∫ K dA over the triangle (positive K), or falls short by the same amount (negative K). On a plane the sum is exactly π. Gauss reportedly checked this on a Hannover surveying triangle to test whether real space had non-zero curvature — a precocious empirical test of geometry.
• Mesh processing and CAD. Discrete Gaussian curvature on a polyhedral mesh is encoded by angle defects at vertices (2π minus the sum of incident triangle angles at v). Algorithms for curvature estimation, smoothing, remeshing, and surface reconstruction all use this discrete K as a robust geometric signal.
• Optimal transport and information geometry. The space of probability distributions over a fixed sample space carries the Fisher information metric, a Riemannian structure whose curvature governs convergence of optimization algorithms (natural gradient descent) and statistical-inference rates (Cramér-Rao bounds).
• Hyperbolic networks and embeddings. Hyperbolic surfaces have K < 0 and grow exponentially. Embedding hierarchical or scale-free networks into hyperbolic balls (Poincaré embeddings, Krioukov-Boguñá hyperbolic geometry of complex networks) captures tree-like structure with low distortion that Euclidean (K = 0) embeddings cannot.
• Soap films and minimal surfaces. A minimal surface has mean curvature H = (κ₁ + κ₂)/2 = 0 — but not zero K. Soap films are minimal; their K varies from point to point, going negative wherever the film curls in two opposing directions (most of the time).

Definitions: principal curvatures and K

Let M be a smooth 2-surface in ℝ³ with unit normal vector field n. At a point p ∈ M, the normal curvature κ_v in a tangent direction v ∈ T_p M is the curvature of the planar curve obtained by intersecting M with the plane spanned by v and n(p) (the normal section). Computed via second fundamental form II:

κ_v = II(v, v) / I(v, v)    where I is the first fundamental form, II the second.

As v rotates over unit tangent directions at p, κ_v reaches a maximum κ₁ and a minimum κ₂ — the principal curvatures. The corresponding orthogonal directions are the principal directions.

Gaussian curvature: K = κ₁ · κ₂.
Mean curvature: H = (κ₁ + κ₂)/2.

Examples.

• Sphere of radius r: κ₁ = κ₂ = 1/r at every point. K = 1/r², H = 1/r.
• Plane: κ₁ = κ₂ = 0. K = 0, H = 0.
• Cylinder of radius r: κ₁ = 1/r (around the axis), κ₂ = 0 (along the axis). K = 0, H = 1/(2r).
• Saddle z = xy at origin: κ₁ = 1, κ₂ = −1 (in suitable units). K = −1, H = 0.
• Pseudosphere (tractrix surface of revolution): K = −1 everywhere. Models the hyperbolic plane locally.

Theorema Egregium — K is intrinsic

Gauss's 1827 paper "Disquisitiones generales circa superficies curvas" contains the result he himself called remarkable:

Theorem (Theorema Egregium). The Gaussian curvature K of a smooth surface M ⊂ ℝ³ depends only on the first fundamental form (the induced metric) on M. Equivalently, K is preserved by any local isometry between surfaces.

Why "remarkable." The definition K = κ₁κ₂ uses the second fundamental form, which encodes how M curls in 3-space — extrinsic information about the embedding. The theorem says that despite this extrinsic-looking definition, K is computable using only intrinsic data: distances and angles measured within M.

Costed claim. Theorema Egregium: K is computed from the first fundamental form only. Concretely, for a metric ds² = E du² + 2F du dv + G dv², Gauss gave a formula for K in terms of E, F, G and their first and second partial derivatives. In the conformal case ds² = e^{2λ}(du² + dv²), the formula is the elegant K = −e^{−2λ}Δλ, where Δ is the planar Laplacian. The second fundamental form (encoding embedding-specific curl) is redundant for computing K.

Consequence. Rolling a flat sheet of paper into a cylinder preserves K — both surfaces have K = 0. A cylinder is intrinsically flat; lengths and angles on the paper are unchanged. By contrast, a sphere of radius r has K = 1/r² ≠ 0, so no isometric flattening exists: every map of the Earth must distort.

Examples and signs

Surface	K at a point	Sign / type	Geometric picture	Total ∫∫ K dA (closed)
Plane ℝ²	0	flat	No bending in any direction	0 (non-compact)
Cylinder (∞)	0	flat (intrinsically)	Curls in one direction, flat in the other	0 (non-compact)
Sphere of radius r	1/r²	positive constant	Bends the same way in all directions	4π (χ = 2)
Torus (standard ℝ³)	varies: +out, −inside hole	mixed	Donut: positive on outside, negative inside hole	0 (χ = 0; cancellation)
Saddle z = xy	−1 at origin	negative	Bends up one way, down the perpendicular	non-compact
Pseudosphere	−1	negative constant	Hyperbolic plane locally; horns at infinity	2π (Beltrami)
Genus-2 hyperbolic surface	−1 (intrinsic)	negative constant	Two-holed donut with negative-curved metric	−4π (χ = −2)

Gaussian vs mean vs principal curvatures

Curvature	Formula	Intrinsic?	Sign meaning	Where it appears
Principal κ₁, κ₂	max/min of κ_v over tangent directions	No — extrinsic, depends on embedding	Individual bending rates	Local geometry of normal sections
Mean H = (κ₁ + κ₂)/2	average of principal curvatures	No — extrinsic	H = 0 ⇔ minimal surface (soap film)	Minimal-surface theory, capillary action
Gaussian K = κ₁κ₂	product of principal curvatures	Yes — Theorema Egregium	Local shape type (sphere/plane/saddle)	Riemannian geometry, Gauss-Bonnet
Scalar curvature R	R = 2K for surfaces (in higher dim, trace of Ricci)	Yes — intrinsic	Volume comparison with Euclidean	Einstein equations, scalar-curvature flows
Sectional curvature K(π)	K of the surface spanned by 2-plane π in higher dim	Yes — intrinsic	Plane-by-plane curvature	Riemannian geometry, comparison theorems
Ricci curvature	average of sectional curvatures in n dim	Yes — intrinsic	Volume contraction along geodesics	Einstein equations, Ricci flow (Perelman)

Total curvature and Gauss-Bonnet

For a closed oriented 2-surface M without boundary, the Gauss-Bonnet theorem says

∫∫_M K dA  =  2π · χ(M),

where χ(M) is the Euler characteristic. The total curvature is a topological invariant, fixed by χ regardless of how the surface is bent. Examples.

• Sphere (χ = 2): ∫∫_{S²} K dA = 4π. Consistent with K = 1/r² on a sphere of radius r: (1/r²) · 4πr² = 4π.
• Torus (χ = 0): ∫∫_T K dA = 0. The donut embedded in ℝ³ has positive K on the outside and negative K on the inside of the hole; the contributions cancel exactly.
• Closed genus-g surface (χ = 2 − 2g): total K = 2π(2 − 2g). For g ≥ 2 this is negative, and the surface admits a hyperbolic metric of constant K = −1 by uniformization.
• Klein bottle (non-orientable, χ = 0): if embedded in ℝ⁴ smoothly, ∫∫ K dA = 0. Cannot be embedded in ℝ³ without self-intersection.

The boundary version: for a compact 2-manifold M with smooth boundary ∂M, ∫∫_M K dA + ∮_{∂M} kg ds + Σ (exterior angles) = 2π χ(M), where kg is the geodesic curvature of the boundary and exterior angles are accumulated at corner points. This is the form used in mesh processing and in physical problems with boundary.

Geodesic triangles and curvature

On a surface, a geodesic triangle is a region bounded by three geodesic arcs. Its angle sum α + β + γ deviates from π by exactly the total Gaussian curvature inside:

α + β + γ  =  π + ∫∫_T K dA.

On a sphere (K = 1/r²), the angle sum exceeds π by Area/r². A large triangle on the Earth can have noticeably non-π angle sums; in surveying, this is the spherical excess.

On a hyperbolic surface (K = −1 in normalized units), the angle sum is less than π by exactly the area: α + β + γ = π − Area. Hyperbolic triangle area is therefore bounded by π. An ideal triangle has all three vertices at infinity and angles all zero, with area exactly π.

On a flat plane (K = 0), the angle sum is exactly π — Euclidean geometry's familiar statement.

Computing K from the first fundamental form

Let (u, v) be local coordinates on M and ds² = E du² + 2F du dv + G dv² the first fundamental form, with E, F, G smooth functions of u, v.

Brioschi formula (general):

K = (1 / (EG - F²)) · [   det(matrix involving second partials of E, F, G)
                          − det(matrix involving first partials of E, F, G)  ]

(The full formula is several lines; modern references use the Brioschi or Liouville form depending on the metric structure.)

Special case — orthogonal coordinates (F = 0):

K = − (1 / (2 √(EG))) · ( ∂_u ( G_u / √(EG) ) + ∂_v ( E_v / √(EG) ) ).

Special case — conformal coordinates (E = G = e^{2λ}, F = 0):

K = − e^{−2λ} · Δλ,    where Δ = ∂²_u + ∂²_v.

The conformal form is the cleanest. Every surface has local conformal coordinates (isothermal coordinates), and so this gives a uniform way to compute K. For the standard sphere of radius r in stereographic coordinates, λ = log(2 r / (1 + u² + v²)) and the formula recovers K = 1/r² (a useful exercise).

Common pitfalls

• "K is always non-negative." No. Saddle surfaces, the inside of a torus's hole, the pseudosphere, and any hyperbolic surface have K < 0. The pringle chip and the helicoid are everyday negative-curvature examples.
• "K depends on the embedding in 3-space." No — by Theorema Egregium, K depends only on the metric (first fundamental form). Two isometric surfaces have the same K everywhere. A piece of paper rolled into a cylinder has K = 0 = K of the flat paper.
• "K = 0 means flat plane." A cylinder also has K = 0 — intrinsically flat. The geometry on the cylinder is exactly the geometry on the plane (just with a different global topology). Same goes for cones (except at the apex). Theorema Egregium says intrinsic flatness, not literally-planar, is what K = 0 detects.
• "Mean curvature is intrinsic." No — H = (κ₁ + κ₂)/2 is extrinsic. The cylinder has H = 1/(2r) ≠ 0, but the flat paper has H = 0. Bending the paper into a cylinder preserves K (both 0) but changes H. Only K is intrinsic; H captures embedding-specific curl.
• "K determines the surface." No — many non-isometric surfaces have the same K. All cylinders have K = 0; all flat-paper rollings into cones have K = 0 everywhere except possibly at the apex. K is one invariant; full intrinsic geometry requires the metric tensor (not just its scalar curvature K).
• "K can be measured by a 'flat-lander' in the surface." Yes, and this is the operational content of Theorema Egregium. A 2D inhabitant of the surface can measure K by triangulating: compute the angle excess of a geodesic triangle and divide by area. Or by measuring the circumference of small geodesic circles: a circle of radius r has circumference 2πr − (π K r³)/3 + O(r⁵). The deviation from 2πr at order r³ gives K.

History

Carl Friedrich Gauss developed the theory of surface curvature in his 1827 paper "Disquisitiones generales circa superficies curvas." Motivated by geodesy and the survey of Hanover, he introduced the first and second fundamental forms, the principal curvatures, the Gauss map (sending a point on the surface to its unit normal direction on the unit sphere), and the now-classical formula K = κ₁κ₂. The Theorema Egregium ("Theorem A") was the central result, proved by relating the Gauss map's Jacobian to the metric structure.

Gauss's framework remained essentially the model for differential geometry until Riemann's 1854 habilitation lecture "Über die Hypothesen, welche der Geometrie zu Grunde liegen" (published posthumously in 1868), which generalized to n-dimensional manifolds and introduced the Riemann curvature tensor — a multilinear generalization of K. The Riemann tensor in 2D reduces to a single scalar component, K itself; in higher dimensions it has more independent components (n²(n²−1)/12 in dimension n).

The Theorema Egregium opened the door to general relativity: Einstein's field equations relate the intrinsic curvature of spacetime to its matter-energy content, without reference to any external embedding. The "spacetime is a 4-manifold" picture is conceivable precisely because Gauss showed curvature is intrinsic. Modern uses span machine learning on manifolds (Poincaré embeddings, hyperbolic graph neural networks), the Ricci flow of Hamilton and Perelman, the calculus of variations on Riemannian manifolds, and the geometric analysis of singularities in nonlinear PDE.