Gauss's Theorema Egregium

A sphere cannot be unrolled onto a plane without tears or folds. A cylinder can. What is the difference? Gaussian curvature: K=1/R2 for the sphere, K=0 for the cylinder. Theorema Egregium (1827): K is measured from inside the surface, without stepping into the surrounding space. One hundred ninety-four years later - Facebook Poincare embeddings, Riemannian SGD, geometric deep learning. All work with manifolds of non-zero curvature. Gauss called the result 'worthy of admiration'. Einstein built general relativity on it. Bronstein builds geometric DL on it.

• **Poincare embeddings (Facebook 2017):** hyperbolic space K=-1 for WordNet - 5 dimensions instead of 200+ Euclidean with better quality
• **Geometric deep learning (Bronstein 2021):** CNNs as a special case on flat K=0 manifolds; spherical CNNs and hyperbolic networks generalize this
• **Cartography:** all map projections distort something - Theorema Egregium makes a perfect flat map of a sphere mathematically impossible
• **3D meshes:** discrete Gauss-Bonnet via angle defects - fast topological check for any triangulated model

Предварительные знания

• Gaussian and Mean Curvature

The Theorem: K is an Intrinsic Invariant

**1827. Gauss publishes Disquisitiones generales circa superficies curvas and calls one result 'theorema egregium' - the remarkable theorem.** Eighty-eight years later Einstein builds general relativity on it. One hundred ninety-four years later Facebook uses the same geometry for Poincare embeddings.

The theorem fits in one line: **Gaussian curvature K can be computed from the first fundamental form E, F, G alone - no normal vector, no embedding in R3 required.** A geodesic bug crawling on the surface can measure K from inside, without ever leaving the surface.

**Consequence: isometries preserve K.** One cannot bend a surface without stretching and change K. A flat sheet (K=0) stays K=0 no matter how it is folded. A sphere (K=1/R2) is not isometric to a plane. Plane, cylinder, and cone are all mutually isometric - all have K=0.

**H is NOT an intrinsic invariant.** A plane has H=0, a cylinder H=1/(2R) - yet they are isometric. A bug on the cylinder cannot detect H from intrinsic measurements. In geometric DL: K is an intrinsic property of the data manifold, H depends on the embedding.

Isometries and Map Projections

An **isometry** is a map f: S1 -> S2 that preserves the metric (first fundamental form). Formally: E = E-tilde, F = F-tilde, G = G-tilde in corresponding parametrizations. Isometries preserve curve lengths, angles, and areas.

Map projections are attempts to find some sphere-to-plane mapping that minimizes distortion. Each class sacrifices something: **Mercator** preserves angles (conformal) but distorts areas (Greenland looks as large as Africa). **Goode's** preserves areas but tears shapes. **Azimuthal** is accurate near the center, distorted near edges. All are consequences of the theorem.

**ML connection - the manifold hypothesis:** high-dimensional data (images, texts) lie on a lower-dimensional manifold. If that manifold has K != 0, it cannot be isometrically 'flattened' into Euclidean space without losing distance information. This is why Nickel & Kiela (Facebook AI, 2017) proposed Poincare embeddings: hyperbolic space (K=-1) for hierarchical data instead of flat R^n.

Gauss-Bonnet: Curvature Equals Topology

**Global Gauss-Bonnet theorem:** for a closed orientable surface M: **integral integral_M K dA = 2pi * chi(M)**, where chi(M) = V - E + F is the Euler characteristic. Total curvature depends only on topology - the specific shape is irrelevant.

Sphere (chi=2): total curvature = 4pi, regardless of size. Torus (chi=0): total curvature = 0 - positive and negative regions cancel exactly. Pretzel with three holes (chi=-4): total curvature = -8pi. This is not coincidence - it is the theorem.

**Discrete Gauss-Bonnet for meshes:** angle defect delta_i = 2pi - sum(triangle angles at vertex i). Then sum_i delta_i = 2pi*chi. One pass over vertices yields the topology of any triangulated mesh - used in mesh processing, 3D printing, CAD analysis.

**Geometric deep learning (Bronstein 2021):** generalization of CNNs to manifolds and graphs. An ordinary CNN is a special case of processing on a flat (K=0) Euclidean manifold. On a sphere (K>0) or hyperboloid (K<0) the convolution architecture must change. Gauss-Bonnet explains why copying the Euclidean architecture fails.

Curvature in ML: Poincare to Riemannian SGD

**Poincare embeddings (Nickel & Kiela, Facebook AI, 2017)** - the first large-scale application of hyperbolic geometry (K=-1) in ML. Core insight: hierarchical data (WordNet: 82,115 concepts, 743,241 relations) naturally inhabit hyperbolic space - trees fit into the Poincare disk without exponential distortion. Euclidean R^n requires 200+ dimensions for the same quality; hyperbolic space needs only 5.

**Riemannian SGD (RSGD)** - generalization of gradient descent to manifolds with K != 0. Standard SGD steps in the tangent space and projects back onto the manifold. For hyperbolic space this is non-obvious: geodesics are curved, exp map and log map are not identities. Implementations: geoopt (PyTorch), McTorch.

**Ricci flow** - metric evolution via dg/dt = -2*Ric (Ric - Ricci tensor, trace of Riemann tensor). Used by Perelman in 2003 to prove the Poincare conjecture (Millennium Prize). In ML: Ollivier-Ricci curvature on graphs as a structural measure for community detection and network bottleneck analysis (Lin, Lu & Yau, 2011).

Key Ideas

• **Theorema Egregium:** K is computed from E, F, G alone. Isometries preserve K. Sphere (K>0) and plane (K=0) are not isometric - hence no perfect map exists
• **Isometry** preserves the first fundamental form - lengths, angles, areas. H is not preserved
• **Gauss-Bonnet:** integral K dA = 2pi*chi(M). Total curvature is a topological invariant
• **Poincare embeddings and RSGD:** K=-1 hyperbolic geometry encodes hierarchies exponentially more efficiently than R^n

Related Topics

Theorema Egregium bridges curvature, metric, and topology:

• Gaussian and Mean Curvature — K = kappa1*kappa2 defined extrinsically in dg-04; here its intrinsic nature is revealed
• Riemann Curvature Tensor — Full intrinsic curvature - generalizes K to manifolds of any dimension
• Gauss-Bonnet Theorem (full) — Chern classes, topological insulators, full statement

Вопросы для размышления

• The torus has integral integral K dA = 0 though K changes sign. Can one build a torus with K=0 everywhere? A flat torus embeds isometrically in R4 but not in R3.
• The Poincare disk uses the Euclidean metric in R2 but a different distance function. Is the Poincare disk isometric to the Euclidean plane? Why?
• The loss landscape of a neural network is a manifold in parameter space. What does 'curvature' of this manifold mean? Is it intrinsic or extrinsic, and what does that imply for measuring it?

Связанные уроки

• dg-04 — K = kappa1*kappa2 via second fundamental form - foundation for Theorema Egregium
• dg-10 — Riemann curvature tensor - full generalization of K to arbitrary-dimensional manifolds
• dg-11 — Gauss-Bonnet in full strength: Chern classes, topological insulators
• ig-02-fisher-metric — Fisher information metric - curvature on the manifold of probability distributions
• ig-07-natural-gradient — Natural gradient = Riemannian SGD on the parameter manifold
• de-03 — Geodesic equations - ODEs on a curved manifold
• calc-01-sequences