Differential Geometry (4): Intrinsic Geometry — Theorema Egregium and Geodesics

The previous two chapters set up a clear dichotomy. Chapter 2 introduced the first fundamental form $\mathrm{I}$ — the intrinsic metric, what an ant on the surface can measure. Chapter 3 introduced the second fundamental form $\mathrm{II}$ and the shape operator — the extrinsic data, how the surface bends in $\mathbb{R}^3$ . From $\mathrm{II}$ we computed Gaussian curvature $K = \det S$ and mean curvature $H = \mathrm{tr}\,S/2$ . By all appearances, both $K$ and $H$ should depend on the embedding. Bend the surface (without stretching) and you would expect both to change.

For $H$ , that is correct: the cylinder has $H = 1/(2r)$ , the plane has $H = 0$ , even though they are isometric (you can unroll the cylinder flat without stretching).

For $K$ , something miraculous happens: $K$ does not change. The cylinder has $K = 0$ and so does the plane. Both have zero Gaussian curvature even though only one of them looks flat from the outside. This is the Theorema Egregium — Latin for “remarkable theorem” — proved by Gauss in 1827, and it is the central result of classical differential geometry.

Out of the Theorema Egregium come two enormous consequences. First, intrinsic geometry on surfaces is rich enough to define “straight lines” (geodesics), measure curvature, and classify surfaces — all without mentioning the embedding. Second, the same intrinsic apparatus generalizes seamlessly to higher-dimensional manifolds where no ambient space even exists. The path from the Theorema Egregium to general relativity is conceptually straight.

Christoffel Symbols: How Coordinates Twist on a Curved Surface#

In $\mathbb{R}^n$ with the standard coordinates, the basis vectors $\mathbf{e}_1, \ldots, \mathbf{e}_n$ are constant — they do not change from point to point. On a surface parametrized by $\mathbf{x}(u, v)$ , the coordinate basis vectors $\mathbf{x}_u$ and $\mathbf{x}_v$ do change as we move along the surface. The tangent plane tilts, stretches, and rotates from point to point. When we differentiate a vector field expressed in this basis, we cannot just differentiate the components — we must also account for the basis itself varying. This bookkeeping is the role of Christoffel symbols.

where $L_{ij}$ are the second fundamental form coefficients (extrinsic — measuring the normal component of the acceleration) and the Christoffel symbols $\Gamma^k_{ij}$ are the tangential components (intrinsic — measuring how the basis vectors twist within the surface).

where $g_{ij}$ are the metric components ($g_{11} = E$ , $g_{12} = F$ , $g_{22} = G$ ) and $g^{kl}$ are the entries of the inverse metric matrix.

This is the punchline of the entire section: Christoffel symbols depend only on the first fundamental form and its first derivatives. Despite being defined via the second derivatives of the embedding $\mathbf{x}$ , they are computable from the metric coefficients alone. The normal component $L_{ij}$ carries the extrinsic information; the tangential components $\Gamma^k_{ij}$ are purely intrinsic.

The derivation is instructive. Starting from $\mathbf{x}_{ij}\cdot\mathbf{x}_l = \sum_k \Gamma^k_{ij} g_{kl}$ (dot the decomposition with $\mathbf{x}_l$ ; the $L_{ij}\mathbf{n}$ term vanishes because $\mathbf{n}\perp\mathbf{x}_l$ ). Now use the product rule: $\partial_i g_{jl} = \partial_i(\mathbf{x}_j\cdot\mathbf{x}_l) = \mathbf{x}_{ij}\cdot\mathbf{x}_l + \mathbf{x}_j\cdot\mathbf{x}_{il}$ . By cycling indices and combining three such identities, we isolate $\mathbf{x}_{ij}\cdot\mathbf{x}_l = \frac{1}{2}(\partial_i g_{jl} + \partial_j g_{il} - \partial_l g_{ij})$ . Multiplying by $g^{lk}$ (the inverse metric) and summing gives the formula above. The second fundamental form never appeared in the final answer — it dropped out when we projected onto the tangent plane.

Physically, imagine walking on a curved surface while carrying a coordinate grid painted on the surface. As you move, the grid lines curve and spread. The Christoffel symbols quantify this grid distortion: $\Gamma^k_{ij}$ tells you how much the $k$ -th basis vector changes when you move in the $i$ -th direction, holding the $j$ -th coordinate constant. On a flat surface in Cartesian coordinates, all $\Gamma^k_{ij} = 0$ — the grid does not distort. On a sphere in polar coordinates, $\Gamma$ ’s are nonzero because the meridians converge at the poles.

Worked example: unit sphere in spherical coordinates. Take $(\theta, \varphi)$ with $g_{11} = g_{\theta\theta} = \sin^2\varphi$ , $g_{22} = g_{\varphi\varphi} = 1$ , $g_{12} = 0$ . The only nonzero metric derivative is $\partial_\varphi g_{\theta\theta} = 2\sin\varphi\cos\varphi = \sin 2\varphi$ . Computing the Christoffel symbols:

• $\Gamma^\theta_{\theta\varphi} = \frac{1}{2}g^{\theta\theta}(\partial_\varphi g_{\theta\theta}) = \frac{1}{2}\cdot\frac{1}{\sin^2\varphi}\cdot 2\sin\varphi\cos\varphi = \cot\varphi$ .
• $\Gamma^\varphi_{\theta\theta} = -\frac{1}{2}g^{\varphi\varphi}(\partial_\varphi g_{\theta\theta}) = -\frac{1}{2}\cdot 1 \cdot 2\sin\varphi\cos\varphi = -\sin\varphi\cos\varphi$ .
• All other $\Gamma$ ’s vanish (check by plugging in).

The $\cot\varphi$ term diverges as $\varphi \to 0$ (the north pole), reflecting the coordinate singularity there — not a geometric singularity. The sphere is perfectly smooth at the pole; it is the spherical coordinate system that degenerates. A different chart (stereographic projection) would have bounded Christoffel symbols everywhere in its domain.

The Theorema Egregium: Curvature is Intrinsic#

Armed with Christoffel symbols, we can state and understand the central theorem of classical differential geometry.

Theorem (Gauss, 1827 — Theorema Egregium). The Gaussian curvature $K$ is expressible entirely in terms of the metric coefficients $g_{ij}$ and their first and second partial derivatives. In particular, $K$ does not depend on the second fundamental form $\mathrm{II}$ or on any information about how the surface is embedded in $\mathbb{R}^3$ .

where $\Delta = \partial_u^2 + \partial_v^2$ is the flat Laplacian. This makes the intrinsic nature visually obvious: $K$ is determined by $\lambda$ alone, which is the metric.

The proof of the Theorema Egregium proceeds by comparing two expressions for the quantity $\partial_i\Gamma^k_{jl} - \partial_j\Gamma^k_{il} + \text{quadratic terms in }\Gamma$ ’s. The equality of mixed third partials $\mathbf{x}_{ijk} = \mathbf{x}_{jik}$ (applied to the decomposition into tangential and normal parts) yields two sets of equations: the tangential part gives the Gauss equation (relating $K$ to Christoffel symbols), and the normal part gives the Codazzi-Mainardi equations (relating derivatives of $L, M, N$ to Christoffel symbols). The Gauss equation is the Theorema Egregium in formula form.

The key insight about why the theorem works: the combination $LN - M^2$ that defines $K$ (in the numerator of $K = (LN - M^2)/(EG - F^2)$ ) is precisely the combination that appears in the compatibility condition for the surface to exist. The compatibility condition is purely about the intrinsic metric (since it is about whether the embedding equations are consistent), so the combination it produces must also be intrinsic. This is not a coincidence — it is the geometry forcing the algebra.

Consequences that flow immediately:

1. Isometries preserve $K$ . If two surfaces are isometric (same first fundamental form), they have the same Gaussian curvature at corresponding points. The cylinder and the plane both have $K = 0$ — consistent with being isometric. The sphere ($K = 1/r^2$ ) and the plane ($K = 0$ ) are not isometric — no distance-preserving map between them exists. This is the mathematical reason why every flat map of the Earth distorts.

2. Bending preserves $K$ . Bending a surface (deforming it without stretching) is an isometry. So bending cannot change $K$ . Roll a flat piece of paper into a cylinder: $K$ stays at zero. Try to bend paper into a sphere: impossible without stretching, because the sphere has $K > 0$ .

3. $H$ is genuinely extrinsic. The plane has $H = 0$ ; the cylinder has $H = 1/(2r) \neq 0$ . They are isometric, but $H$ differs. Mean curvature detects the embedding; Gaussian curvature does not.

Gauss himself was reportedly surprised by this result. The definition of $K$ involves the Gauss map and the shape operator — manifestly extrinsic objects — yet the value depends only on the intrinsic metric. He called it “egregium” (remarkable), and the name stuck. It is a rare case in mathematics where the discoverer’s excitement is preserved in the theorem’s official name.

A historical note: Gauss proved this during his work surveying the Kingdom of Hanover (1821-1825). The practical question was: can you make an accurate flat map of a piece of the Earth’s surface? The Theorema Egregium gives the definitive answer: no, because the Earth (approximately a sphere with $K > 0$ ) cannot be flattened (to a surface with $K = 0$ ) without distorting distances somewhere. Every cartographic projection — Mercator, Lambert, etc. — introduces distortion, and the Theorema Egregium explains why this is unavoidable.

This has quantifiable implications for navigation and mapmaking. The Mercator projection preserves angles but distorts areas (Greenland appears as large as Africa on a Mercator map, though Africa is 14 times larger). The equal-area projections preserve areas but distort angles. No projection preserves both simultaneously — this follows directly from the Theorema Egregium, since preserving both angles and areas would be an isometry, and no isometry from the sphere to the plane exists. Gauss’s surveying work (which produced the triangulation of Hanover) was the practical context that led him to these theoretical insights. The most profound theorem of classical differential geometry emerged from the mundane problem of making accurate maps.

A related consequence for geodesy: the shape of the Earth is determined (locally) by the Gaussian curvature of its surface. If you measure $K$ at every point (by measuring angles of triangles, per the Gauss-Bonnet local theorem), you know the intrinsic geometry completely. Two planets with the same curvature distribution have the same intrinsic geometry — even if one is round and the other is wrinkled — because the Theorema Egregium makes the shape operator irrelevant for intrinsic questions.

Verification on the isothermal formula. For the unit sphere with stereographic coordinates, $\lambda = 4/(1 + u^2 + v^2)^2$ . Then $\log\lambda = \log 4 - 2\log(1 + r^2)$ where $r^2 = u^2 + v^2$ . Computing $\Delta\log(1 + r^2)$ : in polar form, $\Delta f(r) = f'' + f'/r$ . With $f = \log(1+r^2)$ : $f' = 2r/(1+r^2)$ , $f'' = (2(1+r^2) - 4r^2)/(1+r^2)^2 = (2 - 2r^2)/(1+r^2)^2$ . So $\Delta f = (2-2r^2)/(1+r^2)^2 + 2/(1+r^2) = 4/(1+r^2)^2$ . Then $\Delta\log\lambda = -8/(1+r^2)^2 = -2\lambda$ . By the isothermal formula: $K = -\frac{1}{2\lambda}(-2\lambda) = 1$ . Constant unit curvature on the sphere, computed purely from the metric. No shape operator, no normal vector, no embedding information used.

Geodesics: The Straight Lines of Intrinsic Geometry#

With Christoffel symbols in hand, we can define the curves that play the role of “straight lines” on a curved surface. In flat space, a straight line has zero acceleration. On a surface, the analog is a curve whose tangential acceleration vanishes — it does not steer within the surface, and any acceleration it has is purely normal (forced by the constraint of staying on the surface).

This is a system of two second-order ODEs. By the Picard-Lindelof theorem, given any point $p \in S$ and any tangent vector $\mathbf{v} \in T_pS$ , there exists a unique geodesic through $p$ with initial velocity $\mathbf{v}$ (at least for short time). Geodesics are determined by initial point and direction, just like straight lines in Euclidean space.

Three equivalent characterizations, each with different intuitive content:

1. Zero tangential acceleration. The curve does not “steer” within the surface. All its acceleration comes from the constraint of staying on the surface (normal acceleration). Imagine a ball bearing sliding on a frictionless surface with no gravity — it traces a geodesic.
2. Locally length-minimizing. Among all nearby curves connecting two close points, the geodesic is the shortest. (Warning: this is only local — geodesics can fail to be globally shortest, like the “long way around” on a great circle.)
3. Zero geodesic curvature. The geodesic curvature $\kappa_g$ — the in-plane turning rate — vanishes identically. The curve goes “straight” in the sense of not turning within the surface, even though it may curve when viewed from $\mathbb{R}^3$ .

On the unit sphere, geodesics are great circles — the curves of maximal radius. On a cylinder, geodesics are helices (including the special cases of straight lines along the axis and circles around the cylinder). On the Poincaré disk model of the hyperbolic plane, geodesics are circular arcs perpendicular to the boundary circle (plus diameters).

The first equation can be written as $\frac{d}{dt}(\sin^2\varphi\,\dot\theta) = 0$ , giving the conservation law $\sin^2\varphi\,\dot\theta = c$ (constant). This is Clairaut’s relation: along a geodesic on a surface of revolution, $\rho\sin\psi = $ const, where $\rho$ is the distance from the axis of revolution and $\psi$ is the angle between the geodesic and the parallel (circle of latitude). It is a conservation of angular momentum, reflecting the rotational symmetry.

Clairaut’s relation immediately tells us qualitative facts about geodesics on the sphere: a geodesic that starts heading “eastward” ($\psi$ near $\pi/2$ ) at a low latitude ($\rho$ large) will maintain the product $\rho\sin\psi$ . As it moves toward the pole ($\rho \to 0$ ), it must have $\sin\psi \to \infty$ … which is impossible, so instead the geodesic turns back before reaching the pole. The only geodesics that reach the pole are the meridians ($\psi = 0$ , giving $c = 0$ ). Great circles indeed.

More worked examples. On a cylinder of radius $r$ (with metric $ds^2 = r^2\,d\theta^2 + dz^2$ ), the Christoffel symbols all vanish (the metric coefficients are constant). The geodesic equations reduce to $\ddot\theta = 0$ and $\ddot z = 0$ , so geodesics are curves with $\theta(t) = at + b$ and $z(t) = ct + d$ — helices, with pitch depending on the ratio $c/a$ . Special cases: $a = 0$ gives vertical lines along the axis; $c = 0$ gives circles around the cylinder. If you unroll the cylinder flat, every helix becomes a straight line — a concrete demonstration that geodesics are “intrinsic straight lines” preserved under isometry.

On a cone with half-angle $\alpha$ , parametrized by $\mathbf{x}(u, v) = (v\sin\alpha\cos u, v\sin\alpha\sin u, v\cos\alpha)$ , the metric is $ds^2 = v^2\sin^2\alpha\,du^2 + dv^2$ . Clairaut’s relation gives $v\sin\alpha\sin\psi = $ const. The geodesics can be understood by unrolling the cone into a flat sector of angle $2\pi\sin\alpha$ : geodesics on the cone correspond to straight lines in the sector. When you roll the sector back into a cone, the straight lines become curves that generally do not close up — they spiral around the cone, approaching the apex and then receding. Only specific initial conditions produce closed geodesics.

The variational perspective ties geodesics to physics. The geodesic equation is the Euler-Lagrange equation for the energy functional $E(\gamma) = \frac{1}{2}\int_0^1 |\gamma'|^2\,dt = \frac{1}{2}\int_0^1 \sum_{ij} g_{ij}\dot u_i\dot u_j\,dt$ . Minimizing energy (with fixed endpoints) yields the same curves as minimizing length, but with the additional property that the parametrization is proportional to arc length. This variational formulation — geodesics as extremals of an action integral — is exactly the principle of least action in mechanics. A free particle on a curved surface moves along geodesics because it minimizes the kinetic energy integral. The connection between geometry (shortest paths) and physics (least action) is deep and recurs throughout modern mathematical physics.

Parallel Transport and the Geometry of Holonomy#

Geodesics are intrinsic “straight lines.” But there is a subtler intrinsic concept: carrying a vector along a curve without “rotating” it within the surface. This is parallel transport.

This is a linear first-order ODE system in $V^k(t)$ , so it has a unique solution for any initial $\mathbf{v}_0$ . Parallel transport from $\gamma(0)$ to $\gamma(1)$ is a linear isomorphism $P_\gamma: T_{\gamma(0)}S \to T_{\gamma(1)}S$ that preserves inner products (since the Christoffel symbols come from a metric-compatible connection).

A geodesic is precisely a curve whose tangent vector is parallel-transported along itself: $D\gamma'/dt = 0$ . A geodesic “goes straight” because it carries its own direction without rotating.

Now here is the phenomenon that makes curved geometry fundamentally different from flat geometry: parallel transport around a closed loop does not return the vector to its starting direction. On a flat surface, carrying a vector around any closed path brings it back unchanged. On a curved surface, the vector comes back rotated. The angle of rotation is the holonomy of the loop.

The canonical demonstration: parallel-transport a tangent vector around the boundary of the right-angled spherical triangle (one-eighth of the unit sphere). Start at the north pole with a vector pointing along a meridian (say, toward $(1,0,0)$ ). Walk south along the meridian to the equator: parallel transport along a geodesic preserves the angle with the geodesic, so the vector stays pointing south. At the equator, turn east and walk a quarter of the equator: the vector maintains its angle with the equator (which is also a geodesic), so it stays pointing south. At the point $(0,1,0)$ , walk north back to the pole: again parallel transport along a geodesic. The vector arrives at the pole… but now it points toward $(0,1,0)$ , which is $90°$ rotated from the initial direction toward $(1,0,0)$ .

The vector has been rotated by $\pi/2$ after traversing a loop enclosing area $\pi/2$ . On the unit sphere with $K = 1$ , the holonomy equals $K \cdot \text{Area} = 1 \cdot \pi/2 = \pi/2$ . This is not a coincidence. For any surface, for a small loop enclosing area $\Delta A$ in a region of approximately constant $K$ , the holonomy angle is approximately $K \cdot \Delta A$ . This is a precise intrinsic characterization of Gaussian curvature: $K$ is the holonomy per unit area.

Foucault’s pendulum provides a physical realization. A pendulum swinging at latitude $\varphi$ on Earth has its plane of oscillation parallel-transported along the daily rotation path (a circle of latitude). The enclosed spherical cap has area $2\pi(1 - \cos\varphi)$ , and with $K = 1/R^2$ for a sphere of radius $R$ , the holonomy per day is $2\pi(1 - \cos\varphi) \cdot K \cdot R^2 = 2\pi(1 - \cos\varphi)$ … actually, more precisely, the daily rotation of the pendulum plane is $2\pi\sin\varphi$ (as observed by Foucault), which is the solid angle subtended by the circle of latitude. The point is: the precession of the pendulum is a physical manifestation of holonomy, and holonomy is curvature integrated over area.

The holonomy phenomenon has deep consequences for physics beyond Foucault’s pendulum. In quantum mechanics, Berry’s geometric phase arises when a quantum state is parallel-transported around a loop in parameter space: the state acquires a phase factor determined by the “curvature” of the parameter-space connection. In gauge theory, the Wilson loop — the holonomy of a gauge connection around a closed path — is a fundamental observable that encodes the field strength. Both are direct generalizations of the surface-level phenomenon we just described: curvature manifests as holonomy, the failure of parallel transport to return to its starting value.

The connection to the Theorema Egregium is now complete. We have three equivalent characterizations of Gaussian curvature, all intrinsic:

1. The formula $K = -\frac{1}{2\lambda}\Delta\log\lambda$ in isothermal coordinates (computed from the metric).
2. The holonomy per unit area (measured by parallel transport around small loops).
3. The angle excess per unit area of geodesic triangles (measured by geodesic triangles).

All three give the same number $K$ , all three are computable from $\mathrm{I}$ alone, and all three make the Theorema Egregium geometrically obvious once you accept that parallel transport and geodesics are intrinsic operations. The surprising thing is that this same $K$ also equals the extrinsic formula $\det(S) = (LN - M^2)/(EG - F^2)$ involving the shape operator. The Theorema Egregium says: these are the same number. The intrinsic characterizations and the extrinsic formula agree — always.

Isometry, Constant Curvature, and the Three Model Geometries#

The Theorema Egregium opens a classification program: since isometries preserve $K$ , surfaces with different $K$ are never isometric. What are the surfaces of constant curvature?

Theorem (Minding, 1839). Any two surfaces with the same constant Gaussian curvature are locally isometric. That is, small patches of one can be mapped isometrically onto small patches of the other.

This means the local geometry of a surface of constant curvature is completely determined by the value of $K$ . There are exactly three cases:

$K > 0$ (spherical geometry). The model space is the sphere of radius $1/\sqrt{K}$ . Geodesics are great circles. Any two geodesics intersect (there are no “parallels”). The angle sum of a geodesic triangle exceeds $\pi$ by an amount equal to the area times $K$ . The total area of the sphere is $4\pi/K$ . Every point looks the same as every other (homogeneity).

$K = 0$ (Euclidean geometry). The model space is the plane. Geodesics are straight lines. The parallel postulate holds. The angle sum of any triangle is exactly $\pi$ . Flat, infinite, familiar.

$K < 0$ (hyperbolic geometry). The model space is the hyperbolic plane $\mathbb{H}^2$ . Geodesics diverge exponentially. Through any point not on a given geodesic, infinitely many geodesics pass that never intersect the given one (violating Euclid’s parallel postulate). The angle sum of a geodesic triangle is less than $\pi$ by an amount equal to the area times $|K|$ . There is more area at a given “radius” than in Euclidean geometry — hyperbolic space is “roomier” than flat space.

The hyperbolic plane was the great geometric discovery of the 19th century. Bolyai and Lobachevsky (independently, around 1830) realized that negating the parallel postulate leads to a consistent geometry. The Theorema Egregium and Minding’s theorem showed this was not just a logical curiosity but a genuine geometric space — the geometry of surfaces with constant negative curvature. The pseudosphere (tractrix of revolution) provides a concrete piece of this geometry embedded in $\mathbb{R}^3$ , though Hilbert (1901) proved no complete smooth embedding exists.

The Poincaré disk model represents $\mathbb{H}^2$ as the unit disk with the metric $ds^2 = 4(du^2 + dv^2)/(1 - u^2 - v^2)^2$ . The isothermal formula gives $K = -1$ . Geodesics are circular arcs perpendicular to the boundary circle, and the boundary itself represents “infinity” (infinitely far away in the hyperbolic metric, though visually finite in the Euclidean picture). The stunning images of M.C. Escher’s “Circle Limit” series are tilings of the Poincaré disk by congruent hyperbolic polygons.

A concrete pair illustrating the Theorema Egregium: the helicoid and catenoid. Both are minimal surfaces with negative Gaussian curvature, and there exists a one-parameter family of isometric deformations between them. You can physically bend one into the other (without stretching) by a continuous motion. Their $K$ functions are identical at corresponding points — as the Theorema Egregium guarantees — even though one is a helicoidal ramp and the other is a neck-shaped surface of revolution. They look completely different extrinsically but are intrinsically the same surface.

A subtler example: two non-isometric surfaces can have the same $K$ function without being isometric. What makes the helicoid-catenoid pair special is not just that they have the same $K$ , but that there is a global isometry between them (an explicit map preserving $\mathrm{I}$ ). Having the same $K$ is necessary for isometry (by the Theorema Egregium) but not sufficient — the full first fundamental form must match, not just its derived quantity $K$ .

The three model spaces also admit beautiful explicit descriptions of their geodesics. On the sphere of radius $R$ : every geodesic is a great circle of circumference $2\pi R$ ; any two geodesics intersect (there are no parallels); the geodesic distance between antipodal points is $\pi R$ (the farthest two points can be on a sphere). On the Euclidean plane: geodesics are infinite straight lines; two geodesics either intersect once or are parallel; there is no maximum distance. On the hyperbolic plane of curvature $-1$ : geodesics are infinite and diverge exponentially from each other; through a point not on a given geodesic, infinitely many geodesics pass that never intersect the given one. The exponential divergence of geodesics is measured by the rate at which nearby geodesics separate: on a surface of constant curvature $K$ , two initially parallel geodesics at distance $d$ apart diverge as $d(t) \sim d(0)\cosh(t\sqrt{|K|})$ when $K < 0$ , remain constant when $K = 0$ , and reconverge as $d(t) \sim d(0)\cos(t\sqrt{K})$ when $K > 0$ . This is the Jacobi equation in action, and it explains why negative curvature makes spaces feel “bigger” (geodesics spread apart) while positive curvature makes them feel “smaller” (geodesics reconverge).

Geodesic Curvature and the Bridge to Gauss-Bonnet#

measuring how fast the tangent vector $\gamma'$ rotates within the tangent plane as we move along $\gamma$ .

A geodesic has $\kappa_g \equiv 0$ : its tangent does not rotate in-plane. A circle of latitude on a sphere (other than the equator, which is a geodesic) has nonzero $\kappa_g$ — it is constantly turning within the surface to maintain its latitude. The equator has $\kappa_g = 0$ because it is a great circle; other circles of latitude have $\kappa_g = \cot\varphi$ (where $\varphi$ is the colatitude), measuring their deviation from being geodesics.

The total curvature of $\gamma$ as a space curve satisfies $\kappa^2 = \kappa_n^2 + \kappa_g^2$ , partitioning into extrinsic and intrinsic parts. Crucially, $\kappa_g$ is computable from $\mathrm{I}$ alone — it is intrinsic. This makes it available on abstract manifolds where there is no ambient space.

This formula connects three things: the curvature of the region’s interior ($K$ ), the turning of the smooth boundary ($\kappa_g$ ), and the angular jumps at corners ($\theta_i$ ). All three are intrinsic. They must conspire to give $2\pi$ — the angle of one full rotation. When $K = 0$ and there are no corners, the boundary must turn through exactly $2\pi$ : this is the flat-space theorem that a simple closed curve has winding number $\pm 1$ . Nonzero $K$ “uses up” some of this turning budget internally, so the boundary can turn less (on a positively curved surface) or must turn more (on a negatively curved surface) to compensate.

A worked example brings this to life. Take a geodesic disk of radius $r$ on the unit sphere (the set of all points within geodesic distance $r$ of the north pole — this is a spherical cap bounded by a circle of latitude). The boundary is a circle of latitude at colatitude $\varphi = r$ . Its geodesic curvature is $\kappa_g = \cot r$ (pointing inward). The integrated geodesic curvature is $\int\kappa_g\,ds = \cot r \cdot 2\pi\sin r = 2\pi\cos r$ . The area of the cap is $2\pi(1 - \cos r)$ , so $\iint K\,dA = 2\pi(1-\cos r)$ (with $K = 1$ ). Check: $2\pi(1-\cos r) + 2\pi\cos r = 2\pi$ . The local Gauss-Bonnet formula is satisfied. As $r \to 0$ , the curvature term shrinks and the boundary turning approaches $2\pi$ (a small circle turns almost all the way around, as in flat space). As $r \to \pi/2$ (a hemisphere), the curvature term gives $2\pi(1 - 0) = 2\pi$ and the boundary turning is $2\pi\cos(\pi/2) = 0$ — the equator is a geodesic ($\kappa_g = 0$ ), and all the “turning” comes from the interior curvature. As $r \to \pi$ (the whole sphere minus a point), the curvature gives nearly $4\pi$ and the boundary shrinks to a point, contributing negligible turning.

This is the local form of the Gauss-Bonnet theorem, and it encapsulates the entire relationship between local curvature and global turning. The global form — integrating over a closed surface — will be the climax of the classical theory.

Deeper Examples and Common Pitfalls#

The earlier sections established Christoffel symbols, the Theorema Egregium, geodesics, parallel transport, and constant-curvature model geometries. This section computes them on concrete surfaces in detail, points out the slip-ups beginners make, and shows where these objects do real work outside pure mathematics.

A worked numerical example: Christoffel symbols on the sphere#

Here $g_{\phi\phi} = 1$ , $g_{\theta\theta} = \sin^2\phi$ , off-diagonal $g_{\phi\theta} = 0$ . Inverse: $g^{\phi\phi} = 1$ , $g^{\theta\theta} = 1/\sin^2\phi$ .

$\Gamma^\phi_{\theta\theta} = \tfrac{1}{2} g^{\phi\phi} (-\partial_\phi g_{\theta\theta}) = -\tfrac{1}{2} \cdot 2\sin\phi \cos\phi = -\sin\phi\cos\phi$ .

$\Gamma^\theta_{\phi\theta} = \Gamma^\theta_{\theta\phi} = \tfrac{1}{2} g^{\theta\theta} \partial_\phi g_{\theta\theta} = \tfrac{1}{2 \sin^2\phi} \cdot 2 \sin\phi\cos\phi = \cot\phi$ .

All others vanish. Check: at $\phi = \pi/2$ (equator), $\Gamma^\phi_{\theta\theta} = 0$ and $\Gamma^\theta_{\phi\theta} = 0$ — the metric is locally Euclidean to first order at the equator (in these coordinates, anyway). At $\phi = \pi/4$ , $\Gamma^\phi_{\theta\theta} = -1/2$ and $\Gamma^\theta_{\phi\theta} = 1$ . These specific numbers determine the geodesic equation: $\ddot{\phi} - \sin\phi\cos\phi\, \dot\theta^2 = 0$ and $\ddot\theta + 2\cot\phi\, \dot\phi\, \dot\theta = 0$ . Plug in $\phi(t) = \pi/2$ (constant), $\theta(t) = t$ (uniform). The first equation: $0 - \sin(\pi/2)\cos(\pi/2) \cdot 1 = 0$ , satisfied. The second: $0 + 2\cot(\pi/2) \cdot 0 \cdot 1 = 0$ , satisfied. So the equator is a geodesic, as it should be.

Now try $\phi = \pi/4$ (a non-great circle): the first equation gives $0 - (1/\sqrt{2})(1/\sqrt{2}) \cdot 1 = -1/2 \neq 0$ . Not a geodesic. Numerically, this is exactly the centripetal acceleration the curve fails to provide intrinsically — it would need to deflect downward, but constant-$\phi$ does not. Good: only great circles are geodesics on a sphere.

A worked numerical example: parallel transport around a triangle#

Take the spherical triangle with vertices at the north pole, $(\phi=\pi/2, \theta=0)$ , and $(\phi=\pi/2, \theta=\pi/2)$ — three right angles, area $\pi/2$ on a unit sphere. Start at the north pole with a tangent vector $V_0 = \partial_\phi|_{\theta=0}$ . Parallel-transport along the $\theta=0$ meridian to the equator: $V$ stays equal to $\partial_\phi$ (since the meridian is a geodesic and $V$ is tangent to it). At the equator, $V = \partial_\phi|_{\phi=\pi/2, \theta=0}$ , which points “south.”

Now parallel-transport along the equator from $\theta=0$ to $\theta=\pi/2$ . On the equator, $V$ is perpendicular to the direction of motion (it points south, the motion is east). Parallel transport keeps $V$ pointing south, so at the second vertex $V = \partial_\phi|_{\phi=\pi/2, \theta=\pi/2}$ .

Finally, parallel-transport up the meridian $\theta = \pi/2$ back to the north pole. Going up the meridian, $V$ stays tangent to it, but at the pole the meridian $\theta=\pi/2$ corresponds to a direction rotated by $\pi/2$ relative to the original meridian $\theta=0$ . So $V$ has rotated by exactly $\pi/2$ after a full traverse of the triangle. The holonomy is $\pi/2$ , equal to the area $\pi/2$ times the curvature $K = 1$ . This is Gauss-Bonnet at the local-holonomy level, and it generalizes to: holonomy around a small loop equals the integral of $K$ over the enclosed region.

Intuition + counterexample: why curvature is intrinsic#

The Theorema Egregium says $K$ is computable from the metric alone. Beginners often try the wrong intuition: “But Gaussian curvature is a determinant of second derivatives in $\mathbb{R}^3$ — surely it depends on the embedding?” The resolution is that $K$ is also expressible as a function of $E, F, G$ and their first and second derivatives. Brioschi’s formula is the explicit expression — a 3x3 determinant of partials of $E, F, G$ .

Counterexample to the wrong intuition: take a piece of paper and a cylinder. Both have $K = 0$ everywhere. They have the same first fundamental form. They have different second fundamental forms. Yet the function $K$ computed either way (from second fundamental form, or from first via Brioschi) gives the same answer 0 in both cases. The Theorema Egregium is the assertion that this is no coincidence: $K$ is determined by the first fundamental form alone, and the second fundamental form contributes no extra information to it. The second fundamental form contributes the mean curvature $H$ , which does differ between plane and cylinder.

A third worked example: geodesics on the hyperbolic upper half-plane#

A vertical line $x = $ const, $y(t) = e^t$ satisfies both: $\dot x = 0$ , $\ddot x = 0$ , $\dot y = e^t$ , $\ddot y = e^t$ , so the second equation reads $e^t + (1/e^t)(0 - e^{2t}) = e^t - e^t = 0$ . Confirmed.

A semicircle centered on the real axis, $(x(t), y(t)) = (a + r\tanh t, r/\cosh t)$ , also satisfies them (verify by direct substitution; the algebra is tedious but mechanical). So the geodesics in $\mathbb{H}$ are vertical lines and semicircles meeting the $x$ -axis perpendicularly. This is the Poincaré model of hyperbolic geometry, and the Christoffel symbols you computed give a complete description of straight lines in this model. Distances are infinite as $y \to 0$ (the boundary is at infinite distance), so a hyperbolic plane is complete even though it looks bounded in the embedding.

Counterexample: when isometries fail to exist globally#

Two surfaces with the same Gaussian curvature need not be globally isometric. Take the unit sphere and the surface obtained by gluing two unit hemispheres along their equators — a pinched sphere. Both have $K = 1$ everywhere away from the pinch. But the pinched sphere has different topology and no global isometry to the round sphere exists. The local-to-global gap is important: the Theorema Egregium gives a necessary condition for isometry (matching $K$ ), not a sufficient one. The Cohn-Vossen rigidity theorem partially addresses this: closed convex surfaces with the same intrinsic metric are congruent in $\mathbb{R}^3$ . But for non-convex or non-closed surfaces, the rigidity fails — there exist isospectral but non-isometric closed surfaces (Sunada’s construction).

Common pitfall for beginners#

Beginners often think Gaussian curvature is extrinsic — after all, we computed it from the shape operator, an extrinsic gadget. The whole point of the Theorema Egregium is that it is intrinsic. Here is a way to feel why: take a flat sheet of paper and try to wrap it around a sphere without cuts, folds, or stretches. You cannot — the paper buckles. Empirically, this is because the paper has $K = 0$ and the sphere has $K = 1/r^2$ , and any isometry (length-preserving map) preserves $K$ . Conversely, the paper can be wrapped around a cylinder or a cone, because those have $K = 0$ . The presence or absence of buckling is an intrinsic property of the target surface, not an aesthetic one — it is the Gaussian curvature voting on whether a map exists.

A second pitfall: confusing geodesics with shortest paths. Geodesics are locally shortest, not globally. On a sphere, the great-circle arc from the north pole to the south pole going east-around is a geodesic, but it is twice as long as the one going west-around, which is also a geodesic. Both satisfy the geodesic equation. The variational principle ($\delta \int ds = 0$ ) selects critical points of arc length, not necessarily minima.

Where this matters in physics and engineering#

In general relativity, the Theorema Egregium generalizes to: the Riemann curvature tensor is intrinsic, computable from the metric alone. Einstein’s field equations relate this intrinsic curvature to matter distribution. The fact that Einstein could write $R_{\mu\nu} - \tfrac{1}{2} R g_{\mu\nu} = 8\pi G T_{\mu\nu}$ depends on the left side being computable from the metric — i.e., a generalized Theorema Egregium. Without it, “spacetime curvature” would require an embedding into a higher-dimensional flat space, and physics would degenerate into bookkeeping about the embedding.

In robotics, the configuration space of a robot arm is a manifold (usually a torus, since each joint angle lives on a circle). Path planning uses geodesics on this manifold with respect to a metric encoding actuator energy. The Christoffel symbols are computed once at startup; the geodesic equation becomes the optimal trajectory ODE. Industrial pick-and-place robots solve this in real time using the same machinery you learned in this article.

In deep learning, the loss landscape of a neural network is a Riemannian manifold (with the Fisher information metric, or natural-gradient metric). Natural-gradient descent — used in advanced optimizers — is gradient descent in the metric, which involves Christoffel-like correction terms. Empirically it converges faster than vanilla SGD because it respects the curvature of the parameter space rather than treating it as flat.

Revisiting “what’s next” with sharper questions#

Article 5 will state and prove the Gauss-Bonnet theorem, which says that for a closed surface, $\int_M K\, dA = 2\pi \chi(M)$ . To prepare:

(1) The Theorema Egregium says $K$ is intrinsic. The Euler characteristic $\chi$ is topological. The Gauss-Bonnet theorem links them. Why should an intrinsic geometric quantity equal a topological one? (2) For a triangle on a sphere, the angle sum exceeds $\pi$ by exactly the area times $K$ . How does this local fact integrate up to the global statement? (3) The proof uses parallel transport and the local Gauss-Bonnet for a single triangle. What is the right notion of triangulation, and why does the global integral not depend on the triangulation chosen?

You now have all the local intrinsic geometry. Article 5 packages it into a global theorem. Read it asking “where does the topological invariant $\chi$ enter the proof?” The answer — Euler’s formula $V - E + F = \chi$ — comes in via combinatorics of triangulations, and the magic is how the geometric integral exactly recovers it.

One last worked example: Brioschi’s intrinsic formula on the sphere#

Apply this to the unit sphere with $E = 1$ , $G = \sin^2 \phi$ , where $\phi$ plays the role of $u$ and $\theta$ plays the role of $v$ . Then $\sqrt{EG} = \sin\phi$ . $G_\phi = 2\sin\phi\cos\phi$ , so $G_\phi / \sqrt{EG} = 2\cos\phi$ . $\partial_\phi(2\cos\phi) = -2\sin\phi$ . $E_\theta = 0$ , so the second term vanishes. Plug in: $K = -(1/(2\sin\phi)) \cdot (-2\sin\phi) = 1$ . Confirmed: $K = 1$ on the unit sphere, computed entirely from $E, G$ — without ever invoking the second fundamental form.

This is exactly the Theorema Egregium in action: $K$ has been extracted from the metric alone, giving the same answer as the second-fundamental-form computation. The formula is ugly but it is the one that survives the move to abstract Riemannian manifolds, where there is no embedding into $\mathbb{R}^3$ and therefore no second fundamental form. On a general 2D Riemannian manifold, $K$ is defined by Brioschi’s formula (or its non-orthogonal generalization). The intrinsic-extrinsic equivalence on embedded surfaces is what makes the abstract definition consistent with the classical one.

One more quantitative example: spherical excess at scale#

To feel how curvature emerges quantitatively, compute the geodesic triangle excess on a sphere of radius $R$ at varying scales. Take an equilateral spherical triangle with side length $s$ (along great circles) on a sphere of radius $R = 6371$ km (Earth-sized). The excess satisfies $E = \text{area}/R^2 = (\sqrt{3}/4 \cdot s^2)/R^2$ in the small-triangle limit, by the spherical-trigonometry formula. For $s = 100$ km: $E \approx (\sqrt{3}/4)(10^4)/(6371)^2 \approx 1.1 \times 10^{-4}$ rad $\approx 0.006°$ . Tiny but measurable: classical geodesy used precisely this excess to determine the Earth’s radius from triangulated surveys before satellites existed.

Now scale up. For $s = 1000$ km: $E \approx 0.55°$ , a meaningful angular discrepancy that must be corrected for in continental-scale surveys. For $s = 10000$ km (close to the maximum where the spherical-triangle approximation holds): $E \approx 55°$ — comparable to the three angles themselves, and the small-triangle approximation breaks down entirely. The curvature term becomes the dominant feature of the geometry.

This scaling — “curvature effects are quadratic in length scale, divided by curvature radius squared” — is the universal heuristic for when curvature matters. In GR, near a black hole of Schwarzschild radius $r_s$ , deviations from flat spacetime become $O(1)$ at distances $\sim r_s$ . In differential geometry, the same heuristic governs everything from cartographic projection error to gravitational lensing.

A second numerical anchor: the holonomy of parallel transport around a small geodesic loop of area $A$ on a surface of Gaussian curvature $K$ is approximately $K \cdot A$ rad — exactly the Gauss-Bonnet relationship at infinitesimal scale. For a $1$ -meter loop on Earth’s surface (treating Earth as a sphere), $K \approx 1/R^2 \approx 2.5 \times 10^{-14}$ m$^{-2}$ , area $\approx 1$ m$^2$ , so holonomy $\approx 2.5 \times 10^{-14}$ rad — utterly negligible at human scale. For a $1000$ -km-radius region, holonomy reaches $0.05$ rad $\approx 3°$ — easily measurable, and the basis for precision navigation systems that account for Earth’s curvature.

What’s next#

The Theorema Egregium has now been thoroughly digested. Gaussian curvature $K$ — defined extrinsically through the shape operator — turns out to depend only on the first fundamental form $\mathrm{I}$ and its derivatives. Three equivalent intrinsic characterizations all converge on the same answer: Brioschi’s formula in isothermal coordinates, the holonomy per unit area around a small loop, and the angle excess per unit area of a geodesic triangle. Three roads, one destination. Each pins $K$ firmly on the intrinsic side of the dichotomy, leaving $H$ , $k_1$ , $k_2$ as the genuinely extrinsic invariants that an ant on the surface cannot recover.

The next chapter promotes this local statement into a global one. The Gauss-Bonnet theorem says that integrating $K$ over an entire closed surface yields $2\pi\chi(S)$ , where $\chi$ is the Euler characteristic — a purely topological invariant. Bend, twist, or smoosh a sphere into any shape you like; the total curvature is always $4\pi$ . Any torus, regardless of metric, has total curvature $0$ . Any double torus has total curvature $-4\pi$ . An analytic quantity (the integral of a smooth function defined by second derivatives of the metric) equals a combinatorial quantity (vertices minus edges plus faces of any triangulation). This is the most beautiful theorem in classical differential geometry, and it is the conceptual ancestor of every “index theorem” in modern mathematics — from the Atiyah-Singer index theorem to the Chern-Gauss-Bonnet formula in higher dimensions.

The proof’s key tool is exactly what this chapter set up: the angle excess of a geodesic triangle equals the integral of $K$ over the triangle’s interior. Gauss-Bonnet triangulates the closed surface, sums the angle excesses, and uses Euler’s combinatorial identity $V - E + F = \chi$ to package the result into a single global formula. The next article walks through this carefully, and along the way reveals why a smooth integral can be forced to take only integer (times $2\pi$ ) values — the rigidity that signals deep structure connecting analysis and topology.