Skip to content
🔬 Scientist Pure mathematics — no exam syllabus

Differential Geometry — The Calculus of Curved Space

Differential geometry applies calculus to study curves, surfaces, and higher-dimensional spaces. It asks: what does it mean for a space to be curved? How do you measure distances on a sphere? The answers led to Einstein's general relativity, the classification of surfaces, and the proof of the Poincaré conjecture. It's where geometry meets analysis, producing some of the most beautiful and useful mathematics ever discovered.

"Differential geometry is the study of those properties of surfaces that are invariant under bending and twisting — it's the calculus of shapes."
— Carl Friedrich Gauss

🌀 Curves — The Geometry of Motion

How a Curve Knows How to Turn and Twist
A curve in space is a continuous path γ(t) = (x(t), y(t), z(t)). At each point, it has a tangent vector γ′(t) pointing in the direction of motion. But a curve can do more than go straight — it can bend (curvature κ) and twist (torsion τ). The Frenet-Serret formulas describe how three orthogonal vectors — tangent T, normal N, and binormal B — change as you move along the curve: T′ = κN, N′ = −κT + τB, B′ = −τN. Curvature measures how sharply the curve turns; torsion measures how it lifts out of the plane. A straight line has κ = 0, a circle has constant κ = 1/r, and a helix has both constant κ and constant τ — it's a curve that simultaneously turns and rises.
Every curve has a fingerprint — curvature and torsion uniquely determine its shape up to rigid motion.

Arc Length and Reparametrization

The arc length of a curve from t=a to t=b is ∫ ‖γ′(t)‖ dt. A curve parametrized by arc length (‖γ′(s)‖ = 1) is called unit-speed. In this parametrization, the Frenet formulas simplify because the tangent is already a unit vector. Curvature is just ‖γ″(s)‖ — the magnitude of acceleration when moving at unit speed. The osculating circle at a point has radius 1/κ and best approximates the curve near that point.

Curvature
κ = ‖dT/ds‖
Torsion
τ = −(dB/ds)·N
Arc Length
s = ∫ ‖γ′(t)‖ dt
Frenet Frame
{T, N, B} orthonormal
✨
Computing Curvature of a Helix ▶
Consider the helix γ(t) = (a cos t, a sin t, bt). The tangent is γ′(t) = (−a sin t, a cos t, b), with speed √(a² + b²).

The curvature comes out to κ = a / (a² + b²). When b = 0 (a circle), κ = 1/a — tighter helix = more curvature.

Torsion τ = b / (a² + b²). When b = 0 (circle), τ = 0 — the curve stays in a plane. The helix is the simplest curve with both bending and twisting.

Plane Curves — The Turning Tangent

For plane curves, torsion is zero, and curvature can be signed: positive when turning left, negative when turning right. The total curvature of a closed plane curve is always 2π times the turning number (how many times the tangent winds around). For a simple closed curve (no self-intersections), the turning number is ±1 — this is the Hopf Umlaufsatz, and it's a precursor to the Gauss-Bonnet theorem.

Total Curvature of a Convex Curve
For any convex closed plane curve, the total signed curvature is exactly 2π. This means the tangent vector rotates exactly once as you go around. For a figure-8 curve, the turning number is 0 — the tangent rotates once clockwise and once counterclockwise, canceling out.

🔮 Surfaces — The First Fundamental Form

The First Fundamental Form — How to Do Geometry Without Leaving the Surface
A surface is a 2D shape sitting in 3D space — like the surface of a sphere, a donut, or a saddle. The first fundamental form tells you how to measure distances, angles, and areas on the surface using only information available to a 2D creature living on it. If you parametrize a surface as X(u,v), then the metric coefficients are E = X_u·X_u, F = X_u·X_v, G = X_v·X_v. The length of a curve on the surface is ∫ √(E u′² + 2F u′v′ + G v′²) dt. This is the intrinsic geometry — it doesn't care how the surface sits in space, only how distances work on the surface itself. Gauss called this the inner geometry of the surface.
A 2D creature on a surface can measure lengths and angles without ever looking into the third dimension.
First Fundamental Form
I = E du² + 2F du dv + G dv²
Area Element
dA = √(EG − F²) du dv
Sphere Metric
R²(dθ² + sin²θ dφ²)
Unit Normal
n = (X_u × X_v) / ‖X_u × X_v‖

Gaussian Curvature — The Theorema Egregium

The second fundamental form measures how the surface bends in space — how the unit normal changes as you move. But Gauss made a stunning discovery: the Gaussian curvature K (product of the principal curvatures) can be computed from the first fundamental form alone! This is the Theorema Egregium (Remarkable Theorem). It means a 2D creature can detect curvature by measuring distances and angles on the surface — without looking into the third dimension. A sphere has K = 1/R² (positive curvature — triangles have more than 180°). A saddle has K < 0 (negative curvature — triangles have less than 180°). A cylinder has K = 0 (intrinsically flat — it's just a rolled-up plane).

Theorema Egregium
Gaussian curvature is intrinsic — it depends only on the first fundamental form
This is why you can't flatten a sphere onto a plane without distortion (map projections always stretch something).
✨
Why a Sphere Can't Be Flattened ▶
A sphere has constant positive Gaussian curvature K = 1/R². A plane has K = 0 everywhere. Since Gaussian curvature is intrinsic (Theorema Egregium), any map from the sphere to the plane would have to preserve K — but it can't, because K differs everywhere.

Real-world consequence: Every world map distorts something. Mercator preserves angles but blows up areas near the poles. Lambert preserves areas but distorts shapes. You literally cannot have a map that preserves both distances and angles — curvature is an obstacle.

This is the deep reason we need different map projections — the geometry of the Earth forces us to choose what to preserve.

📐 Geodesics and Curved Spaces

Geodesics — Straight Lines on Curved Surfaces
On a curved surface, what's the shortest path between two points? A geodesic is the curve that locally minimizes distance — it's the closest thing to a straight line on a curved surface. On a sphere, geodesics are great circles (the equator, lines of longitude). On a cylinder, geodesics are helices. The key property: the acceleration vector of a geodesic is always perpendicular to the surface — there's no sideways acceleration. Geodesics satisfy the geodesic equation, a second-order differential equation involving the Christoffel symbols Γᵏᵢⱼ, which encode how the coordinate basis changes from point to point. Parallel transport moves a vector along a curve keeping it "as parallel as possible." On a curved surface, parallel transport around a closed loop rotates the vector — the amount of rotation is exactly the Gaussian curvature enclosed by the loop.
On a curved surface, going "straight" means following a geodesic — and your direction might change even when you think you're going straight.

Riemannian Manifolds

A Riemannian manifold is a smooth manifold with a metric tensor g that defines an inner product on the tangent space at each point. This lets you measure lengths, angles, volumes, and curvature in any dimension. Riemannian geometry generalizes the surface theory of Gauss to n dimensions. The Riemann curvature tensor Rᵢⱼₖₗ describes curvature completely — it tells you how parallel transport around an infinitesimal loop changes a vector. In 2D, it reduces to Gaussian curvature. In 4D, it describes the curvature of spacetime in general relativity.

Metric Tensor
g = Σ gᵢⱼ dxⁱ dxʲ
Christoffel Symbols
Γᵏᵢⱼ = ½gᵏˡ(∂ᵢgⱼₗ + ∂ⱼgᵢₗ − ∂ₗgᵢⱼ)
Geodesic Equation
d²xᵏ/dt² + Γᵏᵢⱼ dxⁱ/dt dxʲ/dt = 0
Ricci Curvature
Rᵢⱼ = Σ Rᵏᵢₖⱼ
Geodesics on a Sphere
On a sphere of radius R, the metric in spherical coordinates is g = R²(dθ² + sin²θ dφ²). The geodesic equation gives great circles: curves that are intersections of the sphere with planes through the center. Airlines use great circle routes because they're the shortest path — that's why flights from New York to Tokyo go over Alaska.

🌌 Curvature, Relativity, and the Shape of the Universe

The Deepest Theorem in Differential Geometry
The Gauss-Bonnet theorem is one of the most beautiful results in all of mathematics. It says: for a compact 2-dimensional surface M, the total Gaussian curvature equals 2π times the Euler characteristic: ∫_M K dA = 2π χ(M). This connects local geometry (curvature at each point) to global topology (the Euler characteristic, which counts handles). A sphere (χ=2) must have positive total curvature — which it does, K = 1/R² everywhere. A torus (χ=0) must have zero total curvature — and indeed, while a torus has both positively curved (outer) and negatively curved (inner) regions, they cancel exactly. For a surface of genus g (g holes), ∫ K dA = 4π(1 − g). This is the first example of a local-to-global theorem — geometry at every point adds up to determine the overall shape of the universe.
The total curvature of a surface is determined entirely by how many holes it has — geometry and topology are bound together.

General Relativity — Geometry Is Gravity

Einstein's general theory of relativity is built on Riemannian geometry — but with a twist. Spacetime is a 4-dimensional Lorentzian manifold (the metric has a time-like direction). The Einstein field equations relate the curvature of spacetime to its matter and energy content:

Einstein Field Equations
Rᵤᵥ − ½gᵤᵥR + Λgᵤᵥ = 8πGTᵤᵥ
Curvature on the left, matter and energy on the right — mass tells spacetime how to curve, curvature tells matter how to move.

The Ricci curvature Rᵤᵥ (a trace of the Riemann tensor) determines how volumes change along geodesics. Positive Ricci curvature means geodesics converge (gravity is attractive). The Weyl curvature (the traceless part) describes tidal forces and gravitational waves. The metric of a black hole (Schwarzschild solution) has a singularity at the center where curvature becomes infinite — a point where the manifold breaks down. The expansion of the universe is encoded in the Friedmann-Lemaître-Robertson-Walker metric, which has constant curvature depending on the density of matter.

✨
Ricci Flow — Perelman's Proof of the Poincaré Conjecture ▶
Ricci flow is a process that smooths out curvature over time, described by ∂g/∂t = −2 Ric(g). It's like the heat equation for the metric — regions of high curvature spread out and average.

Perelman's idea: start with any simply connected 3-manifold and run Ricci flow. The manifold either becomes round (curvature constant everywhere) or develops singularities. Perelman showed that all singularities can be surgically removed, and the flow eventually produces a round 3-sphere.

This proved the Poincaré conjecture (every simply connected closed 3-manifold is homeomorphic to S³) and earned him a Fields Medal and $1M Clay Prize, both of which he declined.

Modern Differential Geometry — Beyond Einstein

Differential geometry continues to evolve. Calabi-Yau manifolds are compact Ricci-flat manifolds that appear in string theory as the hidden dimensions of space. Geometric analysis uses PDEs to study curvature and minimal surfaces. Optimal transport (the Wasserstein metric) puts a Riemannian structure on the space of probability distributions. Information geometry applies Riemannian geometry to statistical manifolds — the Fisher information metric measures distance between probability distributions. Even machine learning uses differential geometry: manifold learning assumes high-dimensional data lies on a low-dimensional curved surface, and geometric deep learning builds neural networks that respect the geometry of the data.

✨
The Shape of the Universe ▶
Is the universe finite or infinite? The answer depends on its curvature. If the universe has positive spatial curvature (like a 3-sphere), it's finite but unbounded — like the Earth's surface, you can travel forever without reaching an edge. If curvature is negative, the universe is infinite and saddle-shaped. If flat (zero curvature), it's infinite Euclidean space.

Measurements from the Planck satellite show the universe is very close to flat (Ω ≈ 1 within 0.4%). But this raises another question: is the universe simply connected? Could space be a 3-torus (finite but flat)? Current data can't rule it out — we might live in a space where traveling far enough returns you to your starting point.
"The miracle of differential geometry is that the universe speaks its language — from the fall of an apple to the shape of the cosmos."
— Inspired by Albert Einstein

🎯 Practice — Check Your Geometry Intuition

These puzzles build differential geometry intuition — no heavy calculus needed.

← Back to Math Map

Second Fundamental Form and Mean Curvature

The first fundamental form captures intrinsic distances on a surface, but the way a surface bends inside the ambient space is encoded in the second fundamental form. Given a regular parametrisation X(u,v) with unit normal N, the coefficients are e = ⟨Xuu, N⟩, f = ⟨Xuv, N⟩, g = ⟨Xvv, N⟩. Together with the first fundamental form coefficients E, F, G they determine the shape operator S = I⁻¹II, whose eigenvalues are the principal curvatures κ₁, κ₂. The mean curvature H = (κ₁+κ₂)/2 = (eG−2fF+gE)/(2(EG−F²)) measures the average bending, while the Gaussian curvature K = κ₁κ₂ = (eg−f²)/(EG−F²) is intrinsic by Gauss’s Theorema Egregium. A surface is minimal precisely when H ≡ 0, and it is totally umbilic when κ₁ = κ₂ everywhere, which characterises spheres and planes. The Codazzi–Mainardi equations ∂ve − ∂uf = eΓ¹₁₂ + f(Γ²₁₂−Γ¹₁₁) − gΓ¹₁₁ and the Gauss equation K = (eg−f²)/(EG−F²) are the integrability conditions for an immersion to exist with prescribed I and II. In applications, the mean curvature vector appears in the first variation of area: δA = −∫ H ⟨V, N⟩ dA for a normal variation V, which underlies the physical model of soap films.

Minimal Surfaces and the Plateau Problem

A minimal surface is a critical point of the area functional; equivalently its mean curvature vanishes identically. Classical examples include the plane, the catenoid (the only non‑planar minimal surface of revolution), the helicoid, and Scherk’s doubly periodic surface. The Plateau problem asks whether a given Jordan curve in ℝ³ bounds a surface of least area. Douglas and Radó independently solved it in the 1930s using the direct method of the calculus of variations: minimise the Dirichlet integral ∫(|Xu|²+|Xv|²) du dv over conformal parametrisations, which yields a harmonic map whose image is a minimal surface. Regularity theory (Morrey, De Giorgi, Nash) shows that the solution is smooth up to the boundary except possibly at branch points. The Weierstrass–Enneper representation expresses any simply connected minimal surface via holomorphic data (g, dh) on a Riemann surface: X(z) = Re ∫ (½(1−g²), i/2(1+g²), g) dh. This links minimal surface theory to complex analysis and allows the construction of complete embedded minimal surfaces of finite total curvature, such as the Costa surface discovered in 1982. The Morse index of a minimal surface counts the number of independent area‑decreasing variations; the catenoid has index one, while the plane has index zero. Stability (index = 0) forces strong topological restrictions: a complete stable minimal surface in ℝ³ must be a plane (Fischer‑Colbrie–Schoen, 1980).

Worked computation of mean curvature for a graph
Let the surface be given by z = f(x,y) = x² − y². Compute the mean curvature H at the origin.
For a graph X(x,y) = (x, y, f(x,y)) the upward unit normal is N = (−fx, −fy, 1)/√(1+fx²+fy²). Here fx = 2x, fy = −2y, so at (0,0) we have N = (0,0,1). The second derivatives are fxx = 2, fxy = 0, fyy = −2. The second fundamental form coefficients at the origin are e = fxx = 2, f = fxy = 0, g = fyy = −2. The first fundamental form at the origin is I = dx²+dy², so E = G = 1, F = 0. Hence H = (eG−2fF+gE)/(2(EG−F²)) = (2·1 + (−2)·1)/2 = 0. The surface is a saddle (hyperbolic paraboloid) with zero mean curvature at the origin, confirming it is a minimal surface only at that point; globally H ≠ 0.
Exam‑day strategy
When a question asks for mean curvature of a graph, write the formula H = div(∇f/√(1+|∇f|²))/2. At a critical point of f the gradient vanishes, so H reduces to (fxx+fyy)/2 = Δf/2. This shortcut saves time and avoids algebraic errors.

Parallel Transport, Holonomy, and the Ambrose–Singer Theorem

Parallel transport moves a tangent vector along a curve while keeping it covariantly constant: ∇γ̇V = 0. On a Riemannian manifold the Levi‑Civita connection preserves the metric, so lengths and angles are preserved. If a vector is transported around a closed loop, it may return rotated; the linear map from the initial tangent space to itself is the holonomy transformation. The set of all such transformations for loops based at a point forms the holonomy group Holp(M). For a simply connected manifold the holonomy group is independent of the base point. The Ambrose–Singer theorem states that the Lie algebra of Holp(M) is generated by the curvature endomorphisms R(X,Y) evaluated on all tangent vectors X,Y at points reachable by horizontal curves. In particular, if the curvature tensor vanishes identically, the holonomy group is trivial and the manifold is locally isometric to Euclidean space. Berger’s classification (1955) lists the possible irreducible holonomy groups for non‑symmetric Riemannian manifolds: SO(n), U(n/2), SU(n/2), Sp(n/4), Sp(n/4)·Sp(1), G₂ (n=7), Spin(7) (n=8). Each case corresponds to a special geometric structure: Kähler, Calabi‑Yau, hyper‑Kähler, quaternion‑Kähler, and the exceptional geometries. Holonomy also controls the existence of parallel spinors, which is crucial in supersymmetric string compactifications.

Comparison Geometry: Rauch, Toponogov, and Bishop–Gromov

Comparison theorems relate the geometry of a Riemannian manifold to that of a space form of constant curvature. Rauch’s comparison theorem (1951) bounds the differential of the exponential map: if sectional curvatures satisfy K ≤ κ, then Jacobi fields grow at least as fast as in the model space of curvature κ; if K ≥ κ, they grow at most as fast. Toponogov’s triangle comparison (1959) states that for a manifold with K ≥ κ, geodesic triangles are “fatter” than corresponding triangles in the κ‑plane: the hinge angle opposite a side is at least the model angle. This yields the Cartan–Hadamard theorem (complete, simply connected, K ≤ 0 ⇒ universal cover diffeomorphic to ℝⁿ) and the splitting theorem (complete, Ric ≥ 0 with a line ⇒ isometric product ℝ × N). Bishop–Gromov volume comparison (1970) says that for Ric ≥ (n−1)κ, the ratio Vol(B(p,r))/Volκ(r) is non‑increasing in r, where Volκ(r) is the volume of a ball in the κ‑space form. Consequently, the asymptotic volume ratio AVR = limr→∞ Vol(B(p,r))/rⁿ is bounded above by the Euclidean volume. These tools underpin the proof of the sphere theorem (¼‑pinched ⇒ homeomorphic to sphere) and the Cheeger–Gromoll soul theorem (complete, non‑negative sectional curvature ⇒ diffeomorphic to normal bundle of a compact totally convex soul).

Spectral Geometry and the Laplace–Beltrami Operator

The Laplace–Beltrami operator Δ = div grad acts on functions on a Riemannian manifold (M,g). In local coordinates Δf = (1/√|g|) ∂i(√|g| gij ∂jf). Its spectrum on a compact manifold consists of a discrete sequence 0 = λ₀ < λ₁ ≤ λ₂ ≤ … → ∞ with finite multiplicities. The first non‑zero eigenvalue λ₁ satisfies the Lichnerowicz estimate λ₁ ≥ nκ/(n−1) when Ric ≥ (n−1)κ > 0 (Obata’s rigidity characterises equality by the round sphere). The heat kernel p(t,x,y) = Σ e−λkt φk(x)φk(y) encodes geometric information; its short‑time asymptotics p(t,x,x) ∼ (4πt)−n/2 Σ ak(x) tk yields the heat invariants a₀ = 1, a₁ = R/6 (scalar curvature), a₂ involving |Ric|², |Riem|², ΔR. The celebrated “Can one hear the shape of a drum?” question asks whether the spectrum determines the isometry class; counterexamples exist (isospectral non‑isometric manifolds by Milnor, Sunada). Nevertheless, spectral data control many topological invariants: the Minakshisundaram–Pleijel zeta function ζ(s) = Σ λk−s analytically continues and its value at s=0 relates to the Euler characteristic via the Gauss–Bonnet–Chern theorem. The Atiyah–Singer index theorem expresses the analytical index of an elliptic operator (e.g. the Dirac operator) in terms of characteristic classes, linking spectral asymmetry to topology.

  • Confusing the second fundamental form with the shape operator; the shape operator is I⁻¹II, not II itself.
  • Assuming every minimal surface is globally area‑minimising; minimal means critical, not necessarily minimal area.
  • Forgetting that holonomy group classification requires simply connected, irreducible, non‑symmetric hypotheses.
  • Applying Toponogov’s theorem without checking the lower curvature bound K ≥ κ.
  • Misusing the Bishop–Gromov ratio by comparing volumes of balls centred at different points.
  • Thinking the spectrum of Δ always determines the manifold up to isometry; isospectral non‑isometric examples exist.
  • Neglecting the role of the heat kernel coefficients a₁, a₂ in recovering curvature integrals.