From a curve's bending to a surface's bending
In the earlier guide on the Frenet-Serret frame you learned to measure how a single curve bends: one number, the curvature kappa, told you how fast the unit tangent turns as you walk along. A surface is harder, and more interesting, because at each point it can bend by different amounts in different directions. Stand on a saddle: walk one way and the ground curves up, walk the perpendicular way and it curves down. A single number can no longer capture that, so we need new machinery. This guide builds it from two objects, traditionally called the first and second fundamental forms.
Set up the picture concretely. Describe a surface by a map r(u, v) that takes a flat patch of (u, v)-coordinates and places it into three-dimensional space — think of (u, v) as longitude and latitude on a globe, and r as the rule that pins each pair to a point of the sphere. The two partial derivatives r_u and r_v are velocity vectors: r_u points along the surface as you increase u with v held fixed, r_v points along the surface as you increase v. These two tangent vectors span the tangent plane at the point, the flat plane that best matches the surface there — exactly the surface analogue of the tangent line to a curve.
Perpendicular to that tangent plane stands a single direction, the unit normal N. You build it as the cross product r_u cross r_v, divided by its length so it has size one. N is the surface's 'which way is up' at each point, and almost everything about bending will come from watching how N tilts as you move. Keep this trio in mind — r_u, r_v lying in the surface, and N pointing out of it — because the two fundamental forms are nothing more than careful bookkeeping of dot products among them.
The first fundamental form: a ruler living in the surface
The first fundamental form answers one question: if I take a tiny step (du, dv) in coordinates, how long is the path it traces on the surface? A small displacement on the surface is dr = r_u du + r_v dv, and its squared length is the dot product dr . dr. Expanding gives ds^2 = E du^2 + 2F du dv + G dv^2, where E = r_u . r_u, F = r_u . r_v, and G = r_v . r_v. These three functions of position are exactly the line element you met for the metric, written out for a 2D surface — E, F, G are the components of the metric tensor in disguise.
Once you have E, F, G you can do all of intrinsic geometry without ever leaving the surface. Lengths of curves come from integrating sqrt(E du^2 + 2F du dv + G dv^2) — the surface version of the arc-length integral for parametric curves. Angles between two directions come from the dot product the form encodes. Areas come from a double integral of sqrt(EG - F^2) du dv. The quantity EG - F^2 is the squared area of the little parallelogram spanned by r_u and r_v — it is the surface's own Jacobian determinant, the stretch factor between flat coordinates and real area.
The second fundamental form: how the surface bends away
The first form is blind to bending — a rolled-up cylinder fools it completely. To see bending you must watch the surface peel away from its own tangent plane, and that is what the second fundamental form records. Take a small step (du, dv); the surface point moves to second order like a little parabola, and the amount it lifts off the tangent plane, measured along the normal N, is II = L du^2 + 2M du dv + N_2 dv^2. Here L = r_uu . N, M = r_uv . N, and N_2 = r_vv . N, the second partial derivatives of r dotted with the unit normal. Each coefficient asks: along this coordinate direction, how strongly does the surface curl toward N?
There is a cleaner way to feel the second form: watch the unit normal N tilt. On a flat plane N never changes — no bending. On a sphere N fans out steadily as you walk, and the faster it swings, the more sharply curved the surface. The rate at which N turns as you move in a tangent direction is captured by an operator (the shape operator), and the second fundamental form is exactly the bending you read off from it. This is the surface echo of the curve idea from before, where bending was the rate the tangent direction turned; now it is the rate the normal direction turns.
Now slice the surface. Pass a plane through the point containing the normal N and one chosen tangent direction; the intersection is a curve, and its ordinary curvature in that plane is called the normal curvature in that direction. As you spin the slicing plane around N, the normal curvature rises and falls. Its maximum and minimum values, kappa_1 and kappa_2, are the principal curvatures, and the two perpendicular directions that achieve them are the principal directions — the saddle's 'up' valley and 'down' ridge made precise.
Gaussian and mean curvature: two ways to fold two numbers
Two principal curvatures kappa_1 and kappa_2 contain the local bending, but we usually want a single summary number. There are two natural ways to combine them, and both matter. The Gaussian curvature is the product, K = kappa_1 kappa_2; the mean curvature is the average, H = (kappa_1 + kappa_2)/2. The product and the average of the same two numbers behave very differently, and that difference is the whole story of how a surface curves.
Read the sign of K and you read the shape. On a sphere both principal curvatures bend the same way, so K = kappa_1 kappa_2 > 0: the surface is dome-like, sitting entirely on one side of its tangent plane. On a saddle they bend opposite ways, one positive and one negative, so K < 0: the surface crosses through its tangent plane. On a cylinder or a cone one principal curvature is zero (the straight direction does not bend at all), so K = 0 even though the thing is visibly rolled up. Mean curvature H tells a different tale: a surface with H = 0 everywhere is a minimal surface, the shape a soap film takes to minimize its area, like the catenoid spanning two rings.
Both curvatures from the fundamental-form coefficients
first form: E F G second form: L M N2
Gaussian curvature K = (L*N2 - M^2) / (E*G - F^2)
Mean curvature H = (E*N2 - 2*F*M + G*L) / (2*(E*G - F^2))
Principal curvatures kappa_1, kappa_2 are the roots of
kappa^2 - 2*H*kappa + K = 0
so K = kappa_1 * kappa_2 and H = (kappa_1 + kappa_2)/2
Unit sphere of radius a: kappa_1 = kappa_2 = 1/a
-> K = 1/a^2 (>0) H = 1/a
Saddle z = x^2 - y^2 at origin: kappa_1 = +2, kappa_2 = -2
-> K = -4 (<0) H = 0 (a minimal point)Gauss's Theorema Egregium: curvature you can feel from inside
Look again at how we defined things. The first fundamental form (E, F, G) is intrinsic — measurable from inside. The second fundamental form (L, M, N_2) seems hopelessly extrinsic, because every coefficient is a dot product with the normal N, and N points out into the surrounding space the inside-ant cannot reach. Both K and H are built from both forms. So you would bet that neither curvature could possibly be known from within. For mean curvature H that bet is correct. For Gaussian curvature K it is spectacularly wrong, and that is Gauss's Theorema Egregium — Latin for 'remarkable theorem', and he chose the word himself.
The theorem says: Gaussian curvature K depends only on the first fundamental form E, F, G and their derivatives — the extrinsic-looking second form cancels out of the answer entirely. So K is intrinsic: the flat ant, armed only with a ruler and never seeing the third dimension, can compute K by measuring how lengths, angles, and areas in its world deviate from flat Euclidean expectations. Concretely, the ant draws a small geodesic circle of radius rho and measures its circumference; on a flat plane that would be 2 pi rho, but on a curved surface it is 2 pi rho times (1 - K rho^2/6 + ...). The shortfall, read through a Taylor expansion in rho, hands the ant the value of K without any reference to the outside.
This is why you cannot flatten a sphere onto paper without distortion, and why every flat world map lies. A sphere has K = 1/a^2 > 0; a flat plane has K = 0; and since K is intrinsic, no amount of bending without stretching can ever turn one into the other — that would change a number both sides are forced to agree on. The same theorem explains why a cylinder unrolls perfectly flat (K = 0 to start) but an orange peel never does (K > 0), and why a pizza slice held flat at the crust droops, but folded into a U-channel stays rigid: folding cannot create the curvature K = 0 forbids, so the slice borrows rigidity from the geometry it cannot escape.
Why this is the doorway to general relativity
The Theorema Egregium is not just a charming fact about oranges; it is the seed of the entire modern view of curved space. Gauss showed that for a 2D surface, curvature is intrinsic — knowable from the metric alone. His student Riemann asked the natural next question: what if you only ever had the metric, in any number of dimensions, with no surrounding space to fall back on? The answer is the Riemann curvature tensor, which packages curvature purely from the first fundamental form generalized to n dimensions, exactly the intrinsic data Gauss isolated. There is no longer any 'outside' to define a normal N — and the remarkable theorem promises you never needed one.
This is exactly the setting general relativity needs. Spacetime is a four-dimensional thing with a metric, and there is no fifth dimension to look in from; gravity is the intrinsic curvature of that metric, felt by matter as the bending of geodesics — the straightest possible paths, which are the surface generalization of straight lines. The machinery you will meet next, the covariant derivative and the Christoffel symbols, is precisely the tool that lets you differentiate on a curved space using only the metric, with no outside reference. Surfaces in three dimensions were the warm-up where you could still cheat by peeking at the normal; from here on, the intrinsic view Gauss opened is the only view there is.