Walk the curve at unit speed: arc length
Before tensors on curved surfaces, master the simplest curved thing: a single curve threading through space. Write it as a vector-valued function r(t) = (x(t), y(t), z(t)). The parameter t could be anything — clock time, a dial you turn — and that freedom is exactly the problem. The same curve can be traced fast or slow, and any geometric quantity worth its name must not care which. So the first move is to strip the parametrization down to the one choice the curve itself prefers.
The velocity is r'(t), the derivative you know from Volume I applied component by component, and its length |r'(t)| is the speed. Arc length is the running total of distance actually traveled: s(t) = integral from t0 to t of |r'(u)| du. This is just the Volume-I definite integral you already trust, summing tiny straight-line steps |r'| du and letting them shrink. Notice the integrand |r'(u)| du is the line element ds — the infinitesimal ruler-length along the curve, which is the same ds that will later become the metric's line element on a curved surface.
The tangent, and how fast it turns: curvature
At every point the unit tangent T = dr/ds points the way you are heading — the local 'forward.' Because T has constant length 1, it can never grow or shrink; the only thing it can do is rotate. So the entire shape of the curve is locked up in how T swings as you advance. Differentiate again: dT/ds measures the rate at which the heading turns per meter of curve. The magnitude of this vector is the curvature, written kappa = |dT/ds|. Large kappa means a tight, sudden bend; kappa = 0 means dead straight.
There is a beautiful concrete picture. At a given point, look for the circle that hugs the curve most snugly — matching not just the position and the tangent direction but the bending too. That is the osculating circle ('osculating' is Latin for kissing). Its radius is R = 1/kappa, the radius of curvature, and its center sits a distance R off to the side toward which the curve is bending. A gentle highway curve is a huge kissing circle and tiny kappa; a hairpin is a small circle and big kappa. A straight line is the limiting case of an infinite circle, R = infinity, kappa = 0. This is the same idea as a second-order Taylor approximation from Volume I — but matching a circle to the curve instead of a parabola to a graph.
Where does dT/ds point? Differentiate the identity T dot T = 1 with respect to s: the product rule gives 2 (T dot dT/ds) = 0, so dT/ds is always perpendicular to T. This is not a coincidence to memorize — it is forced by T having constant length, and it is the cleanest example of a recurring trick: differentiating a 'unit-length' or 'orthonormality' constraint produces a perpendicularity relation. The unit vector in the direction of dT/ds is the principal normal N = (dT/ds)/kappa. It points straight at the center of the osculating circle — the direction the curve is falling toward.
A frame that rides with you, and the twist it can't capture alone
You now have two perpendicular unit vectors moving with you: forward T and inward N. In three-dimensional space there is exactly one direction left to complete a right-handed set of axes, and you build it with the cross product: the binormal B = T cross N. The three together — T, N, B, mutually perpendicular, each of unit length — form the [[frenet-serret-frame|Frenet-Serret frame]], a tiny coordinate system that you carry along the curve like a backpack of axes that re-aims itself at every step. T and N span the osculating plane (the plane of the kissing circle); B is the normal to that plane, the axle the curve is momentarily spinning around.
Curvature alone cannot tell the whole story. Imagine two curves both bent into the same gentle arc, but one stays flat in a plane while the other lifts out of it like a spiral staircase or a strand of DNA. They can share the same kappa at every point yet be utterly different curves. What distinguishes them is whether the osculating plane itself tilts as you advance — whether the curve is twisting out of its instantaneous plane. That out-of-plane twist is measured by watching how the binormal B turns, and it is captured by a second number: torsion, written tau.
Run the same constant-length trick on B. From |B| = 1 we get dB/ds perpendicular to B; from B dot T = 0 (differentiated, using that dT/ds is along N) we get dB/ds perpendicular to T as well. The only direction left orthogonal to both B and T is N, so dB/ds must be a pure multiple of N. We name that multiple: dB/ds = -tau N. The minus sign is a convention, and tau can be positive or negative — its sign tells you the handedness of the twist, a right-handed corkscrew versus a left-handed one. A purely planar curve never tips its osculating plane, so for it tau = 0 everywhere; that is the precise statement of 'flat.'
The Frenet-Serret equations: how the whole frame tumbles
We have how T turns (toward N, rate kappa) and how B turns (toward N, rate tau). One derivative is left: how does the normal N change? Use the constant-length trick a final time, now on the orthonormal trio. Each derivative of a frame vector must be a combination of the other frame vectors, and antisymmetry of the turning coefficients (forced again by differentiating the dot products T dot N = 0, N dot B = 0) pins dN/ds down completely: dN/ds = -kappa T + tau B. No new constant appears — N's motion is entirely dictated by the same kappa and tau. Two numbers run the whole show.
THE FRENET-SERRET EQUATIONS (derivatives taken with respect to arc length s)
dT/ds = kappa N
dN/ds = -kappa T + tau B
dB/ds = - tau N
As a single matrix equation, d/ds [T; N; B] = M [T; N; B] with
M = [ 0, kappa, 0 ;
-kappa, 0, tau ;
0, -tau, 0 ]
Note M is ANTISYMMETRIC (M transpose = -M). That is no accident:
an antisymmetric matrix generates a rotation, so the frame can only
ROTATE rigidly as it slides along -- it never stretches or shears.
kappa is the turning rate in the T-N plane; tau the turning rate in the N-B plane.Here is the climax, the fundamental theorem of space curves. Give me any two functions kappa(s) > 0 and tau(s) of arc length, and there exists a curve with exactly that curvature and torsion — and it is unique up to where you place it and how you rotate it in space. In other words, kappa and tau are a complete set of intrinsic instructions: 'bend this hard, twist this hard, at each step,' and the curve is determined. The shape lives entirely in those two scalar functions; the position and orientation are mere placement. Bend with no twist (tau = 0) traces a plane curve; constant kappa and constant nonzero tau traces a perfect helix.
Honest limits, and the bridge to curved spaces
One subtle but vital distinction: the curvature kappa here is extrinsic. It measures how the curve bends inside the surrounding space — a circle drawn on a flat sheet of paper has kappa = 1/R, yet an ant confined to the one-dimensional circle, who can only crawl forward and back, has no way to feel any bending at all. Intrinsically a curve is boringly flat; all its bending is borrowed from how it sits in the ambient world. Hold that thought, because it is exactly the question the rest of this rung tackles for surfaces and spacetime: which curvature is felt from inside, and which is only seen from outside?
Everything you just learned is a rehearsal for the bigger stage. The arc length s came straight from the line element ds, which on a curved surface becomes ds^2 = g_ij dx^i dx^j with the metric tensor g — that is how length and angle are measured once flatness is gone. The Frenet idea of differentiating a vector that rides along a curve, and demanding it stay consistent, matures into the covariant derivative, the right way to differentiate vectors when the very axes are turning. And a curve that turns as little as the geometry permits — the straightest possible path, with zero sideways bending felt from inside — is a geodesic, the curved-space replacement for a straight line and the path light and free particles follow in general relativity. T, N, B were your first moving frame; the rest of this rung hands the whole space one.