Why the ordinary derivative breaks
Guide 1 in this rung gave you a Riemannian metric g on M and ended with a promise — the fundamental theorem of Riemannian geometry guarantees exactly one well-behaved way to differentiate vector fields. This guide is where we open that box. But first we have to feel the problem honestly, because the fix only looks inevitable once you have hit the wall it solves.
Here is the wall. Take a vector field Y on M and try to ask how fast it changes in the direction of another vector X at a point p — the naive answer is the directional derivative of each component. But Y(p) lives in T_p M and Y at a nearby point q lives in a different vector space T_q M. Subtracting them, which is what a difference quotient secretly does, is comparing apples in one room with apples in another. On flat R^n we never noticed because every tangent space is silently identified with R^n itself; on a curved M there is no such free identification.
You can see the breakage in coordinates immediately. Differentiate Y = Y^i d/dx^i component by component and the result transforms with an extra term under a change of chart — the second derivatives of the coordinate change leak in, and the naive 'derivative' is not a tensor. It depends on the chart, which means it is not geometry. A connection is precisely the device that adds a correction term to cancel that leak, restoring a derivative that is the same in every coordinate system.
What a connection actually is
Strip away the metric for a moment. An affine connection on M is a rule nabla that takes a direction X and a vector field Y and returns a new vector field nabla_X Y — read 'the covariant derivative of Y in the direction X'. It is required to be linear over functions in the X slot (so nabla_(fX) Y = f nabla_X Y, meaning it really only depends on X at the point), additive in Y, and to obey a Leibniz product rule in Y: nabla_X (fY) = (Xf) Y + f nabla_X Y. That Leibniz rule is the whole soul of a connection — it says nabla behaves like a derivative, not like a tensor, in its second argument.
There are infinitely many affine connections on a given manifold — the choice is genuinely extra data, not forced by the smooth structure. The geometric content of a connection is parallel transport: it tells you how to slide a vector along a curve while keeping it 'as constant as possible', namely so that its covariant derivative along the curve is zero. Different connections give different notions of 'staying parallel', and that is the whole point. The metric will single one of them out.
Two conditions pin down a unique answer
Among all affine connections, the Levi-Civita connection is the unique one satisfying two natural demands. First, metric compatibility: parallel transport should preserve lengths and angles, equivalently X(g(Y,Z)) = g(nabla_X Y, Z) + g(Y, nabla_X Z) — the connection commutes with the metric the way an ordinary derivative respects a product. Second, torsion-freeness: nabla_X Y − nabla_Y X = [X, Y], where [X, Y] is the Lie bracket. The second condition is what makes mixed second covariant derivatives symmetric, the geometric echo of 'partials commute'.
The miracle, which is the content of the fundamental theorem of Riemannian geometry, is that these two conditions together have exactly one solution. You do not get to choose a connection AND a metric independently; once g is fixed, nabla is determined. The proof is constructive and worth seeing once: write metric compatibility three times with the indices cyclically permuted, add two and subtract the third, and torsion-freeness lets the unwanted brackets collapse. What survives is the Koszul formula, an explicit expression for g(nabla_X Y, Z) purely in terms of g and brackets.
Christoffel symbols: the connection in coordinates
In a chart with coordinate fields d/dx^1, ..., d/dx^n, the connection is completely encoded by how it differentiates the basis fields against each other. The Christoffel symbols Gamma^k_ij are defined by nabla_(d/dx^i) (d/dx^j) = Gamma^k_ij d/dx^k — they are the n^3 numbers (functions, really) saying 'the rate at which the j-th basis field rotates as you move in the i-th direction, expressed back in the basis'. They are emphatically not the components of a tensor: that leftover non-tensorial transformation law from section 1 is exactly what they carry, and it is what makes nabla itself tensorial.
For the Levi-Civita connection the Koszul formula collapses into a clean closed form: Gamma^k_ij is built entirely from the metric components g_ij and their first partial derivatives. The shape to memorize is one half, times the inverse metric, times a cyclic sum of three first-derivatives of g. Two facts fall straight out: torsion-freeness shows up as the lower-index symmetry Gamma^k_ij = Gamma^k_ji, and the symbols vanish at a point exactly when the metric's first derivatives vanish there — which is what 'flat to first order' will mean in guide 3.
Gamma^k_ij = (1/2) g^kl ( d_i g_jl + d_j g_il - d_l g_ij ) d_i = partial derivative in x^i, g^kl = inverse metric symmetry: Gamma^k_ij = Gamma^k_ji (torsion-free) Sphere of radius 1, coords (theta, phi), ds^2 = dtheta^2 + sin^2(theta) dphi^2 Gamma^theta_phiphi = - sin(theta) cos(theta) Gamma^phi_thetaphi = Gamma^phi_phitheta = cot(theta) all other Gamma = 0
Covariant differentiation in practice
Now you can differentiate any vector field for real. Write Y = Y^j d/dx^j and apply the Leibniz rule together with the definition of the symbols. The component answer splits into the part you naively expected plus the Christoffel correction — and that correction is precisely the term that cancels the chart-dependence we exposed at the start. Here is the calculation done slowly, the one every geometer runs in their head.
- Set X = d/dx^i and expand by Leibniz: nabla_(d/dx^i) (Y^j d/dx^j) = (d_i Y^j) d/dx^j + Y^j nabla_(d/dx^i)(d/dx^j).
- Replace the second term using the symbols: nabla_(d/dx^i)(d/dx^j) = Gamma^k_ij d/dx^k. Relabel the dummy index in the first term from j to k so everything sits in the d/dx^k slot.
- Collect: the k-th component is (nabla_i Y)^k = d_i Y^k + Gamma^k_ij Y^j. The first piece is the ordinary partial; the second is the geometric correction.
- Sanity check on flat R^n in Cartesian coordinates: every Gamma vanishes, so the formula reduces to plain componentwise partials — exactly the answer you trusted before you ever met a manifold.
The same machine extends to every tensor by demanding Leibniz across tensor products and that nabla of a function is its ordinary differential. Each upper index contributes a +Gamma term, each lower index a −Gamma term, with the appropriate slots contracted. Two structural facts then come for free from metric compatibility: nabla g = 0 (the metric is covariantly constant — you can raise and lower indices freely through a covariant derivative), and a curve is 'as straight as the geometry allows' exactly when nabla of its velocity along itself is zero, which unpacks into the geodesic equation that opens guide 3.
Why this is the gateway to curvature
Everything downstream in this rung runs through nabla. Parallel transport is integration of the equation nabla of a vector along a curve equals zero; the failure of that transport to return a vector to itself after a small loop is the first appearance of curvature. Indeed the Riemann tensor is built directly from nabla by measuring how badly second covariant derivatives fail to commute: R(X,Y)Z = nabla_X nabla_Y Z − nabla_Y nabla_X Z − nabla_([X,Y]) Z. That commutator is identically zero on flat space and nonzero exactly when the manifold curves.
A vivid sanity picture lives on the sphere from the worked example. Parallel transport a vector around a spherical triangle — say up a meridian, along the equator, back down another meridian — and it returns rotated by an angle equal to the area enclosed. The vector never 'turned' in its own frame; it stayed parallel at every instant. The rotation is pure curvature, detected entirely by nabla. The same vector dragged around any loop in the flat plane comes home unchanged.
One honest caveat before guide 3 takes the wheel. We built nabla from a metric, but the connection idea is strictly more general — affine connections live on any manifold with no metric at all, and the fundamental theorem of Riemannian geometry is the special, beautiful fact that a metric picks out one canonically. In the broader theory of fibre bundles a connection is data on a principal bundle, and the Levi-Civita connection is one instance of that machinery; if you have met the connection form elsewhere in this volume, that is the same idea wearing different notation. Keep the two pictures linked, and you will read both literatures fluently.