Where the last guide left a gap
In the previous guide you built a smooth manifold: a space stitched together from overlapping flat coordinate patches, so that near any point it looks like ordinary R^n, even though globally it can twist, close up, or curve. That construction was deliberately spare. A bare manifold knows its own shape in the loosest topological sense — which points are near which, which paths can be deformed into which — but it has no idea how long any path is, and no idea what angle two directions make. Ask a bare manifold 'how far is it from A to B?' and it simply has no answer to give.
This is exactly the situation the surfaces rung met from the other side. There, a surface came pre-embedded in space, and you measured on it by borrowing the ambient dot product — the three numbers E, F, G of the first fundamental form. A manifold has no ambient space to borrow from; it sits inside nothing. So measurement cannot be inherited from outside — it has to be declared as a new piece of equipment, added on top of the smooth structure by hand. That extra equipment is the Riemannian metric. But before we can install a ruler, we need to be clear about what it measures: the directions you could move in. That is the tangent space, and we build it first.
The tangent space: velocities, not arrows in space
Stand at a point p on the manifold and ask the surfaces-rung question again: in what directions, and at what speeds, could I set off from here? On an embedded surface the honest answer was the tangent plane — a flat sheet of arrows kissing the surface at p. The tangent space, written T_p M, is that idea cut loose from any surrounding space. Take any smooth curve gamma(t) on the manifold that passes through p; its velocity at p is a tangent vector. Collect the velocities of all such curves and you get T_p M: a flat vector space of dimension n, the best linear approximation to the manifold right at p, the little world in which calculus happens.
Here is the conceptual leap that makes a manifold harder than a surface, and it is worth slowing down for. On a surface in space, a tangent vector was a literal arrow you could see, pointing off into the third dimension. On an abstract manifold there is no third dimension to point into — the arrow has nowhere external to live. So a tangent vector is defined intrinsically, by what it does rather than where it points. The cleanest definition: a tangent vector at p is a directional-derivative operator, a machine that takes any smooth function f on the manifold and returns its rate of change in that direction at p. A velocity (d/dt) gamma(t) acts on f by the chain rule to give (d/dt) f(gamma(t)) — the rate at which f changes as you move along the curve. Velocities and directional derivatives are two faces of the same object.
In any coordinate chart this becomes wonderfully concrete. If the chart gives coordinates x^1, ..., x^n near p, then the n partial-derivative operators d/dx^1, ..., d/dx^n form a basis of T_p M. Every tangent vector is a combination v = a^1 (d/dx^1) + ... + a^n (d/dx^n), and the numbers (a^1, ..., a^n) are its components in that chart — the manifold version of the (a, b) coefficients that combined x_u and x_v on a surface. Choose a different chart and these components change by a clean linear rule, but the underlying vector — the actual velocity, the actual operator — does not. A smooth choice of one such tangent vector at every point is a vector field, the manifold's analogue of wind arrows across a weather map.
The metric: a ruler-and-protractor for each tangent space
Now we install measurement. A tangent space is a flat vector space, but a bare vector space — like the abstract plane of all pairs (a, b) — has no built-in notion of length or angle either; those came, on a surface, from the dot product. The fix on a manifold is to supply a dot product by hand, separately at every point. A Riemannian metric g is exactly that: at each point p it is an inner product on the tangent space T_p M — a rule g_p(u, v) that eats two tangent vectors and returns a number, behaving in every way like an ordinary dot product. It is symmetric, g_p(u, v) = g_p(v, u); it is linear in each slot; and it is positive-definite, meaning g_p(v, v) is greater than 0 for every nonzero v, so that lengths are never zero or imaginary.
From this single gadget every measurement flows, by copying the dot-product formulas you already trust. The length of a tangent vector is |v| = square root of g_p(v, v). The angle theta between two tangent vectors comes from the same relation as in school vectors: g_p(u, v) = |u| |v| cos theta, so cos theta = g_p(u, v) / (|u| |v|). Two directions are perpendicular exactly when g_p(u, v) = 0. The decisive word is smoothly: the metric must vary smoothly from point to point, so that g is one continuous field of inner products, not a disconnected pile of unrelated rulers. Glue a smoothly-varying ruler-and-protractor to every tangent space and the bare manifold finally becomes a place where you can measure.
Reading the metric in coordinates
To compute with a metric we write it in a chart, and here the surfaces rung pays off directly. Fix the coordinate basis d/dx^1, ..., d/dx^n. Because the inner product is determined by its values on basis vectors, the metric is completely captured by the numbers g_ij = g_p(d/dx^i, d/dx^j) — the inner product of basis direction i with basis direction j. These g_ij are the components of the metric tensor, and they are nothing but the higher-dimensional version of E, F, G. On a 2-dimensional patch, g_11 is E, g_22 is G, and g_12 = g_21 is F; symmetry of the inner product is exactly why g_ij = g_ji and why E, F, G were only three numbers rather than four.
With the components in hand, the length of a tiny step with coordinate displacement (dx^1, ..., dx^n) is the line element ds^2 = sum over i, j of g_ij dx^i dx^j. In two dimensions this is literally ds^2 = E du^2 + 2F du dv + G dv^2 again — the same line element from the surfaces rung, now understood not as borrowed from an ambient dot product but as the manifold's own declared metric. The whole apparatus of arc length, angle, and area carries over verbatim: integrate ds along a curve for its length; use the area element square root of (determinant of g) for volume. The metric tensor is the grammar of all measurement, written from inside.
the metric tensor in a chart g_ij = g(d/dx^i, d/dx^j) symmetric: g_ij = g_ji 2D dictionary: g_11 = E g_12 = g_21 = F g_22 = G length of a tangent vector v : |v| = sqrt( g(v, v) ) angle between u and v : cos theta = g(u, v) / (|u| |v|) line element (tiny step dx^i) : ds^2 = sum_ij g_ij dx^i dx^j
A worked example: the upper half-plane
Let us see a metric reshape a space we think we know. Take the ordinary upper half-plane — all points (x, y) with y greater than 0 — but instead of the usual flat ruler, declare the metric ds^2 = (dx^2 + dy^2) / y^2. In components that is g_11 = g_22 = 1 / y^2 and g_12 = 0. The manifold is just the boring half-plane; the geometry is anything but. Notice the y^2 in the denominator: a coordinate step of fixed size (dx, dy) has true length divided by y. High up (large y) real distances shrink, so big-looking journeys are short; down near the x-axis (small y) the same coordinate step costs an enormous real length — the axis is infinitely far away, an unreachable horizon.
This is not an idle game. The space you have just built is the Poincare upper half-plane model of hyperbolic geometry, met back in the non-Euclidean rung — there it was handed to you as a model; here you have constructed it from a single metric on a flat domain. The same set of points, given a different g, becomes a world where through a point off a line there are infinitely many non-meeting lines, where triangles have angle sum less than 180 degrees. The lesson lands hard: it is the metric, not the underlying set of points, that decides which geometry you are living in.
Why this is the language relativity speaks
Step back and see what you now command. A smooth manifold supplies a curved arena; the tangent space at each point supplies the local linear world where velocities and the differential of a map live; and a Riemannian metric supplies the smoothly-varying inner product that turns the whole thing into a place with genuine distance and angle. Crucially, all of this is intrinsic — defined entirely from inside, with no ambient space anywhere in sight. A creature confined to the manifold, who could never step outside it, owns every one of these tools and can measure with them. That is the decisive break from the surfaces rung, where measurement still leaned on an embedding.
One honest extension closes the loop to the rung's final guide. Einstein's general relativity uses almost exactly this picture, with a single relaxation: it drops the positive-definite demand. Spacetime carries a pseudo-Riemannian metric, where g(v, v) can be positive, negative, or zero — the negative directions are time, the positive ones space, and the zero ones the paths of light. The bilingual reader should hold the term steady: a Riemannian metric proper is positive-definite, as we built it; the relativistic version is its near cousin. Everything else — tangent spaces, the metric tensor g_ij, the line element ds^2 — transfers intact. You have just built the grammar in which the geometry of curved spacetime will be written.