Riemannian Geometry: Differential Geometry on Curved Manifolds

Why Long-Haul Flights Bow Toward the Pole

Open any flight-tracking app and look at a long-haul route between two cities that sit at fairly high, similar latitudes but are far apart in longitude — say, somewhere in northern India and somewhere on the US east coast. On the flat map inside the app, the flight path is not a straight line. It arcs upward, bulging toward the Arctic, before curving back down. Passengers sometimes assume the pilot is avoiding weather or restricted airspace. The real reason is geometric, and it has nothing to do with weather: on a sphere, the shortest path between two points is almost never the path you'd draw as "straight" on a flat map, because a flat map cannot represent a sphere without distorting either distances, angles, or areas.

This is not a minor cartographic inconvenience. It is the central question of an entire branch of mathematics. If you cannot trust a flat map to tell you the shortest route, how do you actually compute the shortest path on a curved surface — using only measurements taken on that surface? How do you even define "curved" in a way that a creature confined to the surface, with no way to look at it from outside, could verify for itself? These questions were answered by Carl Friedrich Gauss and, more generally, by Bernhard Riemann in an 1854 lecture that founded what we now call Riemannian geometry — the mathematics of measuring length, angle, and curvature on spaces that are not flat. It underlies GPS, general relativity, computer graphics, and much of modern machine learning, and this chapter builds it from first principles.

Intrinsic Geometry: What an Ant on a Surface Can Actually Know

Take a flat sheet of paper and roll it into a cylinder, without stretching or creasing it. An ant walking on the cylinder's surface, measuring distances and angles between nearby points, would find exactly the same relationships it would find on the original flat sheet — because rolling a flat sheet into a cylinder is a local isometry: it preserves every length and every angle measured along the surface. The ant cannot detect, using only surface measurements, that its world has been rolled up. A triangle drawn on the cylinder still has angles summing to 180°. This is why you can wrap a flat poster around a cylindrical pillar with no wrinkles.

Now try the same thing with a sphere. Peel an orange and try to flatten the peel onto a table without tearing or stretching it — you cannot do it, and this is not a failure of dexterity; it is a theorem. A sphere's surface has a geometric property, called curvature, that a flat plane or a cylinder simply does not have, and no amount of bending (without stretching) can create or destroy it. This is exactly why every flat map of Earth distorts something: on the widely used Mercator projection, Greenland (about 2.17 million km²) appears roughly the same size as Africa (about 30.4 million km²) — a landmass roughly fourteen times larger. The projection cannot avoid this; it is forced to trade accurate area for straight compass bearings.

The property that distinguishes the sphere from the cylinder — the thing the ant on the sphere can detect using only local measurements, while the ant on the cylinder cannot — is called intrinsic curvature. Building a precise, computable definition of it, one that works using only measurements confined to the surface, is the first goal of Riemannian geometry.

The Metric Tensor: Measuring Distance Without Leaving the Surface

On an ordinary flat plane with coordinates $(x,y)$, the distance between two infinitesimally close points is given by the Pythagorean theorem: $ds^2 = dx^2 + dy^2$. This formula — called the Euclidean metric — is the same at every point of the plane. It is what lets you say, without ambiguity, "this segment is 5 cm long."

A curved surface does not have a single flat coordinate grid you can lay across it globally (that is precisely the orange-peel problem). Instead, you describe points on it using some pair of parameters $(u,v)$ — think of these as a stretchy, distorted coordinate grid drawn on the surface, like lines of latitude and longitude on Earth. Because the grid is stretchy, a small step $du$ in the $u$-direction does not correspond to the same physical length everywhere, and it may not even be perpendicular to a step $dv$. You need a local "conversion table" that tells you, at each point, how to turn coordinate steps $(du, dv)$ into actual length.

That conversion table is the metric tensor. At each point on the surface, form the two tangent vectors $\mathbf{r}_u = \partial \mathbf{r}/\partial u$ and $\mathbf{r}_v = \partial \mathbf{r}/\partial v$ (the directions the surface moves as you nudge $u$ or $v$), and define:

E = r_u . r_u
F = r_u . r_v
G = r_v . r_v

Then the length of an infinitesimal step is $ds^2 = E\,du^2 + 2F\,du\,dv + G\,dv^2$. This is called the first fundamental form, and $(E,F,G)$ — or, in general dimension, the full symmetric matrix $g_{ij}$ — is the metric tensor. It is exactly the generalization of the Pythagorean theorem to curved coordinates: instead of one fixed formula, you get a formula whose coefficients change from point to point, encoding exactly how the local coordinate grid is stretched and skewed relative to true, physical distance. In full generality, for an $n$-dimensional space with coordinates $x^1,\dots,x^n$, you write $ds^2=\sum_{i,j} g_{ij}\,dx^i dx^j$ (mathematicians often drop the summation sign here — repeated upper and lower indices are automatically summed, a shorthand called the Einstein summation convention). Every question in Riemannian geometry — distance, angle, area, curvature — is answered using nothing but $g_{ij}$ and calculus. This is the crucial shift: geometry becomes something you compute from a matrix of functions, not something you read off a picture.

Worked Example: The Metric of a Sphere

Parametrize a sphere of radius $R$ using colatitude $\theta$ (the angle down from the north pole, $0 \le \theta \le \pi$) and longitude $\varphi$ ($0 \le \varphi < 2\pi$):

r(theta, phi) = R (sin(theta) cos(phi), sin(theta) sin(phi), cos(theta))

Differentiate to get the tangent vectors:

r_theta = R (cos(theta)cos(phi), cos(theta)sin(phi), -sin(theta))
r_phi   = R (-sin(theta)sin(phi), sin(theta)cos(phi), 0)

Now take the dot products. For $E = \mathbf{r}_\theta \cdot \mathbf{r}_\theta$, using $\cos^2\varphi+\sin^2\varphi=1$ and then $\cos^2\theta+\sin^2\theta=1$, you get $E = R^2$. For $G = \mathbf{r}_\varphi \cdot \mathbf{r}_\varphi$, the same kind of simplification gives $G = R^2\sin^2\theta$. For the cross term $F = \mathbf{r}_\theta \cdot \mathbf{r}_\varphi$, every term contains a factor of $\cos\varphi\sin\varphi - \sin\varphi\cos\varphi = 0$, so $F = 0$ — the $\theta$ and $\varphi$ coordinate lines meet at right angles everywhere, which matches the familiar picture of latitude and longitude lines on a globe. This gives the metric of the sphere:

ds^2 = R^2 dtheta^2 + R^2 sin^2(theta) dphi^2

Sanity-check this against intuition. At the equator ($\theta=\pi/2$, so $\sin\theta=1$), the formula reduces to $ds^2 = R^2 d\theta^2 + R^2 d\varphi^2$ — moving along the equator, a full trip ($\varphi$ from $0$ to $2\pi$) has length $2\pi R$, the correct circumference of a great circle. Near a pole ($\theta \to 0$), the coefficient of $d\varphi^2$ shrinks toward zero — a given change in longitude corresponds to almost no physical distance, exactly matching the fact that lines of longitude converge and nearly touch at the poles. A formula that failed either check would be wrong; this one passes both.

Geodesics: The Straightest Possible Paths

On a flat plane, "shortest path" and "straight line" mean the same thing. On a curved surface, there generally are no straight lines, so Riemannian geometry replaces the idea with geodesics: curves that are locally length-minimizing, or equivalently, curves along which you never have to steer sideways — at every instant, your acceleration points only along the surface's normal direction, never tangentially. A great circle (the intersection of a sphere with a plane through its center) is a geodesic of the sphere; this is why "great-circle route" and "shortest flight path" are the same phrase.

You can find geodesics on a surface of revolution like a sphere using a symmetry argument, the same logic that gives you conservation of angular momentum in physics: if a system has no torque about an axis, angular momentum about that axis is conserved. A free particle sliding without friction on the sphere, moving along a geodesic, feels no tangential force by definition — including no tangential force in the direction "around" the polar axis. The distance from a point at colatitude $\theta$ to the polar axis is $R\sin\theta$ (read directly off the metric: it's $\sqrt{G}$). The resulting conserved quantity, called Clairaut's relation, states that along any geodesic on a surface of revolution, $R\sin\theta \cdot \sin\psi$ is constant, where $\psi$ is the angle the geodesic makes with the local meridian. Two special cases: along the equator ($\theta=\pi/2$, moving purely in the $\varphi$-direction, $\psi=90°$) the relation holds automatically — the equator is a geodesic. Along any meridian ($\varphi$ fixed, $\psi=0°$) it holds trivially too — meridians are always geodesics. But be careful: satisfying Clairaut's relation is necessary for a curve to be a geodesic, not sufficient. A circle of constant latitude away from the equator also has $\psi=90°$ everywhere and trivially keeps $R\sin\theta\cdot\sin\psi$ constant (since $\theta$ itself never changes) — yet, as the next worked example shows numerically, it is not the shortest path between two of its points at all.

Worked Example: A Path That Looks Straight but Isn't

Consider two points on a unit sphere ($R=1$) at the same colatitude $\theta_0 = 30°$ (latitude $60°$), separated by $\Delta\varphi = 120°$ of longitude. There are two natural candidate paths between them: (1) the "naive" path that keeps latitude fixed and changes longitude at a constant rate — this is exactly the path you'd get by linearly interpolating the coordinates $(\theta,\varphi)$ — and (2) the true geodesic, the great circle through the two points. The code below computes the length of each using the tools just derived: the naive path's length via direct integration of $ds$ using the metric tensor, and the geodesic length via the spherical law of cosines (a direct consequence of the same metric).

import numpy as np

def naive_path_length(theta0, phi1, phi2, n=10000):
    """Length of the path that holds colatitude fixed at theta0 while
    phi changes linearly, found by summing ds = sqrt(E dtheta^2 + G dphi^2)
    over n small steps. Here dtheta = 0 and G = sin(theta0)**2 (unit sphere)."""
    phis = np.linspace(phi1, phi2, n + 1)
    dphi = phis[1] - phis[0]
    G = np.sin(theta0) ** 2
    ds = np.sqrt(G) * abs(dphi)
    return ds * n

def geodesic_distance(theta0, dphi):
    """Great-circle distance between two points at the same colatitude
    theta0, dphi apart in longitude, via the spherical law of cosines:
    cos(c) = cos(a)cos(b) + sin(a)sin(b)cos(C), with a = b = theta0
    (angular sides from the pole) and C = dphi (angle at the pole)."""
    cos_c = np.cos(theta0) ** 2 + np.sin(theta0) ** 2 * np.cos(dphi)
    return np.arccos(np.clip(cos_c, -1, 1))

theta0 = np.radians(30)
dphi = np.radians(120)

naive = naive_path_length(theta0, 0, dphi)
geo = geodesic_distance(theta0, dphi)

print(f"Naive constant-latitude path: {naive:.4f} R")
print(f"True geodesic distance:       {geo:.4f} R")
print(f"Naive path is {100 * (naive / geo - 1):.1f}% longer")

Tracing the arithmetic: $\sin(30°)=0.5$, so $G=0.25$ and $\sqrt{G}=0.5$; the naive path length is $0.5 \times \Delta\varphi = 0.5 \times 2.0944 = 1.0472\,R$. For the geodesic, $\cos c = \cos^2(30°) + \sin^2(30°)\cos(120°) = 0.75 + 0.25 \times (-0.5) = 0.625$, so $c = \arccos(0.625) \approx 0.8957\,R$. The program prints:

Naive constant-latitude path: 1.0472 R
True geodesic distance:       0.8957 R
Naive path is 16.9% longer

The path that "looks" straightforward — hold your latitude, change only longitude — is nearly 17% longer than the true shortest path, which instead dips toward the pole before coming back down. This is exactly the mechanism behind the opening hook: for two points at similar, fairly high latitudes separated by a wide range of longitude, the great-circle route visibly bows poleward on a flat map, precisely because a constant-latitude route wastes distance that a pole-ward detour saves.

Curvature: Gaussian Curvature and the Angle-Sum Rule

Now to make "curved" precise. On a flat plane, the interior angles of any triangle sum to exactly 180°. On a sphere, they do not — draw a triangle with one vertex at the north pole and two vertices on the equator, and each angle is already at least as large as the corresponding flat-plane angle, so the sum exceeds 180°. This excess is not a coincidence; it is governed by the Gauss–Bonnet theorem, which for a small triangle on a surface of constant curvature $K$ states:

(sum of angles) - pi = K * (Area of the triangle)

Here $K$ is the Gaussian curvature: on a sphere of radius $R$, $K = 1/R^2$ everywhere (a bigger sphere is more gently curved — think of standing on Earth's surface, which looks locally flat, versus standing on a basketball). On a flat plane, $K=0$ and the formula correctly reduces to the familiar "angle sum equals 180°." On a saddle-shaped surface, $K<0$ and triangle angle sums fall short of 180° — this is called hyperbolic geometry. Crucially, $K$ is computable directly from the metric tensor $g_{ij}$ alone (a fact known as the Theorema Egregium, "remarkable theorem," which Gauss himself considered one of his best results) — you never need to know how the surface sits inside any larger space to compute it.

Worked Example: Angle Excess on a Spherical Triangle

Take the unit sphere ($R=1$, so $K=1$) and the triangle with vertices at the north pole $N=(0,0,1)$, and two equator points $A=(1,0,0)$ and $B=(0,1,0)$ (longitudes $0°$ and $90°$). At each vertex, the interior angle is the angle between the tangent directions of the two geodesic edges meeting there. The tangent direction from a vertex $V$ toward a point $P$ along their connecting great circle is found by projecting $(P-V)$ onto the plane tangent to the sphere at $V$ (subtracting off the component along $V$) and normalizing:

import numpy as np

def tangent_direction(V, P):
    t = P - V * np.dot(V, P)
    return t / np.linalg.norm(t)

def angle_at(V, P, Q):
    tp = tangent_direction(V, P)
    tq = tangent_direction(V, Q)
    cos_angle = np.dot(tp, tq)
    return np.degrees(np.arccos(np.clip(cos_angle, -1, 1)))

N = np.array([0.0, 0.0, 1.0])
A = np.array([1.0, 0.0, 0.0])
B = np.array([0.0, 1.0, 0.0])

angle_N = angle_at(N, A, B)
angle_A = angle_at(A, N, B)
angle_B = angle_at(B, N, A)
total = angle_N + angle_A + angle_B

print(f"Angle at N: {angle_N:.1f}°")
print(f"Angle at A: {angle_A:.1f}°")
print(f"Angle at B: {angle_B:.1f}°")
print(f"Sum: {total:.1f}°, excess over 180°: {total - 180:.1f}°")

Trace it by hand: at $N$, since $N \cdot A = 0$ and $N \cdot B = 0$, the tangent directions are simply $A$ and $B$ themselves, and $A \cdot B = 0$, giving $\arccos(0)=90°$. By the same reasoning (each pair of these three points is mutually perpendicular as vectors) the angle at $A$ and the angle at $B$ are also exactly $90°$. The program prints:

Angle at N: 90.0°
Angle at A: 90.0°
Angle at B: 90.0°
Sum: 270.0°, excess over 180°: 90.0°

Now check this against Gauss–Bonnet. This triangle is bounded by two quarter-meridians and a quarter of the equator, so it covers exactly one-eighth of the sphere's total surface area $4\pi R^2 = 4\pi$; its area is $\pi/2$. Gauss–Bonnet predicts an excess of $K \times \text{Area} = 1 \times \pi/2 = \pi/2$ radians $= 90°$ — matching the number the code computed independently, from raw vector geometry, with no reference to areas at all. Two completely different calculations agreeing is exactly the kind of consistency check that should make you trust a geometric theorem.

Parallel Transport and Holonomy

There is a second, equally important way to feel curvature, beyond angle sums: transport a vector around a closed loop while keeping it "as parallel as possible" at every step — never letting it rotate relative to the surface, only sliding it along. On a flat plane, carrying a vector around any closed loop this way brings it back pointing in exactly its original direction. On a curved surface, it generally does not. Carry an arrow around the spherical triangle from the previous example — start at $N$ pointing along the edge toward $A$, slide it (without twisting) along $N\to A$, then along $A \to B$, then along $B \to N$ back to the start — and the arrow returns rotated by an angle equal to exactly the angle excess of the triangle: $90°$. This mismatch is called holonomy, and the fact that it equals the Gauss–Bonnet excess is not a second coincidence — angle excess and holonomy are two faces of the same underlying curvature. This is the sense in which curvature is a genuinely physical, measurable quantity: an ant carrying a compass around that triangular loop on the sphere would find its compass needle rotated by $90°$ upon return, with no external reference required.

A Common Misconception

It is tempting to think a surface is "curved" only when you can see it bending in three dimensions — that curvature is fundamentally about how a shape sits inside a larger ambient space. This is wrong, and the cylinder-versus-sphere comparison from earlier is the proof: the cylinder visibly bends in 3D, yet it is intrinsically flat ($K=0$ everywhere) — its ant detects no angle excess and no holonomy, because it is isometric to a flat sheet. The sphere's curvature, by contrast, is detectable with a protractor and a piece of string, entirely within the surface, with no reference to any surrounding space at all — that is exactly the content of Gauss's Theorema Egregium: $K$ is computed purely from $g_{ij}$. Riemannian manifolds, in fact, are defined this way from the start — as a set of coordinate charts glued together plus a metric tensor on each — with no embedding required by the definition. (A separate, much later result, Nash's embedding theorem of 1956, shows that any smooth Riemannian manifold happens to fit inside some sufficiently high-dimensional Euclidean space — but this is a convenience for visualization, never a requirement for doing the geometry.)

From Surfaces to Manifolds: Generalizing the Idea

Everything above was built on a 2-dimensional surface, but nothing in the machinery — tangent vectors, the metric tensor $g_{ij}$, geodesics, curvature — actually depends on there being only two coordinates. An $n$-dimensional manifold is a space that locally looks like ordinary $n$-dimensional space (each small patch admits smooth coordinates, called a chart; overlapping charts stitch together into an atlas), equipped with a metric tensor $g_{ij}$ (an $n\times n$ symmetric matrix at each point) that defines distance and angle. In more than two dimensions, a single number $K$ is no longer enough to capture curvature — you need the full Riemann curvature tensor, a more elaborate object that reduces exactly to the Gaussian curvature $K$ when $n=2$. The most consequential application of this generalization is Einstein's general relativity, where spacetime itself is modeled as a 4-dimensional manifold with a metric (technically pseudo-Riemannian, since time and space enter the metric with opposite signs rather than all being positive-definite), and gravity is reinterpreted not as a force but as the curvature of that manifold, governed by the Einstein field equations relating curvature to the distribution of matter and energy.

Where This Shows Up in Computing

Geodesy and GPS depend on differential geometry directly: Earth is modeled not as a sphere but as an oblate ellipsoid (the WGS84 reference ellipsoid), and computing precise distances or bearings between two coordinates uses formulas — such as Vincenty's iterative method — built on exactly the metric-tensor machinery of this chapter, generalized from a sphere to an ellipsoid. GPS also depends on general relativity in a very concrete, corrective way: because satellites sit in weaker gravity than the ground (making their clocks run fast by roughly 45 microseconds per day) but move at high orbital speed (which by special relativity makes their clocks run slow by roughly 7 microseconds per day), the net effect is about 38 microseconds per day of drift. Left uncorrected, that tiny rate would accumulate into several kilometers of positioning error within a single day — the correction built into every GPS satellite's onboard clock is a direct, working application of curved-spacetime geometry.

In computer graphics and geometry processing, computing the shortest path between two points on a 3D mesh (a character's skin, a scanned object) is the discrete analogue of finding a geodesic — used for texture mapping, mesh segmentation, and animation. The heat method introduced by Keenan Crane, Clarisse Weischedel, and Max Wardetzky in 2013 cleverly computes approximate geodesic distances by solving a heat-diffusion equation on the mesh instead of directly solving the geodesic differential equations, because heat diffusion is much cheaper to compute — a good example of Riemannian ideas driving a practical algorithm design choice.

In machine learning, data that looks high-dimensional often actually lies on or near a low-dimensional curved manifold embedded in that high-dimensional space — the classic teaching example is the "Swiss roll," a 2D sheet curled up in 3D. Straight-line (Euclidean) distance between two points on the sheet can badly misjudge how "close" they really are along the sheet's surface, exactly like the naive constant-latitude path in the worked example above misjudged distance on the sphere. The Isomap algorithm (Tenenbaum, de Silva, and Langford, 2000) instead approximates true geodesic distances along the data manifold, using shortest paths through a graph connecting nearby points, and uses those geodesic distances to unfold the manifold into a faithful low-dimensional map — a direct, computational descendant of the ideas in this chapter.

CBSE and Competitive-Exam Mapping

You have already met a one-dimensional shadow of this chapter's curvature idea: CBSE Class 12's "Application of Derivatives" chapter gives the radius of curvature of a plane curve $y=f(x)$ as $R = \dfrac{(1+(dy/dx)^2)^{3/2}}{|d^2y/dx^2|}$. That formula is the $n=1$ case of exactly the same underlying concept generalized here to surfaces and manifolds — both measure how sharply a space bends, using only derivatives available locally. In JEE Advanced, curvature and parametric-curve problems occasionally appear inside coordinate geometry and calculus questions, though full Riemannian machinery is outside the JEE syllabus. Differential geometry of curved space is directly relevant to India's participation in the International Olympiad on Astronomy and Astrophysics (IOAA), where general-relativistic effects (gravitational time dilation, light bending) are tested at a qualitative-to-intermediate level. At the undergraduate-engineering level feeding into GATE, these tools resurface in robotics (configuration-space geometry), computer vision, and machine-learning electives that deal with curved data or curved control spaces.

Summary

• Curvature is intrinsic — detectable by measurements confined entirely to a surface (angle sums, holonomy) — and does not require, or depend on, any embedding in a larger space (Theorema Egregium).
• The metric tensor $g_{ij}$ generalizes the Pythagorean theorem to curved coordinates: $ds^2 = \sum g_{ij}\,dx^i dx^j$, letting you compute length, angle, and area purely from coordinate data.
• For a sphere of radius $R$: $ds^2 = R^2 d\theta^2 + R^2\sin^2\theta\, d\varphi^2$, derived directly from tangent-vector dot products.
• Geodesics generalize straight lines as locally shortest, "no sideways force" paths; Clairaut's relation gives a necessary (not sufficient) conserved quantity for geodesics on surfaces of revolution.
• Gaussian curvature $K$ controls the excess of a triangle's angle sum over $180°$ via Gauss–Bonnet, and equals the rotation angle (holonomy) picked up by a vector parallel-transported around a closed loop — both verified numerically in this chapter for a spherical triangle giving exactly $90°$.
• The theory generalizes from 2D surfaces to $n$-dimensional Riemannian (and pseudo-Riemannian) manifolds, underlying general relativity, and drives real algorithms in GPS/geodesy, computer graphics, and manifold learning.

Practice

1. Using $ds^2=R^2d\theta^2+R^2\sin^2\theta\,d\varphi^2$, compute the circumference of the parallel (circle of latitude) at colatitude $\theta=60°$ on Earth ($R\approx6371$ km). Why is it shorter than the equator's $2\pi R$?
2. Explain, in the language of this chapter, why a cylinder is visibly curved in 3D yet has Gaussian curvature $K=0$ everywhere. What would an ant confined to the cylinder's surface measure for a triangle's angle sum?
3. A spherical triangle on Earth ($R\approx6371$ km) has an angle sum of $200°$. Using Gauss–Bonnet, find its area.
4. True or false, with justification: "If a curve on a surface of revolution satisfies Clairaut's relation, it must be a geodesic." Use the constant-latitude example from this chapter to support your answer.
5. In your own words, explain why Isomap uses graph-shortest-path (geodesic) distances rather than straight-line Euclidean distances when the underlying data lies on a curved manifold like a Swiss roll.

Think About It

Think about this: How would you explain riemannian geometry: differential geometry on curved manifolds to a friend who has never seen a computer? What real-world analogy would you use? Imagine you had to build a system using these concepts — what would be your first step? Try this: before moving on, write down three things you learned and one question you still have.