Riemannian Metrics, Geodesics, and the Exponential Map

A Riemannian metric $g$ equips a smooth manifold with a smoothly varying inner product on each tangent space, turning it into a measuring instrument for lengths, angles, areas, and volumes. Geodesics are the locally length-minimizing curves — the curved-space analogue of straight lines — characterized as zero-acceleration paths of the Levi-Civita connection. The exponential map $exp_{p}$ packages all geodesics through a point $p$ into a single diffeomorphism (locally) from the tangent space onto the manifold, providing the canonical "straightening" coordinates. Foundational to differential geometry (Riemann 1854, Levi-Civita 1917); ubiquitous in general relativity, optimization on manifolds, and shape analysis.

A Riemannian metric on a smooth manifold $M$ is a smooth assignment $p \mapsto g_{p}$ of a positive-definite symmetric bilinear form $g_{p}$ on each tangent space $T_{p} M$ . From this single datum one derives all metric geometry: the length of curves, the Riemannian distance, the unique torsion-free metric-compatible (Levi-Civita) connection, geodesics, curvature, and volume. Geodesics arise in two equivalent guises — as critical points of the length (or energy) functional and as solutions of a second-order ODE — and the exponential map encodes their flow, mapping a tangent vector $v$ to the point reached by traveling unit time along the geodesic with initial velocity $v$ . The surprising fact: geodesics that locally minimize length can globally fail to do so (think great-circle arcs longer than a semicircle), and the exponential map ceases to be injective exactly at the cut locus — geometry's way of encoding global topology and curvature in a single map.

The Core Insight

A Riemannian metric is infinitesimal data — an inner product at each point — yet it determines global geometric structure. The bridge from infinitesimal to global is integration: lengths integrate $g (\overset{γ}{˙}, \overset{γ}{˙})$ , the connection differentiates the metric, geodesics integrate the connection's ODE, and the exponential map integrates geodesics. Each step trades local linear-algebraic data for global differential-geometric structure. The exponential map is the keystone: it lets you transport the flat linear structure of $T_{p} M$ onto the curved manifold, and the failure of this transport (Jacobi fields, conjugate points, cut locus) is precisely what curvature measures.

Formal Statement

Riemannian Metric

Let $M$ be a smooth $n$ -manifold. A Riemannian metric $g$ is a smooth $(0, 2)$ -tensor field such that for every $p \in M$ , $g_{p} : T_{p} M \times T_{p} M \to R$ is symmetric and positive-definite. In local coordinates $(x^{1}, \dots, x^{n})$ with coordinate basis $\partial_{i} = \partial / \partial x^{i}$ ,

g = g_{ij} d x^{i} \otimes d x^{j}, g_{ij} (p) = g_{p} (\partial_{i}, \partial_{j}), g_{ij} = g_{j i}, (g_{ij}) > 0.

⟳Expand a general

(0, 2)

-tensor in the coordinate cobasis

d x^{i}

; symmetry forces

g_{ij} = g_{j i}

, positive-definiteness forces the matrix

(g_{ij})

to have all positive eigenvalues at each point.

The pair $(M, g)$ is a Riemannian manifold. Symbols: $g_{ij}$ is the metric matrix (smooth functions of $p$ ); $g^{ij}$ denotes its matrix inverse, satisfying $g^{ik} g_{k j} = δ_{j}^{i}$ (Einstein summation throughout).

Length, Energy, and Riemannian Distance

For a piecewise-smooth curve $γ : [a, b] \to M$ , define its length and energy:

L (γ) = \int_{a}^{b} g_{γ (t)} (\overset{γ}{˙} (t), \overset{γ}{˙} (t)) d t, E (γ) = \frac{1}{2} \int_{a}^{b} g_{γ (t)} (\overset{γ}{˙} (t), \overset{γ}{˙} (t)) d t .

⟳

L

is the integral of the speed

∥ \overset{γ}{˙} ∥_{g} = g (\overset{γ}{˙}, \overset{γ}{˙})

;

E

replaces the square root by the square, removing reparametrization invariance — by Cauchy–Schwarz

L^{2} \leq 2 (b - a) E

with equality iff constant speed.

The Riemannian distance is

d (p, q) = in f {L (γ) : γ piecewise-smooth, γ (a) = p, γ (b) = q} .

⟳Take the infimum of length [2] over all admissible curves joining

p

q

; smoothness of

g

guarantees the integrand is positive on nonconstant curves, so

d (p, q) = 0 ⟺ p = q

This $d$ is a genuine metric inducing the manifold topology.

Levi-Civita Connection

There exists a unique affine connection $\nabla$ on $(M, g)$ that is torsion-free ( $\nabla_{X} Y - \nabla_{Y} X = [X, Y]$ ) and metric-compatible ( $X g (Y, Z) = g (\nabla_{X} Y, Z) + g (Y, \nabla_{X} Z)$ ). Its Christoffel symbols are

Γ_{ij}^{k} = \frac{1}{2} g^{k ℓ} (\partial_{i} g_{j ℓ} + \partial_{j} g_{i ℓ} - \partial_{ℓ} g_{ij}) .

⟳Write metric compatibility for the three cyclic permutations

X g (Y, Z)

Y g (Z, X)

Z g (X, Y)

; add the first two, subtract the third, and use torsion-freeness

[\partial_{i}, \partial_{j}] = 0

to solve algebraically for

g (\nabla_{\partial_{i}} \partial_{j}, \partial_{ℓ})

— this is the

Koszul formula.

Here $Γ_{ij}^{k} = Γ_{j i}^{k}$ (symmetry from torsion-freeness), and $\nabla_{\partial_{i}} \partial_{j} = Γ_{ij}^{k} \partial_{k}$ .

Geodesic

A curve $γ$ is a geodesic if its velocity is parallel along itself, $\nabla_{\overset{γ}{˙}} \overset{γ}{˙} = 0$ . In coordinates this is the geodesic equation:

\overset{γ}{¨}^{k} (t) + Γ_{ij}^{k} (γ (t)) \overset{γ}{˙}^{i} (t) \overset{γ}{˙}^{j} (t) = 0, k = 1, \dots, n .

⟳Expand

\nabla_{\overset{γ}{˙}} \overset{γ}{˙}

with

\overset{γ}{˙} = \overset{γ}{˙}^{i} \partial_{i}

using the Leibniz rule and

\nabla_{\partial_{i}} \partial_{j} = Γ_{ij}^{k} \partial_{k}

; the chain rule on

\overset{γ}{˙}^{k} (t)

produces

\overset{γ}{¨}^{k}

, the connection produces the quadratic

Γ

-term.

Geodesics have constant speed: $\frac{d}{d t} g (\overset{γ}{˙}, \overset{γ}{˙}) = 2 g (\nabla_{\overset{γ}{˙}} \overset{γ}{˙}, \overset{γ}{˙}) = 0$ .

Existence, Uniqueness, and the Exponential Map

For each $p \in M$ and $v \in T_{p} M$ there is a unique maximal geodesic $γ_{v}$ with $γ_{v} (0) = p$ , $\overset{γ}{˙}_{v} (0) = v$ , depending smoothly on $(p, v)$ . Geodesics rescale: $γ_{λ v} (t) = γ_{v} (λ t)$ . Define the exponential map on the open star-shaped domain $E_{p} = {v \in T_{p} M : γ_{v} defined on [0, 1]}$ by

exp_{p} (v) = γ_{v} (1), exp_{p} (t v) = γ_{v} (t) .

⟳Equation [5] is a second-order ODE; recast it as a first-order system on

T M

and apply Picard–Lindelöf for existence/uniqueness and smooth dependence on initial data. The homogeneity

γ_{λ v} (t) = γ_{v} (λ t)

follows because both sides solve [5] with initial velocity

λ v

Then $d (exp_{p})_{0} = id_{T_{p} M}$ (under the canonical identification $T_{0} (T_{p} M) ≅ T_{p} M$ ), so by the inverse function theorem $exp_{p}$ is a diffeomorphism from a neighborhood of $0 \in T_{p} M$ onto a neighborhood of $p$ .

Gauss Lemma & Local Minimization

Let $v \in T_{p} M$ with $exp_{p}$ defined at $v$ , and let $w \in T_{p} M ≅ T_{v} (T_{p} M)$ . Then

g_{e x p_{p} (v)} (d (exp_{p})_{v} (v), d (exp_{p})_{v} (w)) = g_{p} (v, w) .

⟳Differentiate the energy of a one-parameter family of geodesics (a geodesic variation) and apply the first variation formula; the radial geodesics meet the geodesic spheres

exp_{p} ({∣ v ∣ = r})

orthogonally precisely because the cross term vanishes.

Consequence: within a geodesic ball

exp_{p} (B_{r} (0))

, the radial geodesics are the unique shortest paths from

p

, so

d (p, exp_{p} (v)) = ∣ v ∣_{g}

for small

v

Hopf–Rinow

For a connected Riemannian manifold $(M, g)$ the following are equivalent: (i) $(M, d)$ is a complete metric space; (ii) $exp_{p}$ is defined on all of $T_{p} M$ for some (hence every) $p$ (geodesic completeness); (iii) closed bounded subsets are compact. Any of these implies any two points are joined by a minimizing geodesic.

⟳From geodesic completeness, follow the minimizing radial geodesic of length

d (p, q)

and show via the Gauss lemma and a continuity argument that the set of

r

for which the geodesic minimizes up to distance

r

is open, closed, and nonempty, hence all of

[0, d (p, q)]

Why It Holds / Derivation

Geodesics as Critical Points of Energy

We show that solutions of [5] are exactly the critical points of the energy functional $E$ — establishing the variational characterization.

Step 1 — Set up a variation. Let

γ : [a, b] \to M

with fixed endpoints and let

γ_{s} (t)

s \in (- ε, ε)

, be a smooth variation with

γ_{0} = γ

and

γ_{s} (a) = γ (a)

γ_{s} (b) = γ (b)

. Write the variation field

V (t) = \partial_{s} γ_{s} (t)_{s = 0}

, which vanishes at endpoints. Step 2 — Differentiate energy. With

T = \partial_{t} γ_{s}

and

V = \partial_{s} γ_{s}

\frac{d}{d s} E (γ_{s}) = \int_{a}^{b} g (\nabla_{s} T, T) d t .

Because the variation is smooth and

\nabla

is torsion-free with

[\partial_{s}, \partial_{t}] = 0

pulled back, the symmetry

\nabla_{s} T = \nabla_{t} V

holds. Step 3 — Integrate by parts. Using metric compatibility,

g (\nabla_{t} V, T) = \partial_{t} g (V, T) - g (V, \nabla_{t} T)

. Hence

\frac{d}{d s}_{0} E (γ_{s}) = [g (V, T)]_{a}^{b} - \int_{a}^{b} g (V, \nabla_{t} \overset{γ}{˙}) d t .

Step 4 — Apply boundary conditions. Since

V (a) = V (b) = 0

, the boundary term drops:

\frac{d}{d s}_{0} E (γ_{s}) = - \int_{a}^{b} g (V, \nabla_{\overset{γ}{˙}} \overset{γ}{˙}) d t .

Step 5 — Fundamental lemma. This vanishes for all admissible

V

iff

\nabla_{\overset{γ}{˙}} \overset{γ}{˙} = 0

, which is [5].

■

Length critical points coincide with energy critical points after constant-speed reparametrization (Cauchy–Schwarz $L^{2} \leq 2 (b - a) E$ , equality iff $∥ \overset{γ}{˙} ∥$ constant).

Bernhard Riemann | 1826–1866

In his 1854 Göttingen Habilitationsvortrag, "Über die Hypothesen, welche der Geometrie zu Grunde liegen" (delivered before Gauss, who was reportedly astonished), Riemann introduced the notion of an $n$ -dimensional manifold equipped with a quadratic differential form measuring infinitesimal length — the birth of Riemannian geometry. He defined curvature in arbitrary dimensions, foreseeing that geometry need not be a global a-priori but could vary point to point, governed by a metric. He died of tuberculosis at 39, having also revolutionized complex analysis, number theory, and the theory of integration. His vision of intrinsic, variable-curvature geometry became, sixty years later, the mathematical skeleton of Einstein's general relativity.

Building Intuition

Think of the metric $g_{ij}$ as a position-dependent ruler-and-protractor. At each point it tells you how to convert coordinate displacements into true lengths and how to read off angles. The Christoffel symbols [4] are the correction terms needed because your coordinate grid is, in general, stretched and curved — they tell a moving frame how to stay "parallel" as it slides.

A geodesic is what a particle does when no force acts on it intrinsically: it keeps its velocity vector "as constant as the curved space allows" — $\nabla_{\overset{γ}{˙}} \overset{γ}{˙} = 0$ says the acceleration tangent to the manifold is zero. On a sphere these are great circles; in the plane, straight lines.

The exponential map is "fire an arrow and see where it lands." Stand at $p$ , choose a direction-and-speed vector $v$ in the flat tangent space, shoot a geodesic for unit time, and $exp_{p} (v)$ is the landing point. Because $d (exp_{p})_{0} = id$ , near $p$ the arrow's landing spot is faithfully proportional to the arrow's length and direction — normal coordinates make the manifold look Euclidean to first order, with all $Γ_{ij}^{k} (p) = 0$ and $g_{ij} (p) = δ_{ij}$ . Curvature is exactly the obstruction to extending this Euclidean illusion to second order.

Geodesics radiating from $p$ via the exponential map; a geodesic sphere meets radial geodesics orthogonally (Gauss Lemma).

Worked Examples

Geodesics and Distance on the Round Sphere $S^2(R)$

Given: The sphere of radius

R

with the induced (round) metric. In spherical coordinates

(θ, ϕ)

θ \in (0, π)

colatitude,

ϕ \in [0, 2 π)

longitude:

g = R^{2} d θ^{2} + R^{2} sin^{2} θ d ϕ^{2}, g_{θ θ} = R^{2}, g_{ϕϕ} = R^{2} sin^{2} θ .

Step 1 — Christoffel symbols. From [4], the nonzero ones are

Γ_{ϕϕ}^{θ} = - sin θ cos θ, Γ_{θ ϕ}^{ϕ} = Γ_{ϕθ}^{ϕ} = cot θ .

Step 2 — Geodesic equations [5]:

\ddot{θ} - sin θ cos θ \dot{ϕ}^{2} = 0, \ddot{ϕ} + 2 cot θ \dot{θ} \dot{ϕ} = 0.

Step 3 — Solve. The equator

θ = π /2

ϕ = t

satisfies both (

\dot{θ} = 0

cos θ = 0

). By the rotational symmetry of the round metric, every great circle is a geodesic; these are the only geodesics. Step 4 — Distance. Take two points separated by central angle

Δ σ

. The minimizing geodesic is the shorter great-circle arc:

d (p, q) = R Δ σ, Δ σ \in [0, π] .

Numbers: On Earth (

R \approx 6371

km), London

(51. 5^{\circ} N, 0^{\circ})

to New York

(40. 7^{\circ} N, 7 4^{\circ} W)

gives

cos Δ σ = sin ϕ_{1} sin ϕ_{2} + cos ϕ_{1} cos ϕ_{2} cos Δ λ \approx 0.563

, so

Δ σ \approx 0.974

rad.

d \approx 6371 \times 0.974 \approx 6206 km

(matching the real great-circle flight distance to within rounding).

Exponential Map and Normal Coordinates on the Hyperbolic Plane $\mathbb{H}^2$

Given: The upper half-plane model

H^{2} = {(x, y) : y > 0}

with

g = \frac{d x ^{2} + d y ^{2}}{y ^{2}}, g_{xx} = g_{y y} = y^{- 2} .

This has constant curvature

K = - 1

and is geodesically complete. Step 1 — Geodesics. They are vertical lines and semicircles centered on the

x

-axis (the boundary at infinity). Along the vertical geodesic

x = x_{0}

, parametrize by arclength:

y (s) = y_{0} e^{s}

since

d s = d y / y

. Step 2 — Distance along the vertical. From

p = (x_{0}, y_{0})

q = (x_{0}, y_{1})

d (p, q) = \int_{y_{0}}^{y_{1}} \frac{d y}{y} = ln \frac{y _{1}}{y _{0}} .

Step 3 — Exponential map at $p = (0, 1)$ . A unit tangent vector pointing straight up,

v = (0, a)

in coordinates with

∣ v ∣_{g} = a /1 = a

(since

g = I

(0, 1)

), generates the geodesic

exp_{p} (t v) = (0, e^{t a})

. So

exp_{(0, 1)} (0, a) = (0, e^{a}) .

Step 4 — Check the Gauss-lemma length. The geodesic distance from

p

exp_{p} (v)

equals

∣ v ∣_{g}

d ((0, 1), (0, e^{a})) = ∣ ln e^{a} ∣ = a = ∣ v ∣_{g} . ✓

Numbers: Take

a = ln 5 \approx 1.609

. Then

exp_{(0, 1)} (0, ln 5) = (0, 5)

, and the Riemannian distance is

d = ln 5 \approx 1.609 (dimensionless metric units)

Note: because $K < 0$ , $exp_{p}$ is a global diffeomorphism (Cartan–Hadamard) — no conjugate points, no cut locus.

Common Misconceptions

"Geodesics minimize length." Only locally. A geodesic is a critical point of energy, not necessarily a minimum. On the sphere, the great-circle arc from the North Pole the long way around past the South Pole is a perfectly good geodesic but is not the shortest path. Globally, a geodesic minimizes only up to its first cut point. Beyond the cut locus, it either stops being shortest or stops being unique. " $exp_{p}$ is a global diffeomorphism." False in general. It is only a local diffeomorphism near

0

. On the sphere

S^{n}

exp_{p}

wraps all vectors of length

π R

onto the single antipodal point — a catastrophic failure of injectivity. The largest radius on which

exp_{p}

is injective is the injectivity radius. "The metric is the matrix $g_{ij}$ ." The numbers

g_{ij}

depend on coordinates; the metric is the coordinate-free tensor. Off-diagonal

g_{ij}

encode non-orthogonal coordinate axes, not intrinsic curvature. "Christoffel symbols are tensor components." They are not tensorial — under coordinate change they acquire an inhomogeneous second-derivative term. This is exactly why one can always arrange

Γ_{ij}^{k} (p) = 0

at a chosen point via normal coordinates, even though curvature (built from

\partial Γ

and

ΓΓ

) cannot be made to vanish.

A Geodesic That Fails to Minimize Past Its Conjugate Point

On the unit sphere $S^{2}$ , fix the North Pole $N$ . The geodesic shooting south along a meridian minimizes distance up to the South Pole $S$ , reached at arclength $π$ . The South Pole is the conjugate point of $N$ (all meridians refocus there — a one-parameter family of geodesics of equal length connects $N$ to $S$ ). Continue the geodesic past $S$ : now it is climbing back up another meridian and is strictly longer than the direct route on the other side. Concretely, a geodesic of length $π + 0.1$ from $N$ ends at a point whose true distance from $N$ is only $π - 0.1$ . Thus $exp_{N}$ has a critical point (singular differential) exactly at vectors of length $π$ : the Jacobi field measuring geodesic spreading vanishes there, signaling the conjugate point. This is the prototype for how positive curvature focuses geodesics and destroys global minimization.

Connections

Riemannian metrics sit at the hub of differential geometry. The Levi-Civita connection [4] is the metric specialization of a general Affine Connection; its non-commutativity yields the Riemann Curvature Tensor, whose traces give Ricci and scalar curvature. The second variation of energy is governed by Jacobi Fields, whose zeros mark conjugate points and detect curvature via the Jacobi equation $\nabla_{t}^{2} J + R (J, \overset{γ}{˙}) \overset{γ}{˙} = 0$ . The exponential map provides Normal Coordinates, the cleanest local model. Globally, the interplay of curvature and topology is the subject of Comparison Geometry (Bonnet–Myers, Rauch, Toponogov) and the Cut Locus. The energy functional and its critical points generalize to Morse Theory on Path Spaces. Replacing positive-definite $g$ by signature $(-, +, +, +)$ yields a Lorentzian Manifold, the arena of General Relativity, where geodesics are free-fall worldlines. The metric also defines the canonical Riemannian Volume Form $d V_{g} = det g_{ij} d x^{1} \dots d x^{n}$ and the Laplace–Beltrami Operator.

Real-World Applications

General Relativity. Spacetime is a Lorentzian manifold; the metric $g_{μν}$ is the gravitational field. Planets and light follow timelike/null geodesics ( $\nabla_{\overset{γ}{˙}} \overset{γ}{˙} = 0$ is Einstein's law of free fall). GPS satellite clocks are corrected for metric-induced time dilation (~38 µs/day).

Geodesy & Navigation. Great-circle routing on the (nearly) spherical Earth, and geodesics on the WGS-84 ellipsoid (Vincenty's and Karney's algorithms) for shortest air/sea routes and surveying.

Optimization on Manifolds. Riemannian gradient descent and trust-region methods on the sphere, Stiefel, Grassmann, and SPD-matrix manifolds use $exp_{p}$ (or cheaper retractions) to move along the manifold — used in matrix completion, ICA, dictionary learning (Manopt, Pymanopt).

Medical Imaging / Diffusion MRI. Diffusion tensors at each voxel form a Riemannian manifold of symmetric positive-definite matrices; the affine-invariant metric gives geodesic interpolation and statistics resistant to the "swelling effect" of Euclidean averaging.

Computer Graphics & Geometry Processing. Geodesic distances on triangle meshes (heat method, fast marching) drive texture mapping, mesh parameterization, and shortest-path computation on surfaces.

Shape Analysis / Computational Anatomy. Spaces of shapes and diffeomorphisms (LDDMM) are infinite-dimensional Riemannian manifolds; geodesics interpolate between anatomical shapes and quantify morphological differences.

Robotics & Motion Planning. Configuration spaces with metrics encoding kinematic costs; geodesics give energy-optimal trajectories, and the exponential map underlies manifold-aware sampling-based planners.

Historical Development

1827 — Gauss's Disquisitiones generales circa superficies curvas introduces intrinsic curvature and the Theorema Egregium: curvature depends only on the first fundamental form (the metric), not the embedding.
1854 — Riemann's Habilitationsvortrag generalizes Gauss to $n$ dimensions, defining manifolds with a quadratic line element and sectional curvature — the conceptual founding of Riemannian geometry.
1869–1900 — Christoffel (1869) introduces the symbols $Γ_{ij}^{k}$ and covariant differentiation; Ricci-Curbastro and Levi-Civita develop the absolute differential calculus (tensor calculus), publishing the synthesis in 1900.
1900s–1915 — Tensor calculus matures into the indispensable computational language; Einstein, guided by Grossmann, adopts it for gravitation.
1915–1916 — Einstein's general relativity casts gravity as spacetime curvature; geodesics become free-fall trajectories.
1917 — Levi-Civita defines parallel transport, giving the geometric meaning of the connection and clarifying the geodesic equation; independently Schouten and Hessenberg contribute.
1925–1928 — Élie Cartan develops the method of moving frames, the modern theory of connections, and symmetric spaces; Cartan–Hadamard theorem on complete nonpositively-curved manifolds.
1931 — Hopf and Rinow prove the equivalence of metric completeness, geodesic completeness, and the existence of minimizing geodesics.
1930s–1950s — Synge, Myers, Rauch, and Bonnet establish comparison theorems linking curvature to topology; Morse theory of geodesics (Morse, 1934) and de Rham, Hodge develop global analysis on Riemannian manifolds.

Subtleties & Limits

Cut locus and injectivity radius remain analytically delicate: the cut locus can be non-smooth (a "skeleton" of the manifold), and computing or even estimating it sharply for general metrics is hard. Lower bounds on the injectivity radius in terms of curvature and volume (Cheeger's lemma) are central yet quantitatively subtle. Convergence and collapse. Gromov–Hausdorff limits of Riemannian manifolds with bounded curvature can degenerate (collapsing), and characterizing limit spaces (e.g., Ricci limit spaces, Cheeger–Colding theory, RCD spaces) is an active frontier; whether all such limits are "rectifiable" in suitable senses is partly open. Closed geodesics. Whether every compact Riemannian manifold carries infinitely many geometrically distinct closed geodesics is proven for the 2-sphere (Bangert–Franks–Hingston) but open in general for higher-dimensional spheres with arbitrary metrics — a celebrated unresolved problem. Infinite dimensions. On infinite-dimensional Riemannian manifolds (shape spaces, diffeomorphism groups) the exponential map need not be a local diffeomorphism, Hopf–Rinow can fail, and metric completeness vs. geodesic completeness diverge — requiring substantial functional-analytic care. Prescribed-curvature and isometric-embedding problems (Nash–Kuiper, the smooth Nash embedding, the regularity of isometric embeddings of negatively curved surfaces) blend geometry with hard nonlinear PDE and retain open regularity questions.