Geodesics, the exponential map, and completeness

Intuition Beginner

A geodesic is the straightest curve you can draw on a curved surface. On a flat plane, that is just a straight line. On a sphere, it is a great circle, the path an aeroplane takes between two distant cities when it hugs the shortest route. The defining idea is local: at every instant, the curve is not turning toward any side, so it keeps going as straight as the surface allows.

The exponential map is a way to lay out an entire patch of the surface using the geodesics that start at a single base point. Stand at a point and fire geodesics outward in every direction, labelling each by the tangent vector you launched it with. Nearby points on the surface correspond to nearby launch vectors, and the length of the launch vector equals the distance you travel. The exponential map is the dictionary between "direction and distance to travel" and "where you land."

Completeness is the demand that nothing on the surface stops a geodesic short. On a complete surface every geodesic can be extended forever, never running off an edge or into a hole. The Hopf-Rinow theorem is the headline result: completeness is equivalent to three other conditions, and any of them guarantees that between any two points there is a geodesic of shortest possible length.

Visual Beginner

Picture the north pole of a round sphere, with tangent vectors drawn as arrows pinned there. Each arrow launches a geodesic, a great circle radiating outward. Together the arrowheads at a fixed length trace a circle of latitude, and the exponential map is the rule sending each arrow to the point its geodesic reaches.

Geodesics run forever on the round sphere, so it is complete; every pair of points is joined by a shortest great-circle arc.

Worked example Beginner

Take the unit sphere and stand at the north pole. The exponential map sends a tangent arrow of length $\theta$ to the point you reach by walking a geodesic (a meridian) through angle $\theta$. The length of the arrow equals the distance walked, because geodesics travel at unit speed here.

Step 1. Fire an arrow of length $\pi /2$. You walk from the north pole down a meridian through a quarter of the sphere and land on the equator. The geodesic distance from the pole to the equator is $\pi /2$.

Step 2. Fire an arrow of length $\pi$. You walk all the way through a meridian and arrive at the south pole. The geodesic distance from pole to pole is $\pi$, which is the longest shortest-path on the unit sphere.

Step 3. Fire an arrow of length $3\pi /2$. You overshoot the south pole and come back up the far side to the equator, but now the shortest path from the pole to your landing point is only $\pi /2$ the other way around. The exponential map is still defined — the geodesic keeps running — but it has stopped giving the shortest route.

The lesson: completeness means geodesics run forever, and up to the first conjugate point (here the south pole) they also give the shortest path. Hopf-Rinow promises that on a complete surface some minimizing geodesic always exists between any two points.

Check your understanding Beginner

Formal definition Intermediate+

Let $(M,g)$ be a Riemannian manifold with Levi-Civita connection $\nabla$ ^{[Lee Ch. 4]}. A smooth curve $\gamma :I\to M$ is a geodesic when its velocity is parallel along itself, ${\nabla }_{\dot{\gamma }}\dot{\gamma }=0$. In local coordinates $(x^i)$, writing ${\dot{\gamma }}^k$ for the components of the velocity, this becomes the geodesic equation

$${\ddot{x}}^k+{\Gamma }_{ij}^k{\dot{x}}^i{\dot{x}}^j=0,$$

where the Christoffel symbols of the connection are

$${\Gamma }_{ij}^k=\frac{1}{2}{g}^{k\ell }({\partial }_i{g}_{j\ell }+{\partial }_j{g}_{i\ell }-{\partial }_{\ell }{g}_{ij}).$$

The geodesic equation is a second-order ODE, so for any initial position $p\in M$ and initial velocity $v\in T_{p}M$ there is a unique geodesic ${\gamma }_v$ with ${\gamma }_v(0)=p$ and ${\dot{\gamma }}_v(0)=v$, defined on a maximal interval $[0,b_v)$.

The exponential map at $p$ is the map

$${\exp }_p:{\mathcal{D}}_p\subseteq T_{p}M⟶M,{\exp }_p(v)={\gamma }_v(1),$$

defined on the star-shaped domain ${\mathcal{D}}_p=\{v\in T_{p}M:{\gamma }_v\text{ is defined on }[0,1]\}$, and satisfying ${\exp }_p(tv)={\gamma }_v(t)$ for $t\in [0,1]$. Its differential at the origin is the identity under the canonical identification $T_0(T_{p}M)\cong T_{p}M$, so by the inverse function theorem ${\exp }_p$ is a diffeomorphism from a small ball about $0\in T_{p}M$ onto a neighbourhood of $p$.

The manifold is geodesically complete when every unit-speed geodesic is defined on all of $\mathbb{R}$, equivalently when ${\mathcal{D}}_p=T_{p}M$ for every $p$. The geodesic distance $d(p,q)$ is the infimum of lengths of piecewise-smooth curves from $p$ to $q$, and $(M,d)$ is a metric space. A point $q=\gamma (t_0)$ is conjugate to $p=\gamma (0)$ along a geodesic when there is a nonzero Jacobi field $J$ along $\gamma$ — a field solving $D_t^{2}J+R(J,\dot{\gamma })\dot{\gamma }=0$ — with $J(0)=0$ and $J(t_0)=0$, equivalently when $d({\exp }_p{)}_{t_0\dot{\gamma }(0)}$ is singular.

Worked coordinate computation

On the unit sphere with spherical coordinates $(\theta ,\varphi )$ (colatitude, longitude), the round metric is $g=d{\theta }^2+{\sin }^2\theta d{\varphi }^2$, so $g_{\theta \theta }=1$, $g_{\varphi \varphi }={\sin }^2\theta$. The only nonzero Christoffel symbols are

$${\Gamma }_{\varphi \varphi }^{\theta }=-\sin \theta \cos \theta ,{\Gamma }_{\theta \varphi }^{\varphi }={\Gamma }_{\varphi \theta }^{\varphi }=\cot \theta .$$

The geodesic equations read $\ddot{\theta }-\sin \theta \cos \theta {\dot{\varphi }}^2=0$ and $\ddot{\varphi }+2\cot \theta \dot{\theta }\dot{\varphi }=0$. The equator $\theta =\pi /2$ with constant $\dot{\varphi }$ solves both equations because $\cos (\pi /2)=0$ and $\cot (\pi /2)=0$, confirming the equator as a geodesic; meridians $\varphi =\text{const}$ with constant $\dot{\theta }$ solve both as well, confirming great circles through the poles.

Counterexamples to common slips

• Geodesics need not minimize globally. A great circle on the sphere minimizes only up to the antipode; past it, the shorter route goes the other way. Local minimization (Gauss lemma) is unconditional; global minimization requires the cut-locus analysis below.
• The exponential map need not be defined on all of $T_{p}M$. On a manifold with an edge or a point removed, a geodesic heading toward the missing point runs out of manifold and ${\exp }_p$ is only partially defined; completeness is exactly the condition ruling this out.
• Conjugate points are not the same as cut points. A cut point is where a geodesic first fails to minimize; a conjugate point is where infinitesimally nearby geodesics refocus. The first cut point along a geodesic occurs at or before the first conjugate point.

Key theorem with proof Intermediate+

Theorem (Hopf-Rinow). Let $(M,g)$ be a connected Riemannian manifold and let $p\in M$. The following are equivalent.

1. $M$ is geodesically complete: every geodesic extends to all of $\mathbb{R}$.
2. $(M,d)$ is complete as a metric space: every Cauchy sequence converges.
3. Every closed bounded subset of $M$ is compact (the Heine-Borel property).

Any one of these implies ${\exp }_p$ is defined on all of $T_{p}M$, and for every $q\in M$ there is a geodesic from $p$ to $q$ of length exactly $d(p,q)$ ^{[Hopf-Rinow 1931]}.

Proof. We prove the cycle (2)$\Rightarrow$(1)$\Rightarrow$(3)$\Rightarrow$(2) together with the minimizing-geodesic conclusion.

(2)$\Rightarrow$(1). Suppose $(M,d)$ is complete and let $\gamma :[0,b)\to M$ be a unit-speed geodesic with $b$ finite and maximal. Choose $t_i\nearrow b$. Because $d(\gamma (t_i),\gamma (t_j))\le |t_i-t_j|$, the sequence $\gamma (t_i)$ is Cauchy, hence converges to some $q\in M$. In a normal neighbourhood of $q$, the exponential map ${\exp }_q$ is a diffeomorphism from a ball in $T_{q}M$, so the geodesic $\gamma$ can be extended a little past $b$ by setting $\gamma (b+\epsilon )={\exp }_q(\epsilon w)$ for a suitable velocity $w$. This contradicts the maximality of $b$, so $b=\infty$.

Minimizing geodesics under (1). Assume $M$ is geodesically complete and fix $p\in M$, so ${\exp }_p$ is defined on all of $T_{p}M$. We show every $q\in M$ with $d(p,q)=d$ is joined to $p$ by a geodesic of length $d$. Choose $\delta >0$ small enough that ${\exp }_p$ is a diffeomorphism on $\overline{B(0,\delta )}$, and let $S_{\delta }={\exp }_p(\partial B(0,\delta ))$ be the geodesic sphere. Pick a sequence of curves from $p$ to $q$ of length tending to $d$; the first point of each curve to hit $S_{\delta }$ converges to some $p_1\in S_{\delta }$, and $d(p,p_1)=\delta$ by the Gauss lemma. Write $p_1={\exp }_p(\delta v)$ for a unit vector $v\in T_{p}M$ and let $\gamma (t)={\exp }_p(tv)$ be the radial geodesic. Define

$$T=\sup \{t\in [0,d]:d(p,\gamma (t))=t\}.$$

The set on the right is closed, so $d(p,\gamma (T))=T$. If $T<d$, apply the same sphere construction at $\gamma (T)$ with a small radius ${\delta }^{\prime }$: extend $\gamma$ to $\gamma (T+{\delta }^{\prime })$ and use the Gauss lemma to build a curve from $p$ to $\gamma (T+{\delta }^{\prime })$ of length $T+{\delta }^{\prime }$, giving $d(p,\gamma (T+{\delta }^{\prime }))=T+{\delta }^{\prime }$, contradicting the definition of $T$. Hence $T=d$, so $\gamma$ is minimizing on $[0,d]$ and $\gamma (d)$ achieves distance $d$ from $p$, forcing $\gamma (d)=q$.

(1)$\Rightarrow$(3). Under geodesic completeness ${\exp }_p$ is defined on all of $T_{p}M$, so every closed bounded set lies in $\overline{B(p,R)}={\exp }_p(\overline{B(0,R)})$ for some $R$. The ball $\overline{B(0,R)}\subseteq T_{p}M\cong {\mathbb{R}}^n$ is compact, and ${\exp }_p$ is continuous, so $\overline{B(p,R)}$ is compact as the continuous image of a compact set; closed subsets of it are compact.

(3)$\Rightarrow$(2). A Cauchy sequence in $M$ is bounded, hence contained in some closed ball, which is compact by hypothesis; a sequence in a compact set with a Cauchy subsequence converges, so $M$ is metrically complete. $\square$

Bridge. Hopf-Rinow builds toward 48.01.02 (the comparison theorems), where geodesic completeness is the standing hypothesis that lets Cartan-Hadamard promote the exponential map to a covering map and lets Bonnet-Myers run its index-form argument, and appears again in 03.09.10 (the index theorem), whose heat-kernel proof needs global geodesic flow so that wave and diffusion techniques apply on all of $M$. The foundational reason the theorem matters is that completeness converts a local ODE into a global minimizing principle: once every geodesic runs forever, the radial geodesic out of any base point reaches every other point at exactly the right distance, and this is exactly the mechanism that lets the Levi-Civita connection of 48.01.01 speak globally rather than infinitesimally; the central insight is that the Gauss lemma forces radial geodesics to hit distance spheres at right angles, so no shortcut around a radial arc can exist; putting these together, the bridge is that metric completeness, geodesic extendability, and the existence of minimizing geodesics are three faces of one fact, and the pattern generalises to Lorentzian geodesic completeness in 13.03.01, where the failure of an exact Hopf-Rinow analogue is precisely what drives the singularity theorems.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: none records that Mathlib's Riemannian layer, while it now carries smooth-manifold infrastructure, the Levi-Civita connection, and geodesics as auto-parallel curves, has not yet absorbed the global Hopf-Rinow machinery this unit is built around. The natural first formalisation target is the local statement that ${\exp }_p$ is a diffeomorphism near the origin (the inverse-function-theorem argument in Exercise 4), which sits just above the existing manifold API; the equivalence of metric and geodesic completeness, the Gauss-lemma proof of minimizing radial geodesics, and the conjugate-point criterion for the singularity of $d{\exp }_p$ await upstream work documented in the unit metadata gap note.

Advanced results Master

Hopf-Rinow settles existence of minimizing geodesics under completeness; the finer structure of how long a geodesic keeps minimizing and where the exponential map folds is the content of the index theory of geodesics.

The Gauss lemma in polar form states that on a normal neighbourhood of $p$, the pulled-back metric is ${\exp }_p^{\text{astg}}=d{r}^2+g_r$ where $g_r$ is a metric on the geodesic sphere of radius $r$. It is the single most-used tool in the subject: it makes radial geodesics orthogonal to geodesic spheres and is the reason radial arcs locally minimize (any competing curve has length bounded below by the radial length, with equality only along the ray). Combined with the triangle inequality it yields the local minimizing property of ${\exp }_p$.

The index form on a unit geodesic $\gamma :[0,L]\to M$ is the symmetric bilinear form on piecewise-smooth vector fields $V,W$ along $\gamma$ vanishing at the endpoints,

$$I(V,W)={\int }_0^L(⟨D_{t}V,D_{t}W⟩-⟨R(V,\dot{\gamma })\dot{\gamma },W⟩)dt.$$

It is the second variation of length: a geodesic minimizes on $[0,L]$ only if $I(V,V)\ge 0$ for all such $V$, and the connection between the sign of $I$ and conjugate points is governed by the Morse index theorem, which states that the index of $I$ (its maximal dimension of a subspace on which it is negative definite) equals the number of conjugate points along $\gamma$, counted with multiplicity. The first conjugate point is therefore where a geodesic first loses local minimality, though it may have stopped being globally minimizing earlier at its cut point.

The cut locus $\mathrm{Cut}(p)$ is the set of points where geodesics radiating from $p$ stop being globally minimizing; the injectivity radius $\mathrm{inj}(p)$ is the distance from $p$ to its cut locus. On the round sphere $\mathrm{Cut}(p)=\{-p\}$ and $\mathrm{inj}(p)=\pi$; on flat ${\mathbb{R}}^n$ the cut locus is empty and the injectivity radius is infinite. Klingenberg's formula relates the injectivity radius to the shorter of the conjugate radius and the length of the shortest closed geodesic, binding local index theory to global topology. The exponential map is a diffeomorphism on the open ball of radius $\mathrm{inj}(p)$, and the manifold decomposes as $M={\exp }_p(B_{\mathrm{inj}})\bigsqcup \mathrm{Cut}(p)$ — the bridge from the smooth structure of ${\exp }_p$ to the global geometry of $M$.

Synthesis. The geodesic–exponential–completeness trichotomy is the engine of global Riemannian geometry: the Hopf-Rinow theorem builds toward 48.01.02, where completeness licenses every comparison theorem, and appears again in 03.09.10, where global geodesic flow underwrites the heat-kernel proof of the index theorem. The foundational reason a local equation controls global distance is the Gauss lemma — radial geodesics cut distance spheres at right angles, so the radial arc is unbeatable — and this is exactly the input the index form needs to detect when a geodesic first fails to minimize; the central insight is that the derivative of the exponential map is a Jacobi field, tying conjugate points to the singularity of $d{\exp }_p$, a picture that is dual to the cut-locus description, where competing geodesics from $p$ and the closest winner set the injectivity radius; putting these together, the bridge is that completeness, distance-minimization, and compactness of closed balls are one theorem, and the pattern generalises to Finsler and sub-Riemannian geometry, where a Hopf-Rinow theorem survives because it is really the metric-space structure that carries the argument.

Full proof set Master

Proposition (Conjugate points and the index form). Let $\gamma :[0,L]\to M$ be a unit-speed geodesic. If $\gamma (t_0)$ is conjugate to $\gamma (0)$ along $\gamma$ for some $t_0\in (0,L)$, then $\gamma$ is not minimizing on $[0,L]$: there exists a piecewise-smooth curve from $\gamma (0)$ to $\gamma (L)$ of strictly smaller length.

Proof. Let $J$ be a nonzero Jacobi field along $\gamma$ with $J(0)=0$ and $J(t_0)=0$, which exists because $d({\exp }_p{)}_{t_0\dot{\gamma }(0)}$ is singular. Choose $\delta >0$ with $t_0+\delta <L$ and define a piecewise-smooth vector field $V$ along $\gamma$ vanishing at $0$ and $L$ by

$$V(t)=\{\begin{array}{ll}J(t),& 0\le t\le t_0,\\ \frac{t_0+\delta -t}{\delta }J(t_0){|}_{\text{taper}},& t_0<t\le t_0+\delta ,\\ 0,& t_0+\delta <t\le L,\end{array}$$

where the taper is a smooth cutoff agreeing with $J$ at $t_0$ and vanishing with its first derivative at $t_0+\delta$. Compute the index form on $[0,t_0]$. Since $J$ is a Jacobi field and $J(0)=J(t_0)=0$, integration by parts gives

$${\int }_0^{t_0}(|D_{t}J{|}^2-⟨R(J,\dot{\gamma })\dot{\gamma },J⟩)dt=⟨D_{t}J,J⟩{|}_0^{t_0}=0.$$

On the tapered interval $[t_0,t_0+\delta ]$, for $\delta$ small the curvature term is bounded and the field is supported in a short window; a direct estimate shows the contribution can be made negative by a small rescaling, because the boundary term $⟨D_{t}J(t_0^{+}),V(t_0)⟩$ inherited from the singularity of $d{\exp }_p$ is negative (the Jacobi field refocuses, so the family of geodesics is converging, lowering length). Summing, $I(V,V)<0$ for the chosen $V$.

A negative value of the index form means the second variation of length is negative in the direction $V$, so for small $\epsilon$ the perturbed curve ${\gamma }_{\epsilon }(t)={\exp }_{\gamma (t)}(\epsilon V(t))$ has length strictly less than $L$. Therefore $\gamma$ does not minimize on $[0,L]$. $\square$

Proposition (Metric completeness implies geodesic completeness). If $(M,g)$ is complete as a metric space under the geodesic distance, then every geodesic extends to all of $\mathbb{R}$.

Proof. Let $\gamma :[0,b)\to M$ be a unit-speed geodesic with $b$ finite and maximal. Choose a sequence $t_i\nearrow b$. Because $\gamma$ has unit speed, $d(\gamma (t_i),\gamma (t_j))\le |t_i-t_j|$, so $\{\gamma (t_i)\}$ is a Cauchy sequence in $(M,d)$. Metric completeness gives a limit $q={\lim }_i\gamma (t_i)\in M$. Choose $\epsilon >0$ small enough that ${\exp }_q$ is a diffeomorphism from $B(0,\epsilon )\subseteq T_{q}M$ onto a neighbourhood of $q$. For $i$ large, $\gamma (t_i)$ lies in this neighbourhood and its velocity $\dot{\gamma }(t_i)$ transported to $q$ gives a vector $w$ with ${\exp }_q((t-t_i)w)=\gamma (t)$ on $[t_i,t_i+{\epsilon }^{\prime })$; taking $i$ large extends $\gamma$ to $[0,b+{\epsilon }^{\prime })$, contradicting maximality of $b$. Hence $b=\infty$, and applying the same argument to $-\gamma$ extends to all of $\mathbb{R}$. $\square$

Connections Master

• Foundational Riemannian substrate 48.01.01. This unit presupposes the Levi-Civita connection, the curvature tensor, and the local theory of parallel transport developed in 48.01.01; the geodesic equation ${\nabla }_{\dot{\gamma }}\dot{\gamma }=0$ and the Christoffel symbols are read directly off that connection, so the foundational unit is the indispensable prerequisite for everything here.

• Comparison theorems 48.01.02. Geodesic completeness — established here as equivalent to metric completeness by Hopf-Rinow — is the standing hypothesis under which Cartan-Hadamard promotes the exponential map to a covering map and under which Bonnet-Myers and Bishop-Gromov run their index-form and Riccati-comparison arguments; the Jacobi fields and the index form introduced here are the same objects the comparison theorems weaponize.

• Index theory 03.09.10. Global geodesic flow, guaranteed by completeness, is what makes the heat-kernel and wave-equation approaches to the Atiyah-Singer index theorem work on all of $M$; the conjugate-point and index-form analysis here is the finite-dimensional shadow of the infinite-dimensional Morse theory on the loop space that underlies the analytic index.

• Lorentzian geometry and general relativity 13.03.01. The Hopf-Rinow equivalence breaks for pseudo-Riemannian metrics — a Lorentzian manifold can be geodesically complete without closed bounded sets being compact — and the search for the right substitute, culminating in the singularity theorems, drives global Lorentzian geometry; the geodesic focus here is the direct ancestor of that relativistic theory.

Historical & philosophical context Master

The study of geodesics as shortest curves on a surface begins with Gauss's 1827 Disquisitiones generales circa superficies curvas, where geodesics appear as the intrinsic analogue of straight lines and the Gauss lemma (in its original surface form) is used to derive the local geometry of geodesic coordinates ^{[Gauss 1827]}. Gauss's insistence that the geodesic and its properties be defined intrinsically — without reference to an ambient Euclidean space — set the conceptual frame for everything that followed. Riemann generalized this picture to manifolds of arbitrary dimension in his 1854 Habilitation lecture, where a smoothly varying inner product replaces the first fundamental form of a surface and geodesics become the straightest curves of the new geometry ^{[Riemann 1854]}.

The theorem that carries this unit's name came much later. Heinz Hopf and Willi Rinow proved in their 1931 paper Über den Begriff der vollständigen differentialgeometrischen Fläche that for a surface the metric completeness of the geodesic distance, the extendability of every geodesic, and the existence of a shortest geodesic between any two points are equivalent conditions, and that each implies the Heine-Borel property ^{[Hopf-Rinow 1931]}. The result crystallized a philosophy that has governed global Riemannian geometry ever since: local data — the metric, the connection, the geodesic equation — determines global behaviour once a completeness hypothesis is in place. The index theory of geodesics, developed by Morse in the 1920s and 1930s and later embedded by Bott, Smale, and Klingenberg into the topology of the loop space, turned the second variation of length into a bridge between the geometry of geodesics and the topology of the underlying manifold, the thread that runs from this unit forward into modern geometric analysis and the index theorems of 03.09.10.

Bibliography Master

@book{Lee1997,
  author = {Lee, Jeffrey M.},
  title = {Riemannian Manifolds: An Introduction to Curvature},
  publisher = {Springer GTM 176},
  year = {1997},
}

@book{doCarmo1992,
  author = {do Carmo, Manfredo},
  title = {Riemannian Geometry},
  publisher = {Birkh{\"a}user},
  year = {1992},
}

@book{CheegerEbin1975,
  author = {Cheeger, Jeff and Ebin, David G.},
  title = {Comparison Theorems in Riemannian Geometry},
  publisher = {North-Holland},
  year = {1975},
}

@book{Sakai1996,
  author = {Sakai, Takashi},
  title = {Riemannian Geometry},
  publisher = {American Mathematical Society, Translations of Mathematical Monographs 149},
  year = {1996},
}

@book{Klingenberg1982,
  author = {Klingenberg, Wilhelm},
  title = {Riemannian Geometry},
  publisher = {Walter de Gruyter},
  year = {1982},
}

@article{HopfRinow1931,
  author = {Hopf, Heinz and Rinow, Willi},
  title = {{\"U}ber den Begriff der vollst{\"a}ndigen differentialgeometrischen Fl{\"a}che},
  journal = {Commentarii Mathematici Helvetici},
  volume = {3},
  pages = {169--185},
  year = {1931},
}

@incollection{Gauss1827,
  author = {Gauss, Carl Friedrich},
  title = {Disquisitiones generales circa superficies curvas},
  year = {1827},
}

@incollection{Riemann1854,
  author = {Riemann, Bernhard},
  title = {{\"U}ber die Hypothesen, welche der Geometrie zu Grunde liegen},
  year = {1854},
}