Differential geometry
The Hopf-Rinow Theorem: When Geodesics Reach Everywhere
On a connected Riemannian manifold, five superficially different notions of "no holes" all collapse into one: if a manifold is geodesically complete — meaning every geodesic extends for all time — then it is metrically complete, every closed bounded set is compact, and, crucially, any two points are joined by a length-minimizing geodesic. The Hopf-Rinow theorem (Heinz Hopf and Willi Rinow, 1931) makes this equivalence precise.
Concretely: let (M, g) be a connected Riemannian manifold and d its induced distance. Then the following are equivalent — (i) (M, d) is a complete metric space; (ii) M is geodesically complete (the exponential map exp_p is defined on all of T_pM for some, hence every, p); (iii) closed and bounded ⟹ compact (the Heine-Borel property). Any one of these implies the existence-of-minimizing-geodesics conclusion. It is the theorem that lets you say the word "shortest path" and know it exists.
- FieldRiemannian geometry / differential geometry
- First provedHeinz Hopf & Willi Rinow, 1931
- Key hypothesisM connected; geodesic OR metric completeness
- StatementGeodesic ⇔ metric completeness ⇔ Heine-Borel; each ⟹ minimizing geodesics exist
- Proof techniqueContinuity/connectedness argument + Arzelà-Ascoli compactness
- Generalizes toLength spaces (Cohn-Vossen 1935; Gromov)
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
What the theorem claims, precisely
Let (M, g) be a connected Riemannian manifold, with distance d(p, q) = inf over piecewise-C¹ paths γ from p to q of the length L(γ) = ∫ ‖γ′(t)‖_g dt. Fix a point p ∈ M. The Hopf-Rinow theorem asserts the equivalence of:
- (i) The metric space (M, d) is complete.
- (ii) M is geodesically complete: the exponential map exp_p is defined on all of the tangent space T_pM (equivalently every geodesic γ: [0, a) → M extends to ℝ).
- (iii) Closed and bounded subsets of M are compact (the Heine-Borel property).
Moreover, each of these conditions implies the Hopf-Rinow conclusion: for every q ∈ M there is a minimizing geodesic γ from p to q with L(γ) = d(p, q). Notably, geodesic completeness at a single point p already forces all of (i)-(iii) and the global minimizer property. Connectedness is essential merely so that d is finite and M is a metric space in the usual sense.
The picture: no edges to fall off
Think of a geodesic as an ant walking straight ahead, never turning. Geodesic completeness says the ant never runs off an edge: it can walk forever in any direction. The claim is that this local-looking condition — 'straight-ahead trajectories never terminate' — is secretly the same as the global topological statement 'the space has no missing boundary points' (metric completeness) and the compactness statement 'bounded regions are honestly finite' (Heine-Borel).
The minimizing-geodesic conclusion is the payoff. On the sphere S², the shortest route from the north pole to any point is an arc of a great circle — a geodesic — and it always exists. On the flat plane ℝ², it is the straight segment. The theorem promises that on any complete manifold this 'straightest = shortest, and it exists' phenomenon holds globally, no matter how curved or twisted M is. Without completeness, the destination might sit just past a hole, and no straight path reaches it.
Key idea of the proof: grow the reachable radius
The heart is proving that geodesic completeness at p ⟹ every q is joined to p by a minimizer. Set r = d(p, q). Take a small geodesic sphere S(p, δ) = ∂B(p, δ) around p; it is compact, so the continuous function x ↦ d(x, q) attains its minimum at some point x₀ = exp_p(δv), ‖v‖ = 1. Let γ(t) = exp_p(tv), the unit-speed geodesic through x₀ — this exists for all t precisely by geodesic completeness.
Now consider the set A = { t ∈ [0, r] : d(γ(t), q) = r − t }, i.e. 'γ is still on-track to minimize'. One shows A is nonempty, closed, and — via the first variation of arc length and the triangle inequality applied at a fresh small sphere around γ(t) — open in [0, r]. By connectedness of [0, r], A = [0, r], so d(γ(r), q) = 0, meaning γ(r) = q. Thus γ|_[0,r] is a geodesic of length exactly r = d(p, q): a minimizer. The remaining equivalences (i)⇔(ii)⇔(iii) follow, using Arzelà-Ascoli to extract convergent subsequences of unit-speed minimizers and the completeness of ℝⁿ.
Worked example: the sphere, the plane, and the punctured plane
Sphere S²(R). Compact, hence metrically complete, hence geodesically complete by Hopf-Rinow. Every geodesic (a great circle) extends forever — it just wraps around periodically with period 2πR. Between north and south poles there are infinitely many minimizers (any meridian), each of length πR: Hopf-Rinow guarantees existence, not uniqueness.
Plane ℝ². Complete; the unique minimizer between p and q is the straight segment, exp_p(v) = p + v defined on all of T_pℝ² = ℝ².
Punctured plane ℝ² ∖ {0}. Now completeness fails. The geodesic starting at (1, 0) heading toward the origin reaches the missing point at t = 1 and cannot be extended — geodesically incomplete. Consequently no minimizing geodesic joins (−1, 0) to (1, 0): the infimum of lengths is 2, but every path must detour around the hole, and 2 is not attained. This single example shows all three implications and the minimizer conclusion collapsing together the moment completeness breaks.
Why the hypotheses matter, and where it fails
Completeness is non-negotiable. The open unit disk D with the flat metric is geodesically incomplete: radial geodesics hit the boundary in finite time. (It is a clean illustration of incompleteness only — the disk is convex, so between any two interior points the straight segment still lies inside D and is a minimizer; minimizers do not fail to exist there.) To see the minimizer conclusion genuinely break, one needs a non-convex incomplete manifold: in the punctured plane ℝ² ∖ {0}, the distance from (−1, 0) to (1, 0) is 2, yet every path must detour around the hole and the value 2 is not attained, so no minimizing geodesic exists. Drop completeness and the whole conclusion can evaporate, as that example shows.
Finite dimensionality matters. Hopf-Rinow fails in infinite dimensions. On an infinite-dimensional Hilbert manifold one can have a metrically complete, geodesically complete manifold on which exp_p is not surjective and some pairs of points have no minimizing geodesic — Grossman and McAlpin gave explicit examples. The Arzelà-Ascoli compactness step, which needs local compactness (finite dimension), is exactly what breaks.
The result also refuses to hold verbatim for pseudo-Riemannian (Lorentzian) metrics: there d is not a genuine metric, and geodesic completeness no longer implies the existence of maximizing timelike geodesics — a fact central to singularity theorems in general relativity.
Applications and significance
Hopf-Rinow is the license that makes global Riemannian geometry possible: nearly every comparison and rigidity theorem quietly assumes it. The Cartan-Hadamard theorem (complete, simply connected, nonpositive curvature ⟹ diffeomorphic to ℝⁿ via exp_p) needs geodesic completeness to know exp_p is defined everywhere. Bonnet-Myers (Ricci ≥ (n−1)k > 0 ⟹ M compact with diam ≤ π/√k) uses Hopf-Rinow to promote 'bounded' to 'compact'. The Cheeger-Gromoll splitting theorem, sphere theorems, and Toponogov comparison all rest on the guaranteed existence of minimizers.
Beyond geometry: geodesic completeness underlies well-posedness of geodesic flows, optimal-transport (Wasserstein) geometry, and shortest-path questions in shape analysis and robotics. Cohn-Vossen (1935) and later Gromov extended the theorem to locally compact length spaces, where 'complete + locally compact ⟹ geodesic + Heine-Borel + minimizers exist' — the foundation of modern metric geometry and CAT(κ) spaces.
| Condition | Precise meaning | Guarantees minimizing geodesic between ALL pairs? |
|---|---|---|
| Metric completeness | (M, d) complete: every Cauchy sequence converges | Yes |
| Geodesic completeness | exp_p defined on all of T_pM for one (⇔ every) p | Yes |
| Heine-Borel property | Every closed, bounded subset of M is compact | Yes |
| Compactness of M | M is a compact metric space | Yes (compact ⟹ complete; strictly stronger) |
| exp_p surjective at one p | exp_p: T_pM → M is onto for a single p | Only from that p — not between arbitrary pairs |
| None of the above | e.g. the punctured plane ℝ² ∖ {0} with the flat metric | No — minimizers can fail to exist |
Frequently asked questions
Why is completeness necessary — can't I just take shortest paths anyway?
Without completeness, the infimum of path lengths between two points may not be attained. In the punctured plane ℝ² ∖ {0}, the distance from (−1,0) to (1,0) is 2, but every actual path must skirt the missing origin, so no path achieves length exactly 2. Completeness is precisely the condition ruling out these 'holes just past the destination'.
Does geodesic completeness at one point suffice, or do I need it at every point?
One point suffices. If exp_p is defined on all of T_pM for a single p, the theorem forces metric completeness of the whole manifold, from which geodesic completeness at every other point follows. This asymmetric-looking strengthening is one of the theorem's most useful features in practice.
Does Hopf-Rinow hold in infinite dimensions?
No. On infinite-dimensional Hilbert manifolds, metric and geodesic completeness can both hold while exp_p fails to be surjective and some pairs of points admit no minimizing geodesic (Grossman; McAlpin gave explicit counterexamples). The failure traces to the loss of local compactness — the Arzelà-Ascoli step in the proof requires finite dimensionality.
Does the theorem guarantee a UNIQUE minimizing geodesic?
No — only existence. On the sphere S², the north and south poles are joined by infinitely many minimizing meridians. Uniqueness holds only up to the cut locus: for q not in the cut locus of p, the minimizer is unique. Hopf-Rinow is purely an existence statement.
How does it relate to the Heine-Borel theorem for ℝⁿ?
It generalizes it. In ℝⁿ, 'closed and bounded ⟺ compact' is the classical Heine-Borel theorem. Hopf-Rinow shows this property holds on any complete connected Riemannian manifold, and is in fact equivalent there to metric and geodesic completeness. So ℝⁿ is the flat special case.
Why doesn't Hopf-Rinow work in Lorentzian (relativistic) geometry?
In pseudo-Riemannian signature the metric is indefinite, so d is not a genuine distance function and the length functional is not bounded below on causal curves. Geodesic completeness no longer implies existence of maximizing geodesics between causally related points. The breakdown is deliberate and important: Lorentzian geodesic incompleteness is exactly how Penrose-Hawking singularity theorems detect black holes and the Big Bang.