03.02.32 · differential-geometry / manifolds

The Riemannian Hopf–Rinow theorem

shipped3 tiersLean: none

Anchor (Master): do Carmo Riemannian Geometry Ch. 7; Lee Introduction to Riemannian Manifolds Ch. 6; Hopf--Rinow 1931 Comment. Math. Helv. 3; Petersen Riemannian Geometry Ch. 5

Intuition Beginner

Picture a curved surface you can wander across — the round Earth, say. Two questions feel separate at first. Can you keep walking forever in any direction without ever reaching an edge or a hole? And: between any two towns, is there always a genuine shortest route you could pave? The Hopf–Rinow theorem says these two questions have the same answer. No edges to fall off, and you can always walk a straight shortest path.

The everyday version is this. On a globe there is no boundary; head out in any direction and you simply keep going. Because of that, every pair of cities really does have a shortest great-circle road between them. The "no edges" property and the "shortest road always exists" property travel together.

Now spoil the surface. Cut a single point out of a flat tabletop. Walking toward the hole, you march straight at the edge and then run out of floor — your journey cannot continue. And for two towns sitting on opposite sides of the hole, the straight road between them would pass through the missing point, so no actual shortest path exists on the surface that remains. Lose "no edges," and you lose "shortest paths always exist" at the same time.

Visual Beginner

Alt text: A three-panel figure. Left panel: a complete flat plane with two points joined by a straight segment labelled "shortest path exists." Middle panel: a sphere with two points joined by a great-circle arc, also labelled "shortest path exists," with the note "no edge, keep walking forever." Right panel: a flat disc with one interior point removed, and two points on opposite sides of the hole; the straight dashed route between them runs through the missing point, captioned "incomplete: no shortest path exists." The figure links walking forever to shortest paths being guaranteed.

Worked example Beginner

Take the flat plane and the two points $A = (0, 0)$ and $B = (6, 8)$ . The straight segment from $A$ to $B$ has length $6^{2} + 8^{2} = 36 + 64 = 100 = 10$ . No path from $A$ to $B$ can be shorter than $10$ , and the straight one achieves it. So a shortest path exists, and its length is the distance between the points. The plane has no edges and no holes, matching the theorem: complete surface, shortest path guaranteed.

Now remove the single point $(3, 4)$ from the plane and ask again for the shortest route from $A$ to $B$ . The point $(3, 4)$ sits exactly on the old straight segment, at its midpoint. That perfect route is now forbidden, because $(3, 4)$ is missing. You can get the length as close to $10$ as you like by detouring in a tiny arc around the gap, but you can never reach $10$ itself: every allowed path is strictly longer than $10$ , and there is no shortest one. The distance is still $10$ , but nothing achieves it.

What this tells us: punching out a single point destroys both "you can walk forever" (you stall at the hole) and "a shortest path exists" (the best route is forbidden). The two failures arrive together — which is the content of Hopf–Rinow.

Check your understanding Beginner

Formal definition Intermediate+

Let $(M, g)$ be a connected Riemannian manifold with Levi-Civita connection $\nabla$ 03.02.27. For a piecewise-smooth curve $γ : [a, b] \to M$ , its length is $$ L(\gamma) = \int_a^b |\dot\gamma(t)|_g , dt, \qquad |\dot\gamma|g = \sqrt{g(\dot\gamma, \dot\gamma)} . $$ The Riemannian distance between $p, q \in M$ is the infimum of lengths over piecewise-smooth curves joining them, $$ d(p, q) = \inf { L(\gamma) : \gamma(a) = p,\ \gamma(b) = q } . $$ Connectedness makes this infimum finite. The function $d$ is a metric, and the metric topology it induces coincides with the manifold topology of $M$ ; this is the standard fact that on a normal ball the radial geodesics realise distance (the Gauss lemma), proved in 03.02.27. A curve $γ$ is minimising if $L (γ) = d (γ (a), γ (b))$ . A geodesic is a curve with $\nabla{\dot\gamma}\dot\gamma = 0$; geodesics are locally minimising, and a minimising curve, parametrised by arc length, is a geodesic.

The exponential map $exp_{p} : U_{p} \subseteq T_{p} M \to M$ sends $v$ to $γ_{v} (1)$ , where $γ_{v}$ is the geodesic with $γ_{v} (0) = p$ and $\overset{γ}{˙}_{v} (0) = v$ ; its domain $U_{p}$ is the set of $v$ for which $γ_{v}$ exists up to parameter $1$ 03.02.27.

Definition (the completeness conditions). $(M, g)$ is

metrically complete if $(M, d)$ is a complete metric space (every Cauchy sequence converges);
geodesically complete if every maximal geodesic is defined on all of $R$ — equivalently, $exp_{p}$ is defined on all of $T_{p} M$ for every $p$ ;
geodesically complete at $p$ if $exp_{p}$ is defined on all of $T_{p} M$ for that single point; and $(M, d)$ has the Heine–Borel property if every closed and bounded subset is compact (equivalently, $(M, d)$ is a proper metric space).

Sign and parametrisation conventions. Geodesics are parametrised proportionally to arc length, so $∣ \overset{γ}{˙} ∣_{g}$ is constant; distances are non-negative and the metric is positive-definite, in contrast with the indefinite-signature Lorentzian setting of 03.02.17 where the "distance" is a Lorentzian time-separation that is maximised, not minimised. This sign difference is the entire source of the divergence between the two Hopf–Rinow theorems.

Counterexamples to common slips

Geodesic completeness is not local. A single geodesic running off the edge in finite parameter — for instance a straight line approaching the puncture in $R^{2} ∖ {0}$ — already breaks geodesic completeness, even though every small neighbourhood looks like complete Euclidean space. Completeness is a global condition.
Bounded does not mean compact without completeness. In the open unit disc with the flat metric, the closed set ${∣ x ∣ \leq 1 - \frac{1}{n} for all n}$ is closed and bounded but not compact, because the boundary circle is missing. The Heine–Borel property fails exactly when completeness fails.
A minimising geodesic need not be unique. Antipodal points on the sphere are joined by infinitely many minimising geodesics. Hopf–Rinow guarantees existence of a minimiser, not uniqueness; uniqueness is governed by the cut locus and conjugate points 03.02.19.

Key theorem with proof Intermediate+

Theorem (Hopf–Rinow). Let $(M, g)$ be a connected Riemannian manifold with distance $d$ . The following are equivalent.

$(M, d)$ is a complete metric space.
Closed bounded subsets of $(M, d)$ are compact (the Heine–Borel property).
$M$ is geodesically complete: every geodesic extends to all of $R$ .
For some point $p \in M$ , $exp_{p}$ is defined on all of $T_{p} M$ .

Moreover, any of these conditions implies:

Any two points $p, q \in M$ are joined by a minimising geodesic, i.e. a geodesic $γ$ with $L (γ) = d (p, q)$ .

The proof rests on one lemma, which already delivers conclusion (5) from the weak hypothesis (4).

Key lemma (minimising-geodesic existence from $exp_{p}$ ). Suppose $exp_{p}$ is defined on all of $T_{p} M$ . Then for every $q \in M$ there is a unit-speed geodesic $γ$ from $p$ to $q$ with $L (γ) = d (p, q)$ .

Proof of the lemma. Fix $q \neq = p$ and write $r = d (p, q)$ . Choose a geodesic ball $B_{δ} (p) = exp_{p} ({∣ v ∣ < δ})$ small enough that its boundary sphere $S_{δ} = exp_{p} ({∣ v ∣ = δ})$ is compact and every point of $\overline{B_{δ} (p)}$ is joined to $p$ by a radial minimising geodesic (a normal ball, 03.02.27). The continuous function $x \mapsto d (x, q)$ attains a minimum on the compact sphere $S_{δ}$ at some point $x_{0} = exp_{p} (δ v_{0})$ , with $∣ v_{0} ∣ = 1$ . Let $$ \gamma(s) = \exp_p(s, v_0), \qquad s \in [0, \infty), $$ which exists for all $s$ by hypothesis (4). The claim is $γ (r) = q$ , which gives a minimising geodesic of length $r$ .

Consider the set $$ A = { s \in [0, r] : d(\gamma(s), q) = r - s } . $$ The point $s = 0$ lies in $A$ since $d (p, q) = r$ . The set $A$ is closed by continuity of $d$ and $γ$ . It remains to show $A$ is open in $[0, r)$ in the sense that $s_{0} \in A$ , $s_{0} < r$ , forces $s_{0} + δ^{'} \in A$ for a small $δ^{'} > 0$ ; then $A = [0, r]$ and $s = r$ gives $d (γ (r), q) = 0$ , i.e. $γ (r) = q$ .

First handle $s_{0} = 0$ . Any curve from $p$ to $q$ must cross $S_{δ}$ , so $$ r = d(p, q) = \min_{x \in S_\delta} \big( d(p, x) + d(x, q) \big) = \delta + d(x_0, q), $$ hence $d (x_{0}, q) = r - δ$ . Since $x_{0} = γ (δ)$ , this is exactly $δ \in A$ .

Now suppose $s_{0} \in A$ with $0 < s_{0} < r$ . Around the point $p^{'} = γ (s_{0})$ choose a normal ball of radius $δ^{'} < r - s_{0}$ with boundary sphere $S_{δ^{'}}^{'}$ , and let $x_{0}^{'}$ minimise $d (\cdot, q)$ on $S_{δ^{'}}^{'}$ . As before, $$ d(p', q) = \delta' + d(x_0', q), \qquad\text{so}\qquad d(x_0', q) = d(p', q) - \delta' = (r - s_0) - \delta', $$ using $s_{0} \in A$ . Then the triangle inequality gives $$ d(p, x_0') \ge d(p, q) - d(x_0', q) = r - \big( (r - s_0) - \delta' \big) = s_0 + \delta'. $$ But there is a path of exactly this length from $p$ to $x_{0}^{'}$ : follow $γ$ from $p$ to $p^{'}$ (length $s_{0}$ , since $γ$ is unit-speed and minimising up to $s_{0}$ — this uses $d (p, p^{'}) = s_{0}$ , which holds because $γ$ realises $d$ on $A$ ) and then the radial geodesic from $p^{'}$ to $x_{0}^{'}$ (length $δ^{'}$ ). Its total length is $s_{0} + δ^{'} \leq d (p, x_{0}^{'})$ , forcing equality, so this concatenation is minimising. A minimising curve is a geodesic, and a geodesic with no corner is smooth; therefore the concatenation has no corner at $p^{'}$ , which means $x_{0}^{'} = γ (s_{0} + δ^{'})$ . Combining with $d (x_{0}^{'}, q) = r - (s_{0} + δ^{'})$ gives $s_{0} + δ^{'} \in A$ . This proves $A = [0, r]$ , so $γ (r) = q$ and $L (γ ∣_{[0, r]}) = r = d (p, q)$ . $■$

Proof of the equivalences. (3) $\Rightarrow$ (4) is immediate: geodesic completeness for all $p$ in particular gives it at one $p$ . (4) $\Rightarrow$ (2): fix $p$ with $exp_{p}$ defined on all of $T_{p} M$ . A bounded set $K$ lies in some metric ball $\overline{B_{R} (p)}$ . By the lemma every point of $\overline{B_{R} (p)}$ is $exp_{p} (v)$ for some $∣ v ∣ \leq R$ , so $\overline{B_{R} (p)} = exp_{p} ({∣ v ∣ \leq R})$ is the continuous image of a compact ball, hence compact. A closed subset of a compact set is compact, so $K$ (if also closed) is compact: the Heine–Borel property holds. (2) $\Rightarrow$ (1): a Cauchy sequence is bounded, so it lies in a compact set and has a convergent subsequence; a Cauchy sequence with a convergent subsequence converges. (1) $\Rightarrow$ (3): suppose $(M, d)$ is complete but some unit-speed geodesic $γ$ is defined only on a maximal interval $[0, T)$ with $T < \infty$ . For $t_{n} ↑ T$ the points $γ (t_{n})$ form a Cauchy sequence (since $d (γ (t_{n}), γ (t_{m})) \leq ∣ t_{n} - t_{m} ∣$ ), so they converge to some $x \in M$ . Taking a normal ball of uniform radius around $x$ and a totally normal neighbourhood, the geodesic flow extends past $T$ — geodesics starting near $x$ exist for a uniform time — contradicting maximality. Hence $T = \infty$ and $γ$ extends to all of $R$ . The lemma supplies (5) under (4), and (4) holds under any of (1)–(3). $■$

Bridge. This theorem builds toward every global result that needs a shortest path to exist, and it appears again in the diameter and sphere theorems of comparison geometry, where a minimising geodesic between diameter-realising points is the object on which curvature bounds act. The foundational reason the four completeness conditions coincide is that the exponential map at a single point, once globally defined, already parametrises the whole manifold by minimising geodesics — this is exactly the lemma, and the central insight is that metric completeness and geodesic completeness are two readings of the same Arzelà–Ascoli compactness. Putting these together, the positive-definite theorem generalises the Euclidean fact that straight segments minimise, and it is dual to the second-variation analysis of 03.02.19, which decides when such a minimiser stops being a local minimum past a conjugate point.

Exercises Intermediate+

Exercise (hard, short-answer). Give an example of a connected Riemannian manifold that is geodesically complete but on which some pair of points is joined by no unique minimiser, and explain why this is consistent with Hopf–Rinow.

Hint

Antipodes on a sphere; or a flat cylinder. Hopf–Rinow promises existence, not uniqueness.

Answer

Rubric. On $S^{2}$ the north and south poles are joined by every meridian, all of length $π$ , so there is a circle's worth of minimisers. The sphere is compact, hence complete, so Hopf–Rinow applies and guarantees a minimiser exists — it never claims uniqueness. Non-uniqueness is the cut-locus phenomenon: the antipode is the cut point of the north pole, and it is also conjugate to it 03.02.19. Full credit names a complete example, exhibits multiple minimisers, and states that the theorem asserts existence only.

Exercise (hard, short-answer). Prove the implication (4) $\Rightarrow$ (5) directly using only the key lemma, and explain at which single step positive-definiteness of $g$ is essential, so that the argument breaks in Lorentzian signature [03.02.17].

Hint

The lemma already is (4) $\Rightarrow$ (5). Locate where "the boundary sphere $S_{δ}$ is compact and $d (\cdot, q)$ attains a minimum on it" uses positive-definiteness.

Answer

Rubric. The lemma takes $p$ with $exp_{p}$ globally defined and produces a minimising geodesic to any $q$ , which is precisely (5). Positive-definiteness enters where the geodesic sphere $S_{δ} = exp_{p} ({∣ v ∣ = δ})$ is used as a compact set on which the continuous function $d (\cdot, q)$ attains its minimum: in Riemannian signature ${∣ v ∣ = δ}$ is a round sphere in $T_{p} M$ , compact. In Lorentzian signature the "unit" set is a hyperboloid — non-compact — and the time-separation is maximised rather than minimised, so the minimum-on-a-compact-sphere step has no analogue; this is why completeness alone does not yield a maximising causal geodesic and one must instead impose global hyperbolicity. Full credit identifies the compact-sphere/minimum step as the failure point.

Lean formalization Intermediate+

Mathlib has the exponential map, geodesics, and the abstract metric notions of completeness, properness, and the Heine–Borel property, but it does not yet build the Riemannian distance from the length functional or relate it to geodesic completeness, so the theorem is not formalisable end-to-end. The following is the statement-level shape one would target, in Lean-compatible pseudo-Lean. It does not compile against current Mathlib (lean_status: none).

-- Statement target (NOT compiling against current Mathlib):
variable {M : Type*} [RiemannianManifold M] [ConnectedSpace M]

-- Riemannian distance as an infimum of lengths of piecewise-smooth paths:
def riemDist (p q : M) : ℝ :=
  ⨅ γ ∈ piecewiseSmoothPaths p q, pathLength γ

def GeodesicallyComplete : Prop :=
  ∀ p : M, ∀ v : TangentSpace M p, ∀ t : ℝ, (expMap p (t • v)).IsDefined

-- Hopf–Rinow: the four completeness conditions are equivalent.
-- theorem hopf_rinow :
--   (CompleteSpace (withMetric M riemDist))
--     ↔ HeineBorel (withMetric M riemDist)
--     ↔ GeodesicallyComplete M
--     ↔ (∃ p, ∀ v, (expMap p v).IsDefined)

-- and any of them yields a minimising geodesic:
-- theorem exists_minimizing_geodesic (h : GeodesicallyComplete M) (p q : M) :
--   ∃ γ : Geodesic p q, pathLength γ = riemDist p q

the Mathlib gap analysis enumerates the missing primitives: the length-infimum distance, its agreement with the manifold topology via the Gauss lemma, the equivalence of the four completeness conditions, and the minimising-geodesic existence statement together with its Arzelà–Ascoli compactness input.

Advanced results Master

Cartan–Hadamard. When $(M, g)$ is complete, simply connected, and of non-positive sectional curvature, the global $exp_{p} : T_{p} M \to M$ promoted by Hopf–Rinow is in addition a diffeomorphism ^{[do Carmo Ch. 7]}. The completeness half is supplied here — $exp_{p}$ is defined on all of $T_{p} M$ ; non-positive curvature supplies the rest by the infinitesimal argument of 03.02.19, where the index form is positive definite and there are no conjugate points, so $exp_{p}$ is a local diffeomorphism everywhere, and simple connectedness upgrades this to a global one. Hopf–Rinow is the input that turns a curvature hypothesis into a statement about the whole manifold.

Bonnet–Myers and compactness from curvature. If $(M, g)$ is complete and its Ricci curvature satisfies $Ric \geq (n - 1) K g$ with $K > 0$ , then $M$ is compact with diameter at most $π / K$ ^{[do Carmo Ch. 9]}. The mechanism: completeness gives, for any two points, a minimising geodesic; along it the second-variation/index-form estimate forces a conjugate point before length $π / K$ 03.02.19, so no minimising geodesic can be longer, bounding the diameter; a closed bounded manifold of finite diameter is then compact by the Heine–Borel half of Hopf–Rinow. The theorem is what licenses "take the minimising geodesic between far-apart points" — the step that has no Lorentzian analogue without global hyperbolicity.

The role in comparison geometry. Across Rauch, Toponogov, and the sphere theorem, the recurring move is to compare a minimising geodesic in $M$ with a model geodesic in a constant-curvature space. The existence of that minimiser, on which all triangle-comparison estimates act, is a Hopf–Rinow output. Toponogov's globalisation of the local angle-comparison inequality, in particular, is stated for complete manifolds precisely so that the connecting geodesics of a comparison triangle are guaranteed to exist ^{[Petersen Ch. 5]}.

Metric-space completeness without smoothness. The equivalence (1) $\Leftrightarrow$ (2) is purely metric: in any length space, local compactness plus completeness yields the Heine–Borel property and geodesic existence. This is the Hopf–Rinow–Cohn-Vossen theorem, the form used in Alexandrov geometry and on metric measure spaces, where there is no exponential map and the minimising-geodesic continuation is run directly in the metric via midpoint limits ^{[Lee Ch. 6]}.

Synthesis. Hopf–Rinow is the foundational reason the global theory of Riemannian manifolds is possible at all: it is exactly the statement that the four faces of completeness — metric, Heine–Borel, geodesic, and single-point exponential — are one condition, and that this condition is the bridge from local geodesic existence to global minimiser existence. Putting these together, the theorem generalises the Euclidean straight-line minimiser to arbitrary complete curvature, it is dual to the conjugate-point obstruction of 03.02.19 that limits how far a minimiser can run, and the central insight is that the same Arzelà–Ascoli compactness underlies both the metric-completeness reading and the geodesic-completeness reading. This is exactly why every comparison theorem — Cartan–Hadamard, Bonnet–Myers, Toponogov, the sphere theorem — opens with the hypothesis "complete," and why its Lorentzian cousin 03.02.17 must replace completeness with global hyperbolicity: the compact-sphere minimisation step that drives the proof has no indefinite-signature analogue.

Full proof set Master

Proposition (the four completeness conditions are equivalent). For a connected Riemannian manifold, conditions (1)–(4) of the theorem are equivalent.

Proof. (3) $\Rightarrow$ (4) is the restriction of an all-points statement to one point. For (4) $\Rightarrow$ (2): with $exp_{p}$ globally defined, the key lemma gives $\overline{B_{R} (p)} = exp_{p} ({∣ v ∣ \leq R})$ for every $R$ , the continuous image of a compact Euclidean ball, hence compact; a closed bounded set sits inside some such ball and is a closed subset of a compact set, hence compact. For (2) $\Rightarrow$ (1): a Cauchy sequence is bounded, lies in a compact closed ball, has a convergent subsequence, and a Cauchy sequence with a convergent subsequence converges. For (1) $\Rightarrow$ (3): a unit-speed geodesic on a maximal $[0, T)$ with $T < \infty$ has $γ (t_{n})$ Cauchy for $t_{n} ↑ T$ (as $d (γ (t_{n}), γ (t_{m})) \leq ∣ t_{n} - t_{m} ∣$ ), so $γ (t_{n}) \to x$ ; a totally normal neighbourhood of $x$ provides a uniform existence time for geodesics starting near $x$ , extending $γ$ past $T$ and contradicting maximality. $■$

Proposition (existence of a minimising geodesic). If $(M, g)$ satisfies any of (1)–(4), then any two points are joined by a minimising geodesic.

Proof. By the equivalence, (4) holds at every point; in particular pick the source point $p$ and apply the key lemma, which constructs a unit-speed geodesic $γ$ from $p$ to the target $q$ with $L (γ) = d (p, q)$ . The lemma's engine is the closed-and-open set $A = {s : d (γ (s), q) = r - s}$ inside $[0, r]$ ; closedness is continuity, and openness is the normal-ball minimisation $d (p^{'}, q) = δ^{'} + d (x_{0}^{'}, q)$ together with the no-corner regularity of a length-minimising concatenation, which forces $x_{0}^{'} = γ (s_{0} + δ^{'})$ . Connectedness of $[0, r]$ gives $A = [0, r]$ , so $γ (r) = q$ . $■$

Proposition (metric topology equals manifold topology). The topology induced on $M$ by the distance $d$ coincides with the manifold topology.

Proof. On a normal ball $B_{δ} (p) = exp_{p} ({∣ v ∣ < δ})$ the Gauss lemma 03.02.27 gives $d (p, exp_{p} (v)) = ∣ v ∣$ for $∣ v ∣ < δ$ , so the metric ball of radius $ρ < δ$ about $p$ is exactly $exp_{p} ({∣ v ∣ < ρ})$ , an open manifold neighbourhood, and conversely every manifold neighbourhood of $p$ contains such a metric ball. Matching bases of neighbourhoods at every point identifies the two topologies. $■$

The Cartan–Hadamard global-diffeomorphism statement, the Bonnet–Myers diameter bound, and the Hopf–Rinow–Cohn-Vossen length-space generalisation are stated above without full proof — see do Carmo ^{[do Carmo Ch. 7]} for Cartan–Hadamard and Ch. 9 for Bonnet–Myers, and Lee ^{[Lee Ch. 6]} for the metric-space version.

Connections Master

Synge's theorem and positively curved manifolds 03.02.35. Synge's theorem — that a compact orientable even-dimensional manifold of positive sectional curvature is simply connected — runs its second-variation argument on a minimising closed geodesic in a free homotopy class, and the existence of that minimiser in each class is a direct consequence of the completeness packaged here. Hopf–Rinow is the standing hypothesis that makes the Synge variational comparison have an object to vary.
Levi-Civita connection, exponential map, geodesics 03.02.27. This unit consumes the exponential map, the normal balls, and the Gauss lemma built there; the entire proof is run on the geodesic spheres $exp_{p} ({∣ v ∣ = δ})$ and the global flow of geodesics. The local statement "geodesics are locally minimising" is upgraded here to the global "minimisers exist between any two points."
Geodesics and parallel transport 13.02.02. The geodesic equation $\nabla_{\overset{γ}{˙}} \overset{γ}{˙} = 0$ and the constant-speed parametrisation come from there; the distance function whose completeness this theorem characterises is the infimum of lengths of the curves that unit studies, and the Cauchy-sequence-along-a-geodesic argument uses the unit-speed bound $d (γ (s), γ (t)) \leq ∣ s - t ∣$ .
Jacobi fields and conjugate points 03.02.19. Hopf–Rinow guarantees a minimiser exists; the index theory there decides how long it stays minimising — a geodesic ceases to minimise past its first conjugate point. The two results are complementary: existence of minimisers versus the obstruction to remaining one, and together they drive Bonnet–Myers.
Lorentzian Hopf–Rinow and global hyperbolicity 03.02.17. The contrast is structural. In indefinite signature the geodesic "spheres" are non-compact hyperboloids and the separation is maximised, so the compact-sphere minimisation step fails and completeness no longer yields a maximising causal geodesic; global hyperbolicity is the correct replacement hypothesis, and the Avez–Seifert theorem is the correct replacement conclusion. This unit is the positive-definite original of which that unit is the indefinite-signature departure.

Historical & philosophical context Master

The theorem was proved by Heinz Hopf and Willi Rinow in their 1931 paper "Über den Begriff der vollständigen differentialgeometrischen Fläche" (Commentarii Mathematici Helvetici 3, 209–225), which isolated the notion of a complete surface and established the equivalence of metric and geodesic completeness with the existence of shortest joins ^{[Hopf-Rinow 1931]}. The extension from surfaces to manifolds of arbitrary dimension, and the recognition that the metric-space half of the argument requires no smooth structure, is due to Stefan Cohn-Vossen, whose 1935 work placed the result in the setting of inner-metric spaces — the form now standard in Alexandrov geometry and on metric measure spaces.

The conceptual contribution was to make "complete" a single notion with several faces. Before Hopf–Rinow, "can you extend every geodesic" and "is the metric space complete" and "do shortest paths exist" were distinct questions; the theorem collapsed them. This is the reason the global theory of Riemannian manifolds — comparison geometry, the Cartan–Hadamard and Bonnet–Myers theorems, the structure theory of nonnegatively curved spaces — uniformly takes completeness as its standing hypothesis. The indefinite-signature failure of the equivalence, isolated by relativists in the 1950s–70s and codified in the global hyperbolicity of Leray, Geroch, and others, marks the precise boundary of the Riemannian intuition ^{[do Carmo Ch. 7]}.

Bibliography Master

@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   = {209--225},
  year    = {1931}
}

@article{CohnVossen1935,
  author  = {Cohn-Vossen, Stefan},
  title   = {Existenz k{\"u}rzester Wege},
  journal = {Compositio Mathematica},
  volume  = {3},
  pages   = {441--452},
  year    = {1936}
}

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

@book{Lee2018,
  author    = {Lee, John M.},
  title     = {Introduction to Riemannian Manifolds},
  edition   = {2nd},
  series    = {Graduate Texts in Mathematics},
  number    = {176},
  publisher = {Springer},
  year      = {2018}
}

@book{Petersen2016,
  author    = {Petersen, Peter},
  title     = {Riemannian Geometry},
  edition   = {3rd},
  series    = {Graduate Texts in Mathematics},
  number    = {171},
  publisher = {Springer},
  year      = {2016}
}

Prerequisites

03.02.27
13.02.02

Tier anchors

beginner: Needham Visual Differential Geometry (completeness and shortest paths, informal); do Carmo Riemannian Geometry Ch. 7 informal
intermediate: do Carmo Riemannian Geometry Ch. 7 (the Hopf--Rinow theorem); Lee Introduction to Riemannian Manifolds Ch. 6
master: do Carmo Riemannian Geometry Ch. 7; Lee Introduction to Riemannian Manifolds Ch. 6; Hopf--Rinow 1931 Comment. Math. Helv. 3; Petersen Riemannian Geometry Ch. 5

References

do Carmo — Riemannian Geometry (1992) · Ch. 7 (geodesics, completeness, the Hopf--Rinow and Hadamard theorems)
Lee — Introduction to Riemannian Manifolds (2nd ed., 2018) · Ch. 6 (geodesics and distance; the Hopf--Rinow theorem)
Hopf, H. and Rinow, W. — Über den Begriff der vollständigen differentialgeometrischen Fläche (1931) · Commentarii Mathematici Helvetici 3, pp. 209--225
Petersen — Riemannian Geometry (3rd ed., 2016) · Ch. 5 (geodesics and distance; metric completeness)

Estimated time

beginner: 16m
intermediate: 40m
master: 75m