The Cartan–Hadamard Theorem: Nonpositive Curvature and the Exponential Map

Precise statement and terminology

Let (M, g) be a connected, geodesically complete Riemannian manifold of dimension n with sectional curvature K(σ) ≤ 0 for every 2-plane σ ⊂ T_pM at every point p. Fix a base point p and let exp_p : T_pM → M be the Riemannian exponential map, v ↦ γ_v(1), where γ_v is the geodesic with γ_v(0) = p, γ̇_v(0) = v. Completeness guarantees exp_p is defined on the entire tangent space T_pM ≅ ℝⁿ.

Cartan–Hadamard. Under these hypotheses, exp_p is a smooth covering map. If in addition M is simply connected, exp_p is a diffeomorphism from T_pM onto M. Consequently M is diffeomorphic to ℝⁿ, is contractible, has no conjugate points, and any two points are joined by a unique minimizing geodesic. A complete simply connected manifold with K ≤ 0 is called a Hadamard manifold. Note the hypothesis is on sectional curvature, not merely Ricci or scalar curvature — the sign of every sectional curvature is what controls geodesic spreading.

The intuition: geodesics that never refocus

Picture a family of geodesics all leaving p in nearly the same direction. On a sphere (K > 0) they curve back and reconverge at the antipode — that reconvergence point is a conjugate point, and there the exponential map folds over itself, ruining injectivity. Nonpositive curvature does the opposite: neighboring geodesics diverge at least as fast as in flat space, so they never refocus. No conjugate points ever appear.

Because rays from the origin in T_pM spread apart when pushed into M, the map exp_p can only stretch distances, never compress them below the Euclidean rate. It is like unrolling the manifold onto its tangent plane without any pleats or creases. The tangent space becomes a global chart: every point of M is exp_p(v) for some v, and the straight ray t ↦ tv maps to the geodesic reaching that point. Curvature ≤ 0 is precisely the condition that makes 'shoot a geodesic in a straight direction' a globally coherent coordinate system.

Key idea of the proof: Jacobi fields and no conjugate points

The engine is the second-order behavior of geodesic variations, encoded by Jacobi fields. A Jacobi field J along a geodesic γ satisfies J″ + R(J, γ̇)γ̇ = 0, where R is the curvature tensor and ″ is covariant differentiation along γ. Consider f(t) = ‖J(t)‖². A direct computation gives f″(t) = 2‖J′‖² − 2⟨R(J, γ̇)γ̇, J⟩. The curvature term ⟨R(J, γ̇)γ̇, J⟩ equals K times an area factor, so K ≤ 0 makes it ≤ 0, hence f″ ≥ 0: the squared length of a Jacobi field is convex. A Jacobi field vanishing at t = 0 with J′(0) ≠ 0 therefore has ‖J(t)‖ strictly increasing and never returns to zero. No nontrivial Jacobi field vanishes at two points, so there are no conjugate points.

Since d(exp_p)_v is singular exactly at conjugate points, exp_p is a local diffeomorphism everywhere. Now pull g back to T_pM: exp_p becomes a local isometry from a complete manifold, and a standard lemma (a local isometry from a complete space is a covering map) finishes it. Simple connectivity then forces the covering to be trivial — a global diffeomorphism.

Canonical example: the hyperbolic plane and SPD matrices

The hyperbolic plane ℍ² has constant curvature K = −1 and is complete and simply connected, so Cartan–Hadamard applies: ℍ² ≅ ℝ². In the disk model, exp_p sends each tangent ray to a geodesic (a diameter or a circular arc meeting the boundary orthogonally), and every point is reached exactly once. Along a unit-speed geodesic, a Jacobi field with J(0) = 0 grows like sinh(t) rather than the flat t or the spherical sin(t) — exponential spreading, the hallmark of K < 0.

A richer example is P(n), the space of n×n symmetric positive-definite matrices with the affine-invariant metric ⟨X, Y⟩_A = tr(A⁻¹X A⁻¹Y). This is a Hadamard manifold of dimension n(n+1)/2 with K ≤ 0, and exp_I(X) = matrix exponential of X. The theorem guarantees a unique geodesic between any two SPD matrices A, B — namely A^{1/2}(A^{−1/2}BA^{−1/2})^t A^{1/2} — which underlies the geometric mean of matrices used in diffusion-tensor imaging and covariance averaging.

Why each hypothesis is essential — and the connections

Completeness is not optional. Delete the origin from ℝ² (still K = 0, simply connected): exp_p no longer reaches the deleted point and is not surjective; the space is not ℝ². Completeness is what makes exp_p defined on all of T_pM and turns the local isometry into an honest covering map via Hopf–Rinow.

K ≤ 0 is essential: on Sⁿ (K = 1 > 0) conjugate points appear at distance π, exp_p folds, and the manifold is compact — the opposite of ℝⁿ. Simple connectivity only affects the last step: without it you still get a covering ℝⁿ → M, so the universal cover is ℝⁿ. That is exactly how flat tori ℝⁿ/ℤⁿ and closed hyperbolic surfaces of genus ≥ 2 arise — quotients of ℝⁿ by discrete groups of isometries. This links Cartan–Hadamard to the theory of aspherical manifolds (πₖ = 0 for k ≥ 2), Gromov's CAT(0) and δ-hyperbolic spaces, and the Preissmann and Byers rigidity theorems on their fundamental groups.

Applications and significance

Cartan–Hadamard is a keystone of global Riemannian geometry: it is the prototype of a curvature ⇒ topology theorem, showing a purely local sign condition forces a rigid global structure. It certifies that nonpositively curved manifolds are aspherical, making their homotopy type completely determined by the fundamental group — the starting point for the Borel conjecture and geometric group theory. The metric-space generalization (Cartan–Hadamard for CAT(0) and complete simply connected geodesic spaces, due to Gromov, Alexandrov, and Bruhat–Tits) is the foundation of the theory of buildings and of nonpositively curved cube complexes used by Agol and Wise.

Practically, the unique-geodesic and global-convexity consequences give well-posed optimization: on a Hadamard manifold the squared distance is geodesically convex, so gradient descent, the Riemannian mean (Karcher/Fréchet mean), and manifold-valued statistics have unique minimizers. This drives algorithms in computer vision, medical imaging (DTI), robotics, and machine learning on the SPD and hyperbolic-embedding manifolds.

How dropping each hypothesis changes the conclusion of Cartan–Hadamard

Hypotheses on (M, g)	exp_p behavior	Global conclusion
Complete, simply connected, K ≤ 0	Diffeomorphism T_pM → M	M ≅ ℝⁿ; unique geodesic between any two points
Complete, K ≤ 0 (not simply connected)	Smooth covering map, not injective	Universal cover ≅ ℝⁿ; e.g. flat torus, hyperbolic surfaces of genus ≥ 2
Complete, simply connected, K > 0 somewhere	May have conjugate points; not a local diffeo everywhere	Fails: sphere Sⁿ (K = 1) has conjugate points, is compact
Simply connected, K ≤ 0, NOT complete	Defined only on a subset of T_pM	Fails: open disk in ℝ² missing a point is not covered by exp
Complete, simply connected, K ≤ 0, infinite-dim Hilbert manifold	Local diffeo, no conjugate points	Still a diffeomorphism (McAlpin/Grossman); Hopf–Rinow can fail