The sign that forbids reconvergence
Guide 3 of this rung lived under positive curvature, where Bonnet-Myers pulled geodesics back together and capped the diameter. Now flip the sign and watch the opposite world unfold. We assume the sectional curvature K is nonpositive everywhere — K <= 0 on every 2-plane in every tangent space — and on a complete manifold. The slogan of this entire guide is one sentence: nonpositive curvature means geodesics spread apart and never reconverge. Everything below is that single fact, dressed in increasingly powerful clothes.
Recall the engine from guide 2: a Jacobi field J along a geodesic gamma measures how a nearby parallel geodesic drifts away, and it obeys the Jacobi equation J'' + R(J, gamma')gamma' = 0. Take the norm-squared and differentiate twice. The curvature term feeds in a sectional-curvature factor, and when K <= 0 you get (|J|^2)'' >= 0 in a precise sense — the separation between neighboring geodesics is a convex function of arc length. A convex function that starts at zero and has positive initial slope can only grow. So once two geodesics from a common point begin to separate, they keep separating: there is no way back.
No conjugate points, so the exponential map unrolls everything
A conjugate point is where a nontrivial Jacobi field vanishing at p vanishes again — the place where neighboring geodesics from p refocus. But we just argued that under K <= 0 the separation is convex and strictly increasing once it leaves zero, so a Jacobi field that starts at zero with nonzero derivative can never return to zero. Conclusion: a manifold with K <= 0 has no conjugate points at all. This is the technical heart, and it converts directly through the index form of guide 2: with no conjugate points, the second variation of energy is positive, so every geodesic is a local minimizer, never just a critical point that a Jacobi field could perturb downward.
Now recall what conjugate points obstruct: the exponential map exp_p: T_p M -> M fails to be a local diffeomorphism exactly where a conjugate point sits (its differential degenerates there, by the Jacobi-field interpretation of d(exp_p)). No conjugate points means d(exp_p) is everywhere nonsingular, so exp_p is a local diffeomorphism on all of T_p M. Combine this with completeness — which by Hopf-Rinow from guide 1 makes exp_p defined on the whole tangent space — and you have a smooth surjective local diffeomorphism from a vector space onto M. The Cartan-Hadamard theorem upgrades this to a covering map.
K <= 0 ==> (separation of geodesics is convex) ==> no conjugate points
==> d(exp_p) nonsingular everywhere
complete + no conjugate points
|
v
exp_p : T_p M --> M is a smooth COVERING map
|
v
universal cover ~= R^n (diffeomorphic) [Cartan-Hadamard]The Cartan-Hadamard theorem and what it really says
Here is the full statement to hold fixed. Cartan-Hadamard theorem: if M is a complete Riemannian n-manifold with K <= 0 everywhere, then for every point p the exponential map exp_p: T_p M -> M is a smooth covering map. Consequently the universal cover of M is diffeomorphic to R^n, and in particular M is aspherical: all higher homotopy groups vanish, pi_k(M) = 0 for k >= 2. The proof that exp_p is a covering uses that it is a local diffeomorphism which is also distance-nondecreasing (again from the convexity of Jacobi fields), and a local diffeomorphism that does not shrink distances out of a complete space is a covering map.
Read what this costs and what it buys. It does not say M itself is R^n — a flat torus and a compact hyperbolic surface both have K <= 0 yet are compact, nothing like a plane. What unrolls is the universal cover. So all the topology of M is concentrated in the deck group, i.e. the fundamental group pi_1(M): M is a quotient of R^n (or of hyperbolic space) by a discrete group of isometries acting freely. The torus is R^2 / Z^2; a genus-2 surface is the hyperbolic disk modulo a surface group. Nonpositive curvature trades all higher-dimensional topology for a single, group-theoretic invariant.
Convexity: the geometric superpower of K <= 0
The deepest practical consequence of nonpositive curvature is convexity of the distance function. On a Cartan-Hadamard manifold, fix a point q; the function p -> d(p, q)^2 is convex along every geodesic. Equivalently, given two geodesics gamma_1 and gamma_2, the function t -> d(gamma_1(t), gamma_2(t)) is convex. Compare this to a tiny worked picture: on the Euclidean plane two geodesics from a common point separate at a linear rate (the gap is exactly the angle times the distance); under K <= 0 the gap grows at least that fast, and under strictly negative K it grows exponentially. Distances only ever get more convex than flat.
Convexity is not a curiosity; it is what makes these spaces tame. A convex distance function means balls are convex, geodesics between two points are unique, and — crucially — there are no closed geodesics that bound disks, no focusing, no minimal-surface bubbles trapped inside. It also gives fixed-point theorems: a compact group of isometries of a Cartan-Hadamard manifold must fix a point (the circumcenter of any orbit, well-defined precisely because d^2 is convex). That single fixed-point fact is the geometric seed behind rigidity results and behind why nonpositively curved groups are so well-behaved.
From smooth to synthetic: CAT(0) spaces
Everything above used the smooth machinery — Jacobi fields, the exponential map, the index form. But the conclusions were metric: thin triangles, convex distances, unique geodesics. This suggests stripping away the manifold and keeping only the metric content, which is exactly the CAT(0) idea. A geodesic metric space is CAT(0) if every geodesic triangle is at least as thin as its comparison triangle in the Euclidean plane: build a flat triangle with the same three side lengths, and every chord across the real triangle is no longer than the corresponding chord across the flat model. No curvature tensor required — only distances.
This is the synthetic synthesis of the whole guide, and it generalizes far past manifolds — into Alexandrov spaces, trees, and the polyhedral complexes of geometric group theory, none of which carry a metric tensor. A complete simply connected manifold with K <= 0 is CAT(0), so Cartan-Hadamard becomes a special case of a much broader principle: thin triangles force convex distances, unique geodesics, and contractibility (CAT(0) spaces are contractible, the metric echo of pi_k = 0). Conversely, a Gromov-hyperbolic space captures the strictly-negative case coarsely, where triangles are not just thin but uniformly thin up to a constant.
Honest limits and the road to guide 5
Hold the hypotheses tightly, because each is load-bearing. Drop completeness and exp_p may not even be defined on all of T_p M, killing the covering argument — a punctured nonpositively curved disk has no Cartan-Hadamard conclusion. Drop the sectional bound to a mere Ricci bound and you lose control of individual Jacobi fields, so conjugate points can reappear. Weaken K <= 0 to K < 0 (strict) and you gain more — exponential divergence, hyperbolicity, visual boundaries at infinity — but that is a strengthening, not a free lunch. State K <= 0, complete, every time; the slogan alone will mislead you.
Place this guide in the rung's arc. Guide 3 handled K > 0, where geodesics reconverge and the manifold is compact with finite fundamental group (Bonnet-Myers, Synge). This guide handled K <= 0, where they diverge and the manifold unrolls (Cartan-Hadamard). The two are mirror images across the flat case K = 0. What remains is the quantitative bridge between them: Rauch's theorem compares Jacobi-field growth against a constant-curvature model from either side, and Toponogov's theorem does the same with whole triangles — turning the qualitative pictures of guides 3 and 4 into sharp inequalities.
One honest caveat about depth. We proved Cartan-Hadamard at the level of its real ideas — convex Jacobi fields, no conjugate points, exp_p a covering — but the splitting theorem and the full CAT(0) theory are surveyed here, not proved. The splitting theorem (a complete manifold with Ric >= 0 containing a line splits off an R factor isometrically) genuinely belongs to guide 5 and is a hard analytic argument via the geodesic equation and harmonic functions. Treat what you have here as a load-bearing foundation, and meet the sharp comparison machinery next.