The trick: compare a triangle against a model
In Guide 1 you learned to do geometry with only a distance d, and to demand that shortest curves actually exist — a geodesic space. Now we add the missing ingredient: curvature, with no smoothness, no sectional curvature tensor, not even a tangent space. The idea is disarmingly simple. Pick three points in your space and join each pair by a shortest path; you get a geodesic triangle. Now draw, on a familiar model surface of CONSTANT curvature k, a second triangle whose three side lengths match yours exactly. That model triangle is the comparison triangle, and comparing your triangle to it is how curvature sneaks back in.
The three model surfaces are exactly the constant-curvature space forms you already know. For k = 0 the model is the flat Euclidean plane. For k > 0 it is the round sphere of radius 1/sqrt(k), where triangles are fat — their angles sum to MORE than pi. For k < 0 it is the hyperbolic plane of curvature k, where triangles are thin and starved, with angle sum LESS than pi. Call this model surface M_k. Given any three side lengths (small enough, if k > 0, to actually fit on the sphere), there is a triangle in M_k with those sides, and it is unique up to congruence. That uniqueness is what lets the comparison triangle serve as a fixed yardstick.
CAT(k): triangles at least as thin as the model
Here is the precise comparison. Take a geodesic triangle with vertices p, q, r and pick a point x on the side from q to r. In the comparison triangle in M_k, the matching vertices p-bar, q-bar, r-bar have a matching point x-bar on the q-bar-r-bar side at the SAME distance along the edge. A space is a CAT(k) space if, for every such triangle and every such point, the genuine distance d(p, x) is at most the model distance d(p-bar, x-bar). In words: chords inside your triangle are no longer than the corresponding chords in the model. Your triangles are pinched, slim, no fatter than the flat (or spherical, or hyperbolic) reference. The letters honor Cartan, Alexandrov, and Toponogov.
geodesic triangle p, q, r comparison triangle in M_k: p-bar, q-bar, r-bar
x on side [q, r] x-bar on [q-bar, r-bar], same edge-distance
CAT(k): d(p, x) <= d(p-bar, x-bar) ( your triangle is no fatter )
sides match: d(p,q) = d(p-bar,q-bar), d(q,r) = d(q-bar,r-bar), d(r,p) = d(r-bar,p-bar)When k > 0 there is a size limit baked in: triangles on the sphere of radius 1/sqrt(k) cannot have perimeter beyond 2 pi / sqrt(k), so the CAT(k) condition is only imposed on triangles small enough to have a comparison triangle at all. For k <= 0 there is no such ceiling — every triangle, however large, gets compared. This is why CAT(0) and CAT(-1) are global conditions while CAT(k) for k > 0 is inherently local in scale. A surface can be CAT(1) in the small while doing something wild in the large.
CAT(0): the geometry of nonpositive curvature, coarsely
The case k = 0 is the workhorse, so it gets its own name: a CAT(0) space is a geodesic space all of whose triangles are at least as thin as flat Euclidean triangles. This single condition forces a remarkable list of consequences, all provable with nothing but the comparison inequality. Geodesics between two points are UNIQUE. The distance function is convex along geodesics, so the space has no 'shortcuts that fork'. There are no conjugate points and no closed geodesics that bound a disk. Most strikingly, a CAT(0) space is contractible — it deformation-retracts to a point by sliding everything along geodesics toward a fixed center. All of this is the metric shadow of nonpositive curvature.
This is exactly the metric-space face of a theorem you met in the comparison-geometry rung. The Cartan-Hadamard theorem said a complete, simply-connected Riemannian manifold of nonpositive sectional curvature is diffeomorphic to R^n via the exponential map. The metric reformulation drops 'Riemannian' and 'diffeomorphic' entirely: a complete, simply-connected space that is LOCALLY CAT(0) is automatically GLOBALLY CAT(0), hence contractible. Local nonpositive curvature plus simple connectivity globalizes — the same Cartan-Hadamard miracle, now stated with no manifold in sight, valid for singular spaces like trees and cube complexes that are not manifolds at all.
Alexandrov spaces: curvature bounded the other way
Flip every inequality and you get the dual theory. An Alexandrov space of curvature bounded BELOW by k is a geodesic space where triangles are at least as FAT as in M_k: now d(p, x) is at least d(p-bar, x-bar) for every triangle and every point on a side. Triangles bulge outward, angles are wider than the model's, geodesics tend to spread apart slowly. This is the metric incarnation of a LOWER curvature bound, and it is the direct descendant of the Toponogov comparison theorem you saw for Riemannian manifolds with sectional curvature bounded below.
Why bother with the lower-bound side separately? Because it is precisely the class of spaces that survives a limit. This is the deep payoff of Guide 1's Gromov-Hausdorff distance: if a sequence of Riemannian manifolds all have sectional curvature at least k and converge in the Gromov-Hausdorff sense, the limit is an Alexandrov space of curvature at least k — even though the limit can be a cone, can have corners, can drop dimension. Lower curvature bounds are STABLE under Gromov-Hausdorff convergence; this stability is the engine behind Gromov's compactness theorem and the structure theory of collapsing manifolds. CAT(k), by contrast, is the right notion when you want uniqueness and contractibility, not limits.
A concrete picture cements the asymmetry. The surface of a solid cube is an Alexandrov space of curvature bounded below by 0: it is flat on each face but concentrates a positive angle defect at each corner, where a little triangle straddling the vertex is fatter than flat. It is NOT a CAT(0) space — those same corners make some triangles too fat, violating the upper bound. The cone over a circle behaves the same way: positively curved at the cone point in the lower-bound sense, never an upper-bound space there. Cones, corners, and edges are the native habitat of Alexandrov geometry, and the place where the theory is genuinely more general than the smooth one.
Angles without calculus, and where this is honestly hard
One subtlety deserves spelling out: in a bare metric space there is no inner product, so what is an 'angle' between two geodesics leaving a point p? The comparison framework gives a clean answer. For geodesics starting at p, look at two nearby points x and y along them, form the comparison triangle for p, x, y in the flat plane, and read off its angle at p-bar; the comparison angle is this Euclidean angle. The honest Alexandrov angle between the two geodesics is the limit of these comparison angles as x and y slide back to p. In a CAT(k) space these limits exist and behave well (angles are well-defined and the triangle angle sums obey the model bound); that they exist at all is a small theorem, not a definition.
Be honest about the boundaries of all this. First, every definition here PRESUPPOSES a geodesic space — shortest paths must exist, which by Guide 1 needs completeness and (locally) compactness; on a mere length space the comparison game does not even start. Second, comparison gives you a curvature BOUND, an inequality, never a curvature VALUE: CAT(0) says 'curvature at most 0 in this synthetic sense', it does not assign a number to each point the way sectional curvature does. Third, for k > 0 you must constantly mind the diameter restriction, or you compare against a triangle that does not exist. These are not technicalities to wave away; they are where careful proofs live.
Why does all this earn its keep, rather than being abstraction for its own sake? Because it lets curvature reach spaces calculus cannot touch. The boundary of any Riemannian symmetric space, a building, an infinite tree, the Cayley graph of a group, a polyhedral complex glued from flat pieces — none of these is a smooth manifold, yet each can be exactly CAT(0) or CAT(-1), and that single fact controls their large-scale geometry. This is the bridge to the rest of the rung: in the next guide, Gromov-hyperbolic spaces coarsen CAT(-1) into a condition stable under quasi-isometry, and then groups themselves become geometric objects, fat or thin triangles and all.