The slogan: triangles that are uniformly thin
In the previous guide you learned to bound curvature by comparing a geodesic triangle against a model triangle in the constant-curvature space M_k. Gromov's insight was to keep only the crudest shadow of an upper bound and throw away the model entirely. Fix a geodesic triangle with sides A, B, C in a geodesic metric space. Call it delta-thin if every point on side A lies within distance delta of the union of the other two sides B and C. A space X is Gromov-hyperbolic if there is a single delta >= 0 that works for every geodesic triangle at once. That is the whole definition: one number, all triangles.
Picture the hyperbolic plane in the Poincaré disk. A huge ideal triangle there is mostly empty in the middle: its three sides hug the boundary so closely that the inscribed circle has a bounded radius no matter how large the triangle. That bounded inradius is exactly delta-thinness made visible. In Euclidean space the opposite happens — a triangle scaled by a factor of t has an inscribed disk of radius proportional to t, so no finite delta can ever work. The plane R^2 is therefore NOT hyperbolic. Negative curvature, coarsely seen, just means triangles never fatten up as they grow.
Four-point and product views of the same idea
Thin triangles are vivid but they presuppose actual geodesics. There is an equivalent formulation that needs only the distance function and so applies to any metric space at all. Fix a basepoint w and define the Gromov product (x | y)_w = (1/2)( d(x, w) + d(y, w) - d(x, y) ). In a tree this is literally the distance from w to the point where the paths to x and y split apart, so it measures how long x and y travel together before diverging. The four-point condition says: for all x, y, z, w the product (x | z)_w is at least the minimum of (x | y)_w and (y | z)_w, up to an additive constant delta.
Gromov product: (x | y)_w = (1/2) ( d(x,w) + d(y,w) - d(x,y) )
four-point delta-inequality, for all x, y, z, w:
(x | z)_w >= min{ (x | y)_w , (y | z)_w } - delta
in a tree (delta = 0): (x | y)_w = length of the shared initial path from wOn a geodesic space the four-point condition and thin triangles are equivalent, with the two deltas differing by a controlled universal factor — so people freely say a space is delta-hyperbolic without fussing over which delta. Be honest about one thing, though: the constant delta is NOT canonical. Different but equivalent definitions, and different basepoints, give different numerical deltas. What is intrinsic is the existence of some finite delta, not its value. Treat delta the way you treat the implied constant in big-O notation.
Why thinness is robust: the Morse lemma
The deepest payoff of thin triangles is stability of quasi-geodesics, often called the Morse lemma. A (lambda, c)-quasi-geodesic is a path that is a geodesic up to multiplicative error lambda and additive error c — exactly the kind of distorted path a quasi-isometry produces when it drags a true geodesic from one space to another. The lemma says: in a delta-hyperbolic space, every (lambda, c)-quasi-geodesic stays within a bounded distance R(delta, lambda, c) of an honest geodesic with the same endpoints. The wobble is forgiven; the large-scale route is pinned down.
This fails dramatically in Euclidean space, and the failure is worth feeling. In R^2 take a logarithmic spiral that crawls outward forever: it can be a quasi-geodesic yet wander arbitrarily far from the straight segment between any two of its points. Negative curvature kills exactly this kind of drift. The mechanism is the thin triangle: if a quasi-geodesic tried to bulge away from the geodesic, the triangle it forms with that geodesic would have to be fat, contradicting delta-thinness. So thinness is not a curiosity — it is the engine that makes hyperbolicity a quasi-isometry invariant.
The boundary at infinity
A proper geodesic hyperbolic space comes with a boundary at infinity, written dX, that records the directions in which one can escape to infinity. The clean construction: take all geodesic rays from a basepoint and declare two rays equivalent if they stay a bounded distance apart forever. Each equivalence class is one boundary point. In the Poincaré disk this recovers exactly the bounding circle — the unit circle that the disk model draws but never reaches. For the regular tree of valence three, the boundary is a Cantor set: the ends of the tree.
Two features make the boundary powerful. First, it does not depend on the basepoint, because the Morse lemma keeps rays from genuinely different ends from ever being confused. Second, and more strikingly, a quasi-isometry between hyperbolic spaces extends to a homeomorphism of their boundaries. So dX is a quasi-isometry invariant: it sees only the large-scale shape, exactly the regime where geometry survives coarsening. The boundary carries more than a topology — it has a natural family of visual metrics, where the distance between two boundary points decays like e^(-epsilon times the Gromov product) measured from the basepoint.
One honest caveat: the visual metric is only defined up to a quasi-symmetry, and its parameter epsilon must be small relative to delta for the construction to give a genuine metric rather than a mere quasi-metric. So while the topology of dX is robust, its precise metric structure is a more delicate object — the subject of conformal dimension and a notoriously hard invariant to compute. As ever in this rung, the coarse data is stable and the fine data is subtle.
Hyperbolic groups, and what the theory does not promise
The reason geometers care so intensely is that groups can be hyperbolic. A finitely generated group, equipped with the word metric on its Cayley graph, is a metric space, and it is called a hyperbolic group when that space is delta-hyperbolic. Because the Cayley graphs for different finite generating sets are quasi-isometric, and the Morse lemma makes hyperbolicity a quasi-isometry invariant, the definition does not depend on the chosen generators. Free groups (their Cayley graphs are trees) and the fundamental groups of closed hyperbolic manifolds are the founding examples.
- Pick a finite generating set S for the group G and build its Cayley graph; the word metric makes it a geodesic space.
- Check that some single delta makes every geodesic triangle delta-thin (equivalently, the four-point condition holds). If so, G is hyperbolic.
- Form the boundary dG from geodesic rays in the Cayley graph; it is a compact metrizable space and a quasi-isometry invariant of G.
- Read off algebra from geometry: G has solvable word problem, contains no copy of Z^2, and has only finitely many conjugacy classes of finite subgroups.
Now resist the hype, which is the running theme of this whole ladder. Hyperbolicity is a strong, restrictive condition, not a generic one: Z^2 is not hyperbolic (it contains flat planes, hence fat triangles), and many groups you can write down are not hyperbolic either. The theory says little about CAT(0) groups that fail to be hyperbolic, where flats reappear and the boundary loses its good behavior. And being a SURVEY is the honest status of the surrounding landscape — Mostow rigidity, the connection to geometric group theory, and the deep theorems on subgroup structure are motivated here, not proved; each is a course in itself.