Four Points Hyperbolic: Unlocking the Secrets of Hyperbolic Geometry
Search for "four points hyperbolic" and you will find the phrase scattered across wildly different branches of mathematics, from the cross-ratio of the Poincaré disk to the four-point condition that defines Gromov hyperbolic spaces, and even to the elementary problem of fitting a hyperbola through four given points. Each of these uses the same four-word hook to describe something distinct, and confusing them is easy. This article untangles the three main meanings, shows where each one comes from, and explains why the number four — rather than two, three, or five — keeps turning up in these settings.
The starting point is a simple observation about symmetry. Two points define a segment, three points define a triangle, but four points are the smallest configuration that carries genuine combinatorial ambiguity: there are three different ways to pair them up. That ambiguity is the seed of the cross-ratio, of the four-point condition, and of the pencil of conics through four fixed points. Understanding four points, in other words, is understanding the first real structure that a set of points can have.
Four Points in the Hyperbolic Plane
The hyperbolic plane is the canonical surface of constant negative curvature. In the Poincaré disk model it is the interior of the unit circle equipped with the metric that blows up near the boundary, so that distances grow without bound as you approach the rim. Geodesics — the hyperbolic analogues of straight lines — appear as circular arcs meeting the boundary at right angles. Triangles in this geometry are visibly thin, with angle sums strictly less than 180 degrees, and the deficit grows with the size of the triangle.
Distance in this world is not a simple function of Euclidean separation. If you take two points, the invariant that actually matters depends on where the geodesic through them meets the boundary circle, which means any honest statement about the separation of two interior points secretly involves four points: the two you care about and two ideal points sitting at infinity. This is the first sense in which four points are intrinsic to hyperbolic geometry rather than merely convenient.
The Cross-Ratio as the Engine of Hyperbolic Distance
For four distinct points on the extended real line or the Riemann sphere, the cross-ratio is defined as (a, b; z, w) = ((a − z)(b − w)) / ((a − w)(b − z)). Its defining property is invariance: Möbius transformations, the maps z ↦ (αz + β)/(γz + δ), send cross-ratios to themselves. Since the isometries of the hyperbolic plane are exactly the Möbius transformations preserving the disk or the upper half-plane, the cross-ratio is the natural measuring stick for the geometry.
Concretely, let z and w be interior points of the disk, and let a and b be the two ideal endpoints of the geodesic that passes through both of them. Then the hyperbolic distance between z and w is given by the absolute value of the logarithm of the cross-ratio of the four points a, b, z, w, up to the curvature normalization you have chosen. A quick sanity check: in the upper half-plane, the points i and iy lie on the imaginary axis whose endpoints are 0 and infinity, and the formula returns exactly |ln y|, which is the textbook distance. Four points, one logarithm, and the entire geometry falls into place.
Ideal Points and Four-Point Configurations on the Boundary
The boundary itself is a fascinating object. It is not part of the hyperbolic plane — no point on the circle is at finite distance from anywhere else — but it controls the geometry from the outside. A Möbius transformation is determined by where it sends three boundary points, which makes the action triply transitive, so no invariant can be built from three boundary points alone. Add a fourth and something survives: the cross-ratio of four boundary points is a real number, and it is a complete invariant for quadruples under the isometry group.
This fact does real work in the theory of Kleinian and Fuchsian groups. Hyperbolic elements of the Möbius group have two fixed points on the boundary, and their translation length along the invariant geodesic can be read off from a cross-ratio involving those fixed points and any two points they move. When four boundary points strictly interleave around the circle, the corresponding geodesics cross, which is the criterion behind many ping-pong arguments used to prove that a group acts freely and properly.
The Four-Point Condition and Gromov Hyperbolicity
Mikhail Gromov's approach to negative curvature, developed in the 1980s, shifted attention from smooth manifolds to metric spaces and asked a coarse question: do triangles in this space look thin at large scales? The resulting theory of δ-hyperbolic spaces applies to Cayley graphs of finitely generated groups, to trees, and to certain networks, and it has become one of the most productive ideas in geometric group theory. Its cleanest formulation avoids triangles entirely and speaks only about four points.
The statement is disarmingly elementary. Given any four points x, y, z, w in a metric space, form the three sums of opposite distances: d(x,y) + d(z,w), d(x,z) + d(y,w), and d(x,w) + d(y,z). A space is 0-hyperbolic precisely when, for every choice of four points, the two largest of these three sums are equal. Relaxing that equality to an inequality — the two largest sums differ by at most 2δ, depending on the convention you adopt — gives the definition of a δ-hyperbolic space.
Reading the Three Pairings
The intuition behind the condition lives in trees. In any tree, four points have a natural meeting structure: there is a unique minimal subtree connecting them, shaped like an H or a star. Pair the points according to how the tree branches and the two largest sums line up exactly; pair them any other way and you lose. This is why trees are the archetypal 0-hyperbolic spaces, and why the four-point condition is often described as measuring how far a space is from being a tree.
The three sums correspond to the three perfect matchings of four objects, and the table below makes the bookkeeping explicit. Notice that each point appears exactly once in each row, and that the rows together account for the six pairwise distances exactly once each per row.
| Pairing | Sum of distances |
|---|---|
| (x, y) with (z, w) | d(x, y) + d(z, w) |
| (x, z) with (y, w) | d(x, z) + d(y, w) |
| (x, w) with (y, z) | d(x, w) + d(y, z) |
Concrete examples sharpen the contrast. The real line satisfies the condition exactly, as does any weighted tree. The Euclidean plane fails it badly: place four points at the corners of a large square and the two largest sums differ by a quantity proportional to the side length, so no finite δ works. The hyperbolic plane Hⁿ, by contrast, is δ-hyperbolic for every n, with a small universal δ. That single fact is what licenses the slogan that hyperbolic spaces behave like trees at large scales.
Trees, Networks, and Machine Learning on Curved Space
The four-point condition doubles as a practical test. Given a distance matrix from real data, you can check all quadruples and see how close the two largest sums are. If they agree, the data is an additive tree metric and can be drawn as a weighted tree with no loss of information — the basis of methods such as neighbor joining and split decomposition in phylogenetics, where the "four points" are species or gene sequences and the tree is an evolutionary history. Buneman's theorem makes the identification precise: satisfying the four-point condition for all quadruples is equivalent to being an additive tree metric.
When the condition fails, the size of the failure is itself informative. Bounded failure means the data is tree-like up to a scale, which is exactly the setting where embedding points into hyperbolic space pays off. Poincaré embeddings exploit this: hierarchical and taxonomic data fit into hyperbolic space with far less distortion than into Euclidean space of the same dimension, because exponentially many leaves can be packed near the boundary. From Gromov's characterization of hyperbolic groups to modern recommender systems that place products in a curved latent space, the four-point condition is quietly doing the quality control.
Four Points on a Hyperbola in Analytic Geometry
The third meaning of "four points hyperbolic" belongs to classical coordinate geometry rather than metric geometry. Here the question is how many points determine a conic section, and what happens when the points you are handed are supposed to lie on a hyperbola rather than an ellipse or a parabola. Five points in general position determine a unique conic; four points determine a one-parameter family, known as a pencil of conics. Everything interesting about four points and conics comes from that extra degree of freedom.
It is worth flagging a terminological trap before going further. In conic sections, "hyperbola" refers to a curve with eccentricity greater than one, defined by the difference of distances to two foci being constant. In geometry, "hyperbolic" usually refers to negative curvature. The two words share a Greek root meaning "excess," and both describe a kind of overshoot, but the mathematics is unrelated. A reader searching for one should not be surprised to land on the other.
Fitting a Conic Through Four Points
Write a general conic as Ax² + Bxy + Cy² + Dx + Ey + F = 0. Four point conditions impose four linear constraints on the six coefficients, leaving a two-dimensional solution space and hence a pencil of conics through the four points. The geometric type is governed by the discriminant B² − 4AC: negative for an ellipse, zero for a parabola, positive for a hyperbola. As you move through the pencil, the discriminant changes sign, so a single set of four points can typically be joined by ellipses and hyperbolas alike.
If you specifically want a rectangular hyperbola, you add the linear condition A + C = 0, which forces the asymptotes to be perpendicular. One extra linear equation on a two-dimensional pencil generically cuts out a single member, so four points in general position lie on exactly one rectangular hyperbola — provided that condition is not identically satisfied across the whole pencil. Numerically, this is a small eigenvalue problem dressed up in polynomial clothing, and poorly conditioned when the four points are nearly collinear or nearly concyclic.
Orthocentric Systems and Rectangular Hyperbolas
There is a beautiful rigidity result here. Take any triangle, and consider its four-point orthocentric system: the three vertices together with the orthocenter. A classical theorem states that any rectangular hyperbola passing through the three vertices of a triangle also passes through its orthocenter. Reversing the logic, any four points forming an orthocentric system lie on a unique rectangular hyperbola, and that curve is the natural frame in which the four points are symmetric.
The named triangle hyperbolas make this concrete. The Kiepert hyperbola, generated by erecting similar triangles on the sides of a reference triangle, is a rectangular hyperbola passing through the three vertices and the orthocenter. The Jerabek hyperbola, defined through the circumcenter, is another. In each case the four-point configuration is not incidental but structural: the orthocentric relationship among the points is exactly what forces the curve to have perpendicular asymptotes. Algebraic conditions and synthetic configurations agree, as they usually do.
Four points turn out to be the smallest dataset that can carry a genuine invariant, whether that invariant is a cross-ratio on the boundary of the disk, a measure of how far a metric space sits from being a tree, or the unique rectangular hyperbola threading through an orthocentric system. If you want to test the idea on your own material, the most accessible entry point is the four-point condition: take any distance matrix, compute the three pairing sums for a handful of quadruples, and see how close the top two come to matching. The gap you measure is a surprisingly honest summary of how tree-like, how flat, or how curved your data really is.