Hyperbolic Geometry Models via Stereographic and Orthogonal Projections

The geodesic connecting two points u1 and u2 in the upper half-plane is the circular arc (or vertical line) intersecting the real axis orthogonally and passing through u1 and u2. This arc minimizes the hyperbolic distance given by d(u1,u2) = |ln((u1-u2)/(u1-ū2) * (ū2-u2)/(ū2-ū1))|/2. The hyperbolic distance is invariant under SL(2,R) transformations, meaning d(T(u1), T(u2)) = d(u1, u2) for any T in SL(2,R). This invariance follows from the invariance of the cross ratio under Möbius transformations. In hyperbolic geometry, a triangle is a region bounded by three geodesic arcs, with angles determined by hypergeometric function parameters. In the hyperbolic case (sum < π), the area is π - (θ1 + θ2 + θ3) by the Gauss-Bonnet formula.

A two-dimensional space with constant negative curvature (Gaussian curvature = -1) is a hyperbolic space. In the Poincaré half-plane model, geodesics are either vertical lines or semicircles orthogonal to the x-axis, and they diverge from each other. To compute distances, we find the geodesic curve connecting two points and integrate the line element ds along it. Using the Euler-Lagrange equation and Beltrami identity, we derive that geodesics are Euclidean semicircles centered on the x-axis (with vertical lines as special cases where the center is at infinity). The shortest path between points (0,1) and (1,3) in this space is approximately 1.2158, which is about 45.64% less than the Euclidean distance (2.2361) calculated using the Pythagorean theorem.

There are three fundamental geometries with different properties: (1) Euclidean geometry (flat plane) where straight lines are shortest paths and triangle angles sum to 180°; (2) Spherical geometry (surface of sphere) where geodesics are great circles and triangle angles sum to more than 180°, with the excess proportional to area; (3) Hyperbolic geometry (represented by Poincaré disc) where geodesics are arcs of circles meeting the boundary at right angles and triangle angles sum to less than 180°.

This section develops the concept of geodesics (shortest paths) and distance in hyperbolic geometry. Geodesics in the upper half-plane are either vertical lines perpendicular to the real axis or semicircles centered on the real axis that intersect it orthogonally. The hyperbolic distance between two points M and N is given by d(M,N) = |log(cross-ratio(M, N, N₀, M₀))|, where N₀ and M₀ are the intersection points of the geodesic through M and N with the real axis. This distance is invariant under PSL2(R) transformations. The geodesic connecting M and N is the unique arc of a circle orthogonal to the real axis passing through both points. This geometric interpretation connects algebraic invariants (cross-ratio) with geometric intuition (shortest paths).

In hyperbolic geometry, through a point not on a given line, there can be more than one line parallel to the original line. This is impossible on a flat surface but possible on a curved surface. What makes lines 'straight' is that they are the shortest paths between two points on a surface. On curved surfaces, these shortest paths (called geodesics) may appear curved due to the surface's curvature. For example, airplanes flying between cities follow the shortest path, which appears curved on a flat map because the Earth's surface is curved. These geodesics are the 'straight lines' of hyperbolic geometry.

The hyperbolic plane H² is equipped with the metric induced by the density λ(z) = 1/Im(z), and PSL(2,R) acts as a group of isometries on H² because Möbius transformations T(z) = (az + b)/(cz + d) with real coefficients and determinant 1 satisfy the condition that the pullback of the hyperbolic density equals the original density divided by the absolute value of the conformal factor |T'(z)| = 1/|cz + d|², preserving hyperbolic distances.

Möbius transformations z ↦ (az + b)/(cz + d) with ad - bc = 1 are the isometries of H². The group PSL(2,ℝ) (2×2 real matrices with determinant 1, modulo ±I) is the full orientation-preserving isometry group of the hyperbolic plane. The instructor explains how these transformations preserve the hyperbolic metric and act naturally on H². The boundary ∂H² is topologically ℝ ∪ {∞}, the real projective line. PSL(2,ℝ) acts on ∂H² by extending the action from H². This boundary action is crucial for understanding element dynamics.

The isometry group of the hyperbolic plane is PSL(2,R), with Möbius transformations z ↦ (Az+B)/(Cz+D). The unit tangent bundle is identified with PSL(2,R) itself. The geodesic flow corresponds to right multiplication by diagonal matrices [[e^(t/2), 0], [0, e^(-t/2)]]. The stable horocyclic flow corresponds to [[1, τ], [0, 1]], and the unstable horocyclic flow corresponds to [[1, 0], [τ, 1]]. The rotational flow corresponds to rotation matrices. The Lie bracket relations [N-, A] = -N- and [N+, A] = N+ encode the geometric relations between these flows.

The orientation-preserving isometry group of the hyperbolic plane is isomorphic to PSL(2,R), and the full isometry group is PSL(2,R) semidirect product with Z/2Z, where elements are classified by their fixed points: elliptic elements have two complex conjugate fixed points, parabolic elements have one fixed point on the boundary, and hyperbolic elements have two distinct fixed points on the boundary.

The hyperbolic plane H is a two-dimensional manifold with constant negative curvature normalized to -1, equipped with the metric ds² = dx² + dy²/y² in the upper half-plane model. Its orientation-preserving isometry group is PSL(2,R), acting via Möbius transformations. Hyperbolic surfaces are defined as quotients H/Γ where Γ is a discrete subgroup of PSL(2,R). When Γ has finite covolume (finite area), Γ must be infinite and finitely presented. Elliptic elements in Γ produce conical points (elliptic points) on the surface. Hyperbolic triangle groups ΓABC are constructed from three positive integers A,B,C by taking geodesics meeting at angles π/A, π/B, π/C, with area π(1-1/A-1/B-1/C).

Hyperbolic geometry provides a natural framework for embedding hierarchical and tree-like data because hyperbolic space exhibits exponential volume growth, unlike Euclidean space which only shows polynomial growth; this allows hyperbolic embeddings to efficiently represent complex hierarchical structures such as phylogenetic trees or word taxonomies, where geodesics in hyperbolic space better approximate the hierarchical distances between nodes compared to straight-line distances in Euclidean space.

Hyperbolic geometry, discovered in the 19th century as a non-Euclidean geometry where the parallel postulate fails, provides a powerful framework for representing hierarchical data structures that exhibit exponential growth, such as word hierarchies and social networks; this is achieved through models like the Poincaré disk and connections to special relativity via the pseudoeuclidean Minkowski metric, enabling the development of hyperbolic neural networks that replace standard Euclidean operations with hyperbolic transformations for more efficient data representation and analysis.

This section develops hyperbolic geometry applications for hierarchical data. It covers: (1) Exponential volume growth in hyperbolic spaces enabling low-dimensional representation of hierarchical structures, contrasting with Euclidean polynomial growth; (2) The perceptron algorithm adapted to hyperbolic manifolds by learning manifold-Euclidean hyperplane intersections; (3) Theoretical analysis showing hyperbolic space requires only 2 dimensions for tree representation while Euclidean needs logarithmic dimensions; (4) Complexity analysis demonstrating exponential vs constant scaling benefits; (5) Empirical validation on hierarchical image data showing perfect classification in 8D hyperbolic space versus significant error in Euclidean space, confirming theoretical predictions about representation trade-offs.

Hyperbolic geometry's applications extend far beyond theoretical mathematics. M.C. Escher created Circle Limit artwork demonstrating hyperbolic principles, with figures shrinking toward the boundary to maintain equal hyperbolic size. Most significantly, hyperbolic geometry serves as the mathematical foundation for Einstein's special relativity theory, introduced approximately 100 years after Gauss developed non-Euclidean geometry. In modern data science, neural networks classify data by representing inputs as points in space, but hierarchical data like phylogenetic trees face representation challenges in Euclidean space. Species that are very different evolutionarily can cluster together, causing misclassification. Hyperbolic geometry solves this because points near the boundary are actually very far apart in hyperbolic distance. By embedding trees in hyperbolic space, species that are far apart evolutionarily can be placed far apart geometrically. Higher-dimensional hyperbolic spaces provide analogous solutions for complex hierarchical datasets. This demonstrates how abstract mathematical discoveries made centuries ago find profound practical applications in physics and data science, fundamentally describing the geometry of spacetime and enabling better representation of hierarchical data structures.

Hyperbolic geometry, characterized by negative constant curvature and exponential distance growth from the origin, provides a more natural representation space for hierarchical data structures compared to Euclidean geometry. This is because hyperbolic spaces can compactly encode tree-like hierarchies, where distances grow exponentially as one moves away from the center, making them particularly effective for representing discrete entities like words (capturing semantic hierarchies such as 'king - man + woman = queen') and graph nodes (preserving hierarchical relationships in networks). Research has shown that hyperbolic embeddings can achieve comparable or superior performance to Euclidean embeddings in lower dimensions, especially for tasks involving hierarchical data, due to their ability to capture latent tree structures that Euclidean geometry cannot efficiently represent.