Complex Analysis
Möbius Transformations
In 1827 August Möbius introduced fractional linear transformations in *Der Barycentrische Calcul*. Fifty-five years later Poincaré realized those same maps are the isometries of the hyperbolic plane. In 2017, Maximilian Nickel and Douwe Kiela (Facebook AI Research, NeurIPS) published Poincaré embeddings: WordNet's 82,000-noun hierarchy fits into a 5-dimensional Poincaré disk with MAP 0.86, beating a 200-dimensional Euclidean embedding at 0.81. A 19th-century geometric construction became state of the art in 21st-century graph machine learning.
- **Poincaré Embeddings (Nickel-Kiela, NeurIPS 2017):** WordNet's 82,000-node hierarchy in a 5D Poincaré disk reaches MAP 0.86 versus Euclidean's 0.81 at 200D
- **Computer vision (homography):** the cross-ratio of four collinear points is the only projective invariant, used directly in OpenCV's `findHomography`
- **Circle packing (Thurston, 1985):** Stephenson's CirclePack algorithm uses Möbius maps to build conformal flattenings of 3D meshes
- **SL(2,R) monodromy:** Heun and Painlevé ODEs are described by Möbius actions of the fundamental group, central to integrable systems and string theory
Предварительные знания
The Möbius Group
Möbius transformations form a group under composition, isomorphic to PSL(2,C). Every conformal automorphism of the Riemann sphere is a Möbius transformation, making the group fundamental to complex analysis and geometry.
**Möbius transformation:** f(z) = (az+b)/(cz+d), ad−bc ≠ 0 **Group structure:** Möb ≅ PSL(2,C) = SL(2,C)/{±I} - Composition ↔ matrix multiplication: [a,b;c,d] ∘ [e,f;g,h] - Identity: f(z) = z ↔ I = [1,0;0,1] - Inverse: f⁻¹(z) = (dz−b)/(−cz+a) **Subgroups:** - PSL(2,R): automorphisms of the upper half-plane - PSU(1,1): automorphisms of the unit disk
The composition of two Möbius transformations with matrices M₁ and M₂ corresponds to:
Fixed Points and Classification
Every non-identity Möbius transformation has exactly 1 or 2 fixed points on the Riemann sphere. The number and nature of fixed points classifies all Möbius transformations into three geometric types.
**Fixed points:** solve f(z) = z, i.e., (az+b)/(cz+d) = z → cz² + (d−a)z − b = 0 **Classification by trace τ = a+d (in SL(2,C)):** - **Elliptic:** τ ∈ (−2,2) real → two conjugate fixed points - Rotates around fixed points; |τ|<2 - Example: z ↦ e^{iθ}z (pure rotation) - **Hyperbolic:** τ ∈ R, |τ|>2 → two real fixed points - Stretches between fixed points - **Parabolic:** τ = ±2 → one fixed point (double root) - Translation-like: conjugate to z ↦ z+1
A Möbius transformation with trace τ = 2 has:
The Cross-Ratio
The cross-ratio is the unique Möbius-invariant. It encodes the relative positions of four points on the Riemann sphere and is preserved under every Möbius transformation.
**Cross-ratio:** (z₁,z₂;z₃,z₄) = (z₁−z₃)(z₂−z₄) / [(z₁−z₄)(z₂−z₃)] **Möbius invariance:** for any Möbius f: (f(z₁),f(z₂);f(z₃),f(z₄)) = (z₁,z₂;z₃,z₄) **Key uses:** - Three points → one (determine the map) - Real cross-ratio ↔ four points on a circle or line - Hyperbolic distance in the Poincaré model: d(z₁,z₂) = |log(z₁,z₂;p,q)| where p,q are ideal boundary points
The cross-ratio (z₁,z₂;z₃,z₄) is real if and only if:
The Poincaré Disk Model
The Poincaré disk model of hyperbolic geometry lives inside the unit disk. Its isometries are exactly the Möbius transformations preserving the disk: the connection between complex analysis and non-Euclidean geometry.
**Poincaré disk:** H² = {|z| < 1} with metric ds = 2|dz|/(1−|z|²) **Isometries:** Möbius transformations of the form: f(z) = e^{iθ} · (z−z₀)/(1−z̄₀z) **Geodesics:** arcs of circles perpendicular to the unit circle ∂D **Hyperbolic distance:** d(z₁,z₂) = 2·arctanh|(z₁−z₂)/(1−z̄₁z₂)| **Group:** PSU(1,1) = automorphisms of D
Möbius, Poincaré, and non-Euclidean geometry
August Möbius introduced the transformations $(az + b)/(cz + d)$ in *Der Barycentrische Calcul* (1827) as a tool of projective geometry. Fifty-five years later, Henri Poincaré in *Théorie des groupes fuchsiens* (1882) recognized that these same transformations are precisely the isometries of the hyperbolic plane (the Poincaré disk). That insight gave the modern realization of Lobachevsky's non-Euclidean geometry. In 2017 Maximilian Nickel and Douwe Kiela at Facebook AI Research (NeurIPS) revived the disk for machine learning: WordNet's noun hierarchy embeds into a 5-dimensional Poincaré disk more accurately than into a 200-dimensional Euclidean space.
In the Poincaré disk, geodesics are:
Key Ideas
- **Möbius group:** Möb ≅ PSL(2,C); composition = matrix multiplication
- **Classification:** elliptic/hyperbolic/parabolic by trace τ; 2 / 2 / 1 fixed points
- **Cross-ratio:** Möbius-invariant; real ↔ four concyclic points
- **Poincaré disk:** isometries = Möbius transforms preserving |z|<1; geodesics ⊥ ∂D
Connected Topics
Möbius transformations connect complex analysis to geometry and physics:
- Conformal Mappings — Möbius transforms are the simplest and most important conformal maps
- Analytic Continuation — Möbius transforms on the Riemann sphere appear in analytic continuation over complex paths
- Complex Analysis in Physics — PSL(2,C) is the Lorentz group; Möbius transforms act on the relativistic celestial sphere
Вопросы для размышления
- Why does PSL(2,C) rather than GL(2,C) appear as the Möbius group?
- How does the cross-ratio prove that Möbius transformations map generalized circles to generalized circles?
- Why is the Poincaré disk model of hyperbolic geometry relevant to machine learning embedding problems?