Differential Geometry
Comparison Geometry
Comparison geometry asks how closely a Riemannian manifold resembles a sphere or hyperbolic space. Perelman's 2003 proof of the Poincare conjecture relied exactly on Ricci flow combined with the curvature comparison techniques developed in this area.
- Poincare conjecture: Perelman 2003 used Ricci flow and Bishop-Gromov comparison
- Machine learning: loss surface curvature is analogous to Ricci curvature bounds
- Shape analysis: Gromov-Hausdorff metric for comparing geometric shapes
- Computer vision: Gromov-Wasserstein distance for point cloud matching
- Neuroscience: metric geometry of connectome graphs and brain networks
- Materials science: Alexandrov spaces arising in discrete geometric structures
Цели урока
- Understand the Gromov-Hausdorff metric and its properties
- Master the Bonnet-Myers and Bishop-Gromov theorems
- Know the Rauch comparison theorem and the sphere theorem
- Understand Alexandrov spaces as GH-limits of manifolds
Предварительные знания
- Riemannian geometry and curvature
- Geodesics and exponential map
- Convergence theory
Comparison Theorems and the Gromov-Hausdorff Metric
How can one compare the shape of two metric spaces without assuming anything about their structure, and what happens to the geometry of a manifold as curvature pinching increases?
- Topological data analysis (TDA): the Gromov-Hausdorff metric is used to compare point clouds in feature space; persistence diagrams are stable under GH-metric (Cohen-Steiner stability theorem: d_bottleneck(Dgm(f), Dgm(g)) <= ||f-g||_inf).
- Hamilton-Perelman geometrization program: Perelman's proof of Thurston's geometrization conjecture uses Gromov-Hausdorff limit transitions in Ricci flow with surgery - collapsing parts of the manifold converge to Alexandrov spaces.
- ICP (Iterative Closest Point) algorithm for aligning 3D models in computer vision minimizes an approximation of the GH-distance between discrete surfaces, alternating between nearest-neighbor search and least-squares minimization.
- Geometric deep learning: comparison theorems are used to bound the expressiveness of convolutional networks on spaces with a lower curvature bound; Lott-Villani-Sturm defined synthetic Ricci curvature via optimal transport.
Comparison geometry studies how lower (or upper) curvature bounds constrain the topology and geometry of a manifold. The key tool is comparing volumes of balls and Jacobi field lengths with reference spaces of constant curvature: sphere S^n(K>0), Euclidean space (K=0), and hyperbolic space H^n(K<0). The Gromov-Hausdorff metric allows one to speak about convergence of sequences of manifolds to limits that may not be smooth manifolds. Classical results: Bonnet-Myers theorem (Ric>0 compactifies), sphere theorem (1/4 < sec <= 1 => homeomorphic to S^n), Cheeger finiteness theorem (finitely many diffeomorphism types under bounded curvature, diameter, and volume). The Cheeger-Colding theory extends these results to metric measure spaces, proving that manifolds with Ric >= -(n-1) and non-collapsed volume converge to spaces with a well-defined measure and a notion of heat flow, enabling quantitative estimates on singular sets.
Connections to Other Topics
Comparison geometry unifies Riemannian geometry, topology, and modern analysis on spaces with singularities.
- Geometrization Conjecture — Perelman's proof (2003) uses GH-limits in Ricci flow with surgery; collapsing parts converge to Alexandrov spaces, which must be classified topologically
- Optimal Transport — The Wasserstein metric W_2 on measure spaces is analogous to GH; Lott-Villani-Sturm defined synthetic Ricci curvature via convexity of entropy in (W_2, Prob(M))
- Topological Data Analysis — Stability theorem: persistence diagrams are stable under GH-metric; d_bottleneck <= ||f-g||_inf; this makes TDA robust to noise in point cloud data
- Cheeger-Colding Theory — Manifolds with Ric >= -(n-1) and volume bounded below converge to metric measure spaces; this is the foundation for Naber-Valtorta regularity and quantitative estimates
Итоги
- GH-metric d_GH makes the space of isometry classes of compact metric spaces into a metric space; d_GH = 0 iff X and Y are isometric
- Bonnet-Myers theorem: Ric >= (n-1)K > 0 => diam(M) <= pi/sqrt(K), compactness, and finiteness of pi_1(M)
- Bishop-Gromov theorem: lower Ricci bound controls ball volume growth from above via the model space of curvature K
- Rauch theorem: lower sectional curvature bound slows geodesic divergence; sphere theorem: 1/4 < sec <= 1 => M ~ S^n
- GH-limit of manifolds with lower curvature bound is an Alexandrov space (singular generalization with curvature defined via triangles)
- Applications: Poincare conjecture proof (Perelman), TDA, optimal transport
The Bonnet-Myers theorem states: if Ric >= (n-1)K > 0, then:
Bonnet-Myers: Ric >= (n-1)K > 0 compactifies the manifold (completeness + bounded diameter). Finiteness of pi_1 follows from the estimate on the universal cover. Flatness requires Ric = 0, and vanishing b_1 follows from Bochner's theorem but not directly from Bonnet-Myers.
Итоги
- GH-metric: d_GH(X,Y) = inf_{Z,i_X,i_Y} d_H^Z(i_X(X), i_Y(Y)); metric space of isometry classes of compact metric spaces
- Bonnet-Myers: Ric >= (n-1)K > 0 => diam(M) <= pi/sqrt(K), pi_1(M) finite
- Bishop-Gromov: vol(B_r(p))/v_K(r) non-increasing under Ric >= (n-1)K
- Rauch: |J(t)| >= |J_delta(t)| for K <= delta; sphere theorem for 1/4 < sec <= 1
- GH-limit is an Alexandrov space with curvature >= K