Topological Data Analysis
Core Definitions
Key Formulas
Important Theorems
Worked Examples
Practice Problems
Common Mistakes
| Mistake | Correct Approach |
|---|---|
| Assuming more data always gives better topology | Noise creates short-lived topological features; focus on persistence |
| Confusing homotopy equivalence with homeomorphism | Homotopy equivalence is weaker; a disk and a point are homotopy equivalent but not homeomorphic |
| Using Vietoris-Rips without understanding its relation to Čech | Rips is a superset of Čech; Rips may create false features that Čech avoids |
| Ignoring the choice of distance metric | TDA results depend on the metric; different metrics yield different persistence diagrams |
| Assuming Betti numbers capture all topology | Betti numbers are invariants, but they don't capture the full homotopy type (e.g., lens spaces) |
| Confusing persistent homology with clustering | Clustering gives partition; persistent homology captures higher-dimensional topology (loops, voids) |
| Forgetting that persistence is invariant under rotation/translation | TDA captures intrinsic geometry, not extrinsic position |
Connections to Machine Learning
Exam/Interview Questions
Q1: What are Betti numbers, and what do they measure?
Answer: Betti numbers count the number of -dimensional "holes" in a topological space: = connected components, = independent loops, = enclosed voids. They are topological invariants—preserved under homeomorphism. For a circle: , . For a sphere: , . For a torus: , , .
Q2: Explain the difference between the Vietoris-Rips and Čech complexes.
Answer: The Vietoris-Rips complex includes a simplex if all pairwise distances are . The Čech complex includes a simplex if the balls of radius around the vertices have a common intersection point. Čech is topologically more accurate (by the Nerve Theorem) but harder to compute. Rips is always a superset of Čech and depends only on pairwise distances, making it computationally efficient.
Q3: Why is persistent homology "stable"?
Answer: The stability theorem states that , where is the bottleneck distance. This means if we perturb the data (or the function defining the filtration) by a small amount in the supremum norm, the persistence diagram changes by at most that amount. This guarantees that TDA is robust to noise—small perturbations cause small changes in the topological summary.
Q4: What is the Euler characteristic and how does it relate to Betti numbers?
Answer: The Euler characteristic is . For surfaces: sphere (), torus (), genus- surface (). It can also be computed from any triangulation: where is the number of -simplices. It is a topological invariant (independent of triangulation) and relates geometry to topology via the Gauss-Bonnet theorem.
Q5: How can persistent homology be used as features for machine learning?
Answer: Persistence diagrams can be converted to feature vectors via: (1) Persistence landscapes—transform diagrams into functions amenable to averaging and statistics. (2) Persistence images—rasterize diagrams into fixed-size images. (3) Betti curves—plot as a function of scale. (4) Persistence statistics—compute mean, variance, and other statistics of birth-death pairs. These features capture topological structure (holes, loops, voids) that traditional features miss, and have been used in shape classification, time series analysis, and molecular biology.
Q6: Why is compactness important in topology and analysis?
Answer: Compactness ensures that: (1) Every sequence has a convergent subsequence (sequential compactness), guaranteeing existence of solutions. (2) Continuous functions on compact sets are bounded and attain their extrema (extreme value theorem). (3) Open covers have finite subcovers, enabling finite-dimensional arguments. In TDA, compactness of the underlying space ensures persistence diagrams are well-defined (features have finite birth and death times). In analysis, compactness justifies convergence of iterative algorithms and existence of optimal solutions.
Quick Reference
| Concept | Definition | Key Property |
|---|---|---|
| Topological Space | with open set axioms | Generalizes metric spaces |
| Open Set | Closed under arbitrary union | |
| Closed Set | open | Contains all limit points |
| Connected | Cannot split into disjoint open sets | counts components |
| Compact | Every open cover has finite subcover | Sequentially compact in metric spaces |
| Homeomorphism | Continuous bijection with continuous inverse | Preserves topology |
| Homotopy | Continuous deformation between maps | Preserves "shape" |
| Simplicial Complex | Collection of simplices closed under subsets | Discrete model of topology |
| Betti Number | Counts -dimensional holes | |
| Euler Characteristic | Topological invariant | |
| Persistent Homology | Track feature birth/death across scales | Stable under perturbation |
| Vietoris-Rips | simplex iff | Pairwise distance condition |
| Čech Complex | simplex iff balls have common intersection | Nerve Theorem applies |
| Bottleneck Distance | Stability metric |
Cross-References
- 096-advanced-tensor-calculus — Tensor products in algebraic topology; chain complexes as tensor products
- 097-advanced-differential-geometry — Manifolds are topological spaces with smooth structure; de Rham cohomology relates to simplicial homology
- 098-advanced-functional-analysis — Function spaces are topological vector spaces; weak topologies connect analysis to topology
- 099-advanced-measure-theory — Borel σ-algebras are generated by open sets; measurability relates to topological structure
- Graph Theory: Graphs are 1-dimensional simplicial complexes; graph connectivity relates to