Video summary

DSA2023.1 -- Lecturer : Raphaël Kevin Tinarrage (FGV)- An introduction to Topological Data Analysis

Main summary

Key takeaways

Science and Nature

Scientific Concepts, Discoveries, and Nature Phenomena Presented

Core Field: Topological Data Analysis (TDA)

Topological Data Analysis (TDA) is presented as a framework that uses ideas from algebraic topology to analyze data. Its purpose is to uncover hidden structure in datasets using topological invariants.

Key philosophy: Data clouds reflect underlying topological spaces. The goal is to recover meaningful structure—such as connected components, holes, and the overall shape—from noisy and/or high-dimensional observations.

Examples Motivating TDA (Data → Point Clouds → Topology)

Chemistry: Molecular Conformation Spaces

Problem: Characterize the configuration/conformation space of molecules where physical constraints restrict possible atom positions.

Method described:

  • Consider a cyclic molecule with:
    • 8 carbon atoms
    • 16 hydrogen atoms
    • Total: 24 atoms
  • Each atom has 3 Euclidean coordinates.
  • Each molecule is mapped to a single point in:
    • (\mathbb{R}^{72}) because (3 \times 24 = 72).
  • A large set of such molecules produces a point cloud in (\mathbb{R}^{72}).
  • TDA reveals the point cloud is close to a topological shape described as:
    • a union of a sphere and a cylinder-like (“cylindrical model”) component embedded in high-dimensional space.

Interpretation: Some molecular configurations correspond to critical configurations, linked to intersections of components in the geometric/topological model.

Biology / Neuroscience: Grid Cells and Spatial Mapping

Phenomenon: Grid cells encode spatial structure, supporting spatial visualization/mapping.

Hypothesis stated: The interaction pattern among grid cells should have torus-like (torus) topology.

Method described:

  • Record spike activity from many grid cells using Neuropixels.
  • Each neuron yields a time series.
  • Compute a distance between grid cells based on their time series.
  • Build a distance-based matrix (implied).
  • Compute the topology of the resulting structure, reported as supporting a torus.

Medicine / Oncology: Breast Cancer Subtyping via Topology

Problem: Breast cancer contains many subtypes, making categorization challenging.

Method described:

  • Convert each patient’s genomic information into a feature vector with 262 coordinates.
  • Represent each patient as a point in (\mathbb{R}^{262}), forming a point cloud.
  • Topological analysis yields a point cloud shaped like a tree with three main branches, corresponding to different breast cancer types.
  • One branch corresponds to a previously not-known patient group characterized by expression of cMYB (“cMYB plus”).
  • This branch group is described as having favorable outcomes (patients reportedly “do not die from the cancer,” i.e., survival is observed).

The Mathematical Framework: Topology and Invariants

Topological Spaces and Continuous Maps

The transcript emphasizes:

  • Topological spaces defined via open sets.
  • Continuity defined purely in topological terms (stated as equivalent to the usual (\varepsilon)–(\delta) definition when viewed via Euclidean embedding).

Equivalence Relations on Spaces

Homeomorphism (bicontinuous bijection)

Definition: A map that is:

  • bijective,
  • continuous, and
  • has a continuous inverse.

This is described as bicontinuous.

Examples:

  • A circle is homeomorphic to other “closed loop” shapes (e.g., polygons approximating a circle).
  • An interval is not homeomorphic to a circle since going from an interval to a circle requires “cutting/opening,” which is not allowed under continuous bijections.

Brouwer’s theorem (as stated in the transcript): (\mathbb{R}^n) and (\mathbb{R}^m) are homeomorphic iff (n=m).

Surface classification via genus:

  • Sphere: genus 0
  • Torus: genus 1
  • Double torus: genus 2 (and so on) under deformation/homeomorphism-like equivalence.

Homotopy

Definition: Maps are continuously deformable through a parameter (t \in [0,1]).

Example contrast:

  • Two maps may be homotopic when the codomain is (\mathbb{R}^2).
  • They may fail to be homotopic when the codomain is (\mathbb{R}^2 \setminus {0}) because removing the origin creates a topological obstruction.

Homotopy Equivalence

Definition: There exist mutual maps:

  • (F: X \to Y)
  • (G: Y \to X)

such that:

  • (F \circ G) and (G \circ F) are homotopic to the relevant identity maps.

Key point: Homotopy equivalence is weaker than homeomorphism, meaning more spaces can fall into the same equivalence class. The transcript also mentions contractible spaces (everything contracts to a point) as a class.

Topological Invariants (to Distinguish Equivalence Classes)

Invariance under homeomorphism

  • Embeddability: If two spaces are homeomorphic, then either both can be embedded into (\mathbb{R}^n) or neither can (if one embeds, so does the other).

Invariance under homotopy equivalence

  • Number of connected components
  • Euler characteristic
    • presented as an alternating-sum invariant,
    • also described as “weak” compared with homology.
  • Betti numbers
    • invariants of homotopy type computed via homology.

Simplex Complexes and Euler Characteristic

Methodology described:

  • Represent a space with a simplicial complex:
    • choose a vertex set (V),
    • include a collection of simplices such that if a simplex is included, all its faces are included.
  • Compute an Euler-like alternating sum:
    • in the transcript: “number of vertices − number of edges + number of triangles − …”
  • This generalizes Euler characteristic from polyhedra and relates to Euler characteristic of triangulations.

Betti Numbers (Homology-Based Invariants)

Betti numbers (\beta_0, \beta_1, \beta_2, \dots) summarize:

  • (\beta_0): number of connected components
  • (\beta_1): number of holes (1D cycles)
  • (\beta_2): number of voids (2D cavities)
  • etc.

Examples listed:

  • Circle: (\beta_0 = 1), (\beta_1 = 1)
  • Torus: stated as (\beta_0 = 1), (\beta_1 = 2), (\beta_2 = 1)
  • Sphere: stated as (\beta_0 = 1), (\beta_1 = 0), (\beta_2 = 1)

Additional Application Example: Natural Image Topology

Problem: Analyze the space of natural images using TDA.

Method described:

  • Extract small grayscale patches (e.g., 3×3 patches).
  • Each patch becomes a point in (\mathbb{R}^9) (grayscale values normalized to ([0,1])).
  • The patches form a point cloud.
  • Compute Betti numbers of the underlying topology.

Claimed observation:

  • Betti numbers reported: (\beta_0 = 1), (\beta_1 = 2), (\beta_2 = 1).
  • This matches torus topology but not uniquely, since the Klein bottle is also described as having the same Betti numbers as the torus.

Later claim in transcript:

  • Empirical checking suggests the embeddings actually correspond to a Klein bottle.
  • This motivates Klein-bottle-inspired algorithms for image analysis.

Lists / Methodology Outlines

General Pipeline Implied for TDA in the Examples

  1. Convert each object (molecule/patient/neural recording/image) into a feature representation.
  2. Map each sample to a point in high-dimensional space (\mathbb{R}^d).
  3. Treat the dataset as a point cloud.
  4. Use topological invariants (e.g., connected components, holes, Betti numbers, Euler characteristic) to infer the underlying topology.
  5. Interpret invariant-based structure in the original scientific domain.

Neuroscience Example Pipeline (Grid Cells)

  • Record spikes from many grid cells (Neuropixels).
  • Build a time series for each neuron.
  • Compute pairwise distances between grid cells based on their time series.
  • Build a structure from the distance matrix and compute topology, yielding a torus-like result.

Researchers / Sources Featured (As Named or Explicitly Referenced)

  • Raphaël Kevin Tinarrage (lecturer)
  • Stefanella (appears as a named person in the opening/intro)
  • Leonhard Euler (referred to as “Eiler”; early topology context, including the Konigsberg bridge)
  • August Ferdinand Möbius (surfaces / Möbius-type ideas)
  • Bottleneck / “remaining Boutique” (transcript unclear; intended historical contributors to topological invariants)
  • Poincaré (referenced via “pumpkin picture” and homology-related discussion)
  • Bower (transcript form of a dimension theorem about (\mathbb{R}^n) vs (\mathbb{R}^m) homeomorphism; standard attribution likely Brouwer)
  • Jordan (“Jordan’s theorem” about closed curves in the plane)
  • Atlas / “He said” book (“math is the art of giving the same name to two different things”; source not fully specified in transcript)
  • Neuropixels (technology)
  • Euler characteristic / homology content referenced via a homology paper and a “famous paper” (author not uniquely cited in transcript)
  • Stephen / “Step Up” (author of the natural image space study; unclear naming in transcript)
  • cMYB (protein mentioned; not a researcher)

Original video