Connectedness and Compactness

In 2023 a team at Ayasdi analyzed breast cancer records - thousands of patients, hundreds of parameters. Classical clustering found nothing. Then came TDA: build a Vietoris-Rips complex, watch connected components merge and split as the scale grows - persistent homology. The result: a new tumor subtype with high survival rates, invisible to k-means and PCA. The mathematical foundation of that discovery is connectedness and compactness. H_0 homology is the count of connected components. Compactness guarantees the algorithm terminates. Abstractions deliver.

• **TDA / Giotto-TDA** - persistent homology on medical and genomic data: connected components (H_0) and cycles (H_1) as topological features; compactness of the finite point cloud guarantees a finite persistence barcode
• **Manifold hypothesis (UMAP, t-SNE)** - data lies on a compact path-connected manifold in R^n; local connectedness is used to build a fuzzy topological structure before projecting to 2D
• **Extreme value theorem in ML** - if the hyperparameter search domain is compact (closed and bounded), a loss minimum is guaranteed to exist; this is why learning rate schedules often operate on closed intervals

Предварительные знания

• Continuity and Homeomorphism

Connectedness

In 2023, the company Ayasdi applied Topological Data Analysis to medical records and found a subtype of breast cancer that had escaped every classical method. The tool: persistent homology. Take a point cloud, build a Vietoris-Rips complex, watch connected components emerge and merge as the scale grows. H_0 homology is literally the count of connected components. The entire machinery starts from a single definition.

A space is **connected** if it cannot be split into two non-empty open sets. Formally: X is connected if the only subsets that are simultaneously open and closed (clopen) are the empty set and X itself. No middle ground.

**Three equivalent definitions of connectedness:** 1. X is not the union of two non-empty disjoint open sets. 2. The only clopen subsets are the empty set and X. 3. Every continuous function f: X -> {0, 1} with the discrete topology is constant.

The key consequence, used throughout mathematics: **the continuous image of a connected space is connected**. If f: X -> Y is continuous and X is connected, then f(X) is connected in Y. This immediately gives the intermediate value theorem: a continuous function on [a,b] takes every value between f(a) and f(b). In TDA: if data lies on a connected manifold, its projection via UMAP is also connected.

Path Connectedness

The manifold hypothesis in ML states: natural data (images, text, audio) does not live in all of R^40000 but on a compact manifold of dimension 50-100. UMAP and t-SNE exploit this - both assume the manifold is **path-connected**: between any two data points there exists a continuous path along the surface. This is strictly stronger than connectedness.

**Path-connected => connected** (always). **Connected => path-connected** (not always!). Counterexample: the topologist's sine curve - connected but not path-connected. One of the most counterintuitive constructions in analysis.

For manifolds (open subsets of R^n, surfaces) connectedness and path connectedness coincide. Pathological examples like the topologist's sine curve do not arise in ML. This is exactly why the manifold hypothesis holds in practice: real data lies on manifolds, not on theoretical pathologies.

Compactness

Here is what happens inside Giotto-TDA when analyzing medical data: a Vietoris-Rips complex is built from thousands of points, then the radius grows from zero to its maximum. At every step the algorithm is guaranteed to terminate. The reason: compactness. A finite point cloud gives a compact complex - one where **every open cover has a finite subcover**. Without this property the persistence barcode would be infinite.

An **open cover** of X is a collection of open sets {U_alpha} whose union contains X. **Compactness**: for ANY open cover there exist finitely many U_1, ..., U_n that already cover X. This is a statement about all covers, not one specific cover.

The Weierstrass theorem follows: a continuous function on a compact set attains its maximum and minimum. In ML this is the extreme value theorem - if the search domain for parameters is compact, a loss minimum is guaranteed to exist. This is why optimization problems are often stated on closed bounded domains rather than all of R^n.

Heine-Borel and Continuous Maps

The cover definition is abstract. For R^n there is a practical criterion - the **Heine-Borel theorem**: a subset of R^n is compact if and only if it is **closed AND bounded**. This is the working tool: checking closedness and boundedness is far easier than inspecting all possible covers.

**Heine-Borel works ONLY in R^n!** In a general metric space, closed + bounded does not imply compact. Example: R with the metric d(x,y) = min(|x-y|, 1) - the whole space is bounded (diameter = 1) and closed, but NOT compact. Infinity can hide in the topology, not in the distances.

Two key facts about continuous maps. First: **the continuous image of a connected space is connected** - this proves the Bolzano intermediate value theorem. Second: **the continuous image of a compact space is compact** - this gives the Weierstrass theorem immediately. Both hold because continuous maps preserve topological properties.

Key Ideas

• **Connectedness**: cannot be split into two non-empty open sets; continuous image of a connected space is connected - giving the intermediate value theorem
• **Path-connected**: between any two points there is a path; strictly stronger than connected (topologist's sine curve is the counterexample); manifold hypothesis assumes path-connectedness
• **Compactness**: every open cover has a finite subcover; guarantees max/min (Weierstrass), convergent subsequences (Bolzano), finite TDA barcodes
• **Heine-Borel**: in R^n, compact = closed + bounded; fails in general metric spaces

Related Topics

Connectedness and compactness are the main invariants for distinguishing spaces:

• Continuity and Homeomorphism — Connectedness and compactness are preserved under homeomorphisms - tool for distinguishing spaces
• Topological Spaces — Connectedness and compactness are defined via open sets
• Metric Spaces — In metric spaces compactness equals sequential compactness
• Fundamental Group — Connectedness is a prerequisite for a non-trivial fundamental group

Вопросы для размышления

• Why does the Heine-Borel theorem fail for arbitrary metric spaces? Build a counterexample using the metric d(x,y) = min(|x-y|, 1).
• The Cantor set is a closed subset of [0,1]. Is it compact? Connected? Path-connected? Justify each answer.
• Persistent homology (Giotto-TDA) builds a barcode on a finite point cloud. Why does compactness guarantee that the barcode contains only finitely many intervals?

Связанные уроки

• top-02 — Open sets and homeomorphisms are the language of connectedness
• top-04 — Metric spaces give geometric meaning to compactness
• top-05 — Fundamental group counts holes in connected spaces
• fa-01 — Compact operators on Banach spaces - direct generalization
• calc-01-sequences — Sequential compactness is Bolzano-Weierstrass in topological form
• calc-15-convergence
• calc-14-improper