Continuity and Homeomorphism

2018. McInnes, Healy, and Melville publish UMAP - an algorithm that visualises 100 000 points in 2D in seconds. The first line of the abstract: 'the theoretical foundations are in algebraic topology'. UMAP searches for a homeomorphism between the topological structure of the data and the plane. Normalizing flows - the generative models powering high-end speech synthesis and density estimation - are homeomorphisms between probability distributions. Not a metaphor. A definition.

• **Normalizing flows (RealNVP, Glow, NICE):** homeomorphism between a Gaussian and a complex distribution - the basis of invertible generative models in speech synthesis (WaveGlow) and density estimation
• **UMAP and t-SNE:** preserving the topological structure of data when projecting to 2D - the manifold hypothesis in action; StyleGAN maps the 512-dimensional W-space as a chart of the manifold of face images
• **Topological Data Analysis (TDA):** persistent homology tracks topological invariants (components, loops, voids) in data - applied in genomics (cancer subtypes), neuroscience, and molecular structure analysis

Continuity

Normalizing flows - the class of generative models behind RealNVP and Glow - work precisely because they can build continuous bijective maps between probability distributions. Their mathematical foundation reduces to one sentence: **f: X -> Y is continuous if the preimage of every open set is open**. No epsilons or deltas. No metric. Just open sets.

**f^{-1}(V) = {x in X : f(x) in V}** - the preimage of the set V. Continuity via preimages is equivalent to epsilon-delta for metric spaces, but applies to ANY topological space - even those without a notion of distance.

A consequence that breaks intuition: if X has the discrete topology, every function from X is continuous - all subsets are already open, there is nothing to check. If Y has the trivial topology, the same holds: the preimages of the empty set and all of Y are always open. Continuity is a property of the pair (space, topology), not of the function alone.

Homeomorphism

2018. McInnes, Healy, and Melville publish UMAP - an algorithm that visualises 100 000 points in 2D in seconds. The abstract opens with: *'the theoretical foundations are in algebraic topology'*. UMAP constructs a fuzzy topological structure of the data and searches for a homeomorphic map of that structure into the plane. Homeomorphism is not an abstraction here. It is the algorithm.

**Homeomorphism = topological isomorphism.** A bijection f: X -> Y where both f and f^{-1} are continuous. Equivalently: f is a continuous bijection that sends open sets to open sets (an open map). Notation: X ≅ Y.

Normalizing flows ARE homeomorphisms. RealNVP, Glow, NICE - each model builds a homeomorphism f: R^n -> R^n between a simple distribution (Gaussian) and a complex one (the data). Invertibility is not an implementation detail. It is the definition. Without f^{-1}, computing the log-likelihood is impossible, and there is nothing to train.

A critical trap: a continuous bijection is not necessarily a homeomorphism. The inverse must also be continuous. Classic counterexample: f: [0, 2pi) -> S^1, f(t) = (cos t, sin t). Continuous and bijective. But f^{-1} is discontinuous at (1, 0): points just below have preimage ~2pi, points just above have preimage ~0. A genuine discontinuity.

Topological Invariants

A topological invariant is a property preserved by every homeomorphism. If two spaces differ in some invariant, they cannot be homeomorphic - no exhaustive search over maps needed. This is the only practical way to prove non-homeomorphism. Persistent homology (TDA) is precisely a machine for computing topological invariants across scales in data.

**Main topological invariants:** 1. **Connectedness** - is the space in one piece? 2. **Compactness** - does every open cover have a finite subcover? 3. **Genus** - number of holes in a surface. 4. **Fundamental group** - algebraic structure of loops. 5. **Number of connected components.** 6. **Euler characteristic** chi = V - E + F.

An invariant is a one-sided tool. Agreement on an invariant does not prove homeomorphism. No complete invariant distinguishing all spaces exists - that is a fundamental result. In TDA this means: persistent homology captures certain invariants (components, loops, voids) but cannot distinguish spaces whose persistence diagrams coincide.

Classical Examples

The manifold hypothesis in ML states: natural images are not random points in a 200x200=40 000-dimensional space. They live on a manifold of dimension roughly 50-100. StyleGAN exploits this: the 512-dimensional W-space is a chart of that manifold. t-SNE and UMAP project the manifold into 2D while preserving its topological structure. Understanding why this works means understanding homeomorphisms in practice.

(0, 1) is homeomorphic to R - a paradoxical fact. A bounded interval is topologically identical to the infinite line. The homeomorphism f(x) = tan(pi(x - 1/2)) continuously stretches (0,1) onto all of R. Length and boundedness are not topological properties - they are metric ones. Topology does not see them.

**Non-topological properties:** length, area, curvature, boundedness. **Topological properties:** connectedness, compactness, number of holes (genus), number of components, fundamental group. Topology is 'rubber-sheet geometry': any continuous stretching or compressing is allowed, but no tearing or gluing.

The surface classification theorem (early 20th century): every compact connected surface is homeomorphic either to a sphere with g handles (orientable, genus g) or to a sphere with k cross-caps (non-orientable). A complete classification of two-dimensional spaces. For manifolds of dimension 3 and higher the analogous problem turned out to be orders of magnitude harder - the Poincare conjecture remained open for 100 years.

Key Ideas

• **Continuity** in topology: the preimage of every open set is open - no metric needed, just open sets
• **Homeomorphism** = bijective continuous map with continuous inverse; X ≅ Y means topological indistinguishability
• Continuous bijection ≠ homeomorphism: the inverse must also be continuous (counterexample: $f: [0, 2\pi) \to S^1$)
• **Topological invariants** (connectedness, compactness, chi, pi_1) are preserved under homeomorphism - the only way to prove non-homeomorphism
• **Normalizing flows** are homeomorphisms; **UMAP** finds a homeomorphism of data into 2D; **TDA** works with invariants

Related Topics

Homeomorphism is the central concept of topology:

• Topological Spaces — Open sets are the language for defining continuity and homeomorphism
• Connectedness and Compactness — The main topological invariants used to distinguish spaces
• Fundamental Group — Algebraic invariant - the next level beyond compactness and connectedness
• Metric Spaces — The setting where continuity is most intuitive and coincides with epsilon-delta

Вопросы для размышления

• Why do length and area not matter in topology? For which problems do they still matter?
• Normalizing flows build homeomorphisms between distributions. Why does continuity of the inverse matter - what breaks without it?
• The classification theorem completely describes compact surfaces. Why did the analogous problem for 3-manifolds take another 100 years (the Poincare conjecture)?

Связанные уроки

• top-01 — Open sets are the language for defining continuity
• top-03 — Connectedness and compactness - the main invariants after homeomorphism
• top-05 — Fundamental group - the next level of topological invariants
• la-15-svd — SVD preserves metric structure - an analogue of homeomorphism in linear algebra
• top-04 — Metric spaces - the setting where continuity is most intuitive
• calc-05-continuity
• calc-03-limits-intro