Skip to content

Topology Landscape

The current toolkit starts with topological data analysis: point clouds, Vietoris-Rips filtrations, Betti numbers, persistence diagrams, and time-delay embeddings. That is useful, but it is only one part of topology.

This page is the expansion map. It records the major topology families the toolkit should teach, prototype, or eventually support as stable APIs.

For an auditable status table, see Topology Family Coverage Matrix.

What The Current Stack Covers

The current code and documentation are strongest in:

  • persistent homology over bounded Vietoris-Rips and witness complexes;
  • Betti numbers for connected components, loops, and voids;
  • point-cloud and manifold embeddings;
  • spherical and hyperbolic locality;
  • Voronoi-style routing and memory layout;
  • spectral contraction and Lyapunov-style stability language;
  • topology-guided scheduling and routing as roadmap material.

That is a strong base, but it overrepresents "holes in point clouds" and underrepresents topology as a full language of continuity, covers, maps, symmetry, gluing, dynamics, and computation.

Broad Topology Taxonomy

Wikipedia's overview frames topology as the study of properties preserved under continuous deformation, such as connectedness, compactness, and dimension, and lists major subfields including general topology, algebraic topology, differential topology, geometric topology, and generalized/categorical directions. This project uses that as a broad scaffold, then translates each family into ML and systems terms.

Missing Or Underused Families

Family Object / invariant Why ML or systems should care Track
Point-set / general topology Open sets, neighborhoods, continuity, compactness, separation, connectedness Makes "topological" mean more than holes. Gives precise language for convergence, observability, finite covers, data validity, and continuity of preprocessing pipelines. Docs-only first
Metric, uniform, and proximity spaces Entourages, Cauchy behavior, total boundedness, nearness without one fixed metric Useful for embedding APIs, approximate nearest-neighbor search, drift detection, cache locality, and tolerance contracts across backends. Active prototype for metric covers
Domain theory / Scott topology Posets, directed suprema, monotone maps, fixed points, computable continuity Natural fit for compiler semantics, schedulers, streaming dataflow, abstract interpretation, and monotone runtime proofs. Core later
Homotopy theory Fundamental group, higher homotopy groups, path classes, deformation classes Distinguishes optimization paths and system-state trajectories that homology can collapse into the same count. Useful for "safe deformation" of plans, policies, or schedules. Prototype
Cohomology beyond Betti counts Cocyles, cup products, obstruction classes Encodes interactions and global consistency, not just feature counts. Candidate for feature interaction diagnostics and distributed consistency checks. Prototype
Sheaves and cosheaves Local data glued by restriction maps; sheaf cohomology residuals Direct model for layers, heads, shards, feature stores, sensor networks, and telemetry systems where local views must agree. Active prototype for residual diagnostics
Mapper / Reeb / contour topology Level-set graphs of scalar filters; critical changes in shape Gives interpretable maps of datasets, activations, loss surfaces, entropy filters, and routing states. More approachable to practitioners than raw chain complexes. Active prototype for Mapper graphs
Morse / Conley / dynamical topology Critical points, gradient flows, invariant sets, attractor decompositions Turns training dynamics, online learning regimes, and optimizer stability into objects that can be graphed and falsified. Prototype
Stratified and singular spaces Spaces with strata, boundaries, corners, singularities Models ReLU regions, decision boundaries, failure surfaces, discontinuities, and non-manifold data instead of pretending every dataset is smooth. Prototype
Nerve / Cech / cover calculus Covers and nerves; topology-preserving summaries Generalizes Vietoris-Rips into explicit cover-based batching, cache regions, privacy-preserving summaries, and routing cells. Core later
Knot, braid, and link topology Embedded curves, crossings, linking and braid invariants Candidate for thread interleavings, trajectory entanglement, attention-head coupling, robotics traces, and multi-agent coordination. Docs-only then prototype
Fiber bundles and characteristic classes Fibers over base spaces, sections, holonomy, Chern/Stiefel-Whitney-style obstructions Language for gauge/equivariant ML, layerwise parameter transport, distributed model alignment, and symmetry defects. Docs-only first
Topological groups, group actions, quotients Orbits, stabilizers, quotient spaces Makes symmetry-aware model tests and equivariance contracts explicit; formalizes SO(3), rotations, and quotient embeddings. Prototype
Low-dimensional / geometric topology Orientability, genus, handles, surgery, 3/4-manifold decompositions Good for mesh/scene systems, memory-layout surgery, graph rewrites, and explaining where "topological surgery" claims need real handle calculus. Docs-only first
Categorical, Grothendieck, and pointless topology Sites, locales, topoi, open-set lattices instead of point sets Gives compositional foundations for observability, privacy, distributed knowledge, and sheaf-ready APIs. Docs-only first
Topological vector spaces / function-space topology Weak/strong topologies, convergence of functions and distributions Relevant to model convergence, kernel methods, infinite-width limits, optimizer stability, and distribution shift. Docs-only then prototype

Implementation Priority

The library should not try to implement all of topology at once. That would produce a broad but useless catalog. The right path is:

  1. Document the full landscape. Make each topology family visible with a plain-language explanation and at least one ML/systems use case.
  2. Prototype where a graph helps. Mapper/Reeb graphs, sheaves, metric covers, and dynamical topology should get visual notebooks before hard APIs.
  3. Promote only falsifiable wins to core. A family becomes core only when it has a stable object, invariant, baseline, failure mode, and benchmark gate.

Candidate Milestones

Milestone Topology object First deliverable Promotion gate
Mapper data maps Cover + nerve graph Gallery page for activations and tabular data Beats PCA/UMAP-only diagnostics on interpretability task
Sheaf consistency Local sections over graph Prototype residual over heads/layers/shards Detects injected inconsistency better than scalar variance
Cech/nerve covers Explicit covers Cover-based batching API Same-budget comparison vs Vietoris-Rips and random covers
Dynamical topology Attractors and recurrent sets Training trajectory gallery Separates stable/unstable runs before final metric diverges
Symmetry topology Group actions and quotients Equivariance test utilities Catches symmetry violations missed by ordinary unit tests
Function-space topology Weak convergence summaries Distribution-shift diagnostics Correlates with held-out degradation better than norm drift

Rule For Future Code

Every new topology family must answer five questions before implementation:

  1. What is the topological object?
  2. What invariant or construction do we compute?
  3. What ML or systems decision does it change?
  4. What baseline or null model can falsify it?
  5. What graph makes the idea understandable to a non-topologist?

If any answer is missing, the family stays in docs or gallery until it earns code.