Topology and Algebra for ML

Topology studies shape that survives stretching and algebra studies structure preserved by operations, and together they name what data and networks keep under change.

Stretching preserves holes but not distances, and algebraic operations classify the shapes that stretching cannot tell apart

Why Does This Exist?

A coffee cup and a donut share one hole: stretch either and you get the other, though no rigid motion matches them. Data analysis meets the same split constantly. Clustering and manifold learning care about connectivity and holes (topology), while symmetry-aware networks care about operations that leave the task unchanged (algebra). Treating both as "advanced math" misses that they answer different questions: what stays the same under deformation versus under transformation.

This page maps the block. Point-set topology gives open sets and continuity; homotopy and homology count holes; persistent homology tracks holes across scales; group theory axiomatizes symmetry; representation theory linearizes it; category theory compares structures; algebraic geometry describes network function spaces. Manifold foundations used daily are under vectors and tensors.

The one idea you need first

A continuous map preserves closeness: nearby inputs land nearby. Topology keeps exactly the properties such maps preserve (connectedness, holes); algebra keeps the properties operations preserve (closure, inverses, associativity).

Think of It Like This

Sorting buttons two ways

Dump a box of buttons on a table. One sort groups by holes: no-hole discs, two-hole shirt buttons, four-hole coats. Stretching a rubber button never changes its hole count, so the sort survives deformation.

The other sort groups by symmetry: round buttons spin into themselves any amount, square ones only quarter-turns, patterned ones never. Spinning tests operations, not stretching. Both sorts ignore size, color, and position: each keeps one kind of sameness while discarding the rest. Data shapes and network symmetries need the same two independent sorts.

How It Actually Works

Topology half. Open sets define continuity without distances; homeomorphisms (continuous bijections with continuous inverses) are the allowed stretchings. Homotopy deforms paths into each other; homology counts holes per dimension (components, loops, voids). Persistence watches holes appear and die as a distance threshold grows, keeping long-lived ones as signal.

Algebra half. Groups package one operation with identity and inverses (rotations, permutations); group actions move data (rotate the image) while representations realize the moves as matrices on feature vectors. Equivariant layers commute with the action: rotate input, output rotates predictably. Categories lift the view to objects plus structure-preserving maps, so a construction in one domain transports to another.

Worked example

Eight points on a circle of radius $1$ plus one outlier at the center. At threshold $r = 0.4$ , neighbors connect into $8$ separate edges (no loop). At $r = 0.8$ , consecutive ring points link into one $8$ -cycle: one loop is born ( $H_1 = 1$ ). At $r = 1.1$ , the center connects to all ring points, filling the loop: it dies. Persistence pair $(0.8, 1.1)$ , lifetime $0.3$ : a real ring, not noise (noise pairs live under $0.05$ here). Algebra side: the ring's symmetry group is the $8$ -rotation group $C_8$ ; a layer commuting with $C_8$ shares one filter across all rotations instead of learning eight.

Code

1def betti_loop(n_ring: int, threshold: float) -> int:2    # Toy ring test: loop exists once neighbors link but the center has not filled it.3    neighbor_gap = 2 * 3.14159 / n_ring  # arc gap on unit circle4    return 1 if 0.7 * neighbor_gap < threshold < 1.0 else 05
6
7for t in (0.4, 0.8, 1.1):8    print(t, betti_loop(8, t))  # -> 0.4 0, 0.8 1, 1.1 0

Born at $0.8$ , dead at $1.1$ : lifetime $0.3$ marks signal against sub- $0.05$ noise.

Watch Out For

Reading persistence diagrams as cluster counts

Long-lived $H_0$ bars count connected components merging, but the merge order depends on the metric: Euclidean versus cosine filtrations on the same embeddings give different diagrams. Fix the metric from the task (cosine for normalized embeddings) before interpreting any bar as ground truth.

Imposing symmetry the data does not have

Equivariance to rotations helps natural images and harms digit reading (6 versus 9). The group must match the task's true invariances; a larger group is not a stronger result. Test with transformed validation inputs before committing the architecture.

The Quick Version

Topology keeps deformation invariants (components, loops, voids); algebra keeps operation invariants (closure, inverses).
Point-set foundations lead to homotopy, homology, and multiscale persistence.
Groups plus representations give symmetry-aware layers that share weights over orbits.
An $8$ -ring shows a loop with persistence $(0.8, 1.1)$ and symmetry group $C_8$ .
Metric choice shapes every persistence diagram; symmetry choice shapes every equivariant net.
Use the two sorts independently: holes for shape, groups for transformation behavior.