CW Complexes: Building Spaces by Attaching Cells

The precise definition: skeleta, attaching maps, and the CW axioms

A CW complex is a Hausdorff space X together with a filtration by skeleta X⁰ ⊆ X¹ ⊆ X² ⊆ ⋯ with X = ⋃ₙ Xⁿ, constructed inductively:

• X⁰ is a discrete set of points (the 0-cells).
• Xⁿ is formed from Xⁿ⁻¹ by attaching a collection of n-cells {eⁿ_α}: pick attaching maps φ_α: Sⁿ⁻¹ → Xⁿ⁻¹ and set Xⁿ = (Xⁿ⁻¹ ⊔ ⨆_α Dⁿ_α) / (x ∼ φ_α(x) for x ∈ ∂Dⁿ_α = Sⁿ⁻¹).

Each n-cell comes with a characteristic map Φ_α: Dⁿ_α → X restricting to φ_α on the boundary and to a homeomorphism onto the open cell eⁿ_α on the interior. The two defining conditions are: (C) closure-finiteness — the closure of each cell meets only finitely many cells; and (W) weak topology — a set A ⊆ X is closed iff A ∩ Φ_α(Dⁿ_α) is closed for every cell. When there are finitely many cells (a finite CW complex), (C) and (W) are automatic and X is compact.

The picture: growing a space one dimension at a time

Think of it as sculpting. You start with a scatter of dots (X⁰). You glue in arcs — each 1-cell is an interval whose two endpoints are slapped onto existing dots — turning the space into a graph (X¹). Then you fill in membranes: each 2-cell is a disk whose rim is wrapped around a loop in the graph. A 3-cell is a solid ball whose boundary sphere is mapped into the 2-skeleton, and so on.

The freedom lives entirely in the attaching maps. Wrapping a 2-cell's boundary once around a circle gives a disk (contractible); wrapping it twice around gives ℝP², the real projective plane; wrapping it around a figure-eight along the word aba⁻¹b⁻¹ gives the torus. Same cells, different glue, wildly different topology.

The weak topology axiom is what lets you test continuity cell by cell: a map out of X is continuous exactly when its restriction to each closed cell is continuous. This is the engine behind nearly every construction of maps out of CW complexes.

The key mechanism: induction over skeleta and cellular approximation

Almost every theorem about CW complexes is proved by induction on the skeleta, using one structural fact repeatedly: attaching an n-cell is a pushout. The square with ⨆ Sⁿ⁻¹ → Xⁿ⁻¹ and ⨆ Sⁿ⁻¹ ↪ ⨆ Dⁿ giving Xⁿ is a pushout of topological spaces, and because Sⁿ⁻¹ ↪ Dⁿ is a cofibration, so is every inclusion Xⁿ⁻¹ ↪ Xⁿ — hence every subcomplex inclusion A ↪ X.

The second engine is the Cellular Approximation Theorem (Whitehead): any continuous map f: X → Y between CW complexes is homotopic to a cellular map (one sending Xⁿ into Yⁿ). The proof pushes the image off the interiors of high-dimensional cells using the fact that a map Dⁿ → Dᵏ with n < k, being smooth after approximation, must miss a point of the open k-cell, and one deformation-retracts away from that point. This is why πₙ(Sᵏ) = 0 for n < k — the cells force it. Skeletal induction plus cellular approximation reduce homotopy-theoretic questions to finite combinatorics of cells and degrees.

Worked example: the CW structures on the sphere Sⁿ

The sphere illustrates how the same space admits very different, equally valid CW structures.

• Two-cell model. Give Sⁿ one 0-cell and one n-cell: Sⁿ = e⁰ ∪ eⁿ, with the n-cell's boundary Sⁿ⁻¹ crushed to the single point by the constant attaching map. Concretely Sⁿ = Dⁿ / ∂Dⁿ. This is the minimal CW structure — just two cells.
• Equatorial model. Alternatively, in each dimension 0 ≤ k ≤ n put two k-cells (upper and lower hemispheres), attaching each along the equator Sᵏ⁻¹. This exhibits the tower S⁰ ⊂ S¹ ⊂ ⋯ ⊂ Sⁿ and generalizes to the infinite-dimensional S^∞ = ⋃ Sⁿ.

The two-cell model makes cellular homology trivial to read off: the chain complex is ℤ (degree 0) → 0 → ⋯ → 0 → ℤ (degree n) with all boundary maps zero, giving H₀ = Hₙ = ℤ and Hₖ = 0 otherwise — exactly the homology of Sⁿ, computed with essentially no work. Real projective space ℝPⁿ = e⁰ ∪ e¹ ∪ ⋯ ∪ eⁿ (one cell per dimension) yields the famous complex ⋯ →2→ ℤ →0→ ℤ →2→ ℤ, recovering its ℤ/2 torsion.

Why the axioms matter: what breaks without C and W

Both letters are load-bearing. Drop the weak topology (W) and pathologies appear. The classic case is the Hawaiian earring — nested circles of radius 1/n through a common point, with the subspace topology from ℝ². It is not a CW complex: a genuine countable wedge of circles ⋁ₙ S¹ (the CW version) has the weak topology, in which every neighborhood of the basepoint contains all but finitely many whole circles only if... precisely the opposite holds — the two spaces have different topologies and non-isomorphic fundamental groups (the earring's π₁ is uncountable and not free).

Drop closure-finiteness (C) and the local structure can degenerate: a cell's closure meeting infinitely many other cells can destroy local contractibility. With both axioms, one proves the crucial regularity facts: every CW complex is normal, compactly generated, locally contractible, and paracompact; compact subsets meet only finitely many cells (the 'compactly supported' lemma); and every CW pair (X, A) has the homotopy extension property — i.e. is a cofibration. These are exactly the hypotheses later theorems silently rely on.

Significance: Whitehead's theorem and what CW structure unlocks

CW complexes are the native habitat of homotopy theory. The headline result is Whitehead's Theorem (1949): a map f: X → Y between connected CW complexes that induces isomorphisms on all homotopy groups πₙ(f) is a homotopy equivalence. This is spectacularly false for general spaces (a weak homotopy equivalence need not be an equivalence), so restricting to CW complexes is what makes 'same homotopy groups ⇒ same homotopy type' true.

Everything downstream leans on it:

• Cellular homology gives chain groups free on the cells, matching singular homology — the standard practical computation tool.
• CW approximation: every space admits a weak equivalence from a CW complex, so homotopy-invariant functors lose nothing by working with CW complexes.
• Obstruction theory, Eilenberg–MacLane spaces K(π,n), classifying spaces BG, spectra, and Postnikov towers are all built cell by cell.
• The Whitehead product, cofiber sequences, and the entire model-category structure on Top take CW/cofibrant objects as their well-behaved core.

In short: CW structure converts topology into algebra you can actually compute with.

CW complexes versus neighboring notions of 'built from pieces'

Notion	Building block	Gluing rule	Topology / regularity
Simplicial complex	Simplices (n-dim triangles)	Faces meet along common subfaces; a simplex is determined by its vertices	Combinatorially rigid; every space needs many simplices
Δ-complex	Simplices with ordered vertices	Attach via affine face maps; faces may be identified	Fewer cells than simplicial; still simplex-shaped cells
CW complex	Cells eⁿ ≅ open n-disks	Attach n-cell by any continuous φ: Sⁿ⁻¹ → Xⁿ⁻¹	Weak topology; closure-finite; Hausdorff
Arbitrary quotient of disks	Disks	Any identification whatsoever	Can fail Hausdorff / weak-topology axioms — not a CW complex