From three model surfaces to the dimension-three question
Guide 1 built the general apparatus: a (G,X)-structure dresses a manifold in a rigid model geometry, a pair (G, X) of a homogeneous space X with its group G of motions, and records its global twisting through the holonomy representation. The natural next question is brutally simple to ask and astonishingly hard to answer: in a fixed dimension, WHICH model geometries (G, X) are there? You cannot allow every (G, X) you like, or the question is empty; you must demand that X be the most symmetric, most efficient model of its kind. The dimension-two answer you already know sets the template, so start there.
For closed surfaces the answer is the uniformization theorem, the deep classical fact that every closed orientable surface carries a metric of constant curvature, and there are exactly three model worlds it can be modeled on: the round sphere S^2 (curvature K = +1, spherical), the flat plane R^2 (K = 0, Euclidean), and the hyperbolic plane H^2 (K = -1, hyperbolic). The Euler characteristic chi decides which: chi > 0 forces spherical (only S^2 and RP^2), chi = 0 forces flat (the torus and Klein bottle), and chi < 0 — every surface of genus g >= 2 — forces hyperbolic. So three geometries partition all surfaces, and 'most' surfaces (every higher genus) are hyperbolic. Hold that pattern: a short finite list, with the negatively curved case dominant.
Maximality: the rule that makes the list finite
Before naming the eight, pin down exactly what counts as a Thurston geometry, because the precise definition is what keeps the list finite — without it you could invent infinitely many. A model geometry is a pair (G, X) where X is a connected, simply connected smooth manifold and G is a group acting transitively on X with compact point stabilizers, equivalently X is a homogeneous space G/H with H compact carrying a G-invariant Riemannian metric. Compact stabilizers is the load-bearing clause: it is exactly what guarantees an invariant metric exists (average any metric over the compact H), so 'geometry' really means 'Riemannian geometry with lots of symmetry,' not some weaker structure.
The decisive extra demand is maximality: we require G to be as large as possible — there must be no strictly larger group G' acting on the SAME X with compact stabilizers. This single clause is what does all the pruning. Take flat 3-space R^3: as a bare homogeneous space you could equip it with the smaller symmetry group of the Nil or the Sol geometry, but those groups sit inside the full Euclidean group, so only the maximal one, Euclidean E^3, counts as a Thurston geometry on R^3. Maximality also forces a finiteness convention: a geometry only earns a place on the list if it admits at least one compact quotient, ruling out symmetric spaces that no closed 3-manifold can model.
The eight geometries, named
Now meet all eight, the Thurston geometries. Three are the constant-curvature models you would guess by analogy with surfaces: spherical S^3 (curvature +1), Euclidean E^3 (curvature 0), and hyperbolic H^3 (curvature -1). Two are honest products of lower-dimensional pieces: S^2 x R (a sphere times a line) and H^2 x R (a hyperbolic plane times a line) — these are the first geometries with no single curvature, since a plane slice tangent to the S^2 or H^2 factor is curved while one containing the R direction is flat. The last three are the genuinely new beasts, twisted bundles that a surface analogy could never predict.
The three exotic geometries are Lie groups wearing left-invariant metrics — this is exactly where the homogeneous-space and Lie-group machinery from earlier in Vol II earns its keep. Nil is the Heisenberg group, the 3x3 upper-triangular matrices with 1's on the diagonal; it is a nontrivial line bundle over the flat plane, the geometry of a discrete Heisenberg lattice. Sol is the lowest-symmetry geometry of the eight, a solvable group in which two directions stretch and shrink exponentially — the model for torus bundles whose monodromy is an Anosov map. And SL~(2,R), the universal cover of the unit-tangent-bundle group of H^2, is a twisted line bundle over the hyperbolic plane. These last three are unit-tangent or line bundles, and their twisting IS the geometry.
GEOMETRY MODEL X curvature / type typical 3-manifold ---------------------------------------------------------------------------- 1 S^3 3-sphere K = +1 (isotropic) finite pi_1 (lens, etc.) 2 E^3 R^3 K = 0 (isotropic) flat: 6 closed types 3 H^3 hyp 3-space K = -1 (isotropic) GENERIC: most 3-mfds 4 S^2 x R sphere x line product, mixed sign S^2-bundle, S^2 x S^1 5 H^2 x R hyp.plane x line product, mixed sign circle bdl over hyp surf 6 SL~(2,R) univ. cover twisted line bundle unit-tangent-bdl type 7 Nil Heisenberg grp nilpotent, K <= 0 nontriv. circle bdl, chi=0 8 Sol solvable group anisotropic, K <= 0 torus bundle, Anosov mono 3 isotropic (S^3,E^3,H^3) + 2 products + 3 Lie-group bundles = EIGHT
Geometrization: cutting a 3-manifold into geometric pieces
The hard truth is that almost no closed 3-manifold wears a single geometry — unlike surfaces, where uniformization gives one geometry per surface, a generic 3-manifold is a mongrel. The geometrization conjecture, Thurston's grand 1982 vision and now a theorem, says you can always cut it into pieces that ARE geometric. The cutting happens in two canonical stages, and both are purely topological — no geometry yet. First, the prime decomposition (Kneser-Milnor): every closed orientable 3-manifold is a connected sum M = P_1 # P_2 # ... # P_k of prime pieces, unique up to order, where prime means it is not itself a nontrivial connected sum. Cutting a connected sum means cutting along essential 2-spheres.
Second, the JSJ decomposition (Jaco-Shalen-Johannson): inside each prime piece there is a canonical, minimal collection of disjoint incompressible 2-tori, unique up to isotopy, and cutting along them splits the prime piece into chunks that are each either Seifert-fibered or atoroidal (admitting no further essential torus). Geometrization then asserts the payoff: after these two cuts, every resulting chunk carries one of the eight geometries as a complete finite-volume structure. The Seifert-fibered chunks use the six geometries with a circle-bundle or product flavor (E^3, S^3, S^2 x R, H^2 x R, Nil, SL~(2,R)); the atoroidal chunks are hyperbolic H^3; and Sol shows up for the special torus bundles. The whole 3-manifold is thus reassembled from eight kinds of geometric brick.
Why does this dignify the hyperbolic 3-manifolds of Guide 3 as the heart of the subject? Because the atoroidal pieces — the generic ones, the chunks left after you have stripped away all the spheres and tori — are exactly the hyperbolic ones, and by the Mostow rigidity coming in Guide 3 their geometry is uniquely forced by their topology. The six 'small' geometries describe rigidly structured pieces (Seifert fibrations, torus bundles) that are well understood; the interesting topological complexity of 3-manifolds concentrates entirely in the hyperbolic part. Geometrization is the statement that there is nowhere else for the complexity to hide.
How it was proved: Ricci flow with surgery, honestly
Thurston conjectured geometrization and proved big cases (the Haken manifolds) by hand, but the full conjecture was settled by Grigori Perelman in 2002-2003 using Hamilton's Ricci flow. The idea is irresistibly physical: treat a Riemannian metric g on M as a heat-like quantity and evolve it by the equation d/dt g = -2 Ric(g), letting the metric diffuse so that curvature smooths out. Where the manifold is too curved it shrinks, where it is too floppy it expands; the hope, vindicated for surfaces, is that the metric flows toward a constant-curvature one. The Ricci flow is, loosely, a nonlinear heat equation for the geometry itself.
But in dimension three the flow does NOT just run forever and settle — it crashes. Regions can pinch off in finite time, forming a neck-pinch singularity where a thin neck shrinks to a point, like a sausage being strangled in the middle. This is not a bug; it is precisely how the topology decomposes. Perelman's deep technical achievement was to understand these singularities completely (his entropy and no-local-collapsing arguments), then perform Ricci flow with surgery: when a neck pinches, you cut the manifold at the neck, cap off the two new boundary spheres with smooth caps, and restart the flow on the resulting pieces. Iterating, the singularities of the flow execute exactly the prime-and-JSJ cutting by themselves, and the surviving pieces converge to the eight geometric structures.
- Start with any Riemannian metric g on the closed 3-manifold M and evolve it by the Ricci flow d/dt g = -2 Ric(g), which diffuses curvature like heat.
- Run until a singularity threatens; classify it using Perelman's entropy and no-local-collapsing estimates — in dimension three the model singularity is a neck pinch on a shrinking S^2 x R neck.
- Perform surgery: cut along the thin neck, cap each new boundary sphere with a standard smooth cap, and discard any small pieces that disappear (these are the S^3 summands of the prime decomposition).
- Restart the flow on the surgered pieces and repeat; on any region surviving to long time the rescaled metric converges to one of the eight Thurston geometries — geometrization, read off from the limit.
Resist the hype, and keep three honest caveats. First, this guide STATES and motivates the result; the actual proof is a year-long graduate course and several hundred pages of hard PDE and comparison geometry — never mistake the slogan for the argument. Second, Ricci flow did not 'solve topology'; it solved geometrization (and as a special case the Poincaré conjecture, where every prime piece is S^3), a precise statement about cutting 3-manifolds, nothing wider. Third, geometrization is sharply a dimension-three theorem: in four dimensions no analogous finite list of geometries exists, and the smooth 4-dimensional Poincaré conjecture remains famously open. The triumph is real and bounded — celebrate it for what it is, not for a victory it did not win.