The bridge from a group to a space
In the previous guide you learned to view a finitely generated group G as a geometric object: pick a finite generating set, build the Cayley graph, and the resulting word metric makes G into a geodesic space. You also saw the catch — change the generating set and the word metric changes, but only up to quasi-isometry, the equivalence that ignores bounded errors and bounded multiplicative stretch. So the honest geometric data of G is its quasi-isometry class, not any one metric. The natural next question is: which familiar spaces share that class? The Švarc-Milnor lemma answers it.
The slogan, sometimes called the fundamental observation of geometric group theory, is this: if a group acts nicely enough on a space, the group is quasi-isometric to that space. Concretely, suppose G acts on a metric space X by isometries, and the action is (1) cocompact — some bounded region's translates cover all of X — and (2) properly discontinuous in a metric sense. Then G with any word metric and X with its metric are quasi-isometric. The group can borrow the large-scale geometry of the space outright.
Why it works: orbits as a net
The proof is more illuminating than the statement, so picture it. Fix a basepoint x_0 in X and look at its orbit, the set of points g·x_0 as g ranges over G. Cocompactness means these orbit points are spread through X like a net with no large empty gaps: every point of X lies within some fixed radius R of an orbit point. So the map g -> g·x_0 from G into X is coarsely onto — its image R-covers X, which is exactly the surjectivity-up-to-bounded-error half of being a quasi-isometry.
The harder half is that this map neither stretches nor crushes distances by more than a bounded factor. Word distance in G counts generators; the trick is to choose a generating set FROM the geometry: let S be the set of group elements that move x_0 by at most 2R + 1, say. Properness guarantees S is finite, so it is a legitimate finite generating set. A short word in these generators moves x_0 a controlled distance (upper bound), and conversely a path in X from x_0 to g·x_0 can be chopped at orbit points spaced under 2R apart, each hop realized by one generator in S (lower bound). Stack the two bounds and you have a quasi-isometry.
The lemma at work: three examples
The cleanest source of nice actions is the deck action of a fundamental group. If M is a compact Riemannian manifold, then pi_1(M) acts on the universal cover tilde-M by deck transformations, and that action is by isometries (lift the metric), cocompact (the cover of a compact base is tiled by translates of a fundamental domain), and proper. The Švarc-Milnor lemma instantly gives: pi_1(M) is quasi-isometric to the universal cover tilde-M. The algebra of the group and the large-scale Riemannian geometry of the cover are the same coarse object.
- Flat torus: the n-torus T^n has universal cover R^n with the deck group Z^n. So Z^n is quasi-isometric to Euclidean R^n — the integer lattice and the continuous plane look identical from far away.
- Hyperbolic surface: a genus g >= 2 surface has universal cover the hyperbolic plane H^2. Its fundamental group is therefore quasi-isometric to H^2, hence is a Gromov-hyperbolic group (guide 3) — negative curvature is exported from the cover to the group.
- A finite-index subgroup H of G is quasi-isometric to G: H acts on G's Cayley graph cocompactly (finitely many cosets) and properly. Quasi-isometry cannot see finite index — a powerful coarseness, and also a real limitation when you want finer invariants.
Growth: counting the ball in the group
Now we mine the geometry for an invariant. The growth function beta(n) counts how many group elements lie within word distance n of the identity — equivalently, how many vertices sit inside the ball of radius n in the Cayley graph. It is the most elementary measurement of how fast the group spreads out. Change the generating set and beta(n) changes, but only by a controlled rescaling of the input; so its growth type — polynomial, exponential, or something between — is a quasi-isometry invariant, hence an honest property of the group.
beta(n) = #{ g in G : |g|_S <= n } ( |g|_S = word length of g )
Z^k : beta(n) ~ n^k polynomial, degree k
free F_2 : beta(n) ~ 3 * 2^n - 2 exponential
Heisenberg : beta(n) ~ n^4 polynomial, degree 4 (not 3!)Walk through the examples. In Z^k the ball of radius n is a discrete diamond with on the order of n^k lattice points, so growth is polynomial of degree k — matching the quasi-isometry Z^k -> R^k from the torus above, since Euclidean balls also grow like n^k. In the free group F_2 on two generators, from each vertex you can step to 3 new vertices without backtracking, so the count roughly triples each step: beta(n) is exponential. The free group is a tree, the most spread-out geometry there is.
The Heisenberg group is the instructive surprise. It is generated by two elements x, y whose commutator z = [x, y] is central, and z requires about n^2 generators to reach (you build z out of an n-by-n grid of x, y moves). Counting carefully, beta(n) grows like n^4, not n^3 as the naive three generators x, y, z would suggest. So growth degree need not be the obvious dimension — a healthy warning against reading the answer off the generating set.
Gromov's theorem and the spectrum of growth
The deepest result here is Gromov's polynomial growth theorem (1981): a finitely generated group has polynomial growth — beta(n) <= C·n^d for some C, d — if and only if it is virtually nilpotent, meaning it has a nilpotent subgroup of finite index. One direction is elementary bookkeeping; the converse is a landmark, and its proof is the prototype of the whole Gromov-Hausdorff limit method from guide 1: rescale the Cayley graph by 1/n, take a limit, and show polynomial growth forces the limit to be a nice space (a nilpotent Lie group with a left-invariant metric) on which the group acts. This is a theorem we STATE and motivate, not prove — the real proof is a course in itself.
Between polynomial and exponential lies a famous subtlety. Milnor asked in 1968 whether any group has intermediate growth — faster than every polynomial yet slower than exponential. For fifteen years no example was known; then in 1984 Grigorchuk built one, a group of automorphisms of a binary tree whose growth sits strictly between n^d and 2^n. So the growth spectrum is genuinely richer than the polynomial/exponential dichotomy, and resist any source that presents only two boxes.
Growth is one of several large-scale invariants the Švarc-Milnor lemma makes geometric. Another is amenability — roughly, the existence of an invariant finitely-additive mean, or equivalently of Følner sets that are almost invariant under translation. Subexponential growth implies amenability, while a free group F_2 is the classic NON-amenable group (its exponential growth and tree structure leave no room for Følner sets). These invariants are why geometric group theory can answer purely algebraic questions with pictures — and they are the reason this rung treated groups as spaces in the first place.