Smooth Manifolds
A sphere, a torus, the surface of a doughnut, the configuration space of a robot arm, the set of
all rotations of space — none of these is flat, yet each one looks flat if you stand close enough.
Zoom in on any patch of the Earth and it looks like a piece of a plane; that is exactly why old
maps work at all. A smooth manifold is the precise mathematical object that
captures this idea: a space that is locally like \mathbb{R}^n,
stitched together smoothly, even though globally it can be curved, closed up, or knotted.
This is the setting for all of differential geometry, general relativity, gauge theory and modern
mechanics. The whole trick is to do calculus on a curved space by borrowing the calculus we already
know on \mathbb{R}^n — one flat patch at a time — and to make sure the
patches agree where they overlap. Everything on this page is machinery for that single promise.
A chart: local coordinates on a patch
Start with a topological space
M. A chart is a way of laying flat coordinates onto a
patch of it. Formally, a chart is a pair (U, \varphi) where
U \subseteq M is an open set and
\varphi : U \longrightarrow \varphi(U) \subseteq \mathbb{R}^n
is a homeomorphism onto an open subset of \mathbb{R}^n —
a continuous bijection with a continuous inverse. The map \varphi hands
every point p \in U a list of n real numbers
\varphi(p) = (x^1, \dots, x^n), its local coordinates.
Latitude and longitude are a chart on (most of) the globe; the n numbers
are what a mapmaker actually prints on the flat page.
The number n — how many coordinates each chart uses — is the
dimension of the manifold. A curve is 1-dimensional, a
surface 2-dimensional, spacetime 4-dimensional.
An atlas: charts that cover everything
A single chart rarely covers all of M — a flat map cannot show the whole
round Earth without tearing or badly distorting it. So we use a whole book of maps. An
atlas is a collection of charts
\mathcal{A} = \{\, (U_\alpha, \varphi_\alpha) \,\}
whose domains cover the space: every point lies in at least one chart,
\bigcup_\alpha U_\alpha = M. Just like a real world atlas: no single page
holds the planet, but between them the pages leave nothing out, and neighbouring pages overlap along
their edges so you can trace a route from one to the next.
Those overlaps are where the real content lives. If a point p sits in two
charts (U, \varphi) and (V, \psi) at once, it
gets two sets of coordinates. To do consistent calculus we need a rule for translating
between them — and that rule must itself be smooth.
Transition maps: changing coordinates smoothly
On the overlap U \cap V, compose one chart with the inverse of the other.
This gives the transition map
\psi \circ \varphi^{-1} : \varphi(U \cap V) \longrightarrow \psi(U \cap V),
a map from an open piece of \mathbb{R}^n to another open piece of
\mathbb{R}^n. Crucially it lives entirely in flat space: it takes
the first chart's coordinates for a point and returns the second chart's coordinates for the same
point. It is exactly a change of variables. An atlas is called smooth (or
C^\infty) when every transition map, in both directions, is
infinitely differentiable wherever charts overlap.
That single requirement is what lets us say a function or a curve on M is
"smooth" without ambiguity: check it in any convenient chart, and the smooth transition maps
guarantee the verdict is the same in every other chart. Smoothness becomes a property of the space,
not of the particular map we happened to open.
A smooth n-manifold is a topological space
M together with a smooth atlas, such that:
-
M is Hausdorff (distinct points have disjoint
neighbourhoods) and second-countable (it has a countable base of open sets) —
the two "no pathologies" hypotheses;
-
it is covered by charts (U_\alpha, \varphi_\alpha), each a
homeomorphism onto an open subset of \mathbb{R}^n — so every point has
n local coordinates;
-
all transition maps \varphi_\beta \circ \varphi_\alpha^{-1} are
C^\infty on the overlaps, so "smooth" is chart-independent;
-
the integer n is the dimension
\dim M = n, the same for every chart on a connected manifold.
See two charts overlap on a circle
The circle S^1 is the smallest interesting example. No single chart can
wrap the whole circle onto an open interval without a break, so we use two arcs, each carrying its
own coordinate. They overlap in two little regions — and on that overlap the
transition map \psi \circ \varphi^{-1} is a smooth
relabelling from one arc's coordinate to the other's. Step through the figure: chart
U, then chart V, then the transition map that
glues their coordinate lines together.
A gallery of manifolds
\mathbb{R}^n itself is a manifold with a single global
chart: the identity map. Nothing to glue, one page suffices — flat space is the trivial example the
whole theory generalises.
The circle S^1 and sphere S^2
genuinely need at least two charts. Stereographic projection from the north pole flattens all of the
sphere except that one pole onto a plane; a second projection from the south pole catches the point
the first one missed. Two charts, overlapping everywhere except the two poles, and their transition
map (an inversion, \mathbf{x} \mapsto \mathbf{x}/|\mathbf{x}|^2) is
smooth. So \dim S^1 = 1 and \dim S^2 = 2.
The torus T^2 — the doughnut surface — is a
2-manifold too, most cleanly described as the product
S^1 \times S^1. Which leads to the tidiest dimension rule of all: for a
product manifold,
\dim(M \times N) = \dim M + \dim N.
A point of M \times N is a pair, so its coordinates are just the
M-coordinates followed by the N-coordinates —
you concatenate the two lists. Hence \dim T^2 = 1 + 1 = 2, and the
three-torus T^3 = S^1 \times S^1 \times S^1 has dimension
3.
It feels like it should be possible to find one clever coordinate system for the entire sphere —
but it never is. Any single chart is a homeomorphism onto an open subset of the plane,
and an open planar region cannot be homeomorphic to the whole sphere, which is compact and has no
boundary. Something always has to be left out or torn.
There is a hairier version of the same obstruction. The hairy ball theorem says
you cannot comb a coconut flat: every continuous tangent vector field on S^2
must vanish somewhere. A single smooth coordinate chart covering the whole sphere would hand you
a nowhere-zero field (the coordinate direction), so its very existence would contradict the hairy
ball theorem. The deep reason a globe needs more than one page is the same reason you always end
up with a cowlick — and it is a fact about the sphere's global shape, invisible from inside any
one flat patch.
The commonest misconception: that a manifold "is" a curved surface sitting inside some big
ambient \mathbb{R}^N, like a sphere floating in three-dimensional space.
It is not. A smooth manifold is an abstract, intrinsic object, defined by nothing
but its topology and its atlas. The charts map from the manifold to
\mathbb{R}^n; the manifold itself lives nowhere but in its own
definition.
This matters enormously. Spacetime in general relativity is a
4-manifold that is not sitting inside a bigger flat space — there
is no "outside" to embed it in. Whitney's embedding theorem does guarantee that any abstract
n-manifold can be embedded in some
\mathbb{R}^{2n} if you want to, but that is a convenience, never part of
the definition. Always think intrinsically: the geometry belongs to the space, not to any room you
might place it in.