Tangent Spaces and Vector Fields
On a smooth
manifold there is no ambient space to point arrows into — the manifold is all there is.
Yet we still want directions: which way does a particle move, which way does temperature
increase, which way is "downhill"? The answer at each point p is a whole
vector space of directions, the tangent space T_pM. Glue
one such space to every point and let it vary smoothly and you get a vector field —
the object behind fluid flow, force fields, and the differential equations of motion.
The subtlety is defining "a direction" without ever leaving the surface. There are three standard
ways to do it, and — pleasingly — they all describe the same space.
Three pictures of one tangent vector
(i) Velocities of curves. The most physical picture. Take a smooth curve
\gamma : (-\varepsilon, \varepsilon) \to M that passes through
p at time 0, so
\gamma(0) = p. Its velocity there, \gamma'(0),
is a tangent vector. Every direction you can leave p in is the velocity of
some curve; two curves define the same tangent vector when they agree to first order.
(ii) Derivations. The most algebraic picture. A tangent vector is an operator that
eats a smooth function and returns the rate of change of that function in its direction — a
directional derivative. Abstractly, a derivation at
p is a linear map v on smooth functions
obeying the Leibniz product rule
v(fg) = f(p)\,v(g) + g(p)\,v(f).
Anything satisfying linearity and this rule is a tangent vector — no picture of an arrow
required.
(iii) The coordinate basis. The most computational picture. Pick a chart with local
coordinates (x^1, \dots, x^n) near p. The
partial-derivative operators
\left. \frac{\partial}{\partial x^1} \right|_p, \ \dots, \ \left. \frac{\partial}{\partial x^n} \right|_p
form a basis of T_pM. Every tangent vector is a unique combination of
them,
v = \sum_{i=1}^{n} v^i \, \frac{\partial}{\partial x^i} \Bigg|_p ,
with real components v^1, \dots, v^n. There are exactly
n basis vectors — one per coordinate — which is why the tangent space has
dimension n.
At each point p of a smooth n-manifold
M:
-
the tangent space T_pM is a real
vector
space — you can add tangent vectors at p and scale them;
-
its dimension equals the dimension of the manifold,
\dim T_pM = \dim M = n;
-
the coordinate partials \partial/\partial x^i|_p form a basis, so
there are exactly n of them;
-
the three definitions — curve velocities, derivations, and coordinate combinations — all yield
this same space.
Worked examples
The plane. For M = \mathbb{R}^2 with coordinates
(x, y), the tangent space at any point has basis
\partial/\partial x,\ \partial/\partial y, so
T_p\mathbb{R}^2 \cong \mathbb{R}^2. Here every tangent space looks like a
copy of the plane itself — which is exactly why flat space lets you get away with thinking of vectors
as arrows you can slide around freely.
The sphere. For M = S^2 and a point
p, the tangent space T_pS^2 is the plane just
grazing the sphere at p — the flat sheet touching it there. It is
2-dimensional, matching \dim S^2 = 2, and it
tilts as p moves.
A vector field. A vector field X assigns
to each point p a tangent vector X_p \in T_pM,
varying smoothly. In coordinates,
X = \sum_{i=1}^{n} X^i(x) \, \frac{\partial}{\partial x^i},
with smooth component functions X^i. On the plane, the spinning field
X = -y\,\partial/\partial x + x\,\partial/\partial y gives an arrow at
every point circulating counter-clockwise — the velocity field of a rigid rotation. On the circle
S^1, the unit tangent field points the same way all the way round and
never vanishes.
See a tangent vector and a field
Take the circle S^1. At a point p the tangent
space T_pS^1 is the line just touching the circle there — a
1-dimensional space. A curve running around the circle has a velocity
\gamma'(0) lying in that line. Do this at every point and the velocities
assemble into a smooth vector field that never vanishes — something a sphere, by the
hairy-ball theorem, can never manage. Step through: the tangent line, then the velocity vector, then
the whole field.
Moving vectors between points — the differential
A smooth map F : M \to N carries curves to curves, and so carries their
velocities to velocities. This is the pushforward (or differential)
dF_p : T_pM \longrightarrow T_{F(p)}N,
a linear map between tangent spaces. If v = \gamma'(0) then
dF_p(v) = (F \circ \gamma)'(0): push the curve forward, then take its
velocity. In coordinates dF_p is exactly the
Jacobian matrix of F at p. It
is the manifold version of "the derivative is the best linear approximation" — the derivative of a
map between curved spaces is a linear map between their tangent spaces.
A vector field on S^2 is a way of laying a hair flat at every point of a
sphere. The hairy ball theorem (Poincaré, Brouwer) says this is impossible without
a bald spot: every continuous tangent vector field on an even-dimensional sphere must
vanish somewhere. Comb a coconut and you always get at least one cowlick.
The circle escapes this — you can comb S^1 smoothly all the way
round, which is why the figure above has an unbroken field. The obstruction on
S^2 is topological, measured by the Euler characteristic
(\chi(S^2) = 2 \neq 0); the torus, with
\chi(T^2) = 0, can be combed flat. Meteorologically, the theorem
promises that somewhere on Earth the horizontal wind is always exactly zero — the calm eye of a
cyclone that must exist.
A tangent vector belongs to one point. The velocity of a curve at
p lives in T_pM; a velocity at a different
point q lives in a different vector space
T_qM. These are separate spaces, and there is no canonical way
to add a vector at p to a vector at q, or even
to ask whether they are "the same direction".
In flat \mathbb{R}^n we cheat: we slide arrows around and compare them,
because all the tangent spaces are secretly identified with one another. On a curved manifold that
identification does not exist for free. Comparing tangent vectors at different points requires extra
structure — a connection (parallel transport) — and the answer generally
depends on the path you carry the vector along. That path-dependence is precisely
curvature. So resist the urge to subtract X_p from
X_q: without a connection, the expression is meaningless.