OMN 201612 110645 1 Course - Notes v1

J.M.
Sullivan, TU Berlin A: Curves Diff Geom I, SS 2015
This course is an introduction to the geometry of smooth The length of an arbitrary curve can be defined (Jordan) as
curves and surfaces in Euclidean space (usually R3 ). The lo- total variation:
cal shape of a curve or surface is described in terms of its n
curvatures. Many of the big theorems in the subject such
X
len() = TV() = sup (t ) (t ).
i i1
as the GaussBonnet theorem, a highlight at the end of the t0 <<tn I i=1
semester deal with integrals of curvature. Some of these in-
tegrals are topological constants, unchanged under deforma- This is the supremal length of inscribed polygons. (One can
tion of the original curve or surface. show this length is finite over finite intervals if and only if
Usually not as level sets (like x2 + y2 = 1) as in algebraic has a Lipschitz reparametrization (e.g., by arclength). Lips-
geometry, but parametrized (like (cos t, sin t)). chitz curves have velocity defined a.e., and our integral for-
Of course, by Euclidean space [DE: euklidischer Raum] we mulas for length work fine.)
mean the vector space Rn 3 x = (x1 , . . . , xn ) with the stan- If J is another interval and : J I is an orientation-
dard scalar product [DE: Skalarprodukt] (also called an in- preserving homeomorphism, i.e., a strictly increasing surjec-
ha, bi = a b = ai bi and its associated norm tion, then : J Rn is a parametrized curve with the same
P
ner product)

|a| = ha, ai). image (trace) as , called a reparametrization of . (Note: re-
verse curve : J Rn , (t) := (t) has the same trace in
reverse order orientation reversing reparam.)
A. CURVES For studying continuous curves, its sometimes helpful to
allow reparams that stop for a while (monotonic but not
strictly) or that remove such a constant interval.
Given any interval I R, a continuous map : I Rn is
We instead focus on regular smooth curves . Then if
called a (parametrized) curve [DE: parametrisierte Kurve] in
: J I is a diffeomorphism (smooth with nonvanishing
Rn . We write (t) = 1 (t), . . . , n (t) .

derivative, so 1 is also smooth) then is again smooth
We say is C k if it has continuous derivatives of order up and regular. We are interested in properties invariant under
to k. Here of course C 0 means nothing more than continu- such smooth reparametrization. This is an equivalence rela-
ous, while C 1 is a minimum degree of smoothness insufficient tion. An unparametrized (smooth) curve can be defined as an
for many of our purposes. Indeed, for this course, rather than equivalence class. We study these, but implicitly.
tracking which results require, say, C 2 or C 3 smoothness, we
will use smooth [DE: glatt] to mean C and will typically as- R t For a fixed t0 I we define the arclength function s(t) :=
sume that all of our curves are smooth. t0
|(t)| dt. Here s maps I to an interval J of length len(). If
Examples (parametrized on I = R): is a regular smooth curve, then s(t) is smooth, with positive
derivative s = || > 0 equal to the speed. Thus it has a smooth
(t) = (a cos t, a sin t, bt) is a helix in R3 ; inverse function : J I. We say = is the arclength
parametrization [DE: Parametrisierung nach Bogenlnge] (or
(t) = (t2 , t3 ) is a smooth parametrization of a plane unit-speed parametrization) of . We have (s) = ((s)),
curve with a cusp; so (s(t)) = ((s(t))) = (t). It follows that has constant
speed 1, and thus that the arclength of |[a,b] is b a.
(t) = (sin t, sin 2t) is a figure-8 curve in R2 ; The arclength parametetrization is hard to write down ex-
plicitly for most examples we have to integrate a square root,
(t) = (t, t2 , . . . , tn ) is called the moment curve in Rn . then invert the resulting function. (There has been some work
in computer-aided design on so-called pythagorean hodo-
Define simple [DE: einfache] and closed [DE: geschlossene] graph curves, curves with rational parametrizations whose
curves (and simple closed [DE: einfach geschlossene] curves). speed is also a rational function, with no square root. But this
A smooth (or even just C 1 ) curve has a velocity vector still doesnt get us all the way to a unit-speed parametrization.)
[DE: Geschwindigkeitsvektor] (t) Rn at each point. The The fact that the arclength parametrization always exists,
Rb
fundamental thm of calculus says a (t) dt = (b) (a). however, means that we can use it when proving theorems,
The speed [DE: Bahngeschwindigkeit] of is |(t)|. We say and this is usually easiest. (Even when considering curves
is regular [DE: regulr] if the speed is positive (never van- with less smoothness, e.g., C k , there is a general principle
ishes). Then the speed is a (smooth) positive function of t. that no regular parametrization is smoother than the arclength
(The cusped curve above is not regular at t = 0; the others are parametrization.)
regular.) End of Lecture 13 Apr 2015
The length [DE: Lnge] of is len() = I |(t)| dt. We see
R
Although for an arbitrary parameter we have used the name
Z b Z b t (thinking of time) and written d/dt with a dot, when we use
|(t)| dt (t) dt = (b) (a). the arclength parametrization, well call the parameter s and
a a write d/ds with a prime. Of course, for any function f along
the curve, the chain rule says
That is, a straight line is the shortest path. (To avoid using
vector version of integral inequality, take scalar product with d f ds d f
(b) (a).) = , i.e., f 0 = f/ s = f/||.
ds dt dt
AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

J.M. Sullivan, TU Berlin A: Curves Diff Geom I, SS 2015
Suppose now that is a regular smooth unit-speed curve. Solving for ~ we get the formula
Then its velocity 0 is everywhere a unit vector, the (unit) tan-
gent vector [DE: Tangenten(einheits)vektor] T (s) := 0 (s) to h, i
||2 ~ =
the curve. (In terms of an arbitrary regular parametrization, ||2
we have of course T = /||.)
We should best think of T (s) as a vector based at p = (s), for the curvature of a curve not necessarily parametrized at
perhaps as an arrow from p to p + T (s), rather than as a point unit speed.
in Rn . The tangent line to at the point p = (s) is the line Any three distinct points in Rn lie on a unique circle (or
{p + tT (s) : t R}. (Note about non-simple curves.) (As line). The osculating circle to at p is the limit of such circles
long as has a first derivative at s, this line is the limit (as through three points along approaching p. (Variants with
h 0) of secant lines through p and (s + h). If is C 1 near tangent circles, note on less smoothness, etc.)
s, then it is the arbitrary limit of secant lines through (s + h) Considering T : I Sn1 Rn , we can think of this as
and (s + k).) While velocity depends on parametrization, the another curve in Rn called the tantrix (short for tangent in-
tangent line and unit tangent vector do not. dicatrix) of which happens to lie on the unit sphere. As-
We are really most interested in properties that are also in- suming was parametrized by arclength, the curve s 7 T (s)
dependent of rigid motion. It is not hard to show that a Eu- has speed . Thus it is regular if and only if the curvature of
clidean motion of Rn is a rotation A SO(n) followed by a never vanishes. (Note on curves with nonvanishing curvature
translation by some vector v Rn : x 7 Ax + v. Thus could in R2 versus R3 .)
be considered equivalent to A + v : I Rn , t 7 A(t) + v. End of Lecture 16 Apr 2015
Of course, given any two lines in space, there is a rigid mo-
tion carrying one to the other. To find Euclidean invariants
of curves, we need to take higher derivatives. We define the A1. Plane Curves
curvature vector [DE: Krmmungsvektor] ~ := T 0 = 00 ; its
length is the curvature [DE: Krmmung] := |~|. Now lets consider in particular plane curves (n = 2). We
Recall the Leibniz product rule for the scalar product: if v equip R2 with the standard orientation and let J denote the
and w are vector-valued functions, then (v w)0 = v0 w + v w0 . counterclockwise rotation by 90 so that J(e1 ) = e2 and for
In particular, if v w (i.e., v w 0) then v0 w = w0 v. any vector v, J(v) is the perpendicular vector of equal length
And if |v| is constant then v0 v. (Geometrically, this is just such that {v, Jv} is an oriented basis.
saying that the tangent plane to a sphere is perpendicular to Given a (regular smooth) plane curve , its (unit) normal
the radius vector.) In particular, we have ~ T . vector [DE: Normaleneinheitsvektor] N is defined as N(s) :=
Example: the circle (t) = (r cos t, r sin t) of radius r J(T (s)). Since ~ = T 0 is perpendicular to T , it is a scalar
(parametrized here with constant speed r) has multiple of N. Thus we can define the (signed) curvature
of by N := ~ (so that = |~| = ). For an arbitrary
T = ( sin t, cos t), ~ =
1
(cos t, sin t), 1/r. regular parametrization of , we find
r
det(, )
Given regular smooth parametrization with speed := = .
||3
s = ||, the velocity is T , so the acceleration vector is
From N T and T 0 = N, we see immediately that N 0 =
= T = T + T = T + 2 T 0 = T + 2~.

T . We can combine these equations as
Note the second-order Taylor series for a unit-speed curve T
!0
0 T
! !
around the point p = (0) (we assume without further com- = .
N 0 N
ment that 0 I):
Rotating orthonormal frame, infinitesimal rotation (speed
s2
(s) = p + sT (0) + ~(0) + O(s3 ). ) given by skew-symmetric matrix. The curvature tells us
2 how fast the tangent vector T turns as we move along the curve
These first terms parametrize a parabola agreeing with to at unit speed.
second order (i.e., with the same tangent and curvature vec- Since T (s) is a unit vector in the plane, it can be expressed
tor). Geometrically, it is nicer to use the osculating circle [DE: as (cos , sin ) for some = (s). Although is not uniquely
Schmiegekreis], the unique circle agreeing to second order (a determined (but only up to a multiple of 2) we claim that we
line if ~ = 0). It has radius 1/ and center p + ~/2 . Thus we can make a smooth choice of along the whole curve. Indeed,
can also write if there is such a , its derivative is 0 = . Picking
R s any 0
such that T (0) = (cos 0 , sin 0 ) define (s) := 0 + 0 (s) ds.
(s) = p + cos(s)~/2 + sin(s)T/ + O(s3 ). This lets us prove what is often called the fundamental the-
orem of plane curves [DE: Hauptsatz der lokalen Kurventhe-
(Constant acceleration or second derivative gives a parabola orie] (although it really doesnt seem quite that important):
its points are equivalent by shearing. Constant curvature gives Given a smooth function : I R there exists a smooth
a circle its points are equivalent by rotation.) unit-speed curve : I R with signed curvature ; this

curve is unique up to rigid motion. First note that integrat- 3. thus c K contains at least two points, and if there are
ing gives the angle function : I R (uniquely up to a only two they are antipodal on c.
constant of integration), or equivalently gives the tangent vec-
tor T = (cos , sin ) (uniquely up to a rotation). Integrating T For the case of a curve , by our previous remark, at any point
then gives (uniquely up to a vector constant of integration, of c the curvature of is at least that of c.
that is, up to a translation).
Now suppose is a closed plane curve, that is, an L- End of Lecture 20 Apr 2015
periodic map R R2 . As above, we get an angle func-
tion : R R but this is not necessarily periodic. Instead, Now lets prove the theorem. Let the curvature of the cir-
(L) = (0)+2n for some integer n called the turning number cumcircle c be k. If c includes an arc, there is nothing
[DE: Umlaufzahl] (or rotation index or . . . ) of . (It follows to prove. Otherwise suppose c includes at least n 2
that (s + kL) (s) = kn for any integer k and any s.) Note points pi . (At these points i.) We claim each arc i be-
tween consecutive pi and pi+1 contains a point with < k.
RL
that the total signed curvature 0 ds of is 2n. (If we
reverse the orientation of we negate the turning Then it also contains a vertex qi (a local minimum of ) with
R number.) < k. Since the arc from qi1 to qi includes the point pi with
We will later prove Fenchels theorem that ds 2 for
any curve in Rn (with equality only for convex plane curves). k is also includes a vertex p0i (a local maximum of
A standard result is the Jordan curve theorem, saying that a with k. Thus we have found 2n 4 vertices as desired.
simple closed plane curve divides the plane into two regions, To prove the claim, consider the one-parameter family of
one of which (called the interior [DE: Innere]) is bounded. circles through pi and pi+1 (with signed curvatures decreasing
Assuming the curve is oriented so that its interior is on the left, from k). The last one of these to touch i is tanget to i at at
then the theorem on turning tangents, more often known least one interior point, and since i stays to the right of that
even in English by the German name Umlaufsatz, say that its circle, its signed curvature is even less.
turning number is always +1. (This is a special case of the Where did we use the fact that the curve is simple? (Re-
GaussBonnet theorem, needed as a lemma for the general call that the theorem fails without this assumption!)
case, so we will give a proof later.) When two curves are tangent at p and dont cross locally,
we got an inequality between their signed curvatures. But this
assumes their orientations agree at p. By the Jordan curve the-
A2. The Four-Vertex Theorem orem, a simple curve bounds a compact region K. Clearly,
and K have the same circumcircle c. If both curves are
oriented to have the compact regions to the left, then these
A vertex [DE: Scheitelpunkt] of a plane curve is an extremal
orientations agree. Similarly, further application of the Jordan
point of , that is a point where achieves a local minimum
curve theorem ensure that the oriented circular arc from pi to
or maximum, so that 0 = 0. Since any real-valued function
pi+1 used above agrees in orientation with i .
on a compact set achieves a global minimum and maximum,
any curve has at least two vertices.
Note that the osculating circle to at p crosses at p un-
less p is a vertex. Most peoples sketches of osculating cicles
A3. Evolutes and the Nesting Theorem
are wrong! More generally, suppose and are two regular
curves with the same tangent at p. Then stays to the left of
in some neighborhood of p if > at p. And conversely of Given a curve : I Rn with nonvanishing curvature, its
course if stays to the left, then at least the weak inequality evolute [DE: Evolute] : I Rn is the curve of centers of
holds at p. osculating circles: (t) := (t) + ~(t)/(t)2 . Let us consider
The Four-Vertex Theorem [DE: Vierscheitelsatz] says that in particular a unit-speed plane curve with = > 0 and
any simple closed plane curve has at least four vertices. write r = 1/ for the radius of curvature. Then the evolute is
(Note counterexample r = 1 + 2 sin in polar coords if curve (s) = (s)+r(s)N(s). Its velocity is 0 = T +r0 N +rN 0 = r0 N,
not embedded.) We give a proof due to Bob Osserman (1985). so its speed is |r0 (s)|. (The evolute is singular where has
a vertex.) The acceleration of the evolute is r00 N + r0 N 0 =
Lemma: Given a compact set K in the plane (which might
r00 N r0 T/r, so its curvature is r|r10 | .
be the trace of a curve ) there is a unique smallest circle
c enclosing K, called the circumscibed circle [DE: Umkreis]. Now consider a planar arc with strictly monotonic, non-
(Existence via compactness of an appropriately bounded set of vanishing curvature. By the formula above, its evolute also
circles; uniqueness by constructing smaller circle containing has nonvanishing curvature, so in particular,R the distance
s2
intersection of two given ones.) |(s1 ) (s2 )| is stricly less than the arclength s |r0 (s)| ds =
R 1
Properties: r0 ds = r(s1 ) r(s2 ). This simply says the distance be-

tween the centers of two osculating circles to is less than the
1. c must touch K (for otherwise we could shrink c). difference of their radii, that is, the circles are strictly nested.
This is the nesting theorem of Tait (1896) and Kneser (1914):
2. cK cannot lie in an open semicircle of c (for otherwise the osculating circles along a planar arc with stricly mono-
we could translate c to contradict (1). tonic, nonvanishing curvature are strictly nested.

A4. The Isoperimetric Inequality for line segments with some constant c in place of /2. (We
could easily compute c = /2 by integrating a trig function,
Another global result about plane curves is the isoperimet- but wait!) Next, by summing, it holds for all polygons (with
ric inequality [DE: isoperimetrische Ungleichung]. If an em- the same c). Finally, it holds for smooth curves (or indeed
bedded curve of length L bounds an area A then 4A L2 . for all rectifiable curves) by taking a limit of inscribed poly-
(Equality holds only for a circle.) gons. (To know we can switch the averaging integral with
If R R2 is the region enclosed by the simple closed (C 1 ) the limit of ever finer polygons, we can appeal for instance
curve : [a, b] R, (t) = (x(t), y(t)), then we have by to Lebesgues monotone convergence theorem.) To compute
Greens theorem c = /2 it is easiest to consider the unit circle with length 2
Z Z Z b Z b and constant projection length 4.
A= dx dy = x dy = xy dt = y x dt. Note that everything we have said also works for curves in
R a a Rn (projected to lines in different directions) only the value
of c will be different. Similarly, for an appropriate c = cn,k
(Actually, the formula gives an appropriately defined alge-
we get that the length of a curve in Rn is c times the average
braic area even if the curve is not simple; no change if
length of projections to all different k-dimensional subspaces.
parametrization backtracks a bit.)
The trick suggested by Erhard Schmidt (1939) to prove the For any plane curve , the length of u is at least twice the
isoperimetric inequality is to consider an appropriate compar- width of in the direction u. If is a convex plane curve, we
ison circle. We deal with a smooth curve . First find two have equality, so CauchyCrofton says the length is times
parallel lines tangent to such that lies in the strip between the average width. For instance any curve of constant width 1
them. Choose coords so make them the vertical lines x = r. (like the Reuleaux triangle on an equilateral triangle of side
(Here 2r is the width of in the given direction.) Parametrize length 1, named after Franz Reuleaux, Rector at TU Berlin in
by arclength over [0, L] by (x(s), y(s)) and parametriz the the 1890s) has length . A unit square has minimum width 1
and maximum width 2; since its length is 4, the average
p (s) = (x(s), y(s)): same x(s)
circle of radius r over [0, L] by
width is 4/.
as for , and thus y(s) = r2 x(s)2 . (Note about non-
convex curves, etc.) Writing the various different lines perpendicular to u as
End of Lecture 23 April 2015
Rù,a := {x : hx, ui = a} for a R, we see that len u =
#( ù,a ) da. Thus CauchyCrofton can be formulated as
Note that the unit normal vector to is
R LN = (y , x ), so
0 0 R
hN(s), (s)i = xy + yx . We have A = 0 xy ds and r2 =

0 0 0 2Z
Z
1
len = # ù(),a da d.

RL
0 yx0 ds. Thus 4 0 R
Z L Z L Z L
A + r2 = xy0 yx0 ds = hN, i ds |N| || ds = Lr.
0 0 0 A6. Fenchels theorem
Thus by the arithmetic-geometric mean inequality,

Fenchels theorem says the total curvature of any closed
Ar2 (A + r2 )/2 Lr/2. curve in Rn is at least 2. (Equality holds only for convex
plane curves.) To prove this for C 1 curves , recall that the
Squaring and dividing by r2 gives the isoperimetric inequality. tantrix T (s) has speed (s) and thus its length is the total cur-
It is not hard to check that if all these inequalities hold with vature of . On the other hand, the tantrix lies in no open
equality, then must be a circle. hemisphere of Sn1 , for if we had hT (s), ui > 0 for all s then
we would get
A5. The CauchyCrofton Formula Z L D Z L E

0< hT (s), ui ds = u, T (s) ds
0 0
Given a unit vector u = u() = (cos , sin ) S1 R2 , the = u, (L) (0) = hu, 0i = 0,

orthogonal projection to the line in direction u is u : R2
R2 , x 7 hx, ui u. If : I R2 is a smooth plane curve, then a contradiction. Fenchels theorem is thus an immediate corol-
u = u is its projection (usually not regular!). lary of the theorem below on spherical curves.
The CauchyCrofton formula says the length of is /2
We will state all results for general n, but on first reading
time the average length of these projections. By average
one should probably think of the case n = 3 where lies on
length we mean
the usual unit sphere S2 R3 .
? ? 2 Z 2 To investigate spherical curves in more detail note first that
1
len(u ) du = len(u() ) d := len(u() ) d for points A, A0 Sn1 Rn the spherical distance (the length
S1 0 2 0
of the shortest spherical path, a great circle arc) between them
To prove this, first note that if is a line segment, the aver- is
age projected length is independent of its position and orienta-
(A, A0 ) = arccos A, A0 = 2 arcsin |A A0 |/2 .

tion and proportional to its length. That is, the theorem holds

The points are antipodal [DE: antipodisch] if A = A0 (i.e., one max. At each intermediate height, there are exactly two
= , . . . ). A nonantipodal pair is connected by a unique points of . Joining these pairs by horizontal segments gives
shortest arc, with midpoint M = (A + A0 )/|A + A0 |. an embedded disk spanning , showing it is unknotted. For
Lemma: suppose A, A0 nonantipodal with midpoint M; sup- a knotted curve, every height function must have at least four
pose (X, M) < /2. Then 2(X, M) (X, A) + (X, A0 ). critical points, meaning four intersections of the tantrix with
(This can be used to show that the distance from X to a great every great circle.
circle is a convex function when resticted to the semicircle
where the distance is at most 2.)
To prove this, first note that A, A0 , X all lie in some three di- A7. Schurs comparison theorem
mensional subspace of Rn , so we work there, and in particular and Chakerians packing theorem
on S2 . Consider a 2-fold rotation around M, taking A to A0
and X to some point X 0 . Using the triangle inequality and the Schurs theorem is a precise formulation of the intuitive
symmetry, we get idea that bending an arc more brings its endpoints closer to-
gether.
2(X, M) = (X, X 0 ) (X, A) + (A, X 0 ) = (X, A) + (X, A0 ) Suppose is an arc in Rn of length L, and consider a com-
parison arc in R2 Rn of the same length, such that with
as desired.
respect to a common arclength parameter s, the curvature of
End of Lecture 27 Apr 2015
is positive and everywhere at least that of of : (s) (s).
Theorem: Suppose is a closed curve on Sn1 of length Assuming that with its endpoints joined by a straight seg-
L < 2. Then is contained in some spherical cap {x Sn1 : ment gives a convex (simple closed) curve, we conclude that
(X, M) L/4} of (angular) radius L/4 < /2, and in particu- its endpoints are closer:
lar in some open hemisphere. (Note as promised that Fenchel
is an immediate corollary.) (L) (0) (L) (0).
To prove this, pick two points A, A0 on dividing the ar-
clength in half. Then (A, A0 ) L/2 < . Let M be the Proof: by convexity, we can find s0 such that the tangent
midpoint and let X be any point on . If (X, M) < /2, then T 0 := T (s0 to is parallel to (L) (0). Move by a rigid
by the lemma, motion so that (s0 ) = (s0 ) and they share the tangent vector
T 0 there. We have
(X, M) (X, A) + (X, A0 ) /2 len(AXA0 )/2 = L/4.

Z L
(L) (0) (L) (0), T 0 =

Thus the distance from M to any point on is either at most T (s), T 0 ds,
0
L/4 or at least /2. By continuity, the same possibility holds
for all X; picking X = A we see it is the first possibility. while for , our choice of T 0 gives equality:
There are of course other approaches to proving Fenchels Z L
theorem. One goes through an integral geometry formula
(L) (0) = (L) (0), T 0 =

analogous to our last version of CauchyCrofton. (Well state T (s), T 0 ds,
0
it just for curves in S2 but it holds with the same constant
in any dimension.) For u S2 , the great circle u is the set

Thus it suffices to show T (s), T 0 T (s), T 0 (for all s).
of points orthogonal to u. Then the formula says the length We start from s0 (where both sides equal 1) and move out in
of equals times the average number of intersections of either direction. While T moves straight along a great circle
with these great circles. (When itself is a great circle, this is with speed , a total distance less than , we see that T moves
clear, since there are always 2 intersections.) at slower speed and perhaps not straight. Thus is geometri-
First note that the length of a spherical curve is the limit cally clear that T is always closer to the starting direction. In
of the lengths of spherical inscribed polygons (made of great formulas,
circle arcs). (Indeed the spherical inscribed polygon always Z s Z s
has length larger than the euclidean polygon with the same
T (s), T 0 = cos ds cos ds T (s), T 0 .

vertices, which is already approaching the lenth of the curve s0 s0
from below.) Then just as for CauchyCrofton, we check this
ds is the length of the
R
formula first for great circle arcs, then for polygons and then (The last inequality
follows since

by a (trickier) limiting argument for smooth curves. tantrix, while arccos T (s), T 0 is the distance between its end-
With this formula, one can prove Fenchels theorem for points.)
smooth curves by considering height functions h(s), ui. Each Note that this same proof can be made to work for arbitrary
has at least two critical points (min, max), but critical points curves of finite total curvature. The case of polygonal curves
satisfy T (s) u . That is, the tantrix intersects every great is known as Cauchys arm lemma and was used in his proof
circle at least twice, and thus has length at least 2. (1813) of the rigidity of convex polyhedra, although his proof
Without giving precise definitions about knots, we can of the lemma was not quite correct.
understand the FryMilnor theorem: a nontrivally knotted Chakerian proved the following packing result (which again
curve in R3 has total curvature at least 4. For suppose for can be generalized to all curves although we consider only
some height function h(s), ui there was only one min and smooth curves): A closed curve of length L in the unit ball

in Rn has total curvature at least L. To check this, simply very useful, for instance in computer graphics when drawing
integrate by parts: a tube around a curve. One disadvantage is that along a closed
Z Z Z Z curve, a parallel framing will usually not close up.
len = hT, T i ds =

,~ ds || ds ds.
The second special framing comes from prescribing V = 0,
i.e., ~ = U U. That is, U should be the unit vector in the direc-
tion ~. Here the disadvantage is that things only work nicely
What about nonclosed curves? We just pick up a boundary
for curves of nonvanishing curvature , 0. Assuming this
term in the integration by parts, and find that length is at most
condition, we rename U as the principal normal [DE: Haupt-
total curvature plus 2.
normaleneinheitsvektor] N and V as the binormal [DE: Bi-
normaleneinheitsvektor] B and call {T, N, B} the Frenet frame
A8. Framed space curves
[DE: Frenet-Rahmen]. We have
0
T 0 0 T

We now specialize to consider curves in three-dimensional N = 0 N ,

space R3 . Just as for plane curves we used the 4-fold rotation B 0 0 B
J, in 3-space we will use its analog, the vector cross product.
Recall that v w = w v is a vector perpendicular to both v where (s) is the curvature and (s) is called the torsion [DE:
and w. Torsion] of . In terms of a unit-speed parametrization, we
A framingRahmen along a smooth space curve is a have 0 = T , 00 = T 0 = ~ = N, so N = ~/. Finally,
(smooth) choice of a unit normal vector U(s) at each point N 0 = T + B so = hN 0 , Bi = |N 0 + T |. The expansion of
(s). Defining V(s) := T (s) U(s) we have an (oriented) or- the third derivative in the Frenet frame is
thonormal frame {T, U, V} for R3 at each point of the curve,
000 = (N)0 = 0 N + N 0 = 2 T + 0 N + B.
and the idea is to follow how this frame rotates. As before,
expressing the derivatives in the frame itself gives a skew- Expressions in terms of an arbitrary parametrization of
symmetric matrix: with speed (t) are left as an exercise. Here the nonvanishing
0 curvature condition just says that and are linearly indepen-
U V T

T 0 dent, so that {, , } is an oriented basis. The orthonormal
U 0 U U .
U =
frame {T, N, B} is the result of applying the GramSchmidt
V V U 0 V process to this basis.
Of course N and B span the normal plane [DE: Nor-
Here U , V and U are functions along the curve which de-
malebene] to at p = (s). The curve stays to second order in
pend on the choice of framing. We see that T 0 = ~ =
the osculating plane [DE: Schmiegeebene] spanned by T and
U U + V V, so these are just the components of the curva-
N, which contains the osculating circle. The plane spanned by
ture vector in the chosen basis for the normal plane. (And
T and B is called the rectifying plane [DE: Streckebene] (since
2 = U2 + V2 .) The third function U measures the twisting or
the projection of to that plane has curvature vanishing at p).
torsion of the framing U.
The Taylor expansion of to third order around p = (0) is
End of Lecture 30 Apr 2015
3
2 3
3
Sometimes in physical problems a framing is given to us by (s) p + s s6 2 T + s2 + s6 0 N + s6 B
material properties of a bent rod. Mathematically, the curve
might lie on a smooth surface in space; then we often choose where of course T , N, B, , and 0 are all evaluated at s = 0.
U to be the surface normal so that the conormal V is (like T ) Exercise: look at the projections to the three planes above,
tangent to the surface. (We will explore such Darboux frames and see which quadratic and cubic plane curves approximate
[DE: Darboux-Rahmen] in detail when we study surfaces.) them.
But when no external framing is given to us, there are two The fundamental theorem of space curves says that given
ways to choose a nice framing such that one of the entries in functions , : I R with > 0 determine a space curve
the matrix above vanishes. The first has no twisting (U = 0), (uniquely up to rigid motion) with that curvature and tor-
and such a {T, U, V} is called a parallel frame or Bishop frame. sion. This is basically a standard theorem about existence and
Given any U0 at (s0 ) we want U 0 to be purely tangential, uniqueness of solutions to an ODE. For any given {T 0 , N0 , B0 }
indeed the matrix ODE above has a solution, which stays orthonor-
mal and thus gives a framing. (Changing the initial condition
U 0 = U T = ~, U T. just rotates the frames by a constant rotation.) As in the case of

plane curves, integrating T (s) recovers the curve (uniquely
But this ODE has a unique solution. Since it prescribes U 0 up to translation).
U the solution will have constant length, and since hU 0 , T i = Example: a curve with constant curvature and torsion is a
hT 0 , Ui, the solution will stay normal to T . If we rotate a helix. Its tantrix traces out a circle on S2 at constant speed .
parallel framing by a constant angle in the normal plane (that Any curve whose tantrix lies is a circle on S2 (i.e., makes con-
is, replace U by cos U + sin V) then we get another parallel stant angle with some fixed vector u) is called a generalized
framing (corresponding to a different U0 ). Indeed any two helix. Exercise: this condition is equivalent to / being con-
parallel framings differ by such a rotation. Parallel frames are stant.

Suppose now is a unit-speed curve with > 0. If {T, N, B} In a parallel framing, these are the entries of the top row.
0
is the Frenet frame and {T, U, V} is a parallel frame, then how That
P is, the curvature vector T is an arbitrary combination
are these related? We have of course i Ei of the normal vectors Ei , but each of them is parallel
with derivative i T only in the tangent direction. Given any
cos sin U
! ! !
N framing at an initial point, solving an ODE gives us a parallel
=
B sin cos V frame along the curve.
The generalized Frenet frame exists only under the (some-
for some = (s). Then ~ = N = cos U + sin V what restrictive) assumption that the first n 1 derivatives ,
meaning that U = cos and V = sin . Differentiating , . . . , (n1) are linearly independent, and {T, E2 , . . . , En } is
B = sin U + cos V gives then the GramSchmidt orthonormalization of these vectors.
For this frame, it is only the matrix entries just above the di-
N = B0 = 0 cos U + sin V +0T = 0 N

agonal that are nonzero. Thus
so that 0 = or = ds. (The constant of integration
R
Ei0 := i Ei+1 i1 Ei1
corresponds to the freedom to rotate the parallel frame.) We
see that the twisting or torsion of the Frenet frame really In particular T 0 = 1 E2 so 1 = is the usual curvature and E2
does give the rate 0 at which it rotates relative to the twist- is the principal normal (the unit vector in the direction of ~).
free Bishop frame. Sometimes it is useful to use a complex The i are called Frenet curvatures. A fundamental theorem
curvature (s)ei(s) = U (s) + iV (s). Well defined up to global says that for any functions i (s) with i > 0 for i < n 1, there
rotation by ei0 in the complex plane (corresponding again to is a curve with these Frenet curvatures; it is unique up to rigid
the freedom to rotate the parallel frame). motion.
It is clear that a space curve lies in a plane if and only if
0, if and only if is constant, if and only if the complex
curvature stays on some fixed line through 0.
As another example, the complex curvature of a helix traces
out the circle |z| = at constant speed.
Bishop (1975) demonstrated the usefulness of the parallel
frame by characterizing (C 2 regular) space curves that lie on
some sphere. Indeed, lies on a sphere of radius 1/d if and
only if its complex curvature lies on a line at distance d from
0 C. In an appropriately rotated parallel frame, this line
will be the line U d. (The characterization in terms of the
Frenet frame is more awkward, needing special treatment for
points where and 0 vanish.)
To prove this, note that by translating and rescaling we can
treat the case of S2 , i.e., h, i 1. It follows that T
so U := is a framing of itself. From 0 = T we see that this
framing is parallel. That is, U = , V = T is a Bishop
frame. The equation U 0 = T means U 1, as desired. (Note
that since the position vector on S2 is also the normal vector to
the spherical surface, {T, U, V} is also the Darboux frame for
S2 !) Conversely, suppose has a parallel frame {T, U, V}
with U 1, i.e., U 0 = T . Then U is a constant point P,
meaning lies on the unit sphere around P.
End of Lecture 4 May 2013
A9. Framings for curves in higher dimensions
A framing along a smooth curve in Rn is a choice of ori-

ented orthonormal basis {E1 , E2 , . . . , En } at each point of ,
where E1 (s) = T (s) is the unit tangent vector, and each Ei (s)
is a smooth function. Of course the other Ei (for i 2) are
normal vectors. The infinitesimal rotation of any framing is
given, as in the three-dimensional
case, by a skew-symmetric
matrix, here determined by the n2 entries above the diagonal.
Again it is helpful to choose special framings where only n1
of these entries are nonzero.

J.M. Sullivan, TU Berlin B: Surfaces Diff Geom I, SS 2015
B. SURFACES spiral examples). If it is, then of course its image is a smooth

submanifold. Given an immersion x, any point in U has a
Given an open set U Rm and a map f : U Rn we write neighborhood V such that x|V is a homeomorphism.
D p f for the derivative of f at p U, the linear map Rk Rn Given an immersion x parametrizing a surface M, the span
such that f (p + v) f (p) + D p f (v). We say p U is a of xu and xv (the image of Dx) is two-dimensional, and is
critical point of f if D p f is not surjective (which is automatic called the tangent plane T p M at p = x(u, v). Orthogonal to
if m < n); and we say q = f (p) Rn is a critical value. this is the normal line N p M, spanned by xu xv .
Any other q Rn is called a regular value of f . We say f Note that we typically blur the distinction between a point
is an immersion if D p f is injective at every p (which requires (u, v) U and its image p = x(u, v) M = x(U). We write,
m n). for instance, xu (p) = xu (u, v) interchangeably.
Intuitively, a subset M Rn is a smooth embedded k- Example (surfaces of revolution): Suppose we have a regu-
dimensional submanifold if every point p M has an open lar curve (t) = (r(t), 0, z(t)) in the x > 0 half of the xz-plane.
neighborhood U Rn in which M looks like an open set in Consider the map
Rk . From analysis we recall several equivalent precise formu-
x(t, ) = r(t) cos , r(t) sin , z(t) .

lations:
1. Diffeomorphism: There is an open V Rn and a This is an immersion (domain I R, but injective only on
diffeomorphism : U V Rn taking U M to smaller pieces), parametrizing a surface of revolution. Con-
V (Rk {0}). sider injectivity issues, tangent, normal, etc. see homework.
Example (ruled surface): A surface swept out by straight
2. Level set: There is a smooth map h : U Rnk such lines (a ruled surface) can be parametrized by a base curve
that 0 is a regular value of h and U M = h1 (0). (t) and a director field (t) by setting x(t, u) = (t) + u(t).
End of Lecture 7 May 2015
3. Parametrization: There is an open set V Rk and a
smooth immersion f : V Rn that is a homeomor-
phism from V onto U M. B1. Curves, length and area
k
4. Graph: There is an open set V R and and a smooth
map h : V Rnk such that U M is the graph of h How do we describe a curve in a surface M? If x : U R3
up to permutation of coordinates in Rn . is a surface patch and : I U R2 is a (regular smooth)
curve in U then = x is a (regular smooth) curve in R3
Here of course, a diffeomorphism gives a level set represen- lying on the surface M. Conversely, any curve on M can be
tation, and a graph is a special kind of parametrization. described this way. We postpone a discussion of the details
For the rest of this semester, we will consider surfaces and of the effects of changing coordinates (to an overlapping
(k = 2) in R3 . An example is the graph of a smooth func- surface patch).
Writing (t) = (u(t), v(t)) we have = (u, v) and by the

tion f : U R, parametrized by (u, v) 7 u, v, f (u, v) or
given as the zero-set of F(x, y, z) := f (x, y) z. chain rule the velocity vector of = x is thus uxu +
Another example would be the unit sphere S2 , the level vxv . We see that the tangent plane T p M spanned by {xu , xv } is
set x2 + y2 + z2 = 1. It can be covered by six open hemi- exactly the set of all velocity vectors to curves in M through
spheres on which it is a graph in one of the coordinate direc- p = x(u, v).
tions. Using stereographic projection we can parametrize all The speed of is of course given by the Euclidean norm
but a single point of the sphere by an immersion from R2 . of its velocity vector; the tangent space T p M inherits an inner
The usual geographic coordinates (lattitude and longitude) product h, i as a subspace of T p R3 = R3 . The basis {xu , xv }
(, ) 7 cos cos , sin cos , sin give an immersion

is of course in most cases not orthonormal. The inner product
(/2, /2) R R3 , which is injective if restricted to is a symmetric bilinear form and is expressed in this basis by
(, ). the symmetric matrix
We will typically use parametrizations to describe our sur-
hxu , xu i hxu , xv i
! !
faces. Let x : U R3 be a smooth map defined on an open g p :=
E F
:= .
subset U R2 . At a point (u, v) U we write F G hxv , xu i hxv , xv i
x x (Here, E, F and G are traditional names for the entries of this

xu (u, v) = = D(u,v) x(u ), xv (u, v) = = D(u,v) x(v ) matrix.) The matrix representation means that if a = au xu +
u v
av xv and b = bu xu + bv xv are two tangent vectors, then their
for the partial derivatives. (Here {u , v } form the standard ba- inner product is
sis for R2 , thought of as the tangent vectors at (u, v).) The
derivative D(u,v) x is injective if and only if xu and xv are lin- hxu , xu i hxu , xv i! bu !
ha, bi = au av .
early independent (if and only if xu xv , 0). If this is true at hxv , xu i hxv , xv i bv
all points of U, then x is an immersion.
Note that an immersion need not be one-to-one; even if it Of course the associated quadratic form a 7 ha, ai is given
is, it need not be a homeomorphism onto its image (figure 8, by the same matrix. This is called the first fundamental form

of the surface, and we use g = g p as a name for the matrix and B2. Smooth maps, change of parametrization, differentials
for the bilinear/quadratic form. (Often I = I p , from the roman
numeral one, is used instead.) We usually talk about smoothness of maps defined on an
Returning to the curve , we get open subset of Rm . If A Rm is an arbitrary subset, then a map
u
! f : A Rn is said to be smooth if it has a smooth extension
|| = g p () = u v g p
2
= E u2 + 2F uv + Gv2 . f to some open U A. (It suffices to check this locally in a
v
neighborhood of each point. Standard properties like the fact
The length of Rthep curve is of course then the integral of that the composition of two smooth maps is smooth follow
speed: len = g() dt. immediately.) In the case when A is a surface, wed like to
Note that the velocity (u, v) of has the same expres- check that is the same as requiring smoothness in coordinates.
sion (in the standard basis of R2 ) as the velocity of (in Lemma: If x : U M is a regular parametrization then
our basis {xu , xv }). We often blur the distinction between x1 : x(U) U is smooth. (Thus we say x is a diffeomor-
T (u,v) U = T (u,v) R2 and T p M, and that between and , etc. We phism onto its image.)
can think, for instance, of g p as defining a new inner product Proof: Assume (w/o l.o.g.) that (0, 0) U. We will
on T (u,v) R2 (whose matrix is FE GF with respect to the standard check smoothness near p = x(0, 0). Consider the function

basis {u , v }). y : (t, u, v) 7 x(u, v) + t p , y : R U R3 . At the origin, its
We can use the first fundamental form to measure not only partial derivatives ( p , xu , xv ) are linearly independent. That
length but also area. The parallelogram spanned by xu and xv is, D0 y is bijective. By the inverse function theorem, y is in-
has area |xu xv | and we note jective and has a smooth inverse on some neighborhood of
p = y(0). But of course this inverse is locally the desired
|xu xv |2 = |xu |2 |xv |2 hxu , xv i2 = EG F 2 = det g. extension of x1 , showing that x1 is smooth.
The area of the surface patch is then Suppose M R3 is a surface parametrized by x : U M =
Z Z p x(U). The fact that x and x1 are both smooth immediately
|xu xv | du dv = det g du dv. shows:
U U
1. f : M Rn is smooth f x is smooth.
Note also that a surface patch x(u, v) is regular (an immersion)
if and only if EG F 2 = det g is nonvanishing; this is often 2. f : Rn M is smooth x1 f : Rn R2 is
the easiest way to test the linear independence of xu and xv . smooth.
Although it is easy to arrange that {xu , xv } is an orthonormal
basis so that g is the identity matrix at one given point of Combining these facts, if N R3 is a second surface
interest (say, (0, 0) U), it is too much to hope that a general parametrized by y : V N = y(V) then we can also consider
surface have a parametrization in which {xu , xv } is an orthonor- a map f : M N. It is smooth if and only if y1 f x is a
mal basis everywhere. (We will later classify the intrinsically smooth map U V. (Note also that in this case, the smooth
flat surfaces for which this is possible. As an example think extension f to a neighborhood of M in R3 can be chosen to
of generalized cylinders ruled surfaces with constant direc- take values in N.)
tor .) Now suppose we have two parametrizations x : U M
There are, however, various special classes of parametriza- and y : V M with overlapping images. That means on the
tions which have some of the same advantages. We say a sur- open subset W = x(U) y(V) of M we have two different
face patch x is orthogonal if xu xv , that is if F = hxu , xv i = 0 systems of coordinates. Then the map := y1 x : x1 (W)
or equivalently if g = E0 G0 is a diagonal matrix. An orthogo- y1 (W) is a diffeomorphism between these open subsets of U
nal parametrization is conformal if |xu | = |xv |, that is, if E = G and V (with inverse 1 = x1 y). Being a composition of
or equivalently if g is a scalar multiple of the identity matrix. homeomorphisms, is a homeomorphism. But we also see
This means exactly that the map x : U R3 preserves angles that (and symmetrically 1 ) is smooth.
between tangent vectors (or equivalently between curves). It If f : Rn M R3 with f (a) = p then of course Da f is a
is known that any surface admits a conformal parametriza- linear map from Rn to T p M R3 . Similarly, given a smooth
tion locally. (This is a version of the uniformization theo- map f : M Rn we get a differential D p f : T p M Rn ,
rem from complex analysis.) Conformal coordinates are also the restriction of D p f for any extension f. (Different exten-
called isothermal coordinates. sions will have different derivatives in the normal direction p
We have already mentioned the normal line N p M spanned but not in tangent directions, since the derivative in any tan-
by xu xv . The parametrization x has an implicit orientation gent direction can be computed as the derivative along a curve
which allows us to pick out a unit normal vector in M, where f = f is determined.)
xu xv End of Lecture 11 May 2015
= p := .
xu xv
Note that a different parametrization (like y(u, v) := x(v, u)) B3. The Gauss map and the shape operator
may give the opposite normal vector . Some surfaces are
globally nonorientable, meaning that no continuous choice of A key tool for studying curves was the unit tangent vector
across the whole surface is possible. and its derivatives. A similar role for surfaces is played by the

unit normal . Given a smooth surface M, the map : p 7 The first and second fundamental forms are emphasized in
p S2 R3 is called the Gauss map of the surface M, and many textbooks because they are easiest to compute in coor-
is a smooth map : M S2 . Note that the Gauss map of S2 dinates. But the shape operator S p at a point p M is more
itself is the identity map. (Or the antipodal map, if we oriented directly meaningful. It encodes all the different notions of cur-
S2 with the inward normal.) vature of the surface M at the point p, capturing the second-
Going back to the general case, the differential of the Gauss order behavior of the surface, or more precisely, exactly those
map at p M is a linear map D p : T p M T p S2 . But parts which are independent of parametrization and invariant
these are the same plane the plane in R3 with normal . (Of under rigid motion.
course any two planes are isomorphic vector spaces, but these
are naturally isomorphic.) Thus we can view D p as a linear
operator on T p M. Its negative S p := D p : T p M T p M is B4. Curvatures of a surface
called the shape operator (or Weingarten operator).
Recall that an operator A : V V on an inner product Recall a few facts about a self-adjoint linear operator A on
space V is called self-adjoint if hAv, wi = hv, Awi for all an inner product space V. Its eigenvalues are all real; its eigen-
v, w V. This is equivalent to saying that the bilinear form vectors are perpendicular (since hv, wi = hAv, wi = hv, Awi =
(v, w) 7 hAv, wi is symmetric (and thus induces a quadratic hv, wi implies hv, wi = 0 for , ). That is, we can choose
form v 7 hAv, vi). Note that if A is 2-dimensional with basis an orthonormal basis of eigenvectors, and then A is of course
{e, f } then it suffices to check hAe, f i = he, A f i. represented by a diagonal matrix. The largest and smallest
Proposition: The shape operator S p on T p M is self-adjoint. eigenvalues are the minimum and maximum of the quadratic
Proof: Consider a parametrization x : U M of a neigh- form hAv, vi over the unit sphere in V. (Of course, if v is a unit
eigenvector with eigenvalue then hAv, vi = .)
borhood D of p and useE {x D u , xv } as a basis for T p M. The claim
Especially important are the symmetric functions of the
E
is that D p (xu ), xv = D p (xv ), xu . Write u := D p xu and
v := D p (xv ) for these partial derivatives of x (and xuv for eigenvalues. (These are the coefficients of the characteristic
the mixed second partial of x). Differentiating h, xu i 0 in polynomial det(I A), whose roots are the eigenvalues.) In
the xv direction gives hv , xu i = h, xuv i, while differentiat- particular, the product of the eigenvalues is the determinant
ing h, xv i 0 in the xu direction gives hu , xv i = h, xuv i. det A and their sum is the trace tr A. The average eigenvalue
Thus hu , xv i = hv , xu i, proving the claim. tr A/ dim V is also the average of hAv, vi over the whole unit
Given this proposition, sphere.

the shape operator S p defines a Now lets consider the shape operator S p on T p M. Its
quadratic form v 7 S p v, v on T p M, called the second funda-
mental form h p of M, often written using the Roman numeral eigenvalues k1 and k2 are called the principal curvatures (of M
as II p (v) := h p (v). Note that arguments as in the proof show at p); the eigenvectors are the principal curvature directions,
forming two orthogonal lines in T p M. We can choose unit
hu , xu i = h, xuu i , hv , xv i = h, xvv i . eigenvectors e1 and e2 such that {e1 , e2 , } is an oriented or-
thonormal basis. We define the Gauss curvature
Thus the matrix of the second fundamental form w.r.t. the ba-
K := k1 k2 = det S p
sis {xu , xv } is
and the mean curvature.
hu , xu i hu , xv i
! ! !
L M h, xuu i h, xuv i
h p := := = . k1 + k2 1
M N hv , xu i hv , xv i h, xvu i h, xvv i
H := = tr S p .
2 2
We see that to compute the first and second fundamental
Note that K is independent of orientation, while H changes
forms of a parametrized surface, we start by computing the
sign if we switch the sign of ; more intrinsic is the mean
first and second partial derivatives (xu , xv ; xuu , xuv , xvv ), then
~ = H. (Note also that some authors define
curvature vector H
compute the cross product xu xv and its length |xu xv |. The
scalar products among the first derivatives give the matrix g p . the mean curvature with the opposite sign and/or without the
The scalar products of the second derivatives with xu xv , factor 1/2 despite the name mean.)
divided by the length of this normal vector, give the matrix h p . End of Lecture 18 May 2015
As usual, it is easier to find the matrix for the bilin- As we will see, the intrinsic local shape of the surface is de-
ear/quadratic form h p than to find the matrix for the associ- termined by the Gauss curvature K, in particular, qualitatively
ated operator, the shape operator S p . (Since {xu , xv } is not by its sign. We say p M is an elliptic point if K(p) > 0 (that
generally orthonormal, it is easier to find the scalar products is, the principal curvatures have the same sign) or a hyperbolic
of u with the basis elements than to find its expression in the point if K(p) < 0 (. . . opposite signs). A point where K(p) = 0
basis.) But by linear algebra we know h p = g p S p , or equiva- is called a parabolic point.
lently S p = g1
p h p . Of course the inverse of a 2 2 matrix is A point where k1 = k2 is called an umbilic point; in par-
easy to compute: ticular a planar point has k1 = k2 = 0. (Nonplanar umbilic
!1 ! points are of course elliptic.) At umbilic points, the princi-
E F 1 G F pal directions are not uniquely defined and the normal curva-
g1
p = = .
F G EG F 2 F E ture defined below is constant. Note that many authors use
10

parabolic to mean what we call parabolic and nonplanar; Thus we see that at any point p M there is a uniquely
this no longer depends just on K. determined paraboloid (z = k1 x2 + k2 y2 in the rotated coordi-
The Gauss curvature K and the mean curvature H are nates) that has second-order contact with M at p. Two surfaces
smooth functions on any smooth surface. The principal curva- tangent at p have second-order contact if and only if they have
tures k1 and k2 are the two roots H H 2 K of the charac- the same principal curvatures and directions there (i.e., have
teristic polynomial k2 2Hk + K of the shape operator. These the same ocsulating paraboloid). Up to rigid motion, two sur-
are smooth functions only away from umbilic points (where faces agree to second order (at given points) if and only if they
they coincide because the square root vanishes). have the same Gauss and mean curvatures there.
For a unit tangent vector w T Note: for curves we preferred to talk about osculating cir-

p M, the
normal curvature
of M in direction w is h p (w) = S p w, w . Note that for an cles (with constant curvature ) rather than osculating parabo-
arbitrary nonzero vector w T p M, las (with constant acceleration vector). For surfaces, we might
!
want to use a surface with constant principal curvatures (or
w S p w, w h p (w) equivalently, constant K and H). But we will see later this
hp = = .
|w| hw, wi g p (w) happens only for spheres and planes and cylinders. (Similarly,
you will show in homework that spheres and planes are the
We can write any unit normal vector as w = cos e1 + sin e2 only totally umbilic surfaces.) Although surfaces are in some
and we find the normal curvature of M in this direction is sense determined by their curvatures, this is much more com-
cos2 k1 + sin2 k2 , a weighted average of the principal cur- plicated than saying space curves are determined by (s) and
vatures. Of course, the mean curvature is the average normal (s). First there are compatibility conditions (PDE not ODE:
curvature over the whole circle of directions; the principal cur- compatibility basically says uv = vu ) and second theres no
vatures are the minimum and maximum of the normal curva- standard parametrization (like arclength).
ture. From the Taylor series or osculating paraboloid, we do see
The intersection of M with a normal plane at p a plane for instance that near an elliptic point p, the surface is locally
spanned by p and some unit tangent vector w T p M is convex it stays to one side of its tangent plane. This is not
a curve with T = w and ~ = h p (w). Orienting this nor- true at a hyperbolic point; instead T p M cuts M locally in two
mal plane such that {w, } is an oriented basis, we thus have curves crossing at p; their tangent vectors at p are exactly the
h p (w) = : the normal curvature is the (signed) curvature of directions with vanishing normal curvature, called asymptotic
the normal slice . Later we will see how the normal curva- directions.
ture is the normal component of the curvature of an arbitrary We could also consider the intersections with nearby par-
curve through p in direction w. allel planes (say at distances 2 to either side of the tangent
Let us now consider a surface given as a graph: x(u, v) = plane). Unless p is a planar point, these planes will intersect
u, v, f (u, v) , in particular at a point p where grad f = 0 so M approximately in the curves k1 x2 + k2 y2 = 2 , which are
that the surface is horizontal there. We have xu = (1, 0, fu ), scaled (by ) versions of the Dupin indicatrix, defined to be
xv = (0, 1, fv ) so that at p this is an orthonormal basis for the set of w T p M such that S p w, w = 1. This is an el-

T p M, meaning g p = I. Of course p = (0, 0, 1). Since the lipse at an elliptic point, a pair of hyperbolas (with common
normal (vertical) components of the second derivatives of x asymptotes in the asymptotic directions) at a hyperbolic point,
are the second derivatives of f , we see that and a pair of parallel lines at a nonplanar parabolic point.
!
fuu fuv End of Lecture 21 May 2015
h p = hess f = .
fvu fvv
Of course, since we have an orthonormal basis, this is also B5. Curves on surfaces and the Darboux frame
the matrix of the shape operator. Thus we have K = fuu fvv
2
fuv and 2H = tr(hess f ) = fuu + fvv = f . (In general, one
Lets now consider an arbitrary curve on a surface M.
should think of mean curvature as a geometric version of the
The Darboux frame along the curve is {T, , } where T is
Laplacian; in terms of the intrinsic LaplaceBeltrami operator
~ of course the unit tangent to , is the surface normal, and
M , we have for instance M x = 2H.) := T is called the conormal. As for our other frames, the
If we start with any point p on an arbitrary surface M, we derivative (with respect to an arclength parameter for ) gives
can apply a rigid motion (or equivalently, choose new Eu- a skew-symmetric matrix:
clidean coordinates) to put it into a standard position as fol-
lows. First translate so that p = 0 is the origin, then rotate 0
T 0 g n T

so that p = (0, 0, 1) is vertical. Note that the surface is then =
g 0 g .

locally a graph z = f (x, y) with grad f = 0 at p as above. Fi- n g 0
nally, rotate around the vertical axis until the x- and y-axes are
principal directions. As above, with respect to the standard Of course depends only on the surface M, and so 0 =
basis for T p M = R2 , the matrix for S p is hess f ; this is now D p (T
) = S p(T ). Comparing with
the above, we find
the diagonal matrix k01 k02 . This means the second-order Tay- n = S p (T ), T and g = S p (T ), . Thus n = h p (T ) is

lor expansion of f around 0 is f (x, y) = 0 + k1 x2 + k2 y2 + , the normal curvature of M in the direction T . It is the normal
where the remainder terms are third-order. part of the curvature ~ = g + n of . The tangential part
11

g is called the geodesic curvature. The remaining derivative Theorem: Let X be a smooth vector field on U R2 . For
g , the twisting of the Darboux frame, is called the geodesic any p U there exists a neighborhood V 3 p, a time interval
torsion. Like the normal curvature n , the geodesic torsion I = (, ), and a smooth map : V I R2 (called the
g = h p (T, ) = h p (, T ) depends only on M (and T ). Curves local flow of X) such that for each q V the curve t 7 (q, t)
in M having tangent T at p differ (to second-order) only in is the integral curve of X through q, meaning (q, 0) = q and
having different geodesic curvatures. (Walking in the moun- (q, t)/t = X(q,t) .
tains, we have a choice of turning left or right; whether we For fixed t I the map q 7 (q, t) is called the flow of X by
curve up or down is fixed by staying on the earths surface.) time t. Note that the uniqueness of integral curves implies that
Curves on M for which one of these quantities vanishes flowing by time s and then by time t is the same as flowing by
special. Curvature lines are curves for which g = 0, meaning time s + t, i.e., (q, s), t = (q, s + t) (whenever both sides
that T is always a principal direction. (This is no condition are defined). Taking s = t we see that the flow by time t is
at an umbilic point.) We have 0 = n T , where n is one invertible: it is a diffeormorphism from V to its image in U.
of the principal curvatures. The Darboux frame is a paral- End of Lecture 28 May 2015
lel (Bishop) frame along a curvature line. Curvature-line co-
Corollary: If X is a vector field on U and X p , 0 for some
ordinates are (necessarily orthogonal) coordinates for which
p U, then there exists a neighborhood W 3 p and a smooth
the coordinate lines are curvature lines. Locally away from
function f : W R which is constant along each flow line
umbilic points, this happens exactly when both the first and
of X but has d f , 0 everywhere. (Such an f is called a local
second fundamental forms have diagonal matrices in the basis
first integral for X.)
{xu , xv }. (The coordinates we used for surfaces of revolution
Pf: Assume without loss of generality that p = (0, 0) and
had this form.) This makes many computations much easier.
X p = (1, 0). Let : V I U be a local flow and con-
Asymptotic curves are those for which n 0, that is, T is
sider its restriction to the two-dimensional cross-section
always an asymptotic direction. (This of course requires K
{u = 0} (transverse to X p ). This restriction (v, t) has non-
0.) The conormal is the principal normal N of an asymptotic
singular derivative at 0, so it locally has an inverse on some
line; the Darboux frame is the Frenet frame. Locally near
open neighborhood of (0, 0), mapping this diffeomorphically
a hyperbolic point, asymptotic coordinates always exist, and
to some W 3 p. We can take f to be the v-coordinate of this
are characterized by the second fundamental form having an inverse.
off-diagonal matrix h = M0 M0 .
Definition: A (smooth) vector field X on a surface M is
Geodesics are curves for which g = 0. Given a starting
a function M R3 such that X p = X(p) T p M for each
point p and a starting direction T , there is always a unique
p M. Note that X p = a(p)xu + b(p)xv = D(u,v) x(a, b) for
geodesic, the solution to the ODE T 0 = n . We will consider
smooth real-valued functions a, b on M. (Smoothness in these
this in more detail later. (The surface normal is the principal
coordinates is equivalent to smoothness in R3 .) We see that,
normal = N and the curvature equals the normal curvature.)
just as Dx gives a pointwise isomorphism between T (u,v) R2
Any straight line contained in a surface (for instance, the and T p (M), it gives a one-to-one correspondance between vec-
rulings on a ruled surface) has constant T and is both an tor fields on U R2 and those on x(U) M.
asymptotic line and a geodesic. (The geodesic torsion g gives Thus all local results about vector fields hold also on sur-
the speed at which the surface normal and conormal rotate faces. (Literally just replace U by M in the theorem or corol-
around the line.) lary above.) A stronger way to express the corollary is to say
A curvature line which is also a geodesic has constant that around a point where X p , 0 there are coordinates such
conormal ; equivalently it is the intersection of M with a that X = xu . (The restriction (v, t) gives a new parametriza-
plane meeting M perpendicularly (like the generating curves tion of W in terms of coordinates (v, t); the function f used
on a surface of revolution). above is the v coordinate and we rename t as u to get X = xu .)
The surface normal is constant along a curvature line We can say: any nonzero vector field is locally constant in
which is also asymptotic; such a curve is the intersection of M appropriate coordinates.
with a plane always tangent to M and consists of course of Theorem: Suppose X and Y are two vector fields on M
parabolic points. (Example: top or bottom of round torus or which are linearly indpendent at some point p. Then there
of tube around any plane curve.) exists a parametrization x : U M of some neighborhood
W 3 p such that xu ||X and xv ||Y on W.
Note: It is too much to ask that xu = X and xv = Y; coordi-
B6. Vector fields and line fields nate vector fields always commute (in the sense that the time-t
flow along X and the time-s flow along Y are commuting dif-
A (smooth) vector field X on U R2 is a smooth map feomorphisms). But general vector fields X and Y do not have
U R2 interpreted as p 7 X p T p R2 = R2 . That is, we this property.
think of X p as an arrow based at p. A flow line (or integral Pf: Let u and v be first integrals of Y and X (respectively)
curve or trajectory) of X is a curve in U whose velocity is on some neighborhood V 3 p. The map (u, v) : V R2
given by X, that is, (t) = X(t) for all t. This is a system of has nonsingular differential at p. (Since u = 0 only along
ODEs in the two variables u and v. The standard theorems curves tangent to X p while v = 0 only tangent to Y p , these can
on ODEs say give not only existence and uniqueness of flow never both vanish.) Thus it locally has an inverse, the desired
lines, but also smooth dependence on the initial point. That is: parametrization.
12

Corollary: Any p M has a neighborhood with an orthog- We find the velocity of t is 0t = T + t0 + O(t2 ), and hence
onal parametrization.
|0t |2 = |T + t0 |2 + O(t2 ) = 1 + 2t 0 , T + O(t2 ).

Pf: We just need to find a pair of orthogonal vector fields.
Start with an arbitrary parametrization x of some neighbor- RL
hood of p. Set X = xu and Y = p xu and apply the theorem. Therefore d
dt t=0 |0t | = h0 , T i. Using len(t ) = 0
|0t | ds we
The new parametrization has coordinate lines in the (orthogo- find
nal) X and Y directions. Z
d d
In the theorem, note that X and Y stay linearly indepen- len(t ) = |0t | ds
dent (in particular nonvanishing) on the parametrized neigh- dt t=0 dt t=0
Z Z
borhood. The theoremm only depends on equivalence classes d
= |0t | ds = , T ds.

0
of X and Y under X f X (where f is a nonvanishing scalar dt t=0
function). These are line fields (or direction fields). Note:
globally, a line field may be nonorientable. Locally, however, Here smoothness justifies interchanging the derivative and in-
we can always pick a consistent orientation for the lines, so the tegral. Next we can integrate by parts, recalling that T 0 = ~ =
line fields always arises as above from a nonvanishing vector N; our assumptions mean that the endpoint terms vanish. We
field. The theorem is really about a pair of (nowhere equal) find
line fields. (Line fields have unparametrized integral curves.) d
Z
len() := len(t ) = ,~ ds.

As further applications of the theorem, we can derive local
dt t=0
existence of asymptotic and curvature-line coordinates.
Cor: If p M is a nonumbilic point, then some neighbor- We can think of the right-hand side as the inner product on
hood of p can be parametrized by curvature-line coordinates. L2 (I, R3 ) (the tangent space at to the space of curves, where
Pf: In a orientable neighborhood without umbilics, we can the vector fields and ~ along live). Thus the formula can
distinguish the principal curvatures (say k1 < k2 ). Then we be thought of as saying that ~ is the negative gradient of the
get two line fields along the eigenspaces for k1 and k2 re- length functional.
spectively. Then just apply the theorem. A curve is length-critical if no variation changes its length
Cor: If p M is a hyperbolic point, then some neighbor- to first order. That is, we have len = 0 for all , which
hood of p can be parametrized by asymptotic coordinates. happens if and only if 0. Of course we know that straight
Pf: In an orientable neighborhood where K < 0 we have lines minimize length.
two asymptotic lines at each point, and can distinguish them We see that (as claimed above) the derivative len only
globally (one is to the left of the negative principal curvature depends on and not on the higher-order terms. Also, it is
direction). Thus we get two line fields and can apply the the- independent of the tangential part of if we keep the same
orem. family of curves t but change their parametrizations, that cor-
responds to changing by a tangential field but clearly has no
effect on (the derivative of) length. Any family of curves can
B7. First variation of length be reparametrized so that the variation field is normal.
Note furthermore that len depends only the component
We want to understand the geometric meaning of the mean of in the principal normal direction. The so-called Hasi-
curvature H. In particular, if we consider variations of a sur- moto flow is the PDE = T ~ = B for a moving curve t in
face, we will see how to express the derivative of area in terms R3 , which physically is an approximation to smoke-ring flow.
of H. If a surface has H 0 then it is a critical point for area, Since we move only in the binormal direction, this flow pre-
called a minimal surface. serves length; indeed is it a so-called integrable system which
First we consider the simpler case of the length of a (com- also preserves a whole hierarchy of other invariants.
pact) curve. Suppose t (s) is a smoothly varying family Curve-shortening flow is the PDE = ~ for a moving curve
of smooth curves in Rn . We assume that (s) = 0 (s) is t whose length decreases as fast as possible, since we follow
parametrized by its arclength s. (But s is not arclength for the the (negative) gradient direction. It is one of the earliest ex-
other curves t .) If is not closed, we assume the variation amples of a geometric flow, and has interesting properties like
is supported on a compact region away from the endpoints. preserving embeddedness of plane curves.
(That is, t (s) is independent of t for s outside this region.)
We can take a Taylor series in t and get
B8. Minimal surfaces
t (s) = (s) + t(s) + O(t2 ),
We now want to do the similar calculation to find the first
where is a variation vector field along . We will see that the variation of surface area. A more sophisticated approach
derivative of length depends only on this infinitesimal varia- would use the characterization 2H ~ = M x and an intrisic ver-
tion, and not on the higher order terms we have omitted. (One sion of Stokes theorem. We will take a more hands-on ap-
could think of the vector field as being a tangent vector to proach in coordinates.
the infinite dimensional space of curves at the point .)
Consider an initial surface with an orthogonal parametriza-
End of Lecture 1 June 2015 tion x : U M. Let : U R have compact support in U
13

and describe a normal variation of M. That is, we consider the B9. Isometries
family of surfaces xt := x + t. (Guided by our experience
with curves, we realize that nothing would change if we added What do we mean when we talk about intrinsic properties
tangential terms or higher-order terms.) of a surface, properties that only depend on the intrinsic ge-
We find xtu = xu + tu + tu and xtv = xv + tv + tv . ometry of the surface and not on how it sits in space? More
Recalling that we have assumed hxu , xv i = 0, this gives precisely, these are properties that are the same for any two
xu , xu = hxu , xu i + 2t hxu , u i + O(t2 ),

t t isometric surfaces.
Def: An isometry is a diffeomorphism : M N between
xu , xv = t hxu , v i + t hxv , u i + O(t2 ),

t t
two surfaces which preserves the scalar product on tangent
xv , xv = hxv , xv i + 2t hxv , v i + O(t2 ).

t t
spaces. any p M and any v, w T p M, we have
That is, for
D p (v), D p (w) = hv, wi. It follows that preserves the

This can be written as gt = g 2th + O(t2 ). length of curves: len( ) = len() for any curve in M.
Since in our orthogonal coordinates g = E0 G0 is a diagonal Example: If is a rigid motion of Rn then of course it re-
matrix, the off-diagonal entries are irrelevant for the first-order stricts to any surface M to give an isometry M M. Less
calculation of det gt . Writing h = ML M

N we get in fact trivially, R(, ) is isometric to the unit cylinder in R3 with
det gt = (E 2tL)(G 2tN) + O(t2 ) one vertical line removed. (We see from this simple example
that the mean curvature H, for instance, is not an intrinsic no-
= EG 2t(LG + NE) + O(t2 ). tion; the surprising result later will be that K is intrinsic.)
Taking the square root gives Note: If x : U x(U) = M is a parametrization and
: M N is a diffeomorphism then of course y :=
LG + NE
det gt = EG 1 t + O(t2 ), x : U N is a parametrization. We see that = y x1
p
EG is an isometry if and only if, at corresponding points, the first
but we recognize the fraction LG+NE
EG = 2H as twice the mean fundamental forms for x and y have the
same matrix with
re-
spect to the coordinate bases. That is, xu , xv = yu , yv , etc.

curvature. Thus

d p Def: We say surfaces M and N are locally isometric if
det g = det gt = 2H det g.
p p
dt t=0 each point in either surface has a neighborhood isometric to
an open subset of the other surface.
To find the first variation of area, we simply integrate this Note: we can assume the neighborhoods are small enough
over U: to be parametrized patches. Then we test local isometry by
Z Z p
finding parametrizations with the same first fundamental form
area(x) = dA = det g du dv
g(u, v) = g = FE GF .

Z M p U
Example: The plane and the cylinder are locally isometric.
Z
= det g du dv =
p
2H det g du dv Relaxing the condition of isometry, we can consider con-
U U
Z formal (angle-preserving) maps : M N. Here the con-
= 2 H dA dition is that there is a positive function : M R (called
M the such that for any v, w T p M we have

conformal factor)
D p (v), D p (w) = (p)2 hv, wi.

We see that the mean curvature is the negative gradient for
area to save area one should move the surface in the direction We have seen the sense in which two surfaces are locally
of the mean curvature vector. A surface is area-critical if and isometric if and only if they have the same first fundamental
only if area is zero for every variation , that is, if and only form gi j . A somewhat surprising result is that all surfaces are
if H 0. Such a surface is called a minimal surface. Least- locally conformal. (By transitivity, it suffices to prove that any
area surfaces spanning a given boundary are known to exist surface has conformal parametrizations. This is a PDE result
and be smooth; they are thus minimal surfaces. (It can also that we wont try to prove here.) The analog is not true for
be shown that, given a minimal surface M, any sufficiently higher-dimensional manifolds only certain special metrics
small piece of M indeed any piece which is a graph in some are conformally flat.
direction is the least-area way to span its boundary.)
Minimal surfaces have many interesting properties. For in-
stance the Gauss map : M S2 is (anti)conformal, since B10. Covariant derivatives
its differential has matrix 0k k0 in an orthonormal basis of
principal directions at a point with K = k2 . One can check If M is a surface, and Y is a (vector-valued) function on M,
that in any conformal parametrization x : U M of a mini- then we know what the directional derivative of Y at p M in
mal surface, the coordinate functions or more generally all direction w T p M means: w Y(p) = D p Y(w) is the derivative
height functions hx, ui : U R for constant u S2 are har- of Y along any curve (in M through p) with velocity vector w.
monic functions. Thus they can be thought of as the real parts Now suppose Y : M R3 is a tangent vector field (meaning
of complex holomorphic functions, leading to the so-called Y p T p M for each p). In general, its directional derivatives
Weierstrass representation. (Thinking of S2 as the Riemann w Y will not be tangent to M. Indeed, since the tangent spaces
sphere C, the Gauss map itself is a meromorphic function.) change as p M moves, a tangent vector is forced to change
End of Lecture 4 June 2015 in the normal direction just to stay tangent. We can make this
14

precise using the second fundamental form since hY, i 0, fields in terms of the covariant derivatives of the coordinate
we get vector fields.
End of Lecture 8 June 2015
w Y, = DY(w), = D(w), Y = S (w), Y = h(w, Y).

If X and Y are two vector fields on M, then in general X Y
This normal change in Y is forced by the geometry of M. and Y X are unequal. The difference is called the Lie bracket:
The intrinsic change in Y is given by the tangential parts
of its directional derivatives, called covariant derivatives. We [X, Y] := X Y Y X.
write
A special property of coordinate vector fields is that they have
w Y = (w Y)|| = w Y w Y, = D p Y(w)h(w, Y p ) T p M. vanishing Lie bracket: [xu , xv ] = 0. To verify this, note that

xu xv is by definition the tangential part of xuv . Thus [xu , xv ] =
To define w Y and w Y it of course not necessary that Y be 0 is a trivial consequence of xuv = xvu . Given two linearly
defined on a whole neighborhood of p (in M) it suffices if Y independent vector fields X and Y, we discussed the fact that
is a vector field (tangent to M) along some curve : I M we cannot always find coordinates with X = xu and Y = xv .
through p with velocity (parallel to) w. If Y is defined along Indeed the condition [X, Y] = 0 is exactly what is needed

sometimes we write dt Y := (t) Y for the tangential part of the this is part of the Frobenius Theorem (covered next semester).
derivative Y = dt Y (t) . (Here again, when we talk about t-
d
derivatives of Y, we are really differentiating the composition
Y .) B11. Christoffel symbols
If is parametrized at unit speed, then ds
Y = T Y is the
tangential part of Y 0 . In particular, choosing Y = T and com-
For ease of writing equations in coordinates, we will change
paring with the Darboux equation T 0 = g + n , we see that
notation a bit: we write (u1 , u2 ) := (u, v) and use the subscript i
T T = g gives the geodesic curvature. The equation g = 0
for a partial derivative with respect to ui , so for instance xi :=
for a geodesic can be written T T = 0. (A curve satisfies
xui = x/ui . Then we can write the entries of the matrices
= 0 if and only if it is a geodesic parametrized at con-
forms as gi j = xi , x j

for the first
and second

fundamental
stant not necessarily unit speed.)
and hi j = , xi j = i , x j .

We say the vector field Y is parallel along if dt
Y 0.
Given Y p at any initial point p = (0), there is a unique way We now want to explicitly calculate covariant derivatives in
to extend it to a parallel field Y along (solving the ODE coordinates. We start with the covariant derivatives xi x j of
T Y = 0). Any curve from p to q thus gives a map T p M the coordinate vector fields; since these are tangent vectors,
T q M called parallel transport taking an initial vector at p they can be expressed in terms of the coordinate basis. We in-
to the value at q of the parallel field along . It is important troduce the Christoffel symbols kij as their components. That
to note that this parallel transport from p to q does depend on is, the kij are defined by
the choice of its not a natural identification of the distinct
tangent spaces. 2
X
Important properties are that a parallel vector field has con- xi x j =: kij xk = 1i j x1 + 2i j x2 .
stant length; a pair of parallel fields make constant angle; thus, k=1
parallel transport is an orthogonal map T p M T q M, a map
respecting the scalar products. This is easy to confirm: par- (The vanishing of the Lie bracket can now be expressed as
allel fields have derivatives only in the normal direction , so the symmetry kij = kji .) Since we already know the normal
component hi j = xi j , of xi j , we could write the equations

hX, Yi0 = hX 0 , Yi + hX, Y 0 i = 0.
Given two vector fields X and Y on M, the covariant deriva- above as
tive X Y is defined at every point of M, giving a new vector
field. Note that X Y is R-linear in each argument: xi j = xi x j = 1i j x1 + 2i j x2 + hi j .
X+X 0 Y = X Y + X 0 Y, X (Y + Y 0 ) = X Y + X Y 0 , (This is called the Gauss formula.)

aX Y = aX Y = X aY, a R. At every point in the surface, we have a (nonorthonormal)
frame {x1 , x2 , }. The derivatives of expressed in this frame
If f : M R is a smooth function, then of course f X means and the normal components of the derivatives of the xi
the vector field whose value at p M is ( f X) p := f (p)X p . are given by the second fundamental form and shape operator.
Since X Y at p depends only on X p , we find that f X Y = The tangential parts of the derivatives xi j are new given by
f X Y. But on the other hand, the Leibniz product rule gives the Christoffel symbols.
Of course, as we saw with S and h, when dealing with
X ( f Y) = X f Y + f X Y,

nonorthonormal bases, it is easier to compute scalar products
than components. Here we have
where we adopt the convention that X f := X f . Since any
vector field is a combination f xu + gxv , these formulas will DX ` E X
x ji , xk = xi x j , xk = i j x` , xk = ì j g`k =: i jk .

allow us to express covariant derivatives of arbitrary vector ` `
15

Here we multiply by the matrix g = (gi j ) to lower an index. B12. Compatability conditions
It is customary to write the inverse matrix as
! Suppose we are given symmetric matrices gi j and hi j vary-
1 G F
g := g =
i j 1
. ing smoothly on a given domain U R2 . Is there some
EG F 2 F E parametrization x : U R3 with these as first and second
fundamental form, respectively? Of course from g and h we
Then multiplying by g1 raises the index again: kij =
know the Christoffel symbols and the matrix for the shape op-
` g i j` . Of course we still have the symmetry iP jk = jik .
P k`
erator. Thus we try to solve the Gauss/Weingarten system
Suppose we write X and Y in coordinates as X = i xi and
Y = xi (for smooth functions , ). Then by linearity and
P i i i
xi j = 1i j x1 + 2i j x2 + hi j , i = S (xi ).
the product rule, we have
X X i X (Of course we also need that is the unit normal vector to the
X Y = i xi j x j = xi k + j kij xk surface given by x. As long as that holds at some initial point,
i, j i,k j the Gauss/Weingarten system is set up to ensure it stays true,
X k X since the scalar products of with itself and with the xi will
= i + j kij xk . be constant.)
i,k
ui j Unlike for ODEs, solutions to PDEs exist only if compat-
Now we consider derivatives of the coefficients of the first ibility equations are satisfied. (The basic idea is that given
fundamental form: functions g and h, the system f x = g, fy = h can have a solu-
tion f only if gy = f xy = fyx = h x .) We will write down the
gi j,k := xk gi j = xk xi , x j = xk xi , x j + xk x j , xi compatibility equations for our system (equating xi jk = xik j );

= xik , x j + x jk , xi = ki j + k ji .

these are clearly necessary.
Using the notation k := xk = /uk for partial derivatives,
This is a set of eight equations. We could write them out differentiating the Gauss formula gives
explicitly in the classical notation, getting Eu = 2111 , etc. X
xi jk = k hi j + hi j k + k ì j x` + ì j x`k .

But lets just cyclically permute the equation above and use
the symmetry of the Christoffel symbols: `
gi j,k = ki j + k ji = ki j + jki , The Gauss/Weingarten system shows us how to write the

right-hand side in the basis {x1 , x2 , }. Each xi jk = xik j then
g jk,i = i jk + ik j = i jk + ki j ,
gives us three scalar compatibility equations; it turns out to
gki, j = jki + jik = jki + i jk . be enough to consider a few of these, since the rest are equiv-
alent by symmetries. First note that the normal component
Subtracting the top equation from the sum of the other two
is k hi j + ì j h`k . The two cases of interest are x112 = x121
P
gives g jk,i + gki, j gi j,k = 2i jk . More important than the exact
and x221 = x212 ; the normal parts give the (Mainardi)Codazzi
form of this equation is the fact that it confirms the intrinsic
equations (evidently first discovered by Peterson), the first of
nature of the covariant derivative: The Christoffel symbols,
our compatibility conditions:
and thus all our formulas for covariant derivatives, can be ex-
pressed in terms of the first fundamental form (and its deriva- 2 h11 + 111 h12 + 211 h22 = 1 h12 + 112 h11 + 212 h21 ,
tives) alone.
After this excursion into very abstract notation, lets look 1 h22 + 122 h11 + 222 h12 = 2 h12 + 112 h12 + 212 h22 .
concretely at these equations in more classical notation. We
specialize to the case of an orthogonal parametrization (g12 Here of course, the Christoffel symbols should be viewed as
0); this makes all the equations a bit simpler. The equations functions of the gi j .
2i jk = g jk,i + gki, j gi j,k become: Recalling that the shape operator has matrix g1 h, we write
this in index notation as hij := k g jk hki , so that i = S (xi ) =
P
2111 = g11,1 = Eu , 2112 = g11,2 = Ev , j hij x j . This lets us express the x2 component of x112 =
P
2121 = g11,2 = Ev , 2122 = g22,1 = Gu , x121 . We get
2221 = g22,1 = Gu , 2222 = g22,2 = Gv . X X
h11 h22 + 2 211 + `11 2`2 = h12 h21 + 1 212 + `12 2`1 .
Of course multiplying by the inverse of a diagonal matrix ` `
is easy, so kij = ` gk` i j` becomes 1i j = i j1 /E and 2i j =
P
i j2 /G. That is, we get This can be written as
111 = Eu /2E, 112 = Ev /2E 122 = Gu /2E 2 211 1 212 +111 212 +211 222 112 211 212 221 = h11 h22 h21 h21 .
211 = Ev /2G, 212 = Gu /2G 222 = Gv /2G Expanding h11 = h11 g11 + h21 g21 and h21 = h12 g11 + h22 g21 , the
right-hand side becomes
g11 h11 h22 h21 h12 = g11 det S = g11 K.

End of Lecture 11 June 2015
16

Since the left-hand side is intrinsic (expressible in terms of the unspecified parameter t gives hardly any information about
first fundamental form alone), so is the Gauss curvature K. the shape of the curve. Any arc of monotonic curvature can
That is, we have Gausss Theorema Egregium (remarkable be parametrized by = t for instance.
theorem): The Gauss curvature K is an intrinsic notion, re- What is the situation for surfaces? Two surfaces are lo-
maining unchanged under local isometries, as when a surface cally isometric (intrinsically equivalent, one might say) if and
is bent without stretching. only if they have the same first fundamental form g in corre-
The equation above in the form sponding coordinates. This implies that they have the same
Gauss curvature K(u, v) in such coordinates. By Bonnets the-
det h
2 211 1 212 + 111 212 + 211 222 112 211 212 221 = g11 , orem, two surfaces are related by a rigid motion (extrinsically
det g equivalent, one might say) if and only if they have the same
first and second fundamental forms g and h in corresponding
again with the understanding that the Christoffel symbols
coordinates. This implies that they have the same Gauss and
should be expressed as functions of the gi j and their deriva-
mean curvatures K and H (or equivalently, the same principal
tives, is the Gauss equation, the last compatibility condition.
curvatures) in such coordinates.
A theorem of Bonnet (basically using standard results about
If we ask when surfaces with the same K are isometric, then
first-order PDEs) now says these compatibility conditions are
we are faced with the same problem as for curves of not know-
also sufficient. If symmetric matrix functions gi j and hi j (with
ing what parametrization is being used to compare the curva-
g positive definite) satify the Gauss and Codazzi equations,
ture functions. For instance, as long as p M is not a critical
then there is a surface with these fundamental forms, unique
point of the function K, then a neighborhood of p is foliated
up to rigid motion. We will not go into the details of the proof.
by lines of constant K and we can use v := K as one of the two
Let us return to the special case of orthogonal coordinates,
coordinates in a regular parametrization of this neighborhood.
and write the intrinsic formula for Gauss curvature more ex-
One case where this problem doesnt arise is that of surfaces
plicitly. We get
with constant Gauss curvature.
Ev Gu EuGu EvGv Ev2 G2u Suppose we have two isometric surfaces with the same
2EK = v u + + , mean curvature (or equivalently the same principal curva-
G G 2EG 2G2 2EG 2G2
tures). This does not always imply that that they are related by
or equivalently a rigid motion. What happens is that, even though the eigen-
values of the shape operator are the same on both surfaces,
EvGv G2 the eigenvectors (the principal directions) can rotate. An im-
2EGK = Evv + Guu u
G G portant example is that of minimal surfaces: it turns out that
EuGu EvGv Ev2 G2u any minimal surface has an isometric conjugate minimal sur-
+ + face. Here the curvature directions have become asymptotic
2E 2G 2E 2G
Ev Gu directions and vice versa. (In fact, these sit in a one-parameter
= EG v + u . family of isometric minimal surfaces with all possible asymp-
EG EG
totic directions.)
For conformal coordinates with conformal factor = e >
0 we have E = G = 2 and the formula becomes
B13. Surfaces of constant curvature
1 v u
K = 2 v + u = e2 .

There are many interesting facts about surfaces with con-
stant Gauss curvature or constant mean curvature. Our first
Bonnets theorem leaves many related open questions. One goal is a thereom of Minding saying any two surfaces with
is the following: to what extent the compatibility conditions the same constant K c are locally isometric. Then we turn
determine the second fundamental form from the first funda- to theorem of Liebmann that characterize the round sphere are
mental form? This is a rigidity question: when does a sur- the unique closed surface of constant K and the unique convex
face admit a unique isometric embedding into R3 ? Unlike hy- surface of constant H.
persurfaces in higher dimensions see Khnels book sur- We will use one further special kind of parametrization:
faces can always be locally embedded in many ways. Glob- Def: An orthogonal parametrization x : U R3 gives
ally, however, there are often rigidity results (like for convex geodesic parallel coordinates if |xu | 1.
surfaces) and other (abstract) surfaces cannot be globally em- Using the notation a = |xv |, we have in geodesic parallel
bedded at all. We may return to these results later. coordinates E 1 and G = a2 . Specializing our formula for K
End of Lecture 15 June 2015 to this case gives 2a2 K = au (2aau /a), that is, K = auu /a.
For curves, of course, there is no intrinsic geometry any Clearly, in geodesic parallel coordinates, the u-curves are
two curves are locally isometric, for instance by choosing ar- parametrized at unit speed: any pair of v-curves cut segements
clength parametrizations for both of them. Two plane curves of equal length from all the u-curves. We claim the u-curves
are related by a rigid motion if and only if they have the same (for each constant v) are geodesics, so that the v-curves should
(extrinsic) curvature (as a function of arclength). Note, how- really be considered as parallel curves at constant distance
ever, that specifying curvature as a given function (t) of an from each other. (The name geodesic parallel coordinates
17

In curvature-line coordinates with g =

then comes from the fact that the u- and v-coordinate lines are Lemma: E 0 and
0 G
geodesics and parallels, respectively.) h = L0 N0 , the Codazzi equations become
To check the claim, we must show xu xu = 0. This is equiv-
alent to 111 = 0 = 211 , or to 111 = 0 = 112 . But in orthogoal Lv = HEv , Nu = HGu .
coordinates, we had 111 = Eu /2 and 112 = Ev /2.
In terms of the principal curvatures, with L = k1 E, N = k2G,
Lemma: Any surface locally admits geodesic parallel coor-
the equations can be written as
dinates.
Proof: In fact we can start with any curve (v) in the surface 2Ev k1 2Gu k2
to be the curve u 0. Then each u-curve v c is determined Ev = , Gu = .
k2 k1 k1 k2
as the geodesic starting at (c) in the conormal direction. The
only thing that one needs to check is that the coordinates stay Proof: Dropping the terms involving h12 , the Codazzi equa-
orthogonal. But since the v-curves are geodesics, xuu is a nor- tions are
mal vector, so hxuu , xv i = 0. Thus
2 h11 + 211 h22 = 112 h11 , 1 h22 + 122 h11 = 212 h22 .
0 = Ev = v hxu , xu i = 2 hxu , xuv i
Substituting the values we computed for the Christoffel sym-
= 2 hxv , xuu i + hxu , xuv i = 2u hxu , xv i = 2Fu .

bols in arbitrary orthogonal coordinates gives
This implies F 0 since we know F vanishes along . Lv NEv /2G = LEv /2E, Nu LGu /2E = NGu /2G.
Note that the special case where is itself a unit-speed
geodesic gives what are called Fermi coordinates (along ), Recalling that the mean curvature is H = L/2E + N/2G gives
often used in Lorenzian geometry for general relativity the first form.
(choosing to be the wordline of some particle). In this case, Now differentiating the equations L = k1 E, N = k2G yields
not only is the first fundamental form the identity matrix along (v k1 )E + k1 Ev = Lv = HEv , etc., which simplifies to the final
the whole starting curve, but also its derivative in the conor- equations in the statment.
mal direction vanishes, so all the Christoffel symbols vanish Lemma (Hilbert): Suppose p is a nonumbilic point with
along that curve. (In particular our claim is that Eu = 0 = Gu k1 > k2 and suppose k1 has a local maximum while k2 has a
along and this follows from calculations like the one above local minimum at p. Then K(p) 0.
for Fu , using the additional fact that is a geodesic.) End of Lecture 18 June 2015
Thm (Minding): Two surfaces with the same constant Proof: Choose curvature-line coordinates in a neighbor-
Gauss curvature K are locally isometric. Indeed, give a point hood of p. By assumption, the and use the classical notation
in each surface and orthonormal frames at these points, the lo- of the last lemma. We have L = k1 E, N = k2G. Differen-
cal isometry can be chosen to map the one frame to the other. tiating these equations and comparing the Codazzi equations
Pf: Construct Fermi coordinates in each surface, starting gives (v k1 )E + k1 Ev = Lv = HEv , etc., that is,
with a unit-speed geodesic (through the given point in the di-
rection of the first frame vector). The first fundamental form 2Ev k1 2Gu k2
in these coordinates will be given by 10 a02 and we have that Ev = , Gu = .
k2 k1 k1 k2
K = auu /a with a(0, v) 1. This can be viewed as the (ordi-
nary!) differential equation auu = Ka for a, which of course But by assumption the derivatives of the principal curvatures
has a unique solution given the initial conditions a(0) = 1, vanish at p, so by the final formulas of the last lemma, Ev =
au (0) = 0. Thus a is independent of v and is the same on both 0 = Gu there. Differentiating these formulas (and dropping
surfaces. terms involving first derivatives of ki to evaluate at p), we find
Note that of course we can solve this ODE explicitly. For that
K = 0 we have a 1 (as for the standard coordinates on R2 , 2E2vv k1 2G2uu k2
while for K > 0 we get a = cos( Ku), and for K < 0 we get Evv =
k2 k1
, Guu =
k1 k2
a = cosh( Ku).
Note that ordinary spherical coordinates (latitude, longi- at p. By assumption, at p we have k1 > k2 at p and also
tude) are Fermi coordinates (around the equator) for the round 2vv k1 0 2uu k2 . Thus Evv 0 and Guu 0 at p. The
sphere with K 1. The pseudosphere is an example of a sur- Gauss equation inorthogonal coordinates gave a nice formula
face with K 1, but no such surface is complete. for K involving EG. At a point where Ev = 0 = Gu , it is
Lemma: Any closed surface in R3 has at least one point easy to reduce this formula to 2EGK = Evv + Guu . It follows
(and hence an open set) where K > 0. immediately that K(p) 0.
Proof: Since M is compact, it is contained in some ball Note: Recall our earlier claim that the only surfaces with
around the origin. Let BR (0) be the smallest such ball. Its constant H and K (or equivalently, with constant principal cur-
boundary sphere (with normal curvatures 1/R) must be tan- vatures) are planes, spheres and cylinders. If k1 = k2 then we
gent to M. Since M stays inside, its normal curvatures in are in the totally umbilic case of planes or spheres. Otherwise,
particular both principal curvatures are at least 1/R. Thus at we can use the calculations above. We have 2vv k1 = 0 =
the point of tangency K > 1/R2 > 0. Evv = 0 and similarly Guu = 0, giving K 0. Surfaces with
18

K 0 are called developable; we will study these later. Our choice of at the basepoint (0), there is a unique smooth an-
first few results will suffice to conclude that the developable gle function along . For consideration of the angle, we might
surfaces here with constant H are round cylinders. as well assume that both vector fields have unit length.
Thm (Liebmann 1899): A smooth closed surface M R3 So suppose X is a unit parallel field along , meaning dt

X=
with constant Gauss curvature K is necessarily a round sphere. 0, and Y is any unit field. Let (t) be the angle from X to Y
Proof: By the first lemma, K > 0; the sphere we seek has at (t). As we observed earlier, parallel transport preserves
radius 1/ K. Denote the two principal curvatures by k1 k2 . the angle between two vectors, so is constant if Y is also
By compactness k1 attains a maximum at some p M (where parallel. In general, we claim that the rate of change in angle
k2 has a minimum, since k1 k2 K). If k1 (p) = K = k2 (p), measures the covariant derivative of Y.
the surface is totally umbilic and as we have seen already, it Any unit Y is perpendicular to its derivative. In particular,
is thus a piece of a sphere, indeed the whole sphere since M is

the covariant derivative dt Y, being perpendicular to as well,
closed. Thus we may assume k1 > k2 at p. But then we are in is a scalar multiple Y,

the scalar
of being given by the triple
Y, Y = Y, Y .

the situation of Hilberts lemma, so K 0, a contradiction. product dt
Thm (Liebmann 1900): A smooth closed surface M R3 In terms of the parallel field X, we can write Y = cos X +
with K > 0 and constant mean curvature H is necessarily a sin ( X). Taking the covariant derivative, we get
round sphere.

Proof: We proceed exactly as before, letting k1 attain its Y = sin X + cos ( X) = ( Y).
maximum at p. Again we just need to rule out the nonumbilic dt
case k1 > k2 . But here again Hilberts lemma applies to give
Equivalently, = dt Y, Y .

the contradiction K 0.
Note: One might ask if there are any other closed sur-
faces of constant mean curvature (CMC), perhaps even allow-
B15. The Umlaufsatz
ing immersed surfaces with self-intersections. Heinz Hopf
(1955) conjectured not, and proved this in the case the surface
is simply connected (topologically a sphere). A.D. Alexan- Note: from here on, these notes become more sketchy and
drov proved (1962) there are no embedded examples of any less detailed.
genus. It was a surprise then in 1986 when Henry Wente At the beginning of the semester, we mentioned the the-
found an immersed CMC torus. Since then, general methods orem on turning tangents, also known (even in English) as
for finding many examples have been developed. the Umlaufsatz: a simple closed plane curve (oriented with
the bounded region to its left) has turning number 1 (or equiv-
alently total signed curvature 2). This is actually the pla-
B14. More on geodesics nar case of the GaussBonnet Theorem. The local version of
GaussBonnet talks about a simple closed curve enclosing
a disk R on a surface M, and says the total geodesic curva-
We derived earlier the formula forR the
first
variation of ture of in M equals 2 minus the total Gauss curvature of R.
length of a space curve: len() = ,~ ds. Of course,
When M = R2 then of course K 0 and g = , so Gauss
if lies on a surface M, then this formula holds for all varia-
Bonnet does reduce to the Umlaufsatz; we use the Umlaufsatz
tions including those that keep on M. Their variation vector
in our proof of GaussBonnet.
fields are tangent to M. As before, up to reparametrization,
we can assume is normal to . Thus = is a varying R mul- End of Lecture 22 June 2015
tiple of the conormal vector . We get len = g ds. Hopf was not the first to prove the Umlaufsatz, but we will
Although only straight lines are length-critical with respect to sketch his proof. Let the curve be parametrized at unit speed
all variations in space, we see that, considering only varia- as an L-periodic map : R R2 . Shift the parameter if nec-
tions within M, a curve is length-critical if and only if it has essary, to ensure that (0) is an extreme point on the convex
g 0, that is, if and only if it is a geodesic. hull. For convenience, rotate so that (0) is a point with lowest
Here are a few facts without proof. We will return to them y-coordinate along . Then T (0) = 0 (0) = e1 is horizontal.
next semester for more general manifolds. First, sufficiently Now define the secant map
short arcs of any geodesic are length-minimizing. On any sur-
face, we can define a metric by setting d(p, q) to be the infi- (s) (t)
f (s, t) := S1 .
mal length of paths from p to q. (One shows that this infi- (s) (t)
mum never vanishes for p , q and that the metric topology
coincides with the usual topology on M.) If the infimum is Because has no self-intersections, this is well defined on
realized, that is, if a shortest path along M from p to q exists the diagonal strip s < t < s + L in the (s, t)-plane. On the
then it is a geodesic. If M is closed, then shortest paths always lower boundary s = t it is extended smoothly by f (s, s) =
exist. T (s), while on the upper boundary it is extended smoothly by
Now consider two nonvanishing vector fields X and Y, tan- f (s, s + L) = T (s). We will be interested in f restricted to
gent to M along some curve . As in our discussion of the the triangle
total curvature of plane curves, even though the angle be-
:= (s, t) : 0 s t L .

tween X and Y is only defined up to multiples of 2, given a
19

Just as we lifted the S1 -valued map T (s) to a real-valued an- i := i P (, ). The total signed curvature should be
gle function (s) when we defined turning number, here we replaced by i ; this will beP 2 times the turning number.
can lift the map f : S1 to a map : R such that Of course, the fact that i = 2 (or equivalently, i =
P
f (s, t) = cos (s, t), sin (s, t) . (For smooth functions like we (n 2) for an n-gon) is standard in elementary geometry.
have here, this is easiest to do by considering what the deriva- To prove it by induction starting from the known case of a
tives of must be. But such a lift exists for any continuous f , triangle, we just need to cut a larger polygon into pieces. (The
as one learns in algebraic topology.) The lift is unique up to one nontrivial lemma is that any simple polygon has some
adding a constant multiple of 2. We choose the lift for which diagonal that can be drawn without crossing any edges. In
(0, 0) = 0. fact one can always find an ear, such a diagonal that cuts
Along the diagonal, this (s, s) is the lift of T , so (L, L) is of just a triangle. Note that this is also the key lemma for a
by definition 2 times the turning number: our goal is to show polygonal version of the Jordan curve theorem.)
(L, L) = 2. Now consider the other sides of the triangle How about piecewise smooth curves? (Note: important to
, recalling that (0) was chosen to be a lowest point on the have smooth on closed subintervals to get well-defined one-
sided tangents.) Definition: over smooth parts plus turning
R
curve.
Along the vertical side, f (0, t) must point upwards, that is, angles at junctions. But how do we choose the sign at cusps
it stays in the (closed) upper semicircle. Thus the angle func- where = ? (Note: do Carmos definition here doesnt
tion (0, t), starting at (0, 0) = 0 must stay in the interval always work.) Right definition: + if cusp points out from
[0, ]. When we reach t = L, where f (0, L) = e1 we know bounded region, if cusp points into bounded region. From
(0, L) is (modulo 2) so the only possibility in the interval now on, we will write i [, ] for this turning angle at the
is (0, L) = . ith corner (with the correct choice of sign at cusps).
Continuing along the horizontal side, f (s, L) always points To prove the Umlaufsatz in this generality, one could try
downwards, staying in the lower semicircle. Thus, starting to take a limit of smooth or polygonal approximations but
at (0, L) = , we see that (s, L) stays in the interal [, 2]. preserving embeddedness and controlling the limiting value
When we reach (L, L), which must be a multiple of 2, we of total curvature is quite tricky. Hopf does it as follows: the
see it must be 2 as desired. secant map on the open triangle still limits to the tangent di-
One basic idea here is that a continuous function with dis- rection on all smooth points of the diagonal. The lifted map
crete (integer) values is constant. This is a recurring theme will jump at the points (s, s) corresponding to corners. We
in topology. Algebraic topology associates to any topological just need to show that the lifted map jumps by no more than
space X various algebraic objects (fundamental groups, ho- at each corner and that the sign is right where the jump
mology groups, etc.) and to any continuous map f : X Y is . (The sign that naturally comes up here is given by the
homomorphisms between the associated groups. For the cir- orientation of a triangle consisting of the cusp and two nearby
cle S1 , the fundamental group (or the first homology group) points chosen such that the curve avoids the segment between
is the integers; a map f : S1 S1 induces a homomorphism them. One can check that this is equivalent to our definition
f : Z Z this must be given by multiplication by some above.)
d Z, called the degree of f .
Given a smooth closed plane curve , its tangent vector can
B16. GaussBonnet: local form
be viewed as a map T : S1 S1 . The turning number of
is the degree of this map. For smooth maps like this, one can
also bypass the machinery of algebraic topology and define Now we want to turn towards GaussBonnet. As a first
the degree via differential topology, as we did by lifting to the result for curves on surfaces, consider a parametrized surface
angle function. Equivalently, the degree of f can be computed patch x : U M = x(U) and a simple closed (piecewise
by integrating the derivative of f Raround the circle just as we smooth) curve M. Consider the angle that T = 0 makes
computed turning number as 2 1
ds. with xu . Then we claim this changes by 2 as we go around
An alternative approach to degree is to note that almost ev- . (Same convention as above for corners.)
ery value in the range of f is a regular value, attained only at Proof: pull everything back to U. If the metric g were stan-
points where the differential of f is surjective. In particular, it dard, this would just be the Umlaufsatz. But we can con-
is attained only finitely many times, and each time has a well- tinuously deform g to the Euclidean metric (through convex
defined sign 1 depending on the orientation. The degree can combinations) an integer value must remain constant.
be computed at any regular value as the sum of these signs. End of Lecture 25 June 2015
(Another possibly familiar example of degree is the winding Let us define the total geodesic curvature TC() of a piece-
number used in complex analysis. If is a closed curve in wise smooth curve as g ds + i , where the integral is
R P
R2 r {0}, then the winding number of around 0 is the degree taken over each smooth subarc. We have not yet formally de-
of /|| : S1 S1 .) fined integrals over regions in a surface, but for any coordinate
We will want the Umlaufsatz not just for smooth curves but chart x(U) and any region D U, we have (just as for surface
for all piecewise smooth curves. As a warmup, consider what area)
it should say for polygons. The tangent vector T is constant Z "
along each side of a polygon but then jumps at each corner.
K dA =
p
K det g du dv,
If the interior angle is i , then the exterior or turning angle is x(D) D
20

and can check that this is independent of coordinates. The Interpretation in terms of holonomy of parallel transport.
local version of GaussBonnet then says (for a piecewise Gauss curvature as density limit over small disks around p.
smooth curve enclosing a disk R M to its left): Special case of sphere spherical polygons (esp. triangles).
So far we have GaussBonnet in the local form R K dA =
R
Z
K dA = 2 TC(). 2TC(R) for any disk R contained in an orthogonal coordi-
R nate patch and with peicewise smooth boundary. A comment
on orientation is in order. Of course any coordinate patch is or-
Additivity under splitting a region.
intable. If we switch orientation, then K is unchanged, so the
Following do Carmo, we prove this directly under the as-
whole GaussBonnet equation must be unchanged. Indeed,
sumption that R is contained in a orthogonally parametrized
our convention that R is oriented with R to the left depends
neighborhood x(U). (The case of larger disks is then a spe-
on the surface normal, so T switches sign. That means how-
cial case of one of the global versions of GaussBonnet.) We
ever that the conormal is unchanged, so the total geodesic
write D := x1 (R) U and write for the piecewise smooth
curvature of R is unchanged. A better way to express the ori-
boundary curve of D with = x .
entation convention for R might be to simply say the conor-
Let us first examine the right hand side. By the extension
mal should point inwards towards R.
to the Umlaufsatz, 2 is the total turning of the tangent vector
relative to xu (with the appropriate convention at corners). By
definition, TC is the total turning of the tangent vector relative B17. Global topology of surfaces
to a parallel field (with the same convention at corners). Thus
2 TC is the total turning of xu relative to a parallel field.
To consider the global forms of GaussBonnet, we need
Passing
to the orthogonal unit vectors e1 := xu / E and e2 :=
to discuss the topology of surfaces. A regular region R on
xv / G, we find that the rate of turning of xu (relative to a a smooth surface will mean a compact subset R which is the
parallel field) is closure of its interior and whose boundary is a finite disjoint
1
union (possibly empty) of simple closed, piecewise smooth
T e1 , e2 = T xu , xv .

curves. Topologically, R is therefore a compact 2-manifold
EG with boundary, a compact Hausdoff space locally homeomor-
With T = uxu + vxv , we have

phic to the closed half-plane (x, y) : y 0 . (That is, each
point p R has a neighborhood homeomorphic either to the
T xu , xv = u112 + v122 .

plane or to the half-plane.)
One way to build such a topological space is to start with a
But in orthogonal coordinates we computed 112 = Ev /2 and finite collection of triangles and zip certain pairs of edges
122 = Gu /2. Thus the total turning, the integral of the rate of together. (Any edge that is not paired remains part of the
turning, is boundary.) This is called a triangulation of the surface. A
Z difficult theorem of Rad (1925) says that any topological 2-
Gu v Ev u
dt. manifold with boundary can be triangulated. For our smooth
2 EG regions R in space, this is not so difficult. Although we skip
On the other hand, using our formula for K, we get the details, the idea is that if we tile R3 with a fine enough
cubic lattice adjusted to be transverse to R then each small
" cube contains just a single disk of R with piecewise smooth
Z
K dA = K EG du dv boundary: a polygon. Of course it is easy to cut an n-gon into
D
"R n 2 triangles. Given an atlas of coordinate charts for a sur-
Ev Gu
= v + u du dv. face, note that we may assume the triangulation is fine enough
R 2 EG 2 EG that each triangle lies in one of the charts.
Now we recall Greens theorem in the plane: if P and Q are End of Lecture 29 June 2015
two functions on a region R bounded by a (piecewise smooth) We now want to give a topological classification of regu-
curve , then lar regions R, that is, of compact 2-manifolds with boundary,
" Z Z that is, of spaces obtainable by zipping triangles together. For
u Q v P du dv = P du + Q dv = (Pu + Qv) dt. this discussion, we use the word surface to mean such a topo-
R logical space (rather than a smooth surface in R3 as usual).
(This is a special case of Stokes Theorem, of course already We follow the description by Francis and Weeks of Conways
known to Gauss in other forms.) Weapply Greens theorem ZIP proof.
with P = Ev /2 EG and Q = Gu /2 EG to give Let us first describe the statement. To perforate a surface
is to delete an open disk. (A sphere with one perforation is
Ev u + Gu v a closed disk; a sphere with two perforations is an annulus
Z Z
K dA = dt. also called a cylinder.) To add a handle or cross-handle
D 2 EG
to a surface is to perforate it twice, and then sew in an an-
This agrees with the expression we got above for 2 TC, so nulus connecting the new boundaries. (Or equivalently, then
we have proved the GaussBonnet theorem. zip these boundaries to each other.) To add a cross-cap to a
21

surface is to perforate it once and then sew in a Mbius band sum the GaussBonnet relation over all triangles. Each R edge
along the new boundary. (Or equivalently, then zip the two is used twice, with
R opposite orientations, so the terms g ds
halves of the boundary together.) cancel out. Thus M K dA equals the sum over all triangles of
Adding a cross-handle is the same as adding two cross- a + b + c , where a, b, c are the interior angles. This is the
caps. (The Klein bottle is the union of two Mbius bands.) sum of all interior angles minus F. But grouping the angle
Thm: Any surface is a finite union of components, each sum around the vertices, it is 2V. Thus we get
being a sphere with a certain number of perforations, handles, Z
and cross-caps.
K dA = 2(V F/2) = 2(M).
Pf: The starting collection of triangles is of this form. It M
suffices to show that a single zipping (involving one or two
components) preserves this. First consider the case where en- An alternative way of doing the bookkeeping is to start with
tire boundary components are zipped together: either join two total charge 2(M) by putting charges +2 at each vertex and
components or add a (cross-)handle. If the two edges to be in each face and 2 on each edge. Then move the charges
zipped are instead the two halves of one boundary compo- from the vertices and edges into the faces, based on angles and
nent, we add a cross-cap or remove a perforation. Finally, if total curvatures. We are left with chargeR in each face, equal
the edges to be zipped are just subarcs of the cases considered (by the local form of GaussBonnet) to K.
so far, then the effect is the same except that we are left with For a general region R, we do the same thing. But along
boundary edges, g ds does not cancel out. Similarly, at
R
one or two more perforations.
Note that adding a cross-cap (or a cross-handle) makes a boundary vertices, the sum of interior angles is not 2. Also
component non-orientable. An orientable component is thus 2E = 3F must be corrected by the number of boundary edges.
g,k , a sphere with g 0 handles and k 0 perforations. Putting it all together, we get:
On a non-orientable component there is no way to distinguish Z
handles from cross-handles. Thus it is has the form Nh,k , a K dA = 2(R) TC(R).
sphere with h 1 cross-caps and k 0 perforations. We can R
restate the classification as follows.
This is the most general form of GaussBonnet, with the pre-
Thm: A connected surface has the form g,k if orientable or
vious versions (for disks and closed surfaces) as special cases.
Nh,k if nonorientable.
There are several immediate applications.
Note that the closed nonorientable surfaces Nh,0 cannot be
embedded in R3 , while all the other types can be.
Any closed surface in R3 with K > 0 has > 0 so it is
The Euler number of a triangulation is := V E + F.
homeomorphic to a sphere or a projective plane; if it is
This is clearly the sum of the Euler numbers of the compo-
embedded it must be a sphere.
nents. For a single triangle = 1. Zipping a pair of edges
leaves F unchanged and decreases E by one; the effect on V If there are two closed geodesics on a surface with K >
varies. But tracing through the cases considered above shows 0, then they intersect (because otherwise they would
that the Euler number depends only on the topology. Indeed bound an annulus with = 0). (Note that Lyusternik
g,k = 2 2g k while Nh,k = 2 h k. When thought and Shnirelman proved that any sphere has at least three
of as an invariant of a topological space, is called the Euler different simple closed geodesics.)
characteristic.
Note that the topological type of a connected surface is thus A simple closed geodesic on a surface with K 0 can-
determined by its orientability, its Euler number , and its not
R bound a disk to either side (because such a disk has
number k of boundary components. K = 2).
There is no geodesic 1-gon or 2-gon (disk) on a surface
B18. GaussBonnet: global forms with K 0 (because geodesics that are tangent coin-
cide, so the exterior angles are not but strictly less).
Recall that a regular region R on a smooth surface M in R3 The angle excess of a geodesic triangle has the same
is compact with piecewise smooth boundary. Every point in R sign as the average (or total) Gauss curvature in the tri-
has a neighborhood with an orthogonal parametrization, and angle.
R can be triangulated by triangles, each containted in such a
parametrized neighborhood. To integrate a function f over R,
we sum over the triangles: B19. The Gauss image and total absolute curvature
Z XZ XZ
f dA = f dA = f det gk duk dvk .
p
The GaussBonnet theorem shows that K is a density im-
R Tk x1
k (T k )
portant is its integral over a region. There is also an extrinsic
Note that
R our local form of GaussBonnet applies to each tri- interpretation.
angle: T K dA = 2 TC(T ). Near any point p M where K , 0 the Gauss map is
Consider first a closed surface R = M. Counting edges locally an immersion M S2 orientation-preserving if K >
of triangles gives 2E = 3F, so (M) = V F/2. Now we 0 and orientation-reversing if K < 0. Its Jacobian determinant
22

K = det D gives the factor by which area is stretched here where P is, say, a Cantor set (R); and the example of two
we mean an algebraic signed area. (Think about curvature- cones where P is a single line at which the ruling is Lipschitz
line coordinates.) but not C 1 as in Khnel.)
We can say R K dA equals the signed area of (R) S2 .
R
We follow a paper of Massey, as summarized in do Carmos
(And we can recover K(p) as the limit of area (R) / area(R) book.

as the region R shrinks down to the point p.) At each p U there is a unique asymptotic direction. Inte-
UsingRthe appropriate notion of area with multiplicities, we grating these, we foliate U by a unique family of asymptotic
can say R K dA = area(|R ) for any region R M. lines. The first claim is that these curves are straight lines in
space. Of course, along these asymptotic curves which are
End of Lecture 6 July 2015 also lines of curvature, the surface normal is constant. (But
Consider a closed surface M R3 . An analog of the Jor- compare top curve of round torus asymptotic=curvature line
dan curve theorem says that it divides space into one bounded of parabolic points, but not straight need to know conormal
and one unbounded region; we can orient it with the outward derivative of K vanishes.)
unit normal (pointing into the unbounded region). Thus M is Proof: Locally on U we can use curvature-line coordinates
necessarily an orientable surface of some genus g 0. where, say, the u-curves are the asymptotics. Then is a func-
By GaussBonnet, the area of the Gauss image (M) in S2 tion of v alone. Now consider the real-valued function hx, i
equals 2 = 4(1 g). That is, the Gauss map covers the on U. Since xu and u 0, its u-derivative vanishes,
sphere 1 g times (in an oriented sense) the degree of the so it is a also function of v alone: hx, i = (v). Differenti-
Gauss map is 1 g. By degree theory, almost every point ating gives hx, v i = 0 (v). Note that v , 0 since we are at
w S2 has finitely many preimages, 1 g of them if counted nonplanar points; of course v (like ) is constant along each
with signs. That is, if there are k preimages with K > 0 then asymptotic curve. Thus each of these equations is the equa-
there are k + g 1 with K < 0. tion of some plane in space (depending on v); the planes are
(Note that the points with normal = w are exactly the orthogonal. The u-curve v = const. lies in the intersection line
critical points of the function h, wi : M R.) of these planes.
Now consider bringing a plane with given normal w S2 We next claim that we may assume u is the arclength pa-
in from infinity until it first touches M. Any point of contact rameter along each u-curve (asymptotic line).
p M has K 0, since M stays stays to one side of the plane. Proof: Recall that in curvature-line coordinates with g =
E 0 and h = L 0 , the Codazzi equations become L =
This means every point w S2 has at least one preimage with 0 G 0 N v
K 0. (That is, k 1 in the counts above.)
R Let M denote the HEv and Nu = HGu . Here we have L 0, so Ev 0, Rmeaning
regions where K > 0. Then we have M K dA 4. With that E is a function of u alone. Thus defining s = E du
+
R
GaussBonnet, this gives M K dA 4g. Subtracting these (independent of v), this is arclength along each u-curve. Then
(s, v) are equally valid curvature-line coordinates, where g =
gives M |K| dA 4(1 + g).
R
1 0 and h = 0 0 .
0G 0 N
Equality holds in these last three inequalities if and only if The condition E 1 means that the new coordinates are
all points p M with K > 0 are extreme on the convex hull. not only curvature-line coordinates but also simultaneously
Such a surface is called tight. An interesting theory of tight geodesic parallel coordinates. As before we write G = a2
surfaces characterizes them, for instance, as exactly those sur- (with a = |xv |) and have K = auu /a. Thus here auu = 0,
faces having the two-piece property of being cut into no more meaning that a is a linear function along each asymptotic line
than two pieces by any plane. (that is, for each v).
Now consider the case that K > 0R holds on all
R of M. (So by Now consider again the Codazzi equation Nu = HGu =
GaussBonnet, M is a sphere and K dA = |K| dA = 4.) 2Haau , where the mean curvature satisfies 2H = N/G =
The Gauss map is bijective, indeed a diffeomorphism M N/a2 . Combining these gives
S2 . Each point p M is an extreme point on the convex hull,
that is, M is globally convex (as we mentioned before. Nu au
The results on total absolute curvature do not hold for ab- = ,
N a
stract surfaces. For instance, R2 /Z2 is a torus with a flat
(K 0) metric. An abstract surface of genus g > 1 can be meaning that for each v (i.e., along each asymptotic curve), N
given a hyperbolic metric with K 1. By GaussBonnet its is a constant multiple of a. That is N = c(v)a. Finally consider
area is then 2 = 4(g 1). the principal radius of curvature r = 2H1 2
= aN = c(v)
a
in the xv
direction. This (like a and N) is a linear function along each
asymptotic line.
B20. Developable Surfaces Lemma: An asymptotic line through a point p U, even if
extended indefinitely, never reaches P.
We now want to consider a surface M R3 with K 0. Proof: The mean curvature H is continuous on M. It van-
Each point on M is parabolic a closed subset P M consists ishes on P but along any asymptotic line in U is the (nowhere
of planar points; the open complement U := M r P consists vanishing) reciprocal of a linear function.
of nonumbilic points. (Note the example of a triangle joined Prop: The boundary U = P M is a union of open line
to three cylinders as in do Carmo; the example of a cylinder segments. (Note: might be uncountable union!)
23

Proof: Consider a boundary point p and some neighbor-

hood V parametrized as a graph over T p M. The set U V is
foliated by lines, which in the projection do not cross. Thus in
a smaller neighborhood V 0 of p their directions are given by a
Lipschitz function. Thus (consider the lines ` pi through points
in U V 0 approaching p) there is a a well-defined limiting
direction at p. Any point in U along the resulting line ` p is a
limit of points along the ` pi and in particular is in U.
Thm: A complete surface with K = 0 is a cylinder over
some plane curve.
Proof: First we claim that the direction of the asymptotic
lines is locally constant on U. Along any line, the radius of
curvature is a linear function, which can never vanish on a
smooth surface. Thus r is constant on each line. It follows
from the equations above (in local coordinates) that a = |xv |
is also constant, implying that xv is constant along each line.
That is, 0 = xvu = xuv . Thus xu (which we know is constant
along each line, of course) is locally constant as claimed.
At points of U we defined lines whose direction was a
limit; using the Lipschitz condition, we find the direction is
actually locally constant on U U. (We dont get, say, a
Cantor function!)
Finally consider the interior PrU of P. It consists of open
pieces of planes, bounded by complete lines. Each such piece
must then be an infinite strip, bounded by two parallel lines.
We can foliate it by further parallel lines.
Thus, as claimed, all of M is foliated in this way by parallel
lines.
24

J.M. Sullivan, TU Berlin A: Manifolds Diff Geom II, WS 2015/16
A. MANIFOLDS These conditions are familiar from metric spaces (X, d),
where the open balls B (x) := {y : d(x, y) < } form a base
for the metric topology. The topological spaces that will
This course is about manifolds. An m-manifold is a space
arise for us are all metrizable, meaning the topology arises
that looks locally like Euclidean space Rm .
from some metric. In particular, the standard topology
Examples of manifolds include a circle S1 , a sphere S2 or on Rn comes, of course, from the Euclidean inner prod-
a torus T 2 = S1 S1 or any surface in R3 . These generalize uct (scalar product)phx, yi = x y = xi yi via the metric
P
for instance to an m-sphere Sm Rm+1 or an m-torus T m = d(x, y) = |x y| = hx y, x yi.
S1 S1 .
In multivariable calculus, one studies (smooth) m-sub- Definition A1.6. A space X is Hausdorff if any two dis-
manifolds of Rn ; these have several equivalent characteri- tinct points x , y X have disjoint (open) neighborhoods.
zations (locally being level sets or images of smooth func- It is regular if given a nonempty closed set A X and a
tions). These are the motivating examples of (smooth) point x X r A, there are disjoint (open) neighborhoods
manifolds. Indeed, we will see any manifold can be em- of A and x. (These are just two examples of the many
bedded as a submanifold of some (high-dimensional) Rn . separation axioms in point-set topology.) A space X is
But it is important to give an abstract definition of mani- second countable if there is a countable base for the topol-
folds, since they usually dont arise as submanifolds. ogy.
The idea of manifolds is that they are spaces on which one Metric spaces are Hausdorff and regular (take metric
can do analysis (derivatives, integrals, etc.). This means neighborhoods of radius d(x, A)/2). Eucldiean space is
we are talking here not simply about topological mani- second countable (take balls with rational centers and
folds, but about smooth (differentiable) manifolds. This radii). The importance of these notions is clear from the
is a distinction we explain soon. Urysohn metrization theorem, which says that X is separa-
ble (that is, has a countable dense subset) and metrizable if
and only if it is Hausdorff, regular and second countable.
A1. Topological manifolds
Definition A1.7. We say a space M is locally homeomor-
phic to Rm if each p M has an open neighborhood U
Definition A1.1. A topology on a set X is a collection U that is homeomorphic to some open subset of Rm . If
of subsets of X (called the open subsets) such that : U (U) Rm denotes such a homeomorphism,
, X U, then we call (U, ) a (coordinate) chart for M. An atlas
for M is a collection (U , S

) of coordinate charts which
U, V U = U V U, covers M, in the sense that U = M.
{U } U = U U.
S
Clearly we can rephrase the definition to say M is locally
All other topological notions are defined in terms of open homeomorphic to Rm if and only if it has an atlas of charts.
sets. Less obvious (but an easy exercise) is that it is equivalent to
Definition A1.2. A subset A X is closed if its com- require each p M to have a neighborhood homeomorphic
plement X r A is open. Any subset Y X naturally to Rm .
becomes a topological space with the subspace topology: Although one might expect this to be a good topological
{U Y : U X open}. That is, the open sets in Y are definition of an abstract manifold, it turns out that there are
exactly the interestions of Y with open sets in X. some pathological examples that we would like to rule out.
Certain properties from point-set topology are not auto-
Definition A1.3. A space X is connected if and X are matically inherited. For instance, examples like the line R
the only subsets that are both open and closed. A space with the origin doubled (or with all x 0 doubled) fail
is compact if every open cover has a finite subcover. (If to be Hausdorff. The long line (obtained by gluing un-
we talk about a subset Y X being connected or compact, countably many unit intervals) fails to be second countable
etc., we mean with respect to the subspace topology.) it is sequentially compact but not compact. The Prfer
Definition A1.4. A map f : X Y between topological surface is separable but not second countable. (Note there
spaces is continuous if the preimage of any open set in Y are also much weirder examples, for instance in papers of
is open in X. A continuous bijection f : X Y whose Alexandre Gabard.) For technical reasons, we also prefer
inverse is also continuous is called a homeomorphism an manifolds to have at most countably many components, as
equivalence of topological spaces. is guaranteed by second countability.
Thus we are led to the following:
Definition A1.5. Usually, one specifies a topology not by
listing all open sets, but by giving a base B. This is a col- Definition A1.8. A topological m-manifold is a second-
lection of basic open sets sufficient to generate the topol- countable Hausdorff space M = M m that is locally homeo-
ogy: an arbitrary U X is defined to be open if and only morphic to Rm .
if it is a union of sets from B. The requirements on B to
form a base are (1) that B covers X and (2) that intersec- Regularity then follows, ensuring that our manifolds are
tions B1 B2 of two basic open sets are open, that is, for metrizable spaces. Indeed, we will later put a (Rieman-
any x B1 B2 , there exists B3 B with x B3 B1 B2 . nian) metric on any (smooth) manifold. It is also straight-

forward to check various other local properties: A topo- Let us look at the lowest dimensions. A 0-manifold is a
logical manifold M is locally connected, locally compact, countable discrete set. Equivalently, we can say a con-
normal and paracompact (defined later when we need it). nected 0-manifold is a point. In general, of course, any
Being separable and locally compact, it is also globally the manifold is a countable union of connected components,
union of countably many compact subsets. so it makes sense to classify connected manifolds. It is
Note: For m , n, it is easy to see there is no diffeomor- not hard to show that any connected 1-manifold is (home-
phism Rm Rn It is also true that there is no homeomor- omorphic to) either R1 or S1 . If we allow manifolds with
phism, but this requires the tools of algebraic topology like boundary, there are just two more examples: the compact
homology theory. Any Rm is contractible, so they all have interval, and the ray or half-open interval.
the same (trivial) homology. The trick is to first remove While the complete classification of noncompact surfaces
a point. Then Rm r {0} contracts to Sm1 , and spheres of is known, we will only consider the compact case. We have
different dimension have different homology. This was the seen the examples S2 , T 2 and RP2 . It turns out that any
start of topological dimension theory, and shows that every connected closed surface is a connect sum of these: either
(nonempty) manifold has a well-defined dimension. an orientable surface g of genus g or a nonorientable sur-
Examples A1.9. face Nh . (And if we allow compact surfaces with boundary,
then all we get are these examples with some number k of
Rm is an m-manifold (with a single chart). open disks removed, denoted g,k or Nh,k .)
An open subset U M m of an m-manifold is itself The uniformization theorem, a classical result in complex
an m-manifold (restricting charts to U). analysis and Riemann surface theory, implies that any sur-
Any smooth surface M 2 R3 is a 2-manifold. (Get face admits a metric of constant Gauss curvature. Indeed,
a chart around p M by projecting orthogonally to the sphere and the projective plane have spherical (K 1)
T p M.) metrics, the torus and Klein bottle have euclidean (flat,
Other surfaces like polyhedral surfaces are also K 0) metrics, and all other closed surfaces have hyper-
topological manifolds. bolic (K 1) metrics.
More generally, any smooth m-submanifold in Rn is Guided by this, Bill Thurston conjectured a method to
an m-manifold. (We will consider such examples in understand compact 3-manifolds (with boundary). In
general later.) 2003, Grigory Perelman proved this geometrization con-
jecture, establishing that any 3-manifold can be cut into
RPm := Sm / = {lines through 0 in Rm+1 } is an m- pieces, each of which admits one of eight standard geome-
manifold called real projective space. tries. There is interesting work remaining to be done to
M m N n is an (m + n)-manifold (using product better understand the case of hyperbolic 3-manifolds.
charts). In dimensions four and higher, there is in some sense no
For any smooth surface M 2 R3 , the tangent bundle hope of classifying manifolds. Given any finite group
T M = {(p, v) : p M, v T p M} is a 4-manifold presentation, one can build a closed 4-manifold with that
fundamental group. Since the word problem is known
One of our first tasks will be to define T p M for an abstract to be undecidable, it is impossible in general to decide
smooth manifold M m ; in general we will find that it is a m- whether 4-manifolds are homeomorphic. Much interest-
dimensional vector space and that these can be put together ing research thus restricts attention to the case of simply
to form a 2m-manifold, the tangent bundle. connected manifolds (with trivial fundamental group).
In some cases it is important to consider also manifolds In certain other ways, higher dimensions are easier to un-
with boundary, modeled on the halfspace derstand. One reason is that two generic 2-disks will have
empty intersection in dimensions five and above. Thus, for
H m := (x1 , x2 , . . . , xm ) Rm : x1 0 ,

instance, the Poincar conjecture was first proved in these
whose boundary is H m = {x1 = 0} Rm1 . dimensions.
Definition A1.10. An m-manifold with boundary is then a

second-countable Hausdorff space locally homeomorphic A2. Smooth structures
to H m . If a point p M is mapped to the boundary in
one chart, then this is true in every chart. Such points form
the boundary M M of M; it is an (m 1)-manifold If (U, ) and (V, ) are two charts for a manifold M m , then
(without boundary, and perhaps empty). The complement
MrM is called the interior and is an m-manifold (without 1 : (U V) (U V)
boundary).
is a homeomorphism between open sets in Rm , called a
We will use manifolds with boundary later when we study change of coordinates or transition function. The inverse
integration and Stokes theorem. Until then, we will ba- homeomorphism is of course 1 .
sically neglect them, with the understanding that all our Since the transition functions are maps between Euclidean
theory extends in the obvious way. The following ter- spaces, we know how to test how smooth they are. Sup-
minology is standard even if confusing at first: a closed pose U Rm is open and f : U Rn . To say f is C 0
manifold is a compact manifold without boundary. just means that it is continous. If f is differentiable at each

p U, then its derivative is a function D f : U Rnm . later. Note that it is easy to tell if a diffeomorphism is
We say f is C 1 if D f is continuous. By induction, we say orientation-preserving; for C 0 manifolds one needs homol-
f is C r if D f is C r1 , that is, if Dr f is continuous. If f has ogy theory.)
(continuous) derivatives of all orders, we say it is C . If Given a C r -structure on any manifold, for any s r, by the
a C map f is real analytic, meaning that its Taylor series lemma it extends to a unique C s -structure. On the man-
around any p U converges to f , then we say f is C . ifold m r
m R , the standard C structure arises from the atlas
Definition A2.1. Fix r {0, 1, 2, . . . , , }. We say (R , id) consisting of a single chart. If we let U r denote
two charts (U, ) and (V, ) for a manifold M m are C r - the collection of all charts C r -compatible with this one,
compatible if the transition functions 1 and 1 then we have
are C r maps. (They are then inverse C r diffeomorphisms.)
U0 U1 U2 U U.
A C r -atlas for M is a collection of C r -compatible coor-
dinate charts which covers M. A C r -structure on M is a
maximal C r -atlas, that is an atlas U = (U , ) such that

any coordinate chart (V, ) which is compatible with all End of Lecture 12 Oct 2015
the (U , ) is already contained in U. A C r -manifold is a The point of a smooth structure is to know which mappings
topological manifold M m with a choice of C r -structure. A are smooth. Suppose f : M m N n is a (continuous) map
chart for a smooth manifold will mean a chart in the given between two smooth manifolds. Given p M, we can find
smooth structure (unless we explicitly refer to a topolog- (smooth) charts (U, ) around p M and (V, ) around
ical chart). f (p) N. Then the composition
Of course the case r = 0 is trivial: any atlas is C 0 and the f 1 : U f 1 (V) Rn

C 0 -structure is the set of all possible topological charts. (In
this case, of course, one should use the term homeomor- is called the expression of f in these coordinates. (Writ-
phism instead of C 0 -diffeomorphism.) ing (x1 , . . . , xm ) for a typical point in (U) Rm and
This course is about smooth manifolds, where we use the (y1 , . . . , yn ) for a typical point in (V) Rn , then we can
word smooth to mean C . When we say manifold think of f 1 very explicitly as n real-valued func-
we will mean smooth manifold unless we explicitly say tions, giving the y j as a function of (x1 , . . . , xm ).)
otherwise. Of course many of our results will be valid even Now it is easy to define smoothness:
for lower degrees of smoothness (usually C 1 or C 2 or C 3
would suffice) but we will not attempt to keep track of this. Definition A2.3. The map f is smooth if for each p we can
find charts (U, ) and (V, ) as above such that f 1 ,
Lemma A2.2. Any C r -atlas is contained in a unique max- the expression of f in these coordinates, is smooth (as a
imal one. map between euclidean spaces). (If M and N are only C r -
manifolds, then it makes sense to ask if f : M N is C s
Proof. Given an atlas U = (U , ) , let V be the col-

for s r but not for s > r.) A diffeomorphism f : M N
lection of all charts (V, ) that are compatible with every between two smooth manifolds is a homeomorphism such
(U , ). We just need to show that V is a C r -atlas, that that both f and f 1 are smooth. The set of all smooth maps
is that any charts (V1 , 1 ) and (V2 , 2 ) are compatible with M N is denoted by C (M, N); we write C (M) :=
each other. But any p V1 V2 is contained in some U , C (M, R).
and on 1 V1 V2 U we can write
Exercise A2.4. If f : M m N n is smooth, then its ex-
1 = 2 1 .
2 1 1 1
pression f 1 in any (smooth) coordinate charts is
smooth.
At the end of this proof, we implicitly use three properties:
Note two special cases: if M = Rm (of course with the
r
The composition of two C functions is C . r
standard smooth structure), then we can take = id and
The restriction of a C r function to an open subset is thus consider f ; if N = Rn then we can take =
Cr . id and consider f 1 . For a map f : Rm Rn , we
take = id and = id and see that our new definition of
A map that is C r is some neighborhood of each point
smoothness agrees with the one we started with for maps
in U is C r on U.
between euclidean spaces.
Without getting into formal details, these properties mean The basic constructions of new manifolds from old open
that the class of C r diffeomorphisms form a pseudogroup subsets and products can be adapted to the smooth set-
(of homeomorphisms on the topological space Rm ). ting.
Although we have only defined C r structures, other kinds If U M m is an open subset of a smooth manifold, then we
of structures on manifolds arise from other pseudogroups. can restrict the smooth structure on M to a smooth struc-
For instance, a projective or Mbius structure on M arises ture on U (which we already know is a topological man-
from an atlas where the transtion functions are all projec- ifold). In particular, each chart (V, ) for M gives a chart
tive or Mbius transformations (respectively). An orienta- (V U, |VU ) for U. (If we talk about a smooth map on an
tion on M arises from an atlas where all transition func- open subset U of a smooth manifold M, then we implic-
tions are orientation-preserving. (We will return to this itly mean smooth with respect to this restricted structure.)

smooth structure on Rm for m , 4.

Suppose we have a cover U of M and a map f : M N.
Then f is smooth if and only if its restriction to each U
is smooth. (This is a version of the pseudogroup property The Hauptvermutung (known by that name even
above.) in English) of geometric topology (formulated 100
If M m and N n are smooth manifolds, then it is a straight- years ago) suggested that every topological manifold
forward exercise to put a smooth structure on the manifold should have a unique piecewise linear (PL) structure
M N. (Hint: Use only product charts (U V, ) essentially given combinatorially by a triangula-
obtained from smooth charts (U, ) and (V, ).) tion and a unique smooth structure. This is now
known to be false.
Every smooth manifold has a (PL) triangulation. For

A3. Exotic smooth structures
every dimension m 4, there are topological m-
manifolds that admit no triangulation and in par-
Suppose h : M m N m is a homeomorphism between ticular no PL or smooth structures. (For m = 4 this
topological manifolds (so that M and N are really the has been known since the 1980s, but for m > 4 it
same topological manifold). Then h can be used to was just proven in 2013!)
move other structures between M and N. A trivial exam-
ple would be a real-valued function f : N R; it can be There are uncountably many different smooth struc-
pulled back to give the real-valued function f h on M. tures on R4 . It is unknown if there is any exotic
Of interest to us is the case of a C r -structure U on N (a smooth structure on S4 .
maximal atlas). We can use the homeomorphism h to pull
In higher dimensions, some things get easier. In
it back to give a C r -structure h (U) on M: the pull-back of
dimensions m 7, for instance, there are exotic
a chart (U, ) U is the chart h1 (U), h for M. Almost
spheres Sm , but these form a well-understood finite
by definition, h: M N is then a C r -diffeomorphism
group (e.g., there are 28 for m = 7). In general, the
from M, h (U) to (N, U).
differences between smooth and PL manifolds (and
Since M and N are homeomorphic, they are really the same to some extent between PL and topological mani-
topological manifold, and we might as well be considering folds) can be analyzed for m 5 by means of alge-
self-homeomorphisms h : M M. If h = id then clearly braic topology.
h (U) = U; more generally this is true any time h is a dif-
feomorphism from the smooth manifold (M, U) to itself. For compact, simply connected topological 4-mani-
But suppose h is a homeomorphism that is not a diffeo- folds, Freedman showed how to use invariants from
morphism. Then h (U) is a distinct smooth structure on algebraic topology to check when they are homeo-
the manifold M. Consider a couple of examples on the morphic. In most (but not all) cases we know which
line M = R, starting with its standard smooth structure of these topological manifolds admit smooth struc-
U; the pull-back structure h (U) is the one generated by tures; it is not known how to classify the smooth
the single coordinate chart (R, h). If h : x 7 x3 then h is structures when they do exist.
smooth but its inverse is not, so with respect to h (U) it is
easier for maps into M to be smooth, but harder for maps Especially since we know there are exotic spheres in cer-
from M to be smooth.
(The reverse is true of course if we tain dimensions, it is important to say what we mean by
start with x 7 3 x.) If on the other hand h : x 7 2x + |x|, the standard sphere Sm as a smooth manifold. The right
then neither h nor its inverse is smooth. (Note that in all answer is that it inherits a smooth structure as a smooth
these examples, the meaning of smoothness changes only submanifold of Rm+1 , but since we havent developed that
near 0.) theory yet, we use explicit charts. Any obvious atlas will
Such examples are weird, but in fact they are all trivial. give the same standard smooth sphere, for instance the two
As we noted charts of stereographic projection:
above, h is always a diffeomorphism from
M, h (U) to (M, U). Thus the two smooth manifolds
x
are diffeomorphic to each other really the same smooth U = Sm r {em+1 }, (x, z) = , x Rm , z R,
manifold. We have merely put on strange eyeglasses the 1z
map h to relabel the points of M.
or the 2m + 2 charts of orthogonal projection:
More interesting is the question of existence of exotic
smooth structures can two different (nondiffeomorphic) U j = x Sm Rm+1 : sgn x j = 1 ,

smooth manifolds have a homeomorphism between them
j (x) = x1 , . . . , xbj , . . . , xm+1 .

(meaning that their underlying topological manifolds are
the same). There are still many interesting open questions
here, especially in dimension 4. The following facts are It is a good exercise to check that all these charts are C -
known: compatible.
With our basic constructions, we then get many further
Up to diffeomorphism, there is a unique smooth examples of smooth manifolds, like the m-torus T m =
structure on any topological manifold M m in dimen- S1 S1 (a product of circles) or the n2 -dimensional
sion m 3. Up to diffeomorphism, there is a unique matrix group GLn R Rnn (an open subset).

A4. Smooth maps, rank, immersions Corollary A4.3. Suppose U M m is open, f : U R is

smooth and p U is given. Then there is a neighborhood
V U of p and a smooth map g : M R such that g f
Suppose f : M m N n is smooth, and g = f 1 is its
on V and g 0 outside U.
expression in some local coordinates (U, ) around p M
and (V, ) around f (p) N. Then we define the rank of f
Proof. Since M is locally compact, we can choose neigh-
at p to be the rank of g at x = (p), that is, the rank of the
borhoods
Jacobian derivative matrix (g j /xi ) there.
That this is well-defined (independent of coordinates) fol- p V1 V1 V2 V2 U
lows from the chain rule: a change of coordinates (on one
side or the other) would multiply the Jacobian by an in- with Vi compact. (Quite explicitly, we can take any coordi-
vertible square matrix (on one side or the other), leaving nate neighborhood (V3 , ) with p V3 U and (p) = 0
the rank unchanged. Note that the rank can be at most and (V3 ) = B3 (0); then we set Vi := 1 Bi (0) .) Use the
min(m, n). theorem to find with 1 on V1 and 0 outside V2 .
The rank theorem from multivariable calculus can be re- Then define g to equal f on U and to be 0 outside V2 .
stated most nicely for smooth manifolds. (When stated for (Each of these is a smooth function on an open set; they
Euclidean spaces it needs to mention diffeomorphisms.) agree on the intersection. Thus they define a smooth g on
the union, which is all of M.)
Theorem A4.1 (Rank Theorem). Suppose f : M m N n
is a smooth map of constant rank k. Then for each p M
there are coordinate neighborhoods (U, ) of p and (V, ) A5. Tangent vectors and tangent spaces
of f (p) such that f 1 is the orthogonal projection
map
We know that the tangent space T p Rn at a point p Rn is
(x , . . . , x ) 7 (x , . . . , x , 0, . . . , 0),
1 m 1 k
a copy of the vector space Rn ; a vector v T p Rn can be
viewed intuitively as an arrow from p to p+v. (Technically,
where of course there are n k zeros at the right. (Note: of course, a tangent vector knows where it is based, so we
if we want, we can require that (U) = B1 (0) and (V) = could set T p Rn = (p, v) : v Rn , but usually we write

B1 (0).) just v for (p, v).)
In particular, maps of maximum rank are important. We If is a curve through p := (0) in Rn , then its velocity
say f is a submersion if f has constant rank n m. We 0 (0) is best viewed as a vector in T p Rn . If M m is an m-
say f is an immersion if f has constant rank m n. For submanifold through p Rn , then the tangent space T p M m
m = n these notions coincide. A map f : M m N m is a is an m-dimensional linear subspace of T p Rn , consisting of
diffeomorphism if and only if it is bijective and has con- all velocity vectors to curves lying in M.
stant rank m. For an abstract manifold M m , its tangent space T p M m
Now recall that smooth functions (unlike analytic func- should still be the collection of velocity vectors to curves
tions) are quite flexible. Starting with a function like through p M. But of course, there are always many
curves with the same tangent vector. One approach would
be to define tangent vectors as equivalence classes of

0,
x 0,
h : R R, h(x) =

e1/x , x > 0,
curves, but when are two curves equivalent? One could
say: when they agree to first order, but this begs the
we can use it to give a smoothed step function (0 for x 0 question.
and 1 for x 1), and then a smoothed bump function, A good approach is to think about what we use tangent
which can also be rotated into higher dimensions. vectors for: to take directional derivatives! If g : Rn R
For the next proof, we thus fix a smooth function g on Rm is a real-valued function, and is a curve through p = (0)
with g(x) 1 for |x| 1 and g(x) 0 for |x| 2. with velocity v = 0 (0) there, then the directional deriva-
tive of g is the derivative along :
Theorem A4.2. Suppose F and K are disjoint closed sub-
sets of a (smooth) manifold M m , with K compact. Then d
v g = D p g(v) = (g ).
there is a smooth function : M R with values in [0, 1] dt t=0
such that 0 on F and 1 on K.
Let us think about v as a map taking g to the real number
Proof. Each point p K has a coordinate neighborhood v g R. This is linear:
U p M r F such that (p) = 0 and p (U p ) = B3 (0)
Rm . Then we can define a smooth function g p : M R v (g + h) = v g + v h
by g p := g p on U p , extending it to be 0 outside U p .
and satisfies the Leibniz product rule:
The open sets V p := 1

p B1 (0) cover K M r F (since
there is one for each p K). By compactness, we get v (gh) = (v g)h(p) + (v h)g(p).
a finite subcover V p1 , . . . , V pk . Then if we define :=

1 ki=1 1 g pi , it is easy to check 1 on K and 0
Q
Such a map is called a derivation. Furthermore, it is local
on F. in the sense that v g only depends on values of g in an

arbitrarily small neighborhood of p. A fact we will check star is used to indicate a pull-back, a map associated to f
later is that there are no other local derivations at p Rn acting in the opposite direction.)
besides these directional derivatives. Thus we can use this Like any linear map between vector spaces, f induces a
as the definition of tangent vector. dual map between the dual spaces; here we claim this re-
So fix a point p in a (smooth) manifold M m . What is the stricts to a map f : T p M T q N. Working out what the
right domain for a derivation (think of a directional deriva- dual map means, we find that for X p T p M and g C (q)
tive) at p M? Consider the class we have
[
C= C (U) f (X p ) (g) = X p f (g) = X p (g f ).

U3p
This linear map f is called the differential of f at p and
of all real-valued functions g defined on some open neigh- we will usually write it as D p f . (Other common notations
borhood of p. If two functions agree on some neighbor- include d p f or simply f 0 .)
hood, then they must have the same derivatives at p, so we
consider them to be equivalent. More precisely, g : U R Theorem A5.1. Given a smooth map f : M m N n of
is equivalent to h : V R if there is some open W 3 p manifolds and a point p M, the construction above in-
(with W U V) such that g|W = h|W . An equivalence duces a linear map f = D p f : T p M T f (p) N, the differ-
class is called a germ (of a smooth function) at p. The set ential of f .
of germs at p is the quotient space C/ =: C (p). If g is a
function on a neighborhood of p, we often write simply g Proof. The many claims we made during the construction
for its germ (which might more properly be called [g]). are all routine to check. We gives just two examples. To
see that f (X p ) is actually a tangent vector at q := f (p), we
Note that if g C (p) is a germ, we can talk about its
need to check the Leibniz rule:
value g(p) R at p, but not about its value at any other
point. (For M = Rm , a germ g at p also encodes all dervia- f (X p )(gh) = X p (gh) f = X p (g f )(h f )

tives at p that is, the Taylor series of g but also much
= X p (g f ) h(q) + X p (h f ) g(q)

more information, since g is not necessarily analytic.)
= f (X p )g h(q) + f (X p )h g(q).

The set C (p) of germs is an (infinite dimensional) algebra
over R, that is, a vector space with multiplication. (Exer- To see that f is linear, we compute:
cise: check that multiplication of germs makes sense, etc.)
We now define a tangent vector X p at p M to be a deriva- f (X p + Y p )(g) = (X p + Y p )(g f )
tion on this algebra. That is, X p : C (p) R is a linear
= X p (g f ) + Y p (g f )
functional:
= f (X p ) + f (Y p ) (g).

X p (g + h) = X p g + X p h

satisfying the Leibniz rule:
It is now a straightforward exercise to check the functori-
X p (gh) = (X p g)h(p) + (X p h)g(p).
ality of the operation f 7 f , that is, the following two
We let T p M denote the tangent space to M at p, that is, the properties:
set of all such tangent vectors X p .
For f = id : M M, the maps f and f are also
End of Lecture 19 October 2015 the identity maps.
Clearly T p M is a vector space, with the obvious operations If h = g f (for maps between appropriate mani-
folds), then h = f g and h = g f .
(X p + Y p ) f := X p f + Y p f.
(The second of these is of course the chain rule from cal-
(In fact, this is just exhibiting T p M as a linear subspace culus.)
within the abstract dual vector space C (p) of all linear
functionals on C (p). Corollary A5.2. If f : M N is a diffeomorphism, then
for any p M, the map D p f : T p M T f (p) N is an iso-
Note that if U M is open with p U, then T p U = T p M
morphism. In particular, if (U, ) is a coordinate chart
since the set C (p) of germs is the same whether we start
for M m , then : T p M T (p) Rm is an isomorphism
with M or U.
Now suppose f : M m N n is a smooth map of manifolds Of course, this refers to T q Rm in the sense we have just de-
and consider a point p M and its image q := f (p) N. fined for abstract manifolds. It is time to go back and prove
If g is a germ at q, then g f is a germ at p. (Here of the claim we made early on, that there are no derivations
course, we really compose f with any of the functions in on Rm other than the usual directional derivatives, that is,
the equivalence class g.) This gives a map that T p Rm Rm .
f : C (q) C (p), f (g) := g f We know we have a map Rm T p Rm which associates to
each v Rm the directional derivative v at p. This map is
between these algebras of germs, which we claim is linear, clearly linear and is easily seen to be injective. Indeed, if
indeed an algebra homomorphism. (Note that the upper i : Rm R denotes the projection p 7 pi , then v i =

vi = i (v); now two distinct vectors v , w must differ in P i that if {e

Note i } is the standard basis of Rm (so that v =

some component vi , wi , meaning v i , w i , so v , v ei ), then ei = xi is the corresponding standard ba-
w . The claim that now remains is just that this map is sis of T p Rm . Recall that, given a coordinate chart (U, )
surjective there are no other derivations. around p M m , the differential D p = : T p M
T (p) Rm Rm is an isomorphism. Under this isomor-
Lemma A5.3. Suppose X p T p Rm and g C (p) is
constant (in some neighborhood of p). Then X p g = 0. phism, the x i correspond to the elements of a basis for
T p M, which we write as
Proof. By linearity of X p it suffices to consider g 1. By

!
the Leibniz rule, i = i,p :=
1
.
xi
X p (1) = X p (1 1) = X p (1) 1 + X p (1) 1 = 2X p (1)
Suppose a function f C (U) has coordinate expression
which clearly implies X p (1) = 0. f 1 : (U) R, then at p U we get
Our next lemma can be thought of as a version of Taylors
!
theorem. (Note that one could let B be an arbitrary star- i f =
1
( f ) = f 1
.
xi xi
shaped region around p.)
In particular, if we consider the individual components
Lemma A5.4. Let B := B (p) where p Rm and > 0.
x j = j of the coordinates as real-valued functions,
For any g C (B), we can find a collection of m functions
g we find i j ) = j /xi = ij . We can express any
hi C (B) with hi (p) = x i (p), such that on B we have
X p T p M in terms of our basis {i } as follows:
X
g(x) = g(p) + (xi pi )hi (x). m
X
Xp = X p (i ) i

i
i=1
Proof. If we set
Consider a smooth map f : M m N n between manifolds,
1
g
Z
i
p + t(x p) dt and choose local coordinates (U, ) around p M and

h (x) :=
0 x i
(V, ) around q := f (p) N. (We write xi = i and
then the desired properties follow from the fundamental y j = j .) In these coordinates, f is represented by the
theorem of calculus in the form map f 1 of Euclidean spaces, or more explicitly as
Z 1 functions y j = f j (x1 , . . . , xm ). Here the derivative is given
d
g(x) = g(p) + g p + t(x p) dt, by the Jacobian matrix

0 dt
y j f j
! !
noting that this t-derivative is the directional derivative of J= = .
g in direction x p. xi xi
Theorem A5.5. The map v 7 v is a (natural) isomor- Let us write {i } as usual for the coordinate frame of T p M,
phism Rm T p Rm . where i = 1 /x . For T q N, we use the notation
i
j = /y . We find that J is the matrix of D p f with

1 j
Proof. As noted above, all that remains is to prove surjec- respect to these bases. That is,
tivity. Given X p T p Rm , define v Rm by vi := X p (i ).
We claim v = X p . By definition, these agree on (the germs X y j !
D p f i,p ) = j,q ,
of) the projections i . Now suppose g C (p) is any xi (p)
j
germ. Finding a representative g C B (p) for some
> 0, we can use the second lemma to write or equivalently, if X p = vi i,p and f (X p ) = w j j,q ,
P P
then we have
X
g = g(p) + (i pi )hi .
i y
j
!
i X
w =
j
v .
Then by the definition of derivation, i
xi (p)
X
X p g = X p g(p) + X p i X p pi hi (p)

We should have waited until now to define the rank at p
M of a map f : M m N n . It is simply the rank of the
X
+ (X p hi ) i (p) pi .

linear map D p f : T p M T f (p) N. In coordinates this is,
Here the last sum (which would seem to involve second of course, the rank of the Jacobian matrix above, as we
derivatives of g) vanishes simply because i (p) = pi . And defined before.
i
One special case is when f = id M is the identity map. That

the terms X p p and X p g(p) vanish by the first lemma.
Thus we are left with is, we have overlapping coordinate charts (U, ) and (V, )
X X i g for M m . At any point p U V, we have two different
Xpg = (X p i ) hi (p) = v i (p) = v (g) coordinate bases for T p M, which we write as {i } (with
x
i
respect to ) and { i } (with respect to ). Then the change-
as desired. of-basis matrix is just the Jacobian matrix of the coordinate

expression of id M , which here is just the transition function by a smooth function f C (M) pointwise: ( f X) p =
1 . (This is the basis for a definition of tangent vectors f (p)X p . That is, the vector fields X(M) form not just a real
still popular among physicists: a tangent vector is its ex- vector space, but in fact a module over the ring C (M) of
pression in a coordinate base, with the rules for changing smooth functions.
this covariantly when we change coordinates.) Given a vector field X and a function f , we can also define
As usual, we also consider the special cases where one of X f C (M) by (X f )(p) := X p f R. Note the distinc-
the manifolds M or N is (a submanifold of) R. For R of tion between the vector field f X (given by pointwise scalar
course we use the standard chart (the identity map), and we multiplication) and the function X f (given by directional
write t for the corresponding basis vector for the tangent derivatives of f ).
space to R at any point. Exercise A6.4. Each of the following conditions is equiv-
A map : (a, b) M m is a curve in M. Its tangent vector alent to the smoothness of a vector field X as a section
at p := (t) M is 0 (t) := Dt (t ) T p M. X : M T M:
The opposite case is a real-valued function f C (M).
For X T p M, we have D p f (X) T f (p) R, so D p f (X) =
For each f C (M), the function X f is also
t for some R. Of course, this is just the direc- smooth.
If we write X|U =: v i in a coordinate chart
P i
tional derivative X f . For instance, in local coordinates,
(D p f )(i ) = (i f )t . We write d f = d p f : T p M R for (U, ), then the components vi : U R are smooth.
the linear map X 7 X f . The dual vector space T p M is
called the cotangent space and its elements are cotangent End of Lecture 26 October 2015
vectors (or covectors for short). Thus d p f T p M is the
covector given by d f (X) := X f ; this is just another way to
view the differential since we have D p f (X) = d p f (X)t . A7. Submanifolds
A6. The tangent bundle The canonical example of an m-dimensional submanifold

of an n-manifold is Rm 0 Rn , the set of vectors whose
last n m components vanish.
Definition A6.1. The tangent bundle T M = T (M) to
a smoothSmanifold M m is, as a set, the (disjoint) union Definition A7.1. Given a manifold N n , we say a subset
T M := pM T p M of all tangent spaces to M; there is M N is an m-submanifold if around each point p M
obviously a projection : T M M with 1 {p} = T p M. there is a coordinate chart (U, ) for N in which M looks
We can equip T M in a natural way with the structure of a like Rm 0 Rn . That is, in such a preferred chart we
smooth 2m-manifold. Start with a (smooth) atlas for M. have
Over any coordinate chart (U, ), therePis a a bijection
D : T U (U) (M U) = (U) (Rm 0).
R R sending i v i T p U =
m 2m i
T p M to (p), v . We define the topology on T M by speci-

fying that these maps D are homeomorphisms; they then It is straightforward then to check that M (with the sub-
form an atlas for T M as a topological 2m-manifold. space topology) is an m-manifold. Indeed, the preferred
Exercise A6.2. These charts for T M are C -compatible, charts form a C atlas for M m .
and thus define a smooth structure on T M. Two alternative local characterizations as for submani-
Note that T M is an example of a vector bundle, which is folds in Rn are then immediate. A submanifold M m N n
a special kind of fiber bundle to be defined later. Without can be described locally (that is, in some neighborhood
going into details, a fiber bundle with base B and fiber F is U N of any point p M) as
a certain kind of space E with projection : E B such
that the preimage of any point b B is isomorphic to F. 1. the zero level set of a submersion N n Rnm (here
A trivial bundle is E = F B projecting to the second composed with projection onto the last n m co-
factor. Any fiber bundle is required to be locally trivial in ordinates), or
the sense that B is covered by open sets U over which the
2. the image of an immersion Rm N n (here the stan-
bundle is trivial (F U). A section of a bundle : E B
dard inclusion Rm , Rn composed with 1 ).
is a continuous choice of point in each fiber, that is, a map
: B E such that = idB .
We now want to consider in more detail the description
Definition A6.3. A (smooth) vector field X on a manifold of submanifolds via immersions. Immersions from Rm are
M m is a smooth choice of a vector X p T p M for each also known as regular parametrizations. Recall that last
point p M. That is, X is a (smooth) section of the bundle semester we used such regular parametrizations to describe
: T M M, meaning a smooth map X : M T M such curves and surfaces in R3 . It is of course important that the
that X = id M . We write X = X(M) = (T M) for the set parametrization be an immersion, in order to be sure that
of all vector fields. the image is a smooth submanifold.
We define addition of vector fields pointwise: (X + Y) p = Examples A7.2. Consider the following examples of im-
X p +Y p . Similarly, we can multiply a vector field X X(M) mersions based on smooth plane curves.

1. t 7 (cos 2t, sin 2t) is a periodic parametrization We will show that the image of a smooth embedding is a
of a simple closed curve, a 1-submanifold. This im- submanifold (and vice versa); the embedding is then not
mersion is not injective, but becomes injective if we merely a homeomorphism but indeed a diffeomorphism
consider the domain to be the circle R/Z S1 . onto its image.
Above we saw many examples of injective immersions of
2. If r : R (1, 2) is strictly monotonic, then manifolds which were not embeddings. However, there is
one standard result from point-set topology which guaran-
t 7 r(t) cos 2t, r(t) sin 2t tees that this never happens when M is compact.
is an injective immersion whose image is a submani- Proposition A7.4. If X is compact and Y is Hausdorff,
fold, a spiral curve in the plane. Note that this image then any continuous bijection f : X Y is a homeomor-
is not a closed subset of R2 (because the immersion phism.
is not proper).
For the proof, recall that a map f : X Y is open if the
3. t 7 (cos 2t, sin 4t) is again a closed curve, this image of every open set U X is open in Y, and f is closed
time a figure-eight. It descends again to the quotient if the image of every closed set A X is closed in Y. If f
cirle R/Z, but is not injective even there. The image is a bijection, then these notions are equiavalent, and also
is not a submanifold. equivalent to f 1 being continuous.
4. If we restrict this last example to the open interval Proof. We need to show f 1 is continuous, or equivalently
(1/4, 3/4), which of course is diffeomorphic to R, that f is a closed map. So suppose A X is closed; we
we get an injective immersion whose image is still need to show f (A) is closed in Y. Since X is compact,
the whole figure-eight curve, not a submanifold. A is also compact. Since f is continuous, f (A) is then
compact. But a compact subset of the Hausdorff space Y
5. One can build an injective immersion whose image
is necessarily closed.
is not even locally connected. For instance, join the
topologists sine curve, the curve t 7 (1/t, sin t)
Two examples related to the quotient map I S1 (where
for t 2, to a downward ray in the y-axis, the curve
I = [0, 1] and the quotient identifies the endpoints {0, 1} to
t 7 (0, t) for t 1, via a smooth intermediate arc
a single point) show why the two conditions are necessary.
for t [1, 2].
First, we can get a bijection by restricting this map to the
6. For any slope R, we can project the line t 7 noncompact interval [0, 1). Second, we can get a bijection
(t, t) of slope from R2 to the quotient torus T 2 = by replacing S1 by a non-Hausdorff circle with a doubled
R2 /Z2 . For = p/q Q, this gives a periodic curve, basepoint (like our line with doubled origin).
that is, a circle submanifold in T 2 called a (p, q)- Corollary A7.5. If X is compact and Y is Hausdorff, then
torus knot. For irrational on the other hand, the any (continuous) injection f : X Y is a topological em-
immersion is injective but its image X is dense in T 2 bedding.
and thus is not a submanifold (in our sense).
Corollary A7.6. If M is a compact manifold, then any in-
In some other contexts, mainly that of Lie groups, exam- jective immersion f : M m N n is a smooth embedding.
ples like this last one can be considered as submanifolds. A
Lie group is smooth manifold with the structure of an alge- Now we show that the concepts of smooth embedding and
braic group, where the group operations are smooth maps; submanifold coincide in the following sense:
we will discuss these later. The torus T 2 is an example of a
(compact, 2-dimensional) Lie group (under addition). The Theorem A7.7. If f : M m N n is an embedding, then
dense subset X consisting of points of the form (t, t) f (M) N is a submanifold and f : M f (M) is a dif-
is a subgroup; from the point of view of Lie groups this feomorphism. If M m N n is any submanifold, then the
is a 1-dimensional Lie subgroup. Of course X T 2 with inclusion i : M , N is an embedding.
the subset topology is not a manifold. Instead we simply
use the bijective immersion R X to transfer the stan- Proof. For the first statement, consider a point p M and
dard smooth manifold structure from R to X. (Indeed, any its image q = f (p) f (M) N. Because f has constant
time we have an injective immersion f : M m N n , it is rank m n, by the rank theorem, we can find coordinates
a bijection onto its image, and could be used to transfer (U, ) around p M and (V, ) around q N in which f
the topology and smooth structure from M to that image, looks like the embedding Rm , Rn . It is tempting to hope
making f by definition a diffeormorphism, though not to a that (V, ) is the preferred chart we seek in the definition
subspace of N.) of submanifold but we have not yet used the fact that
f is an embedding and the problem is that other parts of
Definition A7.3. A continuous injection f : X Y of f (M) might enter V, while we want f (U) = f (M) V.
topological spaces is a topological embedding if it is a But since f is an embedding, f (U) is open in f (M) in
homeomorphism onto its image f (X). A (smooth) embed- the subspace topology from N. By definition of subspace
ding f : M m N n of manifolds is an immersion that is a topology, this means there is an open subset W N such
topological embedding. that f (U) = W f (M). Now we simply restrict (V, ) to

W V and we find these are preferred coordinates showing It follows that each t : M M is a diffeomorphism, with
f (M) as a submanifold of N (around q). inverse t . Note that since R is an abelian group (s + t =
We essentially proved the second statement when we put a t + s), these diffeomorphisms all commute:
smooth structure on the submanifold M N: we remarked t s = s+t = s t .
then that the manifold topology on M was the subspace
topology from N, which exactly means the inclusion is a This is often simply called a one-parameter group action
topological embedding. The fact that it is an immersion is or a (global) flow on M.
also obvious in a preferred coordinate chart. Definition A8.1. We say a vector field X is invariant under
m n the action if it is invariant under each t , that is, if t X =
Suppose M N is a submanifold. Then at any p M
X for all t.
we can view T p M T p N in a natural way as a vector
subspace (using the injective differential D p i of the inclu- This may seem like a very special situation, but we will see
sion map i). More generally, if M is described (locally) it is quite natural.
as the image of a regular parameterization, an immersion
Our notation t (p) := (t, p) emphasizes the diffeomor-
Rm U N, then T p M is the image of its differen-
phisms t obtained by fixing t R. If instead, we fix a
tial. If instead M is (locally) the zero set of a submersion
point p M, we of course get a curve p : R M defined
f : N Rnm , then T p M is the kernel of D p f .
by p (t) := t (p). The trace of this curve is the orbit of
p M under the action . It is helpful to rewrite the defin-
ing property s t = s+t of a flow in terms of these flow
A8. Vector fields and their flows curves. For any point q := p (s) = s (p) along p , we find
that the curve q is just a reparametrization of p ; indeed
Suppose f : M m N n is a smooth map and X is a vector
q (t) = t (q) = t s (p) = t+s (p) = p (s + t).

field on M. For any point p M, we can use f = D p f
to push a vector X p T p M to a vector at f (p) N. If Equivalently, q = s p :
there exists a vector field Y on N such that for each p M
q (t) = s t (p) = s p (t) .

we have Y f (p) = D p f (X p ), then we say Y is f -related to X.
Of course, when f is not injective, it might be impossible Definition A8.2. The infinitesimal generator of the flow
to find an f -related vector field; when f is not surjective, is the vector field X on M defined by X p := 0p (0), the
Y is not uniquely determined away from f (M). But when velocity vector of the curve p at p = p (0).
f is a diffeomorphism, there clearly is a unique Y that is
f -related to any given X, and then we write Y = f (X). End of Lecture 2 November 2015
If f : M M is a diffeomorphism it can happen that a An equivalent way to define the infinitesimal generator X
vector field X is f -related to itself: X = f (X). In this case, comes from looking at a standard vertical vector field on
we say X is f -invariant. As a simple example, consider R M, defined by
the radial field X p = p on Rm . It is invariant under any
homothety f : p 7 p. This may seem like a very spe- T (t,p) (R M) = T t R T p M 3 (t , 0) =: V(t,p) .
cial situation, but in fact our goal now, given an arbritrary
Then it is easy to check that X p = D(0,p) (V).

vector field X, is to construct a one-parameter family of
diffeomorphisms t : M M under which X is invariant. Theorem A8.3. Suppose is a flow on M with infinitesi-
Recall that if G is an algebraic group and X is any set, then mal generator X. Then X is -invariant. That is, for any
an action of G on X is a map : G X X, often written s R and p M we have
as s (X p ) = Xs p .

(g, x) 7 g x := g (x),
Proof. Write q = s (p) so that X p = 0p (0) and Xq = q0 (0).
satisfying the following properties: Then the desired formula follows immediately from the
e = idX , gh = g h . observation above that q = s p .
(That is, in the typical group theory notation, e x = x and Corollary A8.4. If X p = 0 then the curve p is the con-
(gh) x = g (h x).) Each g : X X is a bijection (with stant map p (t) p. If X p , 0, then the curve p is an
inverse g1 ). The action partitions the set X into orbits immersion. If it is not injective on R then it is s-periodic
and injective on R/sZ for some s > 0.
G x := {g x : g G},
which are the equivalence classes under the equivalence Proof. First note that
relation x g x. 0p (t) = X p (t) = t X p .

We are interested in smooth actions of the (1-dimensional
Since t is a linear isomorphism, these vectors either re-
Lie) group (R, +) on a smooth manifold M m . Such an ac-
main zero or remain nonzero. Finally, if we have a non-
tion is a smooth map : R M M, again written as
injective curve with p (t + s) = p (t) for some s and t,
(t, p) 7 t (p), satisfying
then the same holds for this s and every t, that is, p is
0 = id M , s t = s+t . s-periodic.
10

Recall that these curves are the orbits of the flow and Like any local result, this can be transferred immediately to
form a partition of M. (Each orbit is topologically an open the context of an arbitrary manifold, where its restatement
arc or a closed loop or a single point.) has a more geometric flavor.
Note also that the vector field X is -related to the vector
Definition A8.8. A local flow around p M is a map
field V = (t , 0) on R M since
X(t,p) = D(t,p) V(t,p) .
: (, ) V M
As we have seen, any flow has an infinitesimal genera- (for some > 0 and some open V 3 p) such that 0 (q) = q
tor X. What if we start with a vector field X on M: does it for all q V, and
generate a flow? We will see that the answer is always yes
t s (q) = t+s (q)

when M is compact; in general the flow might exist only
for small t (depending on p).
whenever both sides are defined. The flow lines are the
Definition A8.5. Given a vector field X on a manifold M, curves q (t) := t (q); the infinitesimal generator is the vec-
a curve : J M (where J R is some open interval) is tor field Xq = q0 (0) tangent to the flow lines.
called an integral curve of X if 0 (t) = X(t) for all t J.
Theorem A8.9. Any vector field X on a manifold M has a
As we have seen, any R-action has an infinitesimal gen- local flow around any point p M.
erator X. Then each orbit p (t) = t (p) is an integral curve
of X, defined on J = R. In other cases, the integral curve Note that if we prove this theorem by appealing to the pre-
does not exist for all time, since it flows out of M in finite vious theorem on Rm , then the neighborhood V we con-
time. For instance, consider the flow t (x) = x + te1 on Rm , struct (and even the values of ) will be in some coordinate
whose infinitesimal generator is X = 1 . If we replace Rm chart around p. But this doesnt affect the statement of the
by an open subset (like Rn r {0} or Rn r B1 (0)) then we theorem.
sometimes leave this open subset in finite time.
Theorem A8.10. On a compact manifold M m , any vector
Think for a minute about dimension m = 1. Up to dif- field X has a global flow.
feomorphism, there is no difference between reaching the
end of a finite open interval like (0, 1) and reaching .
Proof. For any p M we have a local flow, defined on
A classical example is the flow of t2 t on R, that is, the
some ( p , p ) V p . By compactness, finitely many of the
solution of the ODE du/dt = u2 , which blows up in finite
V p suffice to cover M. Let > 0 be the minimum of the
time. So it is too much to hope for a global solution in
corresponding (finitely many) p . Then we know that the
general. But of course, standard theorems on ODEs guar-
flow of X exists everywhere for a uniform time t (, ).
antee local existence and uniqueness of solutions, which
But then, for instance using nt = t t , we can
can be viewed as integral curves of a vector field. (In an
construct flows for arbitrary times.
ODE course, you might learn about minimal smoothness
conditions for existence and for uniqueness; certainly C
or even C 1 suffices for both.)
A9. Lie brackets and Lie derivatives
Theorem A8.6. Suppose U Rm is open and f : U Rm
is smooth. Then for each p U, there is a unique solution
to the equation dx/dt = f (x) with initial condition x(0) = Definition A9.1. A Lie algebra is a vector space L with
p; it is smooth and is defined on some maximal open time an antisymmetric (or skew-symmetric) product
interval (a p , b p ) 3 0.
L L L, (v, w) 7 [v, w] = [w, v]
A proof of this basic result (using the Banach fixed-
point theorem for contraction mappings) can be found in that is bilinear (i.e., linear in v and in w, where it suffices
Boothby. Somewhat more subtle is the smooth depen- to check one of these:
dence on parameters as given in the next theorem. (See
Conlons textbook for a proof.) In our version, there are no [v + v0 , w] = [v, w] + [v0 , w],
parameters other than the initial point p.
noting this could also be called a distrubutive law) and sat-
Theorem A8.7. Suppose U Rm is open and f : U Rm isfies the Jacobi identity
is a smooth function. For any point p U there exists
> 0, a neighborhood V U of p, and a smooth map u, [v, w] + v, [w, u] + w, [u, v] = 0.

x : (, ) V U A trivial example is any vector space with the zero prod-

satisfying uct [v, w] := 0. The Jacobi identity may not at first seem
intuitive, but in fact there are some familiar nontrivial ex-
x amples.
(t, q) = f x(t, q) , x(0, q) = q

t Example A9.2. Three-space R3 with the usual vector cross
for all t (, ) and q V. product [v, w] := v w = v w is a Lie algebra.
11

Example A9.3. The ordinary matrix product on Rnn is bi- is a vector field, that is, that it is local and satisfies the
linear but neither symmetric nor antisymmetric. But the Leibniz product rule. But locality the fact that the value
matrix commutator [A, B] := AB BA is clearly antisym- of [X, Y] f at p depends only on the germ of f at p and not
metric. To check the Jacobi identity, we compute on its values elsewhere is clear, giving
A, [B, C] = ABC ACB BCA + CBA [X, Y] p : C (p) R, f 7 X p (Y f ) Y p (X f ).

and then cyclically permute. The bracket notation for Lie To show [X, Y] p T p M we now check the Leibniz rule:
algebras comes from this earlier use of brackets for com-
[X, Y] p ( f g) = X p Y( f g) Y p X( f g)

mutators.
= X p f Yg + gY f Y p f Xg + gX f

Example A9.4. More generally (and more abstractly), sup-
pose V is any vector space, and consider the set End(V) := = (X p f )(Y p g) + f (p)X p (Yg)
L(V, V) of linear endomorphisms (self-maps) on V. Then + (X p g)(Y p f ) + g(p)X p (Y f )
the commutator (Y p f )(X p g) f (p)Y p (Xg)
[A, B] := A B B A (Y p g)(X p f ) g(p)Y p (X f )
= f (p)[X, Y] p g + g(p)[X, Y] p f
is again a Lie product.
Exercise A9.6. Of course the Lie bracket [X, Y] is R-
Now consider the set X(M) of smooth vector fields on a bilinear, but it is not C (M)-bilinear. Instead we have
manifold M m . As we have observed, this is an (infinite-
dimensional) vector space over R and indeed a module [ f X, gY] = f g[X, Y] + f (Xg)Y g(Y f )X.
over C (M), where for X, Y X and f C the vector
field f X + Y is defined pointwise: (Hint: consider first the case [ f X, Y] = f [X, Y] (Y f )X
and then use that twice, with the antisymmetry.)
( f X + Y) p = f (p)X p + Y p .
Lemma A9.7. Let i X(U) denote as usual the coordi-
But we also recall that a vector field X X gives (or indeed nate basis vector fields on a chart (U, ). Then their Lie
can be viewed as) a map C (M) C (M) via f 7 X f , brackets vanish: [i , j ] = 0 for all i, j = 1, . . . , m.
taking directional derivatives of f in the directions X p .
That is, we can view vector fields as endomorphisms of Proof. Let f be a germ at p U, and write f for the
C (M): germ f 1 at (p) Rm . By definition of i , we have
(i f )(p) = f/x (p) . Then for the Lie bracket we
i
X(M) End C M . get:

As Lie observed, the commutator product on End(C M) [i , j ] p f = i ( j f )(p) j (i f )(p)

in fact restricts to the subspace X: 2 f 2 f
= (p) (p) = 0,

i j
j i
Theorem A9.5. The space of vector fields X(M) is a Lie dx dx dx dx
algebra with the Lie bracket using the fact that the mixed partials commute.
[X, Y] f := X(Y f ) Y(X f ). Exercise A9.8. Using this lemma and the result of the pre-
vious exercise, compute the formula for the Lie
P bracket
[X, Y] in coordinates, if X = i i and Y = i i in a
P
Note that the ordinary composition product does not re-
strict: the mapping f 7 X(Y f ) is an endomorphism of chart (U, ).
C (M) which should be thought of as taking a second Of course, vector fields are used to take derivatives of func-
derivative of f in particular directions; this does not cor- tions. If X X(M) and f C (M) then (X f )(p) = X p f is
respond to a vector field, because second derivatives do a directional derivative of f . If is an integral curve of X
not satisfy the Leibniz product rule. But partial derivative through p and its (local) flow, then
commute; in the commutator above the second-order terms
cancel, leaving only first-order terms, that is, a vector field

d d
X p f = f p (t) = f t (p) .

[X, Y]. To understand why there can be first-order terms dt t=0 dt t=0
remaining, recall the formula for the second derivative of a
function f along a curve in Rn passing through (0) = p Now suppose we want to take a derivative of a vector
with velocity 0 (0) = v and acceleration 00 (0) = a: field Y along a curve . The problem is that for each
q = (t), the vector Yq lives in a different tangent space
d2 T q (M). So we cannot compare these vectors or ask for
f (t) = D p f (v, v) + D p f (a).
2
2
dt t=0 their rate of change along without some sort of addi-
tional information. Later in the course, we will introduce
Proof. We need to check that the endomorphism the notion of a connection (for instance coming from a
Riemannian metric), which does allow us to differentiate
[X, Y] : C (M) C (M) a vector field along a curve. But Lie suggested a different
12

approach. Suppose we have not just one (integral) curve A10. Vector bundles
p but a whole vector field X and its associated local flow
around p. Then we can use
When we defined the tangent bundle T M of a manifold
D p t = t : T p M T t p M M m , we mentioned that it is a specific example of a smooth
to identify the tangent spaces along the integral curve p . vector bundle over M, with fibers the tangent spaces T p M.
In particular, for each t (in the interval (, ) of definition) In general, a bundle over a base space M consists of a total
we have t Yt p T p M.
space E with a projection : E M. When there is no
confusion, we often refer to E as the bundle. The fiber over
Definition A9.9. If X, Y X(M) then the Lie derivative p M is simply the preimage E p := 1 {p}. If S M, we
LX Y of Y with respect to X is the vector field defined by write ES for the restriction of E to S , that is, the bundle
|S : ES = 1 (S ) S . If : E M and 0 : E 0 M

d
LX Y) p := t Yt p T p M.

dt t=0 are two bundles, then : E E 0 is fiber-preserving if
0 = , that is, if (E p ) E 0p for each p.
It would be straightforward but tedious to check in coor-
Of course, we only call : E M a bundle if the fibers are
dinates that this is a smooth vector field. For us, that will
all isomorphic (in an appropriate sense, to some F); and if
follow from the theorem below, saying that the Lie deriva-
E is locally trivial, locally looking like a product with F.
tive is nothing other than the Lie bracket:
We now give a precise definition for the case of interest
LX Y = [X, Y] = [Y, X] = LY X. here.
For this we first need the following lemma, a modification
Definition A10.1. A (smooth) vector bundle of rank k is a
of the Taylor-type lemma we used to prove T p Rm Rm .
map : E m+k M m of manifolds such that
Lemma A9.10. Suppose a vector field X X(M) has local
flow : (, ) V M around p M. Given any f each fiber is a vector space of dimension k,
C (M), there exists a smooth function g : (, ) V R, each point in M has a trivializing neighborhood U,
which we write as (t, q) 7 gt (q), such that meaning there is a fiber-preserving diffeomorphism
f t (q) = f (q) + tgt (q),

Xq f = g0 (q). EU U Rk that is a vector space isomorphism on
each fiber.
Proof. First we define ht (q) := dtd f t (q) = Xt q f , and

R1 Exercise A10.2. If M m N n is a submanifold and : E
then we set gt (q) := s=0 h st (q) ds. For t = 0 this clearly N is a vector bundle of rank k over N, then the restriction
means g0 (q) = h0 (q) = Xq f . Using a change of variables E M is a vector bundle of rank k over M.
and the fundamental theorem of calculus, for arbitrary t we
get tgt (q) = f t (q) f (q) as desired. Remark A10.3. The tangent bundle T M is a rank-m bun-
dle over M m . Any coordinate chart (U, ) is a trivializing
Theorem A9.11. For any vector fields X, Y on M, the Lie neighborhood where the fiber-preserving diffeomorphism
derivative and Lie bracket coincide, that is, we have LX Y = is
[X, Y].
(, D) : X p 7 p, D p (X p ) .

Proof. Suppose f C (p) is a germ at p M. We want
to show LX Y p f = [X, Y] p f . Choose a representative The restriction D p : T p M Rm to each fiber is indeed
f C (U) for the germ and use the lemma (applied to the linear.
manifold U) to find gt such that g0 = X f and f t = f +tgt ,
or negating t as we will, f t = f tgt . Then starting Suppose : E M is a vector bundle. We can cover M
from the definition of LX Y we find by open sets U that are (small enough to be) both coordi-
nate charts for the manifold M and trivializing neighbor-
1
LX Y p f = lim t Yt p f Y p f hoods for the bundle E. That is, we have diffeomorphisms

t0 t
: U (U) Rm and : EU U Rk , the latter being
1
= lim Yt p f t Y p f fiber-preserving and linear on each fiber. Composing these

t0 t gives a diffeomorphism
1
= lim Yt p f tgt Y p f

t0 t (, id) : EU (U) Rk Rm+k ,

d
=

Y p f lim Yt p gt which is a coordinate chart for the manifold E.
dt t=0 t t0

d
= Y f t (p) Y p g0 Definition A10.4. A section of a vector bundle : E M

dt t=0 is a smooth map : M E such that = id M , meaning
= X p (Y f ) Y p (X f ) = [X, Y] p f (p) E p for each p M. The space of all (smooth)
m n
Exercise A9.12. Suppose f : M N is a smooth map sections is denoted (E).
and Y X(N) is f -related to X X(M) while Y 0 is f -
related to X 0 . Then [Y, Y 0 ] is f -related to [X, X 0 ]. Vector fields, for example, are simply sections of the tan-
gent bundle: X(M) = (T M). As in that example, (E)
End of Lecture 9 Nov 2015 is always a module over C (M), using pointwise addition
13

and scalar multiplication: if , (E) and f C M, write 1 (M) = (T M) for the space of all sections. A
then + f is defined pointwise by one-form 1 (M) acts on a vector field X X(M)
to give a smooth function (X) C (M) via (X)(p) =
( + f ) p = p + f (p) p E p . p (X p ) R. In a coordinate chart (U, ) we have the ba-
sis vector fields i ; taking the dual basis pointwise we get
Sometimes we talk about local sections (EU ) that the basis one-forms dxi satisfying dxi ( j ) = ij . Any one-
are not defined globally on all of M but only on a subset
form 1 (U) can be written as = i dxi , where the
P
U M. A trivialization of E over U is equivalent to a
frame, that is, a set of k sections i (EU ) such that at components i = (i ) C U are smooth functions.

each p U, the i (p) form a basis for E p . An important way to construct one-forms is as the differen-
Operations on vector spaces yield corresponding opera- tials of functions. If f C M then D p f : T p M T f (p) R,
tions (acting fiberwise) on vector bundles. For instance, under the identification T t R R, can be thought of as a
if E M and F M are two vector bundles over M (of covector d f p at p. If X X(M) is a vector field, then
ranks k and l, respectively) then their direct sum (or Whit- d f (X) = X f , meaning d f p (X p ) = X p f for each p M.
ney sum) E F M is a vector bundle of rank k +l, where While X p f depends on the germ of f at p, it only depends
we have (E F) p = E p F p fiberwise. Any neighborhood on d f at p: the covector d f p encodes exactly all the direc-
which trivializes both E and F will trivialize their sum. tional derivatives of f at p.
Note that the notation dxi we used above for the coordinate
basis one-forms in a coordinate chart (U, ) is consistent:
A11. Dual spaces and one-forms these are indeed the differentials of the coordinate compo-
P x := i : U R. In coordinates, we
i i
nent functions
have d f = (i f )dx , where we recall that i f is the ith
We next turn to various constructions on a single vector partial derivative of the coordinate expression f 1 of f .
space V. Even though much of what we say could extend
An interesting example of a one-form is the form d on S1 .
to arbitrary spaces, we assume V is a real vector space of
We cover S1 with coordinate charts of the form
finite dimension k; later it will be a tangent space to a man-
ifold. (cos , sin ) 7 (0 , 0 + 2).
The dual space V := L(V, R) is defined to be the space of On each such chart, we write d for the coordinate basis
all linear functionals V R (also called covectors). The one-form (which would also be called dx1 ); then we note
dual space V is also k-dimensional. that on the overlaps, these forms d agree (independent of
As we have mentioned before, a linear map L : V W the omitted point 0 ).
induces a dual linear map L : W V in the opposite Unlike for most manifolds, the tangent bundle T S1 is
direction, defined naturally by (L )(v) = (Lv). This (globally) trivial; thus T S1 is also trivial. Any one-form
construction is functorial in the sense that id = id and is written as f d for some smooth function f . The form d
0
L L = (L0 ) L . One can check that L is surjective if can be thought of as dual to the vector field ( sin , cos )
and only if L is injective, and vice versa. on S1 , which is 1 in any of the charts above.
While there is no natural isomorphism V V , any ba- The notation d is slightly confusing, since this one-form
sis {e1 , . . . , ek } for V determines a dual basis {1 , . . . , k } is not globally the differential of any smooth function
for V by setting on S1 . Thus on S1 , this is a one-form which is closed
but not exact, meaning that d looks like a differential
0,
i , j,
(e j ) = j =
i i locally but not globally. This shows, in a sense we may

1,
i = j. explore later, that the space S1 has nontrivial first coho-
mology, that is, that it has a one-dimensional loop.
The covector i is the functional that gives the iP th
compo-
One forms are in some sense similar to vector fields, but
nent of a vector in the basis {ei }, that is, v = i (v)ei ,
we will see later how they (as well as differential forms of
P we could also write as v = (v). Similarly,
i i
which
higher degree) are often more convenient. This is mainly
= (ei )i .
because, while a map f : M N does not in general act
There is a natural isomorphism V V , where v V on vector fields, it can be used to pull one-forms on N back
induces the linear functional 7 (v) on V . Given a to M. To see this, note that
linear map L : V W, we have L = L.
f = D p f : T p M T f (p) N
Applying duality to each fiber E p of a vector bundle gives
the dual bundle E with (E ) p = (E p ) . It is trivialized at each p M induces a dual map
over any trivializing neighborhood for E, as one sees by
f := (D p f ) : T f (p) N T p M.
choosing a frame and taking the dual frame.
Applying duality to T p M gives the cotangent space T p M Given 1 (N), we define f 1 (M) by ( f ) p =
to M at p. These fit together to form the cotangent bundle, f ( f (p) ) T p M.
which (just like the tangent bundle) is trivialized over any A special case is the restriction | M of a form 1 N
coordinate neighborhood. to a submanifold M m N n , which is simply the pullback
A (smooth) section of the cotangent bundle is called a under the inclusion map. For any X p T p M T p N at any
covector field or more often a (differential) one-form. We p M N we of course simply have | M (X p ) = (X p ).
14

A12. Bilinear forms and Riemannian metrics The matrix (gi j ) is of course symmetric and positive defi-
nite at each p U.
A bilinear map b : V V R is called a bilinear form The standard Riemannian metric g on the manifold Rm
on V. If {e1 , . . . , ek } is a basis comes from putting the standard Euclidean inner product
P i for V, thenPb isi given on each T p Rm = Rm . That is, in the standard chart (Rm , id)
by bi j = P b(ei , e j ): if v = v ei and w = w ei then
b(v, w) = i, j bi j vi w j . The bilinear form b is called sym- we have gi j = i j .
metric if b(w, v) = b(v, w) (i.e., if bi j = b ji ) and an- If f : M m N n is a smooth map, then we can pull back
tisymmetric or alternating if b(w, v) = b(v, w) (i.e., if sections of Q(T N) to sections of Q(T M) in the natural
bi j = b ji ). Any bilinear form b can be uniquely decom- way:
posed b = b+ + b into a symmetric part b+ and an anti-
( f g)(X p , Y p ) := g( f X p , f Y p ).
symmetric part b , defined by 2b (v, w) = b(v, w)b(w, v).
Given a linear map L : V W, we can pull back any bi- If g is a Riemannian metric, then of course f g will be
linear form b on W to a bilinear form L b on V, defined by positive semidefinite at each p M, but it will be a Rie-
(L b)(v, v0 ) = b(Lv, Lv0 ). mannian metric if and only if f is an immersion (meaning
We will consider antisymmetric differential forms later. that D p f is injective for every p M, and in particular
For now we restrict attention to symmetric bilinear forms m n). Again, an important special case of this pull-back
on V. These are in one-to-one correspondance with metric is when f is the inclusion map of a submanifold;
quadratic forms q : V R: of course q(v) := b(v, v) de- then we speak of restricting the Riemannian metric g on N
pends only on the symmetric part of b and we can recover b to g| M on the submanifold M N.
from q via the formula In particular, the standard metric on Rn restricts to give a
Riemannian metric on any submanifold M m Rn . Last
2b(v, w) := q(v + w) q(v) q(w). semester, we studied the case m = 2, n = 3, and called this
metric g(v, w) = hv, wi the first fundamental form.
(Note that much of our discussion would fail for vector
spaces over fields of characteristic 2, where 2 = 0.) End of Lecture 16 Nov 2015
A symmetric bilinear form b (or the associated quadratic

form q) is called positive semidefinite if q(v) = b(v, v) 0 A13. Partitions of unity
for all v V. It is called positive definite if q(v) = b(v, v) >
0 for all v , 0. A positive definite form on V is also
called an inner product (or scalar product) on V. An inner The sections of any vector bundle E M form a vec-
product is what we need to define the geometric notions of tor space. In particular, there is always the zero section
length (or norm) kvk := b(v, v) and angle p = 0 E p . An interesting question is whether there is
a nowhere vanishing section. A trivial bundle M Rr of
b(v, w) course has constant nonzero sections. Sometimes it turns
(v, w) := arccos
kvk kwk out that the tangent bundle T M is trivial this happens for
instance for S1 or more generally the m-torus T m . Other
between vectors in V. The pullback L b of a positive defi- times, there is no nonvanishing section of T M. For in-
nite form is always positive semidefinite, but it is positive stance the Hopf index theorem shows this is the case for
definite if and only if L is injective. any closed orientable surface M 2 other than the torus T 2 .
Of course the standard example ofPan inner product is the Over S1 there is also a nontrivial line bundle, whose total
Euclidean inner product b(v, w) = vi wi on Rm . space is topologically a Mbius strip. This bundle has no
The quadratic nonvanishing section.
forms on V form a vector space Q(V) of
dimension m+1 2 . The positive definite forms form an open
From this point of view, it might be surprising that the
convex cone in this vector space, whose closure consists bundle Q(T M) has nonvanishing sections for any mani-
of all positive semidefinite forms. (A convex cone is a set fold M m , indeed sections which are positive definite ev-
closed under taking positive linear combinations.) erywhere. In other words, any manifold M can be given
Again, we can apply this construction to the fibers of any a Riemannian metric. This follows from the fact that any
manifold can be embedded in Rn for sufficiently large n,
vector
k+1
bundle E. If E has rank k, then Q(E) has rank or can be proven more directly by taking convex combi-
2 . In case of the tangent bundle T M, we get a vec-
m+1 nations of standard metrics in different coordinate charts.
tor bundle Q(T M) of rank 2 . A positive definite sec- Either of these approaches requires the technical tool of a
tion g Q(T M) is called a DRiemannian

E metric on M. partition of unity, a collection of locally supported func-
It consists of an inner product X p , Y p := g p (X p , Y p ) on tions whose sum is everywhere one. This gives a general
each tangent space T p M, which lets us measure length and method for smoothly interpolating between different local
angles between tangent vectors at any p M. definitions. We do not want to get into questions of sum-
In a coordinate chart (U, ), the metric g is given by com- ming infinite sequences; thus we impose a local finiteness
ponents gi j := g(i , j ) C U so that condition.
X X X Definition A13.1. A collection {S } of subsets of a topo-
g i i , j j = gi j i j . logical space X is called locally finite if each p X has a
i, j neighborhood that intersects only finitely many of the S .
15

Example A13.8. The cover (n, n) : n N+ of R is not

Definition A13.2. The support supp f of a function
f : X Rn is the closure of the set where f , 0. locally finite, but any covering of
R by bounded open sets
is a refinement, so for instance (k 1, k + 1) : k Z is a
Lemma A13.3. If f : M R are P smooth functions such
locally finite refinement.
that {supp f } is locally finite, then f defines a smooth
function M R. A standard result in point set topology says that any sec-
ond countable, locally compact Hausdorff space is para-
Proof. The local finiteness means that each p M has a compact. It is also true that any metric space is paracom-
neighborhood U which P meets only a finite number of the pact. Some authors replace second countable by para-
supp f . On U, the sum f is thus a sum of a fixed finite compact in the definition of manifold, which makes no
collection of smooth functions, hence smooth. difference except for allowing uncountably many compo-
nents. (The long line, for instance, is not paracompact.)
Note that saying the collection of supports supp f is lo- It is known that a topological space X admits a continu-
cally finite is stronger than saying each p M is contained ous partition of unity subordinate to any given open cover
in finitely many supp f , which in turn is stronger than say- if and only if X is paracompact and Hausdorff. We will,
ing only finitely many f P are positive at each p M. This however, explicitly prove what we need for manifolds.
would suffice to evaluate f (p) as a finite sum at each p.
The stronger conditions ensure that the sum is a smooth Lemma A13.9. Every manifold M has a countable base
function. consisting of coordinate neighborhoods with compact clo-
sure.
Definition A13.4. A (smooth) partition of unity on a man-
ifold M is a collection of functions : M R such that
Proof. Start with any countable base {Bi } and let B be the
0, subcollection of those Bi that are contained in some co-
ordinate neighborhood and have compact closure. Now
{supp } is locally finite, suppose we are given an open subset W M and a point
p W. Choose a coordinate chart (U, ) around p such
1.
P
that
A trivial example is the single constant function 1. The
interest in partitions of unity comes from examples where U W, (p) = 0, B2 (0) (U)
the support of each is small in some prescribed sense.
and set V := 1 B1 (0) . Then V has compact closure.

Definition A13.5. Given an open cover {U }, a partition Since {Bi } is a base, we have p Bi V U for some i.
of unity { } is subordinate to the cover {U } if for each But then this Bi is also contained with compact closure
there exists = () such that the support supp is in the coordinate neighborhood U; thus Bi B. Since p
contained in U . and U were arbitrary, this shows B is a base.
Note that if we want, we can then define a new partition Lemma A13.10. Every manifold M has a compact ex-
of unity { } also subordinate to {U } and now indexed by haustion, indeed a nested family of subsets
the same index set. Simply set to be the sum of those
for which = (). (Note that this is not the sum of all , W1 W 1 W2 W 2
supported in U .)
with Wk open and W k compact, whose union is M.
To give the flavor of results about partitions of unity,
we start with the easy case of a compact manifold. All Proof. Choose a countable base {Bi } as in the last S lemma.
manifolds have a related property called paracompactness, We will choose 1 = i1 < i2 < and set Wk := ii=1 k
Bi .
which will be enough to extend this result. These automatically have compact closure and are nested.
Proposition A13.6. Given any open cover {U } of a com- We just need to choose each ik large enough that Wk
pact manifold M m , there exists a partition of unity subor- W k1 . But this is possible, since W k1 is compact and thus
dinate to this cover. covered by some finite collection S of the S
Bi s. Finally, since
ik k we have that Wk Bk so Wk Bk = M.
Proof. For each p M, we have p U for some
= (p), and we can choose a smooth nonnegative func- Corollary A13.11. Given a manifold M, we can find
tion f p supported in U with f p > 0 on some neighbor- countable families of subsetsSKi Oi M, for i N+ ,
hood V p 3 p. Since M is compact, a finite subcollection with Ki compact, Oi open, Ki = M, and {Oi } locally

V p1 , . . . , V pk covers M. Then f := i f pi is a positive
P finite.
smooth function on M, so we can define a finite collec-
tion of smooth functions i := f pi / f . These form a (finite) Proof. Using the nested Wi from the lemma, simply set
partition of unity, subordinate to the given cover. Ki := W i r Wi1 and Oi := Wi+1 r W i2 (where we take
W0 = = W1 ). The local finiteness follows from the fact
Definition A13.7. An open cover {V } is a refinement of that any p is contained in some W j+1 r W j1 , which meets
another open cover {U } if each V is contained in some only four of the Oi .
U . A space X is paracompact if every open cover has a
locally finite refinement. Corollary A13.12. Any manifold M is paracompact.
16

Proof. Suppose {U } is an open cover. Choose Oi and Ki We will restrict to compact manifolds and not attempt to
as in the last corollary. For each i, the compact set Ki is get an optimal n. Rather than using a partition of unity
covered by the sets Oi U , and thus by a finite subcol- directly, we will repeat the easy proof of the compact case,
lection, which we name Oij for j = 1, . . . , ki . The union of using some of the functions involved in the construction
these finite collections, over all i, is a locally finite refine- directly.
ment. The theorems are also true in the noncompact case; the
proof uses decompositions like the Ki and Oi above but re-
We are now set up to adapt the construction of partitions quires knowing that each compact piece can be embedded
of unity from the compact case to the general case. in the same dimension, say in R2m+1 .
Theorem A13.13. Given any covering {U } of a manifold Theorem A13.15. Any compact manifold M m can be em-
M, there exists a partition of unity {i } subordinate to this bedded in some Euclidean space Rn .
covering.
Proof. For each point p M, find a nonnegative function
Proof. Find Ki Oi as above. Fixing i, for each p Ki we f p : M R with f 1 in some neighborhood V p 3 p and
have p U Oi for some = (p). Choose a smooth with support in a coordinate chart (U p , p ). By compact-
nonnegative f p with support in U Oi such that f p > 0 ness, a finite number of the V p suffice to cover M. Call
on some neighborhood V p 3 p. Finitely many of these these points p1 , . . . , pk and simply use the indices 1, . . . k
neighborhoods cover the compact set Ki . Now letting i for the associated objects. Define a map g : M Rk(m+1)
vary, we have a countable family of bump functions fi j , as follows:
whose supports form a locally finite family. Thus dividing
g(p) := f1 (p)1 (p), . . . , fk (p)k (p), f1 (p), . . . , fk (p) .

by their well-defined, positive, smooth sum gives a parti-
tion of unity. On Vi we have fi 1, so the ith block in g equals i , with
injective differential. Thus g is an immersion on each Vi ,
Note that if K M is compact, then for any partition of thus on all of M. By compactness, it only remains to show
unity {i } for M, only finitely many i have support meet- that g is injective. If g(p) = g(q) then in particular we have
ing K. (Each p K has a neighborhood meeting only fi (p) = fi (q) for all i. Choose i such that p Vi Ui .
finitely many supp i ; by compactness finitely many such Then fi (q) = 1 implies q Ui . But then we also have
neighborhoods cover K.)
Now we turn to some applications of these ideas. Note that i (p) = fi (p)i (p) = fi (q)i (q) = i (q).
if { } is a partition of unity subordinate to a cover {U }
and we have functions f C (U ), then f P defines Since i is injective on Ui , it follows that p = q.
a smooth function on M supported in U . Then f
makes sense as a locally finite sum of smooth functions.
A14. Riemannian manifolds as metric spaces
The same works for sections of any vector bundle E M:
local sections E U
can be combined to get a global
section (E).
P
We fix a Riemmanian manifold (M, g), that is, a smooth
m manifold M m with a
fixed Riemannian metric g. Where
Theorem A13.14. Any manifold M admits a Riemannian
convenient, we write X p , Y p := g p (X p , Y p ) for the inner

metric. q
X p , X p for the length of a tangent

product and kX p k :=
Proof. Let g0denote the
standard (flat) Riemannian metric vector.
on Rm . Let (U , ) be an atlas for M. In each chart
the pullback g := (g0 ) is a Riemannian metric on U . Definition A14.1. Suppose : [a, b] M is a piecewise
Now let { } be a partition of unity subordinate to {U }. smooth curve. The length of (with respect to the Riem-
We can consider each g as a global section of Q(T M), manian metric g) is
supported of course in U . Then g is a Riemannian
P
Z b
metric on M, since locally near any point it is a convex len() := k0 (t)k dt.
combination of finitely many Riemannian metrics g . a
Note that by the chain rule, this length is invariant under

An alternative proof of the existence of Riemannian met-
reparametrization. The arclength function along is
rics simply uses the fact that any manifold M m can be em-
bedded in Rn for large enough n. Z t
s(t) := len |[a,t] = k0 (t)k dt.

It is not too hard to show that n = 2m + 1 actually suffices a
a generic projection from higher dimensions to dimension
2m + 1 will still give an embedding but we omit such and assuming is a piecewise immersion we can
discussions. Harder is the Whitney trick used to get down reparametrize by arclength so that k0 k 1.
to n = 2m. Most manifolds actually embed in R2m1 Note that if the standard Riemannian metric on Rn is re-
the only exceptions (besides S 1 ) are closed nonorientable stricted to a submanifold M m , then the length of a curve
manifolds of dimension m = 2k , like closed nonorientable in M as defined above is the same as its length in Rn as
surfaces. considered last semester.
17

Definition A14.2. The distance between two points p, q by at most a constant factor. For finite dimensional vector
M is the infimal length spaces, all norms are equivalent. We sketch a proof of the
case we need.
d(p, q) := inf len()

Lemma A14.4. Any two inner products on Rm induce
taken over all piecewise smooth curves in M from p to q. equivalent norms.
Note that we could easily apply our definition of length P Let kvk denote the standard Euclidean norm, and
Proof.
to more general curves, say to all rectifiable or Lipschitz let gi j vi w j denote an arbitrary inner product on Rm . On
curves. Since any curve can be smoothed, in the infimum the compact
pP unit sphere S
m1
= {v : kvk = 1} the other
defining d it is not important whether we allow all rectifi- norm gi j v v achieves its minimum c > 0 and its max-
i j
able curves or restrict to smooth curves. We have chosen imum C. Then by homogeneity, we have
an option in the middle. qX
End of Lecture 23 Nov 2015
ckvk gi j vi v j Ckvk
Note that (no matter which smoothness class is chosen) the for all v, as desired. We note that the optimal constants
infimum is not always realized, as one sees for instance if depend smoothly on g.
M = R2 r {0}.
Corollary A14.5. Suppose g is a Riemannian metric on
The theorem below will show that (M, d) is a metric space an open set U Rm and K U is compact. Then there
compatible with the given topology on M. Of course when exist constants 0 < c C such that
M is not connected, points p and q in different components
are not connected by any path, so d(p, q) = + by the
p
ckvk g(v, v) Ckvk
above definition. It is easiest to use a definition of metric
spaces that allows infinite distance. If this is not desired, for all p K and all v T p U = T p Rm Rm . In particular,
the following discussion should be restricted to connected for any curve in K from p to q we have
manifolds. Note that a connected component of a man-
ifold is automatically path-connected; any pair of points ckp qk c len0 leng C len0 ,
can actually be joined by a smooth path, whose length is
necessarily finite. where len0 is the length relative to the standard Euclidean
A few properties of d are immediate. The constant path metric and leng is the length relative to g.
shows d(p, p) = 0. The inverse path shows d(p, q) =
Proof. For each p K the lemma gives us c p C p . As-
d(q, p). Concatenating paths gives the triangle inequality
suming we choose the optimal constants at each point, they
d(p, r) d(p, q) + d(q, r). (This is one reason we chose to
depend smoothly on g p thus smoothly on p. By compact-
allow piecewise smooth paths.) That is, we see easily that
ness we can set c := minK c p and C := maxK C p . Integrat-
d is a pseudometric, and to see it is a metric we just need
ing the bounds for tangent vectors gives the final statement
to show that d(p, q) = 0 holds only for p = q.
for any curve .
Lemma A14.3. Consider Rm with the standard Rieman-
nian metric. Then d(p, q) = kp qk. This corollary gives the key uniformity needed for the fol-
lowing theorem.
Proof. Since both sides are clearly translation invariant, it
suffices to consider q = 0. It is easy to compute the length Theorem A14.6. Let (M, g) be a Riemannian manifold.
of the straight path from p to 0 as kpk. We must show no With the distance function d above, it is a metric space
other path has less length (and may assume p , 0). So (M, d). The metric topology agrees with the given manifold
suppose (0) = p and (1) = 0. We may assume (t) , 0 topology on M.
for t < 1, since otherwise we replace by |[0,t] , which is
Proof. We have noted that d is symmetric and satisfies the
not longer. Thus for t < 1 we can write (t) = r(t)(t)
triangle inequality. We must prove d(p, q) = 0 = p = q
where k(t)k = 1 and r(t) > 0. We have r(0) = kpk and
and show that the topologies agree.
r(t) 0 as t 1. Since h, i 1, we get h, 0 i 0.
The product rule 0 = r0 + r0 then gives Let D denote the closed unit ball in Rm . Given p , q
in M, we can find coordinates (U, ) around p such that
k0 k2 = |r0 |2 kk2 + r2 k0 k2 |r0 |2 . (p) = 0, (U) D and q < 1 (D). On D use the last
corollary to get c, C comparing (1 ) g with the standard
Thus metric. Any path from p to q must first leave 1 D. Its g-
Z 1 Z 1 Z 1 length is at least the g-length of this initial piece, which is

k0 k dt |r0 | dt r0 dt = r(0) r(1) = kpk the 1 g-length of its image . Since connects 0 to D,

0 0 0 it has Euclidean length at least 1, so 1 g-length at least c.
as desired. Since this is true for any , we find d(p, q) c > 0.
We have just seen that a Euclidean ball in a coordinate
Two norms on a vector space induce the same topology if chart contains a small metric ball. Thus open sets in the
and only if they are equivalent in the sense that they differ manifold topology are open in the metric topology. To
18

get the converse, consider again coordinates around p with

D (U), and find C as in the corollary. If (q) is in the
-ball around 0 = (p), then they can be joined by a path
of Euclidean length less than < 1. Thus p and q can be
joined by a path of g-length less than C.
19

J.M. Sullivan, TU Berlin B: Differential Forms Diff Geom II, WS 2015/16
B. DIFFERENTIAL FORMS instance, if S has k elements this gives a k-dimensional

vector space with S as basis.
We have already seen one-forms (covector fields) on a Given vector spaces V and W, let F be the free vector
manifold. In general, a k-form is a field of alternating k- space over the set V W. (This consists of formal sums
ai (vi , wi ) but ignores all the structure we have on the set
P
linear forms on the tangent spaces of a manifold. Forms
are the natural objects for integration: a k-form can be in- V W.) Now let R F be the linear subspace spanned by
tegrated over an oriented k-submanifold. We start with ten- all elements of the form:
sor products and the exterior algebra of multivectors. (v + v0 , w) (v, w) (v0 , w),
(v, w + w0 ) (v, w) (v, w0 ),
(av, w) a(v, w), (v, aw) a(v, w).
B1. Tensor products
These correspond of course to the bilinearity conditions
Recall that, if V, W and X are vector spaces, then a map we started with. The quotient vector space F/R will be the
b : V W X is called bilinear if tensor product V W. We have started with all possible
v w as generators and thrown in just enough relations to
b(v + v0 , w) = b(v, w) + b(v0 , w), make the map (v, w) 7 v w be bilinear.
b(v, w + w0 ) = b(v, w) + b(v, w0 ), The tensor product is commutative: there is a natural linear
isomorphism V W W V such that vw 7 wv. (This
b(av, w) = ab(v, w) = b(v, aw).
is easiest to verify using the universal property simply
factor the bilinear map (v, w) 7 w v through V W to
The function b is defined on the set V W. This Cartesian
give the desired isomorphism.)
product of two vector spaces can be given the structure of
a vector space V W, the direct sum. But a bilinear map Similarly, the tensor product is associative: there is a natu-
b : V W X is completely different from a linear map ral linear isomorphism V (W X) (V W) X. Note
V W X. that any trilinear map from V W X factors through this
triple tensor product V W X.
The tensor product space V W is a vector space designed
exactly so that a bilinear map b : V W X becomes a Of special interest are the tensor powers of a single vector
k
linear map V W X. More precisely, it can be charac- space V. We write V := V V. If {ei } is a basis
terized abstractly by the following universal property. for V, then ei1 eik is a basis for V k . In particular
if V has dimension m, then V k has dimension mk . There
Definition B1.1. The tensor product of vector spaces V is a natural k-linear map V k V k and any k-linear map
and W is a vector space V W with a natural bilinear map V k W factors uniquely through V k .
V W V W, written (v, w) 7 v w, with the prop- One can check that the dual of a tensor product is the tensor
erty that any bilinear map b : V W X factors uniquely product of duals: (V W) = V W . In particular,
through V W. That means there exists a unique linear we have (V )k = (V k ) . The latter is of course the set
map L : V W X such that b(v, w) = L(v w). of linear functionals V k R, which as we have seen is
exactly the set of k-linear maps V k R.
This does not yet show that the tensor product exists, but
uniqueness is clear: if X and Y were both tensor products, Definition B1.2. A L graded algebra is a vector space A de-

then each defining bilinear map would factor through the composed as A = k=0 Ak together with an associative
other we get inverse linear maps between X and Y, show- bilinear multiplication operation A A A that respects
ing they are isomorphic. the grading in the sense that the product of elements
Ak and A` is an element of Ak+` . Often we consider
Note that the elements of the form v w must span V W,
graded algebras that are either commutative or anticommu-
since otherwise L would not be unique. If {ei } is a basis
tative. Here anticommutative has a special meaning: for
for V and { f j } a basis for W then bilinearity gives
Ak and A` as above, we have = (1)k` .
X X X
vi ei wj fj = vi w j ei f j . Example B1.3. The tensor algebra of a vector space V is
i j i, j
M
V := V k .
Clearly then {ei f j } spans V W indeed one can check k=0
that it is a basis. This is a valid construction for the space
V W as the span of the ei f j but it does depend on Here of course V 1 V and V 0 R. Note that the tensor
the chosen bases. If dim V = m and dim W = n then we product is graded, but is neither commutative nor anticom-
note dim V W = mn. mutative.
A much more abstract construction of V W goes through
a huge infinite dimensional space. Given any set S , the free
B2. Exterior algebra
vector spaceP on S is the set of all formal finite linear com-
binations ai si with ai R and si S . (This can equally
well be thought of as the set of all real-valued functions We now want to focus on antisymmetric tensors, to de-
on the set S which vanish outside some finite subset.) For velop the so-called exterior algebra or Grassmann algebra
20

of the vector space V. is, alternating k-linear maps from V k correspond to linear
Just as we constructed V V = V 2 as a quotient of a huge maps from k V. (One can also phrase the universality for
vector space, adding relators corresponding to the rules for all k together in terms of homomorphisms of anticommu-
bilinearity, we construct the exterior power V V = 2 V tative graded algebras.)
as a further quotient. In particular, letting S V V denote So far we have developed everything abstractly and alge-
span of the elements v v for all v V, we set V V := braically. But there is a natural geometric picture of how k-
(V V)/S . We write v w for the image of v w under the vectors in k V correspond to k-planes (k-dimensional lin-
quotient map. Thus v v = 0 for any v. From ear subspaces) in V. More precisely, we should talk about
simple k-vectors here: those that can be written in the form
(v + w) (v + w) = 0 v1 vk . We will see that, for instance, e12 +e34 2 R4
is not simple.
it then follows that v w = w v. If {ei : 1 i m} is a
basis for V, then A nonzero vector v V lies in a unique oriented 1-plane
(line) in V; two vectors represent the same oriented line if
{ei e j : 1 i < j m} and only if they are positive multiples of each other. Now
suppose we have vectors v1 , . . . , vk V. They are linearly
is a basis for V V. independent if and only if 0 , v1 vk k V. Two
Higher exterior powers of V can be constructed in the same linearly independent k-tuples (v1 , . . . , vk ) and (w1 , . . . , wk )
way, but formally, it is L
easiest to construct the whole ex- represent the same oriented k-plane if and only if the
terior algebra V = k V at once, as a quotient of wedge products v1 vk and w1 wk are pos-
the tensor algebra V, this time by the two-sided ideal itive multiples of each other, that is, if they lie in the same
generated by the same set S = {v v} V V V. ray in k V. (Indeed, the multiple here is the ratio of k-
This means the span not just of the elements of S but also areas of the parallelepipeds spanned by the two k-tuples,
of their products (on the left and right) by arbitrary other given as the determinant of the change-of-basis matrix for
tensors. Elements of V are called multivectors and ele- the k-plane.)
ments of k V are more specifically k-vectors. We let Gk (V) denote the set of oriented k-planes in V,
called the (oriented) Grassmannian. Then the set of simple
End of Lecture 30 Nov 2015 k-vectors in k V can be viewed as the cone over Gk (V). (If
Again we use to denote the product on the resulting (still we pick a norm on k V, say induced by an inner product
graded) quotient algebra. This product is called the wedge on V, then we can think of Gk (V) as the set of unit sim-
product or more formally the exterior product. We again ple k-vectors, say those arising from an orthonormal basis
get v w = w v for v, w V. More generally, for any for some k-plane.)
v1 , . . . , vk V and any permutation k of {1, . . . , k}, (Often, especially in algebraic geometry, one prefers to
this implies work with the unoriented Grassmannian Gk (V)/. It is
most naturally viewed as lying in the projective space
v1 vk = (sgn ) v1 vk .
P(V) := V r {0} / R r {0} .

A special case is the product of a k-vector with an `-
vector where we use a cyclic permutation to get the anti- In algebraic geometry one typically also replaces R by C
commutative law = (1)k` . throughout.)
If {ei : 1 i m} is a basis for V, then If we give V an inner product, then any k-plane has a
unique orthogonal (m k)-plane. This induces an isomor-
{ei1 ik := ei1 eik : 1 i1 < < ik m} phism between Gk V and Gmk V. It extends to a linear,
norm-preserving isomorphism
is a basis for k V. In particular, dim k V = mk ; we have
0 V = R but also m V R, spanned by e12m . For k > m ? : k V mk V
there are no antisymmetric tensors: k V = 0. The exterior
P m
algebra has dim V = m k=0 k = 2m . The determinant called the Hodge star operator. (Recall
that both these
has a natural definition in terms of the exterior algebra: if m
spaces have the same dimension k .) If v is a simple k-
we havePm vectors v j V given in terms of the basis {ei } vector, then ?v is a simple (m k)-vector representing the
as v j = i vij ei then orthogonal complement. In particular, if {ei } is an oriented
orthonormal basis for V, then
v1 vm = det vij e12m .

? e1 ek = ek+1 em

(The components of the wedge product of k vectors vi are
by the various k k minor determinants of the matrix
given and similarly each other vector in our standard basis for
vij .) k V maps to a basis vector for mk V, possibly with a
The exterior powers of V with the natural k-linear maps minus sign.
V k k V are also characterized by the following univer- Classical vector calculus in three dimensions uses the
sal property. Given any alternating k-linear map V k X to Hodge star implicitly: instead of talking about bivectors
any vector space X, it factors uniquely through k V. That and trivectors, we introduce the cross product and triple
21

product: 1 V = V we set
v w := ?(v w), [u, v, w] := hu, v wi = ?(u v w). := .
But even physicists noticed that such vectors and scalars More generally, for k V and ` V we use an
transform differently (say under reflection) than ordinary alternating sum over all permutations k+` :
vectors and scalars, and thus refer to them as pseudovec-
tors and pseudoscalars. ( )(v1 , . . . , vk+` ) :=
For dim V = m, we can use these terms as follows: 1 X
(sgn ) (v1 , . . . , vk ) (v(k+1) , . . . , v(k+`) ).
k!`!
scalars are elements of R = 0 V,
vectors are elements of V = 1 V, The factor is chosen so that if {ei } is a basis for V and {i }
pseudovectors are elements of ?V = m1 V, and is the dual basis for 1 V = V then
pseudoscalars are elements of ?R = m V. := i1 ik
i1 ik

Of course, these are in a sense the easy cases. For these k, is the basis of k V dual to the basis {ei1 ik } for k V.
any k-vector is simple. We can identify both G1 V and Putting these spaces together, we get an anticommutative
Gm1 V as the unit sphere in V = 1 V m1 V. For graded algebra
2 k m 2 on the other hand, not all k-vectors are
m
simple, and Gk V has lower dimension than the unit sphere M
in k V. Indeed, it can be shown that the set of simple k- V := k V.
k=0
vectors (the cone over Gk V) is given as the solutions to a
certain set of quadratic equations
P called the Grassmann

Again the dimension of each summand is mk so the whole
Plcker relations. For instance ai j ei j 2 R4 is a simple algebra has dimension 2m .
2-vector if and only if
If L : V W is a linear map, then for each k we get an
a12 a34 a13 a24 + a14 a23 = 0. induced map L : k W k V defined naturally by
This shows that G2 R4 is a smooth 4-submanifold in the L (v1 , . . . , vk ) = (Lv1 , . . . , Lvk ).

unit sphere S5 2 R4 .
Of course, we have introduced these ideas in order to apply
If we choose an inner product on V, then thinking about them to the tangent spaces T p M to a manifold m
how oriented orthonormal bases for a k-plane and its or- m M . We get
thogonal complement fit together, we seethat we can iden- dual bundles k T M and k T M of rank k .
tify Gk V = S O(m)/ S O(k) S O(m k) . In particular, it Definition B3.1. A (differential) k-form on a manifold M m
is a smooth manifold of dimension k(m k). is a (smooth) section of the bundle k T M. We write
k M = (k T M) for the space of all k-forms, which
is a module over C ML= 0 M. Similarly we write
B3. Differential forms M = ( T M) = k M for the exterior algebra
of M.
Many textbooks omit discussion of multivectors and con-
If k M is a k-form, then at each point p M the
sider only the dual spaces. (This is presumably because the
value p k T p M is an alternating k-linear form on T p M
abstract definition of tensor powers and then exterior pow-
or equivalently a linear functional on k T p M. That is, for
ers as quotient spaces seems difficult.) Recall that vector
any k vectors X1 , . . . , Xk T p M we can evaluate
subspaces and quotient spaces are dual operations, in the
sense that if Y X is a subspace, then the dual (X/Y) p (X1 , . . . , Xk ) = (X1 Xk ) R.
of the quotient can be naturally identified with a subspace
of X , namely with the annihilator Y o of X , consisting of In particular, p naturally takes values on (weighted) k-
those linear functionals on X that vanish on Y: planes in T p M; as we have mentioned, k-forms are the nat-
ural objects to integrate over k-dimensional submanifolds
(X/Y) Y o X . in M.
If f : M m N n is a smooth map and k N is a k-form,
Using this, we find that
then we can pull back to get a k-form f on M defined
k V := (k V) V k
by
( f ) p (X1 , . . . , Xk ) = f (p) ((D p f )X1 , . . . , (D p f )Xk ).
is the subspace of those k-linear maps V k R that are
alternating. (Of course this vanishes if k > m.) As a special case, if
While it is easy to construct the wedge product on multi- f : M N is the embedding of a submanifold, then f =
vectors as the image of the tensor product under the quo- | M is the restriction of to the submanifold M, in the
tient map, the dual wedge product on V requires con- sense that we consider only the values of p (X1 , . . . , Xk )
structing a map to the alternating subspace. For , for p M N and Xi T p M T p N.
22

Exercise B3.2. Pullback commutes with wedge product in Theorem B4.3. For any manifold M m , the differential map
the sense that d : 0 M 1 M has a unique R-linear extension to an
antiderivation d : M M satisfying d2 = d d = 0.
f ( ) = ( f ) ( f ) This antiderivation has degree 1 in the sense that it sends
k M to k+1 M; it is called the exterior derivative.
for f : M N and , N.
Proof. First suppose g, f i C M so that g d f 1 d f k
In a coordinate chart (U, ) we have discussed the coordi-
k M. The two conditions on d together automatically im-
nate bases {i } and {dxi } for T p M and T p M, respectively,
ply that
the pullbacks under of the standard bases on Rm . Simi-
larly, d g d f 1 d f k = dg d f 1 d f k k+1 M.

n o
dxi1 dxik : 1 i1 < < ik m In a coordinate chart (U, ) of course every k-form can
be expressed as a sum of terms of this form. The propo-
forms the standard coordinate basis for k-forms; any sition above shows we can work locally in such a chart.
k (M) (or more properly its restriction to U) can be ex- Thus we know the exterior derivative (if it exists) must be
pressed uniquely as given in coordinates by
X
|U = i1 ik dxi1 dxik
X X XX
d I dxI = dI dxI = i I dxi dxI
i1 <<ik I I I i
XX
for some smooth functions i1 ik CP
U. To simplify = i I dx dx dxik .
i i1
notation, we often write this as |U = I I dxI in terms I i
of the multi-index I = (i1 , . . . , ik ). (Note that terms here where i = i j will vanish; for the other
terms, reordering the factors in this last wedge product
to put i in increasing order with the i j s and thus obtain a
B4. Exterior derivative standard basis element will introduce signs.)
Now a straightforward calculation shows that the operator
Zero-forms are of course just scalar fields, that is, smooth d defined by this formula really is an antiderivation locally:
functions. We have also already considered one-forms,
d (a dxI ) (b dx J )

which are simply covector fields. In particular, given
f 0 M, its differential d f 1 M is a one-form with = d(ab) dxI dx J = (da)b + a(db) dxI dx J

d f (X) = X f for any vector field X. We now want to gener- = (da dxI ) (b dx J ) + (1)k (a dxI ) (db dx J ),
alize this to define for any k-form its exterior derivative,
a (k + 1)-form d. where I = (i1 , . . . , ik ) is a k-index. Clearly this antideriva-
tion has degree 1 as claimed.
Definition B4.1. An antiderivation on the graded alge-
bra ( M, ) is a linear map D : M M satisfy- Now since d is determined uniquely, if we have overlap-
ing the following version of the Leibniz product rule for ping charts (U, ) and (V, ), then on U V we must get
k M and ` M: the same result evaluating d in either chart. Finally, since
the exterior algebra operations + and are defined point-
D( ) = (D) + (1)k (D). wise, to check that d is an antiderivation and d2 = 0 it
suffices that we know these hold locally.
To remember the sign here, it can help to think of D as Proposition B4.4. The pullback of forms under a map
behaving like a one-form when it moves past . f : M m N n commutes with the exterior derivative. That
is, for N we have d( f ) = f (d).
Proposition B4.2. Any antiderivation on M is a local
operator in the sense that if = on an open set U then
Proof. It suffices to work locally around a point p M.
D = D on U.
Let (V, ) be coordinates around f (p). By linearity we can
assume = a dyi1 dyik in these coordinates. For
Proof. By linearity it suffices to consider the case 0, k = 0 we have = a C N. For any X p T p M we have
k M. Given any p U, we can find a function f
C M supported in U with f (p) = 1. Then f 0 on M ( f da)(X p ) = (da)( f X p ) = ( f X p )a
and it follows that = X p ( f a) = (d f a)(X p ).
0 = D( f ) = (D f ) + f (D).
Note that if ( f 1 , . . . , f n ) = f is the coordinate expres-
At p this gives 0 = D f 0 + 1(D) p = (D) p as desired. sion of f (on some neighborhood of p) then the formula
above gives f (dyi ) = d f i . Since pullback commutes with
wedge products, for k > 0 we then have
End of Lecture 7 Dec 2015 f = ( f a) d f i1 d f ik
23

and so An equivalent definition of orientation (analogous to that

of smooth structures) is through a coherently oriented atlas
d( f ) = d( f a) d f i1 d f ik for M. Here two charts (U, ) and (V, ) are coherently
= f (da) d f i1 d f ik oriented if the transition function 1 is an orientation-
preserving diffeomorphism of Rm .
= f da dyi1 dyik = f (d).

Suppose now M m is an oriented Riemannian manifold. At
any p M there is a unique p m T p M such that
The contraction of a form with a vector field (also known p (e1 , . . . , em ) = +1 for any oriented orthonormal basis
as interior multiplication) has a seemingly trivial defini- {e1 , . . . , em } for T p M. These fit together to give the Rie-
tion: if k M and X X(M) then X k1 is given mannian volume form m M. In terms of the Hodge
by star, we have = ?1.
X (X2 , . . . , Xk ) := (X, X2 , . . . , Xk ). Given an oriented coordinate chart (U, ) then at any p
U we have the coordinate basis {i } for T p M but can also
choose an oriented orthonormal basis {ek }. P
Then of course
First note that this is a purely pointwise operation, so we
for some matrix A = aki we have i = k aki ek . Since

could define it on k V for a single vector space even
hek , e` i = k` , we get
proving the next proposition at that level but we wont
bother. (It is the adjoint of the operator on k V given by DX k X ` E X k k
gi j = i , j = ai ek , a j , e` = ai a j .

left multiplication by X.)

k
Next note that for a 1-form, X () = (X) 0 M. For a
0-form f 0 M = C M we set X f = 0. As a matrix equation, we can write gi j = AT A, which

implies det(gi j ) = (det A)2 . Since both bases are positively
Proposition B4.5. For any X, the operation X is an an- oriented, we know det A > 0, so det A = + det g. (Note
p
tiderivation on M of degree 1 whose square is zero. that while abbreviating det(gi j ) as det g is common, it un-
fortunately hides the fact that this is an expression in par-
Proof. It is clear that X X = 0 since ticular coordinates.)
X X (. . .) = (X, X, . . .) = 0. Now we compute
p (1 , . . . , m ) = (det A) p (e1 , . . . , em ) = det A = det g.

p
The antiderivation property is
Equivalently, we have the coordinate expression
X ( ) = (X ) + (1)k (X )
= det g dx1 dxm .
p
for k M; we leave the proof as an exercise.
On an oriented Riemannian manifold (M m , g), any m-form
We will later discuss Cartans Magic Formula, relating this
is a multiple = f = ? f of the volume form , with
contraction to exterior and Lie derivatives.
f C M.
B5. Volume forms and orientation

B6. Integration
An orientation on an m-dimensional vector space V is a

We will base our integration theory on the Riemann inte-
choice of component of m V r {0} R r {0}, that is a
gral. Recall that given an arbitrary real-valued function f
choice of a nonzero m-form in m V (up to positive
on a box B = [a1 , b1 ] [am , bm ] Rm we define up-
real multiples). If V is oriented by , then an ordered ba-
per and lower Riemann sums over arbitrary partitions into
sis {e1 , . . . , em } for V is said to be positively oriented if
small boxes the function f is Riemann integrable if these
(e1 , . . . , em ) > 0. Often an orientation on V is defined
have the same limiting value, which we call
through such a basis (to avoid the machinery of the exte-
rior algebra).
Z
f dx1 dxm .
A volume form on a manifold M is a nowhere vanishing B
m-form m M. We say M is orientable if it admits a
volume form. (The Mbius strip and the Klein bottle are Recall also that A Rm has (Lebesgue) measure zero if
examples of nonorientable 2-manifolds.) An orientation for each > 0 there is a covering of A by countably many
of M is a choice of volume form, up to pointwise multipli- boxes of total volume less than . The image of a set of
cation by positive smooth functions > 0 C M. This measure zero under a diffeomorphism (or indeed under any
is the same as a continuous choice of orientations of the locally Lipschitz map) again has measure zero. Thus we
tangent spaces T p M. A connected orientable manifold has can also speak of subsets of measure zero in a manifold M.
exactly two orientations. Given a function f : D R with D Rm , we define its
The standard orientation on Rm is given by dx1 dxm , extension by zero f : Rm R by setting f = f on D
so that {e1 , . . . , em } is an oriented basis for each T p M. and f = 0 elsewhere. Lebesgue proved the following: A
24

bounded function f : D R defined on a bounded do- First consider a single (oriented) chart (U, ) and assume
main D Rm is Riemann integrable if and only if f is m c U. Then we define
continuous almost everywhere, meaning that its set of dis- Z Z
continuities has measure zero. := 1 .

(U)
For instance, the characteristic function D is Riemann in- U
tegrable if D is bounded and its boundary D has measure We claim this is independent of : if (U, ) is another ori-
zero. Then we call D a domain of integration. ented chart, then using the diffeomorphism 1 we find
Because a continuous function f on a compact set is Z Z Z
bounded, we find: If U Rm is open and f : U R 1 = 1 1 = 1 .

has compact support in U, then f is Riemann integrable. (U) (U) (U)
We write kc M k M for the subspace of k-forms with In general, we choose a partition of unity { f } subordinate
compact support. (If M is compact, then of course kc = to an oriented atlas (U , ) . For any m c M, note
k .) that = f is a finite sum and each summand has
P
compact support in the respective U . We define
Definition B6.1. If m c U is an m-form with compact
support in U Rm then of course we can write uniquely
Z XZ
:= f .
= f dx1 dxm . We define M U

Z Z Z
We just need to check this is independent of the choice of
= f dx1 dxm := f dx1 dxm .
U U U atlas and partition of unity.
So suppose {g } is a partition
of unity subordinate to an-
Note that we use the standard basis element for m Rm other oriented atlas (V , ) . Then we have
dx1 =
R
2
here.
R Otherwise we have for instance f dx XZ XZ X
f dx1 dx2 . f = f g
U U
Lemma B6.2. If : U V is a diffeomorphism of con- XXZ XXZ
nected open sets in Rm and an m-form with compact = f g = f g .
support in V, then U U V
Z Z But byPsymmetry, we see that the last expression also

= , equals V g , as desired.
R
U V
Note: If M denotesRthe manifold R M with opposite orien-

where the sign depends on whether is orientation- tation, then we have M = M
preserving or not.
Note: for m = 0, a compact oriented 0-manifold is aPfinite
collection of points with signs 1: we write M = pi
Proof. Use xi for the standard coordinates on U and y j for P
q j . (Here we cannot use charts to test orientation.) The
those on V. Then = f dy1 dym for some func-
tion f . Writing i = yi , the Jacobian matrix of is
integral of P
a zero-formP(function) f : M R is defined to
be M f = i f (pi ) j f (q j ).
R
J := (i /x j ). We have di = dyi and so
We have developed this theory for smooth forms, partly
d1 dm = det J dx1 dxm . just because we have no notation for possibly discontin-
uous sections of m T M. As long as is bounded, van-
Thus ishes outside some compact set and is continuous almost
everywhere, we can R repeat the calculations above with no
changes to define M .
Z Z
= ( f ) d1 dm (1)
U U
Z End of Lecture 14 Dec 2015
= ( f ) det J dx1 dxm . (2) On an oriented Riemannian manifold M (or any manifold
U
with a specified volume form ), we define the volume
On the other hand, the standard change-of-variabes for- integral of a function f C M as
mula says Z Z Z
Z Z Z f d vol := f = ? f.
M M M
= f dy dy =
1 m
( f ) | det J| dx1 dxm .
V V U Note that if we switch orientation, the volume form on M
Since U is connected, det J has a constant sign, depending Ris , so the volume integral is independent of orientation:
=
R
f d vol f d vol.
on whether is orientation-preserving. M M
For a domain D M (compact with boundary of measure
zero) we define its volume to be
Now suppose M m is an oriented manifold, and m
c M Z Z Z
is
R a compactly supported m-form. Then we will define
R. vol D := 1 d vol = := D 0.
M D D M
25

The volume of the manifold is vol(M) := M 1 d vol =

R
These special cases are of course normally formulated not
. This works directly if M is compact; for a non-
R
with differential forms and the exterior derivative, but with
M
compact manifold we can take a limit over an appropriate gradients of functions, and divergence and curl of vector
compact exhaustion and reach either a finite value or +. fields. More precisely, on any Riemannian manifold, we
use the inner product to identify T p M and T p M and thus
vector fields with one-forms. The gradient f of a function
B7. Manifolds with boundary f C M is the vector field corresponding in this way to
d f . In particular, for any vector field X, we have
Suppose M m is an manifold with boundary; its boundary g( f, X) = h f, Xi = d f (X) = X f.

M is an (m 1)-manifold. At p M M we see
that T p M T p M is a hyperplane, cutting T p M into two On R3 we further use the Hodge star to identify vectors
parts, consisting of the inward- and outward-pointing vec- with pseudovectors and thus one-forms with two-forms,
tors at p. and to identify scalars with pseudoscalars and thus zero-
forms with three-forms. Then div, grad and curl are all
An orientation on M induces an orientation on M as fol- just the exterior derivative. Explicitly, we identify both the
lows. Suppose (v, v1 , . . . , vm1 ) is an oriented basis for one-form p dx + q dy + r dz and the two-form p dy dz +
T p M, where v is outward-pointing and vi T p M. Then q dz dx + r dx dy with the vector field p x + qy + rz ,
(v1 , . . . , vm1 ) is by definition an oriented basis for T p M. and the three-form f dxdydz with the function f . Then
(There are four obvious possible conventions here either d : 0 1 is the gradient as above, d : 1 2 is the
and inward- or outward-pointing vector could be put either curl, and d : 2 3 is the divergence.
before or after the basis for T p M. Our convention works
best for Stokes Theorem. Our version of Stokes theorem is (as mentioned above)
certainly not the most general. For instance, we could eas-
Equivalently, suppose the orientation of M is given by a ily allow manifolds with corners, like compact domains
volume form , and we pick a vector field X X(M) with piecewise smooth boundaries. (It should be clear that
which is outward-pointing along M. Then the contraction the divergence theorem in R3 is valid for a cube as well as
X () restricted to M is a volume form on the boundary a sphere.)
which defines its orientation.
Theorem B8.2 (Stokes). Suppose M m is an oriented man-
ifold with boundary and is an (m 1)-form on M with
B8. Stokes Theorem compact support. Then
Z Z
d = .
Suppose M m is an oriented manifold with boundary and M M
is an (m 1)-form Rwith compact support on M. Stokes
Theorem then says M d = M . We see d2 = 0 is dual
R
Proof. Both sides are linear and integrals are defined via
to the condition that (M) = : partitions of unity. In particular
Z Z Z X X X X
0= d =
2
d = = 0. d = d( f ) = d f + f d = f d,
M M M
we see that it suffices to consider the case when is
Stokes Theorem is quite fundamental, and can be used
compactly supported inside one oriented coordinate chart
for instance to define d for nonsmooth forms, or M for
(U, ). We may also assume that (U) = Rm or (U) =
generalized surfaces M.
H m , depending on whether U is disjoint from M or not.
Remark B8.1. Of course in M , the integrand is really
R
Since the statement of the theorem is invariant under pull-
the restriction or pullback |M = i of to M. This is back by a diffeomorphism, we have shown it suffices to
now a top-dimensional form on the (m 1)-manifold M. consider the cases (a) M = Rm and (b) M = H m .
When M is a manifold without boundary (M = ) of After scaling, we can assume that is compactly sup-
course Stokes Theorem reduces to M d = 0. It turns out
R
ported within the cube (a) Q := (1, 0)m or (b) Q :=
that on a connected orientable closed manifold M m , anR m- (1, 0] (1, 0)m1 . In either case, we write
form can be written as d for some if and only if M m
X
vanishes; we will return to such questions after proving the = (1) j1 j dx1 dx
cj dxm
theorem. j=1
Stokes himself would probably not recognize this gener-
alized version of his theorem. The modern formulation in with supp j Q, so that
terms of differential forms is due mainly to lie Cartan. X j
The classical cases are those in low dimensions. For M an d = dx1 dxm ,
interval (m = 1), we just have the fundamental theorem x j
of calculus; for a domain in R2 , we have Greens theorem; meaning that
for a domain in R3 , we have Gausss divergence theorem;
and for a surface with boundary in R3 we have the theorem
Z X Z j
d = dx1 dxm .
Q x
attributed to Stokes. j
M
26

Now for each j we have Theorem B9.2. If M m is an orientable closed manifold

with n components, then H 0 (M) Rn .
j j j 1
Z Z 0 Z 0 Z 0
= cj dxm .
dx dx dx
Q x 1 x Proof. Note that B0 = 0 so H 0 = Z 0 , which is the space
j j
1 1
of functions with vanishing differential. But these are just
By the fundamental theorem of calculus, the inner inte-
the locally constant functions, so it is clear this space is
gral in parentheses equals j (. . . , 0, . . .) j (. . . , 1, . . .).
n-dimensional.
Since has compact support in Q, this vanishes for j > 1.
R (a) it vanishes even for j = 1, completing the proof
In case For orientable closed manifolds M m , Poincar duality (re-
that M d = 0. lated to the Hodge star operation) gives a connection be-
In case (b) we have obtained tween co/homology in complementary dimensions. As
Z Z 0 Z 0 an example, if such a manifold has n components, then
d = 1 (0, x2 , . . . , xm ) dx2 dxm . H m (M) Rn . We prove the dimension is at least this big.
Hm 1 1
Theorem B9.3. If M m is an orientable closed manifold
Now consider the restriction of to H m , the pullback with n components, then H m (M) has dimension at least n.
under the inclusion map i. Since i dx1 = 0 we immediately
get Proof. Denote the
R components by Mi . By Stokes, inte-
i = dx dx .
1 2 m gration 7 M over each component gives a map
i
R m = Z m R vanishing on Bm , and thus a map H m R;
Comparing this to the formula for M d shows we are together these give a map to Rn . Choosing a Riemannian
done. metric on Mi , its volume form has positive integral; these
n forms show that our map H m Rn is surjective.
B9. De Rham cohomology

B10. Lie derivatives
Definition B9.1. We say a k-form on M is closed if m
d = 0; we say is exact if there is a (k 1)-form such Earlier we defined the Lie derivative of a vector field Y
that d = . For clarity, write dk := d|k : k k+1 . with respect to a vector field X. This is a derivative along
We write Bk (M) for the space of exact forms and Z k (M) the integral curves of X, where we use (pushforwards un-
for the space of closed forms. That is, Z k = ker dk and der) the flow t of X to move vectors of Y between different
Bk = Im dk1 . points along these curves.
Since by definition d2 = 0, it is clear that exact forms are The Lie derivative of a differential k-form is defined in
closed. (Algebraically, we have Bk Z k k .) An in- the same way, except that the pushforward under t is
teresting question is to what extent the converse fails to be replaced by a pullback under t . That is, we define:
true. The answer is measured by the de Rham cohomology d

d

LX p := t t (p) = (t ) p .

H k (M) := Z k /Bk , the quotient vector space (over R). A dt t=0 dt t=0
typical element is the equivalence class [] = { + d} of
a closed k-form . Note that this is again a k-form. In the particular case of
If we consider all degrees k together, we set k = 0 where = f C M we can ignore the pullback
LX f is simply the derivative of f along the integral curve,
Z := Z 0 Z m = ker d, B := B0 Bm = Im d. that is, LX f = X f .
Defining Proposition B10.1. The Lie derivative LX on forms satis-
fies the following properties:
H := Z/B = H 0 H m
1. it is a derivation on M, that is, an R-linear map
we find this cohomology ring is not just a vector space but satisfying
indeed an algebra under the wedge product. To check the
details, start by noting that if 0 is closed, then LX ( ) = (LX ) + (LX );
( + d) 0 = 0 + d( 0 ). 2. it commutes with the exterior derivative, that is,
An important theorem in the topology of manifolds says LX (d) = d(LX );

that this cohomology agrees with other standard defini-
tions, in particular that it is dual to singular homology. 3. it satisfies the product formula for a k-form
(This is defined via cycles of simplices modulo boundaries, applied to k vector fields Yi X(M) we have
and can be thought of as counting loops or handles in di-
LX (Y1 , . . . , Yk )

mension k.) The key here is Stokes Theorem: a closed
form integrates to zero over any boundary, so closed forms k
X
= (LX )(Y1 , . . . , Yk ) + Y1 , . . . , LX Yi , . . . , Yk .

can be integrated over homology classes. Furthermore an
exact form integrates to zero over any cycle. i=1
27

Proof. 1. This follows directly from the fact that pull- Proof. We know that LX is a derivation commuting with d.
back commutes with wedge product and from the Since d2 = 0, it is easy to check the right-hand side also
product rule for d/dt: commutes with d. Furthermore it is a derivation: for
k M we get
d
LX ( ) p = ( ) p

dt t=0 t dX ( ) + X d( )

d
= (t ) p (t ) p = d((X ) ) + (1)k d( X )
dt t=0
! ! + X ((d) ) + (1)k X ( d)
d d
= ( ) p p + p

( ) p

= (dX ) + (1)k (X ) (d) +
dt t=0 t dt t=0 t

= (dX ) + (X d) + (dX ) + (X d).
= (L x ) p p + p (LX ) p .
Thus if the formula holds for and , it also holds for
2. This follows from the fact that d is linear and com- and for d. By linearity and locality, this means it is
mutes with pullback: enough to check it for 0-forms:

d d (dX + X d) f = X d f = (d f )(X) = X f = LX f.
LX d = t d = dt
dt t=0 dt t=0
d Proposition B10.3. Suppose X and Y are vector fields on
=d = dLX .
M m and is a 1-form. Then
dt t=0 t
3. The proof follows (as for the product rule for d/dt) d(X, Y) = X(Y) Y(X) ([X, Y]).
from a clever splitting of one difference quotient into
two or more. We will write out the proof only for Proof. We use Cartans Magic Formula and the product
k = 1, considering (Y). We find rule for LX :
d(X, Y) = (X d)(Y)

LX (Y) p = lim 1t t p Yt p p (Y p )

t0
= (LX )(Y) (dX )(Y)
= lim 1t t p Yt p p t Yt p

= X (Y) [X, Y] d (X) (Y)

t0
= X (Y) [X, Y] Y (X) .

+ lim 1t p t Yt p p (Y p ) .

t0
Note that by linearity and locality it suffices to con-

Here the second limit clearly gives
sider = f dg. So an alternate proof simply computes
d each term for this case, getting for instance X(Y) =
p t Yt p = p LX Y . X( f dg(Y)) = X( f Yg) = (X f )(Yg) + f XYg.

dt t=0
Theorem B10.4. Suppose k (M m ) is a k-form and
For the first limit, we can rewrite the first term as
X0 , . . . , Xk X(M) are k + 1 vector fields. Then
t t p t Yt p , so that both terms are applied to
the same vector. The limit becomes (d)(X0 , . . . , Xk )
t (t p ) p
X
lim t Yt p ,
= (1)i Xi (X0 , . . . , X
bi , . . . , Xk )
t0 t 0ik
X
where the form clearly limits to (LX ) p and the vec- + (1)i+ j [Xi , X j ], X0 , . . . , X
bi , . . . , Xbj , . . . , Xk ).
tor to Y p . 0i< jk

Note that the case k = 0 is simply d f (X) = X f , and the
case k = 1 is the last proposition. The general proof by
End of Lecture 4 Jan 2016 induction on k is left as an exercise; the hint is to use Car-
tans magic formula as in the proof of the proposition to
Since the Lie derivatives of functions and vector fields are write
known, we can rewrite the product formula as a formula
for LX as follows:
(d)(X0 , . . . , Xk )
(LX )(Y1 , . . . , Yk ) = (LX0 )(X1 , . . . , Xk ) (dX0 )(X1 , . . . , Xk ).
k
X
= X (Y1 , . . . , Yk ) Y1 , . . . , [X, Yi ], . . . , Yk .

i=1
Proposition B10.2 (Cartans magic formula). For any

vector field X we have LX = dX + X d.
28

J.M. Sullivan, TU Berlin C: Riemannian geometry Diff Geom II, WS 2015/16
C. RIEMANNIAN GEOMETRY Example C1.2. Consider the round sphere S2 R3 and

let (t) := (cos t, sin t, 0) be the equator parametrized by
arclength. Consider the vector field X along given by the
To take derivatives of a vector field along a curve requires
tangent vector
comparing tangent spaces at different points. The Lie
derivative uses diffeomorphisms to do this, which is not X(t) := 0 (t) = ( sin t, cos t, 0).
entirely satisfactory since we need not just a curve but a
whole vector field. Since dX/dt = 00 (t) = (t) is normal to S2 , we find
Another approach is through connections or covariant DX/dt 0.
derivatives. In particular, there is a natural connection on In general, a parametrized curve on M is called a
any Riemannian manifold, which is the starting point for geodesic if its velocity vector field X = satisfies DX/dt
studying its geometry. 0. On the sphere, the geodesics are exactly the great circles
parametrized at constant speed.
C1. Submanifolds in Euclidean space Now we want to work out coordinate expressions for the
covariant derivative. So let (U, ) be a coordinate chart
for M m Rn and write V := (U) Rm . Write {ui :
Suppose M m Rn is a submanifold. A map X : M T Rn , i = 1, . . . , m} for the coordinates on Rm . Because M is
p 7 X p T p Rn is called an Rn -valued vector field along embedded in Rn , we can also write the inverse map
M. Of course T p Rn Rn , so we can identify X with a
function X : M Rn . 1 =: = (1 , . . . , n ) : V U M Rn
But T p Rn also has an orthogonal decomposition (with re-
spect to the standard Euclidean inner product) into spaces explicitly in coordinates. (Here we use {x : = 1, . . . , n}
tangent and normal to M: for the coordinates on Rn and have = x .) The
standard coordinate frame for T U is of course given by
T p Rn = T p M N p M. X
i = = .
We let k and denote the orthogonal projections onto ui
ui x
these subspaces, so that X p = k X p + X p .
Now if : [a, b] M is a curve (embedded) in M, then A curve in M will be given in coordinates as
we have the function X : [a, b] Rn and can take its
(t) = u1 (t), . . . , um (t)

derivative. We can view this derivative as an Rn -valued
function on the 1-submanifold M Rn instead of on
[a, b] (technically we compose with 1 ). Again such a for some real-valued functions ui (t).
map to Rn can be identified with an Rn -valued vector field A vector field Y (tangent to M) along can be expressed
along (viewing its value at each point p as lying in T p Rn ). in the coordinate basis as
We call this vector field the derivative dX/dt of X along . X
Y(t) = Y(t) = bi (t)i
Both the original field X and its derivative dX/dt can be
i
decomposed (via k and ) into parts tangent and normal
to M. These decompositions are not in any definite rela- for some real-valued functions bi (t). Its derivative and co-
tion to each other. Consider for instance vector fields along variant derivative along are then
a surface in R3 as we studied last semester. The deriva-
tives of the unit normal vector field are tangent vectors; the dY X dbi di
derivatives of tangent vector fields will usually have both = i + bi ,
dt i
dt dt
tangent and normal components.
DY X dbi Di
Definition C1.1. Suppose X is a smooth vector field on = i + bi .
dt dt dt
M m Rn and is a curve in M. Then the vector field i
DX dX To compute the covariant derivative Di /dt of the coordi-

:= k nate basis vectors, we recall that a time derivative along
dt dt
is a directional derivative in direction , so we get
along , which is tangent to M, is called the covariant d X X X 2 du j
derivative of X along . Di
= k = k .
dt dt ui x j
u j ui dt x
Note that we only need X to be defined along . Note also
that we could apply this definition to any Rn -valued field Here the /x are the standard basis vectors in Rn . Their
X, but there is little reason to do so our goal is to focus on tangent parts can of course be expressed in the coordinate
the geometry of M. Indeed, we will see that this covariant basis:
derivative can be defined in a way depending only on the X
Riemannian metric on M and independent of the particular k = ck k
embedding M Rn . x k
29

for some smooth functions ck C (U). it is C -linear in X:

We now define the so-called Christoffel symbols
f X Y = f X Y;
X 2
kij := ck ,
j ui
it satisfies a product rule in Y:

u
X ( f Y) = (X f )Y + f X Y;
noting the symmetry kij = kji . We have kij C (U) for
1 i, j, k m. it is symmetric in the following sense:
Using these, the expression above for the covariant deriva-
X Y Y X = [X, Y] = LX Y;
tive of i becomes
it is compatible with the Riemannian metric:
Di X k du j
= i j k .
dt dt X Y, Y 0 = X Y, Y 0 + Y, X Y 0 .

j,k
We can consider in particular the covariant derivative along The first two properties are easily verified. The symmetry
a u j -coordinate curve, where u j = t and each other ui is is equivalent to the fact that kij = kji . The metric property
constant. We write this as is left as an exercise.
Di X k
= i j k .
u j k C2. Connections
That is, the Christoffel symbol kij is the k component of

Let us now move to a very general situation. Suppose E is
the covariant derivative of i in direction j .
a vector bundle over a manifold M. A connection on E
We can now return to the general case of the covariant allows us to take covariant derivatives of sections of E.
derivative of Y along ; our formula becomes These are directional derivatives in the direction of some
vector field X X(M) and are again sections of the same
DY X dbk X k i du j bundle E. That is, given a section (E), its covariant
= + i j b k .
dt k
dt i, j
dt derivative (with respect to ) in direction X is the section
X (E). The formal definition is as follows:
Note here that we dont see the coordinates in Rn at all; the
vector field Y and curve on M are expressed in the stan- Definition C2.1. Given a vector bundle E M, a con-
dard instrinsic ways in the coordinate chart (U, ). The em- nection on E is a bilinear map : X(M) (E) (E),
bedding of M Rn enters only in the computation of the written (X, ) 7 X , which is C (M)-linear in X and
Christoffel symbols kij , and our goal is to show these re- satisfies a product rule for :
ally only depend on the Riemannian metric induced on M
by the embedding. f X = f X , X ( f ) = (X f ) + f X .
Now suppose Y = bk k is a vector field defined on all We call X the covariant derivative of .
P
of M (rather than just along ) its components bk are
now functions on U. We note that the covariant derivative Note that the tensoriality (C M-linearity) implies that the
DY/dt at a point p = (t0 ) doesnt depend on the whole dependence on X is pointwise: X ) p depends only on
curve but only on its velocity vector X p := P (t0 ) there. X p and can be written as X p . This covariant deriva-
In particular, if we set a j := du j /dt then X p = a j j , and tive of course depends on more than just p , but as for the
the time derivative dbk /dt appearing in the formula is the other derivatives we have studied, the product rule means
directional derivative X p (bk ). that the definition is local: if and have the same germ
To emphasize this viewpoint, we introduce new notation at p (that is, agree in some open neighborhood U) then
and write this covariant derivative of Y at p in the direction X p = X p . The trick is again to pick a bump function
X p as X p Y. If X and Y are vector fields on M, we write f supported within U with f 1 on some smaller neigh-
X Y for the vector field whose value at p is X p Y. The borhood of p, so then f = f . We calculate
formulas above mean that if X = a j j and Y = bk k
P P
X p = 1X p + 0 p = f (p)X p + (X p f ) p = X p ( f )
in some coordinate chart, then
X X with the same for . Indeed, it suffices that and agree
X Y = a j ( j bk ) + kij bi a j k . locally along some curve with (0) = X p ; this can per-
j,k i
haps most easily be seen in coordinates as below.
We have thus defined a connection, meaning an operation There are many other ways to rephrase this definition, for
instance in terms of sections of various induced bundles.
: X(M) X(M) X(M), : (X, Y) 7 X Y. For any fixed , we can consider as a map taking a
vector field X to the section X . But the pointwise de-
This is clearly bilinear (over R) and we claim it satisfies pendence on X means that this acts pointwise as a linear
the following four properties: map T p M E p . That is, can be viewed as a section
30

of the bundle L(T M, E) = E T M; such a section can be of the properties we observed for the connection induced
called a vector-valued one-form. In this picture, the con- from an embedding M Rn .) In terms of Christoffel sym-
nection is a map from (E) to (E T M). While this bols in a coordinate basis (where [i , j ] = 0), we find that
is the approach taken in many books, we will stick to our is torsion-free if and only if kij = kji .
more down-to-earth approach. On a Riemannian manifold (M, g) we can also ask whether
End of Lecture 11 Jan 2016 a connection on T M is compatible with the metric. A
metric connection is one satisfying
We have already seen one example of a connection: the
one on T M induced by an embedding M Rn , which X Y, Z = X Y, Z + Y, X Z .

satisfied not only the properties in this definition but also
two further properties. As in that case, any connection can One interpretation of this equation is as saying that the
be expressed in coordinates via Christoffel symbols. metric tensor g is parallel with respect to . Just as we
Suppose U is a coordinate neighborhood for M and a triv- saw for the Lie derivative, a connection on one bundle nat-
ializing neighborhood for E, with {i : 1 i dim M} the urally induces connections on the dual bundle and its ten-
coordinate frame for T M and {ea : 1 a rk E} a frame sor powers such that product rules hold. In particular, we
for E. Then a connection is expressed in coordinates by could define the covariant derivative X g via
bia defined by Pi ea = b bia eb , so
P
the Christoffel symbols P
that in general for X = vi i and = a ea we have X g(Y, Z) = (X g)(Y, Z) + g(X Y, Z) + g(Y, X Z).

X X
X = vi i b + bia a eb . Then clearly is a metric connection if and only if for
i,b a all X we have X g = 0.
We will now show that any Riemannian manifold has a
Any collection of smooth functions bia describes a connec-
unique torsion-free metric connection ; this is called the
tion.
Levi-Civita connection. Note that we have already con-
Any connection induces a notion of parallel transport structed such a connection on any manifold M Rn em-
along a curve. If is a curve (from p to q) in M, then bedded in Euclidean space. We give a proof due to Koszul,
a section of E (defined at least on more formally a using the fact that a vector field X Y on a Riemannian
section of E) is said to be parallel along (with respect manifold is specified by its inner products with arbitrary
to the connection ) if (t) 0 along . This corre- vector fields Z.
sponds to a first-order ODE , which has a unique solution
given any initial value. That is, given any p E p , there is Theorem C3.1. Any Riemannian manifold (M, g) has
a unique extension to a parallel section along . In par- a unique Levi-Civita connection, characterized by the
ticular, looking at its value q at the endpoint q, we get a Koszul formula
linear map P : E p Eq called parallel transport along
2g(X Y, Z) = X g(Y, Z) + Y g(X, Z) Z g(X, Y)

(w.r.t. ).
+ g [X, Y], Z g [X, Z], Y g [Y, Z], X .

If is a closed curve a loop based at p then we call
P : E p E p the holonomy of around . Note that
this need not be the identity; instead it demonstrates the Proof. Because the metric is fixed, we write g(, ) as h, i.
curvature of the connection . The uniqueness amounts to checking that any Levi-Civita
connection does satisfy the Koszul formula. We use the
The tangent space to E at each point has a natural vertical
metric property to expand each of the first three terms; the
subspace of dimension k at each point: the tangent space to
first (for instance) becomes hX Y, Zi + hX Z, Yi. We use
the fiber E p or equivalently the kernel of the differential of
the symmetry to expand each of the last three terms; the
the projection E M. Another way to view a connection
first (for instance) becomes hX Y, Zi hY X, Zi. Adding
is as a choice of a complementary horizontal subspace of
everything we find that most terms cancel out; we are left
dimension m. A section is parallel along if D() lies
with 2 hX Y, Zi.
in these horizontal subspaces.
It remains to show that the formula does define a Levi-
Civita connection. First, we claim that the right-hand side
C3. The Levi-Civita Connection is tensorial (meaning C M-linear) in Z:
X Y, f Z + Y X, f Z f Z X, Y

Specializing to the case of connections on the tangent bun- + [X, Y], f Z [X, f Z], Y [Y, f Z], X

dle E = T M, we can compare X Y and Y X. It is too much
= (X f ) Y, Z + f X Y, Z + (Y f ) X, Z + f Y X, Z

to hope that these are the same for any vector fields X and
f Z X, Y + f [X, Y], Z

Y the behavior when we replace X by f X is different. But
this kind of effect is captured also in the Lie bracket of the

f [X, Z], Y (X f ) Z, Y
vector fields. We define the torsion of the connection as

f [Y, Z], X (Y f ) Z, X

T (X, Y) := X Y Y X [X, Y]. This expression is C M-

= f X Y, Z + Y X, Z Z X, Y

linear in each of its arguments. The connection is said
to be symmetric or torsion-free if T (X, Y) = 0 for all X and
+ [X, Y], Z [X, Z], Y [Y, Z], X

Y, that is, if X Y Y X = [X, Y]. (This is of course one
31

This means for any fixed X and Y, the right hand side along , meaning X Y = 0 = X Z. Then the metric con-
is (Z) for some one-form . But using the metric g, dition implies
this one-form equivalent to a vector field W defined by
X Y, Z = X Y, Z + Y, X Z = 0,

2g(W, Z) = (Z).
This construction : (X, Y) 7 W is clearly bilinear in X that is, Y, Z is constant along . In particular, the length

and Y. The facts that it is tensorial in X and satisfies the of any parallel field is constant, as is the angle between two
product rule in Y are verified by calculuations similar to parallel fields.
the one above. The the Koszul formula defines a connec-
tion . What remains to show is that it is symmetric and It follows that parallel fields along with respect to two
compatible with g. different metric connections will thus differ from one an-
other by some rotation the torsion-free Levi-Civita con-
To check the symmetry, note that the right-hand side of nection is in some sense the one for which parallel fields
the Koszul

formula
is symmetric in X and Y except for the rotate the least.
term [X, Y], Z . Thus
End of Lecture 18 Jan 2016
2 X Y, Z 2 Y X, Z = [X, Y], Z [Y, X], Z

The metric condition means parallel transport P : T p M
= 2 [X, Y], Z .

T q M is an orthogonal transformation between these inner
product spaces; the holonomy around any loop based at
Since this holds for all Z, we conclude the connection is p is an element of O(T p M). (If M is orientable, then the
torsion-free: X Y Y X = [X, Y]. holonomy actually lives in S O(T p M). On an oriented sur-
To check the metric property, note that the right-hand

side face, for instance, the holonomy around any loop is rota-
tion by some angle .) Note that it is easy to extend the no-

is antisymmetric in Y and Z except for the term X Y, Z .
Thus tion of parallel transport to the case of a piecewise smooth
curve , simply by transporting along each smooth piece
2 X Y, Z + 2 Y, X Z = 2X Y, Z

in order.
Consider the example of the round sphere S2 R3 . Look
as desired.
at a right-angled triangle; at a lune of angle ; at a general
Now we want to consider what the Levi-Civita connection triangle. In each case we find the holonomy is 2 minus the
looks like in coordinates. We know the Christoffel symbols sum of the exterior angles, which also equals the enclosed
for a torsion-free connection will be symmetric: kij = kji . area.
In terms of the components gi j := g(i , j ) of the metric These are special cases of the GaussBonnet theorem: for
tensor, we can express the metric property of as follows: any disk D on any surface, the holonomy around D is
Z
k gi j = k g(i , j ) 2 TC(D) = K dA.
= g k i , j + g i , k j D

On the sphere we have K 1 so the right-hand side is just
X
= `ki g` j + `k j gì .
`
the area.
The Gauss curvature at a point p on a surface can be mea-
We can express this more simply in terms of another form sured by measuring the holonomy angle around a small
of Christoffel symbols. If we define loop based at p and dividing by the area of the loop.
X
i jk := ì j gk` = g i j , k ,

`
C5. Riemannian curvature
then we get k gi j = ki j + k ji . Using the symmetry i jk =
jik , we can solve this system to give The idea of Riemannian curvature is that given a two-plane
in T p M, the holonomy around an infinitesimal loop in this
2i jk = i g jk + j gik k gi j . plane will give an infinitesimal rotation of T p M. The two-
plane is specified by a two-vector, and the infinitesimal ro-
Writing gk` for the matrix inverse of gi j , we have

tation is given by an operator on T p M saying in which di-
X X gk` rection each vector moves.
kij = gk` i j` = i g j` + j gi` ` gi j . So suppose we have vector fields X and Y near p M. We
` `
2
consider parallel transport for time s along X followed by
time t along Y, and compare this with going the other way
around. Of course if [X, Y] , 0 this isnt even a closed
C4. Parallel transport and holonomy loop, but lets assume [X, Y] = 0. Then the holonomy
around this loop will be approximately st times what we
Suppose is any metric connection on a Riemannian man- call the curvature R(X, Y).
ifold M, and is a smooth curve from p to q in M with In general, of course we need to correct by [X, Y]. Re-
velocity vector X = . Suppose Y and Z are parallel fields call that this Lie bracket is the commutator of directional
32

derivatives: Add these two and subtract the remaining two cyclic per-
mutations. Using the antisymmetries (1) and (2), the result
0 = X(Y f ) Y(X f ) [X, Y] f. follows.
This inspires the definition of the Riemannian curvature End of Lecture 25 Jan 2016
operator
Proof. It remains to show properties (2) and (3). By ten-

R(X, Y)Z := X Y Z Y X Z [X,Y] Z. soriality, it suffices to prove (3) for the commuting basis
vector fields X = i , Y = j , Z = k . We will abbreviate
Lemma C5.1. On any Riemannian manifold the curva- i := i . First note that R(i , j )k = i ( j k ) j (i k ).
ture operator R(X, Y)Z is tensorial its value at p depends Thus the sum of three terms can be written as
only on X p , Y p and Z p . In particular, R(X p , Y p ) is a linear
i j k k j + j k i i k + k i j j i .

operator on T p M.
The proof proceeds by checking that Because the connection is torsion-free, each of the expres-
sions in parentheses is a Lie bracket like [ j , k ], but these
R( f X, Y)Z = R(X, f Y)Z = R(X, Y)( f Z) = f R(X, Y)Z, all vanish.
For (2) it also suffices to consider X = i , Y = j . Since
which follows from the product rules we have for the Lie the symmetric part of a bilinear form in determined by its
bracket and covariant derivative. The details are left as an associated quadratic form, to show the antisymmetry (2) it
exercise. suffices to prove
The definition implies directly that R(X, Y) = R(Y, X).
0 = R(i , j )Z, Z = i ( j Z) j (i Z), Z .

This antisymmetry means that we can think of this linear
operator as depending on X Y.

That is, it suffices to prove that i ( j Z), Z is symmetric
On an inner product space like T p M, a linear operator is in i and j. To do so, consider second derivatives of the
equivalent to bilinear form. Hence we get the Riemannian function hZ, Zi:
curvature tensor
j i hZ, Zi = j 2 Z, i Z

R(X, Y, Z, W) := R(X, Y)Z, W .

= 2 Z, j (i Z) + 2 j Z, i Z .

In coordinates (U, ), with respect to the coordinate frame The last term is clearly symmetric, and the left-hand side
{i }, the curvature has components given by is symmetric since [i , j ] = 0, so we are done.
X
R(k , ` )i =: Rij k` j , The antisymmetry properties (1) and (2) mean that the cur-
vature tensor really can and should be thought of as a bi-
X
R(k , ` )i , j =: Ri jk` = i k` .
g jm Rm

m
linear form on the space of two-vectors:

Theorem C5.2. The Riemannian curvature satisfies the S (X Y, Z W) := R(X, Y)Z, W
following symmetries:
(extended by bilinearity to nonsimple two-vectors). Prop-
(1) R(X, Y) = R(Y, X), erty (4) is then simply the symmetry of S :
R(X, Y)Z, W = R(X, Y)W, Z ,

(2) S (X Y, Z W) = S (Z W, X Y);
(3) R(X, Y)Z + R(Y, Z)X + R(Z, X)Y = 0,
(4)

R(X, Y)Z, W = R(Z, W)X, Y .

this symmetry of course holds for arbitrary two-vectors,
not just simple ones. In these terms, property (3) gets no
Note that we have already proved the antisymmetry (1). simpler:
The further antisymmetry (2) is equivalent to saying that
R(X, Y) is an infinitesimal rotation. S (X Y, Z W) + S (Y Z, X W) + S (Z X, Y W) = 0.
All four symmetries involve different permutations of the If X, Y is an orthonormal basis for a two-plane T p M,
vector fields X, Y, Z, W, and are related to each other. (A then K() := S (X Y, X Y) = R(X, Y, Y, X) is called
more sophisticated approach would study them in terms of the sectional curvature of . It turns out that this equals
representations of the symmetric group S 4 .) For instance, the Gauss curvature of the flattest surface N M with
it is easy to see that, given (4), properties (1) and (2) are T p N = , say foliated by geodesics through p tangent
equivalent. to .
Instead, we start by observing that (4) is an algebraic con- Since any symmetric bilinear form is determined by the
sequence of the first three. For this, write (3) as associated quadratic form, it is not surprising that the sec-
tional curvatures of two-planes determine R completely
R(X, Y)Z, W + R(Y, Z)X, W + R(Z, X)Y, W = 0.

but note that here we are considering only simple two-
Then cyclically permute XYZW to get vectors. The following lemma (applied to the difference
of two possible tensors with the same sectional curvatures)
R(Y, Z)W, X + R(Z, W)Y, X + R(W, Y)Z, X = 0.

takes care of this problem.
33

Lemma C5.3. Suppose S is a symmetric bilinear form on Suppose the geodesic starting at p with velocity X p exists
2 V satisfying for at least unit time. Then we define exp(X p ) := (1) to be
the point at time 1 along this geodesic. For any Rieman-
S (X Y, Z W) + S (Y Z, X W) + S (Z X, Y W) = 0. nian manifold M, this exponential map is defined on some
open neighborhood W of the zero-section of T M and is a
If S (X Y, X Y) = 0 for all X and Y, then S = 0. smooth map exp : W M. Note that t 7 exp(tX p ) is
the parametrized geodesic with constant speed kX p k; the
Proof. First compute length of this curve from p to exp(X p ) is kX p k.
0 = S (X + Y) Z, (X + Y) Z We use the notation exp p for the restriction of exp to T p M

W. Since the differential D p exp p is the identity map, for
= S (X Z, Y Z) + S (Y Z, X Z)
small > 0, exp p is a diffeormorphism from the -ball
= 2S (X Z, Y Z). in T p M to a neighborhood B (p) called a geodesic ball.
Choosing an orthonormal basis for T p M to identify it with
Now using this we get Rm , the map exp1 p : U T p M = R is called normal
m
coordinates around p on the normal neighborhood U.

0 = S X (Z + W), Y (Z + W)

Similar considerations show that the map (, exp) : X p 7
= S (X Z, Y W) + S (X W, Y Z).
(p, exp X p ) is a local diffeomorphism T M M M. For
Then we use this to show any p M and > 0, we deduce the existence of a neigh-
borhood N such that any two points in N are joined by a
S (Y Z, X W) = S (X W, Y Z) = S (X Z, Y W) unique geodesic of length less than .
= S (Z X, Y W). In normal coordinates around p, all geodesics through p
are the images of straight lines through the origin. Fur-
That is, S is invariant under a cyclic permutation of XYZ. thermore, at p we have gi j (0) = i j and kij = 0. The image
But we have assumed the sum of all three cyclic permuta- S r := exp p X p : kX p k = r of a sphere in T p M is called

tions is zero, so we find S = 0. a geodesic sphere around p M. It is easy to check that
geodesics through p meet each of these sphere orthogo-
Note that actually there is a formula with 16 terms giving nally and that (for small r < r0 ) any curve from S r to S r0
has length at least r0 r. Thus geodesics are locally shortest
6S (X Y, Z W) curves.
= S (X + Z) (Y + W), (X + Z) (Y + W)

Note that for small r, the geodesic sphere S r (p) is topo-
S (X + Z) Y, (X + Z) Y + . logically a sphere and is the boundary of the geodesic ball

Br (p), which is a topological ball. For larger r, the expo-
nential map may still exist but no longer be a diffeomor-
C6. The exponential map phism; these spheres and balls will start to overlap and in-
tersect themselves. The geodesic ball Br (p) is always the
metric ball in (M, d), the set of points at distance less than
Any tangent vector X p T p M determines a unique r from p.
geodesic starting at p with velocity X; this is the solution
So far, we have only discussed local existence of
to a second order ODE and will exist at least for some time
geodesics. A manifold is called geodesically complete if
> 0.
every geodesic can be extended indefinitely, that is, if exp
More precisely, we can consider the equation for a is defined on all of T M. The HopfRinow theorem says
geodesic x(t) in a coordinate chart (U, ). In terms of the this happens if and only if the metric space (M, d) is met-
Christoffel symbols we get rically complete. In particular, every compact manifold is
X geodesically complete.
xk + kij (x) xi x j = 0.
i, j
C7. Ricci and scalar curvatures

The existence theorem we have used before then says: For
any q U there is a neighborhood V 3 q and , > 0 such
that, given any initial conditions p V and X p T p M with If T p M is a two-plane, its image under the exponen-
kX p k < (meaning of course x(0) = phi(p) and x(0) = tial map exp p is locally a two-dimensional submanifold
X p ), a unique solution as above exists within U for |t| < through p. One can show that its Gauss curvature at p
, and it depends smoothly on the initial conditions. is the sectional curvature K(). (See for instance Boothby,
Of course if we rescale X p to X p , the geodesic is still Theorem VIII.4.7.)
the same curve, simply parametrized at times the speed. Suppose X T p M m is a unit vector. We define the Ricci
Thus there is a tradeoff between and ; the statement curvature in direction X to be the average sectional curva-
above is also true say for = 2. Note also that we can ture of two-planes including X:
piece together geodesic arcs and the resulting curve is still
Ric(X, X) = (m 1) ave3X K().
a geodesic, so we can apply the existence results indepen-
dent of any coordinate chart. A manifold with constant Ricci curvature is called an Ein-
34

stein manifold, because of the way this condition arises in

general relativity.
The Ricci curvature is a quadratic form; the associated
symmetric bilinear form on T p M is called the Ricci ten-
sor; if {ei } is an orthonormal frame, we have
m
X
Ric(X, Y) = S (X ei , Y ei ).
i=1
In coordinates, Rici j = k Rki k j . On an Einstein manifold,

P
Ric(X, Y) = cg(X, Y).
For m = 3, but not in higher dimensions, the Ricci curva-
ture determines the full Riemannian curvature tensor.
End of Lecture 1 Feb 2016
The Ricci flow dgi j /dt = 2 Rici j is a nonlinear heat

flow of Riemannian manifolds which tries to smooth out
the Ricci curvature. Perelmans proof of Thurstons ge-
ometrization conjecture (including the Poincar conjec-
ture) used Ricci flow on 3-manifolds.
The average Ricci curvature P in all directions at p gives the
scalar curvature S (p) = gi j Ri j .
The images under the exponential map of round balls and
spheres in T p M are called geodesic balls and spheres
Br (p) is the set of all points at distance less than r from p.
The scalar curvature measures the volume growth rate of
geodesic balls or spheres around p positive curvature
means the volume grows more slowly than in Euclidean
space, which negative curvature means it grows faster.
The Ricci curvature Ric(X, X) measures in a similar sense
the rate of spreading of geodesics emerging from p in di-
rections near X.
35

OMN 201612 110645 1 Course - Notes v1

Uploaded by

Copyright:

Available Formats

OMN 201612 110645 1 Course - Notes v1

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

OMN 201612 110645 1 Course - Notes v1

Uploaded by

Copyright:

Available Formats

J.M.

Sullivan, TU Berlin A: Curves Diff Geom I, SS 2015

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

Properties: r0 ds = r(s1 ) r(s2 ) . This simply says the distance be-

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

hN(s), (s)i = xy + yx . We have A = 0 xy ds and r2 =

Thus by the arithmetic-geometric mean inequality,

A5. The CauchyCrofton Formula Z L D Z L E

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

(L) (0) = (L) (0), T 0 =

T (s), T 0 = cos ds cos ds T (s), T 0 .

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

A9. Framings for curves in higher dimensions

A framing along a smooth curve in Rn is a choice of ori-

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

B. SURFACES spiral examples). If it is, then of course its image is a smooth

x x (Here, E, F and G are traditional names for the entries of this

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

X+X 0 Y = X Y + X 0 Y, X (Y + Y 0 ) = X Y + X Y 0 , (This is called the Gauss formula.)

allow us to express covariant derivatives of arbitrary vector ` `

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

gi j,k = ki j + k ji = ki j + jki , The Gauss/Weingarten system shows us how to write the

g11 h11 h22 h21 h12 = g11 det S = g11 K.

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

In curvature-line coordinates with g =

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

Proof: Consider a boundary point p and some neighbor-

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

Definition A1.10. An m-manifold with boundary is then a

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

smooth structure on Rm for m , 4.

Every smooth manifold has a (PL) triangulation. For

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

A4. Smooth maps, rank, immersions Corollary A4.3. Suppose U M m is open, f : U R is

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

vi = i (v); now two distinct vectors v , w must differ in P i that if {e

j = /y . We find that J is the matrix of D p f with

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

A6. The tangent bundle The canonical example of an m-dimensional submanifold

T p M to (p), v . We define the topology on T M by speci-

AMS Open Math Notes: Works in Progress 2016-12-15 09:56:30

Properties: r0 ds = r(s1 ) r(s2 ). This simply says the distance be-

DX dX To compute the covariant derivative Di /dt of the coordi-