You are on page 1of 19

Classnotes: Geometry & Control of Dynamical Systems, M. Kawski.

April 20, 2009 119

4.5 Gauss curvature


This section gives an introduction to curvature of manifolds. It begins with a survey of classical
notions of curvature of two dimensional surfaces embedded in Euclidean three-space, followed
in the next section by a generalization to general Riemannian manifolds. Subsequently, a very
concise and elegant characterization in terms of connections is given. The emphasis is on the
geometric character of the notion of curvature, i.e. its independence of choices of coordinates
and imbeddings in ambient spaces. Following the historical development it is natural to build on
familiar concepts of curvature of plane and space curves to develop useful notions of curvature
of surfaces. This is an opportune time to think about why one may want a quantitative notion
of curvature, and what its properties should be:

Exercise 4.34 Make lists of desirable properties of a notion of curvature, and of possible pur-
poses. [[Suggestions: How should curvature of a surface relate to curvature of curves, consider
e.g. cylinders. Should curvature be a scalar function or what else? How should it transform, if
at all, under coordinate changes? Under what other operations should it be invariant? What
manifolds should be distinguished from each other by curvature? ]]

The natural start point is to consider smooth hypersurfaces M 2 ⊆ R3 (with the induced metric).
At any point p ∈ M one may consider the curvatures κσ (0) of all smooth curves σ : 7→ M 2 with
σ(0) = p. The first significant step is to go from this large collection of all possible curves to those
that are intersections of M 2 with various planes. One might consider (suitably parameterized)
curves of intersection of M 2 with planes Πj = {q ∈ M 2 : xj (q) = xj (p)} through p that are
parallel to the coordinate planes. Alternatively, in the case that the surface is the graph of a
function x3 = f (x1 , x2 ) one might consider the curves of intersection with all planes through p
that contain the line `(t) = (x1 (p), x2 (p), t x3 (p)). Or consider all normal planes through p, i.e
those that contain the normal line ν(t) = p + tN (p) where N (p) is a normal vector to M 2 at p –
i.e. in the case that M 2 is the graph of a function f one might take N = (−(D1 f ), −(D2 f ), 1)
evaluated at (x1 (p), x2 (p)), with the usual identification of R3 with Tp R3 .

Exercise 4.35 Which of the above approaches is not geometric?


Before reading on, try to come up with explicit formulas for the curvatures of the intersection
curves with the various planes suggested above. Explore whether any information is lost when
only considering the curvature with curves of intersection with planes as opposed to all curves.
Explore if either one of the suggested alternatives allows one to recover the information of any
of the others – e.g. if all the curvatures at p of the curves of intersection with the normal planes
are known, can one use these to calculate all the other curvatures mentioned above.
Instead of considering all normal planes, might it be enough to consider two orthogonal planes,
or any two different normal planes.
[[Utilize the MAPLE worksheets to get a better feeling for these curves of intersection.]]

The historical developments over the last 250 years are highly recommended for reading. There
are many publications which discuss excerpts – the following notes are very close to Spivak II
which provides an enlightening discussion contrasting and comparing the original works, with
modern terminologies and tools. For notational convenience and clarity, the following survey of
classical result liberally shall identifies at any point p ∈ R3 the tangent spaces Tp R3 with R3 ,
and also frequently suppresses any mention of inclusion maps such as ı∗p : Tp M ,→ Tı(p) R3 . The
first major theorem is due to Euler (about 1760).
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 120

Theorem 4.23 Suppose M 2 ⊆ R3 is a smooth surface (2-manifold), p ∈ M and {Y1p , Y2p , ν(p)}
is a positively oriented orthonormal basis for Tp R3 such that {Y1p , Y2p } is a basis for Tp M . For
each θ ∈ R let σθ : (−εθ , εθ ) 7→ M (for suitable εθ ) be the unique curve parameterized by arc-
length with σθ (0) = p and such that hσθ0 (s) , − sin θ · Y1p + cos θ · Y2p i ≡ 0 for all s. (Technically,
σθ0 (s) is parallel-transported back to Tp M using the Euclidean connection.) Then there exist
κmin , κmax ∈ R and θ0 ∈ R such that the signed curvature κθ of σθ at t = 0 satisfies

κθ = cos2 (θ − θ0 )κmax + sin2 (θ − θ0 )κmin . (228)

In general it makes no sense to talk about signed curvatures of curves in R3 – but here each
curve σθ lies in a plane that inherits a natural orientation by requiring that the basis {cos θY1p +
sin θY2p , νp } is positively oriented.
This classical result identifies two principal curvatures κmin and κmax together with two or-
thogonal principal directions that determine the curvatures κθ of the curves σθ in all possible
directions. This situation hints that there is a symmetric quadratic form – the second derivative
(of something to be identified) lurking behind. In linear algebra language, the principal direc-
tions should be the eigenvectors, and the principal curvatures the eigenvalues associated with
this form. Naturally, one expects the trace and the determinant of associated with this quadratic
form be natural invariants – indeed they will later be called mean curvature and Gaussian cur-
vature. However, this theorem is really a statement about curves that lie in a surface. Its proof
is a straightforward calculation requiring little more than multi-variable calculus – provided
one judiciously chooses a good set of coordinates. (The subsequent discussion of Gauss and
Riemann’s work will recover these curvatures from a geometric viewpoint that really describes
curvature of surfaces as opposed to curvature of curves in a surface.)
Proof. Choose coordinates (x, y, z) for R3 such that p = 0 and νp = D3,(0,0,0)) = (0, 0, 1). Note
that this can always be achieved by using only translations and rotations – this is important
since the curvature of a space curve is invariant under such coordinate changes, compare exercise
(1.10). Furthermore, the rotation allows enough freedom such that if near p the surface is given
as the graph of a function z = f (x, y) and D1 D2 f (p) = 0.
Exercise 4.36 Suppose that f ∈ C ∞ (R2 ) satisfies f (0) = (D1 f )(0) = (D2 f )(0) = 0. Explicitly
calculate the coordinate change (rotation) in R2 such that in the new coordinates D1 D2 f (p) = 0.
In these coordinates the curve σθ is given as some reparameterization of

γ(t) = (t cos θ, t sin θ, f (t cos θ, t sin θ)) (229)

Use that (D1 f )(0) = (D2 f )(0) = 0 to see that γθ0 (0) = (cos θ, sin θ, 0). Similarly, (D1 D2 f )(0) = 0
(and the induced orientation on the plane containing the image of the curve) yields that

κθ = γθ00 (0) = (0, 0, cos2 θ(D12 f )(0) + sin2 θ(D22 f )(0)). (230)

A closely related theorem of Meusnier in 1776 obtains the curvatures of the curves of intersection
of the surface with arbitrary planes.
Theorem 4.24 Suppose p ∈ M 2 ⊆ R3 and νp is a normal vector to M at p. Suppose Πφ and
Π⊥ are planes that intersect in a line through p that is perpendicular to νp , and such that νp in
contained in Π⊥ and the angle between (the normals of ) the planes is φ. If κφ and κ⊥ are the
curvatures of the curves of intersection of M with Πφ and Π⊥ at p, respectively, then

κφ · cos φ = κ⊥ . (231)
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 121

For a simple, modern proof of this theorem, and to prepare for the subsequent discussions it is
useful to introduce the notion of a smoothly varying unit normal vector field:
Lemma 4.25 Suppose M n ⊆ Rm+1 is an oriented immersed hypersurface. Then there exists a
unique smooth function ν : M 7→ T Rn+1 , called a unit normal vector field, with π ◦ ν = ı, (i.e.
νp ∈ Tı(p) R3 for all p ∈ M ) and kν(·)k ≡ 1, such that whenever {X1p , . . . Xmp } ⊆ Tp M is a
positively oriented basis, then {ı∗p X1p , . . . ı∗p Xmp , νp } ⊆ Tı(p) Rm+1 is a positively oriented basis.
This existence is an immediate consequence of the basic linear algebra theorem that any basis
for a subspace can be extended to a basis for the whole space. The smoothness is immediate,
e.g. from working in a chart. (One also may argue using the Gram-Schmidt algorithm).

The following discussions almost entirely address only (imbedded) smooth surfaces in R3 , and
it is convenient to be a little less precise: Throughout most of the following discussions we shall
omit mention of maps such as ı and ı∗ and conveniently identify e.g. p ∈ M with ı(p) ∈ RM +1 ,
Xp ∈ Tp M with ı∗p Xp ∈ Tı(p) Rm+1 and the Euclidean inner product h· , ·iı(p) on Tı(p) Rm+1 with
its pullback ı∗ ( h· , ·iı(p) ) to Tp M . Moreover, if τ : Rm+1 7→ Rm+1 is a translation τ (p) = p + q
for some fixed q ∈ Rm+1 , then also identify Xp ∈ Tp Rm+1 with τ∗p Xp ∈ Tτ (p) Rm+1 .

Check carefully: The sequel uses Euclidean connection ∇.


Should this be first introduced in section on geodesics and Γkij ?

Exercise 4.37 Explain what this latter translation has to do with connections and parallel trans-
port, i.e. how the identifications described above make use of the trivial Euclidean connection
∇Di Dj ≡ 0 for all i, j ≤ m + 1 on Rm+1 .

Suppose M 2 ⊆ R3 is an immersed smooth surface and σ : (−ε, ε) 7→ M is a smooth curve


parameterized by arc-length (i.e. kσk ≡ 1) with σ(0) = p. If necessary, working locally, assume
that M has an orientation and let ν be the associated unit normal field. Since the curve lies in
M it follows that hσ 0 , ν ◦ σi ≡ 0. Differentiating this identity and evaluating at t = 0 yields, in
modern notation
hσ 00 (0) , (ν ◦ σ)(0)i = −hσ 0 (0) , dt
D
(ν ◦ σ)i (232)
t=0

[[Here D
dt is the covariant derivative along σ associated
to the Euclidean connection on R3 , i.e.
if σ 0 (t) = ai (t)Di and ν(q) = bj (q)Dj then dt D
(ν ◦ σ) = i,j (ai (0) · (Di bj )(p))Dj .
P P P
t=0
The following pages frequently will use the language of connections, but it will always be the
trivial Euclidean connections: It provides a very concise notation even for various expressions
arising in classical calculations, and it is meant to suggest the obvious generalizations to not
necessarily imbedded manifolds. ]]

The right hand side is uniquely determined by the value of σ 0 at t = 0, and writing Q(σ 0 (0)) for
this expression, it defines a function Q : Tp M 7→ R.

Proof (of Meusnier’s theorem). Using the conventions and notation of the preceding discussion
and of the theorem statement suppose that σ⊥ : (−ε, ε) 7→ M ∩ Π⊥ and σφ : (−ε, ε) 7→ M ∩ Πφ
are smooth curves parameterized by arc-length such that σ⊥ (0) = σφ (0) = p and Xp = σ⊥ 0 (0) =
0 00
σφ (0). The signed curvature of either curve is the magnitude of the second derivative σ⊥ (0) or
σφ00 (0). Since σ⊥ (t) ∈ Π⊥ for all t it is clear that σ⊥
00 (0) is a multiple of ν , and hence
p

00
κ⊥ (0) = hσ⊥ (0) , νp i = Q(Xp ). (233)
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 122

Similarly σφ (t) ∈ Πφ for all t implies that σφ00 (0) = κφ (0)Nφ is a multiple of a vector Nφ ∈ Πφ
characterized by hXp , Nφ i = 0 (and respecting the orientation convention). Since Xp ∈ P⊥ ∩ Πφ
this means that hNφ , νp i = cos φ, and thus hσφ00 (0) , νp i = κφ (0) cos φ The result follows from
combining this with (232)
00 0 !
κ⊥ (0) = hσ⊥ (0) , νp i = Q(σ⊥ (0)) = Q(σφ0 (0)) = hσφ00 (0) , νp i = κφ (0) · cos φ. (234)
The constructions leading to the proof given above are mostly elementary – except for the key
observation that the right hand side of (232) only depended on the single tangent vector σ 0 (0).
While one may establish this fact using only basic calculus, we consider this as a sign post
leading to the modern view point in terms of covariant derivatives.

The first step is to recognize the utility of the normal field ν, or rather its rate of change to give
a new characterization (definition!) of the (or rather “a”) curvature of a surface. Define the
Gauss map, and, – using modern terminology – its tangent map/differential:
Definition 4.11 Suppose that M m ⊆ Rm+1 is an oriented smooth hypersurface and ν : M 7→
T Rm+1 is the unit normal vector field (as constructed above) that provides the compatibility of the
orientations of M m and Rm+1 . With the natural identification, considered as a map ν : M 7→ S 2
this map is called the Gauss map. Its tangent map (or differential, see below) considered at each
p ∈ M as a map ILp = ν∗p : Tp M 7→ Tp M is called the Weingarten map.
Pictorially there are no problems with these maps, but it takes a little patience to sort out all of
the details if one wants to be precise about the natural identifications. Clearly they are all about
the trivial Euclidean connection, which allows one to identify tangent vectors Xp ∈ Tp Rm+1
Xq ∈ Tq Rm+1 at different points p, q ∈ Rm+1 with each other, and even with points in Rm+1 .
While not essential, it is a useful, and somewhat challenging exercise to walk through the details:
To begin with, νp ∈ Tı(p) Rm+1 has unit norm and the unit sphere in Tı(p) Rm+1 is identified with
the sphere S m ,→ Rm+1 via a suitable restriction Φ of the linear map P Φ : Tp Rm+1 7→ Rm+1 to
m+1
the set {Xı(p) ∈ Tı(p) R : kXı(p) k = 1} that takes the tangent vector j aj Dj,ı(p) to the point
(a1 , a2 , . . . am+1 ) ∈ S m ⊆ Rm+1 .
The tangent map νp∗ takes the tangent space Tp M to the tangent space Tν(p) (T Rm+1 ). The
tangent map of Φ then serves to identify Tν(p) (T Rm+1 ) with T(Φ◦ν)(p) S m . It remains to identify
T(Φ◦ν)(p) S m back with Tp M . For this, note that Tp M ⊥ νp ∈ Tı(p) Rm+1 means that the image
of Xp ∈ Tp M satisfies (Φ ◦ ν)∗p (Xp ) ∈ T(Φ◦ν)(p) S 2 ⊥ (Φ ◦ ν)(p) ∈ Rm+1 , where in this last step
(Φ ◦ ν)(p) is identified with a tangent vector in T(Φ◦ν)(p) Rm+1 .
An alternative viewpoint considers dν, or more accurately d(Φ ◦ ν)p as a map from Tp M to
Rm+1 : If ν(p) = j ν j (p)Dj ∈ Tı(p) Rm+1 and Xp = Xpi Di,p ∈ Tp M then
P P

dν(Xp ) = (dν 1 (Xp ), dν 2 (Xp ), . . . , dν m+1 (Xp ))


= (Xp (ν 1 ), Xp (ν 2 ), . . . Xp (ν m+1 ))
(235)

= Xp (ν 1 )D1 + Xp (ν 2 )D2 + . . . + Xp (ν m+1 )Dm+1
def
= ∇Xp ν
may again be identified with a tangent vector first in Tı(p) Rm+1 , but actually also with a tangent
vector in Tp M because dν(Xp ) ⊥ ν(p) in Tı(p) Rm+1 as is seen from differentiating the constant
function q 7→ kνq k2 = hνq , νq i ≡ 1 (as a function in C ∞ (M )), i.e.
0 = Xp (1) = Xp hν , νi = h νp , Xp (ν 1 )D1 + . . . + Xp (ν m+1 )Dm+1 i. (236)
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 123

These technical details are not essential when only considering curvature of smooth hypersur-
faces in Euclidean spaces – but they are helpful for the development of notions of curvature
in general manifolds, not necessarily Riemannian. The key is to keep in mind which parts of
the construction made above are only possible in Euclidean spaces, and which ones may be
generalized to general manifolds using connections.

Intuitively the rate of change of the normal vector field ν provides a measure of curvature. It
is worthwhile to first get a better intuitive understanding of this map, i.e. consider the images
ν(U ) ⊆ S m of subsets U ⊂ M m .

Exercise 4.38 Identify the images of familiar surfaces under the Gauss-map: Verify that on
the unit 2 2
p sphere the map ν is the identity. Find the images of the cylinder x +y = 1, of the cone
z = α x2 + y 2 (how does the image depend on α?), of any plane ax + by + cz = k. Continue
with describing the images of the paraboloid z = x2 + y 2 , of the saddle z = x2 − y 2 , and of the
graph of z(1 + x2 + y 2 ) = α (how does the image depend on α?).
[[Use the provided MAPLE worksheet to confirm your descriptions and to enjoy the images. . . ]]

Definition 4.12 Define the total curvature of an oriented smooth hypersurface M m ⊆ Rm+1 to
the volume (area) of the image ν(M m ) ⊆ S m .

Exercise 4.39 Calculate the total curvatures for each of the surfaces of exercise (4.38).

A natural proposal for a quantitative measure of the curvature at a point p ⊆ M m is to consider


sequences of neighborhoods Uk ⊆ M m that shrink [?] to the point p and propose

def ? (signed) area of ν(U )


k(p) = lim (237)
U →p area of U
This is basically what Gauss’ idea was. Compare this to defining curl and divergence in multi-
variable calculus as
~ · ds ~ ~
R RR
def ? F def ? ∂R F · dN
curlF~ (p) · N
~ = lim ∂R
and divF~ (p) = lim (238)
R→p ( area of R ) R→p ( volume of R )

where e.g. in the first case R is a planar surface element with normal N ~ , and in the second case
~
dN is the outward normal on the boundary of the region R. The main advantage is that such
definitions provide immediate understanding of what these objects measure, and they make the
associated integral theorems rather trivial: With these definitions (?) Stokes’ theorem (in its
narrow sense and as the divergence theorem) are immediate, and similarly will be a theorem that
relates the total curvature to the curvature at a point via an integral. On the other hand it takes
substantial, and rather unpleasant, efforts to make these into real definitions. One way is to
restrict the subsets U or regions R to be of a very special form (but if this is too restrictive, the
proofs of the integral theorems immediately become much harder). What is even less attractive
is the necessarily rather arbitrary character of what one calls admissible for U or R.
The alternative is to characterize quantities like the curvature k(p), curl and divergence in al-
gebraic terms, and show that the quantities so defined may be interpreted in the above way.
However, in all fairness, the algebraic formulas do not just fall out of the sky, but rather ought to
be inspired by the proposed preliminary definition given above. Historically one has informally
taken the above as a preliminary definition, and under whatever suitable niceness conditions
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 124

work out what a candidate algebraic definition ought to be. . . Where Gauss worked with in-
finitesimal surface elements, the modern machinery provides an elegant alternative that is at
the same time precise and leads to all algebraic formulas while at the same time being close to
the intuitive ideas. The key innovation is to consider of subsets of the tangent space Tp M and
their images instead of small neighborhoods U ⊆ M and their images ν(U ) ⊆ S 2 .
Using the Riemannian metric on M 2 induced from the Euclidean metric on R3 the area of the
parallelogram spanned by Xp and Yp is

hXp , Xp iM hXp , Yp iM
 
p p
area( {λXp + µYp ∈ Tp M : λ, µ ∈ [0, 1]}) = det (239)
hYp , Xp iM
p hYp , Yp iM
p

Similarly, using the Riemannian metric on S 2 induced from the Euclidean metric on R3 the area
of the parallelogram spanned by ν∗p (Xp ) and ν∗p (Yp ) is
2 2
!
hν∗p Xp , ν∗p Xp iSν(p) hν∗p Xp , ν∗p Yp iSν(p)
det 2 2 (240)
hν∗p Yp , ν∗p Xp iSν(p) hν∗p Yp , ν∗p Yp iSν(p)

To simplify notation we note that the determinants are nothing else than the volume forms
on the Riemannian manifolds M 2 and S 2 , with metrics inherited from their immersions into
2
R3 , and therefore may write dV M and dV S , respectively. To further simplify notation use the
2
pullback of dV S by ν ∗ and formally define:

Definition 4.13 Suppose : M 2 ⊆ R3 is a smooth surface. The Gaussian curvature at p ∈ M is


2
νp∗ (dV S )
k(p) = (241)
dVpM

Note that despite its strange appearance this division makes sense as every 2-from on M 2 is a
scalar multiple of the natural volume form dV M , and thus k ∈ C ∞ (M ).

Exercise 4.40 Verify that the Gaussian curvature is independent of the orientation on M 2 .

It is appropriate to take a short detour and obtain more explicit formulas for graphs of functions
in R3 . One key simplification is based on the observation that since the tangent plane Tp M is
parallel to the tangent plane Tν(p) S 2 the ratio of the areas of the respective parallelograms is
equal to the ratio of the areas of their respective projections onto the coordinate planes!
Hence if M 2 is the graph of a function g : R2 7→ R3 , a natural choice of two independent tangent
vectors at p = F (s, t) ∈ M 2 ⊆ R3 is Xp = F∗ (D1 ) = D1 + (D1 g)D3 and Yp = F∗ (D2 ) =
D2 + (D2 g)D3 . The area of the projection of the parallelogram spanned by these vectors onto
the (x1 , x2 ) coordinate plane is 1.
On the other hand
ν∗p Xp = ν∗ ◦ F∗ (D1 ) = (ν ◦F )∗ (D1 ) = (D1 (ν 1 ◦F ))D1 + (D1 (ν 1 ◦F ))D2 + (D1 (ν 1 ◦F ))D3
ν∗p Yp = ν∗ ◦F∗ (D2 ) = (ν ◦F )∗ (D2 ) = (D2 (ν 1 ◦F ))D1 + (D2 (ν 1 ◦F ))D2 + (D2 (ν 1 ◦F ))D3
(242)

Exercise 4.41 Identify where each of the expressions, functions and tangent vectors, in the
preceding calculations is to be evaluated.
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 125

The area of the projection of the parallelogram spanned by these tangent vectors to S 2 onto the
(x1 , x2 ) coordinate plane, and hence the Gaussian curvature at p is equal to

k(p) = D1 (ν 1 ◦F ) · D2 (ν 2 ◦F ) − D2 (ν 1 ◦F ) · D1 (ν 2 ◦F ) (243)

evaluated at F −1 (s, t). Finally, explicitly calculate the unit normal νp from Xp and Yp as

−(D1 g)D1 − (D2 g)D2 + D3


(νp ◦ F ) = 1 (244)
(1 + (D1 g)2 + (D2 g)2 ) 2

Exercise 4.42 Combine the expressions in the equations (243) (244) to derive by direct calcu-
lation (use a computer algebra system as needed):

(D12 g) · (D22 g) − (D1 D2 g)



k(p) = (245)
(1 + (D1 g)2 + (D2 g)2 )2 F −1 (p)

Briefly return to Euler’s theorem: In the first proof we chose the coordinates so that (D1 g), (D2 g),
and (D2 D1 g) all were zero at F −1 (p), and the formula (245) immediately yields that the Gaussian
curvature is the product of the extremal curvatures kmin and kmax of Euler’s theorem.

Exercise 4.43 Use the formula (245) to explicitly calculate the Gaussian curvatures of a plane
z = mx + ny, the paraboloid z = ax2 +√by 2 , the saddles z = x2 − y 2 and z = xy, and the graphs
of z(1 + x2 + y 2 ) = 1 and z = Re(x + −1y)3 .
In each case identify all regions where the curvature is positive, zero, or negative. [[Compare
with the visual images in the MAPLE worksheet gaussmap.mws.]]

Exercise 4.44 Use the expression (245) to prove that the curvature k(p) of the graph of every
polynomial converges to zero as kpk grows to infinity.

Exercise 4.45 Verify by direct calculation (computer algebra system) that the expression (245)
is invariant under rotations and reflections.

Exercise 4.46 Develop an explicit formula for the Gaussian curvature on a smooth parameter-
ized surface F : (s, t) 7→ (F 1 (s, t), F 2 (s, t), F 3 (s, t)) (not necessarily F 1 (s, t) = s, F 2 (s, t) = t).

Exercise 4.47 Verify by direct calculation (computer algebra system) that the expression ob-
tained in exercise (4.46) is invariant under rotations and reflections.

Exercise 4.48 Verify by direct calculation (computer algebra system) that the expression ob-
tained in exercise (4.46) yields constant curvature for the sphere. Also calculate the Gaussian
curvatures of a cone and cylinder using cylindrical coordinates, and of the torus F : (s, t) 7→
((R + r cos(s)) · cos(t), (R + r cos(s)) · sin(t), r sin(s)). In each case identify all regions where the
curvature is positive, zero, or negative. [[Utilize a computer algebra system as appropriate, and
compare with the visual images in the MAPLE worksheet gaussmap.mws.]]
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 126

Returning to the more theoretical discussion, we now make use of the prior identifications ,
considering the tangent map or differential of the Gauss map ν as the Weingarten map ILp =
(dν)p : Tp M 7→ Tp M . There is an elegant way to express some classical terminology in terms of
the Weingarten map, and thereby obtain new insights about the character of the Gauss curvature
as an magnification factor of infinitesimal area, or rather area in the tangent spaces.
Definition 4.14 Suppose M m ⊆ Rm+1is an oriented smooth manifold (with the induced metric).
Define the 1st , 2nd and 3rd fundamental forms I, II, III pointwise by Ip , IIp , IIIp : Tp M × Tp M 7→ R,

Ip (Xp , Yp ) = hXp , Yp i
IIp (Xp , Yp ) = hILp (Xp ) , Yp i (246)
IIIp (Xp , Yp ) = hIL2p (Xp ) , Yp i

In the case of surfaces M 2 in R3 the classical notation for the components of the fundamental
forms in a chart ((x, y), M 2 ) uses the letters F, G, H and `, m, n for the components of I and II:

I = F dx ⊗ dx + Gdx ⊗ dy + Gdy ⊗ dy + Hdy ⊗ dy


(247)
II = `dx ⊗ dx + mdx ⊗ dy + mdy ⊗ dy + ndy ⊗ dy.

Exercise 4.49 Use the precise definition in terms of dν to verify that the map II is bilinear
over C ∞ (M ). Demonstrate that I and II depend on the orientation on M (hence II/I does not).

It is straightforward to recognize an interpretation of the second fundamental form in terms of


the curvatures of the curves of intersection of M with normal planes as discussed earlier: If Π⊥
is the plane in R3 through p ∈ M that is spanned by νp and a unit vector Xp ∈ Tp M , and
σ : (−ε, ε) 7→ M ∩ Π⊥ is a smooth curve parameterized by arc length, with σ 0 (0) = Xp , then the
signed curvature (with the same orientation convention) of σ at t = 0 equals

κ(0) = hσ 00 (0) , νp i = IIσ(0) (σ 0 (0), σ 0 (0)) (248)

Before further analyzing the Weingarten map we carry out some basic technical constructions
which demonstrate that, locally, vector fields along a submanifold may be smoothly extended.
We check some basic properties of such extensions, which will allow for some elegant proof of
important properties of the Weingarten map.
Suppose ı : M m ,→ N n is an imbedded submanifold, and (u, U ), ū, Ū ) are charts (about p) of
M and N , respectively such that U = Ū ∩ M = {q ∈ Ū : um+1 (q) = . . . = un (q) = 0} and
ui = ūi |U for i = 1, 2, . . . , m. W.l.o.g. assume that ū(Ū ) = B0n (r) is an open ball in Rn so that
the map π : U 7→ U , π(q) = u−1 (u1 (q), u2 (q), . . . , um (q), 0, . . . 0)) is well-defined.
If X ∈ C ∞ (M, T N ) such that πT N ◦ X = ı is a smooth vector field along M , then it may be
smoothly extended to a vector field X̄ ∈ Γ∞ (Ū ) by defining

X̄q = m ∂
P i

i=1 (Xu )π(q) ∂ ūi q . (249)

Lemma 4.26 Suppose M, N, p, U, Ū are as above and ı : U 7→ Ū is the inclusion map. and
Z ∈ Γ∞ (U ) and Z̄ ∈ Γ∞ (Ū ) are tangent vector fields.
Then Z̄ is an extension of Z if and only if (Z̄ f¯)|U = Z(f¯|U ) for every Z̄ ∈ C ∞ (Ū ).
Moreover, if X̄ and Ȳ extend X, Y ∈ Γ∞ (U ) to Ū , then [X̄, Ȳ ] extends [X, Y ] to Ū .
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 127

Proof. If p ∈ U , Z̄p = ı∗ Zp , and f¯ ∈ C ∞ (Ū ), then (Z̄ f¯)(p) = Z̄p f¯ = Zp (f¯ ◦ ı) = (Z(f |U ))(p).
Conversely, if (Z̄ f¯)|U = Z(f¯|U ) then Z̄p f¯ = (Z̄ f¯)p = Zp (f |u ) = Zp (f¯ ◦ ı) = (ı∗ Zp )f¯.
Now suppose p ∈ U , f¯ ∈ C ∞ (Ū ) and that X̄, Ȳ extend X, Y ∈ Γ∞ (U ) to Ū . Then using the
first part, the following shows that [X̄, Ȳ ] extends [X, Y ].

[X̄, Ȳ ]p f¯ = X̄p (Ȳ f¯) − Ȳp (X̄ f¯) = Xp (Y (f¯|U )) − Yp (X(f¯|U )) = [X, Y ](f |U ). (250)

Proposition 4.27 Suppose M m ⊆ Rm+1 is a smooth oriented hypersurface and p ∈ M .


The Weingarten map ILp : Tp M 7→ Tp M is self-adjoint.

Proof. The preceding lemma allows for an elegant proof that uses the Euclidean connection ∇.
Suppose Xp , Yp ∈ Tp M . First extend to tangent vector fields X, Y in a neighborhood on p ∈ M ,
then to vector fields X̄, Ȳ on a neighborhood of ı(p) ∈ Rm+1 . Calculate

hX , IL(Y )ip − hIL(X) , Y ip = hX , ∇Y νip − h∇X ν , Y ip


= hX̄ , ∇Ȳ ν̄ip − h∇X̄ ν̄ , Ȳ ip
(251)
= − h∇Ȳ X̄ , ν̄ip + hν̄ , ∇X̄ Ȳ ip
= hν̄ , [X̄ , Ȳ ]ip = hνp , [X , Y ]p i = 0

using in the third step that X̄p hν̄ , Ȳ i = Ȳp hȲ , ν̄i = 0 since hν , Y i ≡ hX , νi ≡ 0.
In a nutshell the self-adjointness of the Weingarten map may be seen as an immediate con-
sequence of the compatibility of the Euclidean connections (directional derivatives) with the
Euclidean metric (standard inner product).

Note that Euler’s theorem about principal curvatures now becomes a simple corollary: As a
self-adjoint operator IL : Tp M 7→ Tp M has real eigenvalues and is orthogonally diagonalizable.
The principal curvatures and directions are mere consequences of elementary linear algebra.

Exercise 4.50 Fill in the details to derive Euler’s theorem as a corollary of IL being self-adjoint.

Theorem 4.28 The Gaussian curvature k(p) at p ∈ M 2 ⊆ R3 is the determinant of the Wein-
garten map IL : Tp M 7→ Tp M .
k(p) = det ILp (252)

Corollary 4.29 Suppose p ∈ M 2 . If Xp , Yp ∈ Tp M are linearly independent or orthonormal,


then the Gaussian curvature at p is, respectively equal to
 
IIp (Xp , Xp ) IIp (Xp , Yp )
det
IIp (Yp , Yp ) IIp (Yp , Yp )
 
IIp (Xp , Yp ) IIp (Xp , Yp )
k(p) =  and k(p) = det (253)
IIp (Yp , Yp ) IIp (Yp , Yp )

Ip (Xp , Xp ) Ip (Xp , Yp )
det
Ip (Yp , Yp ) Ip (Yp , Yp )

Lemma 4.30 Suppose that M 2 ⊆ R3 is an immersed manifold with the metric inherited from
R3 . Then in each coordinate chart (u, U ) of M 2 the Gaussian curvature may be written as a
∂ ∂
smooth function of the components gij = h ∂u i , ∂uj i of the Riemannian metric and its first and
∂gij ∂ 2 gij
second derivatives uk
and ∂u` ∂uk
.
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 128

Proof (outline). This outline follows Opreah, but more highlighting the role of the connection,
and with details of the calculations provided in the MAPLE worksheet egeregium.mws. (For a
different approach using determinants and vector cross products see Spivak II, pp.108–111.)
Suppose (u, U ) is a chart of M with p ∈ U . Let ν : U 7→ R3 be the unit normal vector field along
U that provides the compatibility of the orientation between U and R3 . Then the Euclidean
connection ∇ defines the coefficients in:

After reorganization: How much of Euclidean connection ∇ need to be defined before?



∇∂ ∂u1
= Γ111 ∂u∂ 1 + Γ211 ∂u∂ 2 + `ν
∂u1

∇∂ ∂u1
= Γ112 ∂u∂ 1 + Γ212 ∂u∂ 2 + mν (254)
∂u2

∇∂ ∂u2
= Γ121 ∂u∂ 1 + Γ222 ∂u∂ 2 + nν
∂u2

using the symmetry of ∇ as a Riemannian (or Levi-Civita) connection:

Γk12 = Γk21 since ∇ ∂ ∂


∂u1
− ∇∂ ∂
∂u1
= [ ∂u∂ 1 , ∂u∂ 2 ] ≡ 0. (255)
∂u2 ∂u2

In this basis the Weingarten map L : Tp M 7→ Tp M may be expanded in the form

IL( ∂u∂ 1 ) = ∇ ∂ ν = w11 ∂ ∂ 1 ∂ ∂



∂u1
+ w12 ∂u2
+0ν = EG−F 2
· ` ∂u1
+m ∂u2
∂u1
(256)
IL( ∂u∂ 2 ) = ∇ ∂ ν = w21 ∂ ∂ 1 ∂ ∂

∂u1
+ w22 ∂u2
+0ν = EG−F 2
· m ∂u1
+n ∂u2
∂u2

Each of these identities may be differentiated – in particular, the equality of the mixed partial
derivative (this is the Euclidean connection!) provides two different expressions from equating
the derivatives    
∂ ∂
∇∂ ∇∂ ∂u 1 = ∇∂ ∇∂ ∂u 1 (257)
∂u2 ∂u 2 ∂u2 ∂u2
As shown in the MAPLE worksheet egregium.mws, it is straightforward to differentiate the
corresponding right hand sides, re-using the identities (254) and (256) to rewrite all derivatives
of the vector fields ∂u∂ 1 , ∂u∂ 2 and ν again as linear combinations of these vector fields. The crucial
feature is that while now also derivatives of the coefficients `, m and n appear, these only appear
in the coefficients of ν. Consequently, one obtains two equations from equating the coefficients
of each of ∂u∂ 1 , ∂u∂ 2 in (257) to zero. The surprising feature is that each of these will yield an
equation that only depends smoothly on the combination (`n − m2 ), but otherwise is indepen-
dent of `, m and n. This allows one to solve for (`n − m2 ) in terms of E, F, and G and their
first and second order derivatives (in the directions ∂u∂ 1 and ∂u∂ 2 ). (Alternatively, these may also
be written in terms of the gij and their first and second order derivatives, or in terms of the
Christoffel-symbols Γkij and their first derivatives (in the directions ∂u∂ 1 and ∂u∂ 2 ).

The resulting formulas look unappealing (actually they are a preview of the components of
the Riemann curvature tensor!). But what matters is that these formulas express the Gaussian
curvature entirely in terms of the intrinsic metric and its (covariant) derivatives (on the manifold
M ). This proves that even though the Gaussian curvature was defined in terms of the normal
vector field ν, it is independent of the imbedding into the ambient space:
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 129

Theorem 4.31 (Gauss, 1827: “Theorema Egregium”) If Φ, Ψ : M 2 ⊆ R3 are immersions


such that Φ∗ h· , ·i = Ψ∗ h· , ·i then for all p ∈ M the Gaussian curvatures of Φ(M ) and Ψ(M ) at
Φ(p) and Ψ(p), respectively, are equal.

A few remarks about geodesic triangles and the Gauss Bonnet theorem. With the tools
developed so far it is basically a just a lengthy exercise, involving some trickier calculations, to
connect the total curvature (the integral of the Gauss curvature) to the angle sum in a triangle:

Theorem 4.32 Suppose M 2 is a Riemannian manifold, and U ⊆ U 0 ⊆ M are neighborhoods


of p ∈ M such that every pair of points in U is joined by a unique geodesic that lies entirely
in U 0 , and such that U is contained in the image of expp , compare theorem (4.18). Suppose
q, q 0 ∈ U and R ⊆ U is a geodesic triangle whose sides are the geodesics connecting the three
points p, q, q 0 , respectively. If the interior angles are αi , i = 1, 2, 3, then
Z
k dV = α1 + α2 + α3 − π. (258)
R

In the standard proof one uses the map expp to pull back the volume form (area form) dV
from U to Tp M , thereby reducing the integration to an integral over a subset of the linear space
Tp M . Note that the preimages under expp of the geodesics connecting p to q and q 0 , respectively,
are straight rays emanating from 0p ∈ Tp M . Thus the integral is particularly easy to evaluate
using polar coordinates (basically these are the pullbacks of the geodesic coordinates on U to
Tp M ). Via a simple application of Green’s theorem it reduces to a single integral involving the
direction of the velocity vector of the preimage of the geodesic connecting q to q 0 as a function
of the angular variable.
So far everything beyond the pullback via expp is basic calculus. What remains is to explicitly
calculate the pullback of k in the geodesic, or polar coordinates. Thus write (r, θ) for the
standard polar coordinates on Tp M and u = (r, θ) ◦ exp−1 for are the associated geodesic
coordinates on U ⊆ M . By definition of the geodesics it is clear that in these coordinates
g11 = E = h ∂u∂ 1 , ∂u∂ 1 i = 1. Gauss’ lemma (4.19) implies that g12 = g21 = F = 0. Consequently
in these coordinates the metric has the form h· , ·i = du1 ⊗ du1 + G · du2 ⊗ du2 for some smooth
positive function G : U 7→ (0, ∞).
2
Exercise 4.51 For the metric as given above verify that k = − √1G ∂u∂1 ∂u
G
1.

On the other hand one needs properties of the geodesics connecting q to q 0 . It is straightforward
to write down the geodesic equation given the data for gij as above.

Exercise 4.52 Given the metric in coordinates as above, calculate the Christoffel symbols Γkij
of the second kind and write out the geodesic equations in these geodesic coordinates. [[The
coefficients will be in terms of derivatives of the given function G = g22 ]].

To complete the proof it remains to find a convenient expression for the integral along the third
side, the geodesic connecting q to q 0 . It appears most convenient to rewrite (reinterpret) the
geodesic equation to obtain a differential equation for the rate of change of the angle between
its velocity vector and either of the coordinate directions. There are several technical details
involved and thus leave it as an extended project:
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 130

Project 4.53 Fill in the missing details into the outline of the proof of the theorem. As reference
consult e.g. Spivak II.pp.114-124.
in progress

Definition 4.15 A triangulation of what ?? is union... intersections are edges or vertices...

in progress

The final highlight of this classical development is without doubt:

Theorem 4.33 (Gauss-Bonnet) : Suppose M 2 is a compact, orientable Riemannian mani-


fold, k is the Gaussian curvature, and χE (M ) is the Euler characteristic of M . Then
Z
k · dV = 2π · χE (M ) (259)
M

The Euler characteristic is a fundamental topological invariant. In the case that M is a tri-
angulated C 0 manifold (think of a polyhedron, and subdivide the faces if necessary) the Euler
characteristic is equal to the alternating sum χE (M ) = V − E + F of the number of vertices,
edges, and faces.

Exercise 4.54 Calculate the Euler characteristic of the regular polyhedra (tetrahedron, cube,
octahedron, dodecahedron, and icosahedron). Are they the same (as topologically manifolds), i.e.
are they homeomorphic? Are they homeomorphic to the sphere?
Given that the Euler characteristic is a topological invariant calculate the Euler characteristics
of a torus, of a double-torus (figure eight shaped tube), and a bretzel-like surface with three holes.
[[Construct, simply draw, a polyhedron that is homeomorphic, e.g. start with a solid rectangular
block and remove one, two, or three appropriate rectangular blocks. Count the faces, edges, and
vertices of the surface of the solid.]] Does it matter for this count whether the faces are allowed
to have any polygonal shape or whether they must first be fully triangulated?

The Euler characteristics belongs into the realm of algebraic topology – the desire for a solid
proof may be taken as a good motivation to develop the tools of cohomologies. Arguably the
closest differential geometric analogue is the de-Rham cohomology which starts with an algebraic
description of in how many ways can closed differential forms on the manifold fail to be exact
(technically these are the dimensions of the quotient spaces of closed differential k-forms by
the exact differential k-forms. [[Look under Betty-numbers in most any advanced book on the
interfaces of differential and algebraic, geometry and topology.]]

What makes the Gauss Bonnet theorem so intriguing that on one side is the integral of an
important geometric invariant, but on the other side is a purely topological invariant. This has
very profound consequences: E.g. given that χE (S 2 ) = 2 it follows immediately that the total
curvature of the sphere is 4π no matter what Riemannian metric S 2 (or, for that, any other
compact simply connected 2-manifold) is equipped with! Similarly, if the Gaussian curvature is
negative at all points on a compact manifold M 2 then M cannot be simply connected.

A complete proof of the Gauss Bonnet theorem requires the development of a substantial machin-
ery of algebraic topology. But if one takes the topological invariance of the Euler characteristic
as a given, then modulo one technical detail the proof is amazingly simple:
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 131

Exercise 4.55 Suppose that M 2 is a Riemannian manifold that is triangulated into a finite
number of geodesic triangles. Use the theorem (4.32) to derive the Gauss-Bonnet theorem for
this special case. [[Sum the left and right hand sides of the identity in theorem (4.32) over all
triangles and note that at each vertex all the interior angles sum to 2π. However, in a case
of general polygons, and also for the generalization to Not necessarily geodesic triangles it is
better to replace the interior αi angles with the change of direction (∆σ 0 )i = limt→ti + σ 0 (t) −
limt→ti − σ 0 (t) = π − αi . Note that this angle only involves the inner product in Tσ(ti ) M 2 . ]]

Thus aside form the topological background (that χE is indeed an invariant) the main missing
part is the existence of a (finite!) triangulation by geodesic triangles. The general result that
every C ∞ -manifold has a triangulation is a difficult theorem (cf. Spivak I. p.579, for a proof
see e.g. Munkres, Elementary Differential Topology). In our case the compactness hypothesis
makes things much easier, but still nontrivial.
On the other side it is relatively straightforward to weaken the demand that the sides of the
triangles are geodesics – in that case one merely has to account for the change of direction of
the velocity vector along the edges – this is amounts to integrating the geodesic curvature of the
edges. This may be done using classical formulations (calculus) in Euclidean spaces, but modern
approach in terms of connections, covariant derivatives along the edges and parallel transports
appears much more appealing.

Project 4.56 Work out an outline to fill in the steps to the more general version of the Gauss-
Bonnet formula given in Schaum’s Outline (see below) and the modern approach in terms of
connections.
Suppose C is a C 2 curvi-linear polygon and a C 3 manifold, which is positively
oriented and its interior is simply connected. Then
Z ZZ X
κg ds + kdV = 2π − αi (260)
R i

where κg is the geodesic curvature along C, R is the union of C with its inte-
-rior [[??]], k is the Gaussian curvature, and αi are the exterior angles of C.
There are many closely related theorems, some almost immediate corollaries of the Gauss Bonnet
theorem. We just mention a few.

Theorem 4.34 (Puiseux, 1848) (compare Spivak II, p.129) Let C(r) be the circumference of
the geodesic sphere (circle) with radius r about p. Then

πr2 − A(r)
k(p) = lim 12 · (261)
r→0 πr4
Theorem 4.35 (Diquet, 1848) (comp. Spivak II, p.131) Let A(r) be the area of the region
inside the geodesic sphere (circle) with radius r about p. Then

2πr − C(r)
k(p) = lim 3 · (262)
r→0 πr3
The following two result from relating the Gaussian curvature to the minimality of geodesics.

Theorem 4.36 (Bonnet) (comp. Sternberg p.292) If M 2 is a connected Riemannian manifold


with k ≥ 1/C 2 , then for any two points p, q ∈ M the distance dR (p, q) is at most πC.
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 132

Theorem 4.37 (compare Sternberg p.292) If M 2 is a connected Riemannian manifold such


that k ≤ 0, then no geodesic on M has any conjugate points, and if M is simply connected then
M is diffeomorphic to the plane R2 .

(Here a conjugate point along a geodesic is defined as the first point along the geodesic where
the tangent map of the map exp : Tp M 7→ M becomes singular.)
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 133

4.6 Riemann curvature tensor


Since Gauss the concept of curvature has been generalized to apply to general manifolds, not
necessarily hypersurfaces or Riemannian. The most prominent notion is due to Riemann who
in 1861 asked the simple question – stated here in modern language
Riemann’s question: When is a Riemannian manifold flat, i.e. locally isometric to Rm ? Find
conditions for the existence (and an algorithm for the construction) of local coordinates (u, U )
∂ ∂
such that in these coordinates gij = h ∂u i , ∂uj i = δij .

The following exposition loosely follows Spivak’s discussion of what basically was Riemann’s
original approach. However, we shall use modern terminology throughout, and sprinkle in some
reference to connections which eventually will suggest a perfectly elegant definition in terms of
Koszul connections.

Thus suppose (M m , h· , ·i) is a Riemannian manifold, and (v, V ) is a chart about a point p ∈ M .
Let gij = h ∂v∂ i , ∂v∂ j i denote the components of the metric in these given coordinates. The
objective is to find conditions for the existence of a local coordinate change u◦−1 that transforms
∂ ∂
the given gij into h ∂u i , ∂uj i = δij . The basic strategy is to derive a set of partial differential
equations for the coordinate change, and massage them so that eventually we may apply standard
integrability theorems (some version of Frobenius’ theorem).

The transformation formulas for the components of the metric provide the basic conditions.
Write them out in both directions. Note the ranges of the summations.
m
∂ ∂
X ∂v k ∂v `
δij = h ∂u i , ∂uj
i = · · gk` (263)
∂ui ∂uj
k,`=1

m
X ∂uk ∂uk
gij = h ∂v∂ i , ∂
∂v j
i = · (264)
∂v i ∂v j
k=1
1
Interpret these as a set of 2 m(m+ 1) quadratic, first order inhomogeneous partial differential
equations for m functions uk : U⊆ V 7→ Rm . Basically the only operation one can always
use is to differentiate [[From the late Robert Gardener I learnt the phrase “apply the Pavlov-
operator”, i.e. take the exterior derivative.]] Here we take partial derivatives in the directions
of the coordinate vector fields (clearly, working with d comes to mind as a more algebraic
alternative. In order to obtain familiar formulas that eventually will lead to linear partial
differential equations, add and subtract the following, carefully noting the cancelations.
m
∂ 2 uk ∂uk ∂uk ∂ 2 uk
 
∂gi` X
(+) = · + ·
∂v j ∂v j ∂v i ∂v ` ∂v i ∂v j ∂v `
k=1

m
∂ 2 uk ∂uk ∂uk ∂ 2 uk
 
∂gj` X
(+) = · + ·
∂v i ∂v i ∂v j ∂v ` ∂v j ∂v i ∂v `
k=1
m
∂ 2 uk ∂uk ∂uk ∂ 2 uk
 
∂gij X
(−) = · + ·
∂v ` ∂v ` ∂v i ∂v j ∂v i ∂v ` ∂v j
k=1
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 134

m
∂ 2 uk ∂uk
 
1 ∂gj` ∂gi` ∂gij X
[i j , ` ] = i
+ j
− = · (265)
2 ∂v ∂v ∂v ` ∂v i ∂v j ∂v `
k=1
k
The plan is to use the identity (264) in order to move the first derivative ∂u
∂v `
to the left side,
ij
noting that a product of two such terms yields the g and thus raises the indices, leading to
Christoffel symbols of the second kind on the left hand side. Specifically use:
m m
X X ∂v ` ∂v k
Γkij = [ij ,`] · g k`
= [ij ,`] · · (266)
∂ur ∂ur
`=1 `, r=1

∂us
Multiply both sides by ∂v k
and sum over k. Reuse the preliminary result (265), and contract
the sums as usual:
m m
X ∂us X ∂v ` ∂v k ∂us
Γkij · = [ij ,`] · · ·
∂v k ∂ur ∂ur ∂v k
k=1 k, `, r=1
m
X ∂v `
= [ij ,`] ·
∂us
`=1
m
X ∂ 2 uk ∂uk ∂v `
= · · (267)
∂v i ∂v j ∂v ` ∂us
k, `=1

∂ 2 uk
=
∂v i ∂v j
Consider these as a set of m3 first order explicit (i.e. solved for the highest derivative) linear
∂ui
homogeneous partial differential equations for m2 unknown functions ∂v j , i.e. think of them as

 
∂uk m
∂ ∂v i X ∂uk
= Γ`ij · (268)
∂v j ∂v `
`=1

Note that these are actually the same partial differential equations for each index k – the
eventual objective is to find (conditions for the existence of) m linearly independent solutions uk
∂ ∂

satisfying e.g as boundary conditions uk (p) = 0 and ∂uk p = ∂vk p . The Frobenius integrability
test requires the equality of mixed partial derivatives, i.e. for all i, j, k, s = 1, . . . , m
 k  
∂ 2 ∂u ∂ 2 ∂uk m m
∂uk ∂uk
   
∂v i
! ∂v i X ∂ ` !
X ∂ `
= i.e. Γij · = Γis · (269)
∂v s ∂v j ∂v j ∂v s ∂v s ∂v ` ∂v j ∂v `
`=1 `=1

Differentiate the products and reuse the partial differential equations (268) to express every
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 135

second order derivative of the uk as a linear combination of first order derivatives.


m
!
∂Γ ` k ` k 2 uk 2 uk
!
X ij ∂u ∂Γ ∂u ∂ ∂
0 = · − is
· + Γ`ij · s ` − Γ`is · j `
∂v s ∂v ` ∂v j ∂v ` ∂v ∂v ∂v ∂v
`=1
m m m
!
X ∂Γ`ij ∂uk ∂Γ`is ∂uk `
X
r ∂uk `
X
r ∂uk
= · − · + Γij · Γ s` · − Γis · Γj` ·
∂v s ∂v ` ∂v j ∂v ` ∂v r ∂v r
`=1 r=1 r=1
m m 
!
X ∂Γ`ij ∂Γ`is ∂uk X
` r ` r
 ∂uk
= − · + Γ ij · Γ s` − Γ is · Γj` · (270)
∂v s ∂v j ∂v ` ∂v r
`=1 `,r=1
m m
!
X ∂Γ`ij ∂Γ`is X
r ` r ` ∂uk
= − + Γij · Γ sr − Γ is · Γ jr · (271)
∂v s ∂v j ∂v `
`=1 r=1
| {z }
`
Rsij

So much for an exposition of the historical origin of the Riemann curvature tensor . . .

Theorem 4.38 (Riemann, 1861) A Riemannian manifold (M m , h· , ·i) is flat if and only if
i
in every coordinate chart if Rjk` ≡ 0 for all i, j, k, ` = 1, . . . m,

The prior calculations essentially establish the necessity. For sufficiency, including a construction
of the new coordinates uk see e.g. Spivak II, pp.210 – 205 (referred to as the “test case”).

 
i 3
Theorem 4.39 The functions Rjk` as above are the components of a tensor of type .
1

A purely mechanical proof checks that these functions transform as required, i.e. suppose that
(u, U ) and (v, V ) are coordinate charts with functions Γkij , Rjk`
i and Γ̃trs , R̃stq
r , respectively. Then

on U ∩ V it is to be verified that for r, s, t, q = 1, . . . m


m
r
X
i ∂v r ∂uj ∂uk ∂u`
R̃stq = Rjk` · · · (272)
∂ui ∂v s ∂v t ∂v q
i,j,k,`=1

This is a straightforward calculation that uses the transformation formula for the Christoffel
symbols of the second kind derived earlier
 
m t i m
X
Γkij ∂v · ∂u ∂uj X 2
∂ u q ∂v t
Γ̃trs = · s + · q (273)
∂uk ∂v r ∂v r s
∂v ∂v ∂u
i,j,k=1 q=1

Thus the Riemann curvature tensor is a map

R : Γ∞ (M ) × Γ∞ (M ) × Γ∞ (M ) 7→ Γ∞ (M ) (274)
i
and in a chart (u, U ) there are functions Rjk` ∈ C ∞ (U ) such that
m
X
i
R( ∂u∂ k , ∂u
∂ ∂
` ) ∂uj = Rjk` ∂
∂ui
(275)
i
Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 136

 
4
For convenience one also introduces a tensor of type , as the map
0

R : Γ∞ (M ) × Γ∞ (M ) × Γ∞ (M ) × Γ∞ (M ) 7→ C ∞ (M ) (276)

and in a chart (u, U )


m
X
i
Rijk` = gis Rjk` ∂
= hR( ∂u∂ k , ∂u ∂
` ) ∂uj ,

∂ui
i (277)
s=1

In the case of a two dimensional manifold M 2 on recovers the Gaussian curvature as


hR(Xp , Yp )Yp , Xp i
k(p) = (278)
kXp , Yp k2

where Xp , Yp ∈ Tp M are linearly independent and kXp , Yp k is the area of the∂parallelogram in




Tp M spanned by Xp and Yp . In the special case that Xp = ∂u1 p are Yp = ∂u2 p orthonormal

one obtains
k(p) = R1221 = hR( ∂u∂ 1 , ∂u∂ 2 ) ∂u∂ 2 , ∂u∂ 1 i (279)

i
Exercise 4.57 Verify the assertion that k = R1221 by direct calculation. [[Write Rjk` in terms
k
of Γij , then in terms of gij or E, F, G and their derivatives.]]

Exercise 4.58 Calculate the components of the Riemann curvature tensor for the special case
that M is the graph of a function z = f (x, y) coordinatized by u = (x, y). [[ Use a computer
algebra system – first do it yourself, then compare the results with the tensor-package.]]

Exercise 4.59 Use the results of the previous exercise to calculate the components of the Rie-
mann curvature tensor for familiar graphs such as a plane, a paraboloid, a saddle, the graph
of (x, y) 7→ 1+x12 +y2 . [[ Use a computer algebra system – first do it yourself, then compare the
results with the tensor-package.]]

Exercise 4.60 Calculate the components of the Riemann curvature tensor for the special case
that M is a parameterized surface (x, y, z) = F (s, t) coordinatized by u = (s, t) [[ Use a computer
algebra system – first do it yourself, then compare the results with the tensor-package.]]

Exercise 4.61 Use the results of the previous exercise to calculate the components of the Rie-
mann curvature tensor for familiar surfaces such as spheres and tori of different radii. [[ Use a
computer algebra system – first do it yourself, then compare the results with the tensor-package.]]

Exercise 4.62 Calculate the components of the Riemann curvature tensor for familiar sur-
faces such a plane, a cone, a cylinder, and a paraboloid using polar / cylindrical coordinates.
Especially, check that in these coordinates still R = 0 for the cone and plane.

The following symmetries of Riemann curvature tensor are very useful.


Classnotes: Geometry & Control of Dynamical Systems, M. Kawski. April 20, 2009 137

Proposition 4.40 The Riemann curvature tensor is antisymmetric in the first two, and in the
last two entries, respectively, but it is symmetric in pairs and satisfies an analogue of the Jacobi
identity, called the Bianchi identity. Specifically, for X, Y, Z, W ∈ Γ∞ (M )

R(X, Y )Z = − R(Y, X)Z (280)

hR(X, Y )Z , W i = − hR(X, Y )W , Zi (281)

hR(X, Y )Z , W i = − hR(W, Z)X , Y i (282)

0 = R(X, Y )Z + R(Y, Z)X + R(Z, X)Y. (283)

Exercise 4.63 Prove the assertions made in the proposition.

While we found that in a chart on a two-dimensional manifold the component R1221 recovers
the Gaussian curvature, it is far from easy to understand the geometric roles of all the other
components. The following quite surprising property helps substantially:

Proposition 4.41 The Riemannian curvature tensor is completely determined by the values,
called sectional curvatures, of the quadratic forms

(Xp , Yp ) 7→ hR(Xp , Yp ), Yp , Xp i (284)

Exercise 4.64 Prove the proposition: Suppose that R, S : V × V × V × V 7→ R are multi-linear


maps satisfying the identities of proposition (4.40) – suitably rewritten. Show that then R = S.

The next section will obtain the following more appealing formula for the Riemann curvature
tensor:
R(X, Y )Z = ∇X ( ∇Y Z ) − ∇Y ( ∇X Z ) − ∇[X,Y ] Z (285)

You might also like