MA 2110, Introduction to Manifolds

Semester 2, 2014-2015
Solutions to selected homework problems

March 23, 2015

1. Let G : U → V and F : V → W be maps between open subsets of Euclidean spaces, let p ∈ U be

a point, and assume that the derivatives Dp G and DG(p) F exist. Prove the coordinate-free Chain
Rule: The derivative of F ◦ G at the point p exists and is given by a composition of linear maps:

Dp (F ◦ G) = DG(p) F ◦ Dp G.

Solution: Write L = DG(p) F and M = Dp G, and denote the error terms by

eF (y) = F (G(p) + y) − F (G(p)) − L(y)

and
eG (x) = G(p + x) − G(p) − M (x).
We know that for every  > 0 we have

|eG (x)| ≤ |x|

for all sufficiently small |x|. (Note for future reference that we also know, because of this, that
limx→0 eG (x) = 0.) We also know that for every  > 0 we have

|eF (y)| ≤ |y|

for sufficiently small |y|. We have to make a similar estimate for the error term

eF ◦G (x) = F (G(p + x)) − F (G(p)) − L(M (x)).

If we write
y = G(p + x) − G(p) = M (x) + eG (x),
then
eF ◦G (x) = eF (y) + L(eG (x)).

1
To see that |L(eG (x)| ≤ /2 for small x, note that for the linear map L there is some C > 0
such that for all x we have |L(x)| ≤ C|x|; now if x is small enough then |eG (x)| ≤ /2C so that
|L(eG (x))| ≤ /2.

To see that |eF (y)| ≤ /2 for small x we use the fact that this is so for small enough y, together
with the fact that limx→0 y = limx→0 M (x) + limx→0 eG (x) = 0.

False solution: A number of students made the mistake of dividing by something that could be
zero. This mistake is possible even in trying to prove the Chain Rule for functions of one variable.
In that context it goes like this: Write
f (g(x)) − f (g(a)) f (g(x)) − f (g(a)) g(x) − g(a)
= × .
x−a g(x) − g(a) x−a

Now let x approach 0; this becomes f 0 (g(a)) × g 0 (a). The problem is that there could be values of
x different from a, maybe arbitrarily close to a, such that g(x) = g(a).

3. For which values of R can we say that the solution set of the pair of equations

x2 + y 2 = 1

x2 + z 2 = R 2
is a one-dimensional smooth manifold in R3 ? Sketch the set for several illustrative values of R.

Solution: Probably the easiest way is to differentiate the map R3 → R2 given by

(x, y, z) 7→ (x2 + y 2 − 1, x2 + z 2 − R2 ).

This leads to a two by three matrix. The three two by two determinants are −4xy, 4xz, and 4yz.
When does the matrix have rank < 2? If and only if xy = xz = yz = 0, so if and only if at least
two out of the three numbers x, y, and z are zero. In the presence of the conditions x2 + y 2 = 1
and x2 + z 2 = R2 , we cannot have x = y = 0. We can have x = z = 0 if and only if R = 0. We
can have y = z = 0 if and only if R2 = 1. Thus with the possible exception of R = 0, R = +1,
and R = −1 the solution set will be a smooth 1-dimensional submanifold of R3 . When R = 0 the
set is just two points. When R = ±1 it is the union of two ellipses (one in each plane z = ±y)
intersecting in two points.

Alternatively one can solve the problem by explicitly parametrizing portions of the solution set, for
example solving for y and z as smooth functions of x,
p
y = ± 1 − x2
p
z = ± R 2 − x2 .
However, you have to make sure to get all the points. For example, when R > 1 the equations above
√ be given by −1 < x < 1)
(with all possible choices of signs, and in each case letting the domain
take care of all points except those where x = ±1, y = 0, z = ± R2 − 1. To parametrize a
neighborhood of such a point you can write
p
x = ± 1 − y2

2
 p
z = ± R2 − 1 + y 2 .
√
Or of course there are other ways, such as x = cosθ, y = sinθ, z = ± R2 − cos2 θ.

I omit the similar discussion of the case 0 < |R| < 1.

4. (a) Prove that a bijection h : R → R is a homeomorphism if and only if it is either order-

preserving or order-reversing.
(b) Suppose that M is a topological 1-manifold covered by two charts (U, x) and (V, y), such that
x(U ) = R = y(V ), and such that x(U ∩ V ) = y(U ∩ V ) = R+ , the set of all positive real numbers.
Let h : R+ → R+ be the homeomorphism given by h(x(p)) = y(p) for all p ∈ U ∩ V . Show that h
must be order-reversing (using the fact that a manifold is required to be a Hausdorff space).
(c) Show that M is homeomorphic to R.
(d) Now assume that h is a diffeomorphism, and note that these two charts then make M a smooth
manifold. Show that M is diffeomorphic to R.

Solution:

For one direction of (a), note that if h is order-preserving or order-reversing then it takes open
intervals to open intervals. Since every open set is a union of open intervals, it therefore takes open
sets to open sets.

For the other direction, suppose that h is a homeomorphism. Then it takes connected sets to
connected sets, and we can use this to show that it preserves the “betweenness” relation. If b is
between a and c then every connected set of real numbers that contains a and c must contain b,
so every connected set of real numbers that contains h(a) and h(c) must contain h(b), so h(b) is
between h(a) and h(c). Now we show that a bijection h which preserves betweenness must either
preserve or reverse the order. There are two kinds of pairs a < b: those for which the order is
preserved (h(a) < h(b)) and those for which the order is reversed (h(a) > h(b)). For any a < b < c
all three of the pairs a < b, a < c, and b < c are of the same kind; this follows from the fact that
h preserves betweenness. Thus two pairs that have an element in common are always of the same
kind. But of course if two pairs do not have an element in common then there is a pair that has
an element in common with both of them. Therefore all pairs are of the same kind.

For (b), note that by (a) the homeomorphism h : R+ → R+ must be either order-preserving or
order-reversing. Since M is a Hausdorff space, the points x−1 (0) ∈ U and y −1 (0) ∈ V have disjoint
neighborhoods in X, say A and B. The set x(A ∩ U ) is a neighborhood of 0 in R, so it contains
the interval (0, δ) for some δ > 0. Likewise y(B ∩ V ) contains an interval (0, ). Since A is disjoint
from B, h(0, δ) must be disjoint from (0, ). This cannot be the case if h is order-preserving, for in
that case h(0, δ) will be (0, δ 0 ) for some δ 0 > 0.

For (c) let us make a homeomorphism f : M → R by making homeomorphisms fU : U → (−∞, +1)

and fV : V → (−1, +∞) in such a way that both fU (U ∩ V ) and fV (U ∩ V ) are equal to (−1, +1)
and both fU and fV give the same homeomorphism U ∩ V → (−1, +1). We may choose fU to be,
say, g ◦ x, where g : R → (−∞, +1) is an order-preserving homeomorphism chosen in such a way
that g(0) = −1. Then g gives a homeomorphism from R+ to (−1, +1). Compose this with h−1
to get an order-reversing homeomorphism from R+ to (−1, +1). Now we need to extend this to
an order-reversing homeomorphism k : R → (−1, +∞). By (a) it is enough if we extend it to an

3
order-reversing bijection. So for example we can finish the job by defining k(t) = 1 − t for t ≤ 0.
The composition k ◦ y is a homeomorphism fV of the desired kind; in fact, for p ∈ U ∩ V we have

fV (p) = k(y(p)) = g(h−1 (y(p))) = g(x(p)) = fU (p).

For (d) we have to work a little harder. If the homeomorphism h is a diffeomorphism then in
the construction of (c) above even if we choose g to be a diffeomorphism (so that fU will be a
diffeomorphism and k is smooth in R+ ), there is no reason why f should be smooth at the point
y −1 (0). The difficulty is that when k is extended to all of R it might not be smooth at 0. Let
us modify the method by choosing fU more carefully. If we plan to finish by once again defining
k(t) = 1 − t for t ≤ 0, then we might try to choose g : R+ → (−1, +1) in such a way that when t is
small and positive k(t): = g(h−1 (t)) = 1−t. So we just need to be able to choose an order-preserving
diffeomorphism g : R+ → (−1, +1) such that for t >> 0 it is given by g(s) = 1 − h(s).

Here is one way to achieve that:

Lemma: Let H be an order-preserving diffeomorphism from a right half-line to a right half-line.

Then there exists an order-preserving diffeomorphism G : R → R such that G coincides with H in
a neighborhood of ∞.

Proof: Choose a < b in the domain of H. Use a smooth function φ : R → R such that φ(u) = 0
for u ≤ a, 0 ≤ φ(u) ≤ 1 for all u, and φ(u) = 1 for u ≥ b. Define Φ(u) = 1 − φ(u) + φ(u)H 0 (u)
for u in the domain of H and otherwise Φ(u) = 0. This is smooth because it is 1 for u < a. It
equals H 0 (u) for u > b. It is positive everywhere. Let G be an antiderivative of Φ and, by adding
a suitable constant to it, make it coincide with H for u > b. Then G is a smooth function with
positive derivative, tending to +∞ at +∞ and tending to −∞ at −∞.

Now apply the lemma by choosing order-preserving diffeomorphisms α : R → R+ and β : R →

(−1, +1), defining H by H(u) = β −1 (1 − α(u)), and using the resulting G to define g(s) =
β(G(α−1 (s)).

6. Verify that at any point (x, y, z) ∈ S 2 the vectors (−y, x, 0) and (xz, yz, z 2 − 1) are tangent to
x
S 2 . Write these same tangent vectors down in terms of the (stereographic) coordinates u = 1+z ,
y 2 ∂ ∂
v = 1+z defined in S − {(0, 0, −1)}. The answers should have the form a ∂u + b ∂v with a and b
being functions of u and v.

Solution: Of course the fact that they are tangent to sphere is just the fact that they are both
orthogonal to the vector (x, y, z). (The sphere is a regular level set of the function f = x2 + y 2 + z 2 ,
whose gradient is 2(x, y, z). Or, in the spirit of this course, we can speak of evaluating df on the
vectors, rather than making the inner product with the gradient.)

To write the vectors (−y, x, 0) and (xz, yz, z 2 − 1) in terms of the u, v coordinates, we can use the
∂ ∂ ∂ ∂ ∂
chain rule. First write the vectors as −y ∂x + x ∂y and xz ∂x + yz ∂y + (z 2 − 1) ∂z Now rewrite the
∂ ∂ ∂
functions x, y, and z as well as the vector fields ∂x , ∂y , and ∂z in terms of u and v. For the former
we recall the equations
2u
x=
1 + u2 + v 2

4
 2v
y=
1 + u2 + v 2
1 − u2 − v 2
z= .
1 + u2 + v 2
x y
For the latter we differentiate the equations u = 1+z , v= 1+z to get

∂ 1 ∂ 1 + u2 + v 2 ∂
= =
∂x 1 + z ∂u 2 ∂u

∂ 1 ∂ 1 + u2 + v 2 ∂
= =
∂y 1 + z ∂v 2 ∂v
∂ x ∂ y ∂ ∂ ∂
=− 2
− 2
= −u(1 + u2 + v 2 ) − v(1 + u2 + v 2 ) .
∂z (1 + z) ∂u (1 + z) ∂v ∂u ∂v
We find
∂ ∂ ∂ ∂
−y +x = −v +u
∂x ∂y ∂u ∂v
and also with a bit of algebraic manipulation
∂ ∂ ∂ ∂ ∂
xz + yz + (z 2 − 1) =u +v .
∂x ∂y ∂z ∂u ∂v

(By the way, the two vectors (−y, x, 0) and (xz, yz, z 2 − 1) are orthogonal to each other and also
have the same length:
(−y)2 + x2 = x2 + y 2
(xz)2 + (yz)2 + (z 2 − 1)2 = (x2 + y 2 )z 2 + (z 2 − 1)2 = (1 − z 2 )z 2 + (1 − z 2 )(1 − z 2 ) = 1 − z 2 = x2 + y 2 .
The same is true of the two vectors(−v, u) and (u, v). This reflects something about stereographic
projection: it is a conformal mapping; it preserves infinitesimal angles. This can be seen geometri-
cally, or by calculation.)

7. Define F : R2 → R3 by F (x, y) = (x, y 2 , y 3 − xy). Find all points in R2 at which F fails to be

an immersion. Find all points p ∈ R3 such that there is more than one point q ∈ R2 for which
F (q) = p. For each such p, determine all of the “tangent planes of F (R2 ) at p” (the images of the
linear maps Dq F : R2 → R3 ).

Solution: Differentiation with respect to x yields (1, 0, −y). Differentiation with respect to y yields
(0, 2y, 3y 2 − x). This 3 × 2 matrix has rank < 2 if and only if 2y = 0 and 3y 2 − x = 0. This means
(x, y) = (0, 0). So (0, 0) is the only non-immersion point.

If F (x, y) = F (x0 , y 0 ) then x = x0 , y 2 = y 02 , and y 3 − xy = y 03 − x0 y 0 . Thus, unless (x, y) = (x0 , y 0 ),

we must have (x0 , y 0 ) = (x, −y) (with y 6= 0), with y 3 − xy = −y 3 − xy, so y 3 = xy, so y 2 = x. Thus
the points must have the form p = (y 2 , y) for some y 6= 0, and the only other point with the same
image is q = (y 2 , −y). Note that F (p) = F (q) = (y 2 , y 2 , 0).

The image of Dp F is spanned by the vectors (1, 0, −y) and (0, 2y, 3y 2 − y 2 ) = (0, 2y, 2y 2 ), so by
(1, 0, −y) and (0, 1, y). The image of Dq F is spanned by (1, 0, y) and (0, 1, −y).

5
(By the way, the intersection of those two planes is spanned by (1, 1, 0). And in the limit as y
approaches 0 the two planes both become the plane spanned by (1, 0, 0) and (0, 1, 0). Note that F
maps the curve x = y 2 in a two-to-one way onto a half-line, with the non-immersion point going
to the endpoint of the half-line.)

8. Show that there is a smooth embedding of RP n in the space of real (n+1)×(n+1) real matrices,
defined by e(x0 : . . . : xn ) = |x|1 2 xt x. Here x = (x0 , . . . , xn ) is a nonzero vector in Rn+1 , considered
also as a 1 × (n + 1) matrix, and xt is the transpose.

Solution: The map is well defined, because any other vector representing the same point in RP n
as x must be cx for some non-zero c ∈ R, and
1 1 t
xt x = e(x) = x x.
|c|2 |cx|2 |x|2

The fact that it is smooth follows from the obvious fact that |x|1 2 xt x is a smooth function of
x = (x0 , . . . , xn ). (Consider the usual coordinate charts for RP n . For example, in the open set of
RP n where x0 6= 1 we can set x0 = 1 and use x1 , . . . , xn as coordinates. Then |x|1 2 xt x is a smooth
function of (x1 , . . . , xn ).

1
It is also easy to see that it is injective: The vector space spanned by the rows of |x|2
xt x (or
equivalently by the rows of xt x) is the one-dimensional space spanned by the vector x.

We can finish the job by proving that this one to one smooth map is an immersion, and is a
homeomorphism onto its image. The fact that it is a homeomorphism onto its image follows from
its being a continuous injection from a compact space (RP n ) to a Hausdorff space (the space R(n+1)
of matrices). The fact that it is an immersion can be verified by a derivative computation. Many
of you calculated the partial derivatives of each matrix entry of |x|1 2 xt x, and then did some serious
row or column manipulations to see that the rank was n. This is not much fun. A nice labor-
saving device is to restrict your attention to S n ⊂ Rn+1 − 0, where the map is given by x 7→ xt x,
and now make an rank calculation. This suffices because we are now looking at a composition
2
S n → RP n → R(n+1) of e with a map that is locally a diffeomorphism.

But there is also an even easier way: the proof of injectivity leads to a way of making a smooth
map back from the space of matrices to RP n , locally: Send a matrix to the point in Rn+1 spanned
by its ith row. Call this map Ri . It is well-defined and smooth in the open set Wi consisting of all
matrices with nonzero ith row. Its composition with e (where defined) is the identity. This already
proves that e is an immersion, since (De(x) Ri ) ◦ (D(x) e) = D(x) (Ri ◦ e) is the identity map of the
vector space T(x) RP n . (Here i is chosen so that e(x) ∈ Wi , that is, such that xi 6= 0.) At the same
time this also shows that the map is a homeomorphism onto its image without making use of the
compact/Hausdorff argument.

10. Let O(n) be the set of all n × n real matrices A such that At A = I. (Note that these are
precisely those A such that the corresponding linear operator on Rn preserves the inner product,
since hAv, Awi = hAt Av, wi. Thus O(n) is a subgroup of GLn (R), the group of invertible matrices.)
Show that O(n) is a compact smooth submanifold of the vector space of all n × n matrices, of
dimension n(n−1)
2 .

6
Solution: O(n) is compact because it is closed and bounded (closed because it can be defined by
polynomial equations, bounded because every column of an orthogonal matrix is a vector of length
one).

One way to see that it is a smooth manifold of the given dimension is to consider the polynomial
map A 7→ At A from the n2 -dimensional vector space of all n × n matrices to the n(n+1)
2 -dimensional
vector space of all symmetric matrices. The derivative of this smooth map at A is the linear map

LA : X 7→ X t A + At X = X t A + (X t A)t ,

because this is the best linear approximation of

X 7→ (A + X)t (A + X) − At A.

It will suffice if LA is surjective whenever A ∈ O(n). In fact, it is surjective whenever A is invertible.

To verify this we find the dimension of the nullspace. This consists of those X such that X t A is
skew-symmetric, and this is isomorphic as a vector space to the space of skew-symmetric matrices,
so that its dimension is indeed n(n−1)2 = n2 − n(n+1)
2 .

11. We say that a smooth map F : M → N is transverse to a smooth submanifold P ⊂ N if for

every a ∈ M such that F (a) ∈ P the tangent space TF (a) N is equal to TF (a) P + (Da F )(Ta M ).
Note that when F is the inclusion of a submanifold M of N then this is the same as the condition
defined in class: M is transverse to P in N . Show that if F is transverse to P then F −1 (P ) is a
smooth submanifold of M .

Solution: It is enough if in a neighborhood U of each point a ∈ F −1 (P ) the set U ∩F −1 (P ) is defined

by the vanishing of smooth functions f1 , . . . , fn−p such that the cotangent vectors da fj ∈ Ta∗ M are
linearly independent. By assumption, in a neighborhood V of b = F (a) the set V ∩ P is defined
by the vanishing of smooth functions g1 , . . . , gn−p such that the cotangent vectors db gj ∈ Tb∗ N are
linearly independent. These cotangent vectors all restrict to 0 on Tb P . (In fact they form a basis
for the dual of the normal space Tb N/Tb P .) The transversality implies that the pulled back vectors
F ∗ db gj are linearly independent. But this gives what we want, taking U = F −1 (V ) and fj = F ∗ gj ,
since F ∗ db gj = da F ∗ gj = da fj .

A diagonal point in M × M is a point (p, q) such that p = q. Denote the set of diagonal points by
∆M . For F : M → N a double point is a point (p, q) such that p 6= q and F (p) = F (q). Thus if we
write F (2) : M × M − ∆M → N × N for the restriction of the map F × F : M × M → N × N then
the set of double points is the preimage of ∆N under F (2) . Say that the double points of F occur
transversely if F (2) is transverse to ∆N . Note that in this case the set of double points is a smooth
manifold of dimension 2m − n (in particular empty if 2m < n).

12. Show that the map F × F can never be transverse to ∆N at a diagonal point of M × M if
m < n.

Solution: The neatest way to say this is that F × F takes the submanifold ∆M into the submanifold
∆N and therefore its derivative at a diagonal point (p, p) takes the tangent space of ∆M into the
tangent space of ∆N . Transversality means that the composed map

T(p,p) (M × M ) → T(F (p),F (p)) (M × M ) → T(F (p),F (p)) (M × M )/T(F (p),F (p)) ∆N ∼
= TF (p) N

7
is surjective. But this factors through T(p,p) (M × M )/T(p,p) (∆M ) ∼
= Tp M , which has dimension
m < n.

A non-immersion point of F is a point p ∈ M such that the linear map Dp F : Tp M → TF (p) N is

not injective.

13. Prove that if the diagonal point (p, p) is a limit of double points for F then p is a non-immersion
point for F .

Solution: This is essentially the familiar statement that an immersion is locally an embedding.

Let 0M ⊂ T M be the image of the zero-section M → T M . Denote by D(0) F the restriction of

DF : T M → T N to the complement T M − 0M , so that F is an immersion if and only if the inverse
image of 0N by D(0) F is empty. Say that the non-immersion points of F occur transversely if D(0) F
is transverse to 0N .

14. Show that F must be an immersion if the non-immersion points occur transversely and 2m ≤ n.

Solution: The codimension of 0N in T N is n, so if DF : T M → T N is transverse to 0N at a point

v ∈ Tp M then the dimension of T M (which is 2m) must be at least n. In fact it must be strictly
greater if v ∈
/ 0M , because in that case the kernel of Dv (DF ) is nontrivial, since DF maps the
one-dimensional vector space spanned by v in Tp M to a point in T N . (Incidentally, if v ∈ 0M then
transversely cannot even occur if m < n, by an argument as in Problem 12, since DF takes the
submanifold 0M into 0N .)