Math 1540

Spring 2011
Notes #6
Inverse and Implicit function theorems
1 Introduction
These theorems are fairly complicated, especially in their proofs, so I will start with
an example of the one-dimensional case and then move to higher dimensions. First
I want to recall some terminology about functions.
Suppose that , : 1
n
1
m
and suppose \ 1
m
. We dene the inverse image
of \ under , as follows:
,
1
(\) := r 1
n
[ , (r) = for some in \ .
This denition does not require that the function , has an inverse. There might be
more than one r for each . As an example, if : = : = 1. , (r) = r
2
and \ = (0. 1) .
then ,
1
(\) = (1. 1). The function , is not 1 : 1 on (1. 1) and so doesnt have
an inverse there. However, the function ,[
(0;1)
is 1 : 1 on (0. 1) and does have an
inverse, which is q () =
_
for 0 < < 1.
Now recall Theorem 4.1.4 on page 180. One consequence of this theorem is the
following:
Lemma 1 Suppose that , : 1
n
1
m
. Then , is continuous if and only if ,
1
(\)
is an open set for each open subset \ of 1
m
.
As examples, we can consider again , (r) = r
2
and notice that ,
1
(0. 1) =
(1. 1) . On the other hand, suppose that
, (r) =
_
r if r _
1
2
r if r
1
2
Let \ =
_
1
4
.
3
4
_
. Then ,
1
(\) = (
1
4
.
1
2
] '
_
.
1
4
_
. This is not an open set.
2 Inverse function theorem in one variable
Theorem 2 Suppose , : 1 1 and , C
1
. (That is, ,
0
(r) exists for each r 1
and ,
0
is continuous.) Suppose that , (c) = / and ,
0
(c) ,= 0. Then there are open
neighborhoods l of c and \ of / such that
1
1. , (l) = \ and ,[
U
is one to one.
2. 1, / is the inverse of ,[
U
. then / C
1
.
3. For each r l. /
0
(, (r)) =
1
f
0
(x)
.
Proof. We will rst prove the theorem under the assumptions that c = / = 0 and
,
0
(0) = 1. In this case, let
G(r) = r , (r) .
Then G(0) = G
0
(0) = 0. Since ,
0
is continuous, G C
1
. and so there is an 0
such that if 2 _ r _ 2. then [G
0
(r)[ _
1
2
. Let
\ = (. ) .
and
l = ,
1
(\) (2. 2) = r (2. 2) [ , (r) \ .
Then , (l) = \. \ is open, and since the inverse image of continuous function is
open, l is open as well. (l is the intersection of two open sets.) We must show
that ,[
U
is 1 : 1. This is pretty easy to prove, but we will give a proof which can be
generalized to the : dimensional case.
For each \. let q
y
(r) = +G(r) . Then q
y
C
1
. and q
0
y
(r) = G
0
(r) . Hence,
if r [2. 2] . then

q
0
y
(r)

_
1
2
. This shows that q
y
is a contraction on (2. 2) .
We want to apply the contraction mapping theorem, but to do this we need to show
that q
y
: (2. 2) (2. 2) . If r (2. 2) . then
[q
y
(r)[ = [ + G(r)[ _ [[ +[G(r)[ .
Since \. [[ < . Since [G
0
(c)[ _
1
2
for c (2. 2) . and G(0) = 0. the mean
value theorem implies that [G(r)[ _
1
2
[r[ for r (2. 2) . Hence,
[q
y
(r)[ < + = 2.
Therefore q
y
is a contraction and maps (2. 2) into itself. Hence, there is a unique
r (2. 2) such that
+ r , (r) = r.
which implies that , (r) = . On the other hand, if , (r) = . then r is a xed point
for q
y
.
We get such an r for each \. Dene / : \ (2. 2) by letting /()
be the unique xed point of q
y
in (2. 2) . The uniqueness of /() implies that
2
,[
(2";2")
is 1 : 1. and since l (2. 2) . ,[
U
is 1 : 1 and / = (,[
U
)
1
. We have
left to show that / is dierentiable and /
0
() =
1
f
0
(h(y))
.
We rst show that / is continuous. Suppose
1
and
2
are in \. and let r
1
=
/(
1
) . r
2
= /(
2
) . Since r = , (r) + G(r) . we obtain that
[r
1
r
2
[ = [, (r
1
) , (r
2
) + G(r
1
) G(r
2
)[ _ [, (r
2
, (r
2
))[ +
1
2
[r
1
r
2
[ .
or
1
2
[r
1
r
2
[ _ [, (r
1
) , (r
2
)[ . (1)
In other words,
[/(
1
) /(
2
)[ _ 2 [
1

2
[ .
showing that / is continuous.
Next note that on (2. 2) . ,
0
= 1 G
0
. and since [G
0
(r)[ _
1
2
. ,
0
(r) _
1
2
. In
particular, ,
0
(/()) ,= 0. To show that / is dierentiable we use the denition of
derivative, and compute
lim
y!y
0

/() /(
0
)
1
f
0
(h(y
0
))
(
0
)

[
0
[
for some
0
(. ) . If r and r
0
are in (2. 2) . and /(r) = . /(r
0
) =
0
. then

/() /(
0
)
1
f
0
(h(y
0
))
(
0
)

[
0
[
=
[r r
0
[
1
f
0
(x
0
)
(, (r) , (r
0
))
[, (r) , (r
0
)[
.
Applying the mean value theorem to , (r) , (r
0
) in both the numerator and de-
nominator,

/() /(
0
)
1
f
0
(h(y
0
))
(
0
)

[
0
[
=
[r r
0
[
f
0
(c)
f
0
(x
0
)
(r r
0
)
[,
0
(c) (r r
0
)[
=
[r r
0
[
[,
0
(c)[ [r r
0
[
_
1
,
0
(c)
,
0
(r
0
)
_
.
where c is between r and r
0
. From (1) the rst term is bounded by 2. Since / is
continuous, lim
y!y
0
(r r
0
) = lim
y!y
0
(/() /(
0
)) = 0. Since ,
0
is continuous
and c is between r and r
0
. the second term tends to zero as r r
0
. proving that
/
0
(
0
) exists and that /
0
(
0
) =
1
f
0
(h(y
0
))
. This proves Theorem 2 in the case where
, (0) = 0 and ,
0
(0) = 1.
There are three steps to complete the proof of the theorem.
3
1. We remove the requirement that ,
0
(0) = 1. and only require that ,
0
(0) ,= 0. If
,
0
(0) = c. let q (r) = , (r,c). Then q
0
(0) = 1. Prove the theorem for q and it
easily gives a comparable theorem for ,.
2. Suppose this is done, and we only require that ,
0
(0) ,= 0. We now remove the
requirement that , (0) = 0. Introduce a new q : q (r) = , (r) , (0) . Then
q (0) = 0. and the theorem for q implies the theorem for q.
3. To remove the requirement that c = / = 0. let . = rc and /(.) = , (. + c)
/. then /(0) = 0 and we are back to a previous case.
3 Inverse function theorem for , : 1
:
1
:
.
See that auxiliary notes Notes 6a, which follow the exercises below, where a side by
side comparison is given of the one dimensional and :-dimensional cases. This is
restricted to the case of , (0) = 0. 1, (0) = 1 (the identity transformation). Steps
similar to 1,2,3 above give the general case, in which it is assumed that 1, (r
0
) is
a 1 : 1 map; that is 1, (r
0
) (n) = 0 if and only if n = 0. This is equivalent saying
that the standard matrix of 1, (r
0
) (which is : :) is non-singular. We then get
the theorem in the text (pg. 393), which I will not copy here.
4 Implicit function theorem.
The following is another way of stating the inverse function theorem:
Theorem 3 Suppose that , : 1
n
1
n
satises the conditions of the inverse func-
tion theorem (as given in notes 6a). Suppose that , (r
0
) =
0
. Then in a neighborhood
l \ of the point (r
0
.
0
) of 1
n
1
n
. the equations
,
i
(r
1
. ..... r
n
) =
i
. for i = 1. .... : (2)
can be solved for (r
1
. .... r
n
) in terms of (
1
. .....
n
) .
Note that (2) is a system of : equations for the : unknowns r
1
. .... r
n
. For any
\. the unique solution r of these equations in l is r = ,
1
() . where ,
1
is
the function / found in the inverse function theorem.
The one dimensional version of this is particularly simple: The implicit function
theorem is a generalization of this.
4
Theorem 4 Suppose that 1 : 1
n
1
m
1
m
. and 1 C
1
. Suppose the 1 (r
0
.
0
) =
0 1
m
. for some (r
0
.
0
) 1
n
1
m
. Let
= det
_
_
_
_
_
@F
1
@y
1
(r
0
.
0
)
@F
1
@y
2
(r
0
.
0
)
@F
1
@ym
(r
0
.
0
)
@F
2
@y
1
(r
0
.
0
)
@F
2
@y
2
(r
0
.
0
)
@F
2
@ym
(r
0
.
0
)

@Fm
@y
1
(r
0
.
0
)
@Fm
@y
2
(r
0
.
0
)
@Fm
@ym
(r
0
.
0
)
_
_
_
_
_
.
Assume that ,= 0. Then there are open neighborhoods l of r
0
in 1
m
and \ of
0
in 1
m
. and a unique function , : l \ such that 1 (r. , (r)) = 0 for all r l.
Further, , is dierentiable at each r l.
We will only prove the theorem in the case : = 3. : = 2. in which case we can
write out everything explicitly. Before giving the proof, we give examples in this
case. The easiest examples are linear, such as
r + + . = n +
r . = n
(3)
In this case,
1 (r. . .. n. ) =
_
r + + . n
r . n +
_
.
and (3) is equivalent to 1 (r. . .. n. ) =
_
0
0
_
. We see that
=
_
1 1
1 1
_
.
so det o = 2. We can solve for n and in terms of r. . . by adding and subtracting
the equations, which give
n = r
=
1
2
( + .) .
We can set l = 1
3
. \ = 1
2
. since this solution is valid for all (r. . .) .
Now change the equations a bit:
r + + . = n
2
+
r . = n
(4)
5
Adding the equation gives
n
2
+ n = 2r
We can solve this for n in terms of r by using the quadratic formula:
n
2
+ n 2r = 0
n =
1
_
1 + 8r
2
.
Then we can solve for :
=
1
_
1 + 8r
2
(r .) . (5)
We see that the solution is not unique. This is where neighborhoods come in.
Looking at the statement of the theorem, we see that we need to pick a point (r
0
.
0
)
which is a solution of the set of equations. First, though, lets write down the function
1 :
1 (r. . .. n. ) =
_
r + + . n
2

r . n +
_
.
Next we need to pick a specic point in 1
3
1
2
(the point called (r
0
.
0
) in the
theorem). For instance, we can chose the point (0. 0. 0. 0. 0) 1
3
1
2
. We see that
1 (0. 0. 0. 0. 0) =
_
0
0
_
. Now we nd :
= det
_
2n 1
1 1
_
[
(0;0;0;0;0)
= 1.
This is nonzero, so the theorem applies. We can solve for (n. ) in terms of (r. . .) .
and indeed, we did so. But to get a unique solution we need to choose neighborhoods.
The point r
0
is now (0. 0. 0) . and the point
0
is (0. 0) . The solution = , (r) is now
written (n. ) = , (r. . .) . and we must have (0. 0) = , (0. 0. 0) . Notice in (5) and the
equation just above, that if we choose the minus sign, we get , (0. 0. 0) = (1. 1) .
Therefore we must choose the plus sign. We get
, (r. . .) =
_
1+
p
1+8x
2
1+
p
1+8x
2
(r .)
_
.
This doesnt make sense of r <
1
8
. And the solution to (4) is not unique if we allow
l \ to include the point (1. 1. 0. 0. 0) . For neighborhoods we could choose
l =
_
(r. . .) [
1
8
< r
_
. \ =
_
(n. ) [ n
1
2
_
6
If we take the alternate sign in the solution, the minus sign, then we obtain n <
1
2
.
So in l there is a unique solution for (n. ) which lies in \.
5 Homework due February 16
1. Suppose , : 1
2
1
1
. Suppose that , C
1
. Can ,[
U
have an inverse for some
open set l? Can ,[
U
have a continuous inverse? Can ,[
U
have a dierentiable
inverse?
2. pg. 396, # 1.
3. pg. 396, # 2
4. pg. 400, # 1
5. pg. 401, # 5.
7
Math 1540
Spring 2011
Notes #6a
Inverse and Implicit function theorems
1 Inverse function the-
orem for f : R
n
! R
n
We repeat the statement and proof for
f : R ! R in abbreviated form, for the
case f (0) = 0; f
0
(0) = 1; and then, in the
next column, do the statement and proof
for f : R
n
! R
n
:
Theorem 1 Suppose that f : R ! R;
f 2 C
1
; f (0) = 0; f
0
(0) = 1: Then
there are neighborhoods U and W of 0
such that f (U) = W; fj
U
is 1 : 1, and
if h = (fj
U
)
1
; then h 2 C
1
, and for
y 2 W; h
0
(y) =
1
f
0
(h(y))
:
Proof: Let G(x) = x f (x) : Then
G(0) = G
0
(0) = 0: There exists an " > 0
such that if 2" < x < 2"; then jG
0
(x)j <
1
2
:
Let W = ("; ") and U = f
1
(W) \
(2"; 2") ; where f
1
(W) is the preimage
of W under f:
For y 2 W and x 2 (2"; 2") ; let
g
y
(x) = y + G(x) : Then jg
y
(x)j
jyj + jG(x)j < " +
1
2
jxj < 2"; so g
y
:
(2"; 2") ! (2"; 2") : Also, g
0
y
= G
0
on
(2"; 2") : For x
1
and x
2
2 (2"; 2") ;
jg
y
(x
1
) g
y
(x
2
)j = jG(x
1
) G(x
2
)j

1
2
jx
1
x
2
j :
Theorem and Proof in n-
dimensional case
Read this column in parallel with the
rst column. Here,
B
1
(0; r) = fx 2 R
n
j jjxjj
1
< rg :
For n = 2 this is a square, not a circle.
Theorem 2 Suppose that f : R
n
! R
n
;
f 2 C
1
; f (0) = 0; Df (0) (u) = u.
(Thus, the standard matrix for Df (0)
is the identity matrix I:) Then there are
neighborhoods U and W of 0 such that
f (U) = W; fj
U
is 1 : 1, and if h =
(fj
U
)
1
; then h 2 C
1
, and for y 2 W;
Dh(y) = [Df (h(y))]
1
: (So the corre-
sponding matrices are inverses.)
Proof: Let G(x) = x f (x) : Then
G(0) = 0 2 R
n
; DG(0) = 0 2
L(R
n
; R
n
) : There exists an " > 0
such that if x 2 B
1
(0; 2") ; then
jjDG(x) (u)jj
1
<
1
2
jjujj
1
for all u 2 R
n
:
Let W = B
1
(0; ") and U = f
1
(W) \
B
1
(0; 2") ; where f
1
(W) is the preim-
age of W under f:
For y 2 W and x 2 B
1
(0; 2") ; let
g
y
(x) = y + G(x) : Then jjg
y
(x)jj
1

jjyjj
1
+ jjG(x)jj
1
< " +
1
2
jjxjj
1
< 2";
so g
y
: B
1
(0; 2") ! B
1
(0; 2") : Also,
Dg
y
= DG on B
1
(0; 2") : For x
1
and
x
2
2 B
1
(0; 2")
jjg
y
(x
1
) g
y
(x
2
)jj
1
= jjG(x
1
) G(x
2
)jj
1

1
2
jjx
1
x
2
jj
1
:
1
Hence g
y
is a contraction on (2"; 2")
and g
y
has a unique xed point x. Then
from the denitions of g
y
and G;
y +x f (x) = x;
so f (x) = y: The uniqueness of x implies
that fj
(2";2")
is 1 : 1; and so fj
U
is 1 : 1;
since U (2"; 2") : Hence, fj
U
has an
inverse, which we denote by h:
For y
1
and y
2
in W; let x
1
=
h(y
1
) ; x
2
= h(y
2
). Note that x = f (x) +
G(x) ; so
jx
1
x
2
j jf (x
1
) f (x
2
)j+
1
2
jx
1
x
2
j ;
and
1
2
jx
1
x
2
j jf (x
1
) f (x
2
)j : (1)
To compute the derivative of h we re-
duce " further, if necessary, to insure that
f
0
(x) 6= 0 in U; and consider

h(y) h(y
0
)
1
f
0
(h(y
0
))
(y y
0
)

jy y
0
j
:
If x = h(y) ; x
0
= h(y
0
) ; then we get
from (1) that lim
y!y
0
jx x
0
j = 0; and

x x
0

1
f
0
(x
0
)
(f (x) f (x
0
))

jf (x) f (x
0
)j
=
j(x x
0
)j
jf (x) f (x
0
)j

1
f
0
(x
0
)

(f
0
(x
0
) (x x
0
) (f (x) f (x
0
)))
x x
0

From (1), the rst fraction is bounded by

2; and because f is dierentiable, the sec-
ond tends to zero, showing that h is dif-
ferentiable at x
0
with h
0
(y
0
) =
1
f
0
(h(y
0
))
;
which is continuous in W:
Hence g
y
is a contraction on B
1
(0; 2") :
Hence, g
y
has a unique xed point x.
Then from the denitions of g
y
and G;
y +x f (x) = x;
so f (x) = y: The uniqueness of x implies
that fj
B1(0;2")
is 1 : 1; and so fj
U
is 1 : 1;
since U B
1
(0; 2") : Hence, fj
U
has an
inverse, which we denote by h:
For y
1
and y
2
in W; let x
1
=
h(y
1
) ; x
2
= h(y
2
). Note that x = f (x) +
G(x) ; so
jjx
1
x
2
jj
1
jjf (x
1
) f (x
2
)jj
1
+
1
2
jjx
1
x
2
jj
1
;
so
jjx
1
x
2
jj
1

1
2
jjf (x
1
) f (x
2
)jj
1
:
(2)
To compute the derivative of h we re-
duce " further, if necessary, to insure that
det (Df (x)) 6= 0 in U; and consider

h(y) h(y
0
) [Df (h(y
0
))]
1
(y y
0
)

1
jjy y
0
jj
1
:
Note that Df (h(y
0
)) and its inverse can
both be considered as matrices. The rest
is left to the reader.
2