III )
Section 3.4 Section 3.4 Implicit FUnction Theorems
Then
Xl
=Xo - f'(XO)-ly
=Xo
-(I + B)y =Xo - Y -
By. Hence3 [ 1
2] t [ 1
2]
Xl
(8) = 2 2-
"18 - 8 -10
(38t 2 -"It - t dt b �.0063968616)
=
1
+ 82 +(.0071279315)
8Problems 3.3
135
•
1. For the one-dimensional version of Newton's method, prove that if r is a root of multi
plicity m, then quadratic convergence in the algorithm can be preserved by defining Xn+! = Xn -mI(xn)/I'(xn)
2. Prove the corollaries, giving in each case the precise assumptions that must be made concerning the starting points.
3. Let eO, el . ..
.
be a sequence of positive numbers satisfying en+l � ce�. Find necessary and sufficient conditions for the convergence limn en = O.4. Let I be a function from IR to IR that satisfies the inequalities f' > 0 and I" > O. Prove that if I has a zero, then the zero is unique, and Newton's iteration, started at any point, converges to the zero.
5. How must the analysis in Theorem 1 be modified to accommodate functions from C to C? (Remember that the Mean Value Theorem in its real-variable form is not valid. ) 6 . If r i s a zero o f a function I , then the corresponding "basin o f attraction n is the set
of all x such that the Newton sequence starting at x converges to r. For the function I(z) = z2 + 1, z E C, and the zero r = i, prove that the basin of attraction contains the disk of radius � about r.
3.4 Implicit Function Theorems
In this section we give several versions of the Implicit Function Theorem and prove its corollary, the Inverse Function Theorem. Theorems in this broad cate
gory are often used to establish the existence of solutions to nonlinear equations of the form
f(x)
= y. The conclusions are typicallylocal
in nature, and describe how the solutionX
depends on y in a neighborhood of a given solution(XO , yo).
Usually, there will be a hypothesis involving invertibility of the derivative
f'(xo).
The intuition gained from examining some simple cases proves to be com
pletely reliable in attacking very general cases. Consider, then, a function
F
:IR.2
-+ JR. We ask whether the equationF(x,
y) =0
defines y to be a unique function ofx.
For example, we can ask this question for the equationX
+ y2 -1 =0 (X,
y EJR)
This can be "solved" to yield y = vT=X. The graph of this is shown in the accompanying Figure
3.2.
It is clear that we cannot letX
be the pointA
in thefigure, because there is no corresponding y for which
F(x,
y) ==X
+ y2- 1
= o.One must start with a point
(xo, Yo)
like B in the figure, where we already haveF(xo, Yo)
=O.
Finally, observe that at the pointC
there will be a difficulty, for there are values ofX
nearC
to which no y's correspond. This is a point13
6
Chapter 3 Calculus in Banach Spaceswhere
dy/dx
= 00. Recall that ify
=y(x)
and ifF(x, y(x))
= 0, theny'
can be obtained from the equation{In this equation,
Di
is partial differentiation with respect to the ith argument.) Thus
y'
=-DJ F/ D2F,
and the conditiony'(xo)
= 00 corresponds toD2F(xo, Yo)
= o. In this example, notice t.hat another function arises from Equation ( 1 ), namelyy = -
vT=XIn a neighborhood of ( 1 , 0), both functions solve Equation ( 1 ), and there is a failure of uniqueness.
- 3
Figure 3.2
C A
In the classical implicit function theorem we have a function
F
of two real variables in class C 1 • That means simply thatEJF/EJx
andEJF/EJy
exist and are continuous. It is convenient to denote these partial derivatives byFl
andF2.
Theorem 1 . Classical Implicit Function Theorem. Let
F
bea C 1 -function on the square
{(x, y) : Ix - xol
� 6 ,Iy - Yol
� 6} C ]R2
If
F(xo, Yo)
= 0 andF2(xo, Yo)
I- 0, then there is a continuously differentiable functionJ
defined in a neighborhood ofXo
such thatYo
=J(xo)
andF(x, J(x))
= 0 in that neighborhood. Furthermore,f'(x)
=-Fl (x, y)/ F2(x, y),
wherey
=J(x).
Proof. Assume that
F2(xo, Yo)
> o. Then by continuity,F2(X, y)
> Q > 0 in a neighborhood of(xo, Yo),
which we assume to be the original 6-neighborhood.The function
y
HF(xo, y)
is strictly increasing forYo -
6 �y
�Yo
+ 6. HenceF(xo, Yo
-6) <F(xo, Yo)
= 0 <F(xo, Yo
+ 6)By continuity there is an c in (0, 6) such that
F(x, Yo -
6) < 0 <F(x, Yo +
6) ifIx - xol
< c. By continuity, there corresponds to each suchx
a value of y such thatF(x, y)
= 0 andYo -
6 < y <Yo +
6. If there were two such y's, then by Rolle's theorem,F2(x,y)
= 0 at some point, contrary to hypothesis. Hencey
is unique, and we may puty
=J(x).
Then we haveF(x, J(x)) =
0 andYo
=J(xo).
Section 3.4 Implicit Function The
orems
137 Now fixXI
in the c-neighborhood ofXo.
PutYI
=f(xI).
Let�x
be a small number and letYI + �Y = f(xI + �x).
Then0 = F(xI + �X' YI + �y)
=
FI (XI + ()�x, YI + ()�y)�x + F2(xI + ()�x, YI + ()�y)�y
for an appropriate () satisfying
0 � () �
1. (This is the Mean Value Theorem for a function fromJR2
to lR. See Problem 3.2.8, page 1 24.) This equation gives us�Y FI (xI + ()�x, YI + ()�y)
�x F2(xI + ()�x, YI + ()�y)
As
�x
--+ 0 the right side remains bounded. Hence, so does the left side. This proves that as�x
converges to0, �Y
also converges to o. Hencef
is continuous atXI.
After all,�Y = f(xI + �x) - f(xI ).
Furthermore,1'(xd =
lim�Y =
_FI(xI , yd
�x F2(xI , yJ)
Therefore,
f
is differentiable atXI.
The formula can be written1'(x) = - FI (x, f(x)) I F2 (x, f(x))
and this shows that l' is continuous at
x,
provided thatX
is in the open interval(xo - c,Xo +
c). •Theorem 2. Implicit Function Theorem for Many Variables.
Let F : JRn
xJR
--+JR, and suppose that F(xo, Yo) = 0 for some Xo
ElRn and Yo
E lR.If all
n+
1partial derivatives DiF exist and are continuous in a neighborhood of (xo, Yo) and if Dn+1 F(xo, Yo) i= 0, then there is a continuously differentiable function f defined on a neighborhood of Xo such that F(x, f(x))
=0, f(xo) = Yo, and
D;J(x)
=-DjF(x, f(x))IDn+IF(x, f(x))
( 1�
i� n)
Proof. This is left as a problem (Problem 3.4.4). • Example 1.
F(x, y)
=x2 +y2 +
l orx2 +y2
orx2 +y2 -
1 . (Three phenomenaare illustrated. ) •
If we expect to generalize the preceding theorems to normed linear spaces, there will be several difficulties. Of course, division by
F2
will become multiplication by
F2-1,
and the invertibility of the F'rechet derivative will have to be hypothesized. A more serious problem occurs in defining the value ofy
corresponding to
x.
The order properties of the real line were used in the preceding proofs; in the more general theorems, an appeal to a fixed point theorem will be substituted.Definition. Let X, Y, Z be Banach spaces. Let
F :
X x Y --+ Z be a mapping.The Cartesian product X x Y is also a Banach space if we give it the norm
lI(x, y)1I = IIxli + IIYII·
If they exist, the partial derivatives ofF
at(xo,Yo)
arebounded linear operators
DI F(xo, Yo)
andD2F(xo, Yo)
such thatlim
IIF(xo + h,yo) - F(xo,yo) - DIF(xo, yo)hll/llhll = 0 (h
E X,h
--+ 0)and
lim
IIF(xo, Yo + k) - F(xo, Yo) - D2F(xo, yo)kll/llkil = 0 (k
E Y,k
--+ 0)Thus
DIF(xo, Yo)
E C(X, Z) andD2F(xo, Yo)
E C(Y, Z). We often use the notationFi
in place ofDiF.
138
Chapter 3 Calculus in Banach Spaces
Theorem 3. General Implicit Function Theorem.
Let
X, Y,and Z be normed linear spaces,
Ybeing assumed complete. Let fI be an open set in
X x Y.Let F : fI
-+ Z.Let (xo,Yo)
E fl.Assume that
F is continuous at (xo, Yo), that F(xo, Yo)
==0, that D2F exists in fI, that D2F is continuous at (xo, Yo), and that D2F(xo, Yo) is invertible.
Then there is a function f defined on a neighborhood of Xo such that
F(x,f(x)) == 0, f(xo)
==Yo, f is continuous at xo, and f is unique
in the sense that any other such function must agree with f on some
neighborhood of Xo.
Proof. We can assume that
(xo, Yo)
== (0, 0).
Selecta > °
so that{(x, y) : IIxll �
0,IIYII � a}
c fIPut
A
==D2F(O, 0
).
ThenA
E .c(y, Z) andA
- I E £(Z, Y). For eachx
satisfyingII xii � a
we defineG.,(y)
==y - A-I F(x, y).
HerelIyli � o.
Observe that ifG.,
has a fixed point
y.,
theny.
==G.,(y·)
==y' - A-I F(x, y')
from which we conclude that
F(x,y·)
== O. Let us therefore set about proving thatG.,
has a fixed point. We shall employ the Contraction Mapping Theorem.(Chapter 4, Section 2, page
177).
We haveBy the continuity of
D2F
at(0, 0)
we can reducea
if necessary such thatNow
G.,(O)
==-A-1F(x,O)
==_A-l{F(x,O) - F(O,O)}.
Let°
<e
< O. Bythe continuity of
F
at (0, 0
) we can findo
€ E( 0 ,
0)
so thatIf
/lxll �
o€ andlIyli � e,
then by the Mean Value Theorem III of Section 2, page 123,IIG.,(y)11 � IIG.,(O)/I + IIG.,(y) - G.,(O) II
� �e +
0';;>'';; 1 supIIG�('\Y)II . IIYil
e e
� - + - == e
2 2
Define
U
={y
E Y :IIYII � e}.
We have shown that, for eachx
satisfyingIIxll �
O€ , the functionG.,
mapsU
intoU .
We also know thatIIG�(Y)II � � .
ByI'>roblem 1 ,
G.,
has a unique fixed pointy
inU.
Since this fixed point depends onx,
we writey
=f(x),
thus definingf.
From the observations above we infer thatF(x,f(x))
==0
Section 3.4 Implicit Function Theorems
139 SinceF(O,O)
=0,
it follows thatGo(O)
=O.
By the uniqueness of the fixed point,0
=f(O).
Since c was arbitrary in(0, 8)
we have this conclusion: For eachc in
(0, 8)
there is a8E
such thatOur analysis then showed that
y
=f(x) E U,
orIlf(x)11 !!S;
c. As a consequence,II xII !!S; 8€
=?IIf(x)lI !!S;
c, showing continuity atO.
For the uniqueness, suppose that7
is another function defined on a neighborhood of0
such that7
is continuous at
0, 7(0)
=0,
andF(x,f(x))
=O.
If0
< c <8,
find () >0
suchthat () <
8E
andIIxll !!S;
() =?117(x)lI !!S;
cThen
7(x) E U .
So we have apparently two fixed points,f(x)
and7(x),
for thefunction
Gx•
Since this is not possible,f(x)
=7(x)
wheneverIIxll !!S;
().
•Theorem 4. Second Version of the Implicit Function Theorem.
In the preceding theorem, assume further that F is continuously dif
ferentiable in n and that D2F(xo, Yo) is invertible. Then the function
f will be continuously differentiable and
Furthermore, there will exist a neighborhood of Xo in which f is unique.
This theorem can be found in [Dieu] , page 265.
Theorem 5. Inverse Function Theorem I.
Let f be a continu
ously differentiable map from an open set n in a Banach space into a normed linear space. Ifxo E n and if f'(xo) is invertible, then there is a continuously differentiable function
9defined on a neighborhood N
of f(xo) such that f(g(y))
=y for all Y E N.
Proof.
Forx
in n andy
in the second space, defineF(x,y)
=f(x) - y.
PutYo
=f(xo)
so thatF(xo, Yo)
=O.
Note thatDJ F(x, y)
=f'(x),
and thusDJ F(xo, Yo)
is invertible. By Theorem 4, there is a neighborhoodN
ofYo
anda continuously differentiable function 9 defined on
N
such thatF(g(y),y)
=0,
or
f(g(y)) - y
=0
for allY E N.
•Theorem 6. Surjective Mapping Theorem I.
Let
Xand
Ybe Banach spaces, n an open set in
X .Let f : n
-+ Ybe a continuously differentiable map. Let Xo E n and Yo
=f(xo).
1£f'(xo) is invertible, as an element of .c
(x,
V), thenf(n) is a neighborhood of Yo.
Proof.
DefineF : n»
Y -+ Y by puttingF(x,y)
=f(x)-y.
ThenF(xo, Yo)
=o
andDJF(xo,yo)
=f'(xo). (DJ
is a partial derivative, as defined previously.) By hypothesis,Dl F(xo, Yo)
is invertible. By the Implicit Function Theorem (with the roles ofx
andy
reversed!), there exist a neighborhoodN
ofYo
anda. function
g : N
-+n
such thatg(yo)
=Xo
a.ndF(g(y),y)
=0
for ally E N.
From the definition of
F
we havef (g(y)) - y
=0
for allY E N.
In other words, each elementy
ofN
is the image underf
of some point inn,
namely,g(y).
•140 Chapter 3 Calculus in Banach Spaces
Theorem 1. A Fixed Point Theorem. Let n be an open set in a Banach space X. and let
G
be a differentiable map from n to X.Suppose that there is a closed ball B ==
B(xo.
r) in n such that(
i)
k == supIIG'(x)1I
< 1(
ii) IIG(xo) - xoll
xEB < r( 1 - k)Then
G
has a unique fixed point inB.
Proof. First, we show that
GIB
is a contraction. IfXI
andX2
are in B, then by t.he Mean Value Theorem(
Theorem 4 in Section 3.2, page 123)
IIG(xd - G(x2)11
� 0";..\"; 1 supIIG'(xI
+A(X2 - xdll "XI - x211
Second, we show that
G
mapsB
intoB.
IfX
E B , thenIIG(x) - xoll
�IIG(x) - G(xo)II
+IIG(xo) - xoll
�
kllx - xoll
+ r( 1 -. k)� kr + ( I - k)r = r
Since X is complete,
B
is a complete metric space. By the Contractive Mapping Theorem(
page177). G
has a unique fixed point inB.
•Theorem 8. Inverse Function Theorem II. Let n be an open set in a Banach space X. Let
f
be a differentiable map from n to a normed space Y. Assume that n contains a closed ballB
==B(xo,
r)
such that
(
i)
The linear transformationA
==f'(xo)
is invertible.(
ii)
k == SUPxE BIII - A-I f'(x)II
< 1Then for each
y
in Y satisfyingIIY - f(xo)1I
<( 1
-k)
rIlA-I r
l theequation
f(x)
=y
has a unique solution inB.
Proof. Let
y
be as hypothesized, and defineG(x)
=X - A-I [f(x)
-y].
It is clear thatf(x)
=y
if and only ifX
is a fixed point ofG.
The mapG
isdifferentiable in n, and
G'(x)
=I - A-I f'(x).
To verify the hypothesis (i)
inthe preceding theorem, write
IIG'(x)1I
=III - A-I f'(x)1I
� k(X
EB)
By the assumptions made about
y,
we can verify hypothesis(
ii)
of the preceding theorem by writingIIG(xo) - xoll = IIxo - A-I [J(xo)
-y] - xoll
=
IIA-I II IIf(xo) - YII
� IIA-I II(l - k)rIIA-I II-I
= (1 - k)rBy the preceding theorem,
G
has a unique fixed point inB,
which is the uniquesolution of
f(x)
= y inB.
•Section
3.4
Implicit Function Theorems Example 2. Consider a nonlinear Volterra integral equationx(t} - 2x(0}
+ � lt
cos(st)[x(sWds = y(t}
(0�
t� I)
141
in which
y
E qO,I ]
. Notice that wheny =
0 the integra) equation has the solution x= o.
We ask: Does the equation have solutions whenIIYII
i s small?Here, we use the usual sup-norm on CIO,
1],
as this makes the space complete.(Weighted sup-norms would have this property, too.) Write the integral equation as
f(x} = y,
wheref
has the obvious interpretation. Thenf'(x}
is given by[J'(x}h](t} = h(t} - 2h(0}
+lt cos(st}x(s}h(s}ds
Let
A = I'(O},
so thatAh = h -
2h(0}. One verifies easily thatA2h = h,
fromwhich it follows that
A-I = A.
In order to use the preceding theorem, withXo =
0, we must verify its hypotheses. We have just seen thatA
is invertible.Let
IIxll � r,
wherer
is to be chosen later so thatIII - A-I I'(x} II � k < 1.
From an equation above,
It follows that
I (J'(x}h](t) - (Ah}(t) 1 = 11t
cos( s
t}x(
s}h(
s} ds l
� IIhll llxll
and that
1IJ'(x}h - Ahll � IIhll llxll 1IJ'(x} - All � Ilxll � r
Since
IIAII = IIA-III = 3,
we haveIII - A-IJ'(x}1I = IIA-I(A - J'(x}}11 � IIA-Illr = 3r
The hypothesis of the preceding theorem requires that
3r � k < 1,
wherek
isto be chosen later. By the preceding theorem, the equation
f(x} = y
will havea unique solution if
In order for this bound to be as generous as possible, we let
k = !,
arriving atthe restriction
IIYII < -k.
•Lemma. Let X and Y be Banach spaces. Let n be an open set in X, and let
f
: n � Y be a continuously differentiable mapping. IfXo
E n and E > 0, then there is a 6 > 0 such thatProof. The map x >--+
f'(x}
is continuous from n to .c(X, Y}. Therefore, in correspondence with the given E, there is a 6 > 0 such thatIIx - xoll <
6 ==>1IJ'(x} - !'(xo}11 < E
142 Chapter 3 Calculus in Banach Spaces
(We may assume also that
B(xo,6)
e n.) IfII XI - xoll <
6 andIIx2 - xoll <
6,then the line segment
S
joiningXI
toX2
satisfiesS
CB (xo , S)
c n. By Problem2,
page 145, we haveIIf(xd - f(X2) - !'(XO)(XI - x2)11 � IIxl - x211 .
sup xESII!'(x) - !'(xo) II
� ellXI - X211
•Theorem 9. Surjective Mapping Theorem II.
Let
Xand Y be Banach spaces. Let
nbe an open set in
X.Let f
: n -+Y be
continuously differentiable.
IfXo
E nand f'(xo) has a right inverse in .c
(Y,
X),then f(
n) is a neighborhood of f(xo)·
Proof. Put
A
=f' (xo)
and letL
be a member of.c
(Y,
X) such thatAL
= I,where I denotes the identity map on
Y.
Letc
=IILII.
By the preceding lemma, there exists 6>
0 such thatB(xo,6)
e n and such thatlIu - xoll �
6,IIv - xoll �
6=? IIf(u) - f(v) - A(u - v)11 � to lIu - vII
Let
Yo
=f(xo)
andy
EB(yo, 6/2c).
We wiII findX
E n such thatf(x)
=y.
The point
X
is constructed as the limit of a sequence{xn}
defined inductively as follows. We start with the givenXo.
PutXI
=Xo
+L(y - Yo).
From then on we defineXn+1
=Xn - L[J(xn) - f(xn-d - A(xn - Xn-d
By induction we establish that
IIXn - xn-dl � 6/2n
andIIxn - xoll � 6.
Hereare the details of the induction:
IIxl - xoll
=IIL(Y - Yo II � cllY - Yo II � c6/(2c)
=6/2 . IIXn+1 - xnll � cllf(Xn) - f(xn-d - A(xn - xn-dll
� c(i/2C)IIXn - xn_III � 6/2n+1
IIxn+1 - xoll � IIxn+1 - xnll
+IIXn - Xn-III
+ . . . +IIXI - xoll
6 6 6
�
--
+ - + · · · + - � 6"" 2n+1 2n 2 '"
Next we observe that the sequence
(xn!
has the Cauchy property, since (form > n)
IIXn -xmll � IIxn -Xn+1 II
+ . . . +IIXm-1 -Xmll � 6 ( 2n � 1
+2n � 2
+ . . .) � 6/2n
Since X is complete, we can define
X
= Hmxn
. All that remains is to provex
E n andf(x)
=y.
SinceIIXn - xoll �
6, we haveIIx - xoll �
6 andX
E n.From the equation defining
Xn+1
we haveA(Xn+1 - Xn)
=-AL{J(xn) - f(Xn-l) - A(xn - xn-d}
=
A(xn - xn-d - {J(Xn) - f(xn-d}
By using this equation recursively we reach finally
A(Xn+1 - Xn)
=A(xi - XO) - {J(Xn) - f(Xn-l)}
- {J(xn-d - f(Xn-2)} - ... - {J(xd - f(xo)}
=
AL(y - Yo) - f(xn)
+f(xo)
=
y - Yo - f(xn)
+Yo
=Y - f(xn)
Let