III )
Section 3.4 Implicit Function Theorems
144
Chapter 3 Calculus in Banach SpacesIf
X =
L�IGlX;
and iff
is a mapping from an open set ofX
into a normed space Y, then the partial derivativesD;/(x),
if they exist, are continuous linear maps(D;/)(x)
E£(X;,
Y) such t.hatThe connection between partial derivatives and a "total" derivative is as one expects from multivariable calculus. That relationship is formalized next.
( 1 )
Theorem 10. Let
f
be defined on an open set n in the direct-sum spaceX = L�=I GlXj
and take values in a normed space Y. Assume that all the partial derivativesD;/
exist in n and are continuous at a pointx
in n . Thenf
is Fnkhet differentiable atx,
and its Frechet derivative is given byJ'(x)h = L
;=1n D;/(x)h; (h
E X)Proof. Equat.ion ( 1 ) defines a linear transformation from
X
to Y, andn n
1\J'(x)hll � L
;=1IID;f(x)hdl � L
.=1IID;/(x)II Ilhill
� I';;J';;n
maxIIDjf(x)1I L
•.
=1n Ilhdl
= l�J�n
maxIIDjf(x)II IIhil
Thus Equation
(1)
defines a bounded linear transformation. LetG(h) = f(x + h) - f(x) - L
;=1n D;/(x)hi
We want to prove that
IIG(h)11
=o(llhll).
For sufficiently smallh, x + h
is in n,and the partial derivatives of
G
exist atx + h.
They areD;G(h)
=D;/(x + II) - D;/(x)
If
c
is a given positive number, we use the assumed continuity ofD;/
atx
tofind a positive 0 such that for
Ilhil
< 0 we haveIIDiG(h)1I
<c,
for1 � i �
n.Then, by the mean value theorem,
IIG(h)11 � IIG(hh h2, · · · , hn) - G(O, h2,
• • •, hn)11 n
+ IIG(O, h2,. · . , hn) - G(O, 0, h3,· · · , hn)11
+ . . .
+ IIG(O,O, .. . ;hn) - G(O,O, .. . ,0)1\
� L
;=1cllhdl = cllhll
Section 3.5 Extremum Problems and Lagrange Multipliers 145 Since c was arbitrary, this shows that
II
C(h)1 I
= o(lI
hlD
.Problems 3.4
•
1. Let U be a closed ball in a Banach space. Let F : U+ --t U, where U+ is an open set containing U. Prove that if sup{IIF'(x)1I : x E U} < 1 , then F has a unique fixed point in U.
2. Let X be a Banach space and 0 an open set in X containing a, b, Xo, and the line segment
S joining a to b. Prove that
IIf(b) - f(a) - !,(xo)(b -a)1
I �
IIb -all
sup 1I!'(x) - !'(xo)1I .:rES
[Suggestion: Use the function g(x) = f(x) - f'(xo)x.J Determine whether the same inequality is true when f'(xo) is replaced by an arbitrary linear operator. In this problem, f : X --t Y, where Y is any normed space.
3. Suppose F(xo, YO) = O. If Xl is close to Xo, there should be a Yl such that F(XI , Yl ) = O.
Show how Newton's method can be used to obtain Yl. (Here F : X x Y --t Z, and X, Y, Z are Banach spaces.)
4. Prove Theorem 2.
5. Let f : n --t Y be a continuously differentiable map, where n is an open set in a Banach space, and Y is a normed linear space. Assume that f'(x) is invertible for each X E ii,
and prove that f(n) is open.
6. Let 0 be the point in [0, 1 J where cos 0 = o. Define X to be the vector space of all continuously differentiable functions on [O, IJ that vanish at the point o. Define a norm on X by writing "xII = sUPO:<::.:<:: 1 Ix'(t)l. Prove that there exists a positive number /j
such that if Y E X and lIyiI <
�
,�
hen there exists an x E X satisfying sin ox + x 0 cos = Y7. Let f be a continuous map from an open set n in a Banach space X into a Banach space Y. Suppose that for some Xo in n, f'(xo) exists and is invertible. Prove that f is one-to-one in some neighborhood of Xo.
8. In Example 2, with the nonlinear integral equation, show that the mapping x H f'(x) is continuous; indeed, it satisfies a Lipschitz condition.
9. Rework Example 2 when the term 2x(0) is replaced by ox(O), for an arbitrary constant o. In particular, treat the case when 0 = O.
3.5 Extremum Problems and Lagrange Multipliers
A minimum
point of a real-valued function f defined on a set Af is a point Xo such that f(xo)�
f(x) for all x EM.
IfM
has a topology, then the concept ofrelative
minimum point is defined as a point Xo EM
such that for some neighborhood N of Xo we have f(xo)�
f(x), for all x in N.Theorem 1. Necessary Condition for Extremum. Let n be an open set in a normed linear space, and let f : n -t IR. If Xo is a minimum point of f and if f'(xo ) exists, then f'(xo) =
o.
146 Chapter 3 Calculus in Banach Spaces
Proof.
Let X be the Banach space, and assumef'(xo)
#- 0. Then there existsv
E X such thatf'(xo)v
=- 1.
By the definition off'(xo)
we can takeA >
0and so small that
Xo + AV
is in n andIf(xo + AV) - f(xo) - AJ'(xo)vl / A ll v ll
<(2I1vID- 1
This means that
H f(xo + AV) - f(xo)]
is within distance4
from-1,
and so is negative. This impliesf(xo + AV)
<f(xo).
•In this section we will be concerned mostly with
constrained
extremum problems. A simple illustrative case is the following. We have two nice functions,f
andg,
onJR2
toJR.
We putM
={(x, y) : g(x, y)
=O}
and seek an extremum offiM.
(That meansf
restricted toM.)
If the equationg(x, y)
= 0 definesy
as a function of
x,
sayy
=y(x),
then we can look for anunrestricted
extremum of<J>(x)
=f(x, y(x)).
Hence we try to solve the equation<J>'(x)
=O.
This leads to0 =
h(x,y(x)) + f2(x,y(x))y'(x)
=
h (x, y(x)) - f2(x, Y(X))gl (x, Y(X))/g2(X, y(x))
Thus we must solve simultaneously
( 1 ) g(x,y)
= 0 andThe method of Lagrange multipliers introduces the function
H(x, y, A)
=f(x, y) + Ag(X, y)
and solves simultaneously
HI
=H2
=H3
=O.
ThusfI(x,y) + Agl(X,y)
=f2(x,y) + Ag2(X,y)
=g(x,y)
= 0If
g2(X, y)
f 0, thenA
= -f2(x, y)/g2(X,
V), and we recover system( 1).
Themethod of Lagrange multipliers treats
x
andy
symmetrically, and includes both cases of the implicit function theorem. Thusy
can be a differentiable function ofx,
orx
can be a differentiable function ofy.
Example 1. Let
f
and 9 be functions fromJR2
toJR
defined byf(x, y)
=x2 + y2, g(x, y)
=x - y +
1 . The setM
={(x, y) : g(x, y)
=O}
is the straight line shown in Figure 3.3. Also shown are some level sets off,
i.e., sets of the type((x,y) : f(x,y)
=c}.
At the solution, the gradient off
is parallel to the gradient ofg.
The functionH
isH(x, y, A)
=x2 + y2 + A(X - y + 1),
and thethree equations to be solved are
2 x + A
=2 y - A
=x - y + 1
=O.
The solutionis