In this
hapter we state and prove the
ontra
tion mapping theorem, whi
h is one
of the simplest and most useful methods for the
onstru
tion of solutions of linear
and nonlinear equations. We also present a number of appli
ations of the theorem.
Thus, a
ontra
tion maps points
loser together. In parti
ular, for every x 2 X ,
and any r > 0, all points y in the ball Br (x), are mapped into a ball Bs (T x), with
x Tx
61
62 The Contra
tion Mapping Theorem
s < r. This is illustrated in Figure 3.1. Sometimes a map satisfying (3.1) with
= 1
is also
alled a
ontra
tion, and then a map satisfying (3.1) with
< 1 is
alled
a stri
t
ontra
tion. It follows from (3.1) that a
ontra
tion mapping is uniformly
ontinuous.
If T : X ! X , then a point x 2 X su
h that
T (x) =x (3.2)
is
alled a xed point of T . The
ontra
tion mapping theorem states that a stri
t
ontra
tion on a
omplete metri
spa
e has a unique xed point. The
ontra
tion
mapping theorem is only one example of what are more generally
alled xed-point
theorems. There are xed-point theorems for maps satisfying (3.1) with
= 1,
and even for arbitrary
ontinuous maps on
ertain metri
spa
es. For example,
the S
hauder xed point theorem states that a
ontinuous mapping on a
onvex,
ompa
t subset of a Bana
h spa
e has a xed point. The proof is topologi
al in
nature (see Kantorovi
h and Akilov [27℄), and we will not dis
uss su
h xed point
theorems in this book.
In general, the
ondition that
is stri
tly less than one is needed for the unique-
ness and the existen
e of a xed point. For example, if X = f0; 1g is the dis
rete
metri
spa
e with metri
determined by d(0; 1) = 1, then the map T dened by
T (0) = 1, T (1) = 0 satises (3.1) with
= 1, but T does not have any xed points.
On the other hand, the identity map on any metri
spa
e satises (3.1) with
= 1,
and every point is a xed point.
It is worth noting that (3.2), and hen
e its solutions, do not depend on the
metri
d. Thus, if we
an nd any metri
on X su
h that X is
omplete and T is a
ontra
tion on X , then we obtain the existen
e and uniqueness of a xed point. It
may happen that X is not
omplete in any of the metri
s for whi
h one
an prove
that T is a
ontra
tion. This
an be an indi
ation that the solution of the xed
point problem does not belong to X , but to a larger spa
e, namely the
ompletion
of X with respe
t to a suitable metri
d.
Proof. The proof is
onstru
tive, meaning that we will expli
itly
onstru
t a
sequen
e
onverging to the xed point. Let x0 be any point in X . We dene a
sequen
e (xn ) in X by
xn+1 = T xn for n 0:
To simplify the notation, we often omit the parentheses around the argument of a
map. We denote the nth iterate of T by T n , so that xn = T n x0 .
Fixed points of dynami
al systems 63
d(xn ; xm ) = d(T n x0 ; T m x0 )
m d(T n m x0 ; x0 )
m d(T n m x0 ; T n m 1x0 ) + d(T n m 1 x0 ; T n m 2x0 )
+ + d(T x0 ; x0 )℄
" #
nX
m 1
m k
d(x1 ; x0 )
"
k =0 #
1
X
m
k
d(x1 ; x0 )
k =0
m
1
d(x1 ; x0 );
whi
h implies that (xn ) is Cau
hy. Sin
e X is
omplete, (xn )
onverges to a limit
x 2 X . The fa
t that the limit x is a xed point of T follows from the
ontinuity
of T :
Tx = T lim
n!1 xn = nlim
!1 T xn = nlim
!1 xn+1 = x:
Finally, if x and y are two xed points, then
x_ = f (x); (3.3)
64 The Contra
tion Mapping Theorem
where the dot denotes a time derivative, and f is a ve
tor eld on X . There is little
loss of generality in assuming this form of equation. For example, a se
ond order,
nonautonomous ODE
y = g (t; y; y_ )
may be written as a rst order, autonomous system of the form (3.3) for the state
variable x = (s; y; v ), where s = t and v = y_ , with
s_ = 1; y_ = v; v_ = g (s; y; v ):
For a dis
rete dynami
al system, we may take time t = n to be an integer, and
the dynami
s is dened by a map T : X ! X that relates the state xn+1 at time
t = n + 1 to the state xn at time t = n,
xn+1 = T xn : (3.4)
If T is not invertible, then the dynami
s is dened only forward in time, while if T is
invertible, then the dynami
s is dened both ba
kward and forward in time. A xed
point of the map T
orreponds to an equilibrium state of the dis
rete dynami
al
system. If the state spa
e is a
omplete metri
spa
e and T is a
ontra
tion, then
the
ontra
tion mapping theorem implies that there is a unique equilibrium state,
and that the system approa
hes this state as time tends to innity starting from
any initial state. In this
ase, we say that the xed point is globally asymptoti
ally
stable.
One of the simplest, and most famous, dis
rete dynami
al systems is the logisti
equation of population dynami
s,
xn+1 = 4xn (1 xn ) ; (3.5)
where 0 1 is a parameter, and xn 2 [0; 1℄. This equation is of the form (3.4)
where T : [0; 1℄ ! [0; 1℄ is dened by
Tx = 4x(1 x):
See Figure 3.2 for a plot of x 7! T x, for three dierent values of . We may
interpret xn as the population of the nth generation of a reprodu
ing spe
ies. The
linear equation xn+1 = 4xn des
ribes the exponential growth (if 4 > 1) or de
ay
(if 4 < 1) of a population with
onstant birth or death rate. The nonlinearity in
(3.5) provides a simple model for a spe
ies in whi
h the ee
ts of over
rowding lead
to a de
line in the birth rate as the population in
reases.
The logisti
equation shows that the iterates of even very simple nonlinear maps
an have amazingly
omplex behavior. Analyzing the full behavior of the logisti
map is beyond the s
ope of the elementary appli
ation of the
ontra
tion mapping
theorem we give here.
When 0 1=4, the point 0 is the only xed point of T in [0; 1℄, and T
is a
ontra
tion on [0; 1℄. The proof of the
ontra
tion mapping theorem therefore
Fixed points of dynami
al systems 65
0.9
0.8
0.7
0.6
0.5 µ=1
0.4 µ=.5
0.3
0.2
µ=.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 3.2 The logisti
map for three dierent values of the parameter . For = 0:2, 0 is the only
xed point and it is stable. For = 0:5, 0 is an unstable xed point and there is a nonzero, stable,
xed point. For = 1, the two xed points are unstable and more
omplex (and more interesting)
asymptoti
behavior of the dynami
al system o
urs.
implies that T n x0 ! 0 for an arbitrary initial population x0 2 [0; 1℄, meaning that
the population dies out. When 1=4 < 1, there is a se
ond xed point at
x = (4 1)=4. The appearan
e of a new xed point as the parameter varies
is an example of a bifur
ation of xed points. As in
reases further, there is an
innite sequen
e of more
ompli
ated bifur
ations, leading to
haoti
dynami
s for
0:89 : : :.
As a se
ond appli
ation of the
ontra
tion mapping theorem, we
onsider the
solution of an equation f (x) = 0. One way to obtain a solution is to re
ast the
equation in the form of a xed point equation x = T x, and then
onstru
t approx-
imations xn starting from an initial guess x0 by the iteration s
heme
xn+1 = T xn :
In other words, we are attempting to nd the solution as the time-asymptoti
state
of an asso
iated dis
rete dynami
al system whi
h has the solution as a stable xed
point. Similar ideas apply in other
ontexts. For example, we may attempt to
on-
stru
t the solution of an ellipti
PDE as the time-asymptoti
state of an asso
iated
paraboli
PDE.
There are many ways to rewrite an equation f (x) = 0 as a xed point problem,
some of whi
h will work better than others. Ideally, we would like to rewrite the
equation as a xed point equation in whi
h T is a
ontra
tion on the whole spa
e,
66 The Contra
tion Mapping Theorem
or at least a
ontra
tion on some set that
ontains the solution we seek.
To provide a simple illustration of these ideas, we prove the
onvergen
e of an
p
algorithm to
ompute square roots. If a > 0, then x = a is the positive solution
of the equation
x2 a = 0:
Tx =
pa + (x pa)2 pa (3.7)
2x
p
for all x > 0. Therefore, the restri
tion of T to [ a; 1) is well-dened, sin
e
T [
pa; 1) [pa; 1);
p
and T is a
ontra
tion on [ a; 1) with
= 1=2. It follows that for any x0 a,
p
p
the sequen
e xn = T n x0
onverges to a as n ! 1. Moreover, as shown in the
Integral equations 67
where k : [a; b℄ [a; b℄ !R and g : [a; b℄ !R are given fun
tions. A Fredholm
integral equation of the rst kind is an equation of the form
Z b
k (x; y )f (y ) dy = g (x):
a
The integral equation (3.8) may be written as a xed point equation T f = f , where
the map T is dened by
Z b
T f (x) = g (x) + k (x; y )f (y ) dy: (3.9)
a
Proof. We prove this result by showing that, when (3.10) is satised, the map T
is a
ontra
tion on the normed spa
e C ([a; b℄) with the uniform norm k k1 .
68 The Contra
tion Mapping Theorem
From Theorem 2.4, the spa
e C ([a; b℄) is
omplete. Moreover, T is a
ontra
tion
sin
e, for any f1 ; f2 2 C ([a; b℄), we have
Z
b
kT f1 T f2 k1 = sup k (x; y )(f1 (y )
f2 (y )) dy
axb a
Z b
sup j
k (x; y ) f1 (y ) jj j
f2 (y ) dy
a x b a
(Z )
b
kf1 f2 k1 sup jk(x; y)j dy
a x b a
kf1 f2 k1 ;
where
(Z )
b
= sup jk(x; y)j dy < 1:
a x b a
The result then follows from the
ontra
tion mapping theorem.
From the proof of the
ontra
tion mapping theorem, we
an obtain the xed
point f as a limit,
= lim n
(3.11)
f
n !1 T f0 ;
for any f0 2 C ([a; b℄). It is interesting to reinterpret this limit as a series. We dene
a map K : C ([a; b℄) ! C ([a; b℄) by
Z b
Kf = k (x; y )f (y ) dy:
a
The map K is
alled a Fredholm integral operator, and the fun
tion k is
alled the
kernel of K . Equation (3.8) may be written as
(I K )f = g; (3.12)
where I is the identity map, meaning that If = f . The
ontra
tion mapping T is
given by T f = g + Kf , whi
h implies that
T n f0 = g + K (g + : : : + K (g + Kf0 ))
= g + Kg + : : : + K n g + K n+1 f0 :
This series is
alled the Neumann series. The use of the partial sums of this series
to approximate the inverse is
alled the Born approximation . Expli
itly, we have
I + K + K 2 + : : : f (x)
Z b Z bZ b
= f (x) + k (x; y )f (y ) dy + k (x; y )k (y; z )f (z ) dydz + : : : :
a a a
Proposition 3.4 Let f : [0; 1℄ ! R be a
ontinuous fun
tion. The unique solution
v of the boundary value problem
v 00
= f; (3.20)
v (0) = 0; v (1) = 0; (3.21)
is given by
Z 1
v (x) = g (x; y )f (y ) dy;
0
Boundary value problems for dierential equations 71
where
x(1 y) if 0 x y 1,
g (x; y ) = (3.22)
y (1 x) if 0 y x 1.
Proof. Integrating (3.20) twi
e, we obtain that
Z xZ y
v (x) = f (s) ds dy + C1 x + C2 ;
0 1
where C1 and C2 are two real
onstants to be determined later. Integration by parts
of the integral with respe
t to y in this equation gives
Z y x Z x
v (x) = y f (s) ds + yf (y ) dy + C1 x + C2
1 0 0
Z 1 Z x
= x f (y ) dy + yf (y ) dy + C1 x + C2 :
x 0
We determine C1 and C2 from (3.21), whi
h implies that
Z 1
C1 = yf (y ) dy; C2 = 0:
0
It follows that
Z x Z 1
v (x) = y (1 x)f (y ) dy + x(1 y )f (y ) dy;
0 x
max u(x) 0:
0x1
Similarly, at an interior minimum, we have u00 0, so u 0, and therefore
min
0x1
u(x) 0:
It follows that u = 0, so the solution of the boundary value problem is unique.
Generalizations of this maximum prin
iple argument apply to s
alar ellipti
partial
dierential equations, su
h as Lapla
e's equation.
Initial value problems for dierential equations 73
Proof. We will show that T is a
ontra
tion on the spa
e of
ontinuous fun
tions
dened on a time interval t0 t t0 + Æ , for suÆ
iently small Æ . Suppose that
u; v : [t0 ; t0 + Æ ℄ ! Rn are two
ontinuous fun
tions. Then, from (3.29) and (3.30),
74 The Contra
tion Mapping Theorem
we estimate
kT u Tv k1 = sup kT u(t) k
T v (t)
+
t0 t t0 Æ
Z t
= sup
f (s; u(s)) f (s; v (s)) ds
t0 tt0 +Æ t0
Z t
sup k
f (s; u(s)) k
f (s; v (s)) ds
+
t0 t t0 Æ t0
Z t
sup k
C u(s) k
v (s) ds
+
t0 t t0 Æ t0
CÆ u k v k1 :
It follows that if Æ < 1=C , then T is a
ontra
tion on C ([t0 ; t0 + Æ ℄). Therefore,
there is a unique solution u : [t0 ; t0 + Æ ℄ ! Rn . The argument holds for any t0 2 I ,
and by
overing I with overlapping intervals of length less than 1=C , we see that
(3.25) has a unique
ontinuous solution dened on all of I . The same proof applies
for times t0 Æ < t < t0 .
We may have I = R in this theorem, in whi
h
ase the solution exists globally.
Example 3.8 Linear ODEs with
ontinuous
oeÆ
ients have unique global solu-
tions. Sin
e higher order ODEs may be redu
ed to rst-order systems, it is suÆ
ient
to
onsider a rst-order linear system of the form
u_ (t)
= A(t)u(t) + b(t);
u(0) = u0 ;
Therefore,
kA(t)u A(t)v k C ku k
v ;
for all u; v 2 Rn and t 2 I , so the hypotheses of Theorem 3.7 are satised, and we
have a unique
ontinuous solution on I . Sin
e I is an arbitrary interval, we
on
lude
that there is a unique
ontinuous solution for all t 2 R.
The appli
ations of Theorem 3.7 are not limited to linear ODEs.
Initial value problems for dierential equations 75
ju vj p 2 ju2j + p
jv j
a + u + a2 + v 2
ju v : j
Theorem 3.7 implies that there is a unique global solution of this ODE.
The global Lips
hitz
ondition (3.30) plays two roles in Theorem 3.7. First, it
ensures uniqueness, whi
h may fail if f is only a
ontinuous fun
tion of u. Se
ond,
it implies that f does not grow faster than a linear fun
tion of u as kuk ! 1. This
is what guarantees global existen
e. If f is a nonlinear fun
tion, su
h as f (u) = u2
in Example 2.22, that satises a lo
al Lips
hitz
ondition but not a global Lips
hitz
ondition, then \blow-up" may o
ur, so that the solution exists only lo
ally.
The above proof may be modied to provide a lo
al existen
e result.
Hen
e T u 2 X so that T : X ! X.
Se
ond, we estimate
Z t
kT u Tv k1 = sup
[f (s; u(s)) f (s; v (s))℄ ds
jt t0 j t0
C u k v k1 ;
where C is a Lips
hitz
onstant for f . Hen
e if we
hoose = 1=(2C ) then T is a
ontra
tion on X and it has a unique xed point.
Sin
e depends only on the Lips
hitz
onstant of f and on the distan
e R of
the initial data from the boundary of B R (u0 ), repeated appli
ation of this result
gives a unique lo
al solution dened for jt t0 j < Æ .
A signi
ant feature of this result is that if f (t; u) is
ontinuous for all t 2 R,
then the existen
e time Æ only depends on the norm of u. Thus, the only way in
whi
h the solution of an ODE
an fail to exist, assuming that the ve
tor eld f is
Lips
hitz
ontinuous on any ball, is if ku(t)k be
omes unbounded. There are many
fun
tions that are bounded and
ontinuous on an open interval whi
h
annot be
extended
ontinuously to R; for example, u(t) = sin(1=t) is bounded and
ontinuous
on ( 1; 0) but has no
ontinuous extension to ( 1; 0℄. This kind of behavior
annot happen for solutions of ODEs with
ontinuous right-hand sides, be
ause the
derivative of the solution
annot be
ome large unless the solution itself be
omes
Initial value problems for dierential equations 77
large. If we
an prove that for every T > 0 any lo
al solution satises an a priori
estimate of the form
ku(t)k R for jtj T ;
then the lo
al existen
e theorem implies that the lo
al solution
an be extended to
the interval ( T; T ), and hen
e to a global solution.
3.6 Referen
es
General referen
es for this
hapter are Simmons [50℄ and Marsden and Homan [37℄.
For more about the logisti
map and
haoti
dynami
al systems, see Devaney [8℄,
Gu
kenheimer and Holmes [18℄, and S
hroeder [51℄. Hirs
h and Smale [21℄ dis
usses
gradient
ows.
jT (x) j j
T (y ) < x y j for all x 6= y 2 R :
Why doesn't this example ontradi t the ontra tion mapping theorem?
Exer
ise 3.2 (a) Show that for any y > 0, the
onvergen
e of T n y to x is
p
in fa
t faster
pxjthan exponential. Start from (3.7) to obtain a good estimate
for jT y .
(b) Take y = 1 and p x = 2. Find a good estimate of how large n needs to be
for T 1 and 2 to have identi
al rst d digits. For example, how many
n
p
iterations of the map T are suÆ
ient to
ompute the rst 63 digits of 2?
Exer
ise 3.3 The se
ant method for solving an equation f (x) = 0, where f : R !
R, is a variant of Newton's method. Starting with two initial points x0 , x1 , we
ompute a sequen
e of iterates (xn ) by:
xn+1 = xn f (xn )
xn xn 1 for n = 1; 2; 3; : : :
f (xn ) f (xn 1)
Formulate and prove a
onvergen
e theorem for the se
ant method with f (x) =
x2 2.
Exer
ises 79
Exer
ise 3.4 If T satises (3.1) for
< 1, show that we
an estimate the distan
e
of the xed point x from the initial point x0 of the xed point iteration by
1
d(x; x0 ) d(x ; x ):
1
1 0
Exer
ise 3.5 An n n matrix A is said to be diagonally dominant if for ea
h row
the sum of the absolute values of the o-diagonal terms is less than the absolute
value of the diagonal term. We write A = D L U where D is diagonal, L is
lower triangular, and U is upper triangular. If A is diagonally dominant, show that
kL + U k1 < kDk1 ;
where the 1-norm kk1 of a matrix is dened in (5.9). Use the
ontra
tion mapping
theorem to prove that if A is diagonally dominant, then A is invertible and that the
following iteration s
hemes
onverge to a solution of the equation Ax = b:
xn+1 = D 1 (L + U ) xn + D 1 b; (3.33)
1
xn+1 = (D L) U xn + (D L) b:
1 (3.34)
What
an you say about the rates of
onvergen
e?
Su
h iterative s
hemes provide an eÆ
ient way to
ompute numeri
al solutions
of large, sparse linear systems. The iteration (3.33) is
alled Ja
obi's method, and
(3.34) is
alled the Gauss-Seidel method.
Exer
ise 3.6 The following integral equation for f : [ a; a℄ ! R arises in a model
of the motion of gas parti
les on a line:
Z a
1 1
f (x) =1+
1 + (x y )2
f (y ) dy for a x a:
a
Prove that this equation has a unique bounded,
ontinuous solution for every 0 <
a < 1. Prove that the solution is nonnegative. What
an you say if a = 1?
Exer
ise 3.7 Prove that there is a unique solution of the following nonlinear BVP
when the
onstant is suÆ
iently small,
u00 + sin u = f (x);
u(0) = 0; u(1) = 0:
Here, f : [0; 1℄ ! R is a given
ontinuous fun
tion. Write out the rst few iterates
of a uniformly
onvergent sequen
e of approximations, beginning with u0 = 0.
Hint. Reformulate the problem as a nonlinear integral equation.