Chapter 6

Existence and regularity of

solutions

In this chapter we take up certain matters of mathematical principle: exis-

tence of solutions on an interval, continuation of this interval of existence,
continuity of solutions with respect to initial data and parameters, and re-
lated matters. The linear equations that we have mostly treated up to the
present will appear here only as special cases. It is particularly useful to for-
mulate these general problems in terms of systems of first-order equations.
The general differential equation of order n, written in standard form, is
 
u(n) = f x, u, u0 , . . . , u(n−1) , (6.1)

where, as usual, u(k) signifies the kth derivitave with respect to the inde-
pendent variable x. We can only expect this equation to have a unique
solution if we supply, in addition to the equation, other data specifying the
value of u and its first n − 1 derivatives at some value x0 of the independent
variable. Moreover, in general, we can only expect that a solution exists on
some interval, possibly small, containing the point x0 , as examples show.

Example 6.0.3 The first-order equation u0 = 2xu2 with initial data u (0) =

−1
u0 has the solution u = u0 1 − u0 x2 . If u0 > 0 this becomes infinite at
−1/2
x = ±u0 , which defines the interval of existence for this solution. 

Unless otherwise indicated, we shall assume in this chapter that the vari-
ables, both dependent and independent, are real.

131
132

The standard technique for converting the nth order equation (6.1) to a
system of n first-order equations is to write

u1 = u, u2 = u0 , . . . , un−1 = u(n−2) , un = u(n−1) . (6.2)

Denoting the column vector with components u1 , . . . , un by U , we then

replace equation (6.1) by the system
 0   
u1 u2
 u0   u3 
 2  
U0 ≡  .  =   = F (x, U ) . (6.3)

..
 ..   . 
u0n f (x, u1 , u2 , . . . , un )
In the present chapter, we consider first-order systems of the form

U 0 = F (x, U ) (6.4)

where the vector field F appearing on the right-hand side need not have the
special form of equation (6.3), but can have components each of which is an
arbitrary function of the n + 1 variables x, u1 , . . . , un . Conclusions drawn
for such a general system are then necessarily valid for the special system
(6.3), from which they can be immediately translated into the language of
the nth-order equation (6.1).
Among the advantages of the system approach is that its treatment is
independent of n, and therefore its treatment for arbitrary n closely mimics
that of a single first-order equation, n = 1. We therefore investigate this
relatively simple case first.

6.1 The first-order equation

Consider then the initial-value problem

y 0 = f (x, y) , y (x0 ) = y0 (6.5)

where x and y are real variables. We shall assume that the function f is
continuous in a domain1 D of the xy-plane, and satisfies a Lipschitz condi-
tion there, as defined in Chapter 1, equation (1.49). The basic conclusion,
sometimes called local existence theory, is that there is an interval I of the
x axis, possibly very small, containing the point x0 , on which there is a
solution y (x). We address later the issue of extending this local solution to
a longer interval.
1
By a domain we understand a connected, open set.
 133

6.1.1 Existence of Solutions

The initial-value problem (6.5) is equivalent to the integral equation
Z x
y (x) = y0 + f (s, y (s)) ds (6.6)
x0

in the sense that any continuous solution of the latter is a solution of the
former, and conversely. The plan is to construct a solution of the latter.
A number of ideas from real analysis are used in the proof of the basic
existence theorem below. One of the most important of these is idea of
uniform convergence of a sequence of functions. For a sequence of real
numbers, ak , say, the idea of convergence is expressed as follows: there is
a real number a (the limit of the sequence) with the property that, given
any positive number , there is an integer N such that |an − a| <  for all
n > N . The notion of the convergence of a sequence of functions uk , each
defined on some interval I, to a function u on that interval is the same: for
each x ∈ I, the sequence of real numbers {uk (x)} should converge to the
real number u(x). In this convergence, the number N in general depends on
the choice of x ∈ I. A sequence of functions {uk } of functions defined on I
is said to converge uniformly on I if it converges to some function u there
and, given  > 0, it is possible to find an integer N such that, for any x ∈ I,
|un (x) − u (x)| <  if n ≥ N. The uniformity of the convergence consists in
that the choice of N depends only on  and does not depend on x.
P The definition of uniform convergence for an infinite series of functions
vn (x) follows by considering the sequence of partial sums of the series.
A convenient criterion for the uniform convergence of the infinite
P series on
a set S is the existence of a convergent series of constants Mn such that
|vn (x)| ≤ Mn for each x in S. This criterion, the so-called Weierstrass
M -test, may be found (for example) in the texts by Taylor2 or Rudin3 .
Another idea used below is that of uniform continuity. A function u (x)
is uniformly continuous on an interval I if, given  > 0, there is δ > 0 such
that |u (x) − u (y)| <  for all points x, y ∈ I such that |x − y| < δ. Here
the uniformity consists in that the choice of δ depends only on  and does
not depend on x or y.
Neither of these notions is restricted to functions of a single real variable,
but have immediate generalizations to functions of several variables.
Standard theorems used below, for which we refer the reader to textbooks
on real analysis (or advanced calculus), are the following:
2
Advanced Calculus
3
Principles of Mathematical Analysis
134

• Theorem 6.1.1 If yn (x) → y (x) uniformly on an interval I and if

each function yn is continuous there, then so is the limit function y.

• Theorem 6.1.2 To the assumptions in theR preceding theorem

R adjoin
the assumption that I is bounded. Then lim I yn (x) dx → I y (x) dx.

• Theorem 6.1.3 Suppose the function f (x, y) is continuous on a closed,

bounded set S. Then it is uniformly continuous on S, and attains a
maximum and a minimum value there.

We are now ready to state and prove the existence theorem.

Theorem 6.1.4 Suppose f is defined and continuous in the domain D ⊂

R2 and satisfies there the Lipschitz condition (1.49). Suppose further that
(x0 , y0 ) ∈ D. Then there is an interval I of the x axis containing x0 on
which the initial-value problem (6.5) has a unique solution.

Proof: We seek a continuous solution of the equivalent integral equation

(6.6) in a rectangle R lying in D. We define R by the inequalities x0 − a ≤
x ≤ x0 + a, y0 − b ≤ y ≤ y0 + b where a and b are chosen small enough
that this rectangle lies entirely in D. This is possible since D is an open set
containing the point (x0 , y0 ). Since R is closed and bounded, the function
|f | has a maximum value M there by Theorem 6.1.3. We now define a
sequence of functions {yn (x)} on an interval I : x0 − α ≤ x ≤ x0 + α where

α = min (a, b/M ) ,

as follows:
Z x
y0 (x) = y0 , yn (x) = y0 + f (s, yn−1 (s)) ds , n = 1, 2, . . . . (6.7)
x0

Two things need to be checked: 1) the graphs of the approximating functions

remain in D and 2) the sequence converges to a solution of equation (6.6).
To verify that the graphs of the solutions remain in D, we shall check
that in fact they remain in R on the interval I. This is clearly true for y0 (x).
For y1 note that, if x > x0 ,
Z x Z x
|y1 (x) − y0 | ≤ |f (s, y0 )| ds ≤ M ds = M (x − x0 ) ≤ b, (6.8)
x0 x0

since x − x0 ≤ α ≤ b/M. In the case when x < x0 the integral above would
taken from x to x0 and the conclusion would be the same except that x − x0
 135

should be replaced by |x − x0 |. Thus for x ∈ I, y0 − b ≤ y1 (x) ≤ y0 + b. This

means that the graph (x, y1 (x)) lies in R and hence in D. Now suppose the
graph of yn−1 lies in R when x ∈ I. Then (again supposing x > x0 )
Z x
|yn (x) − y0 | ≤ |f (s, yn−1 (s))| ds.
x0

Since, by the assumption on yn−1 (the induction hypothesis), the argument

of the function f in the preceding integral lies in R for each x ∈ I, the
integrand is bounded by M and we conclude, just as we did for the function
y1 , that the graph of yn remains in R as long as x ∈ I. By the principle of
mathematical induction, this proves that the graph of each approximating
function yn remains in R on that interval.
We next turn to the proof that this sequence converges to a solution of
equation (6.6). As already noted in inequality (6.8), for x0 ≤ x ≤ x0 + α,
|y1 (x) − y0 | ≤ M (x − x0 ). Therefore

Z
|y2 (x) − y1 (x)| ≤ |f (s, y1 (s)) − f (s, y0 (s))| ds

(x − x0 )2
Z x Z x
≤L |y1 (s) − y0 (x)| ds ≤ LM (s − x0 ) ds = M L ,
x0 x0 2

where L is the Lipschitz constant (cf. the definition 1.4.2). We propose the
induction hypothesis that on the interval [x0 , x0 + α]

M (L (x − x0 ))n+1
|yn+1 (x) − yn (x)| ≤ . (6.9)
L (n + 1)!

We have verified this for the cases n = 0 and n = 1. Assume it’s true for a
value n − 1. Then we find

(s − x0 )n
Z x Z
|yn+1 (x) − yn (x)| ≤ L |yn (s) − yn−1 (s)| ds ≤ L M Ln−1 ds
x0 n!
n (x− x0 )n+1
= ML ,
(n + 1)!

which shows that the inequality (6.9) holds for all n = 0, 1, . . ..

Consider now the function yn expressed as follows:

yn (x) = y0 (x) + (y1 (x) − y0 (x)) + · · · + (yn (x) − yn−1 (x)) .

136

P
The sequence {yn (x)} will converge if the corresponding series (yn+1 (x) − yn (x))
converges. But the latter is dominated by the series of positive constants
(M/L) (Lα)n /n! which is the convergent expansion for (M/L) exp (Lα) .
P
The sequence {yn (x)} therefore converges uniformly on the interval [x0 , x0 +
α] by the Weierstrass M -test. The restriction to this subinterval was for con-
venience in expressing the estimates; if we restrict to the interval [x0 − α, x0 ]
we find the same estimates, so we conclude that this sequence is uniformly
convergent on the entire interval I. We denote its limit function by y (x).
It is continuous by Theorem 6.1.1, and it satisfies the condition
Z x
y (x) = y0 + lim f (s, yn (s)) ds.
n→∞ x
0

The limit can be taken inside the integral if the hypotheses of Theorem
6.1.2 hold. But the integrand Fn (s) ≡ f (s, yn (s)) converges uniformly to
f (s, y (s)) by virtue of the continuity of the function f 4 . This verifies the
hypotheses of that theorem, and shows that the function y (x) so constructed
satisfies (6.6).
The uniqueness of this solution was established in Chapter 1 as Theorem
1.4.2. 
In this theorem, the function f need not satisfy a Lipschitz condition on
the entire domain D. It suffices if it satisfies this condition on each closed
rectangle R ⊂ D since only this assumption was used in the proof.

6.1.2 Continuation of Solutions

The solution guaranteed by this local theorem can be extended to a maximal

interval of existence. Consider the set of all solutions of the initial-value
problem (6.5) defined on intervals [x0 , ξ) with ξ > x0 . The union of these
intervals is a maximally extended interval [x0 , b). At any point x of this
interval take y (x) to be the value of any of the solutions defined on any
of the intervals including the point x; this is well defined by virtue of the
uniqueness theorem.
One possibility is that b = +∞, so consider instead the case when b
is finite. It is possible that y(x) is unbounded as x → b. If instead y(x)
is bounded as x → b, consider a sequence {xn } such that xn → b but
xn < b. Since the sequence {(xn , y(xn ))} is likewise bounded, there is a con-
vergent subsequence5 . Let the latter converge to (b, yb ). Since the sequence
4
See the next set of exercises.
5
This is a consequence of the Bolzano-Weierstrass theorem of advanced calculus.
 137

{(xn , y(xn ))} lies in D the point (b, yb ) lies either in D or on ∂D. In fact
it must lie on ∂D for, if it were to lie in D, then solving the initial-value
problem with initial data (b, yb ) would extend the interval of existence to
the right of b, contradicting the assumption that b is right-maximal. We can
summarize this as follows:

Theorem 6.1.5 The solution of the initial-value problem (6.5) can be con-
tinued to the right to a maximal interval of existence [x0 , b) such that at least
one of the following holds:
1. b = +∞,
2. y (x) is unbounded as x → b, or
3. (x, y (x)) approaches the boundary of D as x → b.

Of course, the same considerations apply to the left. This provides a maxi-
mally extended interval of existence (a, b) for the solution of the initial-value
problem.
The uniqueness of solutions as stated in Theorem 6.1.4 applies to solu-
tions defined on possibly small intervals containing the initial point (x0 , y0 ),
but it extends to solutions on arbitrary intervals: either two solutions dis-
agree at every point of their common interval of existence or they are iden-
tical.

6.2 The system of n first-order equations

In the analysis of the first-order equation we made frequent use of estimates
based on the absolute value. To treat systems like the vector system (6.4)
on a similar footing, we need the “absolute value” of an n-component vector
U . In other words, we need a real number |U | associated with U , positive
when U 6= 0 and zero when U = 0, satisfying the homogeneity condition
|αU | = |α| |U | for any real number α (here |α| represents the absolute value
of α), and further satisfying the triangle inequality

|U + V | ≤ |U | + |V | . (6.10)
Such a real-valued function of the vector U is called a norm. There are
various choices when n > 1. Two popular choices are:
• The Euclidean norm
q
|U | = U12 + U22 + · · · + Un2 . (6.11)
138

• A norm based on absolute values:

|U | = |U1 | + |U2 | + · · · + |Un | (6.12)

where, on the right-hand side, |Ui | signifies the usual absolute value of
a real number.

It can be shown that the various choices of norms in Rn are equivalent in

the sense that, if k · k1 and k · k2 are any two choices, there are positive
constants c and C such that

ckxk2 ≤ kxk1 ≤ Ckxk2 ∀x ∈ Rn .

The Lipschitz condition is now stated in terms of the norm (we assume one
has been chosen). It takes the following form:

Definition 6.2.1 A function F (x, U ) defined on the (n + 1)-dimensional

domain D satisfies a Lipschitz condition there with respect to the variable U
if there is a constant L such that, for each pair (x, U ) and (x, V ) in D,

|F (x, U ) − F (x, V )| ≤ L |U − V | . (6.13)

The equivalent integral equation can be written in the same way as it

was previously written for the case n = 1:
Z x
U (x) = U0 + F (s, U (s)) ds, (6.14)
x0

which is the vector expression of the system of integral equations

Z x
ui (x) = ui0 + Fi (s, u1 (s) , . . . , un (s)) ds
x0

for i = 1, 2, . . . , n. It is a straightforward matter to show that the estimate

used for the absolute value of a single integral on an interval I,
Z  Z
 
 f (x) dx ≤ |f (x)| dx
 
I I

continues to hold if the real function f is replaced by a vector U and the

absolute value is replaced by the norm (see Exercise 12 below).
Finally consider Theorems 6.1.1, 6.1.2, and 6.1.3 from advanced calculus,
which were used in the proof of Theorem 6.1.4. The first two of these
are unchanged if the real-valued functions are interpreted as n-component
 139

vectors. The third is also unchanged if the function f is interpreted as

a real-valued function on Rn+1 ; in Theorem 6.1.4 it is used to bound the
absolute value |f |. In the generalization it is needed only to bound the norm
|F (x, U )| . We are now ready to state the generalization. We consider the
system (6.4) together with initial data:

U 0 = F (x, U ) , U (x0 ) = U0 . (6.15)

Theorem 6.2.1 Suppose F is defined and continuous in the domain D ⊂

Rn+1 and satisfies there the Lipschitz condition (6.13) with respect to U .
Suppose further that (x0 , y0 ) ∈ D. Then there is an interval I of the x
axis containing x0 on which the initial-value problem (6.15) has a unique
solution.

Proof: The proof of existence is virtually word-for-word the same as that for
theorem 6.1.4 if the norm is everywhere substituted for the absolute value;
this proof is therefore omitted. The proof of uniqueness is also virtually
word-for-word the same as that given in Chapter 1 for the case n = 1 again
substituting the norm for the absolute value; this proof is therefore likewise
omitted. 

Theorem 6.2.1 is, like Theorem 6.1.4, a local existence theorem. How-
ever, the argument leading to the continuation of solutions to a maximal
interval of existence, Theorem 6.1.5, is independent of n and carries over to
the present case without modification. We state it, for the record:

Theorem 6.2.2 Under the conditions of Theorem 6.2.1, the solution can be
continued to a maximal interval (a, b). The possible behaviors of the solution
at either of these endpoints are described in Theorem 6.1.5.

In Chapter 2 we took for granted the existence of solutions to linear

initial-value problems. We can now prove Theorem 2.28 of that chapter.
This basic existence theorem for linear equations with continuous coeffi-
cients differs from the more general Theorem 6.2.1 in that it asserts that the
solution exists on the full interval [a, b] where the coefficients are defined and
continuous, whereas the theorem above is local. We suppose the initial-value
problem has been converted to the form (cf. equation 2.37)

U 0 = A (x) U + R (x) , U (x0 ) = U0 . (6.16)

140

The right-hand side of this equation has the form F (x, U ) of the existence
theorem 6.2.1, and we can conclude local existence provided that F is con-
tinuous and satisfies a Lipschitz condition with respect to U . If the matrix-
valued function A has entries that are continuous on the finite interval [a, b],
and if the components of R are likewise continuous there, then the function
F = AU + R is indeed continuous. To verify the Lipschitz condition, note
that

F (x, U ) − F (x, V ) = A (x) (U − V ) . (6.17)

Inasmuch as each entry of the matrix A is continuous on the closed, bounded
interval [a, b], each of these is bounded there by Theorem 6.1.3. It is now
easy to prove (see following Exercises) that there is some constant α such
that

|AU | ≤ α |U | , (6.18)
which shows that F satisfies a Lipschitz condition.
We can now infer from the basic existence theorem that the initial-value
problem (6.16) possesses a solution on some interval, and therefore on a
maximal interval. It remains to show that the maximal interval coincides
with the prescribed interval [a, b] on which the coefficients are defined and
continuous.
Since each component of R is bounded on [a, b] we infer that there is a
positive number β such that

|R (x)| ≤ β (6.19)

for each x in [a, b]. Now consider in place of the initial-value problem (6.16)
the equivalent integral equation
Z x
U (x) = U0 + (A (s) U (s) + R (s)) ds. (6.20)
x0

Taking norms and defining φ (x) = |U (x)| leads to the inequality (for x >
x0 )
Z x Z x
φ (x) ≤ φ (x0 ) + (αφ (s) + β) ds ≤ K + α |U (s)| ds, (6.21)
x0 x0

where K = |U0 | + β (b − a) . Gronwall’s lemma, Lemma 1.4.1, now shows

that
φ (x) = |U (x)| ≤ K exp α (b − a) ,
 141

i.e., U is bounded for all x in the interval [x0 , b]. It now follows from Theorem
6.2.2 that the graph (x, U (x)) extends to the boundary of the domain. Since
the boundaries are x = a and x = b in this case, the right-maximal interval
is [x0 , b).
This argument shows that the solution exists on [x0 , b) but, as it stands,
does not continue the solution to the closed interval: it is possible in principle
for the graph to approach the boundary without approaching a particular
point of the boundary. In fact, for the linear system (6.16), it does approach
a point of the boundary. For, from the equivalent integral formulation we
infer
|U (x̃) − U (x)| ≤ α |x̃ − x| .
Therefore, if {xn } is a sequence such that xn → b, then

|U (xn ) − U (xm )| ≤ α |xn − xm | ,

showing, by Cauchy’s criterion, that the sequence {U (xn )} converges. There-

fore (xn , U (xn )) tends to a point of the boundary. Since {xn } is any sequence
tending to b, it follows that (x, U (x)) tends to a boundary point of D. This
continues the solution to the closed interval [x0 , b]. Clearly a similar argu-
ment holds to the left of x0 and we have

Theorem 6.2.3 Let the matrix A and the vector function R be continuous
on the closed, bounded interval [a, b]. Then the solution of the initial-value
problem 6.16 exists on that interval.

PROBLEM SET 6.2.1

1. Prove the statement following equation (6.6) to the effect that a continuous
solution of this equation is equivalent to a solution of the initial-value problem
(6.5).
2. For the initial-value problem y 0 = λy, y (0) = 1, formulate the equivalent
integral equation and find the first three approximations y0 , y1 , y2 in the
successive-approximations approach to a solution.
3. Same as the preceding exercise for the initial-value problem y 0 = y 2 , y (0) =
1.
Rx
4. Solve the integral equation u (x) = 1 + 0 s2 u (s) ds.
5. Suppose that f (x, y) is continuous on a closed, bounded set S, and that
the sequence of functions {yn (x)} converges uniformly on an interval I to
a function y (x). Suppose that (x, yn (x)) ∈ S for x ∈ I. Show that the
sequence Fn (x) = f (x, yn (x)) converges uniformly to f (x, y (x)) .
142

6. Give examples of each of the behaviors enumerated in Theorem 6.1.5.

7. Solve explicitly the one-dimensional initial-value problem

u0 = (2x − 1)u2 , , u(1/2) = 1/4

and find the maximal interal of existence.

8. Consider the linear system (n = 2)

u0 = a (x) u + b (x) v, v 0 = c (x) u + d (x) v

where a, b, c, d are defined on an interval I of the s-axis. Eliminate v to obtain

a second-order, linear equation for u. You may take for granted any amount
of differentiability that you need. Are any other conditions on the coefficients
needed?

9. Take n = 2 and denote the norm (6.11) by k · k and the norm (6.12) by | · |.
Show that these norms are equivalent in the sense that there exist positive
constants a, b, c, d such that, for any vector U = (U1 , U2 ),

akU k ≤ |U | ≤ bkU k and c|U | ≤ kU k ≤ d|U |.

10. Let a norm k · k be given for vectors U ∈ Rn . Let A be a linear operator

from Rn to Rn . Define kAk = supkU k=1 kAU k. Show that kAk possesses the
defining properties for a norm as given in the first paragraph of §6.2

11. Prove that each of the two norms (6.11) and (6.12) satisfies the triangle
inequality.

12. Using the norm 6.12, show that, for a < b

Z  Z
 b  b
U (x) dx ≤ |U (x)| dx.
 

 a  a

13. Consider the initial-value problem

   
w w2 w3
d  1  
ẇ = w2 = −w3 w1  , w (0) = w0 . (6.22)
dt
w3 −µw1 w2

For µ > 0 and arbitrary initial-value vector w0 , show that solutions exist on
(−∞, +∞).
Hint: Find a positive function of w1 ,w2 and w3 that is constant on solutions.
 143

6.3 Continuous dependence

The solution U of the initial-value problem (6.15) depends not only on the
current value of the independent variable x but also on the initial data
U0 . Moreover, the uniqueness theorem ensures that, for a given value of
x, solutions with distinct values of U0 must have distinct values of U , and
conversely, distinct values of U require distinct values of U0 . The solution
may be written U = U (x, U0 ) to emphasize that U is determined by its
initial data, i.e., is a function of its initial data. It is called the solution
map; for a fixed value of x, it maps inital data U0 to U . The initial data
should include the specification of the initial point x0 as well, but in this
section we shall keep x0 fixed.
The uniqueness theorem shows that the solution map is an invertible
function of U0 , but it does not show that it possesses the properties of
continuity, differentiability, etc. that are so useful in analysis. Continuity of
this function is relatively easy to prove; we turn to this next.
Consider, in addition to the initial-value problem (6.15), a second initial-
value problem
V 0 = G (x, V ) , V (x0 ) = V0 (6.23)
where the domain of G is the same as that of F ; it is to be solved on the same
interval [a, b] as the problem (6.15). From the equivalent integral equations
we derive
Z x
U (x) − V (x) = U0 − V0 + (F (s, U (s)) − G (s, V (s))) ds.
x0

We rewrite the integrand as F (s, U ) − F (s, V ) + F (s, V ) − G (s, V ), and

take norms:
Z x
kU (x) − V (x)k ≤ kU0 − V0 k + kF (s, U (s)) − F (s, V (s))k ds
x0
Z x
+ kF (s, V (s)) − G (s, V (s))k ds.
x0

We consider this on an interval such that the graphs of both solutions lie in
D. Let F satisfy a Lipschitz condition there with Lipschitz constant L, and
suppose the difference F (x, U ) − G (x, U ) is bounded there, by a constant
M . Then
Z x
kU (x) − V (x)k ≤ (kU0 − V0 k + M (b − a)) + L kU (s) − V (s)k ds.
x0
144

By Gronwall’s inequality, we have

kU (x) − V (x)k ≤ (kU0 − V0 k + M (b − a)) exp (L (b − a)) . (6.24)

Suppose first that G = F and therefore we can choose M = 0 in equation

(6.24). This formula shows then that

kU (x, U0 ) − U (x, V0 )k ≤ K |U0 − V0 | ,

where K is a constant depending only on the differential equation and the

interval over which its solution is considered. This latter formula clearly
shows that the function U (x, U0 ) is a continuous (in fact, Lipschitz contin-
uous) function of the initial data. We therefore have proved the following:
Theorem 6.3.1 Suppose the problem (6.15) has a solution on an interval
[a, b] for all initial data U0 sufficiently close to V0 . Assuming that F satisfies
a Lipschitz condition in D, we find that the solution map U (x, U0 ) is a
continuous function of U0 .
Here the phrase “U0 sufficiently close to V0 ” is meant in the sense of the cho-
sen norm, namely that there is some positive number δ such that kU0 − V0 k <
δ.
Next consider the case when the vector field depends on a parameter µ,
i.e., F = F (x, U, µ), where (x, U ) ∈ D and µ ∈ J, where J is some open
set. In treating this case we shall suppose for simplicity that the initial
data are fixed so that the first term on the right-hand side of the inequality
(6.24) vanishes and we may concentrate on the effect of varying µ. The
solution U = U (x, µ) will then depend on µ. In the estimate (6.24) we may
take G (x) = F (x, µ̃). The function F is a uniformly continuous function
of its arguments on closed and bounded subsets of D × J, so the difference
kF (x, V (x), µ) − F (x, V (x), µ̃)k may be taken arbitrarily, small, uniformly
for a ≤ x ≤ b, by choosing kµ − µ̃k sufficiently small, i.e., the bound M may
be chosen arbitrarily small if we take kµ − µ̃k suficiently small. This implies
that the solution map is a continuous function of parameters. Applying
again the inequality (6.24), we find

kU (x, µ) − U (x, µ̃)k ≤ M K,

where K = (b − a) exp (L(b − a)) and M → 0 as µ̃ → µ. This proves

Theorem 6.3.2 Consider the initial-value problem

U 0 = F (x, U, µ) , U (x0 ) = U0 . (6.25)

 145

Assume that F is continuous on D×J and satisfies a Lipschitz condition with

respect to U with Lipschitz constant L. Assume further that the solution U
exists on [a, b] for all µ sufficiently near µ0 . Then (for fixed U0 ) the solution
U (x, µ) is a continuous function of µ at µ0 .

6.4 Differentiablity
In order to be able to infer that the solution map is not only a continuous but
also a differentiable function of the initial data or of parameters, we must
place further requirements on the function f of the initial-value problem
(6.5) (in the case n = 1) or of the function F of problem (6.15) (for the
general case). We consider the treatment of the case n = 1, considering
only the differentiability with respect to initial data (but see the problem
set for an extension to differentiability with respect to parameters). Let
y = φ (x, y0 ) be the solution map. It satisfies

φx (x, y0 ) = f (x, φ (x, y0 )) , φ (x0 , y0 ) = y0 , (6.26)

where we have used a partial derivative to emphasize that we now regard

φ as a function of two variables. Formally differentiate either side of this
equation with respect to y0 :

φxy0 (x, y0 ) = fy (x, φ (x, y0 )) φy0 (x, y0 ) , φy0 (x0 , y0 ) = 1.

If these procedures are justified, the function φy0 satisfies the following linear
initial-value problem:

v 0 = a (x) v, v (x0 ) = 1, (6.27)

with a (x) = fy (x, φ (x, y0 )) . We are thus led to a linear equation for the
derivative (this is called the variational equation). Of course, for n = 1 we
can obtain a formula for its solution, but this is not our goal. Our goal is
to justify this procedure. The first observation is that it is not sufficient to
assume that f is continuous in D and satsifies a Lipschitz condition there.
We have taken a partial derivative with respect to y. We therefore make the
stronger assumption that f has a continuous partial derivative with respect
to y in the domain D.
With the aid of this assumption we can complete the proof that φy0
exists, and is equal to the solution of the variational equation, along the
following lines. Denote by ∆φ the difference

∆φ (x, y0 , h) = φ (x, y0 + h) − φ (x, y0 )

146

and by q the difference quotient

q (x, y0 , h) = ∆φ (x, y0 , h) /h.

The object is to show that the limit of q exists as h → 0 and equals the
solution v of the variational equation (6.27). We proceed in two steps, of
which the first is to show that q is bounded as h → 0.
For ∆φ we have the integral equation
Z x
∆φ (x, y0 , h) = h + [f (s, φ (s, y0 + h)) − f (s, φ (s, y0 ))] ds
x0
Z x
=h+ [f (s, φ (s, y0 ) + ∆φ (s, y0 , h)) − f (s, φ (s, y0 ))] ds
y0
Z x
=h+ fy (s, φ (s, y0 ) + θ∆φ) ∆φ ds.
y0

In the last term above, we have used the mean-value theorem. The number
θ = θ (s, y0 , h) lies in the interval (0, 1). We have in this last term suppressed
the arguments of the function ∆φ = ∆φ (s, y0 , h) . From the uniform conti-
nuity of φ (x, y) we know that we may choose |∆φ| as small as we please,
by choosing h sufficiently small. Since by assumption fy is continuous, it
is bounded on any compact subset. We may confine its arguments to a
compact susbset of the domain D by choosing h small enough, and may
therefore assume that |fy | < M for some positive number M . Therefore
Z x
|∆φ| ≤ |h| + M |∆φ (s, y0 , h)| ds,
y0

or, by Gronwall’s lemma,

|∆φ (x, y0 , h)| ≤ |h| exp {M (b − a)} .

This shows that ∆φ tends to zero like h or, equivalently, that q is bounded
as h → 0.
Now form the difference

q (x, y0 , h) − v (x)
Z x 
f (s, φ (s, y0 + h)) − f (s, φ (s, y0 ))
= − fy (s, φ (s, y0 )) v (s) ds
x0 h
Z x Z x
= q [fy (s, φ + θ∆φ) − fy (s, φ)] ds + fy (s, φ) (q (s) − v (s)) ds.
y0 y0
 147

The uniform continuity of fy shows that the term in square brackets in the
first integral can be made less than (say)  if h is chosen sufficiently small,
and since q is bounded, we now find from Gronwall’s lemma that

|q (x, y0 , h) − v (x)| ≤ M (b − a) exp {M (b − a)} .

This proves that φy exists and equals v.

By a fairly straightforward extension of this argument, one can infer,
not only for the one-dimensional case but also for the multi-dimensional
case (6.15),
Theorem 6.4.1 Let F and FU be continuous on D and let U = W (x) be
a solution of (6.15) on the interval [a, b]. Then there exists δ > 0 such that,
for all (x0 , U0 ) in the set

Vδ = {(x0 , U0 ) : a < x0 < b, kU0 − W (x0 )k < δ}

the solution U (x, x0 , U0 ) exists on [a, b] and is C 1 on [a, b] × Vδ . The partial

derivative ∂U/∂U0,k satisfies the ‘variational’ initial-value problem
du
= FU (x, U (x, x0 , U0 ))u, u(x0 ) = ek , (6.28)
dx
where ek is a standard basis vector.

In the statement of the theorem above, we have used the notation U (x, x0 , U0 )
for the solution to emphasize the dependence of the solution on the initial
values of both x and U . However, we have only considered the derivative
with respect to U0 . It is natural to inquire whether the dependence on x0
is continuous and differentiable as well under the appropriate conditions on
the vector field F . The answer is yes, but we do not pursue this further (but
see Problem 7 below).
Theorem 6.4.1 can be generalized to the case when the vector field f de-
pends differentiably on parameters. We easily obtain (see Problem 6 below)
Theorem 6.4.2 Consider the initial-value problem (6.25). Assume that F
is continuous on D × J and possesses continuous partial derivatives with
respect to U and µ there. Assume further that, if µ = µ0 , a solution W (x)
exists on [a, b]. Then there exists δ > 0 such that, for all (x0 , U0 ) in the set

Vδ = {(x0 , U0 , µ) : a < t0 < b, kU0 − W (x0 )k + kµ − µ0 k < δ}

the solution U (x, x0 , U0 , µ) exists on [a, b] and is C 1 on [a, b] × Vδ .

148

It is straightforward to generalize the results of this section to cases when the

vector field can be differentiated more than once with respect to parameters
and to the dependent variables; the solutions then possess the same amount
of differentiablity.

PROBLEM SET 6.4.1

1. Consider example 6.0.3 as a function of initial data u0 in a neighborhood of

u0 = 1, on the interval (−1/2, +1/2). Show that it is a continuous function
of u0 .
2. For the initial-value problem (6.15) the solution map should be written
U (x, x0 , U0 ) to allow for variations in the initial point x0 as well as vari-
ations in U0 . Consider the autonomous system U 0 = F (U ) for which the
right-hand side does not depend explicitly on the independent variable x. If
F is defined on a domain Ω ⊂ Rn , then the domain D ⊂ Rn+1 on which the
vector field is defined is D = Ω × R, i.e., it is defined for all x ∈ R.
Denote the solution of this autonomous system with the special initial data
(x0 , U0 ) = (0, U0 ) by U = φ(x, U0 ). Show that, for general initial data
(x0 , U0 ),
U (x, x0 , U0 ) = φ(x − x0 , U0 ).

3. Consider the linear initial-value problem on [−1, 1]:

u00 + p (x) u0 + q (x) u = 0, u (0) = u0 , u0 (0) = u00 .

Choose a basis u1 , u2 such that u1 (0) = 1, u01 (0) = 0 and u2 (0) = 0, u02 (0) =
1. Write the solution u (x, u0 , u00 ). Is it a continuous function of the inital
data? Does it possess partial derivatives with respect to the initial data?
4. Consider the linear initial-value problem on [1, 2] with parameter β:

x2 u00 + xu0 − β 2 u = 0, u (1) = 1, u0 (1) = 0.

Find the solution u (x, β). Is it a continuous function of β? Can it be differ-

entiated with respect to β?
5. Obtain the variational equation for Example 6.0.3, and solve it. Verify that
the result is the same as differentiating the solution map given explicitly in
that example.
6. Prove Theorem 6.4.2 (hint: augment the vector U to a vector Ũ of length
n+m, and augment the vector f to a vector f˜ by adjoining the m-component
zero vector; then apply Theorem 6.4.1).
7. In the notation of Theorem 6.4.1, suppose that U (x, x0 , U0 ) is a differentiable
function of x0 and write v (x) = ∂U/∂x0 . Proceeding formally, obtain an
initial-value problem determining v.
 149

8. Consider the initial-value problem for n = 2

 
dx 0
= A(δ, µ)x, x(0) = ,
dt 1

where  
−δ 1
A(δ, µ) =
0 −µ
and δ and µ are positive parameters.
(a) Solve this problem explicitly when µ = δ.
(b) Solve it when µ 6= δ.
(c) Show explicitly that, for any fixed t > 0, if we take the limit as µ → δ,
the two solutions become the same.
9. In equation (6.22) of Problem 13 of Problem Set 6.2.1, take
 
0
w0 =  0  .
1

Solve equations (6.22) explicitly for µ = 0. Now consider the variational

problem for v = ∂w/∂µ and solve this also for µ = 0, with arbitrary initial
data for v. Use this to give an approximate expression for the solution w
when µ is small but not zero.
314
Bibliography

[1] G. Birkhoff and G.-C. Rota. Ordinary Differential Equations. John

Wiley and Sons, Inc., New York, 1978.

[2] P.F. Byrd and M. D. Friedman. Handbook of elliptic integrals for engi-
neers and physicists. Springer, Berlin, 1954.

[3] Jack Carr. Applications of Centre Manifold Theory. Springer-Verlag,

New York, 1981.

[4] Earl A. Coddington and Norman Levinson. Theory of Ordinary Differ-

ential Equations. New York: McGraw-Hill, 1955.

[5] J. Guckenheimer and P. Holmes. Nonlinear Oscillations, Dynamical

Systems, and Bifurcation of Vector Fields. Springer-Verlag, New York,
1983.

[6] Einar Hille. Lectures on Ordinary Differential Equations. London:

Addison-Wesley Publishing Company, 1969.

[7] M. Hirsch and S. Smale. Differential Equations, Dynamical Systems,

and Linear Algebra. Academic Press, New York, 1974.

[8] E.L. Ince. Ordinary Differential Equations. Dover Publications, New

York, 1956.

[9] S. MacLane and G. Birkhoff. Algebra. New York: Macmillan, 1967.

[10] Jerrold E. Marsden, Anthony J. Tromba, and Alan Weinstein. Ba-

sic Multivariable Calculus. New York: Springer-Verlag:W.H. Freeman,
1993.

[11] I.G. Petrovski. Ordinary Differential Equations. Prentice-Hall Inc.,

Englewood Cliffs, 1966.

315
316

[12] Walter Rudin. Principles of Mathematical Analysis. McGraw-Hill, New

York, 1964.