Oridinary Differential Equations by Raz Kupferman

Ordinary Dierential Equations
Raz Kupferman
Institute of Mathematics
The Hebrew University
June 26, 2012
2
Contents
1 Existence and uniqueness 3
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Eulers approximation scheme . . . . . . . . . . . . . . . . . . . . . 9
1.3 The Cauchy-Peano existence proof . . . . . . . . . . . . . . . . . . . 11
1.4 Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5 The Picard-Lindl of existence proof . . . . . . . . . . . . . . . . . . 20
1.6 Continuation of solutions . . . . . . . . . . . . . . . . . . . . . . . . 22
1.7 Generalized solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.8 Continuity of solutions with respect to initial conditions . . . . . . 23
2 Linear equations 25
2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Linear homogeneous systems . . . . . . . . . . . . . . . . . . . . . . 26
2.2.1 The space of solutions . . . . . . . . . . . . . . . . . . . . . 26
2.2.2 Fundamental matrices . . . . . . . . . . . . . . . . . . . . . 30
2.2.3 The solution operator . . . . . . . . . . . . . . . . . . . . . . 35
2.2.4 Order reduction . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3 Non-homogeneous linear systems . . . . . . . . . . . . . . . . . . . 37
2.4 Systems with constant coecients . . . . . . . . . . . . . . . . . . . 40
2.4.1 The Jordan canonical form . . . . . . . . . . . . . . . . . . . 40
2.4.2 The exponential of a matrix . . . . . . . . . . . . . . . . . . 40
2.4.3 Linear system with constant coecients . . . . . . . . . . . 43
ii CONTENTS
2.5 Linear dierential equations of order n . . . . . . . . . . . . . . . . 45
2.5.1 The Wronskian . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.5.2 The adjoint equation . . . . . . . . . . . . . . . . . . . . . . 48
2.5.3 Non-homogeneous equation . . . . . . . . . . . . . . . . . . 50
2.5.4 Constant coecients . . . . . . . . . . . . . . . . . . . . . . 52
2.6 Linear systems with periodic coecients: Floquet theory . . . . . . 55
2.6.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.6.2 General theory . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.6.3 Hills equation . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3 Boundary value problems 63
3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2 Self-adjoint eigenvalue problems on a nite interval . . . . . . . . . 67
3.2.1 Denitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2.2 Properties of the eigenvalues . . . . . . . . . . . . . . . . . 69
3.2.3 Non-homogeneous boundary value problem . . . . . . . . . 71
3.3 The existence of eigenvalues . . . . . . . . . . . . . . . . . . . . . . 76
3.4 Boundary value problems and complete orthonormal systems . . . 83
4 Stability of solutions 87
4.1 Denitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.3 Lyapunov functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.4 Invariant manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.5 Periodic solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.6 Index theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.7 Asymptotic limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.8 The Poincar e-Bendixon theorem . . . . . . . . . . . . . . . . . . . . 107
CONTENTS 1
Bibliography
1. Theory of ordinary dierential equations by E.A. Coddington and N. Levin-
son.
2. Ordinary dierential equations by V.I. Arnold.
3. Solving ordinary dierential equations by E. Hairer, N.P Nrsett and G.
Wanner.
2 CONTENTS
Chapter 1
Existence and uniqueness
1.1 Introduction
Denition 1.1 A dierential equation is an equation that relates a function to
its derivative(s). The unknown is the function. A dierential equation is said to be
ordinary ( ) if the function is uni-variate, and more precisely if its domain is
a connected subset of R. We abbreviate ordinary dierential equation into ode.
Example: Find a function y R R that satises the equation
y
= 6y,
or in a dierent notation,
t R y
(t} = 6y(t}.
A number of questions arise right away:

' Does the equation have a solution?
^ If it does, is the solution unique?
What is the solution?
Is there a systematic way to solve such an equation?
4 Chapter 1
In the above example, direct substitution shows that any function of the form
y(t} = ae
6t
, a R is a solution of the equation. Thus, a solution exists and it is not
unique. This is not surprising. We know from elementary calculus that knowing
the derivative of a function does not uniquely determine the function; it does only
up to an integration constant. We could try to obtain a problem for which there
exists a unique solution by imposing an additional condition, like the value of the
function at a certain point, say, y(2} = 8. In this case, the unique solution that
belongs to the above exponential family of solutions is found by setting
y(2} = ae
12
= 8,
i.e., a = 8e
12
. Still, can we be sure that this is the unique solution to the equation
y
= 6y y(2} = 8 ?
We will soon know the answer to this question.
Denition 1.2 An ordinary dierential equation is said to be of order k if the
highest derivative of the unknown function that appears in the equation is the k-th
derivative.
Example: The following equation is a rst-order ode:
y
(t} = t sin
_
3 + y
2
(t},
which we will often write in the more sloppy form:
y
= t sin
_
3 + y
2
.
(It is sloppy because it is not clear in this notation that t is a point in the domain
of y; suppose we wanted to denote the independent variable by x, as we often do.)
First-order ode The most general rst-order ode is of the form

(y
(t}, y(t}, t} = 0.
Such an equation is said to be implicit ( ). We will always assume that
the equation has been brought into explicit form ( ):
y
(t} = f (t, y(t}}.

Existence and uniqueness 5
Second-order ode The most general second-order ode is of the form
(t, y
(t}, y
(t}, y(t}} = 0,
and in explicit form
y
(t} = f (t, y
(t}, y(t}}.
Example: Consider the following equation:
y
(t} + y(t}y
(t} + y(t} sint = 0.

As in the rst example, we ask ourselves whether this equation has a solution and
whether it is unique. At this point it is clear that there wont be a unique solution
unless we specify some extra information about the function at certain points. The
question is how much extra information is needed in order to obtain an equation
that has a unique solution (assuming that such exists).
Systems of dierential equations In the same way as we go from univariate al-
gebraic equations into multivariate ones, so we can extend the idea of a dierential
equation into a system of dierential equations in which more than one function
is unknown.
Example: Find functions y, z, such that
y
(t} = 6tz
z
(t} = z
3
y
+ 4.
More generally, in a rst-order system of n (explicit) odes the unknown is a vector-

valued function, y R R
n
, that satises an equation of the form
y
(t} = f (t, y(t}},

where f R R
n
R
n
. In components,
y
i
(t} = f
i
(y
1
(t}, . . . , y
n
(t}, t}, i = 1, . . . , n.
6 Chapter 1
High-order equations and rst-order systems Every n-th order (scalar) equa-
tion can be turned into a rst-order system of n odes by a simple procedure. Sup-
pose we have an equation of the form
y
(n)
(t} = f (t, y
n1
(t}, . . . , y
(t}, y(t}}.
We then dene the following vector-valued function,
Y =
y
y
y
(n1)
.
It satises the following system of equations,
Y
1
= Y
2
Y
2
= Y
3
Y
n1
= Y
n
Y
n
= f (t, Y
1
, . . . , Y
n
},
which we can rewrite in vector form,
Y
(t} = F(t, Y(t}}.

What is ordinary in an ode? A dierential equation is called ordinary when the
unknown function depends on a single real-valued variable. This is in contrast to
partial dierential equations ( ), in which the unknown function
depends on more than one variable.
Example: An example of a pde is the heat equation (also known as the diusion
equation): the unknown function depends on two variables (position and time),
y R R R, and satises the equation,
y
t
(x, t} =

2
y
x
2
(x, t}.
The theory of pdes is innitely harder than the theory of odes, and will not be con-
sidered in this course (we will eventually explain why there is such a fundamental
dierence between the two).
Why do we care about odes? Dierential and integral calculus were developed
in the 17th century motivated by the need to describe the laws of nature. Classical
mechanics is one big system of odes; if the instantaneous position of all the
particles in the world at time t can be viewed as a huge vector y(t}, then according
to classical mechanics, the second derivative of this vector (the acceleration) is a
god-given function of the current positions and velocities of all the particles. That
is,
y
(t} = f (t, y
(t}, y(t}}.
As we will see, such a system may have a unique solution if we specify both y
(positions) and y
(velocities) at an initial time (the big bang?). By Newtons laws,

the evolution of the universe is determined forever from its initial conditions.
Example: Consider a single point particle that can only move along a single
axis. The particle is attached to a spring that is connected to the origin. Newtons
law is
my
(t} = k y(t},
where m is the mass of the particle and k is the spring constant (the force law is
that the force on the particle is propositional to its distance from the origin and
inverse in sign). As we will learn, the most general solution to this equation is
y(t} = a cos
_
k
m
t + b sin
_
k
m
t,
where a, b are two integration constants. They can be determined by specifying,
say, y(0} and y
(0}.
odes are important not only in physics. Any (deterministic ) process that evolves
continuously can be described by an ode. odes are used generically to describe
evolving systems in chemistry, economics, ecology, demography, and much more.
The oldest ode The rst documented instance of an ode is an equation studied
by Newton (1671), while developing dierential calculus:
y
(t} = 1 3t + y(t} + t
2
+ ty(t} y(0} = 0.
Newtons approach to solve this equation was based on a method that we call
today an asymptotic expansion. He looked for a solution that can be expressed
as a power series ( ),
y(t} = a
1
t + a
2
t
2
+ a
3
t
3
+ .
8 Chapter 1
Substituting into the equation,
a
1
+ 2a
2
t + 3a
3
t
2
+ 4a
4
t
3
= 1 3t + (a
1
t + a
2
t
2
+ a
3
t
3
} + t
2
+ t(a
1
t + a
2
t
2
+ a
3
t
3
}.
He then proceeded to identify equal powers of t,
a
1
= 1 2a
2
= 3 + a
1
3a
3
= a
2
+ 1 + a
1
4a
4
= a
3
+ a
2
,
etc. This yields, a
1
= 1, a
2
= 1, a
3
= 1]3, a
4
= 1]6, so that
y(t} = t t
2
+
t
3
3

t
4
6
+
For xed t, as more terms as added, the closer we are to the true solution, however,
for every xed number of terms the approximation deteriorates for large t.
A problem studied by Leibniz In 1674 Leibniz (Newtons competitor on the
development of calculus) studied a geometrical problem that had been studied
earlier by Fermat. He tried to solve the following dierential equation,
y
(t} =
y(t}
_
a
2
y
2
(t}
.
Remember that at that time, derivatives were not dened as rigorously as you
learned it (it took 200 more years until calculus took the form you know). For
Leibniz, the derivative was roughly,
y
(t}
y
t
=
y(t}
_
a
2
y
2
(t}
.
If we consider y and t as nite numbers, then we may write
_
a
2
y
2
y
y = t.
Suppose that y(0} = y
0
. Then summing over many small steps t and y,
y
y
0
_
a
2

2
d = t.
This integral can actually be calculated analytically (the primitive function is
known) and so Leibniz could get an implicit representation of the solution.
Initial and boundary value problems Recall that any n-th order equation can
be represented as a rst-order system, thus, until further notice we will assume
that we deal with rst order systems,
y
(t} = f (t, y(t}}.

As we saw, even if such an equation has a solution, it will fail to be unique unless
additional conditions on y are prescribed. In general, a system of n rst-order
equations requires n additional data on y. If all the data are prescribed at a single
point we call the resulting problem an initial value problem. The origin of this
name is realizations where t is time, and the data are prescribed as some initial
time. If the data are distributed between two points of more then we call the re-
sulting problem a boundary value problem. The origin of the name is situations
where the independent variable is space (and then often denoted by x), and y(x} is
some function of space that is prescribed at the boundary of the region of interest.
While the distinction between initial and boundary value problems may seem im-
material it is fundamental. The theories of initial and boundary value problems
are very dierent. The rst part of this course will be devoted to initial value
problems.
Integral formulation Consider a rst order system,
y
(s} = f (s, y(s}} y(t

0
} = y
0
.
Integrating both sides from t
0
to t,
y(t} = y
0
+
t
t
0
f (s, y(s}} ds.
This is an integral equation (y(t} is still unknown), which shares a lot in common
with the dierential equation. We will later see in which sense the dierential and
integral systems are equivalent, and in what sense they dier.
1.2 Eulers approximation scheme
Consider an initial value problem,
y
(t} = f (t, y(t}}, y(t

0
} = y
0
,
10 Chapter 1
where y and f are vector-valued, and suppose that we want to nd the function in
some interval |t
0
, T| (we will often refer to t as time as a suggestive interpreta-
tion).
In 1768 Euler proposed a method to approximate the solution. He considered
partitions of the interval:
t
0
< t
1
< t
2
< < t
n
= T,
and approximated y(t
j
} by y
j
given by
y
j+1
y
j
t
j+1
t
j
= f (t
j
, y
j
},
or equivalently,
y
j+1
= y
j
+ f (t
j
, y
j
}(t
j+1
t
j
}, j = 0, . . . , n 1.
Thus, the rst-order dierential equation is approximated by a one-step dier-
ence equation ( ).
Eulers belief was that if we rene the partition suciently, then we will get ar-
bitrarily close to the true solution (whose existence was not questioned). Eulers
method (usually variants of it) is used until today to numerically approximate the
solution of ordinary dierential systems.
In 1820 Cauchy proved that Eulers approximation does indeed converge, as the
partition is rened, to a solution of the dierential equation, and in fact Cauchys
proof is the rst proof that the dierential system has a solution and that this
solution is unique.
Exercise 1.1 Approximate the solution to the equation
y
(t} = 6y(t} y(0} = 1

at the point t = 1 using Eulers scheme for a uniform partition t
j+1
t
j
= 1]n. (i)
Obtain an explicit expression for the approximate solution. (ii) What is the true
solution? Does the approximate solution converge to the true solution as n ?
TA material 1.1 Learn to solve a large class of equations.
1.3 The Cauchy-Peano existence proof
In order to be able to analyze the dierential system, we must have some a priori
assumptions about the function f . Throughout, we will assume that f is continu-
ous in its two arguments in some open and connected domain D R R
n
.
R
n
can be endowed with a norm. As we know from advanced calculus, all norms
on R
n
are equivalent
1
, and without loss of generality we will use the Euclidean
norm, which we denote by .
(2 hrs, ())
Denition 1.3 Let D R
n
R be a domain in which f is continuous. Let I be an
open connected interval. A function I R
n
is called an -approximation to
the dierential system in D if:
' is continuous.
^ is continuously dierentiable, except perhaps at a nite set of points S ,
where
has one-sided limits.

For every t I, (t, (t}} D.
For every t I S ,
(t} f (t, (t}} .

Let (t
0
, y
0
} D and consider a rectangle of the form
R = |t
0
, t
0
+ a| y y y
0
b} D.
Since f is continuous and R is compact, it follows that f is bounded in R; dene
M = max
(t,y)R
f (t, y}.
Then, we dene
= min(a, b]M},
and replace R by another rectangle,
R
0
= |t
0
, t
0
+ | y y y
0
b} D.
1
Two norms
1
and
2
are said to be equivalent if there exist constants c, C > 0 such that
(x) cx
1
x
2
Cx
1
.
12 Chapter 1
!
"
$
"
$
"
%&

$
"
'&

!
"
%(

!
"
%!

Theorem 1.1 There exists for every > 0 an -approximation in R
0
to the equation
y
(t} = f (t, y(t}} on the interval |t

0
, t
0
+ | satisfying the initial condition y(t
0
} =
y
0
.
Comment: A similar theorem can be formulated for an interval on the left of t
0
.
Mathematically, there is no dierence between past and future.
Proof : Since f is continuous on R
0
, it is uniformly continuous ( ).
Thus, given > 0 there exists a > 0 such that if
|t s| < and y z < then f (t, y} f (s, z} < .
Take the interval |t
0
, t
0
+ | and partition it,
t
0
< t
1
< . . . t
n
= t
0
+ ,
such that
max
j
(t
j+1
t
j
} < min(, ]M}.
We then construct the Euler approximation to this solution, and connect the point
by straight segments, so that Eulers approximation is a polygon, which we denote
by (t}. It is a piecewise continuously dierentiable function.
As long as the polygon is in R
0
its derivative, which is equal to some f (t
j
, y(t
j
}},
has norm less than M. Thus, over an interval of length , (t} can dier from y
0
(in norm) by at most M b. This means that all Eulers polygons remain in R
0
on the entire interval |t
0
, t
0
+|, no mater how the partition of the segment is done.
Let t be a point in the j-th interval. Then,
(t} = f (t
j
, (t
j
}},
and further,
(t} f (t, (t}} = f (t

j
, (t
j
}} f (t, (t}}.
By our choice of the partition,
|t t
j
| < .
Because the derivative of is bounded (in norm) by M,
(t} (t
j
} M|t t
j
| M(]M} = .
By our choice of ,
f (t
j
, (t
j
}} f (t, (t}} < ,
which proves that is an -approximation to our system. B
Denition 1.4 A family F of functions I R
n
is said to be equicontinuous
( ) if there exists for every > 0 a > 0, such that for every
|t
1
t
2
| < and every g F,
g(t
1
} g(t
2
} < .
(Equicontinuity is uniform continuity with the same valid for all g F.)
Theorem 1.2 (Arzela-Ascoli) Let F be an innite family of functions I R
n
that
are uniformly bounded, namely,
sup
f F
max
tI
f (t} < ,
and equicontinuous. Then there exists a sequence f
n
F that converges uniformly
on I.
(3 hrs, ())
14 Chapter 1
Proof : Let (r
k
} be a sequence that contains all the rational numbers in I. Consider
the set
f (r
1
} f F} R
n
.
Since F is uniformly bounded, this set is bounded, and therefore has an accumu-
lation point: there exists a sequence f
(1)
j
F for which f
(1)
j
(r
1
} converges.
Consider then the set
f
(1)
j
(r
2
} j N}.
This set is bounded, therefore there exists a subsequence f
(2)
j
( f
(1)
k
} for which
f
(2)
j
(r
2
} converges (and also f
(2)
j
(r
1
} converges).
We proceed inductively, and for every k derive a sequence f
(k)
j
such that f
(k)
j
(r
i
}
converges for every i j.
Consider now the diagonal sequence F
k
= f
(k)
k
. It converges pointwise on all the
rationals in I. Why? Because for every the diagonal sequence is eventually a
subsequence of f
()
j
. This means that
(r I Q}( > 0}(N N} (m, n > N} (F
n
(r} F
m
(r}} < .
Note that since we dont know what the limit is, we use instead the Cauchy crite-
rion for convergence.
Fix > 0. Since the family of functions F is equicontinuous,
( > 0} (t
1
, t
2
, |t
1
t
2
| < }(f F} ( f (t
1
} f (t
2
} < }.
Partition the interval I into subintervals,
I = I
1
I
2
I
K
,
such that each subinterval if shorter than . In every subinterval I
k
select a rational
number r
k
. Since there is a nite number of such r
k
s,
(N N} (m, n > N}(k = 1, . . . , K} (F
n
(r
k
} F
m
(r
k
}} <
Then for every t I
k
, and m, n > N,
F
n
(t} F
m
(t} F
n
(t} F
n
(r
k
}
-- -
equicontinuity
+F
n
(r
k
} F
m
(r
k
}
- - -
convergence
+F
m
(r
k
} F
m
(t}
---
equicontinuity
3.
Thus, the sequence F
n
satises Cauchys criterion for uniformconvergence, which
concludes the proof. B
With this, we can prove Cauchys theorem:
Theorem 1.3 (Cauchy-Peano) Let f (t, y} and the rectangle R
0
be dened as
above. Then the dierential equation y
(t} = f (t, y(t}} has a solution on the

interval |t
0
, t
0
+ | satisfying the initial condition y(t
0
} = y
0
.
Comments:
' Note that our only conditions on f is that it be continuous.
^ At this stage not a word about uniqueness.
This theorem is known as a local existence theorem. It only guarantees the
existence of a solution on some interval. We have no a priori knowledge on
how long this interval is.
Proof : Let (
k
} be a sequence of positive numbers that tends to zero. We proved
that for every k there exists a function
k
I R
n
that is an
k
-approximation on
I = |t
0
, t
0
+ | to the dierential equation, and satises the initial conditions.
The sequence (
k
} is uniformly bounded on I because
(t I}(k N} (
k
(t} y
0
< b}.
It is also equicontinuous because
(t
1
, t
2
I}(k N} (
k
(t
2
}
k
(t
1
} M|t
2
t
1
|},
and therefore, given > 0 we set = ]M, and
(|t
1
t
2
| < }(k N} (
k
(t
2
}
k
(t
1
} }.
It follows from the Arzela-Ascoli theorem that (
k
} has a subsequence (which we
do not relabel) that converges uniformly on I; denote the limit by y. It remains to
show that y is a solution to the dierential equation.
16 Chapter 1
For each k N it follows from the fundamental theorem of calculus that
2
k
(t} = y
0
+
t
t
0
k
(s} ds
= y
0
+
t
t
0
f (s,
k
(s}} ds +
t
t
0
|
k
(s} f (s,
k
(s}}| ds.
The fact that
k
is not dened at a nite number of points is immaterial.
Using the fact that
k
is an
k
-approximation to the dierential equation,
_
t
t
0
|
k
(s} f (s,
k
(s}}| ds_
k
|t t
0
|.
Letting k and using the fact that the convergence of
k
to y is uniform, we
get that
y(t} = y
0
+
t
t
0
f (s, y(s}} ds.
Since y C
1
(I}, we can dierentiate this equation and obtain y
(t} = f (t, y(t}}. B

Exercise 1.2 Let D R R
n
be an open domain, f C(D}, and (t
0
, y
0
} D.
Show that the equation
y
(t} = f (t, y(t}}, y(t

0
} = y
0
has a solution on some open interval I that contains t
0
.
1.4 Uniqueness
Example: Consider the initial value problem,
y
(t} = y
13
(t}, y(0} = 0.
2
The functions
k
are absolutely continuous ( ), which means that there exists
for every > 0 a > 0 such that for every nite sequence of disjoint intervals (x
k
, y
k
) that satises
k
(y
k
x
k
) < ,
k
f (y
k
) f (x
k
) < .
A function is absolutely continuous if and only if it is almost everywhere dierentiable, and its
derivative satises the fundamental theorem of calculus.
By the Cauchy-Peano theorem there exists an > 0 such that this equation has
a solution in the interval |0, |. The Cauchy-Peano theorem, however, does not
guarantee the uniqueness of the solution. A trivial solution to this equation is
y(t} = 0.
Another solution is
y(t} = _
2t
3
_
32
.
In fact, there are innitely many solutions of the form
y(t} =

0 0 t c
[
2(tc)
3
)
32
t > c,
for every c > 0.
0 0.5 1 1.5 2 2.5 3
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
(What would have happened had we approximated the solution with Eulers method?)
The above example shows that existence of solutions does not necessarily go hand
in hand with uniqueness. It turns out that in order to guarantee uniqueness we need
more stringent requirements on the function f .
Proposition 1.1 Let f R R
n
R
n
by continuous in t and Lipschitz continuous
in y with constant L in the domain D. Let
1
and
2
be
1
- and
2
-approximations
to the equation y
(t} = f (t, y(t}} contained in D on the interval I, and moreover

for a certain t
0
I,
1
(t
0
}
2
(t
0
} < .
18 Chapter 1
Then, for all t I,
1
(t}
2
(t} < e
Ltt
0
+

1
+
2
L
{e
Ltt
0
1) .
Comment: By Lipschitz continuity in y we mean that for every (t, y} and (t, z} in
D,
f (t, y} f (t, z} Ly z.
Comment: This proposition states that the deviation between two approximate
solutions can be split into two: (i) an initial deviation that may grow exponentially
fast, and (ii) a deviation due to the fact that both functions are only approximate
solutions to the same equation.
Proof : Suppose, without loss of generality that t > t
0
. By the denition of -
approximations, for every t
0
s t,
1
(s} f (s,
1
(s}}
1
2
(s} f (s,
2
(s}}
2
.
For i = 1, 2,
i
(t} =
i
(t
0
} +
t
t
0
f (s,
i
(s}} ds +
t
t
0
|
i
(s} f (s,
i
(s}}| ds,
hence
_
i
(t}
i
(t
0
}
t
t
0
f (s,
i
(s}} ds_
i
(t t
0
}.
Using the triangle inequality, a b a + b,
_|
1
(t}
2
(t}| |
1
(t
0
}
2
(t
0
}|
t
t
0
| f (s,
1
(s}} f (s,
2
(s}}| ds_ (
1
+
2
}(tt
0
}.
It further follows that
1
(t}
2
(t}
1
(t
0
}
2
(t
0
}
+
t
t
0
f (s,
1
(s}} f (s,
2
(s}} ds + (
1
+
2
}(t t
0
}.
Dene now r(t} =
1
(t}
2
(t}. Then, using the Lipschitz continuity of f and
the bound on the initial deviation,
r(t} + L
t
t
0
r(s} ds + (
1
+
2
}(t t
0
}. (1.1)
This is an integral inequality. Our goal is to get an explicit bound for r(t}. We
proceed as follows: dene
R(t} =
t
t
0
r(s} ds,
so that
R
(t} + LR(t} + (
1
+
2
}(t t
0
}, R(t
0
} = 0,
which is a dierential inequality.
Next, for every t
0
s t:
{e
L(st
0
)
R(s})
= e
L(st
0
)
|R
(s} LR(s}| e
L(st
0
)
( + (
1
+
2
}(s t
0
}} ,
Since integration preserves order, we may integrate this inequality from t
0
to t and
get
e
L(tt
0
)
R(t} R(0}
-
0
t
t
0
e
L(st
0
)
( + (
1
+
2
}(s t
0
}} ds,
and it only remains to integrate the right hand side explicitly
3
and substitute back
into (1.1). B
TA material 1.2 Various Gronwall inequalities.
Theorem 1.4 (Uniqueness) Let
1
and
2
be two solutions of the initial value
problem
y
(t} = f (t, y(t}}, y(t

0
} = y
0
,
where f is continuous in t and Lipschitz continuous in y. Then
1
=
2
.
Proof : This is an immediate corollary of the previous proposition as =
1
=
2
=
0. B
3
t
0
se
Ls
ds =
t
L
e
Lt
1
L
2
e
Lt
1 .
20 Chapter 1
Comment: We have obtained also an estimate for the error of Cauchys approx-
imation. It y is the solution and is an -approximation with exact initial data,
then
(t} y(t} <

L
{e
Ltt
0
1) .
(5 hrs, ())
1.5 The Picard-Lindl of existence proof
There is another standard way to prove the existence of solutions, and it is of suf-
cient interest to justify a second proof. Recall that a solution to the initial value
problem y
(t} = f (t, y(t}}, y(t

0
} = y
0
is also a solution to the integral equation,
y(t} = y
0
+
t
t
0
f (s, y(s}} ds.
Dene the function y
(0)
(t} y
0
, and further dene
y
(1)
(t} = y
0
+
t
t
0
f (s, y
(0)
(s}} ds.
For what values of t is it dened? For the same |t
0
, t
0
+ | as in the previous
sections. We can show that (t, y
(1)
(t}} remains in R
0
as follows: let
= supt I y
(1)
(t} y
0
b}.
If < t
0
+ then
y
(1)
(} y
0
M(t t
0
} < b,
contradicting the fact that is the supremum.
We then dene inductively for every n N:
y
(n+1)
(t} = y
0
+
t
t
0
f (s, y
(n)
(s}} ds,
and show as above that (t, y
(n+1)
(t}} remains in R
0
for all t I.
We are now going to show that the sequence y
(n)
converges uniformly to a solution
to the initial value problem. The dierence with the proof based on Eulers method
is that we need to assume the Lipschitz continuity of f in R
0
. Dene,
(n)
= y
(n+1)
y
(n)
,
then,
(n+1)
(t} =
t
t
0
f (s, y
(n)
(s}} f (s, y
(n1)
(s}} ds,
and
(n+1)
(t} L
t
t
0
(n)
(s} ds,
with
(1)
(t}
t
t
0
f (s, y
0
} ds M|t t
0
|.
Once again, we have an integral inequality, but we may proceed dierently:
(2)
(t} L
t
t
0
(1)
(s}| ds LM
t
t
0
(s t
0
} ds
LM
2
|t t
0
|
2
,
(3)
(t} L
t
t
0
(2)
(s}| ds
L
2
M
2

t
t
0
(s t
0
}
2
ds
L
2
M
3!
|t t
0
|
3
,
and generally,
(n)
(t}
M
L
(L|t t
0
|}
n
n!
.
Now we write the n-th function as follows:
y
(n)
= {y
(n)
y
(n1)
) + {y
(n1)
y
(n2)
) + + y
0
= y
0
+
n1
k=0
(k)
.
By the Weiersta M-test this series converges uniformly to a limit y. All that
remains is to let n in the equation
y
(n+1)
(t} = y
0
+
t
t
0
f (s, y
(n)
(s}} ds,
to obtain
y(t} = y
0
+
t
t
0
f (s, y(s}} ds.
Comment: Usually one puts this proof in the context of the Picards contractive
mapping theorem. I purposely did it here by hand to show how straightforward
the proof is. This method is known as the method of successive approximations.
Note also that we can estimate the error of the successive approximations. We
have
y
(n)
y
(m)
=
n1
k=m
(k)
,
22 Chapter 1
hence
y
(n)
(t} y
(m)
(t}
n1
k=m
(k)
(t}
n1
k=m
L
k1
M
k!
(t t
0
}
k
k=m
L
k1
M
k!
(t t
0
}
k
=

k=0
L
m+k1
M
(m+ k}!
(t t
0
}
m+k
k=0
L
m+k1
M
m!k!
(t t
0
}
m+k
=
L
m1
M(t t
0
}
m
m!
k=0
L
k
k!
(t t
0
}
k
=
M|L(t t
0
}|
m
Lm!
e
L(tt
0
)
.
The right hand side does not depend on n, so we can let n to get an estimate
for y y
(m)
.
1.6 Continuation of solutions
The existence and uniqueness proof shows that there exists an such that a unique
solution exists in the interval |t
0
, t
0
+ |. Does it really mean that the domain of
existence is limited? What prevents the existence of solution for all t R?
Example: Consider the initial value problem:
y
(t} = y
2
(t}, y(0} = 1.
It is easy to check that
y(t} =
1
1 t
,
is a solution on the interval |0, 1}, and we know therefore that it is unique. This
solution however diverges as t 1, indicating that it cant be further continued.
This situation is typical to equations that do not have global solutions. Suppose
that f (t, y} is continuous in t and Lipschitz continuous in y for all (t, y} R R
n
(as is f (t, y} = y
2
), and dene
a = sup > t
0
there exists a solution on |t
0
, |}.
If a is nite and
lim
ta
y(t} = b < ,
then consider the initial value problem
y
(t} = f (t, y(t}}, y(a} = b.

By the existence and uniqueness theorems there exists a such that a unique
solution exists on |a, a + |, i.e., a unique solution exists on |t
0
, a + |, violating
the denition of a as a supremum. Thus, if a is nite then
lim
ta
y(t} = ,
namely, the only way a solution can fail to exist indenitely is if it diverges.
This observation has an important implication: whenever you want to show that a
certain equation has global solutions, you can try to show that as long as a solution
exists it is bounded uniformly.
(6 hrs, ())
1.7 Generalized solutions
1.8 Continuity of solutions with respect to initial con-
ditions
The solution of an ivp is of course a function of time t, but it can also be viewed
as a function of the initial data, namely, we can viewed the solution of
y
(t} = f (t, y(t}}, y(} =

as a function of t,, and ; for now we denote this solution as (t, , }. This is
a dierent point of view in which we have a function that represents the solution
not to a single ivp, but to all ivps together.
We know that is dierentiable in t, as it satises the equations,
t
(t, , } = f (t, (t, , }}, (, , } = . (1.2)
The question is how regular it is with respect to the initial data and . This is an
important question, as if it were, for example, discontinuous with respect to the
initial data, it would not be an appropriate model for any real life problem. In this
24 Chapter 1
context, one speaks of the well-posedness of an equation, which includes exis-
tence, uniqueness, and continuity on initial data (and possibly other parameters in
the equation).
Let D R R
n
be a domain in which f is continuous in t and Lipschitz in y. Let
I = |a, b| and let I R
n
be a solution to the dierential equation, such that
(t, (t}} t I} D.
Theorem 1.5 There exists a > 0 such that for every
(, } U
= (, } I, (} < }
There exists a solution to (1.2) on I with (, , } = . Moreover, is continuous
on I U
.
Chapter 2
Linear equations
2.1 Preliminaries
Recall that for every normed space (X, } we can endow the vector space of
linear operators L(X, X} with an operator norm,
A = sup
x0
Ax
x
.
It can be shown that this is indeed a norm. This norm has the following (addi-
tional) properties:
1. For every x X and A L(X, X},
Ax Ax.
2. For every A, B L(X, X},
AB AB.
3. It is always the case that id = 1.
We will deal with the vector space R
n
endowed with the Euclidean norm. One can
ask then what is the explicit form of the corresponding operator norm for matrices
A R
nn
.
Exercise 2.1 Find the explicit form of the operator norm corresponding to the
Euclidean norm on R
n
.
26 Chapter 2
2.2 Linear homogeneous systems
2.2.1 The space of solutions
Let I R be a closed interval and let A I R
nn
be a continuous n-by-n matrix-
valued function of I. We denote its entries, which are real-valued functions, by
a
i j
. The rst-order system of dierential equations,
y
(t} = A(t}y(t} (2.1)

is called a linear homogeneous system. In component notation,
y
j
(t} =
n
k=1
a
jk
(t}y
k
(t}.
Example: A linear homogeneous system for n = 2:
_
y
1
y
2
_
(t} = _
0 1
1 +
1
2
sint 0
__
y
1
y
2
_(t}.
Proposition 2.1 For every I and R

n
, there exists a unique function y I
R
n
such that
y
(t} = A(t}y(t}, y(} = .

Comment: The main claim is that a (unique) solution exists on the whole of I; in
fact this remains true also if I = R.
Proof : The function f (t, y} = A(t}y is continuous in t and Lipschitz in y, hence
a local solution exists for some interval containing . We need to show that the
solution can be continued on the entire of I, i.e., that it cannot diverge. This is
done as follows. We rst write the corresponding integral equation,
y(t} = +
A(s}y(s} ds,
which holds as long as a solution exists.
Linear equations 27
Taking norms,
y(t} +
A(s}y(s} ds +
A(s}y(s} ds.
Because A is continuous on I, so is A, which means that there exists a uniform
bound
A(s} M, t I.
This implies that
y(t} + M
y(s} ds.
This is an integral inequality that we can solve. Dening Y(t} =
y(s} ds we
get
Y
(t} + MY(t}, Y(} = 0.

Then,
e
M(t)
|Y
(t} MY(t}| e
M(t)
.
Integrating from to t,
e
M(t)
Y(t}

M
{e
M(t)
1) ,
and further
Y(t}

M
{e
M(t)
1) .
Substituting back into the inequality for y(t},
y(t} + M
M
{e
M(t)
1) = e
M(t)
.
This can never diverge in I, which proves that a solution to the ivp exists on I. B
The trivial solution One characteristic of linear homogeneous systems is that
y 0 is always a solution. This solution is called the trivial solution.
Proposition 2.2 Let X be the space of functions y I R
n
satisfying the linear
homogeneous system (2.1). Then X is a vector space over the complex eld C
with respect to pointwise addition and scalar multiplication.
28 Chapter 2
Proof : Let y, z X, i.e.,
y
(t} = A(t}y(t} and z
(t} = A(t}z(t}.
For every , C,
(y + z}
(t} = A(t}y(t} + A(t}z(t} = A(t}(y + z}(t},

which proves that y + z X. B
Comment: The elements of X are functions I R
n
.
Comment: This is what physicists call the principle of superposition. Any linear
combination of solutions is also a solution.
Recall that the dimension of a vector space is dened as the maximal number n
for which there exists a set of n linearly independent vectors, but every (n + 1}
vectors are linearly dependent.
Proposition 2.3 Let X be dened as above, then
dimX = n,
that is there exists a set of functions
1
, . . . ,
n
X, such that
n
k=1
c
k
k
= 0
implies that c
k
= 0 for all k. Moreover, for every set of n+1 function
k
X there
exist scalars c
k
, not all of them zero, such that
n+1
k=1
c
k
k
= 0.
Proof : Take I, and select n independent vectors
1
, . . . ,
n
R
n
; we denote the
k-th component of
j
by
k j
. For each
j
, the ivp
j
(t} = A(t}
j
(t},
j
(} =
j
Linear equations 29
has a unique solution. We claim that these solution are independent. Indeed,
suppose that
n
k=1
c
k
k
0,
then in particular for t = ,
n
k=1
c
k
k
= 0,
but because the
k
are independent, the c
k
are all zero.
Let
1
, . . . ,
n
be linear independent solutions. Take any I and let
j
= (}.
The
n
are linearly independent, for if they were linearly dependent, we would
have
n
=
n1
k=1
c
k
k
,
but then
=
n1
k=1
c
k
k
is a solution to the dierential system satisfying (} =
n
, and by uniqueness
=
n
, which contradicts the linear independence of the
k
.
Suppose now that is another solution of the dierential system. Set
= (}.
Because the
j
span R
n
, there exist coecients c
k
such that
=
n
k=1
c
k
k
.
By the same argument as above,
=
n
k=1
c
k
k
,
which proves that the
k
form a basis. B
30 Chapter 2
2.2.2 Fundamental matrices
Let now
1
, . . . ,
n
be a basis. Let I R
nn
be the matrix-valued function
whose columns are the vectors
j
,
(t} =

11

1n

n1

nn
(t},
that is,
i j
is the i-th component of the j-th basis function. This matrix satises
the matrix-valued dierential equation,
(t} = A(t}(t},
which in components reads
i j
(t} = a
ik
(t}
k j
(t}.
(t} is called a fundamental matrix ( ) for the linear homogeneous
system (2.1); we call a fundamental matrix for (2.1) any solution X I R
nn
of
the matrix equation
X
(t} = A(t}X(t}, (2.2)

whose columns are linearly independent.
Because the columns of a fundamental matrix span the space of solutions to (2.1),
any solution y I R
n
can be represented as
y
i
=
n
j=1
c
j
i j
,
where the c
j
s are constant, and in vector form,
y = c.
Proposition 2.4 For a matrix-valued function to be a fundamental matrix of
(2.1), it has to satisfy the matrix equation (2.2) and its columns have to be inde-
pendent for some I. In particular, if this holds then its columns are independent
for all t I.
Linear equations 31
Proof : Let I R
nn
be a fundamental matrix; we know that it is a solution of
(2.2) and that any solution y I R
n
of (2.1) can be represented as
y = c,
Fix a I. The vector c is determined by the equation
y(} = (}c.
This is a linear equation for c; we know that it has a unique solution (because
the columns of span the space of solutions), hence from basic linear algebra,
det (} 0. Since the same considerations hold for every t I, det (t} 0 for
all t I. B
(8 hrs, ())
Here is another proof that the columns of the matrix remain independent, but it
comes with a quantitative estimate:
Proposition 2.5 Let I R
nn
be a fundamental matrix. Then
det (t} = det (t
0
} exp_
t
t
0
Tr A(s} ds_.
In particular, if det (t
0
} 0 then det (t} 0 for all t I.
Proof : Since
=
11

1n

n1

nn
,
we dierentiate to obtain
11

1n

n1

nn
+ +
11

1n

n1

nn
j
a
1 j
j1

j
a
1j
jn

n1

nn
+ +
11

1n

j
a
n j
j1

j
a
n j
jn
.
32 Chapter 2
Consider the rst determinant. We subtract from the rst line a
12
times the second
line, a
13
times the third line, etc. Then we remain with a
11
det . We do the same
with the second determinant and remains with a
22
det , and so on. Thus,
= (Tr A} .
This is a scalar linear ode, and it remains to integrate it.
B
Example: Consider the matrix
(t} = _
t t
2
0 0
_.
This matrix is singular, but its columns are linearly independent. What does it
imply? That this matrix-valued function cannot be a fundamental matrix of a
linear homogeneous system; there exists no matrix A(t} such that
_
t
0
_
= A(t} _
t
0
_ and _
t
2
0
_
= A(t} _
t
2
0
_.
Theorem 2.1 Let be a fundamental matrix for (2.1) and let C C

nn
be a non-
singular complex (constant) matrix. Then C is also a fundamental matrix. More-
over, if is a fundamental matrix then = C for some nonsingular complex
matrix C.
Proof : By denition
= A and we know that det (t} 0 for all t. Then for

every nonsingular complex matrix C,
(C}
= AC
and det (t}C = det (t} det C 0 for all t I.
Let be a fundamental matrix. Because it is nonsingular for all t,
1
exists, and
0 = (
1
}
= (
1
}
1
,
Linear equations 33
from which we get that
(
1
}
=
1
1
=
1
A
1
=
1
A.
Now,
(
1
}
= (
1
}
+
1
=
1
A+
1
A = 0.
It follows that
1
= C, or = C. B
Comment: The matrix (
}
1
satises the equation,
|(
}
1
|
= A
}
1
.
This equation is known as the equation adjoint ( ) to (2.1).
Proposition 2.6 If is a fundamental matrix for (2.1) then is a fundamental
matrix for the adjoint system if and only if
= C,
where C is a nonsingular constant matrix.
(9 hrs, ())
Proof : Let satisfy the above equation then
(
= 0,
i.e.,
(
A,
and since is nonsingular
(
A,
and it remains to take the Hermitian adjoint of both sides.
Conversely, if is a fundamental solution of the adjoint equation then
(
= 0,
which concludes the proof. B
34 Chapter 2
Corollary 2.1 Let A be anti-hermitian, i.e.,
A(t} = A
(t},
then the linear homogeneous system and its adjoint coincide. In particular, if is
a fundamental matrix for (2.1) then it is also a fundamental matrix for its adjoint,
and
(t}(t} = C.
The matrix
(t}(t} determines the magnitudes of the columns of (t} and the

angles between pairs of vectors. In particular, since any solution is of the form
y(t} =
n
j=1
j
(t},
then
y(t}
2
=
n
i, j=1
j
(
i
,
j
} =
n
i, j=1
j
C
i j
,
i.e., the norm of any solution is independent of t.
Example: Consider the linear system
y
(t} = _
0 1
1 0
_y(t}.
Two independent solutions are
1
(t} = _
sint
cos t
_ and
2
(t} = _
cos t
sint
_.
A fundamental matrix is therefore
(t} = _
sint cos t
cos t sint
_.
It follows that,
(t}(t} = _
sint cos t
cos t sint
__
sint cos t
cos t sint
_ = _
1 0
0 1
_.
Linear equations 35
Any solution of this linear system is of the form
y(t} = _
sint
cos t
_ + _
cos t
sint
_.
Now,
y(t}
2
=
2
{sint cos t) _
sint
cos t
_ + 2{sint cos t) _
cos t
sint
_
+
2
{cos t sint) _
cos t
sint
_ =
2
+
2
.
2.2.3 The solution operator

Consider the initial value problem,
y
(t} = A(t}y(t}, y(t

0
} = given.
If (t} is a fundamental matrix, then the solution is of the form
y(t} = (t}c.
The vector c is found by substituting the initial data, y(t
0
} = (t
0
}c, i.e., c =
1
(t
0
}y(t
0
}, and the solution is therefore,
y(t} = (t}
1
(t
0
}
-- -
R(t,t
0
)
y(t
0
}.
Comments:
' It follows that the solution at time t depends linearly on the solution at time
t
0
, and the linear transformation between the solution at two dierent times
is the matrix R(t, t
0
}. We call the matrix R(t, t
0
} the solution operator
( ) between times t
0
and t.
^ The solution operator seems to depend on the choice of the fundamental
matrix, but it doesnt. It doesnt because the solution is unique. We can also
see it as any other fundamental matrix (t} can be represented as (t} =
(t}C, and then
(t}
1
(t
0
} = (t}CC
1
1
(t
0
} = (t}
1
(t
0
}.
36 Chapter 2
For every t
0
, t
1
, t
2
,
R(t
2
, t
1
}R(t
1
, t
0
} = R(t
2
, t
0
}.
Why? Because for every y(t
0
},
y(t
2
} = R(t
2
, t
0
}y(t
0
},
but also,
y(t
2
} = R(t
2
, t
1
}y(t
1
} = R(t
2
, t
1
}R(t
1
, t
0
}y(t
0
}.
R(t, t} = I for all t.
I R(t, t
0
} is non-singular and
|R(t, t
0
}|
1
= R(t
0
, t}.
Let us recapitulate: given a homogeneous linear system, we need to nd n inde-
pendent solutions. Then we can construct a fundamental matrix and the associated
solution operator. This gives an explicit solution for any initial value problem. The
catch is that in most cases we will not be able to nd n independent solutions.
2.2.4 Order reduction
Consider an n-dimensional linear homogeneous system, and suppose that we have
m linearly independent solutions
1
, . . . ,
m
,
with m < n.
Example: Consider the dierential system in some interval that does not include
the origin:
y
(t} = _
0 1
3]2t
2
1]2t
_y(t}.
A solution to this system is
1
(t} = _
1]t
1]t
2
_,
Linear equations 37
as
1
(t} = _
1]t
2
2]t
3
_ = _
0 1
3]2t
2
1]2t
__
1]t
1]t
2
_.
All we need in order to be able to solve an initial value problem is to nd another
independent solution. The knowledge of one solution enables us to get for the
other solution an equation of lower order (in this case, a scalar equation).
2.3 Non-homogeneous linear systems
Consider now a system of the form
y
(t} = A(t}y(t} + b(t},

where A I R
nn
and b I R
n
are continuous. Such a system is called a linear
non-homogeneous system.
It turns out that if a fundamental matrix for the corresponding homogeneous sys-
tem is known, then we can solve the non-homogeneous system in explicitly form.
Theorem 2.2 Let (t} be a fundamental matrix for the homogenous system, then
the solution for the non-homogeneous system is
y(t} = R(t, t
0
}y(t
0
} +
t
t
0
R(t, s}b(s} ds,
where as above R(t, t
0
} = (t}
1
(t
0
}.
Proof : Note that the solution operator also satises the homogeneous system,
R
(t, t
0
} = A(t}R(t, t
0
}
Dierentiating the above solution,
y
(t} = R
(t, t
0
}y(t
0
} +
t
t
0
R
(t, s}b(s} ds + R(t, t}b(t}

= A(t}R(t, t
0
}y(t
0
} + A(t}
t
t
0
R(t, s}b(s} ds + b(t}
= A(t}y(t} + b(t}.
38 Chapter 2
By uniqueness, this is the solution. B
(10 hrs, ( ))
Look at the structure of the solution. It is a sum of the solution of the solution to
the homogeneous problem and a term that accumulates the non-homogeneous
term.
The space of solutions is an ane space Another point of practical interest is
the following. Suppose that y
in
(t} is a solution to the non-homogeneous equation
and let y(t} by any other solution. Then,
(y y
in
}
= A(t}(y y
in
},
which implies that any solution to the non-homogeneous system can be expressed
as y
in
(t} plus a solution to the homogeneous system.
The method of variation of constants Another way to derive Theorem 2.2 is
the method of variation of constants. The solution to the homogeneous system
is of the form
y(t} = (t}C.
We then look for a solution where C is a function of time,
y(t} = (t}C(t}.
Dierentiating we get
y
(t} =
(t}
-
A(t)(t)
C(t} + (t}C
(t} = A(t}y(t} + b(t} = A(t}(t}C(t} + b(t},

hence
(t}C
(t} = b(t} or C
(t} =
1
(t}b(t}.
Integrating we nd a solution to the non-homogeneous equation,
C(t} =
t
t
0
1
(s}b(s} ds i.e. y(t} =
t
t
0
(t}
1
(s}b(s} ds.
This is not the most general equation: we can always add a solution to the homo-
geneous system:
y(t} = (t}C
1
+
t
t
0
(t}
1
(s}b(s} ds,
and substituting the initial data we get that C
1
=
1
(t
0
}y(t
0
}.
Linear equations 39
Comment: The method shown here is a special case of something known as
Duhammels principle. It is a general approach for expressing the solution of
a linear non-homogeneous system in terms of the solution operator of the homo-
geneous system.
Example: Consider a forced linear oscillator,
y
(t} = _
0 1
1 0
_y(t} + _
0
sint
_,
where is a frequency. This system describes a unit mass attached to a spring
with unit force constant and forced externally with frequency . The entries of
y(t} are the displacement and the momentum.
A fundamental matrix for the homogeneous system is
(t} = _
sint cos t
cos t sint
_.
Since
1
(t} = _
sint cos t
cos t sint
_,
then the solution operator is
R(t, t
0
} = _
sint cos t
cos t sint
__
sint
0
cos t
0
cos t
0
sint
0
_ = _
cos(t t
0
} sin(t t
0
}
sin(t t
0
} cos(t t
0
}
_.
Substituting into Theorem 2.2 we get
y(t} = _
cos(t t
0
} sin(t t
0
}
sin(t t
0
} cos(t t
0
}
_y(t
0
} +
t
t
0
_
cos(t s} sin(t s}
sin(t s} cos(t s}
__
0
sins
_ ds
= _
cos(t t
0
} sin(t t
0
}
sin(t t
0
} cos(t t
0
}
_y(t
0
} +
t
t
0
_
sins sin(t s}
sins cos(t s}
_ ds.
Assume that t
0
= 0. Then
y(t} = _
cos t sint
sint cos t
_y(0} +
1
2
1
_
sint sint
(cos t cos t}
_.
The solution splits into a part that depends on the initial condition plus a part
that depends on the forcing. The expression for the non-homogeneous term holds
40 Chapter 2
only of 1. If = 1, then the natural frequency coincides with the forcing
frequency, a phenomenon known as resonance. If = 1 then we get
y(t} = _
cos t sint
sint cos t
_y(0} +
1
2
_
sint t cos t
t sint
_,
i.e., the solution diverges linearly in time.
2.4 Systems with constant coecients
2.4.1 The Jordan canonical form
Let A be a (constant) n-by-n matrix. Recall that any matrix is similar to a matrix
in Jordan form: that is, there exists a non-singular complex matrix P such that
A = PJP
1
.
and J is of the form
J =
J
0
J
1
J
s
,
where
J
0
=
and J
i
=
q+i
1 0 0 0
0
q+i
1 0 0

0 0 0 0 1
0 0 0 0
q+i
,
for i = 1, . . . , s. If A is diagonalizable then it only has the diagonal part J
0
. (Recall
that a matrix is diagonalizable if it is normal, i.e., it commutes with its adjoint.)
2.4.2 The exponential of a matrix
Let A be a (constant) n-by-n matrix. We dene
e
A
=

n=0
A
n
n!
.
Linear equations 41
Is it well-dened? Yes, because,
k>n
A
k
k!

kn
A
k
k!
=

k=0
A
k+n
(k + n}!

k=0
A
k+n
n! k!

A
n
n!
e
A
n
0,
and by Cauchys criterion the series converges absolutely.
What can we say about the exponential of a matrix? In general,
e
A+B
e
A
e
B
,
unless A and B commute.
If A = PJP
1
, where J is the Jordan canonical form, then
e
A
=

n=0
(PJP
1
}
n
n!
=

n=0
PJ
n
P
1
n!
= Pe
J
P
1
,
which means that it is sucient to learn to exponentiate matrices in Jordan canon-
ical form.
Note that for a block-diagonal matrix J,
e
J
=
e
J
0
e
J
1
e
Js
,
so that it is sucient to learn to exponentiate diagonal matrices and matrices of
the form
J =
1 0 0 0
0 1 0 0

0 0 0 0 1
0 0 0 0
.
The exponential of a diagonal matrix
J
0
=
is e
J
0
=
1
e
e
q
.
42 Chapter 2
The non-diagonal blocks have the following structure:
J
i
=
q+i
I
r
i
+ E
r
i
,
where r
i
is the size of the i-th block and
E
i
=
0 1 0 0 0
0 0 1 0 0

0 0 0 0 1
0 0 0 0 0
.
Because I
r
i
and E
i
commute then
e
J
i
= e
q+i
I
e
E
i
= e
q+i
e
E
i
.
The matrix E
i
is nilpotent. Suppose for example that r
i
= 4, then
E
i
=
0 1 0 0
0 0 1 0
0 0 0 1
0 0 0 0
E
2
i
=
0 0 1 0
0 0 0 1
0 0 0 0
0 0 0 0
E
3
i
=
0 0 0 1
0 0 0 0
0 0 0 0
0 0 0 0
E
4
i
= 0.
Then,
e
E
i
=
1 1
1
2!
1
3!
0 1 1
1
2!
0 0 1 1
0 0 0 1
For later use, we note that if we multiply J

j
by a scalar t, then
e
tJ
i
= e
t
q+i
1 t
t
2
2!
t
3
3!
0 1 t
t
2
2!
0 0 1 t
0 0 0 1
Example: Let
A = _
0 t
t 0
_.
Linear equations 43
One way to calculate e
tA
is to calculate powers of A:
A
2
= t
2
I A
3
= t
3
A A
4
= t
4
I,
hence
e
tA
= I + tA
t
2
2!
I
t
3
3!
A +
t
4
4!
I +
t
5
5!
A
t
6
6!
I = cos t I + sint A,
i.e.,
e
tA
= _
cos t sint
sint cos t
_.
Alternatively, we diagonalize the matrix,
A = _
1
2
_
1 1
i i
__ _
it 0
0 it
__
1
2
_
1 i
1 i
__ ,
and then
e
tA
= _
1
2
_
1 1
i i
__ _
e
it
0
0 e
it
__
1
2
_
1 i
1 i
__
=
1
2
_
1 1
i i
__
e
it
ie
it
e
it
ie
it
_
=
1
2
_
e
it
+ e
it
ie
it
+ ie
it
ie
it
ie
it
e
it
+ e
it
_
= _
cos t sint
sint cos t
_.
2.4.3 Linear system with constant coecients

We now consider homogeneous linear systems in which the coecients are con-
stant,
y
(t} = Ay(t},
where A is a (constant) n-by-n matrix. As we know, solving this system amounts
to nding a fundamental matrix.
44 Chapter 2
Proposition 2.7 The matrix e
tA
is a fundamental matrix for the system with con-
stant coecients A.
Proof : From the denition of the derivative,
lim
h0
e
(t+h)A
e
tA
h
= _lim
h0
e
hA
I
h
_e
tA
= Ae
tA
,
hence y(t} = e
tA
is a fundamental solution. B
Comments:
' Since A and e
tA
commute it is also the case that
d
dt
e
tA
= e
tA
A.
^ It follows that the solution operator is
R(t, t
0
} = e
tA
e
t
0
A
= e
(tt
0
)A
,
where we used the fact that e
tA
and e
sA
commute for every t, s.
It follows that the solution of the inhomogenous equation y
(t} = Ay(t}+b(t}
is
y(t} = e
(tt
0
)A
y(t
0
} +
t
t
0
e
(ts)A
b(s} ds.
Inspired by the case of n = 1, one may wonder whether the exponential is
only a fundamental solution for constant coecients. Isnt the solution to
y
(t} = A(t}y(t},
(t} = e
t
0
A(s) ds
?
Lets try to dierentiate:
(t + h} (t}
h
=
e
t+h
0
A(s) ds
e
t
0
A(s) ds
h
.
However, in general
e
t+h
0
A(s) ds
e
t
0
A(s) ds
e
t+h
t
A(s) ds
,
which is what prevents us from proceeding. These matrices commute either
if A is constant, or if it is diagonal (in which case we have n independent
equations).
Linear equations 45
I If A = PJP
1
then
e
tA
= Pe
tJ
P
1
,
and by the properties of fundamental matrices,
e
tA
P = Pe
tJ
is also a fundamental solution. This means that the t dependence of all the
components of the fundamental matrix is a linear combinations of terms of
the form t
m
e
t
, where is an eigenvalue of A and m is between zero is one
minus the multiplicity of .
2.5 Linear dierential equations of order n
2.5.1 The Wronskian
Let a
0
, a
1
, . . . , a
n
I Cbe continuous complex-valued functions dened on some
interval. We dene the linear dierential operator, L
n
C
1
(I} C
0
(I}.
L
n
( f } = a
0
f
(n)
+ a
1
f
(n1)
+ + a
n1
f
+ a
n
f ,
or in another notation,
L
n
= a
0
d
n
dt
n
+ a
1
d
n1
dt
n1
+ + a
n1
d
dt
+ a
n
.
(It will be assume that a
0
(t} 0 for all t I.)
A (scalar) dierential equation of the form
L
n
y = 0
is called a homogeneous linear equation of order n. It can also be written in the
form
y
(n)
(t} =
a
1
(t}
a
0
(t}
y
(n1)
(t}
a
2
(t}
a
0
(t}
y
(n2)
(t}
a
n1
(t}
a
0
(t}
y
(t}
a
n
(t}
a
0
(t}
y(t}.
As usual, we can convert an n-th order equation into a rst-order system of n
equations,
Y
(t} = A(t}Y(t},
46 Chapter 2
where
A =
0 1 0 0 0
0 0 1 0 0

0 0 0 0 1
a
n
]a
0
a
n1
]a
0
a
n2
]a
0
a
n3
]a
0
a
1
]a
0
.
This is a homogeneous linear system, hence we know a lot about the structure
of its solutions. It is however a very special system, as the matrix A has most of
its elements zero. In fact, every solution of the linear system has the following
structure:
(n1)
.
In particular, a fundamental matrix is determined by n independent scalar func-
tions
j
:
=
1

2

n
(n1)
1

(n1)
2

(n1)
n
.
The determinant of this matrix is called the Wronskian of the system L
n
y = 0
with respect to the solutions
1
, . . . ,
n
. We denote it as follows:
W(
1
, . . . ,
n
}(t} = det (t}.
By Proposition 2.5
W(
1
, . . . ,
n
}(t} = W(
1
, . . . ,
n
}(t
0
} exp_
t
t
0
Tr A(s} ds_
= W(
1
, . . . ,
n
}(t
0
} exp_
t
t
0
a
1
(s}
a
0
(s}
ds_.
As we know, n solutions
j
(t} are linearly independent if the Wronskian (at any
point t) is not zero.
Linear equations 47
The fact that for any n linearly independent solutions
j
,
W(
1
, . . . ,
n
}(t} = c exp_
t
a
1
(s}
a
0
(s}
ds_
is known as Abels theorem.
TA material 2.1 Use Abels theorem to show how given one solution one can
reduce the order of the equation for the remaining solutions. Show it explicitly for
n = 2 and solve an example. Look for example at
http://www.ux1.eiu.edu/wrgreen2/research/Abel.pdf
Thus, given an n-th order linear operator L
n
on I and n solutions
j
I R
satisfying L
n
j
= 0 for which W(
1
, . . . ,
n
}(t} 0, then every function y I R
satisfying L
n
y = 0 is a linear combination of the
j
s.
It turns out that the opposite is also true. To every n linearly independent functions
corresponds a linear dierential operator of order n for which they form a basis:
Theorem 2.3 Let
1
, . . . ,
n
be n-times continuously dierentiable functions on an
interval I such that
t I W(
1
, . . . ,
n
}(t} 0.
Then there exists a unique linear dierential operator of order n (assuming that
the coecient of the n-th derivative is one) for which these functions form a basis.
Proof : We start by showing the existence of such an equation. Consider the equa-
tion
(1}
n
W(y,
1
, . . . ,
n
}(t}
W(
1
, . . . ,
n
}(t}
= 0.
This is an n-th order linear dierential equation for y. Also the coecient of y
(n)
is 1.
Clearly,
W(
j
,
1
, . . . ,
n
}(t} = 0
for all js, which means that the
j
s form a basis for this equation.
48 Chapter 2
It remains to prove the uniqueness of this equation. The solutions
j
provide us
with a fundamental matrix for the associated rst-order system:
Y
(t} = A(t}Y(t}.
I.e., the matrix A(t} is uniquely determined, and this in turn uniquely determines
the ratios a
j
(t}]a
0
(t}.
B
2.5.2 The adjoint equation
In the context of rst-order linear systems we encountered the equation adjoint to
a given equation:
Y
(t} = A
(t}Y(t}.
The adjoint equation will become important later when we deal with boundary
value problems.
Consider now the n-th order operator L
n
, and assume that a
0
(t} = 1. The adjoint
system is
A
0 0 0 0 0 a
n
1 0 0 0 0 a
n1
0 1 0 0 0 a
n2

0 0 0 0 0 a
2
0 0 0 0 1 a
1
Let us write down this equation in components:

y
1
= a
n
y
n
y
2
= y
1
+ a
n1
y
n
y
3
= y
2
+ a
n2
y
n
n1
= y
n2
+ a
2
y
n
y
n
= y
n1
+ a
1
y
n
.
Dierentiate the last equation (n 1} times,
y
(n)
n
= y
(n1)
n1
+ (a
1
y
n
}
(n1)
.
Linear equations 49
Dierentiating (n 2} times the equation for y
n1
and substituting:
y
(n)
n
= y
(n2)
n2
(a
2
y
n
}
(n2)
+ (a
1
y
n
}
(n1)
.
We proceed, and eventually get an n-th order scalar equation for y
n
,
L
n
y = 0,
where the adjoint operator L
n
is given by
L
n
= (1}
n
d
n
dt
n
+ (1}
n1
d
n1
dt
n1
a
1
+ (1}
n2
d
n2
dt
n2
a
2

d
dt
a
n1
+ a
n
.
Theorem 2.4 (Lagrange identity) Suppose that the coecients a
j
in the operator
L
n
are dierentiable suciently many times. Then for every pair of functions u, v
that are n-times dierentiable,
v L
n
u u L
n
v = |uv|
,
where |uv| is dened as follows:
|uv| =
n
m=1
j+k=m1
(1}
j
u
(k)
(a
nm
v}
( j)
.
Proof : We will prove the case n = 2. The general proof is left as an exercise. For
n = 2 (which is what well end up using later anyway):
L
2
=
d
2
dt
2
+ a
1
d
dt
+ a
2
and
L
2
=
d
2
dt
2

d
dt
a
1
+ a
2
.
Thus,
v L
n
u u L
n
v = v(u
+ a
1
u
+ a
2
u} u(v
(a
1
v}
+ a
2
v}
= (u
v + a
1
u
v + a
2
uv} (u(v}
u(a
1
v}
+ ua
2
v}
= (u
v uv
+ ua
1
v}
.
B
50 Chapter 2
Corollary 2.2 (Greens formula) For every t
1
, t
2
:
t
2
t
1
v L
n
u u L
n
v dt = |uv|(t
2
} |uv|(t
1
}.
Proof : Obvious B
2.5.3 Non-homogeneous equation
We know how to express the solution of the non-homogeneous linear equation
in terms of the solution operator of the homogeneous system for a general rst-
order system. In this section we derive an expression for the solution of an non-
homogeneous n-th order linear equation,
(L
n
y}(t} = b(t}.
We may assume without loss of generality that a
0
(t} 1. The corresponding
rst-order system is
Y
(t} = A(t}Y(t} + B(t},

where A(t} is as above and
B(t} =
0
0
0
b(t}
.
The dierences between this problem and the general inhomogenous system is:
' The vector B(t} has all entries but one zero.
^ We are only looking for the 1st component of the solution.
Linear equations 51
Theorem 2.5 Let
1
, . . . ,
n
be n independent solution of the homogeneous equa-
tion L
n
y = 0, then the solution of the initial-value problem
(L
n
y}(t} = b(t} Y(t
0
} = Y
0
,
is
y(t} = y
h
(t} +
n
k=1
k
(t}
t
t
0
W
k
(
1
, ,
n
}(s}
W(
1
, ,
n
}(s}
b(s} ds,
where y
h
(t} is the solution to the homogeneous initial-value problem, and
W
k
(vp
1
, ,
n
} =
1

k1
0
k+1

n
k1
0
k+1

(n1)
1

(n1)
k1
1
(n1)
k+1

(n1)
n
Proof : We knowthat for we can express the non-homogeneous part of the solution
(in vector form) as
t
t
0
(t}
1
(s}B(s} ds.
Since we are only interested in the rst component and by the structure of B(t},
y(t} = y
h
(t} =
n
j=1
t
t
0
1 j
j
(t)
(t}
1
jn
(s}b(s} ds
Thus, it only remains to show that
1
jn
=
W
k
(
1
, ,
n
}
W(
1
, ,
n
}
,
but this follows from basic linear algebra (formula of the inverse by means of
cofactors). B
TA material 2.2 Apply this formula for n = 2.
52 Chapter 2
2.5.4 Constant coecients
We now consider the case of an n-th order equation with constant coecients:
L
n
y = y
(n)
+ a
1
y
(n+1)
+ + a
n
y,
where the a
j
s are now constant. In matrix form,
Y
(t} = AY(t},
where
A =
0 1 0 0 0
0 0 1 0 0

0 0 0 0 1
a
n
a
n1
a
n2
a
n4
a
1
.
Lemma 2.1 The characteristic polynomial of A is
p(} = det(I A} =
n
+ a
1
n1
+ + a
n1
+ a
n
.
Proof : We prove it inductively on n. For n = 2
I A = _
1
a
2
+ a
1
_,
and so
p
2
(A} = ( + a
1
} + a
2
=
2
+ a
1
+ a
2
.
For n = 3,
I A =

1 0
0 1
a
3
a
2
+ a
1
.
Hence,
p
3
(} = p
2
(} + a
3
.
It is easy to see that p
n
(} = p
n1
(} + a
n
. B
Linear equations 53
Theorem 2.6 Let
1
, . . . ,
s
be distinct roots of the characteristic polynomial of A
such that
j
has multiplicity m
j
,
s
j=1
m
j
= n.
Then the space of solutions is spanned by the functions
e
j
t
, t e
j
t
, t
2
e
j
t
, , t
m1
e
j
t
, j = 1, . . . , s.
Proof : We need to show that these n functions are indeed solutions and that they
are linearly independent. First, note that for every (not necessarily a root of the
characteristic polynomial):
L
n
{e
t
) = p
n
(} e
t
.
Let be a root of the characteristic polynomial of multiplicity m, and let 0 k
m 1. Then,
L
n
{t
k
e
t
) = L
n
_
d
k
d
k
e
t
_ =
d
k
d
k
L
n
{e
t
) =
d
k
d
k
{p
n
(}e
t
) .
If has multiplicity m and k < m, then
p
n
(} = p
n
(} = = p
(k)
n
(} = 0,
which proves that
L
n
{t
k
e
t
) = 0.
It remains to prove that these solutions are linearly independent. Suppose that
they were dependent. Then there would have existed constants c
jk
, such that
s
j=1
m
j
1
k=0
c
jk
t
k
- - -
P
j
(t)
e
j
t
= 0.
Dividing by e
1
t
,
P
1
(t} + e
(
2
1
)t
P
2
(t} + + e
(s
1
)t
P
s
(t} = 0.
54 Chapter 2
The P
j
are polynomials. Dierentiate suciently many times until P
1
disappears.
We will get an equation of the form:
e
(
2
1
)t
Q
2
(t} + + e
(s
1
)t
Q
s
(t} = 0.
We then divide by e
(
2
1
)t
, get
Q
2
(t} + e
(
3
2
)t
Q
3
(t} + + e
(s
2
)t
Q
s
(t} = 0.
We proceed this way until we get that Q
s
(t} = 0. It follows then backwards that
all the Q
j
s (or the P
j
s) vanish, i.e., all the c
j
s were zero. B
Example: Consider the equation
y
y = 0.
The characteristic polynomial is
4
1 = 0,
and its roots are = 1, i. This means that the linear space of solutions is spanned
by
e
t
, e
t
, e
it
and e
it
.
Since
cos t =
e
it
+ e
it
2
and sint =
e
it
e
it
2i
we may also change basis and use instead the real-valued functions
e
t
, e
t
, cos t and sint.
Example: Consider the equation

y
+ 2y
+ y = 0.
The characteristic polynomial is
(
2
+ 1}
2
= 0
and it has two roots of multiplicity 2: = i. Thus, the space of solution is
spanned by the basis
e
it
, te
it
, e
it
, and te
it
,
to after a change in bases,
cos t, t cos t, sint, and t sint.
Linear equations 55
2.6 Linear systems with periodic coecients: Flo-
quet theory
2.6.1 Motivation
Consider a linear homogeneous system of the form
y
(t} = A(t}y(t}, (2.3)

where for every t,
A(t + T} = A(t}.
We will assume that T is the smallest number for which this holds. Note that A
is automatically kT periodic for every integer k. Such equations are of interest in
systems in which the dynamics have some inherent periodicity (e.g., dynamics of
some phenomenon on earth that is aected by the position of the moon).
Another important application of linear systems with periodic coecients is the
stability of periodic solutions. Consider a (generally nonlinear) dierential sys-
tem:
y
(t} = f (y(t}}.
Note that f does not depend explicitly on t; such a system is called autonomous.
Suppose that this system has a particular solution, z(t}, that is T-periodic, and
that this solution passes through the point z
0
. Since the system is autonomous,
we can arbitrarily assume that z(0} = z
0
. An important question is whether this
periodic solution is stable. The periodic solution is said to be stable if solutions
with slightly perturbed initial data,
y(0} = z
0
+
0
,
where 1 do not diverge too fast from z(t}.
A standard way to analyze the stability of a particular solution is to represent the
solution as a sum of this particular solution plus a perturbation. In this case we set
y(t} = z(t} + (t}.
Substituting into the equation we get
z
(t} +
(t} = f (z(t} + (t}}.

56 Chapter 2
We then use Taylors theorem (assuming that f is twice dierentiable):
z
(t} +
(t} = f (z(t}} + f (z(t}}(t} + O(

2
}.
Since z
(t} = f (z(t}} we remain with
(t} = f (z(t}}(t} + O(}.

This is a linear equation with T-periodic coecients. The rationale of this ap-
proach is that if (t} remains bounded in time then for small enough , the per-
turbed solution will remain in an neighborhood of z(t}. Otherwise, the perturbed
solution diverges away from the periodic solution.
The above is a very rough analysis intended to motivate the study of linear systems
with periodic coecients.
Might be interesting to read in this context: the van der Pol oscillator. See:
http://www.scholarpedia.org/article/Van_der_Pol_oscillator
2.6.2 General theory
A natural question is whether the periodicity of the equation implies the periodic-
ity of the solution.
Example: Consider the scalar equation:
y
(t} = (1 + sint} y(t},

which is 2-periodic. One can check by direct substitution that the general solu-
tion is
y(t} = c exp(t cos t} ,
which is not periodic (in fact, it diverges as t ).
Theorem 2.7 Let (t} be a fundamental matrix for the periodic equation, then
(t + T} is also a fundamental solution, and moreover, there exists a constant
matrix B, such that
(t + T} = (t}B,
and
det B = exp_
T
0
Tr A(s} ds_.
Linear equations 57
Proof : Let (t} be a fundamental matrix and set (t} = (t + T}. Then,
(t} =
(t + T} = A(t + T}(t + T} = A(t}(t},

which proves the rst part.
As we know, if both (t} and (t} are fundamental matrices, then there exists a
constant matrix B, such that
(t} = (t + T} = (t}B.
Finally, we use Proposition 2.5 to get that
det (t + T} = det (t} _
t+T
t
Tr A(s} ds_,
and since A(t} is periodic, we may as well integrate from 0 to T,
det (t} det B = det (t} _
T
0
Tr A(s} ds_,
Note that in the last step we used the following property:
Lemma 2.2 Let f be periodic with period T, then for every t,
t+T
t
f (s} ds =
T
0
f (s} ds.
The fact that (t + T} = (t}B implies that the fundamental matrix is in general
not periodic. Since this relation holds for every t, it follow that
B =
1
(0}(T}.
Furthermore, we may always choose the fundamental matrix (t} such that (0} =
I, in which case B = (T}, i.e., for every t,
(t + T} = (t}(T}.
This means that the knowledge of the fundamental matrix in the interval |0, T|
determines the fundamental matrix everywhere.
58 Chapter 2
Moreover, it is easy to see by induction that for k N:
(t + kT} = (t}B
k
.
Thus, for arbitrary t,
(t} = (t mod T}
- --
T-periodic
B
tT
,
where
t mod T = t t]TT.
It follows that the behaviour of solution for large t is dominated by powers of the
matrix B. The forthcoming analysis will make this notion more precise.
Denition 2.1 The eigenvalues of B = (T},
i
are called the characteristic
multipliers ( ) of the periodic system. The constants
i
dened by
i
= e
i
T
are called the characteristic exponents ( ) of the periodic
linear system. (Note that the characteristic exponents are only dened modulo
2ik]T.)
Theorem 2.8 Let and be a pair of characteristic multiplier/exponent of the
periodic linear system. There there exists a solution y(t} such that for every t,
y(t + T} = y(t}.
Also, there exists a solution y(t} such that for every t,
y(t + T} = e
t
p(t},
where p(t} is T-periodic.
Proof : Let Bu = u, and dene y(t} = (t}u. By the denition of a fundamental
matrix, y(t} is a solution of (2.3) and
y(t + T} = (t + T}u = (t}Bu = (t}u = y(t},
which proves the rst part.
Linear equations 59
Next dene p(t} = e
t
y(t}, then
p(t + T} = e
(t+T)
y(t + T} = e
T
-
1
e
t
y(t} = p(t},
Suppose now that B is diagonalizable (i.e., has a complete set of eigenvectors).
Then there are n pairs (
i
,
i
}, and the general solution to the periodic linear sys-
tem is
y(t} =
n
k=1
k
e
k
t
p
k
(t},
where the functions p
k
(t} are T-periodic. In other words, every solution is of the
form
y(t} =

p
1
(t} p
n
(t}

1
t
e
nt
- --
fundamental matrix
C,
for some constant vector C.
An immediate corollary is that the system will have periodic solutions if one of
the characteristic exponents of B is zero.
Comment: At this stage the theory we developed is not useful in nding the funda-
mental matrix. Its importance is so far theoretical in that it reveals the structure of
the solution as a combination of exponential functions times periodic functions.
2.6.3 Hills equation
I ended up not teaching this section; makes sense only with a good problem
to study, like the stability of a limit cycle.
We will now exploit the results of the previous section to study a periodic (scalar)
second order equation of the form,
y
(t} + a(t}y(t} = 0,
where a(t + T} = a(t}. Such equations arise in mechanics when then force on the
particle is proportional to its location and the coecient of proportionality is peri-
odic in time (i.e., a pendulum in which the point of support oscillates vertically).
60 Chapter 2
Note that there is no general explicit solution to a linear second-order equation
with non-constant coecients.
As usual, we start by rewriting this equation as a rst order system,
y
(t} = A(t}y(t},
where
A(t} = _
0 1
a(t} 0
_.
Note that Tr A(t} = 0, which implies that det B = 1, i.e.,
1
2
= 1.
We are looking for a fundament matrix
(t} = _
11
(t}
12
(t}
21
(t}
22
(t}
_,
with (0} = I. Since,
1j
(t} =
2 j
(t},
it follows that
(t} = _
11
(t}
12
(t}
11
(t}
12
(t}
_.
Thus,
B = _
11
(T}
12
(T}
11
(T}
12
(T}
_.
Moreover, since the Wronskian is constant, W(t} = det B = 1, it follows that
11
(t}
12
(t}
12
(t}
11
(t} = 1.
The characteristic multiplier are the eigenvalues of B. For any 2-by-2 matrix
B = _
a b
c d
_
the eigenvalues are satisfy the characteristic equation
(a }(d } bc =
2
(Tr B} + det B = 0,
and since det B = 1 we get that
1,2
=
_
2
1,
Linear equations 61
where =
1
2
Tr B =
1
2
|
11
(T} +
12
(T}|. Note that
1
+
2
= 2 and
1
2
= 1.
The characteristic exponents are
1,2
= e
1,2
T
, which implies that
e
(
1
+
2
)T
= 1,
hence
1
+
2
= 0, and
2 = e
1
T
+ e
2
T
= 2cosh
1
T.
Thus, we are able to express the characteristic coecients in terms of the single
parameter :
cosh
1
T and
1
+
2
= 0.
Even though we do not have a general solution for this type of problems, it turns
out that the range of possible solutions can be determined by the single parameter
.
' Case 1: > 1 In this case
1,2
are both real and positive, and furthermore,
0 <
2
< 1 <
1
.
Thus,
1
and
2
=
1
are both real-valued. The general solution is of the
form
y(t} = c
1
e
1
t
p
1
(t} + c
2
e
1
t
p
2
(t},
where p
i
(t} are periodic. It follows that there are no periodic solutions and
that in general the solution diverges as t .
^ Case 2: < 1 In this case both
1,2
are real and negative,
2
< 1 <
1
< 0,
in which case
e
1
T
=
1
implies that
e
(
1
T)T
= |
1
|,
i.e.,
1
=
i
T
+ log|
1
|.
62 Chapter 2
Thus, the general solution is of the form
y(t} = c
1
e
log
1
t
e
itT
p
1
(t}
-- -
period 2T
+c
2
e
log
1
t
e
itT
p
2
(t}
-- -
period 2T
.
Note the doubling of the period. Once again there are no periodic solutions
and that in general the solution diverges as t .
Case 3: 1 < < 1 Now
1,2
are both complex and located on the unit
circle,
1,2
= e
iT
(this relation denes ) and thus
1,2
= i. In this case the general solution
is of the form
y(t} = c
1
e
it
p
1
(t} + c
2
e
it
p
2
(t},
with p
1,2
(t} periodic.
In this case the solutions remain bounded, however they will not be periodic,
unless it happens that T = 2]m for some integer m.
Case 4: = 1
I Case 5: = 1
Chapter 3
Boundary value problems
3.1 Motivation
Many problems in science involve dierential equations with conditions that are
specied at more than one point. Boundary value problems arise also in the study
of partial dierential equations.
Example: The equation that describes the motion of an elastic string is known as
the wave equation ( ),
2
y
t
2
= c
2
2
y
x
2
,
where the unknown is a function y(x, t} of space x and time t. The constant c is
the wave speed that depends on the mass density and tension of the string.
Consider a nite string, x |0, L| tied at its two ends, such that
(t} y(0, t} = y(L, t} = 0.
In addition, one has to specify initial conditions,
y(x, 0} = f (x} and
y
t
(x, 0} = g(x}.
(It is a initial-boundary problem.)
64 Chapter 3
This problem is analytically solvable. Note that this equation is linear, so that if
y
1
(x, t} and y
2
(x, t} are solutions to the wave equation, then so is any linear com-
bination. One method of solution is to look for solutions that have a separation
of variables:
y(x, t} = u(x}v(t}.
Substituting into the wave equation:
u
(x}v(t} = c
2
u(x}v
(t},
and further,
u
(x}
u(x}
= c
2
v
(t}
v(t}
.
Since this equation holds for all x and t, it must be the case that both sides are
equal to the same constant, namely, there exists a constant k, such that
u
(x} = ku(x} and v
(t} =
k
c
2
v(t}.
Note that we may have many such solutions for various values of k, and any linear
combination of these solutions is also a solution to the wave equation.
Consider the equation for u(x}. Because of the boundary conditions, we have a
boundary value problem:
u
(x} = ku(x} u(0} = u(L} = 0.

Does it have a solution? Yes, u 0. Are there other solutions? yes, u(x} =
sin(
kx}, but only if

k = m]L for some integer m. Are there more solutions?
Such questions are going to be addressed in the course of this chapter.
Because the independent variable in boundary value problem is often space rather
than time, we will denote it by x rather than t. Of course, a notation is nothing but
a notation.
Lets get back to the same boundary value problem, this time on the interval |0, 1|:
Ly = y
y(0} = y(1} = 0,
where
L =
d
2
dx
2
.
Boundary value problems 65
The parameter can be in general complex (the reason for the minus sign in L will
be made clear later). We know that the general solution to the equation Ly = y is
y(x} = Asin
12
x + Bcos
12
x.
Since we require y(0} = 0 then B = 0. Since we further require y(1} = 0, it follows
that
=
2
m
2
,
for m = 1, 2, . . . (m = 0 is the trivial solution and we ignore it as of now).
What we see is that the equation Ly = y with boundary values has a solution only
for selected values of . Does it remind us something? L is a linear operator,
is a scalar... yes, this looks like an eigenvalue problem! Indeed, the values of
for which the boundary value problem has a solution are called eigenvalues. The
corresponding eigenvectors are rather called eigenfunctions ( ).
Like for vectors, we endowthe space of functions on |0, 1| with an inner product:
( f , g} =
1
0
f (x}g(x} dx.
Then, the normalized eigenfunctions are
m
(x} =

2 sin(mx}.
Note that this is an orthonormal set,
(
m
,
k
} = 2
1
0
sin(mx} sin(kx} dx =
mk
.
The class of functions that are (well, Lebesgue...) square integrable, i.e,
1
0
| f (x}|
2
dx exists
is known as L
2
|0, 1|. It is a Hilbert space, which is an object that you learn about
in Advanced Calculus 2. In case you didnt take this course, no worry. We will
provide all the needed information.
For a function f L
2
|0, 1|, its inner-product with the elements of the orthogonal
set
k
are called its Fourier coecients,
f (k} = ( f ,
k
} =

2
1
0
f (x} sin(kx} dx.
66 Chapter 3
The series
k=1
f (k}
k
(x}
converges in the L
2
|0, 1| norm to f , namely,
lim
n
_f
n
k=1
f (k}
k
_ = 0.
Finally, the norm of f and its Fourier coecients satisfy a relation known as the
Parseval identity:
f
2
=

k=1
|

f (k}|
2
.
We have stated all these facts without any proof or any other form of justication.
As we will see, these results hold in a much wider scope.
TA material 3.1 Solve the wave equation initial-value problem. Talk about har-
monics. Talk about what happens when you touch the center of a guitar string.
Comment: An interesting connection between these innite-dimensional bound-
ary value problems and standard eigenvalue problems can be made by consid-
ering discretizations. Suppose that we want to solve the boundary value problem
y
(x} = y(x}, y(0} = y(1} = 0.

We can approximate the solution by dening a mesh,
x
i
=
i
N
, i = 0, . . . , N,
and look for y
i
= y(x
i
}, by approximating the dierential equation by a nite-
dierence approximation,
y
i1
+ y
i+1
2y
i
2x
2
, i = 1, . . . , N 1,
where x = 1]N is the mesh size. In addition we set y
0
= y
N
= 0. We get a linear
equation for the discrete vector (y
i
},
2 1 0 0 0
1 2 1 0 0
0 1 2 1 0

0 0 1 2 1
0 0 0 1 2
y
1
y
2
y
N2
y
N1
y
1
y
2
y
N2
y
N1
.
This is a standard eigenvalue problem, for which we expect N eigenvalues. As
we rene the mesh and take N , the number of eigenvalues is expected to
tend to innity.
Exercise 3.1 Find the eigenvalues and eigenvectors of this discretized system,
and verify their dependence on N. Compare your results with the solution of the
continuous equation.
3.2 Self-adjoint eigenvalue problems on a nite in-
terval
3.2.1 Denitions
Consider a linear n-th order dierential operator,
L = p
0
(x}
d
n
dx
n
+ p
2
(x}
d
n1
dx
n1
+ + p
n1
(x}
d
dx
+ p
n
(x},
where p
j
C
nj
(a, b} is in general complex-valued; we assume that p
0
(x} 0 for
all x (a, b}. We also dene a linear operator U C
n
(a, b} R
n
of the form
U
k
y =
n
j=1
{M
k j
y
( j1)
(a} + N
k j
y
( j1)
(b}) , k = 1, . . . , n,
where the M
jk
and N
jk
are constant. That is,
U
k
y = M
k1
y(a} + M
k2
y
(a} + + M
kn
y
(n1)
(a}
+ N
k1
y(b} + N
k2
y
(b} + + N
kn
y
(n1)
(b}.
The equation
Ly = y, Uy = 0
is called an eigenvalue problem.
Example: The example we studied in the previous section is such an instance,
with n = 2,
p
0
(x} = 1, p
1
(x} = p
2
(x} = 0,
68 Chapter 3
and
M
11
= N
21
= 1, and other M
jk
, N
jk
are zero,
i.e.,
U
1
y = M
11
y(a} + N
11
y(b} + M
12
y
(a} + N
12
y
(b} = y(a}
U
2
y = M
21
y(a} + N
21
y(b} + M
22
y
(a} + N
22
y
(b} = y(b}.
Example: Another class of examples for n = 2:

M
11
= N
11
= 1, M
22
= N
22
= 1, and other M
jk
, N
jk
are zero,
i.e.,
U
1
y = M
11
y(a} + N
11
y(b} + M
12
y
(a} + N
12
y
(b} = y(a} y(b}

U
2
y = M
21
y(a} + N
21
y(b} + M
22
y
(a} + N
22
y
(b} = y
(a} y
(b}.
This corresponds to periodic boundary conditions.
Denition 3.1 The eigenvalue problem is said to be self-adjoint ( ) if
(Lu, v} = (u, Lv}
for every u, v C
n
(a, b} that satisfy the boundary conditions Uu = Uv = 0.
Example: The two above examples are self-adjoint. In the rst case Uu = 0
implies that u(a} = u(b} = 0, and then
(Lu, v} =
b
a
u
(x}v(x} dx =
b
a
u
(x}v
(x} dx =
b
a
u(x}v
(x} dx = (u, Lv}.
Recall the adjoint operator L
derived in the previous chapter:

L
= (1}
n
d
n
dx
n
p
0
(x} + (1}
n1
d
n1
dx
n1
p
1
(x} +
d
dx
p
n1
(x} + p
n
(x}.
The Green identity proved in the previous chapter showed that
(Lu, v} (u, L
v} = |uv|(b} |uv|(a}.
Thus, if L
= L and Uu = Uv = 0 implies that |uv|(b} = |uv|(a}, then the problem

is self-adjoint.
A self-adjoint problem always has a trivial solution for any value of . Values of
for which a non-trivial solution exists are called eigenvalues; the corresponding
functions are called eigenfunctions.
3.2.2 Properties of the eigenvalues
Proposition 3.1 Consider a self-adjoint boundary value problem. Then,
' The eigenvalues are real.
^ Eigenfunctions that correspond to distinct eigenvalues are orthogonal.
The eigenvalues form at most a countable set that has no (nite-valued)
accumulation point.
Comment: At this stage there is no guarantee that eigenvalues exist, so as of now
this theorem can hold in a vacuous sense.
Comment: The last part of the proof relies on complex analysis. Fasten your seat
belts.
Proof : Let be an eigenvalue with eigenfunction . Then,
(, } = (L, } = (, L} = (, },
i.e.,
( }(, } = 0,
which proves that is real.
Let be distinct eigenvalues,
L = , L = .
Then,
(, } = (L, } = (, L} = (, },
i.e.,
( }(, } = 0,
which implies that and are orthogonal.
Let now
j
(x, }, j = 1, . . . , n, be solutions of the equation L
j
=
j
with initial
conditions,
(k1)
j
(c, } =
jk
,
70 Chapter 3
where c is some point in (a, b}. That is,
1
(c} = 1
1
(c} =
1
(c} =
1
(c} = =
(n1)
1
(c} = 0
2
(c} = 1
2
(c} =
2
(c} =
2
(c} = =
(n1)
2
(c} = 0
3
(c} = 1
3
(c} =
3
(c} =
3
(c} = =
(n1)
3
(c} = 0
etc.
Note, we are back in initial values problems; there no issue of existence nor non-
uniqueness, i.e., the
j
(x, } are dened for every (complex) . Also, the functions
j
are continuous both in x and . In fact, for xed x the function (x, } is
analytic.
1
The functions
j
are linearly independent, hence every solution of the equation
Ly = y is of the form
y(x} =
n
j=1
c
j
j
(x, }.
For to be an eigenvalue we need
U
y =
n
j=1
c
j
U
j
(, } = 0, = 1, . . . , n,
i.e.,
n
j=1
n
k=1
[M
k
(k1)
j
(a, } + N
k
(k1)
j
(b, })
-- -
A
j
()
c
j
= 0, = 1, . . . , n.
This is a homogeneous linear system for the coecients c
j
. For it to have a non-
trivial solution the determinant of the matrix A
j
(} must vanish. That is, the
eigenvalues are identied with the roots of det A
j
(}.
det A
j
(} is analytic in the entire complex plane. It is not identically zero because
it can only vanish on the real line. It is well-known (especially if you learn it...)
that the zeros of an entire analytic function that it not identically zero do not have
accumulation points (except at innity). B
1
A concept learned in complex analysis; I will say a few words about it.
3.2.3 Non-homogeneous boundary value problem
We turn now to study the non-homogeneous equation
Ly = y + f , Uy = 0,
where f C(a, b}. What to expect? Think of the linear-algebraic analog,
Ax = x + b.
A (unique) solution exists if the matrix A I is not singular, i.e., if is not an
eigenvalue of A. Similarly, we will see that the non-homogeneous boundary value
problem has a solution for values of that are not eigenvalues of the homogeneous
equation.
Example: Such boundary value problems may arise, for example, in electrostat-
ics. The Poisson equation for the electric potential in a domain R
n
with a
metallic boundary is
= 4, |
= 0,
where is the Laplacian and (x} is the charge density. For n = 1 we get
d
2
dx
2
= 4 (0} = (L} = 0,
which is of the same type as above with = 0.
Let
j
(x, } be as in the previous section solutions to L
j
(, } =
j
(, }, subject
to the boundary conditions
(k1)
j
(c, } =
jk
.
Note that
W(
1
, . . . ,
n
}(c} = det I = 1,
hence
W(
1
, . . . ,
n
}(x} = exp_
x
c
p
1
(s}
p
0
(s}
ds_.
Note that p
0
and p
1
are the same for L and L id.
72 Chapter 3
We now dene the following function:
K(x, , } =
1
p
0
(}W(
1
, . . . ,
n
}(}
1
(, }
n
(, }
1
(, }
n
(, }

(n2)
1
(, }
(n2)
n
(, }
1
(x, }
n
(x, }
x
0 x < .
This function depends on x and , both taking values in |a, b|. It is apparently
discontinuous, however note that
K(
+
, , } =
d
dx
K(
+
, , } = =
d
n2
dx
n2
K(
+
, , } = 0,
hence K and its rst n 2 derivatives with respect to x are continuous over all
(x, }; in particular K is continuous for n 2. The (n 1}-st and n-th derivatives
are continuous in every subdomain. Also,
d
n1
dx
n1
K(
+
, , } =
1
p
0
(}
.
Finally, for x > ,
LK(x, , } =
1
p
0
(}W(
1
, . . . ,
n
}(}
1
(, }
n
(, }
1
(, }
n
(, }

(n2)
1
(, }
(n2)
n
(, }
L
1
(x, } L
n
(x, }
= K(x, , }.
This relation holds trivially for x < .
Dene now the function
u(x, } =
b
a
K(x, , } f (} d =
x
a
K(x, , } f (} d.
We calculate its rst n derivatives with respect to n,
u
(x, } =
x
a
d
dx
K(x, , } f (} d + K(x, x
, } f (x}
-- -
0
,
u
(x, } =
x
a
d
2
dx
2
K(x, , } f (} d +
d
dx
K(x, x
, } f (x}
-- -
0
,
and so until
u
(n1)
(x, } =
x
a
d
n1
dx
n1
K(x, , } f (} d +
d
n2
dx
n2
K(x, x
, } f (x}
- - -
0
.
Finally,
u
(n)
(x, } =
x
a
d
n
dx
n
K(x, , } f (} d +
d
n1
dx
n1
K(x, x
, } f (x}
- - -
f (x)p
0
(x)
.
It follows that
Lu(x, } =
x
a
LK(x, , } f (} d + f (x} = u(x, } + f (x},
i.e., u(x, } is a solution to the non-homogeneous equation.
We are looking for a solution of the non-homogeneous equation that satises the
boundary conditions, U
k
y = 0. Dene
G(x, , } = K(x, , } +
n
j=1
c
j
(}
j
(x, },
and set c
j
(} such that for every k = 1, . . . , n and every |a, b|:
U
k
G(, , } = U
k
K(, , } +
n
j=1
c
j
(}U
k
j
(, } = 0.
This linear system has a unique solution if det U
k
j
(, } 0, i.e., if is not an
eigenvalue of the homogeneous boundary value problem.
Set then
u(x, } =
b
a
G(x, , } f (} d
=
b
a
K(x, , } f (} d +
n
j=1
j
(x}
b
a
c
j
(} f (} d.
74 Chapter 3
Then,
Lu(x, } = u(x, },
and
U
k
u(, } =
b
a
_U
k
K(, , } +
n
j=1
c
j
(}U
k
(, }_ f (} d = 0.
Thus, we found a solution to the non-homogeneous boundary value problem for
every value of that is not an eigenvalue of the homogeneous boundary value
problem.
Exercise 3.2 Solve the boundary value problem,
y
(x} = y(x} + f (x}, y(0} = y(1} = 0.

For what values of solutions exist? Repeat the above analysis (including the
calculation of K(x, , } and G(x, , } for this particular example.
Solution 3.2: We start by constructing the two independent solution,
1
(x) =
1
(x),
1
(0) = 1
1
(0) = 0
2
(x) =
2
(x),
2
(0) = 0
2
(0) = 1,
namely,
1
(x) = cos
12
x and
2
(x) =
12
sin
12
x.
Thus,
W(
1
,
2
)(x) =
cos
12
x
102
sin
12
x
12
sin
12
x cos
12
x
= 1.
Next, since p
0
() = 1,
K(x, , ) =
cos
12
sin
12
cos
12
x sin
12
x
= sin
12
( x) x
0 x <
.
Now,
G(x, , ) = K(x, , ) c
1
() cos
12
x c
2
()
12
sin
12
x.
The coecients c
1,2
have to satisfy,
G(0, , ) = 0 c
1
() = 0,
and
G(1, , ) = sin
12
( 1) c
1
() cos
12
c
2
()
12
sin
12
= 0,
from which we get that
G(x, , ) = K(x, , )
sin
12
( 1)
sin
12
sin
12
x
Thus, the solution to the boundary value problem is
u(x, ) =
x
0
sin
12
( x) f () d
sin
12
x
sin
12

1
0
sin
12
( 1) f () d.
It is easy to see that the boundary conditions are indeed satised. You may now check by direct
dierentiation that the dierential equation is indeed satised.
Suppose now that zero is not an eigenvalue of the homogeneous boundary value
problem, and consider the boundary value problem
Ly = f Uy = 0.
The fact the we require zero not to be an eigenvalue of f will turn out not to be
a real restriction since there exists a
0
that is not en eigenvalue and then L
0
I
will be used as an operator for which zero is not en eigenvalue.
We will denote G(x, , 0} simply by G(x, }. Note that by denition LG(x, } = 0
at all points where x .
Dene the linear integral operator:
G f
b
a
G(, } f (} d,
which maps the non-homogeneous term into the solution of the boundary value
problem.
Proposition 3.2 The integral operator G (and the Green function G) satises the
following properties:
' LG f = f and UG f = 0 for all f C(a, b}.
^ (G f , g} = ( f , Gg}.
G(x, } = G(, x}.
GLu = u for all u C
n
(a, b} satisfying Uu = 0.
(Note that L and G are almost inverse to each other.)
76 Chapter 3
Proof : The rst statement follows from the fact that G maps f into the solution
of the boundary value problem. By the self-adjointness of L and the fact that
UG f = UGg = 0,
(LG f , Gg} = (G f , LGg},
hence by the rst statement
( f , Gg} = (G f , g},
which proves the second statement.
It follows from the second statement that
b
a
_
b
a
G(x, } f (} d_g(x} dx =
b
a
f (x}_
b
a
G(x, }g(} d_dx.
Changing the order of integration and interchanging x and in the second integral:
b
a

b
a
[G(x, } G(, x}) f (}g(x} dxd = 0.
Since this holds for every f and g the third statement follows.
Finally, for every g C(a, b} and u C
n
(a, b} satisfying Uu = 0,
(GLu, g} = (u, LGg} = (u, g},
which implies that GLu = u. B
3.3 The existence of eigenvalues
So far we havent shown that the eigenvalue problem
Ly = y, Uy = 0, (3.1)
has solutions. We only know properties of solutions if such exist. In this section
we will show that self-adjoint eigenvalue problems always have a countable set of
eigenvalues.
Assuming that zero is not an eigenvalue we dened the Green function G(x, }
and the operator G. The rst step will be to show that eigenfunctions of L coincide
with eigenfunctions of G:
Proposition 3.3 is an eigenfunction of (3.1) with eigenvalue if and only if it
is also an eigenfunction of G with eigenfunction 1], namely
G =
1
, (3.2)
Proof : Let (, } be eigensolutions of (3.1). Then,
G =
1
G =
1
GL,
and since U = 0, it follows from the fourth statement of Proposition 3.2 that
G =
1
.
Conversely, suppose that (3.2) holds. Then,
L = L
1
= LG = ,
and
U = UG = 0,
where we used the fact that the range of G is functions that satisfy the boundary
conditions. B
Thus, in order to show that (3.1) has eigenvalues we can rather show that G has
eigenvalues.
Lemma 3.1 The set of functions
X = Gu u C(0, 1}, u 1}
is bounded (in the sup-norm) and equicontinuous (but the norm is the L
2
-
norm).
78 Chapter 3
Proof : We will consider the case of n 2; the case of n = 1 requires a slightly
dierent treatment. For n 2 the function G(x, } is continuous on the square
(x, } |a, b|
2
, hence it is uniformly continuous, and in particular,
( > 0}( > 0} ( |a, b|} (x
1
, x
2
|x
1
x
2
| < } |G(x
1
, }G(x
2
, }| < .
Thus,
( > 0}( > 0} (u X} (x
1
, x
2
|x
1
x
2
| < }
it holds that
|Gu(x
1
} Gu(x
2
}|
b
a
|G(x
1
, } G(x
2
, }||u(}| d
b
a
|u(}| d.
Using the Cauchy-Schwarz inequality,
|Gu(x
1
} Gu(x
2
}| _
b
a
|u(}|
2
d_
12
_
b
a
d_
12
= (b a}
12
u,
which proves the equicontinuity of X.
Since G(x, } is continuous, it is also bounded; let K be a bound. Then, for all
u X and x |a, b|,
|Gu(x}|
b
a
|G(x, }||u(}| d K(b a}
12
u,
which proves the uniform boundedness of X. B
It follows from the Arzela-Ascoli theorem:
Corollary 3.1 Every sequence in X has a subsequence that converges uniformly
on |a, b|.
The map G is actually a map from L
2
(a, b} L
2
(a, b}.
2
We therefore dene its
operator norm:
G = sup
u=1
Gu = sup
(u,u)=1
(Gu, Gu}
12
.
It follows from Lemma 3.1 that G < (i.e., G is a bounded linear operator).
2
We dened it as a map from C(a, b) to C
n
(a, b) to avoid measure-theoretic subtleties.
Proposition 3.4
G = sup
u=1
|(Gu, u}|.
Proof : For every normalized u, it follows from the Cauchy-Schwarz inequality
and the denition of the norm that
|(Gu, u}| Guu Gu
2
= G,
i.e.,
sup
u=1
|(Gu, u}| G. (3.3)
By the bilinearity of the inner-product and the self-adjointness of G, for every u
and v,
(G(u + v}, u + v} = (Gu, u} + (Gv, v} + (Gu, v} + (Gv, u}
= (Gu, u} + (Gv, v} + 2R(Gu, v}
sup
w=1
|(Gw, w}|u + v
2
.
and
(G(u v}, u v} = (Gu, u} + (Gv, v} (Gu, v} (Gv, u}
= (Gu, u} + (Gv, v} 2R(Gu, v}
sup
w=1
|(Gw, w}|u v
2
.
Subtracting the second inequality from the rst,
4R(Gu, v} 2 sup
w=1
|(Gw, w}| {u
2
+ v
2
) .
Substitute now v = Gu]Gu. Then,
4R_Gu,
Gu
Gu
_ 2 sup
w=1
|(Gw, w}|
u
2
+ _
Gu
Gu
_
2
,
or,
4Gu 2 sup
w=1
|(Gw, w}| {u
2
+ 1) .
80 Chapter 3
For u = 1 we get
Gu sup
w=1
|(Gw, w}|,
and since this holds for every such u,
G = sup
u=1
Gu sup
w=1
|(Gw, w}|. (3.4)
Eqs. (3.3) and (3.4) together give the desired result. B
Theorem 3.1 Either G or G is an eigenvalue of G.
Comments:
' This proves the existence of an eigenvalue to the boundary value problem.
^ Since there exists an eigenfunction such that
G = G or G = G,
it follows that
G = G,
thus the sup in the denition of G is attained.
Everything we have been doing holds in the nite-dimensional case. If A is
an hermitian matrix then either A or A is an eigenvalue.
Proof : Denote G =
0
. then either
0
= sup
u=1
(Gu, u}
or
0
= sup
u=1
(Gu, u}.
Consider the rst case; the second case is treated similarly.
By the denition of the supremum, there exists a sequence of functions (u
n
} such
that
u
n
= 1 lim
n
(Gu
n
, u
n
} =
0
.
By Corollary 3.1 the sequence (Gu
n
} has a subsequence (not relabeled) that con-
verges uniformly on |a, b|; denote its limit by
0
. In particular, uniform conver-
gence implies L
2
-convergence,
lim
n
Gu
n

0
= 0.
Since
0 (Gu
n

0
, Gu
n

0
} = Gu
n
2
+
0
2
2R(Gu
n
,
0
} Gu
n
2

0
2
,
we get the convergence of the norm,
lim
n
Gu
n
=
0
.
(It is always true that convergence in norm implies the convergence of the norm.)
Now,
Gu
n

0
u
n
2
= Gu
n
2
+
2
0
2
0
(Gu
n
, u
n
}
0
2

2
0
,
from which follows that
0
0 (we need this to claim that
0
is nontrivial).
Moreover, it follows that
0 Gu
n

0
u
n
2
2
2
0
2
0
(Gu
n
, u
n
} 0,
from which follows that
lim
n
Gu
n

0
u
n
= 0.
Finally, by the triangle inequality
G
0

0
0
G
0
G(Gu
n
} + G(Gu
n
}
0
Gu
n
+
0
Gu
n

0
G
0
Gu
n
+ GGu
n

0
u
n
+
0
Gu
n

0
.
Letting n we get that
G
0
=
0
0
,
Theorem 3.2 G has an innite number of eigenfunctions.
82 Chapter 3
Proof : Let
0
and
0
be as above and assume that
0
has been normalized,
0
=
1. Dene a new kernel
G
1
(x, } = G(x, }
0
0
(x}
0
(},
and the corresponding operator,
G
1
f (x} =
b
a
G
1
(x, } f (} d.
This operator is also self-adjoint,
(G
1
u, v} = (Gu, v} +
0
b
a
0
(x}
0
(}u(}v(x} ddx = (u, G
1
v}.
Hence, unless G
1
= 0 (the zero operator), it has an eigenvalue
1
, whose absolute
value is equal to G
1
. Let,
G
1
1
=
1
1

1
= 1.
Note that G
1
0
= 0, hence
(
1
,
0
} =
1
1
(G
1
1
,
0
} =
1
1
(
1
, G
1
0
} = 0,
i.e.,
0
and
1
are orthogonal. Then,
G
1
= G
1
1
+
0
1
(
1
,
0
} = G
1
1
=
1
1
,
which means that
1
is also an eigenfunction of G. By the maximality of
0
,
|
1
| |
0
|.
We continue this way, dening next
G
2
(x, } = G
1
(x, }
1
1
(x}
1
(}
and nd a third orthonormal eigenfunction of G.
The only way this process can stop is if there is a G
m
= 0 for some m. Suppose this
were the case, then for every f C(a, b},
G
m
f (x} = G f (x}
m1
j=1
j
(x}( f ,
j
} = 0.
Applying L on both sides, using the fact that LG f = f and L
j
=
1
j

j
,
f (x} =
m1
j=1
j
(x}( f ,
j
}.
This would imply that every continuous function is in the span of m dierentiable
functions, which is wrong. B
3.4 Boundary value problems and complete orthonor-
mal systems
We have thus proved the existence of a countable orthonormal system,
j
, with
corresponding eigenvalues
|
0
| |
1
| |
2
| .
Lemma 3.2 (Bessels inequality) Let f L
2
(a, b}. Then the sequence
j=0
|( f ,
j
}|
2
converges and
j=0
|( f ,
j
}|
2
f
2
.
Proof : For every m N,
0 _f
m
j=0
( f ,
j
}
j
_
2
= f
2
j=0
|( f ,
j
}|
2
,
which proves both statements. B
84 Chapter 3
Theorem 3.3 Let f C
n
(a, b} and U f = 0, then
f =

j=0
( f ,
j
}
j
,
where the convergence is uniform.
Corollary 3.2 An immediate consequence is Parsevals identity. Multiplying this
equation by

f and integrating over |a, b|, interchanging the order of integration
and summation, we get
f
2
=

j=0
|( f ,
j
}|
2
Proof : For every j Mand x |a, b|,
(G
j
}(x} =
b
a
G(x, }
j
(} d =
j
j
(x}.
Taking the complex conjugate and using the fact that G(x, } = G(, x},
(G(x, },
j
} =
b
a
G(, x}
j
(} d =
j
j
(x}.
Squaring and summing over j,
m
j=0
|(G(x, },
j
}|
2
=
m
j=0
|
j
|
2
|
j
(x}|
2
=
m
j=0
|
j
|
2
.
By Bessels inequality applied to G(x, },
m
j=0
|
j
|
2
G(x, }
2
=
b
a
|G(x, }|
2
d.
Integrating over x,
m
j=0
|
j
|
2
b
a
|G(x, }|
2
ddx K
2
(b a}
2
,
where K is a bound on |G(x, }| It follows at once that
j=0
|
j
|
2
converges and in
particular
j
0.
For m N consider
G
m
(x, } = G(x, }
m1
j=0
j
(x}
j
(}.
By the way we constructed the sequence of eigenfunctions, we know that
G
m
= |
m
|.
Thus, for every u C(a, b},
G
m
u = _Gu
m1
j=0
j
(u,
j
}
j
_ |
m
|u.
Letting m ,
lim
m
_Gu
m1
j=0
j
(u,
j
}
j
_ = 0.
This means that
j=0
j
(u,
j
}
j
converges to Gu in L
2
. We will show that it
converges in fact uniformly. For q > p,
q
j=p
j
(u,
j
}
j
= G _
q
j=p
(u,
j
}
j
_.
Hence,
_
q
j=p
j
(u,
j
}
j
_ K(b a}
12
_
q
j=p
(u,
j
}
j
_ = K(b a}
12
_
q
j=p
|(u,
j
}|
2
_
12
,
where we used the Cauchy-Schwarz inequality. Since the right hand side vanishes
as p, q , it follows by Cauchys criterion that
q
j=p
j
(u,
j
}
j
converges uni-
formly. Since it converges to Gu in L
2
and the latter is continuous, it follows that
j=0
j
(u,
j
}
j
= Gu,
where the convergence is uniform.
Let now f C
n
(a, b}, U f = 0. Then, Lf is continuous, and
j=0
j
(Lf ,
j
}
j
= GLf .
86 Chapter 3
Since GLf = f and L
j
=
1
j

j
,
j=0
( f ,
j
}
j
= f .
B
Chapter 4
Stability of solutions
4.1 Denitions
Consider the following nonlinear system:
y
1
=
1
3
(y
1
y
2
}(1 y
1
y
2
}
y
2
= y
1
(2 y
2
}.
This system does not depend explicitly on t; such systems are called autonomous.
More generally, in a system
y
(t} = f (t, y(t}},

the family of vector elds f (t, } is called autonomous if it does not depend on t,
i.e., f R
n
R
n
.
Note that if is a solution to the autonomous initial data problem,
y
(t} = f (y(t}}, y(0} = y

0
,
then (t} = (t t
0
} is a solution to the initial data problem
y
(t} = f (y(t}}, y(t

0
} = y
0
.
Fixed points For the above system, four points in the y-plane that are of special
interest:
y = (0, 0} y = (2, 2} y = (0, 1} and y = (1, 2}.
88 Chapter 4
These are xed points ( ), or equilibrium points ( ),
or stationary points. Each such point corresponds to a solution to the system that
does not change in time. If any of these points is prescribed as an initial value
then the solution will be constant.
The above system is nonlinear and cannot be solved by analytical means for ar-
bitrary initial values. Yet, in many cases, the global behaviour of solutions can
be inferred by knowing the behaviour of solutions in the vicinity of xed points.
Roughly speaking, a xed point is stable if any solution starting close to it will
be attracted to it asymptotically, and it is unstable if in any neighborhood there
exist points such that a solution starting at this point is repelled from the xed
point.
The notion of stability can be made more precise and more general:
Denition 4.1 Let u(t} be a solution to the autonomous system y
(t} = f (y(t}}.
It is called stable in the sense of Lyapunov if for every > 0 there exists a > 0,
such that for every solution y(t} satisfying y(t
0
} u(t
0
} < ,
y(t} u(t} < t > t
0
.
It is called asymptotically stable if there exists a such that y(t
0
} u(t
0
} <
implies that
lim
t
y(t} u(t} = 0.
Comment: The stability of xed points is a particular case of stability of solutions
as dened above.
We will be concerned in this chapter with the global behaviour of solutions. here
is a list of possible behaviors:
' Fixed point.
^ Asymptotic convergence to a xed point.
Limit cycle (periodic solution).
Asymptotic convergence to a limit cycle.
I Quasi-periodic.
I Chaotic.
I Tends asymptotically to innity.
Stability of solutions 89
4.2 Linearization
Consider an autonomous system
y
(t} = f (y(t}}, (4.1)

and a particular solution u(t}. Since we are interested in understanding the be-
haviour of solutions in the vicinity of u(t}, we set
y(t} = u(t} + z(t}.
Substituting in the equation we get
u
(t} + z
(t} = f (u(t} + z(t}}.

Since u
(t} = f (u(t}}, we get a dierential equation for the deviation z(t} of the
solution y(t} from the solution u(t}:
z
(t} = f (u(t} + z(t}} f (u(t}}.

If f is twice continuously dierentiable, then as long as z(t} is in some compact
set U R
n
.
z
(t} = Df (u(t}}z(t} + O(|z(t}|

2
}.
It is sensible to think that if z(t
0
} is very small, then at least as long as it remains
so, the quadratic correction to the Talor expansion is negligible, hence the devision
of the solution from u(t} is governed by the linearized equation,
y
(t} = Df (u(t}}y(t}, (4.2)

hence the stability of u(t} in the nonlinear equation (4.1) can be determined by
the stability of of the solution y = 0 of the linear equation (4.2).
The task is therefore 2-fold:
' Showthat the zero solution of the linearized equation (4.2) is stable/unstable;
a solution u(t} of (4.1) for which the zero solution of the linearized equa-
tions is stable is called linearly stable ( ).
^ Show that the stability/instability of the zero solution of (4.2) implies the
stability/instability of the solution u(t} of the nonlinear equation (4.1). (That
is, that linear stability implies nonlinear stability.)
90 Chapter 4
The rst task is generally dicult, since we do not have a general way of solving
linear equations with non-constant coecients. The second task is in fact not
always possible; one can nd situations in which a solution is linearly stable but
nonlinearly unstable.
The rst task is however easy if u(t} is a constant solution, i.e., if u is a xed point.
In this case, we know the solution to the linearized equation,
y(t} = e
Df (u) t
y(0}.
Moreover, we know the form of the exponential e
Df (u) t
, and can therefore deter-
mine the following:
Theorem 4.1 A xed point u of (4.1) is linearly stable if and only if (i) all the
eigenvalues of Df (u} have non-positive real parts, and (ii) if an eigenvalue has a
zero real-part then it is of multiplicity one.
While we already know this is true, we will also prove it using a new technique.
Denition 4.2 A xed point u is called hyperbolic if none of the eigenvalues of
Df (u} has real part equal to zero. It is called a sink ( ) if all the eigenvalues
have negative real parts; it is called a source ( ) all the eigenvalues have
positive real parts; it is called a center ( ) if all the eigenvalues are imaginary;
it is called a saddle ( ) if certain eigenvalues have positive real parts while
other have negative real parts.
Example: Consider the unforced Dung oscillator,
y
1
= y
2
y
2
= y
1
y
3
1
y
2
,
where 0. This describes a particle that is repelled from the origin when close
it, but attracted to it when far away from it.
This system has three xed points,
(0, 0} and (1, 0}.
Let u = (u
1
, u
2
} be the xed point. The linearized equation is
y
1
= y
2
y
2
= y
1
3u
2
1
y
1
y
2
,
i.e.,
Df (u} = _
0 1
1 3u
2
1

_.
The eigenvalues satisfy
( + } (1 3u
2
1
} = 0,
which implies that
=
1
2
_
_
2
+ 4(1 3u
2
1
}_.
For the xed point (0, 0} there is a positive and a negative eigenvalue, which im-
plies that it is unstable. For the xed point (1, 0} both eigenvalues have negative
real parts, which implies that these xed points are stable.
4.3 Lyapunov functions
Theorem 4.2 Let f R
n
R
n
be twice dierentiable; let u be a xed point of
f . Suppose that all the eigenvalues of Df (u} have negative real parts. Then u is
asymptotically stable.
Comment: This is a situation in which linear stability implies nonlinear stability.
Comment: Without loss of generality we may assume that u = 0, for dene z(t} =
y(t} u. Then,
z
(t} = f (z + u} F(z}.
Zero is a xed point of F, which is stable if and only if u is a stable xed point of
f ; furthermore,
DF(0} = Df (u}.
92 Chapter 4
Proof : The idea of the proof is based on a so-called Lyapunov function. Suppose
we construct a function V R
n
R that satises the following conditions:
' V(y} 0 for all y.
^ V(y} = 0 i y = 0.
The sets
K
c
= y R
n
V(y} c}
are compact, connected and contain the origin as internal point.
Close enough to the xed point, the vector eld f points inward on the
level sets of V. That is, there exists a neighborhood U of the origin, such
that
y U (DV(y}, f (y}} 0.
Then, solutions are trapped within those level sets.
Back to the proof. There exists a neighborhood U of the origin and a constant
C > 0, such that
y U f (y} Df (0}y Cy
2
.
Dene the Lyapunov function
V(y} =
1
2
y
2
,
and set
g(t} = V(y(t}}.
By the chain rule,
g
(t} = (y(t}, f (y(t}}} = (y(t}, Df (0}y(t}} + (y(t}, f (y(t}} Df (0}y(t}}

- - -
R(t)
.
Since all the eigenvalues of Df (0} have negative real part, there exists a constant
> 0 such that
(y, Df (0}y} < y
2
,
i.e.,
g
(t} < 2g(t} + R(t}.

Moreover, if y(t} U, then
R(t} y(t} Cy(t}
2
C
1
g
32
(t}.
In particular, there exists a constant k > 0 such that K
c
U, and as long as
y(t} U
(t} < g(t}.

If follows that if g(0} < c then g(t} < c for all t > 0.
In addition,
d
dt
(e
t
g(t}} < 0,
and further,
e
t
g(t} < g(0},
which implies that
lim
t
g(t} = 0,
and in turn this implies that y(t} 0, i.e., the solution is asymptotically stable. B
Exercise 4.1 Show that if A is a real matrix whose eigenvalues have all nega-
tive real parts, then there exists a constant > 0, such that
x R
n
(Ay, y} < y
2
.
Example: An important use of Lyapunov functions is in cases where linear stabil-
ity cannot conclusively determine the (nonlinear) stability of a xed point. Con-
sider the following system,
y
1
= y
2
y
2
= y
1
y
2
1
y
2
.
The point (0, 0} is a xed point. The linearized equation about the origin is
y
= _
0 1
1 0
_y,
and the eigenvalues are i. This means that (0, 0} is a center and we have no
current way to determine its (nonlinear) stability.
Consider then the Lyapunov function
V(y} =
1
2
(y
2
1
+ y
2
2
}.
Then,
(DV(y}, f (y}} = (y
1
, y
2
} _
y
2
y
1
y
2
1
y
2
_ = y
2
1
y
2
2
94 Chapter 4
The origin is therefore stable for > 0.
Example: Consider the system

y
1
= 3y
2
y
2
= y
1
+ (2y
3
2
y
2
},
where > 0. The origin is a xed point, and
Df (0} = _
0 3
1 0
_
The eigenvalues are
3i, i.e., the origin is a center. Consider the Lyapunov

function
V(y} =
1
2
(y
2
1
+ 3y
2
2
}.
Setting g(t} = V(y(t}},
g
(t} = y
1
(3y
2
} + 3y
2
y
1
+ (2y
3
2
y
2
} = (3y
2
2
6y
4
2
}.
It follows that g
(t} 0 as long as y
2
2
<
1
2
, i.e., if
g(0} < 3,
then g
(t} 0 for all t.

Theorem 4.3 Let V be a Lyapunov function, i.e., V(y} 0 with V(y} = 0 i y = 0;
also the level sets of V are closed. If
d
dt
V(y(t}} < 0,
then the origin is asymptotically stable.
Proof : Suppose that y(t} does not converge to zero. Since V(y(t}} is monotoni-
cally decreasing it must converge to a limit > 0. That is,
t > 0 V(y(t}} V(y
0
}.
The set
C = y V(y} V(y
0
}}
is closed and bounded and hence compact. Thus,
(DV(y}, f (y}}
must attain its maximum on C,
max
yC
(DV(y}, f (y}} < 0.
Then,
d
dt
V(y(t}} = (DV(y(t}}, f (y(t}} ,
and therefore
V(y(t}} V(y
0
} t,
which leads to a contradiction for large t. B
Exercise 4.2 Consider the nonlinear system
x
= y
y
= ky g(x},
where g is continuous and satises xg(x} > 0 for x 0 and k > 0. (This might
represent the equation of a mass and nonlinear spring subject to friction.) Show
that the origin is stable by using the Lyapunov function
V(x, y} =
1
2
+
x
0
g(s}ds.
Exercise 4.3 Consider the system
x
= x
3
y
2
y
= xy y
3
.
Prove that the origin is stable using the Lyapunov function
V(x, y} = x log(1 x} y log(1 y}.
96 Chapter 4
4.4 Invariant manifolds
Denition 4.3 Let f R
n
R
n
be an autonomous vector eld. This vector eld
induces a ow map ( ),
t
R
n
R
n
via the trajectories of the solutions, namely
y
1
=
t
(y
0
},
if y
1
is the solution at time t of the dierential system,
y
= f (y}, y(0} = y
0
.
Another way to write it is
d
dt
t
(y} = f (
t
(y}}
0
(y} = y.
Denition 4.4 Let f R
n
R
n
be an autonomous vector eld. A set S R
n
is
said to be invariant if
t
(y} S for all y S and for all t R.
That is, a trajectory that starts in S has always been in S and remains in S . It is
said to be positively invariant ( ) if
t
(y} S for all y S and for all t > 0.
It is said to be negatively invariant ( ) if
t
(y} S for all y S and for all t < 0.
Example: Any trajectory,
O(y} =
t
(y} t R}
is an invariant set. Any union of trajectories,
yA
O(y}
is an invariant set, and conversely, any invariant set is a union of trajectories.
Any forward trajectory,
O
+
(y} =
t
(y} t 0}
is a positively-invariant set, and any backward trajectory,
O
+
(y} =
t
(y} t 0}
is a negatively-invariant set.
Example: A xed point is an invariant set (no matter if it is stable or not).
Denition 4.5 An invariant set S is said to be an invariant manifold (
) if looks locally like a smooth surface embedded in R
n
.
1
Consider again an autonomous system,
y
= f (y},
and let u be a xed point. The behavior of the solution near the xed point is
studied by means of the linearized equation,
y
= Ay,
where A = Df (u}. The solution of the linearized system is
y(t} = e
At
y
0
,
i.e., the ow map of the linearized system is given by the linear map
t
= e
At
.
As we know from linear algebra, we may decompose R
n
as follows:
R
n
= E
s
E
u
E
c
,
where
E
s
= Spane
1
, . . . , e
s
}
1
This is of course a non-rigorous denition; for a rigorous denition you need to take a course
in dierential geometry.
98 Chapter 4
is the span of (generalized) eigenvectors corresponding to eigenvalues with nega-
tive real part,
E
u
= Spane
s+1
, . . . , e
s+u
}
is the span of (generalized) eigenvectors corresponding to eigenvalues with posi-
tive real part, and
E
c
= Spane
s+u+1
, . . . , e
s+u+c
}
is the span of (generalized) eigenvectors corresponding to eigenvalues with zero
real part.
These three linear subspaces are examples of invariant manifolds (of the linearized
system!). The characterization of the stable manifold E
s
, is that
lim
t
t
(y} = 0 y E
s
.
The characterization of the unstable manifold E
u
, is that
lim
t
t
(y} = 0 y E
u
.
Example: Let n = 3 and suppose that A has three distinct real eigenvalues,
1
,
2
< 0 and
3
> 0.
Let e
1
, e
2
, e
3
be the corresponding normalized eigenvectors. Then,
A =

e
1
e
2
e
3

1
0 0
0
2
0
0 0
3
e
1

e
2

e
3

1
TT
1
.
As we know,
y(t} = e
At
y
0
= T
1
t
0 0
0 e
2
t
0
0 0 e
3
t
T
1
y
0
.
In this case
E
s
= Spane
1
, e
2
} and E
u
= Spane
3
}
are invariant manifolds. The geometry of the trajectories is depicted in the follow-
ing gure:
E
s
E
u
Example: Suppose now that
1,2
= i and
3
> 0,
where < 0. Then, the corresponding eigenvectors are complex, however
A(e
1
+ e
2
} = ( + i}e
1
+ ( i}e
2
= (e
1
+ e
2
} + i(e
1
e
2
}
A(e
1
e
2
} = ( + i}e
1
( i}e
2
= (e
1
e
2
} + i(e
1
+ e
2
},
or,
A
e
1
+ e
2
2
=
e
1
+ e
2
2

e
1
e
2
2i
A
e
1
e
2
2i
=
e
1
e
2
2i
+
e
1
+ e
2
2
.
Thus, there exists a transformation of variables A = TT
1
, in which
=

0
0
0 0
3
.
The solution to this system is
y(t} = T
e
t
cos t e
t
sint 0
e
t
sint e
t
cos t 0
0 0 e
3
t
T
1
y
0
.
100 Chapter 4
Here again there is a two-dimensional stable subspace and a one-dimensional un-
stable subspace.
SHOW FIGURE.
Example: Consider now the case where
1
< 0 is an eigenvalue of multiplicity
two and
3
> 0. Then, A = TT
1
with
=

1
1 0
0
1
0
0 0
3
,
and the solution is
y(t} = T
1
t
te
1
t
0
0 e
1
t
0
0 0 e
3
t
T
1
y
0
.
In this transformed system E
s
= Spane
1
, e
2
} and E
u
= Spane
3
}. This is a de-
cient system, hence has only one direction entering the xed point.
SHOW FIGURE.
Exercise 4.4 Consider the following dierential systems:

' y
= _
0
0
_y < 0, > 0.
^ y
= _
0
0
_y < 0, < 0.
y
= _

_y < 0, > 0.
y
= _
0 0
0
_y < 0.
I y
= _
0
0 0
_y > 0.
I y
= _
0 0
0 0
_y.
In each case calculate all the trajectories and illustrate them on the plane. Describe
the stable and unstable manifolds of the origin.
Exercise 4.5 Consider the nonlinear systems:
'
1
= y
2
y
2
= y
2
y
1
y
2
1
.
^
1
= y
1
+ y
3
1
y
2
= y
1
+ y
2
.
Find all the xed points and determine their linear stability. Can you infer from
this analysis the nonlinear stability?
The question is what happens in the nonlinear case. How does the structure of the
linearized equation aect the properties of the solutions in the vicinity of the xed
point?
Theorem 4.4 Suppose that the vector eld f is k-times dierentiable and let the
origin be a xed point. Then there exists a neighborhood of the origin in which
there are C
k
-invariant manifold,
W
s
, W
u
and W
c
,
that intersect at the origin are are tangent at the origin to the invariant manifolds
of the linearized invariant manifolds. Moreover, the nonlinear stable and unstable
manifolds retain the asymptotic properties of the linearized manifolds, namely,
y W
s
lim
t
t
(y} = 0,
and
y W
u
lim
t
t
(y} = 0,
Proof : Not in this course. B
Example: Consider again the Dung oscillator,
x
= y
y
= x x
3
y.
102 Chapter 4
The linearized vector eld is
y
= _
0 1
1
_y,
with eigenvalues
1,2
=
1
2
[
2
+ 4 ).
The corresponding (linear) stable and unstable manifolds are
y =
1,2
x.
The origin is a saddle, i.e., it has a one-dimensional stable manifold and a one
dimensional unstable manifold.
!
"
!
$
Example: Consider the following articial system,

x
= x
y
= y + x
2
.
The origin is a xed point and the linearized system is
_
x
y
_
= _
1 0
0 1
__
x
y
_,
i.e., E
s
= Spane
2
} and E
u
= Spane
1
} are the stable and unstable manifolds of
origin in the linearized system.
In this case we can actually compute the stable and unstable manifolds of the
origin in the nonlinear system. We look for trajectories y(x}. These satisfy the
dierential equation,
y
(x} =
y
(t}
x
(t}
=
y
x
+ x.
This is a linear system whose solution is
y(x} =
C
x
+
x
2
3
.
For C = 0, we get a one-dimensional manifold which is invariant (since it is a
trajectory) and tangent to the unstable manifold of the linearized system. The
stable manifold of the nonlinear system is the y-axis, x = 0.
!
"
!
$
4.5 Periodic solutions

Denition 4.6 A solution through a point y is said to be periodic with period
T > 0 if
t
(y} = y.
We will now concentrate on periodic solutions of autonomous systems in the
plane.
104 Chapter 4
Theorem 4.5 (Bendixons criterion) Consider the planar system,
y
= f (y}
where f is at least C
1
. If in a simply connected domain D
div f =
f
1
y
1
+
f
2
y
2
is not identically zero and does not change sign, then there are no closed orbits
lying entirely in D.
Proof : Suppose there was a closed orbit that encloses a domain S D. By
Greens theorem,
S
div f dy =
( f , n} d = 0,
where n is the unit normal to , and ( f , n} = 0 because f is tangent to . If, as
assumed div f is neither identically zero nor changes sign, then this is impossible.
B
Example: For the Dung oscillator,
div f = ,
hence this system cannot have any closed orbit.
4.6 Index theory
In this section we are concerned with autonomous dierential systems in the
plane:
x
= f (x, y}
y
= g(x, y}.
Let be a closed curve in the plane. If the curve does not intersect a xed point,
then at any point along the curve the vector eld makes an angle with the x axis,
(x, y} = tan
1
g(x, y}
f (x, y}
.
When we complete a cycle along the angle must have changed by a multiple
of 2. We dene the index of the curve by
i(} =
1
2
d.
It turns out that the index of a curve satises the following properties:
' The index of a curve does not change when it is deformed continuously, as
long as it does not pass through a xed point.
^ The index of a curve that encloses a single sink, source, or center is 1.
The index of a periodic orbit is 1.
The index of a saddle is 1.
I The index of a curve that does not enclose any xed point is zero.
I The index of a curve is equal to the sum of the indexes of the xed points
that it encloses.
!"#$ !&'()*
!+,,-*
.*("&,")
Corollary 4.1 Every periodic trajectory encloses at least one xed point. If it
encloses a single xed point then it is either a sink, a source, or a center. If all
the enclosed xed points are hyperbolic, then there must be 2n + 1 of those with n
saddles and n + 1 sources/sinks/centers.
106 Chapter 4
Example: According to index theory, possible limit cycles in the Dung oscilla-
tor are as follows:
4.7 Asymptotic limits

We are back to autonomous systems in R
n
; we denote by
t
R
n
R
n
the ow
map. The trajectory through a point y is the set:
O(y} =
t
(y} t R}
We also have the positive and negative trajectories:
O
+
(y} =
t
(y} t 0}
O
(y} =
t
(y} t 0}.
Denition 4.7 Let y R
n
. A point y belongs to the -limit of y,
y (y},
if there exists a sequence t
n
such that
lim
n
tn
(y} = y.
It belongs to the -limit y,
y (y},
if there exists a sequence t
n
such that
lim
n
tn
(y} = y.
Thus the -set ( ) (y} denotes the collection of all future partial
limits of trajectories that start at y, and (y} denotes the collection of all past
partial limits of trajectories that pass through y.
Example: Let u be a xed point and let y belong to the stable manifold of u. Then,
(y} = u}.
If z belongs to the unstable manifold of u, then
(z} = u}.
Example: Let y R R
n
be a periodic solution. Then for every t
0
,
(y(t
0
}} = (y(t
0
}} = O(y(t
0
}}.
4.8 The Poincar e-Bendixon theorem

We consider again two-dimensional systems,
y
= f (y}.
As before we denote the ow function by
t
(x, y}.
Proposition 4.1 Let M be a compact positively invariant set (i.e., trajectories
that start in Mremain in Mat all future times). Let p M. Then,
108 Chapter 4
' (p} .
^ (p} is closed.
(p} is invariant, i.e., it is a union of trajectories.
(p} is connected.
Proof :
' Take any sequence t
n
. Since
tn
(p} M and M is compact this
sequence has a converging subsequence,
lim
k
tn
k
(p} = q M.
By denition q (p}, which is therefore non-empty.
^ Let q ] (p}. By the denition of the -set, there exists a neighborhood U
of q and a time T > 0 such that
t
(p} ] U t > T.
Since U is open, every point z U has a neighborhood V U, and
t
(p} ] V t > T,
which implies that z ] (p}, i.e.,
((p}}
c
is open,
hence (p} is closed.
Let q (p}. By denition, there exists a sequence t
n
, suchthat
lim
n
tn
(p} = q,
Let R be arbitrary. Then,
lim
n
tn+
(p} = lim
n
(
tn
(p}} =
[lim
n
tn
(p}) =
(q},
which proves that
(q} (p} for all R, and hence (p} is invariant.

Suppose by contradiction that (p} was not connected. Then there exist
disjoint open sets V
1
, V
2
such that
(p} V
1
V
2
(p} V
1
and (p} V
2
.
It follows that the trajectory O(p} must pass from V
1
to V
2
and back in-
nitely often. Consider a compact set K that encloses V
1
and not intersect
V
2
. The trajectory must pass innitely often in the compact set KV
1
, hence
must have there an accumulation point, in contradiction to (p} being con-
tained in V
1
V
2
.
!
"
!
$
%
B
Denition 4.8 Let be a continuous and connected arc in R
2
(i.e., a map I
R
2
). It is called transversal to the ow eld f neither vanishes on nor is tangent
to , namely,
(
(t}, f ((t}}
(t} f ((t}.
!
#
110 Chapter 4
Lemma 4.1 Let M R
2
be a positively invariant set and let be a transversal
arc in M. Then for every p M, the positive trajectory O
+
(p} intersects
monotonically, that is, if p
i
= (s
i
} is the i-th intersection of O
+
(p} and , then
the sequence s
i
is monotonic.
Comment: The lemma does not say that O
+
(p} must intersect at all, or that it
has to intersect it more than once.
Proof : Proof by sketch:
!
"
!
"$%
&'(")*+,- ".*/0"/.1 (+1
B
Corollary 4.2 Let M R
2
be a positively invariant set and let be a transversal
arc in M. Then for every p M, (p} intersects at at most one point,
#| (p}| 1.
Proof : The existence of two intersection points would violate the previous lemma,
for suppose that
p
1
, p
2
(p}.
Then, O
+
(p} would intersect at innitely many points in disjoint neighborhoods
of p
1
and p
2
.
Note that this point is somewhat subtle. If p
1
(p} we are guaranteed that
the trajectory will visit innitely often any neighborhood of p
1
, but does it mean
that it has to intersect each time? Yes. Take any neighborhood U of p
1
and
consider
U ,
which is an open neighborhood of p
1
in . By transversality,
t
(U } < t < e}
is an open neighborhood of p
1
in R
2
. Any trajectory visiting this neighborhood
intersects U . B
Proposition 4.2 Let Mbe a positively invariant set and let p M. If (p} does
not contain any xed point, then (p} is a periodic trajectory.
Proof : Let q (p} and let x (q} (we know that both are non-empty sets).
Since O(q} (p} (an -limit set is invariant) and (p} is closed, then
(q} (p}.
It follows that (q} does not contain xed points, i.e., x is not a xed point.
Since x is not a xed point, we can trace through it a transversal arc . The
trajectory O
+
(q} will intersect monotonically at points q
n
x, but since q
n

(p}, then all the q
n
must coincide with x. It follows that O(q} intersects x more
than once, hence it is periodic.
It remains to show that O(q} coincides with (p}. Let now be a transversal
arc at q. We claim that O(q} has a neighborhood in which there are no other
points that belong to (p}, for if every neighborhood of O(q} contained points in
r (p}, O(r} (p}, and there would be such an r for which O(r} intersects .
Since (p} is connected, it follows that (p} = O(q}.
112 Chapter 4
!

#

$ %&'( )' *+),+ -+(.(
/.( '& !&)'- )' !0!1
B
Denition 4.9 A heteroclinic orbit is a trajectory that connects two xed points,
i.e., a trajectory O(y}, such that
(O(y}} = p
1
} and (O(y}} = p
2
},
where p
1
and p
2
are xed points. If p
1
= p
2
(a saddle), then it is called a homo-
clinic orbit.
Proposition 4.3 Let M be a positively invariant set. Let p
1
, p
2
be xed points
and
p
1
, p
2
(p}
for some p M. Since (p} is connected and invariant (a union of trajectories).
There exists a heteroclinic orbit that connects p
1
and p
2
(or vice versa). There
exists at most one trajectory, (p}, such that
p
1
} = (} and p
2
} = (}.
Proof : Proof by sketch:
!
"
!
$
!

B
Proposition 4.4 (Poincar e-Bendixon) Let M be a positively invariant set that
contains a nite number of xed points. Let p M. Then, one and only one of the
following occurs:
' (p} is a single xed point.
^ (p} is a limit cycle.
(p} is a union of xed points p
i
and heteroclinic orbits
j
that connect
them.
Proof : If (p} contains only xed points then it only contains one due to con-
nectedness.
If (p} does not contain any xed point then it is a limit cycle (we proved it).
Remains the case where (p} contains xed points and points that are not xed
points; let q (p} and consider the set (q}. We will show that it contains a
single xed point, hence q is on a trajectory that converges asymptotically to a
xed point in (p}. Similarly, (q} contains a single xed point, so that q lies, as
claimed, on a trajectory that connects between two xed points in (p}.
114 Chapter 4
So let x (q}. Suppose that x was not a xed point, then we would trace a
transversal arc through r. The trajectory O(q} must intersect this arc innitely
often, but since O(q} (p}, it must intersect at a single point, which implies
that O(q} is a periodic orbit, but this is impossible because (p} is connected. It
follows that (q} contains only xed points, but then we know that it can only
contain a single xed point.
B

Oridinary Differential Equations by Raz Kupferman

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Oridinary Differential Equations by Raz Kupferman

Uploaded by

Copyright:

Available Formats

Ordinary Dierential Equations

A number of questions arise right away:

First-order ode The most general rst-order ode is of the form

(t} = f (t, y(t}}.

(t} + y(t} sint = 0.

More generally, in a rst-order system of n (explicit) odes the unknown is a vector-

(t} = f (t, y(t}},

(t} = F(t, Y(t}}.

(velocities) at an initial time (the big bang?). By Newtons laws,

(t} = f (t, y(t}}.

(s} = f (s, y(s}} y(t

(t} = f (t, y(t}}, y(t

(t} = 6y(t} y(0} = 1

has one-sided limits.

(t} f (t, (t}} .

(t} = f (t, y(t}} on the interval |t

(t} f (t, (t}} = f (t

(t} = f (t, y(t}} has a solution on the

(t} = f (t, y(t}}. B

(t} = f (t, y(t}}, y(t

(t} = f (t, y(t}} contained in D on the interval I, and moreover

(t} = f (t, y(t}}, y(t

(t} = f (t, y(t}}, y(t

(t} = f (t, y(t}}, y(a} = b.

(t} = f (t, y(t}}, y(} =

(t} = A(t}y(t} (2.1)

Proposition 2.1 For every I and R

(t} = A(t}y(t}, y(} = .

(t} + MY(t}, Y(} = 0.

(t} = A(t}y(t} and z

(t} = A(t}y(t} + A(t}z(t} = A(t}(y + z}(t},

(t} = A(t}X(t}, (2.2)

Theorem 2.1 Let be a fundamental matrix for (2.1) and let C C

= A and we know that det (t} 0 for all t. Then for

(t}(t} determines the magnitudes of the columns of (t} and the

2.2.3 The solution operator

(t} = A(t}y(t}, y(t

(t} = A(t}y(t} + b(t},

(t, s}b(s} ds + R(t, t}b(t}

(t} = A(t}y(t} + b(t} = A(t}(t}C(t} + b(t},

For later use, we note that if we multiply J

2.4.3 Linear system with constant coecients

Let us write down this equation in components:

(t} = A(t}Y(t} + B(t},

Example: Consider the equation

(t} = A(t}y(t}, (2.3)

(t} = f (z(t} + (t}}.

(t} = f (z(t}} + f (z(t}}(t} + O(

(t} = f (z(t}} we remain with

(t} = f (z(t}}(t} + O(}.

(t} = (1 + sint} y(t},

(t + T} = A(t + T}(t + T} = A(t}(t},

(x} = ku(x} and v

(x} = ku(x} u(0} = u(L} = 0.

kx}, but only if

(x} = y(x}, y(0} = y(1} = 0.

Example: Another class of examples for n = 2:

(b} = y(a} y(b}

(x} dx = (u, Lv}.

Recall the adjoint operator L

derived in the previous chapter:

= L and Uu = Uv = 0 implies that |uv|(b} = |uv|(a}, then the problem

(x} = y(x} + f (x}, y(0} = y(1} = 0.