Professional Documents
Culture Documents
Oridinary Differential Equations by Raz Kupferman
Oridinary Differential Equations by Raz Kupferman
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}.
= 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.)
(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}, 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} = 6tz
z
(t} = z
3
y
+ 4.
(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
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).
Existence and uniqueness 7
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
(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.
Existence and uniqueness 9
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
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
j
, (t
j
}},
and further,
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} = 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.
Existence and uniqueness 17
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
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
}.
Existence and uniqueness 19
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
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
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)
,
Existence and uniqueness 21
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 < ,
Existence and uniqueness 23
then consider the initial value problem
y
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
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}.
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}| 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}z(t}.
For every , C,
(y + z}
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
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
_.
= 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
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
.
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
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
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} = 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
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
_.
(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
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
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.
+ 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} = 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} +
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
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_,
which concludes the proof. B
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},
which concludes the proof. B
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
(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
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
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
.
Boundary value problems 67
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}.
(a} + N
12
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
= (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
(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.
Boundary value problems 71
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
,
Boundary value problems 73
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
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,
Boundary value problems 75
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:
Boundary value problems 77
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.
Boundary value problems 79
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
.
Boundary value problems 81
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
,
which concludes the proof. B
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.
Boundary value problems 83
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.
Boundary value problems 85
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 (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} + z
(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
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}.
Stability of solutions 91
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
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.
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
(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
= 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}
Stability of solutions 97
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:
Stability of solutions 99
E
s
E
u
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.
= _
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.
Stability of solutions 101
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.
!
"
!
$
= 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.
Stability of solutions 103
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.
!
"
!
$
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}
.
Stability of solutions 105
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:
(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.
Stability of solutions 107
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
}}.
= 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
(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}.
Stability of solutions 111
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:
Stability of solutions 113
!
"
!
$
!
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