This action might not be possible to undo. Are you sure you want to continue?
Raz Kupferman
Institute of Mathematics
The Hebrew University
June 26, 2012
2
Contents
1 Existence and uniqueness 3
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Euler’s approximation scheme . . . . . . . . . . . . . . . . . . . . . 9
1.3 The CauchyPeano existence proof . . . . . . . . . . . . . . . . . . . 11
1.4 Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5 The PicardLindl ¨ 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 Nonhomogeneous linear systems . . . . . . . . . . . . . . . . . . . 37
2.4 Systems with constant coeﬃcients . . . . . . . . . . . . . . . . . . . 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 coeﬃcients . . . . . . . . . . . 43
ii CONTENTS
2.5 Linear diﬀerential equations of order n . . . . . . . . . . . . . . . . 45
2.5.1 The Wronskian . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.5.2 The adjoint equation . . . . . . . . . . . . . . . . . . . . . . 48
2.5.3 Nonhomogeneous equation . . . . . . . . . . . . . . . . . . 50
2.5.4 Constant coeﬃcients . . . . . . . . . . . . . . . . . . . . . . 52
2.6 Linear systems with periodic coeﬃcients: Floquet theory . . . . . . 55
2.6.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.6.2 General theory . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.6.3 Hill’s equation . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3 Boundary value problems 63
3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2 Selfadjoint eigenvalue problems on a ﬁnite interval . . . . . . . . . 67
3.2.1 Deﬁnitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2.2 Properties of the eigenvalues . . . . . . . . . . . . . . . . . 69
3.2.3 Nonhomogeneous 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 Deﬁnitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 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´ eBendixon theorem . . . . . . . . . . . . . . . . . . . . 107
CONTENTS 1
Bibliography
1. Theory of ordinary diﬀerential equations by E.A. Coddington and N. Levin
son.
2. Ordinary diﬀerential equations by V.I. Arnold.
3. Solving ordinary diﬀerential equations by E. Hairer, N.P Nørsett and G.
Wanner.
2 CONTENTS
Chapter 1
Existence and uniqueness
1.1 Introduction
Deﬁnition 1.1 A diﬀerential equation is an equation that relates a function to
its derivative(s). The unknown is the function. A diﬀerential equation is said to be
ordinary (הליגר) if the function is univariate, and more precisely if its domain is
a connected subset of R. We abbreviate ordinary diﬀerential equation into ode.
Example: Find a function y ∶ R →R that satisﬁes the equation
y
′
= 6y,
or in a diﬀerent 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.
Deﬁnition 1.2 An ordinary diﬀerential equation is said to be of order k if the
highest derivative of the unknown function that appears in the equation is the kth
derivative.
Example: The following equation is a ﬁrstorder 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.)
▲▲▲
Firstorder ode The most general ﬁrstorder ode is of the form
Φ(y
′
(t}, y(t}, t} = 0.
Such an equation is said to be implicit (המותס הגצה). We will always assume that
the equation has been brought into explicit form (תשרופמ הגצה):
y
′
(t} = f (t, y(t}}.
Existence and uniqueness 5
Secondorder ode The most general secondorder ode is of the form
Φ(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 ﬁrst example, we ask ourselves whether this equation has a solution and
whether it is unique. At this point it is clear that there won’t 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 diﬀerential equations In the same way as we go from univariate al
gebraic equations into multivariate ones, so we can extend the idea of a diﬀerential
equation into a system of diﬀerential 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 ﬁrstorder system of n (explicit) odes the unknown is a vector
valued function, y ∶ R →R
n
, that satisﬁes 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
Highorder equations and ﬁrstorder systems Every nth order (scalar) equa
tion can be turned into a ﬁrstorder system of n odes by a simple procedure. Sup
pose we have an equation of the form
y
(n)
(t} = f (t, y
n−1
(t}, . . . , y
′
(t}, y(t}}.
We then deﬁne the following vectorvalued function,
Y =
⎛
⎜
⎜
⎜
⎝
y
y
′
⋮
y
(n−1)
⎞
⎟
⎟
⎟
⎠
.
It satisﬁes the following system of equations,
Y
′
1
= Y
2
Y
′
2
= Y
3
⋯ Y
′
n−1
= 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 diﬀerential equation is called ordinary when the
unknown function depends on a single realvalued variable. This is in contrast to
partial diﬀerential equations (תויקלח תואוושמ), 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 diﬀusion
equation): the unknown function depends on two variables (position and time),
y ∶ R× R →R, and satisﬁes the equation,
∂y
∂t
(x, t} =
∂
2
y
∂x
2
(x, t}.
▲▲▲
The theory of pdes is inﬁnitely 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
diﬀerence between the two).
Existence and uniqueness 7
Why do we care about odes? Diﬀerential 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
godgiven 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 Newton’s 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. Newton’s
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 ﬁrst documented instance of an ode is an equation studied
by Newton (1671), while developing diﬀerential calculus:
y
′
(t} = 1 − 3t + y(t} + t
2
+ ty(t} y(0} = 0.
Newton’s 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 (תוקזח רוט),
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 ﬁxed t, as more terms as added, the closer we are to the true solution, however,
for every ﬁxed number of terms the approximation deteriorates for large t.
A problem studied by Leibniz In 1674 Leibniz (Newton’s competitor on the
development of calculus) studied a geometrical problem that had been studied
earlier by Fermat. He tried to solve the following diﬀerential equation,
y
′
(t} = −
y(t}
_
a
2
− y
2
(t}
.
Remember that at that time, derivatives were not deﬁned 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 ﬁnite 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
− ξ
2
ξ
dξ = −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 nth order equation can
be represented as a ﬁrstorder system, thus, until further notice we will assume
that we deal with ﬁrst 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 ﬁrstorder
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 diﬀerent. The ﬁrst part of this course will be devoted to initial value
problems.
Integral formulation Consider a ﬁrst 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 diﬀerential equation. We will later see in which sense the diﬀerential and
integral systems are equivalent, and in what sense they diﬀer.
1.2 Euler’s 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 vectorvalued, and suppose that we want to ﬁnd 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 ﬁrstorder diﬀerential equation is approximated by a onestep diﬀer
ence equation (ישרפה תאוושמ).
Euler’s belief was that if we reﬁne the partition suﬃciently, then we will get ar
bitrarily close to the true solution (whose existence was not questioned). Euler’s
method (usually variants of it) is used until today to numerically approximate the
solution of ordinary diﬀerential systems.
In 1820 Cauchy proved that Euler’s approximation does indeed converge, as the
partition is reﬁned, to a solution of the diﬀerential equation, and in fact Cauchy’s
proof is the ﬁrst proof that the diﬀerential 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 Euler’s 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.
Existence and uniqueness 11
1.3 The CauchyPeano existence proof
In order to be able to analyze the diﬀerential 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, (בעשת))
Deﬁnition 1.3 Let D ⊂ R
n
×R be a domain in which f is continuous. Let I be an
open connected interval. A function ϕ ∶ I → R
n
is called an εapproximation to
the diﬀerential system in D if:
' ϕ is continuous.
^ ϕ is continuously diﬀerentiable, except perhaps at a ﬁnite set of points S ,
where ϕ
′
has onesided limits.
· For every t ∈ I, (t, ϕ(t}} ∈ D.
´ For every t ∈ I ∖ S ,
¡ϕ
′
(t} − f (t, ϕ(t}}¡ ≤ ε.
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; deﬁne
M = max
(t,y)∈R
¡ f (t, y}¡.
Then, we deﬁne
α = min(a, b]M},
and replace R by another rectangle,
R
0
= t
0
, t
0
+ α × ¦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
t
0#
y
0#
y
0
+b
#
y
0
'b
#
t
0
+a
#
t
0
+α
#
Theorem 1.1 There exists for every ε > 0 an εapproximation in R
0
to the equation
y
′
(t} = f (t, y(t}} on the interval t
0
, t
0
+ α 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 diﬀerence between “past” and “future”.
Proof : Since f is continuous on R
0
, it is uniformly continuous (הווש הדימב הפיצר).
Thus, given ε > 0 there exists a δ > 0 such that if
t − s < δ and ¡y − z¡ < δ then ¡ f (t, y} − f (s, z}¡ < ε.
Take the interval t
0
, t
0
+ α and partition it,
t
0
< t
1
< . . . t
n
= t
0
+ α,
such that
max
j
(t
j+1
− t
j
} < min(δ, δ]M}.
We then construct the Euler approximation to this solution, and connect the point
by straight segments, so that Euler’s approximation is a polygon, which we denote
by ϕ(t}. It is a piecewise continuously diﬀerentiable 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 α, ϕ(t} can diﬀer from y
0
(in norm) by at most αM ≤ b. This means that all Euler’s polygons remain in R
0
on the entire interval t
0
, t
0
+α, no mater how the partition of the segment is done.
Existence and uniqueness 13
Let t be a point in the jth interval. Then,
ϕ
′
(t} = f (t
j
, ϕ(t
j
}},
and further,
¡ϕ
′
(t} − f (t, ϕ(t}}¡ = ¡ f (t
j
, ϕ(t
j
}} − f (t, ϕ(t}}¡.
By our choice of the partition,
t − t
j
 < δ.
Because the derivative of ϕ is bounded (in norm) by M,
¡ϕ(t} − ϕ(t
j
}¡ ≤ Mt − t
j
 ≤ M(δ]M} = δ.
By our choice of δ,
¡ f (t
j
, ϕ(t
j
}} − f (t, ϕ(t}}¡ < ε,
which proves that ϕ is an εapproximation to our system. B
Deﬁnition 1.4 A family F of functions I → R
n
is said to be equicontinuous
(הדיחא הדימב יפיצר) if there exists for every ε > 0 a δ > 0, such that for every
t
1
− t
2
 < δ and every g ∈ F,
¡g(t
1
} − g(t
2
}¡ < ε.
(Equicontinuity is uniform continuity with the same δ valid for all g ∈ F.)
Theorem 1.2 (ArzelaAscoli) Let F be an inﬁnite family of functions I →R
n
that
are uniformly bounded, namely,
sup
f ∈F
max
t∈I
¡ f (t}¡ < ∞,
and equicontinuous. Then there exists a sequence f
n
∈ F that converges uniformly
on I.
(3 hrs, (בעשת))
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}(∀ε > 0}(∃N ∈ N} ∶ (∀m, n > N} (¡F
n
(r} − F
m
(r}¡} < ε.
Note that since we don’t know what the limit is, we use instead the Cauchy crite
rion for convergence.
Fix ε > 0. Since the family of functions F is equicontinuous,
(∃δ > 0} ∶ (∀t
1
, t
2
, t
1
− t
2
 < δ}(∀f ∈ F} (¡ f (t
1
} − f (t
2
}¡ < ε}.
Partition the interval I into subintervals,
I = I
1
∪ I
2
∪ ⋅ ⋅ ⋅ ∪ I
K
,
such that each subinterval if shorter than δ. In every subinterval I
k
select a rational
number r
k
. Since there is a ﬁnite number of such r
k
s,
(∃N ∈ N} ∶ (∀m, n > N}(∀k = 1, . . . , K} (¡F
n
(r
k
} − F
m
(r
k
}¡} < ε
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
≤ 3ε.
Thus, the sequence F
n
satisﬁes Cauchy’s criterion for uniformconvergence, which
concludes the proof. B
With this, we can prove Cauchy’s theorem:
Existence and uniqueness 15
Theorem 1.3 (CauchyPeano) Let f (t, y} and the rectangle R
0
be deﬁned as
above. Then the diﬀerential equation y
′
(t} = f (t, y(t}} has a solution on the
interval t
0
, t
0
+ α 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 (ε
k
} be a sequence of positive numbers that tends to zero. We proved
that for every k there exists a function ϕ
k
∶ I → R
n
that is an ε
k
approximation on
I = t
0
, t
0
+ α to the diﬀerential equation, and satisﬁes the initial conditions.
The sequence (ϕ
k
} is uniformly bounded on I because
(∀t ∈ I}(∀k ∈ N} (¡ϕ
k
(t} − y
0
¡ < b}.
It is also equicontinuous because
(∀t
1
, t
2
∈ I}(∀k ∈ N} (¡ϕ
k
(t
2
} − ϕ
k
(t
1
}¡ ≤ Mt
2
− t
1
},
and therefore, given ε > 0 we set δ = ε]M, and
(∀t
1
− t
2
 < δ}(∀k ∈ N} (¡ϕ
k
(t
2
} − ϕ
k
(t
1
}¡ ≤ ε}.
It follows from the ArzelaAscoli theorem that (ϕ
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 diﬀerential equation.
16 Chapter 1
For each k ∈ N it follows from the fundamental theorem of calculus that
2
ϕ
k
(t} = y
0
+
t
t
0
ϕ
′
k
(s} ds
= y
0
+
t
t
0
f (s, ϕ
k
(s}} ds +
t
t
0
ϕ
′
k
(s} − f (s, ϕ
k
(s}} ds.
The fact that ϕ
′
k
is not deﬁned at a ﬁnite number of points is immaterial.
Using the fact that ϕ
k
is an ε
k
approximation to the diﬀerential equation,
_
t
t
0
ϕ
′
k
(s} − f (s, ϕ
k
(s}} ds_ ≤ ε
k
t − t
0
.
Letting k → ∞ and using the fact that the convergence of ϕ
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 diﬀerentiate 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 ϕ
k
are absolutely continuous (טלחהב תופיצר), which means that there exists
for every ε > 0 a δ > 0 such that for every ﬁnite sequence of disjoint intervals (x
k
, y
k
) that satisﬁes
∑
k
(y
k
− x
k
) < δ,
k
f (y
k
) − f (x
k
) < ε.
A function is absolutely continuous if and only if it is almost everywhere diﬀerentiable, and its
derivative satisﬁes the fundamental theorem of calculus.
Existence and uniqueness 17
By the CauchyPeano theorem there exists an α > 0 such that this equation has
a solution in the interval 0, α. The CauchyPeano 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 inﬁnitely many solutions of the form
y(t} =
⎧
⎪
⎪
⎨
⎪
⎪
⎩
0 0 ≤ t ≤ c
[
2(t−c)
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 Euler’s 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 ϕ
1
and ϕ
2
be ε
1
 and ε
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,
¡ϕ
1
(t
0
} − ϕ
2
(t
0
}¡ < δ.
18 Chapter 1
Then, for all t ∈ I,
¡ϕ
1
(t} − ϕ
2
(t}¡ < δe
Lt−t
0
+
ε
1
+ ε
2
L
{e
Lt−t
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}¡ ≤ L¡y − 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 deﬁnition of ε
approximations, for every t
0
≤ s ≤ t,
¡ϕ
′
1
(s} − f (s, ϕ
1
(s}}¡ ≤ ε
1
¡ϕ
′
2
(s} − f (s, ϕ
2
(s}}¡ ≤ ε
2
.
For i = 1, 2,
ϕ
i
(t} = ϕ
i
(t
0
} +
t
t
0
f (s, ϕ
i
(s}} ds +
t
t
0
ϕ
′
i
(s} − f (s, ϕ
i
(s}} ds,
hence
_ϕ
i
(t} − ϕ
i
(t
0
} −
t
t
0
f (s, ϕ
i
(s}} ds_ ≤ ε
i
(t − t
0
}.
Using the triangle inequality, ¡a − b¡ ≤ ¡a¡ + ¡b¡,
_ϕ
1
(t} − ϕ
2
(t} − ϕ
1
(t
0
} − ϕ
2
(t
0
} −
t
t
0
 f (s, ϕ
1
(s}} − f (s, ϕ
2
(s}} ds_ ≤ (ε
1
+ε
2
}(t−t
0
}.
It further follows that
¡ϕ
1
(t} − ϕ
2
(t}¡ ≤ ¡ϕ
1
(t
0
} − ϕ
2
(t
0
}¡
+
t
t
0
¡ f (s, ϕ
1
(s}} − f (s, ϕ
2
(s}}¡ ds + (ε
1
+ ε
2
}(t − t
0
}.
Existence and uniqueness 19
Deﬁne now r(t} = ¡ϕ
1
(t} − ϕ
2
(t}¡. Then, using the Lipschitz continuity of f and
the bound on the initial deviation,
r(t} ≤ δ + L
t
t
0
r(s} ds + (ε
1
+ ε
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: deﬁne
R(t} =
t
t
0
r(s} ds,
so that
R
′
(t} ≤ δ + LR(t} + (ε
1
+ ε
2
}(t − t
0
}, R(t
0
} = 0,
which is a diﬀerential inequality.
Next, for every t
0
≤ s ≤ t:
{e
L(s−t
0
)
R(s})
′
= e
L(s−t
0
)
R
′
(s} − LR(s} ≤ e
L(s−t
0
)
(δ + (ε
1
+ ε
2
}(s − t
0
}} ,
Since integration preserves order, we may integrate this inequality from t
0
to t and
get
e
L(t−t
0
)
R(t} − R(0}

0
≤
t
t
0
e
L(s−t
0
)
(δ + (ε
1
+ ε
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 ϕ
1
and ϕ
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 ϕ
1
= ϕ
2
.
Proof : This is an immediate corollary of the previous proposition as δ = ε
1
= ε
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 Cauchy’s approx
imation. It y is the solution and ϕ is an εapproximation with exact initial data,
then
¡ϕ(t} − y(t}¡ <
ε
L
{e
Lt−t
0
− 1) .
(5 hrs, (בעשת))
1.5 The PicardLindl ¨ of existence proof
There is another standard way to prove the existence of solutions, and it is of suf
ﬁcient 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.
Deﬁne the function y
(0)
(t} ≡ y
0
, and further deﬁne
y
(1)
(t} = y
0
+
t
t
0
f (s, y
(0)
(s}} ds.
For what values of t is it deﬁned? For the same t
0
, t
0
+ α as in the previous
sections. We can show that (t, y
(1)
(t}} remains in R
0
as follows: let
τ = sup¦t ∈ I ∶ ¡y
(1)
(t} − y
0
¡ ≤ b}.
If τ < t
0
+ α then
¡y
(1)
(τ} − y
0
¡ ≤ M(t − t
0
} < b,
contradicting the fact that τ is the supremum.
We then deﬁne 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 diﬀerence with the proof based on Euler’s method
is that we need to assume the Lipschitz continuity of f in R
0
. Deﬁne,
∆
(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
(n−1)
(s}}¡ ds,
and
¡∆
(n+1)
(t}¡ ≤ L
t
t
0
¡∆
(n)
(s}¡ ds,
with
¡∆
(1)
(t}¡ ≤
t
t
0
¡ f (s, y
0
}¡ ds ≤ Mt − t
0
.
Once again, we have an integral inequality, but we may proceed diﬀerently:
¡∆
(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
(Lt − t
0
}
n
n!
.
Now we write the nth function as follows:
y
(n)
= {y
(n)
− y
(n−1)
) + {y
(n−1)
− y
(n−2)
) + ⋅ ⋅ ⋅ + y
0
= y
0
+
n−1
k=0
∆
(k)
.
By the Weierstaß Mtest 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 Picard’s 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)
=
n−1
k=m
∆
(k)
,
22 Chapter 1
hence
¡y
(n)
(t} − y
(m)
(t}¡ ≤
n−1
k=m
¡∆
(k)
(t}¡ ≤
n−1
k=m
L
k−1
M
k!
(t − t
0
}
k
≤
∞
k=m
L
k−1
M
k!
(t − t
0
}
k
=
∞
k=0
L
m+k−1
M
(m+ k}!
(t − t
0
}
m+k
≤
∞
k=0
L
m+k−1
M
m!k!
(t − t
0
}
m+k
=
L
m−1
M(t − t
0
}
m
m!
∞
k=0
L
k
k!
(t − t
0
}
k
=
ML(t − t
0
}
m
Lm!
e
L(t−t
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 α such that a unique
solution exists in the interval t
0
, t
0
+ α. 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 can’t 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 deﬁne
a = sup¦τ > t
0
∶ there exists a solution on t
0
, τ}.
If a is ﬁnite and
lim
t→a
y(t} = b < ∞,
Existence and uniqueness 23
then consider the initial value problem
y
′
(t} = f (t, y(t}}, y(a} = b.
By the existence and uniqueness theorems there exists a γ such that a unique
solution exists on a, a + γ, i.e., a unique solution exists on t
0
, a + γ, violating
the deﬁnition of a as a supremum. Thus, if a is ﬁnite then
lim
t→a
y(t} = ∞,
namely, the only way a solution can fail to exist indeﬁnitely 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, (בעשת))
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(τ} = ξ
as a function of t,τ, and ξ; for now we denote this solution as ϕ(t, τ, ξ}. This is
a diﬀerent 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 ϕ is diﬀerentiable in t, as it satisﬁes the equations,
∂ϕ
∂t
(t, τ, ξ} = f (t, ϕ(t, τ, ξ}}, ϕ(τ, τ, ξ} = ξ. (1.2)
The question is how regular it is with respect to the initial data τ and ξ. 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 wellposedness 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 ψ ∶ I →R
n
be a solution to the diﬀerential equation, such that
¦(t, ψ(t}} ∶ t ∈ I} ⊂ D.
Theorem 1.5 There exists a δ > 0 such that for every
(τ, ξ} ∈ U
δ
= ¦(τ, ξ} ∶ τ ∈ I, ¡ξ − ψ(τ}¡ < δ}
There exists a solution ϕ to (1.2) on I with ϕ(τ, τ, ξ} = ξ. Moreover, ϕ is continuous
on I × U
δ
.
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
x≠0
¡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¡ ≤ ¡A¡¡x¡.
2. For every A, B ∈ L(X, X},
¡AB¡ ≤ ¡A¡¡B¡.
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
n×n
.
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
n×n
be a continuous nbyn matrix
valued function of I. We denote its entries, which are realvalued functions, by
a
i j
. The ﬁrstorder system of diﬀerential 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 τ ∈ I and ξ ∈ R
n
, there exists a unique function y ∶ I →
R
n
such that
y
′
(t} = A(t}y(t}, y(τ} = ξ.
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 τ. 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 ﬁrst write the corresponding integral equation,
y(t} = ξ +
t
τ
A(s}y(s} ds,
which holds as long as a solution exists.
Linear equations 27
Taking norms,
¡y(t}¡ ≤ ¡ξ¡ +
t
τ
¡A(s}y(s}¡ ds ≤ ¡ξ¡ +
t
τ
¡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}¡ ≤ ¡ξ¡ + M
t
τ
¡y(s}¡ ds.
This is an integral inequality that we can solve. Deﬁning Y(t} =
∫
t
τ
¡y(s}¡ ds we
get
Y
′
(t} ≤ ¡ξ¡ + MY(t}, Y(τ} = 0.
Then,
e
−M(t−τ)
Y
′
(t} − MY(t} ≤ ¡ξ¡e
−M(t−τ)
.
Integrating from τ to t,
e
−M(t−τ)
Y(t} ≤
¡ξ¡
−M
{e
−M(t−τ)
− 1) ,
and further
Y(t} ≤
¡ξ¡
M
{e
M(t−τ)
− 1) .
Substituting back into the inequality for ¡y(t}¡,
¡y(t}¡ ≤ ¡ξ¡ + M
¡ξ¡
M
{e
M(t−τ)
− 1) = ¡ξ¡e
M(t−τ)
.
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 ﬁeld 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 α, β ∈ C,
(αy + βz}
′
(t} = αA(t}y(t} + βA(t}z(t} = A(t}(αy + βz}(t},
which proves that αy + βz ∈ 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 deﬁned 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 deﬁned as above, then
dimX = n,
that is there exists a set of functions ϕ
1
, . . . , ϕ
n
∈ X, such that
n
k=1
c
k
ϕ
k
= 0
implies that c
k
= 0 for all k. Moreover, for every set of n+1 function ϕ
k
∈ X there
exist scalars c
k
, not all of them zero, such that
n+1
k=1
c
k
ϕ
k
= 0.
Proof : Take τ ∈ I, and select n independent vectors ξ
1
, . . . , ξ
n
∈ R
n
; we denote the
kth component of ξ
j
by ξ
k j
. For each ξ
j
, the ivp
ϕ
′
j
(t} = A(t}ϕ
j
(t}, ϕ
j
(τ} = ξ
j
Linear equations 29
has a unique solution. We claim that these solution are independent. Indeed,
suppose that
n
k=1
c
k
ϕ
k
≡ 0,
then in particular for t = τ,
n
k=1
c
k
ξ
k
= 0,
but because the ξ
k
are independent, the c
k
are all zero.
Let ϕ
1
, . . . , ϕ
n
be linear independent solutions. Take any τ ∈ I and let
ξ
j
= ϕ(τ}.
The ξ
n
are linearly independent, for if they were linearly dependent, we would
have
ξ
n
=
n−1
k=1
c
k
ξ
k
,
but then
ψ =
n−1
k=1
c
k
ϕ
k
is a solution to the diﬀerential system satisfying ψ(τ} = ξ
n
, and by uniqueness
ψ = ϕ
n
, which contradicts the linear independence of the ϕ
k
.
Suppose now that ϕ is another solution of the diﬀerential system. Set
ξ = ϕ(τ}.
Because the ξ
j
span R
n
, there exist coeﬃcients c
k
such that
ξ =
n
k=1
c
k
ξ
k
.
By the same argument as above,
ϕ =
n
k=1
c
k
ϕ
k
,
which proves that the ϕ
k
form a basis. B
30 Chapter 2
2.2.2 Fundamental matrices
Let now ϕ
1
, . . . , ϕ
n
be a basis. Let Φ ∶ I → R
n×n
be the matrixvalued function
whose columns are the vectors ϕ
j
,
Φ(t} =
⎛
⎜
⎝
ϕ
11
⋯ ϕ
1n
⋮ ⋱ ⋮
ϕ
n1
⋯ ϕ
nn
⎞
⎟
⎠
(t},
that is, ϕ
i j
is the ith component of the jth basis function. This matrix satisﬁes
the matrixvalued diﬀerential equation,
Φ
′
(t} = A(t}Φ(t},
which in components reads
ϕ
′
i j
(t} = a
ik
(t}ϕ
k j
(t}.
Φ(t} is called a fundamental matrix (תידוסי הצירטמ) for the linear homogeneous
system (2.1); we call a fundamental matrix for (2.1) any solution X ∶ I → R
n×n
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
ϕ
i j
,
where the c
j
’s are constant, and in vector form,
y = Φc.
Proposition 2.4 For a matrixvalued 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 τ ∈ I. In particular, if this holds then its columns are independent
for all t ∈ I.
Linear equations 31
Proof : Let Φ ∶ I → R
n×n
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 = Φc,
Fix a τ ∈ I. The vector c is determined by the equation
y(τ} = Φ(τ}c.
This is a linear equation for c; we know that it has a unique solution (because
the columns of Φ span the space of solutions), hence from basic linear algebra,
det Φ(τ} ≠ 0. Since the same considerations hold for every t ∈ I, det Φ(t} ≠ 0 for
all t ∈ I. B
(8 hrs, (בעשת))
Here is another proof that the columns of the matrix remain independent, but it
comes with a quantitative estimate:
Proposition 2.5 Let Φ ∶ I →R
n×n
be a fundamental matrix. Then
det Φ(t} = det Φ(t
0
} exp_
t
t
0
Tr A(s} ds_.
In particular, if det Φ(t
0
} ≠ 0 then det Φ(t} ≠ 0 for all t ∈ I.
Proof : Since
Φ =
·
·
·
·
·
·
·
·
·
·
·
·
·
·
ϕ
11
⋯ ϕ
1n
⋮ ⋱ ⋮
ϕ
n1
⋯ ϕ
nn
·
·
·
·
·
·
·
·
·
·
·
·
·
·
,
we diﬀerentiate to obtain
Φ
′
=
·
·
·
·
·
·
·
·
·
·
·
·
·
·
ϕ
′
11
⋯ ϕ
′
1n
⋮ ⋱ ⋮
ϕ
n1
⋯ ϕ
nn
·
·
·
·
·
·
·
·
·
·
·
·
·
·
+ ⋯+
·
·
·
·
·
·
·
·
·
·
·
·
·
·
ϕ
11
⋯ ϕ
1n
⋮ ⋱ ⋮
ϕ
′
n1
⋯ ϕ
′
nn
·
·
·
·
·
·
·
·
·
·
·
·
·
·
=
·
·
·
·
·
·
·
·
·
·
·
·
·
·
∑
j
a
1 j
ϕ
j1
⋯ ∑
j
a
1j
ϕ
jn
⋮ ⋱ ⋮
ϕ
n1
⋯ ϕ
nn
·
·
·
·
·
·
·
·
·
·
·
·
·
·
+ ⋯+
·
·
·
·
·
·
·
·
·
·
·
·
·
·
ϕ
11
⋯ ϕ
1n
⋮ ⋱ ⋮
∑
j
a
n j
ϕ
j1
⋯ ∑
j
a
n j
ϕ
′
jn
·
·
·
·
·
·
·
·
·
·
·
·
·
·
.
32 Chapter 2
Consider the ﬁrst determinant. We subtract from the ﬁrst line a
12
times the second
line, a
13
times the third line, etc. Then we remain with a
11
det Φ. We do the same
with the second determinant and remains with a
22
det Φ, and so on. Thus,
Φ
′
= (Tr A} Φ.
This is a scalar linear ode, and it remains to integrate it.
B
Example: Consider the matrix
Φ(t} = _
t t
2
0 0
_.
This matrix is singular, but its columns are linearly independent. What does it
imply? That this matrixvalued 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 Φ be a fundamental matrix for (2.1) and let C ∈ C
n×n
be a non
singular complex (constant) matrix. Then ΦC is also a fundamental matrix. More
over, if Ψ is a fundamental matrix then Ψ = ΦC for some nonsingular complex
matrix C.
Proof : By deﬁnition Φ
′
= AΦ and we know that det Φ(t} ≠ 0 for all t. Then for
every nonsingular complex matrix C,
(ΦC}
′
= AΦC
and det Φ(t}C = det Φ(t} det C ≠ 0 for all t ∈ I.
Let Ψ be a fundamental matrix. Because it is nonsingular for all t, Ψ
−1
exists, and
0 = (ΨΨ
−1
}
′
= Ψ(Ψ
−1
}
′
+ Ψ
′
Ψ
−1
,
Linear equations 33
from which we get that
(Ψ
−1
}
′
= −Ψ
−1
Ψ
′
Ψ
−1
= −Ψ
−1
AΨΨ
−1
= −Ψ
−1
A.
Now,
(Ψ
−1
Φ}
′
= (Ψ
−1
}
′
Φ+ Ψ
−1
Φ
′
= −Ψ
−1
AΦ+ Ψ
−1
AΦ = 0.
It follows that Ψ
−1
Φ = C, or Φ = ΨC. B
Comment: The matrix (Ψ
∗
}
−1
satisﬁes the equation,
(Ψ
∗
}
−1

′
= −A
∗
(Ψ
∗
}
−1
.
This equation is known as the equation adjoint (הדומצ האוושמ) to (2.1).
Proposition 2.6 If Φ is a fundamental matrix for (2.1) then Ψ is a fundamental
matrix for the adjoint system if and only if
Ψ
∗
Φ = C,
where C is a nonsingular constant matrix.
(9 hrs, (בעשת))
Proof : Let Ψ satisfy the above equation then
(Ψ
∗
}
′
Φ+ Ψ
∗
Φ
′
= 0,
i.e.,
(Ψ
∗
}
′
Φ = −Ψ
∗
AΦ,
and since Φ is nonsingular
(Ψ
∗
}
′
= −Ψ
∗
A,
and it remains to take the Hermitian adjoint of both sides.
Conversely, if Ψ is a fundamental solution of the adjoint equation then
(Ψ
∗
Φ}
′
= 0,
which concludes the proof. B
34 Chapter 2
Corollary 2.1 Let A be antihermitian, i.e.,
A(t} = −A
∗
(t},
then the linear homogeneous system and its adjoint coincide. In particular, if Φ is
a fundamental matrix for (2.1) then it is also a fundamental matrix for its adjoint,
and
Φ
∗
(t}Φ(t} = C.
The matrix Φ
∗
(t}Φ(t} determines the magnitudes of the columns of Φ(t} and the
angles between pairs of vectors. In particular, since any solution is of the form
y(t} =
n
j=1
α
j
ϕ
j
(t},
then
¡y(t}¡
2
=
n
i, j=1
α
i
α
j
(ϕ
i
, ϕ
j
} =
n
i, j=1
α
i
α
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
ϕ
1
(t} = _
sint
cos t
_ and ϕ
2
(t} = _
cos t
−sint
_.
A fundamental matrix is therefore
Φ(t} = _
sint cos t
cos t −sint
_.
It follows that,
Φ
∗
(t}Φ(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} = α_
sint
cos t
_ + β_
cos t
−sint
_.
Now,
¡y(t}¡
2
= α
2
{sint cos t) _
sint
cos t
_ + 2αβ{sint cos t) _
cos t
−sint
_
+ β
2
{cos t −sint) _
cos t
−sint
_ = α
2
+ β
2
.
▲▲▲
2.2.3 The solution operator
Consider the initial value problem,
y
′
(t} = A(t}y(t}, y(t
0
} = given.
If Φ(t} is a fundamental matrix, then the solution is of the form
y(t} = Φ(t}c.
The vector c is found by substituting the initial data, y(t
0
} = Φ(t
0
}c, i.e., c =
Φ
−1
(t
0
}y(t
0
}, and the solution is therefore,
y(t} = Φ(t}Φ
−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 diﬀerent times
is the matrix R(t, t
0
}. We call the matrix R(t, t
0
} the solution operator
(ורתפה רוטרפוא) between times t
0
and t.
^ The solution operator seems to depend on the choice of the fundamental
matrix, but it doesn’t. It doesn’t because the solution is unique. We can also
see it as any other fundamental matrix Ψ(t} can be represented as Ψ(t} =
Φ(t}C, and then
Ψ(t}Ψ
−1
(t
0
} = Φ(t}CC
−1
Φ
−1
(t
0
} = Φ(t}Φ
−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 nonsingular and
R(t, t
0
}
−1
= R(t
0
, t}.
Let us recapitulate: given a homogeneous linear system, we need to ﬁnd 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 ﬁnd n independent solutions.
2.2.4 Order reduction
Consider an ndimensional linear homogeneous system, and suppose that we have
m linearly independent solutions
ϕ
1
, . . . , ϕ
m
,
with m < n.
Example: Consider the diﬀerential 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
ϕ
1
(t} = _
1]t
−1]t
2
_,
Linear equations 37
as
ϕ
′
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 ﬁnd 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 Nonhomogeneous linear systems
Consider now a system of the form
y
′
(t} = A(t}y(t} + b(t},
where A ∶ I →R
n×n
and b ∶ I →R
n
are continuous. Such a system is called a linear
nonhomogeneous system.
It turns out that if a fundamental matrix for the corresponding homogeneous sys
tem is known, then we can solve the nonhomogeneous system in explicitly form.
Theorem 2.2 Let Φ(t} be a fundamental matrix for the homogenous system, then
the solution for the nonhomogeneous 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
} = Φ(t}Φ
−1
(t
0
}.
Proof : Note that the solution operator also satisﬁes the homogeneous system,
R
′
(t, t
0
} = A(t}R(t, t
0
}
Diﬀerentiating 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, (בעשת))
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 nonhomogeneous
term.
The space of solutions is an aﬃne space Another point of practical interest is
the following. Suppose that y
in
(t} is a solution to the nonhomogeneous 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 nonhomogeneous 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} = Φ(t}C.
We then look for a solution where C is a function of time,
y(t} = Φ(t}C(t}.
Diﬀerentiating we get
y
′
(t} = Φ
′
(t}

A(t)Φ(t)
C(t} + Φ(t}C
′
(t} = A(t}y(t} + b(t} = A(t}Φ(t}C(t} + b(t},
hence
Φ(t}C
′
(t} = b(t} or C
′
(t} = Φ
−1
(t}b(t}.
Integrating we ﬁnd a solution to the nonhomogeneous equation,
C(t} =
t
t
0
Φ
−1
(s}b(s} ds i.e. y(t} =
t
t
0
Φ(t}Φ
−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} = Φ(t}C
1
+
t
t
0
Φ(t}Φ
−1
(s}b(s} ds,
and substituting the initial data we get that C
1
= Φ
−1
(t
0
}y(t
0
}.
Linear equations 39
Comment: The method shown here is a special case of something known as
Duhammel’s principle. It is a general approach for expressing the solution of
a linear nonhomogeneous 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
sinωt
_,
where ω is a frequency. This system describes a unit mass attached to a spring
with unit force constant and forced externally with frequency ω. The entries of
y(t} are the displacement and the momentum.
A fundamental matrix for the homogeneous system is
Φ(t} = _
sint cos t
cos t −sint
_.
Since
Φ
−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
sinωs
_ ds
= _
cos(t − t
0
} sin(t − t
0
}
−sin(t − t
0
} cos(t − t
0
}
_y(t
0
} +
t
t
0
_
sinωs sin(t − s}
sinωs cos(t − s}
_ ds.
Assume that t
0
= 0. Then
y(t} = _
cos t sint
−sint cos t
_y(0} +
1
ω
2
− 1
_
ωsint − sinωt
ω(cos t − cos ωt}
_.
The solution splits into a part that depends on the initial condition plus a part
that depends on the forcing. The expression for the nonhomogeneous term holds
40 Chapter 2
only of ω ≠ 1. If ω = 1, then the “natural frequency” coincides with the ”forcing
frequency, a phenomenon known as resonance. If ω = 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 coeﬃcients
2.4.1 The Jordan canonical form
Let A be a (constant) nbyn matrix. Recall that any matrix is similar to a matrix
in Jordan form: that is, there exists a nonsingular complex matrix P such that
A = PJP
−1
.
and J is of the form
J =
⎛
⎜
⎜
⎜
⎝
J
0
J
1
⋱
J
s
⎞
⎟
⎟
⎟
⎠
,
where
J
0
=
⎛
⎜
⎜
⎜
⎝
λ
1
λ
2
⋱
λ
q
⎞
⎟
⎟
⎟
⎠
and J
i
=
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎝
λ
q+i
1 0 0 ⋯ 0
0 λ
q+i
1 0 ⋯ 0
⋅ ⋅ ⋅ ⋅ ⋯ ⋅
0 0 0 0 ⋯ 1
0 0 0 0 ⋯ λ
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) nbyn matrix. We deﬁne
e
A
=
∞
n=0
A
n
n!
.
Linear equations 41
Is it welldeﬁned? Yes, because,
k>n
¡A
k
¡
k!
≤
k≥n
¡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 Cauchy’s 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 suﬃcient to learn to exponentiate matrices in Jordan canon
ical form.
Note that for a blockdiagonal matrix J,
e
J
=
⎛
⎜
⎜
⎜
⎝
e
J
0
e
J
1
⋱
e
Js
⎞
⎟
⎟
⎟
⎠
,
so that it is suﬃcient to learn to exponentiate diagonal matrices and matrices of
the form
J =
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎝
λ 1 0 0 ⋯ 0
0 λ 1 0 ⋯ 0
⋅ ⋅ ⋅ ⋅ ⋯ ⋅
0 0 0 0 ⋯ 1
0 0 0 0 ⋯ λ
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎠
.
The exponential of a diagonal matrix
J
0
=
⎛
⎜
⎜
⎜
⎝
λ
1
λ
2
⋱
λ
q
⎞
⎟
⎟
⎟
⎠
is e
J
0
=
⎛
⎜
⎜
⎜
⎝
e
λ
1
e
λ
2
⋱
e
λq
⎞
⎟
⎟
⎟
⎠
.
42 Chapter 2
The nondiagonal blocks have the following structure:
J
i
= λ
q+i
I
r
i
+ E
r
i
,
where r
i
is the size of the ith 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
λ
q+i
I
e
E
i
= e
λ
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
tλ
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 coeﬃcients
We now consider homogeneous linear systems in which the coeﬃcients are con
stant,
y
′
(t} = Ay(t},
where A is a (constant) nbyn matrix. As we know, solving this system amounts
to ﬁnding a fundamental matrix.
44 Chapter 2
Proposition 2.7 The matrix e
tA
is a fundamental matrix for the system with con
stant coeﬃcients A.
Proof : From the deﬁnition of the derivative,
lim
h→0
e
(t+h)A
− e
tA
h
= _lim
h→0
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
(t−t
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
(t−t
0
)A
y(t
0
} +
t
t
0
e
(t−s)A
b(s} ds.
´ Inspired by the case of n = 1, one may wonder whether the exponential is
only a fundamental solution for constant coeﬃcients. Isn’t the solution to
y
′
(t} = A(t}y(t},
Φ(t} = e
∫
t
0
A(s) ds
?
Let’s try to diﬀerentiate:
Φ(t + h} − Φ(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
λt
, where λ is an eigenvalue of A and m is between zero is one
minus the multiplicity of λ.
2.5 Linear diﬀerential equations of order n
2.5.1 The Wronskian
Let a
0
, a
1
, . . . , a
n
∶ I →Cbe continuous complexvalued functions deﬁned on some
interval. We deﬁne the linear diﬀerential operator, L
n
∶ C
1
(I} ↦C
0
(I}.
L
n
( f } = a
0
f
(n)
+ a
1
f
(n−1)
+ ⋯+ a
n−1
f
′
+ a
n
f ,
or in another notation,
L
n
= a
0
d
n
dt
n
+ a
1
d
n−1
dt
n−1
+ ⋯+ a
n−1
d
dt
+ a
n
.
(It will be assume that a
0
(t} ≠ 0 for all t ∈ I.)
A (scalar) diﬀerential 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
(n−1)
(t} −
a
2
(t}
a
0
(t}
y
(n−2)
(t} − ⋅ ⋅ ⋅ −
a
n−1
(t}
a
0
(t}
y
′
(t} −
a
n
(t}
a
0
(t}
y(t}.
As usual, we can convert an nth order equation into a ﬁrstorder 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
n−1
]a
0
−a
n−2
]a
0
−a
n−3
]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:
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎝
ϕ
ϕ
′
ϕ
′′
⋮
ϕ
(n−1)
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎠
.
In particular, a fundamental matrix is determined by n independent scalar func
tions ϕ
j
:
Φ =
⎛
⎜
⎜
⎜
⎜
⎝
ϕ
1
ϕ
2
⋯ ϕ
n
ϕ
′
1
ϕ
′
2
⋯ ϕ
′
n
⋅ ⋅ ⋯ ⋅
ϕ
(n−1)
1
ϕ
(n−1)
2
⋯ ϕ
(n−1)
n
⎞
⎟
⎟
⎟
⎟
⎠
.
The determinant of this matrix is called the Wronskian of the system L
n
y = 0
with respect to the solutions ϕ
1
, . . . , ϕ
n
. We denote it as follows:
W(ϕ
1
, . . . , ϕ
n
}(t} = det Φ(t}.
By Proposition 2.5
W(ϕ
1
, . . . , ϕ
n
}(t} = W(ϕ
1
, . . . , ϕ
n
}(t
0
} exp_
t
t
0
Tr A(s} ds_
= W(ϕ
1
, . . . , ϕ
n
}(t
0
} exp_−
t
t
0
a
1
(s}
a
0
(s}
ds_.
As we know, n solutions ϕ
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 ϕ
j
,
W(ϕ
1
, . . . , ϕ
n
}(t} = c exp_−
t
a
1
(s}
a
0
(s}
ds_
is known as Abel’s theorem.
TA material 2.1 Use Abel’s 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 nth order linear operator L
n
on I and n solutions ϕ
j
∶ I → R
satisfying L
n
ϕ
j
= 0 for which W(ϕ
1
, . . . , ϕ
n
}(t} ≠ 0, then every function y ∶ I → R
satisfying L
n
y = 0 is a linear combination of the ϕ
j
’s.
It turns out that the opposite is also true. To every n linearly independent functions
corresponds a linear diﬀerential operator of order n for which they form a basis:
Theorem 2.3 Let ϕ
1
, . . . , ϕ
n
be ntimes continuously diﬀerentiable functions on an
interval I such that
∀t ∈ I W(ϕ
1
, . . . , ϕ
n
}(t} ≠ 0.
Then there exists a unique linear diﬀerential operator of order n (assuming that
the coeﬃcient of the nth 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, ϕ
1
, . . . , ϕ
n
}(t}
W(ϕ
1
, . . . , ϕ
n
}(t}
= 0.
This is an nth order linear diﬀerential equation for y. Also the coeﬃcient of y
(n)
is 1.
Clearly,
W(ϕ
j
, ϕ
1
, . . . , ϕ
n
}(t} = 0
for all j’s, which means that the ϕ
j
’s form a basis for this equation.
48 Chapter 2
It remains to prove the uniqueness of this equation. The solutions ϕ
j
provide us
with a fundamental matrix Φ for the associated ﬁrstorder 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 ﬁrstorder 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 nth 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
n−1
0 −1 0 0 ⋯ 0 a
n−2
⋅ ⋅ ⋅ ⋅ ⋯ ⋅ ⋅
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
n−1
y
n
y
′
3
= −y
2
+ a
n−2
y
n
⋮
y
′
n−1
= −y
n−2
+ a
2
y
n
y
′
n
= −y
n−1
+ a
1
y
n
.
Diﬀerentiate the last equation (n − 1} times,
y
(n)
n
= −y
(n−1)
n−1
+ (a
1
y
n
}
(n−1)
.
Linear equations 49
Diﬀerentiating (n − 2} times the equation for y
′
n−1
and substituting:
y
(n)
n
= y
(n−2)
n−2
− (a
2
y
n
}
(n−2)
+ (a
1
y
n
}
(n−1)
.
We proceed, and eventually get an nth 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}
n−1
d
n−1
dt
n−1
a
1
+ (−1}
n−2
d
n−2
dt
n−2
a
2
− ⋯−
d
dt
a
n−1
+ a
n
.
Theorem 2.4 (Lagrange identity) Suppose that the coeﬃcients a
j
in the operator
L
n
are diﬀerentiable suﬃciently many times. Then for every pair of functions u, v
that are ntimes diﬀerentiable,
v L
n
u − u L
∗
n
v = uv
′
,
where uv is deﬁned as follows:
uv =
n
m=1
j+k=m−1
(−1}
j
u
(k)
(a
n−m
v}
( j)
.
Proof : We will prove the case n = 2. The general proof is left as an exercise. For
n = 2 (which is what we’ll 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 (Green’s 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 Nonhomogeneous equation
We know how to express the solution of the nonhomogeneous linear equation
in terms of the solution operator of the homogeneous system for a general ﬁrst
order system. In this section we derive an expression for the solution of an non
homogeneous nth order linear equation,
(L
n
y}(t} = b(t}.
We may assume without loss of generality that a
0
(t} ≡ 1. The corresponding
ﬁrstorder 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 diﬀerences 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 ϕ
1
, . . . , ϕ
n
be n independent solution of the homogeneous equa
tion L
n
y = 0, then the solution of the initialvalue problem
(L
n
y}(t} = b(t} Y(t
0
} = Y
0
,
is
y(t} = y
h
(t} +
n
k=1
ϕ
k
(t}
t
t
0
W
k
(ϕ
1
, ⋯, ϕ
n
}(s}
W(ϕ
1
, ⋯, ϕ
n
}(s}
b(s} ds,
where y
h
(t} is the solution to the homogeneous initialvalue problem, and
W
k
(vp
1
, ⋯, ϕ
n
} =
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
ϕ
1
⋯ ϕ
k−1
0 ϕ
k+1
⋯ ϕ
n
ϕ
′
1
⋯ ϕ
′
k−1
0 ϕ
′
k+1
⋯ ϕ
′
n
⋅ ⋯ ⋅ ⋮ ⋅ ⋯ ⋅
ϕ
(n−1)
1
⋯ ϕ
(n−1)
k−1
1 ϕ
(n−1)
k+1
⋯ ϕ
(n−1)
n
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
Proof : We knowthat for we can express the nonhomogeneous part of the solution
(in vector form) as
t
t
0
Φ(t}Φ
−1
(s}B(s} ds.
Since we are only interested in the ﬁrst component and by the structure of B(t},
y(t} = y
h
(t} =
n
j=1
t
t
0
Φ
1 j
÷
ϕ
j
(t)
(t}Φ
−1
jn
(s}b(s} ds
Thus, it only remains to show that
Φ
−1
jn
=
W
k
(ϕ
1
, ⋯, ϕ
n
}
W(ϕ
1
, ⋯, ϕ
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 coeﬃcients
We now consider the case of an nth order equation with constant coeﬃcients:
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
n−1
−a
n−2
−a
n−4
⋯ −a
1
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎠
.
Lemma 2.1 The characteristic polynomial of A is
p(λ} = det(λI − A} = λ
n
+ a
1
λ
n−1
+ ⋯+ a
n−1
λ + a
n
.
Proof : We prove it inductively on n. For n = 2
λI − A = _
λ −1
a
2
λ + a
1
_,
and so
p
2
(A} = λ(λ + a
1
} + a
2
= λ
2
+ a
1
λ + a
2
.
For n = 3,
λI − A =
⎛
⎜
⎝
λ −1 0
0 λ −1
a
3
a
2
λ + a
1
⎞
⎟
⎠
.
Hence,
p
3
(λ} = λp
2
(λ} + a
3
.
It is easy to see that p
n
(λ} = λ p
n−1
(λ} + a
n
. B
Linear equations 53
Theorem 2.6 Let λ
1
, . . . , λ
s
be distinct roots of the characteristic polynomial of A
such that λ
j
has multiplicity m
j
,
s
j=1
m
j
= n.
Then the space of solutions is spanned by the functions
e
λ
j
t
, t e
λ
j
t
, t
2
e
λ
j
t
, ⋯, t
m−1
e
λ
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 λ (not necessarily a root of the
characteristic polynomial):
L
n
{e
λt
) = p
n
(λ} e
λt
.
Let λ be a root of the characteristic polynomial of multiplicity m, and let 0 ≤ k ≤
m− 1. Then,
L
n
{t
k
e
λt
) = L
n
_
d
k
dλ
k
e
λt
_ =
d
k
dλ
k
L
n
{e
λt
) =
d
k
dλ
k
{p
n
(λ}e
λt
) .
If λ has multiplicity m and k < m, then
p
n
(λ} = p
′
n
(λ} = ⋯ = p
(k)
n
(λ} = 0,
which proves that
L
n
{t
k
e
λt
) = 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
λ
j
t
= 0.
Dividing by e
λ
1
t
,
P
1
(t} + e
(λ
2
−λ
1
)t
P
2
(t} + ⋯+ e
(λs−λ
1
)t
P
s
(t} = 0.
54 Chapter 2
The P
j
are polynomials. Diﬀerentiate suﬃciently many times until P
1
disappears.
We will get an equation of the form:
e
(λ
2
−λ
1
)t
Q
2
(t} + ⋯+ e
(λs−λ
1
)t
Q
s
(t} = 0.
We then divide by e
(λ
2
−λ
1
)t
, get
Q
2
(t} + e
(λ
3
−λ
2
)t
Q
3
(t} + ⋯+ e
(λs−λ
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
λ
4
− 1 = 0,
and its roots are λ = ±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 realvalued functions
e
t
, e
−t
, cos t and sint.
▲▲▲
Example: Consider the equation
y
′′′′
+ 2y
′′
+ y = 0.
The characteristic polynomial is
(λ
2
+ 1}
2
= 0
and it has two roots of multiplicity 2: λ = ±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 coeﬃcients: 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 aﬀected by the position of the moon).
Another important application of linear systems with periodic coeﬃcients is the
stability of periodic solutions. Consider a (generally nonlinear) diﬀerential 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 Tperiodic, 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
+ εη
0
,
where η ≪ 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} + εη(t}.
Substituting into the equation we get
z
′
(t} + ε η
′
(t} = f (z(t} + ε η(t}}.
56 Chapter 2
We then use Taylor’s theorem (assuming that f is twice diﬀerentiable):
z
′
(t} + ε η
′
(t} = f (z(t}} + ε ∇f (z(t}}η(t} + O(ε
2
}.
Since z
′
(t} = f (z(t}} we remain with
η
′
(t} = ∇f (z(t}}η(t} + O(ε}.
This is a linear equation with Tperiodic coeﬃcients. The rationale of this ap
proach is that if η(t} remains bounded in time then for small enough ε, the per
turbed solution will remain in an ε 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 coeﬃcients.
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 2πperiodic. 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 Φ(t} be a fundamental matrix for the periodic equation, then
Φ(t + T} is also a fundamental solution, and moreover, there exists a constant
matrix B, such that
Φ(t + T} = Φ(t}B,
and
det B = exp_
T
0
Tr A(s} ds_.
Linear equations 57
Proof : Let Φ(t} be a fundamental matrix and set Ψ(t} = Φ(t + T}. Then,
Ψ
′
(t} = Φ
′
(t + T} = A(t + T}Φ(t + T} = A(t}Ψ(t},
which proves the ﬁrst part.
As we know, if both Φ(t} and Ψ(t} are fundamental matrices, then there exists a
constant matrix B, such that
Ψ(t} = Φ(t + T} = Φ(t}B.
Finally, we use Proposition 2.5 to get that
det Φ(t + T} = det Φ(t} _
t+T
t
Tr A(s} ds_,
and since A(t} is periodic, we may as well integrate from 0 to T,
det Φ(t} det B = det Φ(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 Φ(t + T} = Φ(t}B implies that the fundamental matrix is in general
not periodic. Since this relation holds for every t, it follow that
B = Φ
−1
(0}Φ(T}.
Furthermore, we may always choose the fundamental matrix Φ(t} such that Φ(0} =
I, in which case B = Φ(T}, i.e., for every t,
Φ(t + T} = Φ(t}Φ(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:
Φ(t + kT} = Φ(t}B
k
.
Thus, for arbitrary t,
Φ(t} = Φ(t mod T}
 
Tperiodic
B
⌊tT⌋
,
where
t mod T = t − ⌊t]T⌋T.
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.
Deﬁnition 2.1 The eigenvalues of B = Φ(T}, ρ
i
are called the characteristic
multipliers (ייניפוא ילפוכ) of the periodic system. The constants µ
i
deﬁned by
ρ
i
= e
µ
i
T
are called the characteristic exponents (ייניפוא יטננופסקא) of the periodic
linear system. (Note that the characteristic exponents are only deﬁned modulo
2πik]T.)
Theorem 2.8 Let ρ and µ 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} = ρy(t}.
Also, there exists a solution y(t} such that for every t,
y(t + T} = e
µt
p(t},
where p(t} is Tperiodic.
Proof : Let Bu = ρu, and deﬁne y(t} = Φ(t}u. By the deﬁnition of a fundamental
matrix, y(t} is a solution of (2.3) and
y(t + T} = Φ(t + T}u = Φ(t}Bu = ρΦ(t}u = ρy(t},
which proves the ﬁrst part.
Linear equations 59
Next deﬁne p(t} = e
−µt
y(t}, then
p(t + T} = e
−µ(t+T)
y(t + T} = e
−µT
ρ

1
e
−µt
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 (ρ
i
, µ
i
}, and the general solution to the periodic linear sys
tem is
y(t} =
n
k=1
α
k
e
µ
k
t
p
k
(t},
where the functions p
k
(t} are Tperiodic. In other words, every solution is of the
form
y(t} =
⎛
⎜
⎝
⋮ ⋯ ⋮
p
1
(t} ⋯ p
n
(t}
⋮ ⋯ ⋮
⎞
⎟
⎠
⎛
⎜
⎝
e
µ
1
t
⋱
e
µnt
⎞
⎟
⎠
 
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 ﬁnding 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 Hill’s 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 coeﬃcient 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 secondorder equation
with nonconstant coeﬃcients.
As usual, we start by rewriting this equation as a ﬁrst 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., ρ
1
ρ
2
= 1.
We are looking for a fundament matrix
Φ(t} = _
ϕ
11
(t} ϕ
12
(t}
ϕ
21
(t} ϕ
22
(t}
_,
with Φ(0} = I. Since,
ϕ
′
1j
(t} = ϕ
2 j
(t},
it follows that
Φ(t} = _
ϕ
11
(t} ϕ
12
(t}
ϕ
′
11
(t} ϕ
′
12
(t}
_.
Thus,
B = _
ϕ
11
(T} ϕ
12
(T}
ϕ
′
11
(T} ϕ
′
12
(T}
_.
Moreover, since the Wronskian is constant, W(t} = det B = 1, it follows that
ϕ
11
(t}ϕ
′
12
(t} − ϕ
12
(t}ϕ
′
11
(t} = 1.
The characteristic multiplier are the eigenvalues of B. For any 2by2 matrix
B = _
a b
c d
_
the eigenvalues are satisfy the characteristic equation
(a − ρ}(d − ρ} − bc = ρ
2
− (Tr B}ρ + det B = 0,
and since det B = 1 we get that
ρ
1,2
= γ ±
_
γ
2
− 1,
Linear equations 61
where γ =
1
2
Tr B =
1
2
ϕ
11
(T} + ϕ
′
12
(T}. Note that
ρ
1
+ ρ
2
= 2γ and ρ
1
ρ
2
= 1.
The characteristic exponents are ρ
1,2
= e
µ
1,2
T
, which implies that
e
(µ
1
+µ
2
)T
= 1,
hence µ
1
+ µ
2
= 0, and
2γ = e
µ
1
T
+ e
µ
2
T
= 2coshµ
1
T.
Thus, we are able to express the characteristic coeﬃcients in terms of the single
parameter γ:
coshµ
1
T and µ
1
+ µ
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
γ.
' Case 1: γ > 1 In this case ρ
1,2
are both real and positive, and furthermore,
0 < ρ
2
< 1 < ρ
1
.
Thus, µ
1
and µ
2
= −µ
1
are both realvalued. The general solution is of the
form
y(t} = c
1
e
µ
1
t
p
1
(t} + c
2
e
−µ
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: γ < −1 In this case both ρ
1,2
are real and negative,
ρ
2
< −1 < ρ
1
< 0,
in which case
e
µ
1
T
= ρ
1
implies that
e
(µ
1
−ıπT)T
= ρ
1
,
i.e.,
µ
1
=
iπ
T
+ logρ
1
.
62 Chapter 2
Thus, the general solution is of the form
y(t} = c
1
e
log ρ
1
t
e
iπtT
p
1
(t}
 
period 2T
+c
2
e
−log ρ
1
t
e
−iπtT
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 < γ < 1 Now ρ
1,2
are both complex and located on the unit
circle,
ρ
1,2
= e
±iσT
(this relation deﬁnes σ) and thus µ
1,2
= ±iσ. In this case the general solution
is of the form
y(t} = c
1
e
iσt
p
1
(t} + c
2
e
−iσt
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 σT = 2π]m for some integer m.
´ Case 4: γ = 1
I Case 5: γ = −1
Chapter 3
Boundary value problems
3.1 Motivation
Many problems in science involve diﬀerential equations with conditions that are
speciﬁed at more than one point. Boundary value problems arise also in the study
of partial diﬀerential equations.
Example: The equation that describes the motion of an elastic string is known as
the wave equation (ילגה תאוושמ),
∂
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 ﬁnite 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 initialboundary 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 = πm]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.
Let’s get back to the same boundary value problem, this time on the interval 0, 1:
⎧
⎪
⎪
⎨
⎪
⎪
⎩
Ly = λy
y(0} = y(1} = 0,
where
L = −
d
2
dx
2
.
Boundary value problems 65
The parameter λ 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 = λy is
y(x} = Asinλ
12
x + Bcos λ
12
x.
Since we require y(0} = 0 then B = 0. Since we further require y(1} = 0, it follows
that
λ = π
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 = λy with boundary values has a solution only
for selected values of λ. Does it remind us something? L is a linear operator, λ
is a scalar... yes, this looks like an eigenvalue problem! Indeed, the values of λ
for which the boundary value problem has a solution are called eigenvalues. The
corresponding “eigenvectors” are rather called eigenfunctions (תוימצע תויצקנופ).
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
ψ
m
(x} =
√
2 sin(mπx}.
Note that this is an orthonormal set,
(ψ
m
, ψ
k
} = 2
1
0
sin(mπx} sin(kπx} dx = δ
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 didn’t take this course, no worry. We will
provide all the needed information.
For a function f ∈ L
2
0, 1, its innerproduct with the elements of the orthogonal
set ψ
k
are called its Fourier coeﬃcients,
ˆ
f (k} = ( f , ψ
k
} =
√
2
1
0
f (x} sin(kπx} dx.
66 Chapter 3
The series
∞
k=1
ˆ
f (k}ψ
k
(x}
converges in the L
2
0, 1 norm to f , namely,
lim
n→∞
_f −
n
k=1
ˆ
f (k}ψ
k
_ = 0.
Finally, the norm of f and its Fourier coeﬃcients satisfy a relation known as the
Parseval identity:
¡ f ¡
2
=
∞
k=1

ˆ
f (k}
2
.
We have stated all these facts without any proof or any other form of justiﬁcation.
As we will see, these results hold in a much wider scope.
TA material 3.1 Solve the wave equation initialvalue problem. Talk about har
monics. Talk about what happens when you touch the center of a guitar string.
Comment: An interesting connection between these inﬁnitedimensional 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} = λy(x}, y(0} = y(1} = 0.
We can approximate the solution by deﬁning a mesh,
x
i
=
i
N
, i = 0, . . . , N,
and look for y
i
= y(x
i
}, by approximating the diﬀerential equation by a ﬁnite
diﬀerence approximation,
−
y
i−1
+ y
i+1
− 2y
i
2∆x
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
N−2
y
N−1
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎠
= λ
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎝
y
1
y
2
⋮
y
N−2
y
N−1
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎠
.
Boundary value problems 67
This is a “standard” eigenvalue problem, for which we expect N eigenvalues. As
we reﬁne the mesh and take N → ∞, the number of eigenvalues is expected to
tend to inﬁnity.
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 Selfadjoint eigenvalue problems on a ﬁnite in
terval
3.2.1 Deﬁnitions
Consider a linear nth order diﬀerential operator,
L = p
0
(x}
d
n
dx
n
+ p
2
(x}
d
n−1
dx
n−1
+ ⋅ ⋅ ⋅ + p
n−1
(x}
d
dx
+ p
n
(x},
where p
j
∈ C
n−j
(a, b} is in general complexvalued; we assume that p
0
(x} ≠ 0 for
all x ∈ (a, b}. We also deﬁne a linear operator U ∶ C
n
(a, b} →R
n
of the form
U
k
y =
n
j=1
{M
k j
y
( j−1)
(a} + N
k j
y
( j−1)
(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
(n−1)
(a}
+ N
k1
y(b} + N
k2
y
′
(b} + ⋯+ N
kn
y
(n−1)
(b}.
The equation
Ly = λy, 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. ▲▲▲
Deﬁnition 3.1 The eigenvalue problem is said to be selfadjoint (ומצעל דומצ) 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 selfadjoint. In the ﬁrst 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}
n−1
d
n−1
dx
n−1
p
1
(x} + ⋯−
d
dx
p
n−1
(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 selfadjoint.
A selfadjoint problem always has a trivial solution for any value of λ. Values of
λ for which a nontrivial solution exists are called eigenvalues; the corresponding
functions are called eigenfunctions.
Boundary value problems 69
3.2.2 Properties of the eigenvalues
Proposition 3.1 Consider a selfadjoint 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 (ﬁnitevalued)
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 λ be an eigenvalue with eigenfunction ψ. Then,
λ(ψ, ψ} = (Lψ, ψ} = (ψ, Lψ} = λ(ψ, ψ},
i.e.,
(λ − λ}(ψ, ψ} = 0,
which proves that ψ is real.
Let λ ≠ µ be distinct eigenvalues,
Lψ = λψ, Lϕ = µϕ.
Then,
λ(ψ, µ} = (Lψ, ϕ} = (ψ, Lϕ} = µ(ψ, ϕ},
i.e.,
(λ − µ}(ψ, ϕ} = 0,
which implies that ψ and ϕ are orthogonal.
Let now ϕ
j
(x, λ}, j = 1, . . . , n, be solutions of the equation Lϕ
j
= λϕ
j
with initial
conditions,
ϕ
(k−1)
j
(c, λ} = δ
jk
,
70 Chapter 3
where c is some point in (a, b}. That is,
ϕ
1
(c} = 1 ϕ
′
1
(c} = ϕ
′′
1
(c} = ϕ
′′′
1
(c} = ⋯ = ϕ
(n−1)
1
(c} = 0
ϕ
′
2
(c} = 1 ϕ
2
(c} = ϕ
′′
2
(c} = ϕ
′′′
2
(c} = ⋯ = ϕ
(n−1)
2
(c} = 0
ϕ
′′
3
(c} = 1 ϕ
3
(c} = ϕ
′
3
(c} = ϕ
′′′
3
(c} = ⋯ = ϕ
(n−1)
3
(c} = 0
etc.
Note, we are back in initial values problems; there no issue of existence nor non
uniqueness, i.e., the ϕ
j
(x, λ} are deﬁned for every (complex) λ. Also, the functions
ϕ
j
are continuous both in x and λ. In fact, for ﬁxed x the function λ ↦ ϕ(x, λ} is
analytic.
1
The functions ϕ
j
are linearly independent, hence every solution of the equation
Ly = λy is of the form
y(x} =
n
j=1
c
j
ϕ
j
(x, λ}.
For λ to be an eigenvalue we need
U
y =
n
j=1
c
j
U
ϕ
j
(⋅, λ} = 0, = 1, . . . , n,
i.e.,
n
j=1
n
k=1
[M
k
ϕ
(k−1)
j
(a, λ} + N
k
ϕ
(k−1)
j
(b, λ})
 
A
j
(λ)
c
j
= 0, = 1, . . . , n.
This is a homogeneous linear system for the coeﬃcients c
j
. For it to have a non
trivial solution the determinant of the matrix A
j
(λ} must vanish. That is, the
eigenvalues λ are identiﬁed with the roots of det A
j
(λ}.
det A
j
(λ} is analytic in the entire complex plane. It is not identically zero because
it can only vanish on the real line. It is wellknown (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 inﬁnity). B
1
A concept learned in complex analysis; I will say a few words about it.
Boundary value problems 71
3.2.3 Nonhomogeneous boundary value problem
We turn now to study the nonhomogeneous equation
Ly = λy + f , Uy = 0,
where f ∈ C(a, b}. What to expect? Think of the linearalgebraic analog,
Ax = λx + b.
A (unique) solution exists if the matrix A − λI is not singular, i.e., if λ is not an
eigenvalue of A. Similarly, we will see that the nonhomogeneous boundary value
problem has a solution for values of λ 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 Ω ⊂ R
n
with a
metallic boundary is
∆ϕ = 4πρ, ϕ
∂Ω
= 0,
where ∆ is the Laplacian and ρ(x} is the charge density. For n = 1 we get
d
2
ϕ
dx
2
= 4πρ ϕ(0} = ϕ(L} = 0,
which is of the same type as above with λ = 0. ▲▲▲
Let ϕ
j
(x, λ} be as in the previous section solutions to Lϕ
j
(⋅, λ} = λϕ
j
(⋅, λ}, subject
to the boundary conditions
ϕ
(k−1)
j
(c, λ} = δ
jk
.
Note that
W(ϕ
1
, . . . , ϕ
n
}(c} = det I = 1,
hence
W(ϕ
1
, . . . , ϕ
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 − λid.
72 Chapter 3
We now deﬁne the following function:
K(x, ξ, λ} =
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
1
p
0
(ξ}W(ϕ
1
, . . . , ϕ
n
}(ξ}
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
ϕ
1
(ξ, λ} ⋯ ϕ
n
(ξ, λ}
ϕ
′
1
(ξ, λ} ⋯ ϕ
′
n
(ξ, λ}
⋅ ⋯ ⋅
ϕ
(n−2)
1
(ξ, λ} ⋯ ϕ
(n−2)
n
(ξ, λ}
ϕ
1
(x, λ} ⋯ ϕ
n
(x, λ}
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
ξ ≤ x
0 x < ξ.
This function depends on x and ξ, both taking values in a, b. It is apparently
discontinuous, however note that
K(ξ
+
, ξ, λ} =
d
dx
K(ξ
+
, ξ, λ} = ⋯ =
d
n−2
dx
n−2
K(ξ
+
, ξ, λ} = 0,
hence K and its ﬁrst n − 2 derivatives with respect to x are continuous over all
(x, ξ}; in particular K is continuous for n ≥ 2. The (n − 1}st and nth derivatives
are continuous in every subdomain. Also,
d
n−1
dx
n−1
K(ξ
+
, ξ, λ} =
1
p
0
(ξ}
.
Finally, for x > ξ,
LK(x, ξ, λ} =
1
p
0
(ξ}W(ϕ
1
, . . . , ϕ
n
}(ξ}
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
ϕ
1
(ξ, λ} ⋯ ϕ
n
(ξ, λ}
ϕ
′
1
(ξ, λ} ⋯ ϕ
′
n
(ξ, λ}
⋅ ⋯ ⋅
ϕ
(n−2)
1
(ξ, λ} ⋯ ϕ
(n−2)
n
(ξ, λ}
Lϕ
1
(x, λ} ⋯ Lϕ
n
(x, λ}
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
= λK(x, ξ, λ}.
This relation holds trivially for x < ξ.
Deﬁne now the function
u(x, λ} =
b
a
K(x, ξ, λ} f (ξ} dξ =
x
a
K(x, ξ, λ} f (ξ} dξ.
We calculate its ﬁrst n derivatives with respect to n,
u
′
(x, λ} =
x
a
d
dx
K(x, ξ, λ} f (ξ} dξ + K(x, x
−
, λ} f (x}
 
0
,
Boundary value problems 73
u
′′
(x, λ} =
x
a
d
2
dx
2
K(x, ξ, λ} f (ξ} dξ +
d
dx
K(x, x
−
, λ} f (x}
 
0
,
and so until
u
(n−1)
(x, λ} =
x
a
d
n−1
dx
n−1
K(x, ξ, λ} f (ξ} dξ +
d
n−2
dx
n−2
K(x, x
−
, λ} f (x}
  
0
.
Finally,
u
(n)
(x, λ} =
x
a
d
n
dx
n
K(x, ξ, λ} f (ξ} dξ +
d
n−1
dx
n−1
K(x, x
−
, λ} f (x}
  
f (x)p
0
(x)
.
It follows that
Lu(x, λ} =
x
a
LK(x, ξ, λ} f (ξ} dξ + f (x} = λu(x, λ} + f (x},
i.e., u(x, λ} is a solution to the nonhomogeneous equation.
We are looking for a solution of the nonhomogeneous equation that satisﬁes the
boundary conditions, U
k
y = 0. Deﬁne
G(x, ξ, λ} = K(x, ξ, λ} +
n
j=1
c
j
(ξ}ϕ
j
(x, λ},
and set c
j
(ξ} such that for every k = 1, . . . , n and every ξ ∈ a, b:
U
k
G(⋅, ξ, λ} = U
k
K(⋅, ξ, λ} +
n
j=1
c
j
(ξ}U
k
ϕ
j
(⋅, λ} = 0.
This linear system has a unique solution if det U
k
ϕ
j
(⋅, λ} ≠ 0, i.e., if λ is not an
eigenvalue of the homogeneous boundary value problem.
Set then
u(x, λ} =
b
a
G(x, ξ, λ} f (ξ} dξ
=
b
a
K(x, ξ, λ} f (ξ} dξ +
n
j=1
ϕ
j
(x}
b
a
c
j
(ξ} f (ξ} dξ.
74 Chapter 3
Then,
Lu(x, λ} = λu(x, λ},
and
U
k
u(⋅, λ} =
b
a
_U
k
K(⋅, ξ, λ} +
n
j=1
c
j
(ξ}U
k
ϕ(⋅, λ}_ f (ξ} dξ = 0.
Thus, we found a solution to the nonhomogeneous boundary value problem for
every value of λ that is not an eigenvalue of the homogeneous boundary value
problem.
Exercise 3.2 Solve the boundary value problem,
−y
′′
(x} = λy(x} + f (x}, y(0} = y(1} = 0.
For what values of λ solutions exist? Repeat the above analysis (including the
calculation of K(x, ξ, λ} and G(x, ξ, λ} for this particular example.
Solution 3.2: We start by constructing the two independent solution,
−ϕ
′′
1
(x) = λϕ
1
(x), ϕ
1
(0) = 1 ϕ
′
1
(0) = 0
−ϕ
′′
2
(x) = λϕ
2
(x), ϕ
2
(0) = 0 ϕ
′
2
(0) = 1,
namely,
ϕ
1
(x) = cos λ
12
x and ϕ
2
(x) = λ
−12
sinλ
12
x.
Thus,
W(ϕ
1
, ϕ
2
)(x) =
cos λ
12
x λ
1−02
sinλ
12
x
−λ
12
sinλ
12
x cos λ
12
x
= 1.
Next, since p
0
(ξ) = −1,
K(x, ξ, λ) =
⎧
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎩
−
cos λ
12
ξ sinλ
12
ξ
cos λ
12
x sinλ
12
x
= sinλ
12
(ξ − x) ξ ≤ x
0 x < ξ
.
Now,
G(x, ξ, λ) = K(x, ξ, λ) − c
1
(ξ) cos λ
12
x − c
2
(ξ)λ
−12
sinλ
12
x.
The coeﬃcients c
1,2
have to satisfy,
G(0, ξ, λ) = 0 − c
1
(ξ) = 0,
and
G(1, ξ, λ) = sinλ
12
(ξ − 1) − c
1
(ξ) cos λ
12
− c
2
(ξ)λ
−12
sinλ
12
= 0,
Boundary value problems 75
from which we get that
G(x, ξ, λ) = K(x, ξ, λ) −
sinλ
12
(ξ − 1)
sinλ
12
sinλ
12
x
Thus, the solution to the boundary value problem is
u(x, λ) =
x
0
sinλ
12
(ξ − x) f (ξ) dξ −
sinλ
12
x
sinλ
12
1
0
sinλ
12
(ξ − 1) f (ξ) dξ.
It is easy to see that the boundary conditions are indeed satisﬁed. You may now check by direct
diﬀerentiation that the diﬀerential equation is indeed satisﬁed.
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 λ
0
that is not en eigenvalue and then L − λ
0
I
will be used as an operator for which zero is not en eigenvalue.
We will denote G(x, ξ, 0} simply by G(x, ξ}. Note that by deﬁnition LG(x, ξ} = 0
at all points where x ≠ ξ.
Deﬁne the linear integral operator:
G ∶ f ↦
b
a
G(⋅, ξ} f (ξ} dξ,
which maps the nonhomogeneous term into the solution of the boundary value
problem.
Proposition 3.2 The integral operator G (and the Green function G) satisﬁes the
following properties:
' LG f = f and UG f = 0 for all f ∈ C(a, b}.
^ (G f , g} = ( f , Gg}.
· G(x, ξ} = G(ξ, 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 ﬁrst statement follows from the fact that G maps f into the solution
of the boundary value problem. By the selfadjointness of L and the fact that
UG f = UGg = 0,
(LG f , Gg} = (G f , LGg},
hence by the ﬁrst statement
( f , Gg} = (G f , g},
which proves the second statement.
It follows from the second statement that
b
a
_
b
a
G(x, ξ} f (ξ} dξ_g(x} dx =
b
a
f (x}_
b
a
G(x, ξ}g(ξ} dξ_dx.
Changing the order of integration and interchanging x and ξ in the second integral:
b
a
b
a
[G(x, ξ} −G(ξ, x}) f (ξ}g(x} dxdξ = 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 haven’t shown that the eigenvalue problem
Ly = λy, Uy = 0, (3.1)
has solutions. We only know properties of solutions if such exist. In this section
we will show that selfadjoint eigenvalue problems always have a countable set of
eigenvalues.
Assuming that zero is not an eigenvalue we deﬁned the Green function G(x, ξ}
and the operator G. The ﬁrst step will be to show that eigenfunctions of L coincide
with eigenfunctions of G:
Boundary value problems 77
Proposition 3.3 ϕ is an eigenfunction of (3.1) with eigenvalue λ if and only if it
is also an eigenfunction of G with eigenfunction 1]λ, namely
Gϕ =
1
λ
ϕ, (3.2)
Proof : Let (λ, ϕ} be eigensolutions of (3.1). Then,
Gϕ =
1
λ
Gλϕ =
1
λ
GLϕ,
and since Uϕ = 0, it follows from the fourth statement of Proposition 3.2 that
Gϕ =
1
λ
ϕ.
Conversely, suppose that (3.2) holds. Then,
Lϕ = λL
1
λ
ϕ = λLGϕ = λϕ,
and
Uϕ = λUGϕ = 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 supnorm) 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
diﬀerent treatment. For n ≥ 2 the function G(x, ξ} is continuous on the square
(x, ξ} ∈ a, b
2
, hence it is uniformly continuous, and in particular,
(∀ε > 0}(∃δ > 0} ∶ (∀ξ ∈ a, b} (∀x
1
, x
2
∶ x
1
−x
2
 < δ} G(x
1
, ξ}−G(x
2
, ξ} < ε.
Thus,
(∀ε > 0}(∃δ > 0} ∶ (∀u ∈ X} (∀x
1
, x
2
∶ x
1
− x
2
 < δ}
it holds that
Gu(x
1
} − Gu(x
2
} ≤
b
a
G(x
1
, ξ} −G(x
2
, ξ}u(ξ} dξ ≤ ε
b
a
u(ξ} dξ.
Using the CauchySchwarz inequality,
Gu(x
1
} − Gu(x
2
} ≤ _
b
a
u(ξ}
2
dξ_
12
_
b
a
dξ_
12
= ε(b − a}
12
¡u¡,
which proves the equicontinuity of X.
Since G(x, ξ} 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, ξ}u(ξ} dξ ≤ K(b − a}
12
¡u¡,
which proves the uniform boundedness of X. B
It follows from the ArzelaAscoli 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 deﬁne 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 deﬁned it as a map from C(a, b) to C
n
(a, b) to avoid measuretheoretic subtleties.
Boundary value problems 79
Proposition 3.4
¡G¡ = sup
u=1
(Gu, u}.
Proof : For every normalized u, it follows from the CauchySchwarz inequality
and the deﬁnition of the norm that
(Gu, u} ≤ ¡Gu¡¡u¡ ≤ ¡G¡¡u¡
2
= ¡G¡,
i.e.,
sup
u=1
(Gu, u} ≤ ¡G¡. (3.3)
By the bilinearity of the innerproduct and the selfadjointness 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 ﬁrst,
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,
4¡Gu¡ ≤ 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 ϕ such that
Gϕ = ¡G¡ϕ or Gϕ = −¡G¡ϕ,
it follows that
¡Gϕ¡ = ¡G¡¡ϕ¡,
thus the sup in the deﬁnition of ¡G¡ is attained.
· Everything we have been doing holds in the ﬁnitedimensional case. If A is
an hermitian matrix then either ¡A¡ or −¡A¡ is an eigenvalue.
Proof : Denote ¡G¡ = µ
0
. then either
µ
0
= sup
u=1
(Gu, u}
or
µ
0
= − sup
u=1
(Gu, u}.
Consider the ﬁrst case; the second case is treated similarly.
By the deﬁnition of the supremum, there exists a sequence of functions (u
n
} such
that
¡u
n
¡ = 1 lim
n→∞
(Gu
n
, u
n
} = µ
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 ϕ
0
. In particular, uniform conver
gence implies L
2
convergence,
lim
n→∞
¡Gu
n
− ϕ
0
¡ = 0.
Since
0 ←(Gu
n
− ϕ
0
, Gu
n
− ϕ
0
} = ¡Gu
n
¡
2
+ ¡ϕ
0
¡
2
− 2R(Gu
n
, ϕ
0
} →¡Gu
n
¡
2
− ¡ϕ
0
¡
2
,
we get the convergence of the norm,
lim
n→∞
¡Gu
n
¡ = ¡ϕ
0
¡.
(It is always true that convergence in norm implies the convergence of the norm.)
Now,
¡Gu
n
− µ
0
u
n
¡
2
= ¡Gu
n
¡
2
+ µ
2
0
− 2µ
0
(Gu
n
, u
n
} →¡ϕ
0
¡
2
− µ
2
0
,
from which follows that ϕ
0
≠ 0 (we need this to claim that ϕ
0
is nontrivial).
Moreover, it follows that
0 ≤ ¡Gu
n
− µ
0
u
n
¡
2
≤ 2µ
2
0
− 2µ
0
(Gu
n
, u
n
} →0,
from which follows that
lim
n→∞
¡Gu
n
− µ
0
u
n
¡ = 0.
Finally, by the triangle inequality
¡Gϕ
0
− µ
0
ϕ
0
¡ ≤ ¡Gϕ
0
− G(Gu
n
}¡ + ¡G(Gu
n
} − µ
0
Gu
n
¡ + ¡µ
0
Gu
n
− µ
0
ϕ
0
¡
≤ ¡G¡¡ϕ
0
− Gu
n
¡ + ¡G¡¡Gu
n
− µ
0
u
n
¡ + µ
0
¡Gu
n
− ϕ
0
¡.
Letting n →∞ we get that
Gϕ
0
= µ
0
ϕ
0
,
which concludes the proof. B
Theorem 3.2 G has an inﬁnite number of eigenfunctions.
82 Chapter 3
Proof : Let ϕ
0
and µ
0
be as above and assume that ϕ
0
has been normalized, ¡ϕ
0
¡ =
1. Deﬁne a new kernel
G
1
(x, ξ} = G(x, ξ} − µ
0
ϕ
0
(x}ϕ
0
(ξ},
and the corresponding operator,
G
1
f (x} =
b
a
G
1
(x, ξ} f (ξ} dξ.
This operator is also selfadjoint,
(G
1
u, v} = (Gu, v} + µ
0 ∬
b
a
ϕ
0
(x}ϕ
0
(ξ}u(ξ}v(x} dξdx = (u, G
1
v}.
Hence, unless G
1
= 0 (the zero operator), it has an eigenvalue µ
1
, whose absolute
value is equal to ¡G
1
¡. Let,
G
1
ϕ
1
= µ
1
ϕ
1
¡ϕ
1
¡ = 1.
Note that G
1
ϕ
0
= 0, hence
(ϕ
1
, ϕ
0
} =
1
µ
1
(G
1
ϕ
1
, ϕ
0
} =
1
µ
1
(ϕ
1
, G
1
ϕ
0
} = 0,
i.e., ϕ
0
and ϕ
1
are orthogonal. Then,
Gϕ
1
= G
1
ϕ
1
+ µ
0
ϕ
1
(ϕ
1
, ϕ
0
} = G
1
ϕ
1
= µ
1
ϕ
1
,
which means that ϕ
1
is also an eigenfunction of G. By the maximality of µ
0
,
µ
1
 ≤ µ
0
.
We continue this way, deﬁning next
G
2
(x, ξ} = G
1
(x, ξ} − µ
1
ϕ
1
(x}ϕ
1
(ξ}
and ﬁnd 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} −
m−1
j=1
µ
j
ϕ
j
(x}( f , ϕ
j
} = 0.
Boundary value problems 83
Applying L on both sides, using the fact that LG f = f and Lϕ
j
= µ
−1
j
ϕ
j
,
f (x} =
m−1
j=1
ϕ
j
(x}( f , ϕ
j
}.
This would imply that every continuous function is in the span of m diﬀerentiable
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, ϕ
j
, with
corresponding eigenvalues
µ
0
 ≥ µ
1
 ≥ µ
2
 ≥ ⋯.
Lemma 3.2 (Bessel’s inequality) Let f ∈ L
2
(a, b}. Then the sequence
∞
j=0
( f , ϕ
j
}
2
converges and
∞
j=0
( f , ϕ
j
}
2
≤ ¡ f ¡
2
.
Proof : For every m ∈ N,
0 ≤ _f −
m
j=0
( f , ϕ
j
}ϕ
j
_
2
= ¡ f ¡
2
−
m
j=0
( f , ϕ
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 , ϕ
j
}ϕ
j
,
where the convergence is uniform.
Corollary 3.2 An immediate consequence is Parseval’s 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 , ϕ
j
}
2
Proof : For every j ∈ Mand x ∈ a, b,
(Gϕ
j
}(x} =
b
a
G(x, ξ}ϕ
j
(ξ} dξ = µ
j
ϕ
j
(x}.
Taking the complex conjugate and using the fact that G(x, ξ} = G(ξ, x},
(G(x, ⋅}, ϕ
j
} =
b
a
G(ξ, x}ϕ
j
(ξ} dξ = µ
j
ϕ
j
(x}.
Squaring and summing over j,
m
j=0
(G(x, ⋅}, ϕ
j
}
2
=
m
j=0
µ
j

2
ϕ
j
(x}
2
=
m
j=0
µ
j

2
.
By Bessel’s inequality applied to G(x, ⋅},
m
j=0
µ
j

2
≤ ¡G(x, ⋅}¡
2
=
b
a
G(x, ξ}
2
dξ.
Integrating over x,
m
j=0
µ
j

2
≤
∬
b
a
G(x, ξ}
2
dξdx ≤ K
2
(b − a}
2
,
where K is a bound on G(x, ξ} It follows at once that ∑
∞
j=0
µ
j

2
converges and in
particular µ
j
→0.
Boundary value problems 85
For m ∈ N consider
G
m
(x, ξ} = G(x, ξ} −
m−1
j=0
µ
j
ϕ
j
(x}ϕ
j
(ξ}.
By the way we constructed the sequence of eigenfunctions, we know that
¡G
m
¡ = µ
m
.
Thus, for every u ∈ C(a, b},
¡G
m
u¡ = _Gu −
m−1
j=0
µ
j
(u, ϕ
j
}ϕ
j
_ ≤ µ
m
¡u¡.
Letting m →∞,
lim
m→∞
_Gu −
m−1
j=0
µ
j
(u, ϕ
j
}ϕ
j
_ = 0.
This means that ∑
∞
j=0
µ
j
(u, ϕ
j
}ϕ
j
converges to Gu in L
2
. We will show that it
converges in fact uniformly. For q > p,
q
j=p
µ
j
(u, ϕ
j
}ϕ
j
= G _
q
j=p
(u, ϕ
j
}ϕ
j
_.
Hence,
_
q
j=p
µ
j
(u, ϕ
j
}ϕ
j
_ ≤ K(b − a}
12
_
q
j=p
(u, ϕ
j
}ϕ
j
_ = K(b − a}
12
_
q
j=p
(u, ϕ
j
}
2
_
12
,
where we used the CauchySchwarz inequality. Since the right hand side vanishes
as p, q →∞, it follows by Cauchy’s criterion that ∑
q
j=p
µ
j
(u, ϕ
j
}ϕ
j
converges uni
formly. Since it converges to Gu in L
2
and the latter is continuous, it follows that
∞
j=0
µ
j
(u, ϕ
j
}ϕ
j
= Gu,
where the convergence is uniform.
Let now f ∈ C
n
(a, b}, U f = 0. Then, Lf is continuous, and
∞
j=0
µ
j
(Lf , ϕ
j
}ϕ
j
= GLf .
86 Chapter 3
Since GLf = f and Lϕ
j
= µ
−1
j
ϕ
j
,
∞
j=0
( f , ϕ
j
}ϕ
j
= f .
B
Chapter 4
Stability of solutions
4.1 Deﬁnitions
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 ﬁelds f (t, ⋅} is called autonomous if it does not depend on t,
i.e., f ∶ R
n
→R
n
.
Note that if ϕ is a solution to the autonomous initial data problem,
y
′
(t} = f (y(t}}, y(0} = y
0
,
then ψ(t} = ϕ(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 yplane that are of special
interest:
y = (0, 0} y = (2, 2} y = (0, 1} and y = (−1, 2}.
88 Chapter 4
These are ﬁxed points (תבש תודוקנ), or equilibrium points (לקשמ יוויש תודוקנ),
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 ﬁxed points.
Roughly speaking, a ﬁxed 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 ﬁxed
point.
The notion of stability can be made more precise and more general:
Deﬁnition 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 ε > 0 there exists a δ > 0,
such that for every solution y(t} satisfying ¡y(t
0
} − u(t
0
}¡ < δ,
¡y(t} − u(t}¡ < ε ∀t > t
0
.
It is called asymptotically stable if there exists a δ such that ¡y(t
0
} − u(t
0
}¡ < δ
implies that
lim
t→∞
¡y(t} − u(t}¡ = 0.
Comment: The stability of ﬁxed points is a particular case of stability of solutions
as deﬁned 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 ﬁxed point.
· Limit cycle (periodic solution).
´ Asymptotic convergence to a limit cycle.
I Quasiperiodic.
I Chaotic.
I Tends asymptotically to inﬁnity.
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 diﬀerential 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 diﬀerentiable, 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 2fold:
' 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 (תיראיניל הביצי).
^ 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 ﬁrst task is generally diﬃcult, since we do not have a general way of solving
linear equations with nonconstant coeﬃcients. The second task is in fact not
always possible; one can ﬁnd situations in which a solution is linearly stable but
nonlinearly unstable.
The ﬁrst task is however easy if u(t} is a constant solution, i.e., if u is a ﬁxed 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 ﬁxed point u of (4.1) is linearly stable if and only if (i) all the
eigenvalues of Df (u} have nonpositive real parts, and (ii) if an eigenvalue has a
zero realpart then it is of multiplicity one.
While we already know this is true, we will also prove it using a new technique.
Deﬁnition 4.2 A ﬁxed point u is called hyperbolic if none of the eigenvalues of
Df (u} has real part equal to zero. It is called a sink (עלוב) if all the eigenvalues
have negative real parts; it is called a source (רוקמ) all the eigenvalues have
positive real parts; it is called a center (זכרמ) if all the eigenvalues are imaginary;
it is called a saddle (כוא) if certain eigenvalues have positive real parts while
other have negative real parts.
Example: Consider the unforced Duﬃng oscillator,
y
′
1
= y
2
y
′
2
= y
1
− y
3
1
− δy
2
,
where δ ≥ 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 ﬁxed points,
(0, 0} and (±1, 0}.
Stability of solutions 91
Let u = (u
1
, u
2
} be the ﬁxed point. The linearized equation is
y
′
1
= y
2
y
′
2
= y
1
− 3u
2
1
y
1
− δy
2
,
i.e.,
Df (u} = _
0 1
1 − 3u
2
1
−δ
_.
The eigenvalues satisfy
λ(λ + δ} − (1 − 3u
2
1
} = 0,
which implies that
λ =
1
2
_−δ ±
_
δ
2
+ 4(1 − 3u
2
1
}_.
For the ﬁxed point (0, 0} there is a positive and a negative eigenvalue, which im
plies that it is unstable. For the ﬁxed point (±1, 0} both eigenvalues have negative
real parts, which implies that these ﬁxed points are stable. ▲▲▲
4.3 Lyapunov functions
Theorem 4.2 Let f ∶ R
n
→ R
n
be twice diﬀerentiable; let u be a ﬁxed 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 deﬁne z(t} =
y(t} − u. Then,
z
′
(t} = f (z + u} ≡ F(z}.
Zero is a ﬁxed point of F, which is stable if and only if u is a stable ﬁxed point of
f ; furthermore,
DF(0} = Df (u}.
92 Chapter 4
Proof : The idea of the proof is based on a socalled Lyapunov function. Suppose
we construct a function V ∶ R
n
→R that satisﬁes the following conditions:
' V(y} ≥ 0 for all y.
^ V(y} = 0 iﬀ 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 ﬁxed point, the vector ﬁeld 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¡ ≤ C¡y¡
2
.
Deﬁne 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
α > 0 such that
(y, Df (0}y} < −α¡y¡
2
,
i.e.,
g
′
(t} < −2αg(t} + R(t}.
Moreover, if y(t} ∈ U, then
¡R(t}¡ ≤ ¡y(t}¡ ⋅ C¡y(t}¡
2
≤ C
1
g
32
(t}.
Stability of solutions 93
In particular, there exists a constant k > 0 such that K
c
⊂ U, and as long as
y(t} ∈ U
′
g
′
(t} < −αg(t}.
If follows that if g(0} < c then g(t} < c for all t > 0.
In addition,
d
dt
(e
αt
g(t}} < 0,
and further,
e
αt
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 β > 0, such that
∀x ∈ R
n
(Ay, y} < −β¡y¡
2
.
Example: An important use of Lyapunov functions is in cases where linear stabil
ity cannot conclusively determine the (nonlinear) stability of a ﬁxed point. Con
sider the following system,
y
′
1
= y
2
y
′
2
= −y
1
− εy
2
1
y
2
.
The point (0, 0} is a ﬁxed 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
− εy
2
1
y
2
_ = −εy
2
1
y
2
2
94 Chapter 4
The origin is therefore stable for ε > 0.
▲▲▲
Example: Consider the system
y
′
1
= −3y
2
y
′
2
= y
1
+ α(2y
3
2
− y
2
},
where α > 0. The origin is a ﬁxed 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
+ α(2y
3
2
− y
2
}¡ = −α(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 iﬀ 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 δ > 0. That is,
∀t > 0 δ ≤ V(y(t}} ≤ V(y
0
}.
Stability of solutions 95
The set
C = ¦y ∶ δ ≤ V(y} ≤ V(y
0
}}
is closed and bounded and hence compact. Thus,
(DV(y}, f (y}}
must attain its maximum on C,
max
y∈C
(DV(y}, f (y}} ≡ −α < 0.
Then,
d
dt
V(y(t}} = (DV(y(t}}, f (y(t}} ≤ −α,
and therefore
V(y(t}} ≤ V(y
0
} − αt,
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 satisﬁes 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
Deﬁnition 4.3 Let f ∶ R
n
→ R
n
be an autonomous vector ﬁeld. This vector ﬁeld
induces a ﬂow map (המירז תקתעה),
Φ
t
∶ R
n
→R
n
via the trajectories of the solutions, namely
y
1
= Φ
t
(y
0
},
if y
1
is the solution at time t of the diﬀerential system,
y
′
= f (y}, y(0} = y
0
.
Another way to write it is
d
dt
Φ
t
(y} = f (Φ
t
(y}} Φ
0
(y} = y.
Deﬁnition 4.4 Let f ∶ R
n
→ R
n
be an autonomous vector ﬁeld. A set S ⊂ R
n
is
said to be invariant if
Φ
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 (תיבויח תיטנאירווניא) if
Φ
t
(y} ∈ S for all y ∈ S and for all t > 0.
It is said to be negatively invariant (תילילש תיטנאירווניא) if
Φ
t
(y} ∈ S for all y ∈ S and for all t < 0.
Example: Any trajectory,
O(y} = ¦Φ
t
(y} ∶ t ∈ R}
is an invariant set. Any union of trajectories,
⋃
y∈A
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} = ¦Φ
t
(y} ∶ t ≥ 0}
is a positivelyinvariant set, and any backward trajectory,
O
+
(y} = ¦Φ
t
(y} ∶ t ≤ 0}
is a negativelyinvariant set. ▲▲▲
Example: A ﬁxed point is an invariant set (no matter if it is stable or not). ▲▲▲
Deﬁnition 4.5 An invariant set S is said to be an invariant manifold (העירי
תיטנאירווניא) 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 ﬁxed point. The behavior of the solution near the ﬁxed 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 ﬂow map of the linearized system is given by the linear map
Φ
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
= Span¦e
1
, . . . , e
s
}
1
This is of course a nonrigorous deﬁnition; for a rigorous deﬁnition you need to take a course
in diﬀerential geometry.
98 Chapter 4
is the span of (generalized) eigenvectors corresponding to eigenvalues with nega
tive real part,
E
u
= Span¦e
s+1
, . . . , e
s+u
}
is the span of (generalized) eigenvectors corresponding to eigenvalues with posi
tive real part, and
E
c
= Span¦e
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→∞
Φ
t
(y} = 0 ∀y ∈ E
s
.
The characterization of the unstable manifold E
u
, is that
lim
t→−∞
Φ
t
(y} = 0 ∀y ∈ E
u
.
Example: Let n = 3 and suppose that A has three distinct real eigenvalues,
λ
1
, λ
2
< 0 and λ
3
> 0.
Let e
1
, e
2
, e
3
be the corresponding normalized eigenvectors. Then,
A =
⎛
⎜
⎝
⋮ ⋮ ⋮
e
1
e
2
e
3
⋮ ⋮ ⋮
⎞
⎟
⎠
⎛
⎜
⎝
λ
1
0 0
0 λ
2
0
0 0 λ
3
⎞
⎟
⎠
⎛
⎜
⎝
⋯ e
1
⋯
⋯ e
2
⋯
⋯ e
3
⋯
⎞
⎟
⎠
−1
≡ TΛT
−1
.
As we know,
y(t} = e
At
y
0
= T
⎛
⎜
⎝
e
λ
1
t
0 0
0 e
λ
2
t
0
0 0 e
λ
3
t
⎞
⎟
⎠
T
−1
y
0
.
In this case
E
s
= Span¦e
1
, e
2
} and E
u
= Span¦e
3
}
are invariant manifolds. The geometry of the trajectories is depicted in the follow
ing ﬁgure:
Stability of solutions 99
E
s!
E
u!
▲▲▲
Example: Suppose now that
λ
1,2
= ρ ± iω and λ
3
> 0,
where ρ < 0. Then, the corresponding eigenvectors are complex, however
A(e
1
+ e
2
} = (ρ + iω}e
1
+ (ρ − iω}e
2
= ρ(e
1
+ e
2
} + iω(e
1
− e
2
}
A(e
1
− e
2
} = (ρ + iω}e
1
− (ρ − iω}e
2
= ρ(e
1
− e
2
} + iω(e
1
+ e
2
},
or,
A
e
1
+ e
2
√
2
= ρ
e
1
+ e
2
√
2
− ω
e
1
− e
2
√
2i
A
e
1
− e
2
√
2i
= ρ
e
1
− e
2
√
2i
+ ω
e
1
+ e
2
√
2
.
Thus, there exists a transformation of variables A = TΛT
−1
, in which
Λ =
⎛
⎜
⎝
ρ −ω 0
ω ρ 0
0 0 λ
3
⎞
⎟
⎠
.
The solution to this system is
y(t} = T
⎛
⎜
⎝
e
ρt
cos ωt −e
ρt
sinωt 0
e
ρt
sinωt e
ρt
cos ωt 0
0 0 e
λ
3
t
⎞
⎟
⎠
T
−1
y
0
.
100 Chapter 4
Here again there is a twodimensional stable subspace and a onedimensional un
stable subspace.
SHOW FIGURE. ▲▲▲
Example: Consider now the case where λ
1
< 0 is an eigenvalue of multiplicity
two and λ
3
> 0. Then, A = TΛT
−1
with
Λ =
⎛
⎜
⎝
λ
1
1 0
0 λ
1
0
0 0 λ
3
⎞
⎟
⎠
,
and the solution is
y(t} = T
⎛
⎜
⎝
e
λ
1
t
te
λ
1
t
0
0 e
λ
1
t
0
0 0 e
λ
3
t
⎞
⎟
⎠
T
−1
y
0
.
In this transformed system E
s
= Span¦e
1
, e
2
} and E
u
= Span¦e
3
}. This is a deﬁ
cient system, hence has only one direction “entering” the ﬁxed point.
SHOW FIGURE.
▲▲▲
Exercise 4.4 Consider the following diﬀerential systems:
' y
′
= _
λ 0
0 µ
_y λ < 0, µ > 0.
^ y
′
= _
λ 0
0 µ
_y λ < 0, µ < 0.
· y
′
= _
λ −ω
ω λ
_y λ < 0, ω > 0.
´ y
′
= _
0 0
0 λ
_y λ < 0.
I y
′
= _
0 λ
0 0
_y λ > 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:
'
⎧
⎪
⎪
⎨
⎪
⎪
⎩
y
′
1
= y
2
y
′
2
= −δy
2
− µy
1
− y
2
1
.
^
⎧
⎪
⎪
⎨
⎪
⎪
⎩
y
′
1
= −y
1
+ y
3
1
y
′
2
= y
1
+ y
2
.
Find all the ﬁxed 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 aﬀect the properties of the solutions in the vicinity of the ﬁxed
point?
Theorem 4.4 Suppose that the vector ﬁeld f is ktimes diﬀerentiable and let the
origin be a ﬁxed 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→∞
Φ
t
(y} = 0,
and
∀y ∈ W
u
lim
t→−∞
Φ
t
(y} = 0,
Proof : Not in this course. B
Example: Consider again the Duﬃng oscillator,
x
′
= y
y
′
= x − x
3
− δy.
102 Chapter 4
The linearized vector ﬁeld is
y
′
= _
0 1
1 −δ
_y,
with eigenvalues
λ
1,2
=
1
2
[±
√
δ
2
+ 4 − δ).
The corresponding (linear) stable and unstable manifolds are
y = λ
1,2
x.
The origin is a saddle, i.e., it has a onedimensional stable manifold and a one
dimensional unstable manifold.
W
u#
W
s#
▲▲▲
Example: Consider the following artiﬁcial system,
x
′
= x
y
′
= −y + x
2
.
The origin is a ﬁxed point and the linearized system is
_
x
y
_
′
= _
1 0
0 −1
__
x
y
_,
i.e., E
s
= Span¦e
2
} and E
u
= Span¦e
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
diﬀerential 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 onedimensional 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 yaxis, x = 0.
W
u#
W
s#
▲▲▲
4.5 Periodic solutions
Deﬁnition 4.6 A solution through a point y is said to be periodic with period
T > 0 if
Φ
t
(y} = y.
We will now concentrate on periodic solutions of autonomous systems in the
plane.
104 Chapter 4
Theorem 4.5 (Bendixon’s 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 Γ that encloses a domain S ⊂ D. By
Green’s theorem,
∬
S
div f dy =
Γ
( f , n} d = 0,
where n is the unit normal to Γ, and ( f , n} = 0 because f is tangent to Γ. If, as
assumed div f is neither identically zero nor changes sign, then this is impossible.
B
Example: For the Duﬃng oscillator,
div f = −δ,
hence this system cannot have any closed orbit. ▲▲▲
4.6 Index theory
In this section we are concerned with autonomous diﬀerential systems in the
plane:
x
′
= f (x, y}
y
′
= g(x, y}.
Let Γ be a closed curve in the plane. If the curve does not intersect a ﬁxed point,
then at any point along the curve the vector ﬁeld makes an angle φ with the x axis,
φ(x, y} = tan
−1
g(x, y}
f (x, y}
.
Stability of solutions 105
When we complete a cycle along Γ the angle φ must have changed by a multiple
of 2π. We deﬁne the index of the curve by
i(Γ} =
1
2π
γ
dφ.
It turns out that the index of a curve satisﬁes 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 ﬁxed 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 ﬁxed point is zero.
I The index of a curve is equal to the sum of the indexes of the ﬁxed points
that it encloses.
Sink% Source%
Saddle%
Periodic%
Corollary 4.1 Every periodic trajectory encloses at least one ﬁxed point. If it
encloses a single ﬁxed point then it is either a sink, a source, or a center. If all
the enclosed ﬁxed 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 Duﬃng oscilla
tor are as follows:
▲▲▲
4.7 Asymptotic limits
We are back to autonomous systems in R
n
; we denote by Φ
t
∶ R
n
→ R
n
the ﬂow
map. The trajectory through a point y is the set:
O(y} = ¦Φ
t
(y} ∶ t ∈ R}
We also have the positive and negative trajectories:
O
+
(y} = ¦Φ
t
(y} ∶ t ≥ 0}
O
−
(y} = ¦Φ
t
(y} ∶ t ≤ 0}.
Deﬁnition 4.7 Let y ∈ R
n
. A point ˜ y belongs to the ωlimit of y,
˜ y ∈ ω(y},
if there exists a sequence t
n
→∞ such that
lim
n→∞
Φ
tn
(y} = ˜ y.
Stability of solutions 107
It belongs to the αlimit y,
˜ y ∈ α(y},
if there exists a sequence t
n
→−∞ such that
lim
n→∞
Φ
tn
(y} = ˜ y.
Thus the ωset (הגמוא תצובק) ω(y} denotes the collection of all “future partial
limits” of trajectories that start at y, and α(y} denotes the collection of all “past
partial limits” of trajectories that pass through y.
Example: Let u be a ﬁxed point and let y belong to the stable manifold of u. Then,
ω(y} = ¦u}.
If z belongs to the unstable manifold of u, then
α(z} = ¦u}.
▲▲▲
Example: Let y ∶ R →R
n
be a periodic solution. Then for every t
0
,
ω(y(t
0
}} = α(y(t
0
}} = O(y(t
0
}}.
▲▲▲
4.8 The Poincar´ eBendixon theorem
We consider again twodimensional systems,
y
′
= f (y}.
As before we denote the ﬂow function by Φ
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
' ω(p} ≠ ∅.
^ ω(p} is closed.
· ω(p} is invariant, i.e., it is a union of trajectories.
´ ω(p} is connected.
Proof :
' Take any sequence t
n
→ ∞. Since Φ
tn
(p} ∈ M and M is compact this
sequence has a converging subsequence,
lim
k→∞
Φ
tn
k
(p} = q ∈ M.
By deﬁnition q ∈ ω(p}, which is therefore nonempty.
^ Let q ] ∈ ω(p}. By the deﬁnition of the ωset, there exists a neighborhood U
of q and a time T > 0 such that
Φ
t
(p} ] ∈ U ∀t > T.
Since U is open, every point z ∈ U has a neighborhood V ⊂ U, and
Φ
t
(p} ] ∈ V ∀t > T,
which implies that z ] ∈ ω(p}, i.e.,
(ω(p}}
c
is open,
hence ω(p} is closed.
· Let q ∈ ω(p}. By deﬁnition, there exists a sequence t
n
→∞, suchthat
lim
n→∞
Φ
tn
(p} = q,
Let τ ∈ R be arbitrary. Then,
lim
n→∞
Φ
tn+τ
(p} = lim
n→∞
Φ
τ
(Φ
tn
(p}} = Φ
τ
[lim
n→∞
Φ
tn
(p}) = Φ
τ
(q},
which proves that Φ
τ
(q} ∈ ω(p} for all τ ∈ R, and hence ω(p} is invariant.
Stability of solutions 109
´ Suppose by contradiction that ω(p} was not connected. Then there exist
disjoint open sets V
1
, V
2
such that
ω(p} ⊂ V
1
∪ V
2
ω(p} ∩ V
1
≠ ∅ and ω(p} ∩ V
2
≠ ∅.
It follows that the trajectory O(p} must pass from V
1
to V
2
and back in
ﬁnitely often. Consider a compact set K that encloses V
1
and not intersect
V
2
. The trajectory must pass inﬁnitely often in the compact set K∖V
1
, hence
must have there an accumulation point, in contradiction to ω(p} being con
tained in V
1
∪ V
2
.
V
1#
V
2#
K#
B
Deﬁnition 4.8 Let Σ be a continuous and connected arc in R
2
(i.e., a map I →
R
2
). It is called transversal to the ﬂow ﬁeld f neither vanishes on Σ nor is tangent
to Σ, namely,
(Σ
′
(t}, f (Σ(t}} ≠ ¡Σ
′
(t}¡ ¡ f (Σ(t}¡.
∑"
f"
110 Chapter 4
Lemma 4.1 Let M ⊂ R
2
be a positively invariant set and let Σ be a transversal
arc in M. Then for every p ∈ M, the positive trajectory O
+
(p} intersects Σ
monotonically, that is, if p
i
= Σ(s
i
} is the ith intersection of O
+
(p} and Σ, then
the sequence s
i
is monotonic.
Comment: The lemma does not say that O
+
(p} must intersect Σ at all, or that it
has to intersect it more than once.
Proof : Proof by sketch:
p
i#
p
i+1#
Posi)vely#invariant#set#
B
Corollary 4.2 Let M⊂ R
2
be a positively invariant set and let Σ be a transversal
arc in M. Then for every p ∈ M, ω(p} intersects Σ at at most one point,
#Σ ∩ ω(p} ≤ 1.
Proof : The existence of two intersection points would violate the previous lemma,
for suppose that
p
1
, p
2
∈ Σ ∩ ω(p}.
Stability of solutions 111
Then, O
+
(p} would intersect Σ at inﬁnitely many points in disjoint neighborhoods
of p
1
and p
2
.
Note that this point is somewhat subtle. If p
1
∈ Σ ∩ ω(p} we are guaranteed that
the trajectory will visit inﬁnitely often any neighborhood of p
1
, but does it mean
that it has to intersect Σ each time? Yes. Take any neighborhood U of p
1
and
consider
U ∩ Σ,
which is an open neighborhood of p
1
in Σ. By transversality,
¦Φ
t
(U ∩ Σ} ∶ −ε < t < e}
is an open neighborhood of p
1
in R
2
. Any trajectory visiting this neighborhood
intersects U ∩ Σ. B
Proposition 4.2 Let Mbe a positively invariant set and let p ∈ M. If ω(p} does
not contain any ﬁxed point, then ω(p} is a periodic trajectory.
Proof : Let q ∈ ω(p} and let x ∈ ω(q} (we know that both are nonempty sets).
Since O(q} ⊂ ω(p} (an ωlimit set is invariant) and ω(p} is closed, then
ω(q} ⊂ ω(p}.
It follows that ω(q} does not contain ﬁxed points, i.e., x is not a ﬁxed point.
Since x is not a ﬁxed point, we can trace through it a transversal arc Σ. The
trajectory O
+
(q} will intersect Σ monotonically at points q
n
→ x, but since q
n
∈
ω(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 ω(p}. Let Σ 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 ω(p}, for if every neighborhood of O(q} contained points in
r ∈ ω(p}, O(r} ⊂ ω(p}, and there would be such an r for which O(r} intersects Σ.
Since ω(p} is connected, it follows that ω(p} = O(q}.
112 Chapter 4
p
"
q
"
A"zone"in"which"there""
are"no"point"in"w(p)"
B
Deﬁnition 4.9 A heteroclinic orbit is a trajectory that connects two ﬁxed points,
i.e., a trajectory O(y}, such that
α(O(y}} = ¦p
1
} and ω(O(y}} = ¦p
2
},
where p
1
and p
2
are ﬁxed 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 ﬁxed points
and
p
1
, p
2
∈ ω(p}
for some p ∈ M. Since ω(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, γ ⊂ ω(p}, such that
¦p
1
} = α(γ} and ¦p
2
} = ω(γ}.
Proof : Proof by sketch:
Stability of solutions 113
p
1#
p
2#
p
#
B
Proposition 4.4 (Poincar´ eBendixon) Let M be a positively invariant set that
contains a ﬁnite number of ﬁxed points. Let p ∈ M. Then, one and only one of the
following occurs:
' ω(p} is a single ﬁxed point.
^ ω(p} is a limit cycle.
· ω(p} is a union of ﬁxed points p
i
and heteroclinic orbits γ
j
that connect
them.
Proof : If ω(p} contains only ﬁxed points then it only contains one due to con
nectedness.
If ω(p} does not contain any ﬁxed point then it is a limit cycle (we proved it).
Remains the case where ω(p} contains ﬁxed points and points that are not ﬁxed
points; let q ∈ ω(p} and consider the set ω(q}. We will show that it contains a
single ﬁxed point, hence q is on a trajectory that converges asymptotically to a
ﬁxed point in ω(p}. Similarly, α(q} contains a single ﬁxed point, so that q lies, as
claimed, on a trajectory that connects between two ﬁxed points in ω(p}.
114 Chapter 4
So let x ∈ ω(q}. Suppose that x was not a ﬁxed point, then we would trace a
transversal arc Σ through r. The trajectory O(q} must intersect this arc inﬁnitely
often, but since O(q} ∈ ω(p}, it must intersect Σ at a single point, which implies
that O(q} is a periodic orbit, but this is impossible because ω(p} is connected. It
follows that ω(q} contains only ﬁxed points, but then we know that it can only
contain a single ﬁxed point.
B
2
Contents
1 Existence and uniqueness 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Euler’s approximation scheme . . . . . . . . . . . . . . . . . . . . . 3 3 9
The CauchyPeano existence proof . . . . . . . . . . . . . . . . . . . 11 Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 The PicardLindl¨ f existence proof . . . . . . . . . . . . . . . . . . 20 o Continuation of solutions . . . . . . . . . . . . . . . . . . . . . . . . 22 Generalized solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Continuity of solutions with respect to initial conditions . . . . . . 23 25
Linear equations 2.1 2.2
Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 Linear homogeneous systems . . . . . . . . . . . . . . . . . . . . . . 26 2.2.1 2.2.2 2.2.3 2.2.4 The space of solutions . . . . . . . . . . . . . . . . . . . . . 26 Fundamental matrices . . . . . . . . . . . . . . . . . . . . . 30 The solution operator . . . . . . . . . . . . . . . . . . . . . . 35 Order reduction . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3 2.4
Nonhomogeneous linear systems . . . . . . . . . . . . . . . . . . . 37 Systems with constant coeﬃcients . . . . . . . . . . . . . . . . . . . 40 2.4.1 2.4.2 2.4.3 The Jordan canonical form . . . . . . . . . . . . . . . . . . . 40 The exponential of a matrix . . . . . . . . . . . . . . . . . . 40 Linear system with constant coeﬃcients . . . . . . . . . . . 43
ii 2.5
CONTENTS Linear diﬀerential equations of order n . . . . . . . . . . . . . . . . 45 2.5.1 2.5.2 2.5.3 2.5.4 2.6 2.6.1 2.6.2 2.6.3 The Wronskian . . . . . . . . . . . . . . . . . . . . . . . . . 45 The adjoint equation . . . . . . . . . . . . . . . . . . . . . . 48 Nonhomogeneous equation . . . . . . . . . . . . . . . . . . 50 Constant coeﬃcients . . . . . . . . . . . . . . . . . . . . . . 52 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 General theory . . . . . . . . . . . . . . . . . . . . . . . . . . 56 Hill’s equation . . . . . . . . . . . . . . . . . . . . . . . . . . 59 63
Linear systems with periodic coeﬃcients: Floquet theory . . . . . . 55
3
Boundary value problems 3.1 3.2
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 Selfadjoint eigenvalue problems on a ﬁnite interval . . . . . . . . . 67 3.2.1 3.2.2 3.2.3 Deﬁnitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 Properties of the eigenvalues . . . . . . . . . . . . . . . . . 69 Nonhomogeneous boundary value problem . . . . . . . . . 71
3.3 3.4 4
The existence of eigenvalues . . . . . . . . . . . . . . . . . . . . . . 76 Boundary value problems and complete orthonormal systems . . . 83 87
Stability of solutions 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8
Deﬁnitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 Lyapunov functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 Invariant manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 Periodic solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 Index theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 Asymptotic limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 The Poincar´ Bendixon theorem . . . . . . . . . . . . . . . . . . . . 107 e
CONTENTS Bibliography
1
1. Theory of ordinary diﬀerential equations by E.A. Coddington and N. Levinson. 2. Ordinary diﬀerential equations by V.I. Arnold. 3. Solving ordinary diﬀerential equations by E. Hairer, N.P Nørsett and G. Wanner.
2
CONTENTS
1 A diﬀerential equation is an equation that relates a function to its derivative(s). A diﬀerential equation is said to be ordinary ( ) if the function is univariate.Chapter 1 Existence and uniqueness 1. We abbreviate ordinary diﬀerential equation into ode.1 Introduction Deﬁnition 1. A number of questions arise right away: Does the equation have a solution? If it does. The unknown is the function. or in a diﬀerent notation. Example: Find a function y R → R that satisﬁes the equation y′ = 6y. ∀t ∈ R y′ (t) = 6 y(t). and more precisely if its domain is a connected subset of R. is the solution unique? What is the solution? Is there a systematic way to solve such an equation? .
direct substitution shows that any function of the form y(t) = a e6t .4 Chapter 1 In the above example. Thus. t) = 0. like the value of the function at a certain point. . 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. suppose we wanted to denote the independent variable by x. Example: The following equation is a ﬁrstorder ode: y′ (t) = t sin 3 + y2 (t). (It is sloppy because it is not clear in this notation that t is a point in the domain of y. a solution exists and it is not unique.e.) Firstorder ode The most general ﬁrstorder ode is of the form Φ(y′ (t). say. y(t). as we often do.2 An ordinary diﬀerential equation is said to be of order k if the highest derivative of the unknown function that appears in the equation is the kth derivative. the unique solution that belongs to the above exponential family of solutions is found by setting y(2) = a e12 = 8. We will always assume that the equation has been brought into explicit form ( ): y′ (t) = f (t. y(t)). We know from elementary calculus that knowing the derivative of a function does not uniquely determine the function. i. a = 8 e−12 . y(2) = 8. Still. which we will often write in the more sloppy form: y′ = t sin 3 + y2 . a ∈ R is a solution of the equation. This is not surprising. In this case. Such an equation is said to be implicit ( ). 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. Deﬁnition 1..
y R → Rn . Example: Find functions y. . . and in explicit form y′′ (t) = f (t. . At this point it is clear that there won’t be a unique solution unless we specify some extra information about the function at certain points. Systems of diﬀerential equations In the same way as we go from univariate algebraic equations into multivariate ones. y(t)). As in the ﬁrst example.Existence and uniqueness Secondorder ode The most general secondorder ode is of the form Φ(t. y′′ (t). yn (t). In components. More generally. . . yi (t) = fi (y1 (t). . that satisﬁes an equation of the form y′ (t) = f (t. where f R × Rn → Rn . such that y′ (t) = 6tz z′ (t) = z3 y′ + 4. so we can extend the idea of a diﬀerential equation into a system of diﬀerential equations in which more than one function is unknown. z. . . The question is how much extra information is needed in order to obtain an equation that has a unique solution (assuming that such exists). y′ (t). n. y(t)) = 0. 5 Example: Consider the following equation: y′′ (t) + y(t)y′ (t) + y(t) sin t = 0. y(t)). we ask ourselves whether this equation has a solution and whether it is unique. t). i = 1. y′ (t). in a ﬁrstorder system of n (explicit) odes the unknown is a vectorvalued function. .
What is ordinary in an ode? A diﬀerential equation is called ordinary when the unknown function depends on a single realvalued variable. . . Yn ). . Y1 . Suppose we have an equation of the form y(n) (t) = f (t. t). and will not be considered in this course (we will eventually explain why there is such a fundamental diﬀerence between the two). . . Example: An example of a pde is the heat equation (also known as the diﬀusion equation): the unknown function depends on two variables (position and time). We then deﬁne the following vectorvalued function. . ∂2 y ∂y (x. yn−1 (t). and satisﬁes the equation. y R × R → R. . ∂t ∂x The theory of pdes is inﬁnitely harder than the theory of odes. This is in contrast to partial diﬀerential equations ( ).6 Chapter 1 Highorder equations and ﬁrstorder systems Every nth order (scalar) equation can be turned into a ﬁrstorder system of n odes by a simple procedure. y(t)). which we can rewrite in vector form. Y= . in which the unknown function depends on more than one variable. ′ Y1 = Y2 ′ Y2 = Y3 ′ Yn−1 = Yn ′ Yn = f (t. . Y ′ (t) = F(t. t) = 2 (x. Y(t)). y′ (t). y y′ y(n−1) It satisﬁes the following system of equations. .
odes are important not only in physics. Classical mechanics is “one big system of odes”. By Newton’s laws. The particle is attached to a spring that is connected to the origin. while developing diﬀerential calculus: y′ (t) = 1 − 3t + y(t) + t2 + ty(t) y(0) = 0. ecology. if the instantaneous position of all the particles in the world at time t can be viewed as a huge vector y(t). the most general solution to this equation is y(t) = a cos k mt + b sin k m t. y(0) and y′ (0). y′′ (t) = f (t. Newton’s approach to solve this equation was based on a method that we call today an asymptotic expansion. economics. then according to classical mechanics. That is. y′ (t). . The oldest ode The ﬁrst documented instance of an ode is an equation studied by Newton (1671). Example: Consider a single “point particle” that can only move along a single axis.Existence and uniqueness 7 Why do we care about odes? Diﬀerential and integral calculus were developed in the 17th century motivated by the need to describe the laws of nature. Any (deterministic ) process that evolves continuously can be described by an ode. y(t) = a1 t + a2 t2 + a3 t3 + . demography. such a system may have a unique solution if we specify both y (positions) and y′ (velocities) at an initial time (the big bang?). say. They can be determined by specifying. where a. He looked for a solution that can be expressed as a power series ( ). the second derivative of this vector (the acceleration) is a godgiven function of the current positions and velocities of all the particles. Newton’s law is my′′ (t) = −k y(t). As we will see. y(t)). the evolution of the universe is determined forever from its initial conditions. odes are used generically to describe evolving systems in chemistry. and much more. b are two “integration constants”. 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.
He tried to solve the following diﬀerential equation. then we may write a2 − y2 ∆y = −∆t. Then summing over many small steps ∆t and ∆y. as more terms as added. a1 = 1. for every ﬁxed number of terms the approximation deteriorates for large t. a4 = −1 6. the derivative was roughly. ξ This integral can actually be calculated analytically (the primitive function is known) and so Leibniz could get an implicit representation of the solution.8 Substituting into the equation. This yields. a3 = 1 3. derivatives were not deﬁned as rigorously as you learned it (it took 200 more years until calculus took the form you know). however. A problem studied by Leibniz In 1674 Leibniz (Newton’s competitor on the development of calculus) studied a geometrical problem that had been studied earlier by Fermat. y′ (t) = − y(t) a2 − y2 (t) . For Leibniz. If we consider ∆y and ∆t as ﬁnite numbers. . y Suppose that y(0) = y0 . so that y(t) = t − t2 + t3 t4 − + 3 6 For ﬁxed t. He then proceeded to identify equal powers of t. y y0 a2 − ξ2 dξ = −t. Chapter 1 a1 + 2a2 t + 3a3 t2 + 4a4 t3 = 1 − 3t + (a1 t + a2 t2 + a3 t3 ) + t2 + t(a1 t + a2 t2 + a3 t3 ). Remember that at that time. a2 = −1. y′ (t) ≈ ∆y =− ∆t y(t) a2 − y2 (t) . the closer we are to the true solution. etc. a1 = 1 2a2 = −3 + a1 3a3 = a2 + 1 + a1 4a4 = a3 + a2 .
As we saw. This is an integral equation (y(t) is still unknown). and the data are prescribed as some initial time. y′ (t) = f (t. y(t)). The origin of the name is situations where the independent variable is space (and then often denoted by x). Integral formulation Consider a ﬁrst order system. which shares a lot in common with the diﬀerential equation. The origin of this name is realizations where t is time. y(t0 ) = y0 . y(s)) ds. The theories of initial and boundary value problems are very diﬀerent. We will later see in which sense the diﬀerential and integral systems are equivalent. it will fail to be unique unless additional conditions on y are prescribed. even if such an equation has a solution. While the distinction between initial and boundary value problems may seem immaterial it is fundamental. 1. Consider an initial value problem. thus. y(s)) Integrating both sides from t0 to t. In general. until further notice we will assume that we deal with ﬁrst order systems. and in what sense they diﬀer. f (s. If the data are distributed between two points of more then we call the resulting problem a boundary value problem. . a system of n ﬁrstorder equations requires n additional data on y. y′ (s) = f (s. y(t) = y0 + t t0 y(t0 ) = y0 .2 Euler’s approximation scheme y′ (t) = f (t. y(t)). The ﬁrst part of this course will be devoted to initial value problems.Existence and uniqueness 9 Initial and boundary value problems Recall that any nth order equation can be represented as a ﬁrstorder system. and y(x) is some function of space that is prescribed at the boundary of the region of interest. If all the data are prescribed at a single point we call the resulting problem an initial value problem.
1 Learn to solve a large class of equations. j = 0. and suppose that we want to ﬁnd the function in some interval [t0 . . < tn = T. Euler’s method (usually variants of it) is used until today to numerically approximate the solution of ordinary diﬀerential systems. . y j ). In 1820 Cauchy proved that Euler’s approximation does indeed converge. (i) Obtain an explicit expression for the approximate solution. as the partition is reﬁned. the ﬁrstorder diﬀerential equation is approximated by a onestep diﬀerence equation ( ). In 1768 Euler proposed a method to approximate the solution. Exercise 1. (ii) What is the true solution? Does the approximate solution converge to the true solution as n → ∞? TA material 1. then we will get arbitrarily close to the true solution (whose existence was not questioned). . . y j )(t j+1 − t j ).1 Approximate the solution to the equation y′ (t) = 6y(t) y(0) = 1 at the point t = 1 using Euler’s scheme for a uniform partition t j+1 − t j = 1 n.10 Chapter 1 where y and f are vectorvalued. t j+1 − t j or equivalently. T ] (we will often refer to t as “time” as a suggestive interpretation). n − 1. Euler’s belief was that if we reﬁne the partition suﬃciently. and in fact Cauchy’s proof is the ﬁrst proof that the diﬀerential system has a solution and that this solution is unique. to a solution of the diﬀerential equation. y j+1 = y j + f (t j . He considered partitions of the interval: t0 < t1 < t2 < and approximated y(t j ) by y j given by y j+1 − y j = f (t j . Thus. .
ϕ(t)) ≤ ε. b M). Let (t0 . R0 = [t0 . (t. and replace R by another rectangle.3 Let D ⊂ Rn × R be a domain in which f is continuous. A function ϕ I → Rn is called an εapproximation to the diﬀerential system in D if: Deﬁnition 1. y0 ) ∈ D and consider a rectangle of the form R = [t0 . Rn can be endowed with a norm. ( )) open connected interval. For every t ∈ I S . (2 hrs. we will assume that f is continuous in its two arguments in some open and connected domain D ⊂ R × Rn . Let I be an ϕ is continuous. and without loss of generality we will use the Euclidean norm.y)∈R Then. t0 + α] × {y 1 y − y0 ≤ b} ⊂ D. ≤ x ≤ C x 1. all norms on Rn are equivalent1 . which we denote by ⋅ . ϕ′ (t) − f (t. (t. ϕ(t)) ∈ D. Since f is continuous and R is compact. For every t ∈ I.Existence and uniqueness 11 1. C > 0 such that (∀x) c x 1 2 . it follows that f is bounded in R. y) . we deﬁne α = min(a. ϕ is continuously diﬀerentiable. deﬁne M = max f (t. Two norms ⋅ 1 and ⋅ 2 are said to be equivalent if there exist constants c. t0 + a] × {y y − y0 ≤ b} ⊂ D. Throughout. we must have some a priori assumptions about the function f . As we know from advanced calculus. where ϕ′ has onesided limits.3 The CauchyPeano existence proof In order to be able to analyze the diﬀerential system. except perhaps at a ﬁnite set of points S .
. has norm less than M. which we denote by ϕ(t). z) < ε. over an interval of length α.12 Chapter 1 y0+b# y0# y0'b# t0# t0+α# t0+a# Theorem 1. j We then construct the Euler approximation to this solution. tn = t0 + α. Comment: A similar theorem can be formulated for an interval on the left of t0 . Thus. given ε > 0 there exists a δ > 0 such that if t−s <δ and y−z <δ then f (t. This means that all Euler’s polygons remain in R0 on the entire interval [t0 . and connect the point by straight segments. It is a piecewise continuously diﬀerentiable function. which is equal to some f (t j . ). y(t j )). As long as the polygon is in R0 its derivative. . t0 + α] and partition it. Proof : Since f is continuous on R0 . δ M).1 There exists for every ε > 0 an εapproximation in R0 to the equation y′ (t) = f (t. t0 + α] satisfying the initial condition y(t0 ) = y0 . . t0 < t1 < . there is no diﬀerence between “past” and “future”. so that Euler’s approximation is a polygon. y) − f (s. Take the interval [t0 . it is uniformly continuous ( Thus. t0 +α]. such that max(t j+1 − t j ) < min(δ. ϕ(t) can diﬀer from y0 (in norm) by at most αM ≤ b. y(t)) on the interval [t0 . Mathematically. no mater how the partition of the segment is done.
ϕ(t j )) − f (t. (3 hrs. namely. Then. which proves that ϕ is an εapproximation to our system. Then there exists a sequence fn ∈ F that converges uniformly on I. By our choice of δ. g(t1 ) − g(t2 ) < ε. ( )) . Because the derivative of ϕ is bounded (in norm) by M. ϕ(t)) = f (t j . f ∈F t∈I and equicontinuous. (Equicontinuity is uniform continuity with the same δ valid for all g ∈ F . 13 I Deﬁnition 1. sup max f (t) < ∞. t − t j < δ. ϕ(t j )). ϕ(t)) < ε. ϕ′ (t) = f (t j .Existence and uniqueness Let t be a point in the jth interval.2 (ArzelaAscoli) Let F be an inﬁnite family of functions I → Rn that are uniformly bounded. ϕ(t)) . ϕ′ (t) − f (t. and further. ϕ(t) − ϕ(t j ) ≤ M t − t j ≤ M(δ M) = δ. By our choice of the partition. such that for every t1 − t2 < δ and every g ∈ F .) Theorem 1. f (t j . ϕ(t j )) − f (t.4 A family F of functions I → Rn is said to be equicontinuous ( ) if there exists for every ε > 0 a δ > 0.
n > N) ( Fn (r) − Fm (r) ) < ε. K) ( Fn (rk ) − Fm (rk ) ) < ε Then for every t ∈ Ik . the sequence Fn satisﬁes Cauchy’s criterion for uniform convergence. equicontinuity convergence equicontinuity Thus. Fix ε > 0. n > N. Since the family of functions F is equicontinuous. Consider the set { f (r1 ) f ∈ F } ⊂ Rn . n > N)(∀k = 1. . Fn (t) − Fm (t) ≤ Fn (t) − Fn (rk ) + Fn (rk ) − Fm (rk ) + Fm (rk ) − Fm (t) ≤ 3ε. This set is bounded. (k) (k) (2) (2) (2) (1) ∈ ( fk ) for which such that f j (ri ) (k) (1) Consider now the diagonal sequence Fk = fk . such that each subinterval if shorter than δ. Since there is a ﬁnite number of such rk s. It converges pointwise on all the rationals in I. I = I1 ∪ I2 ∪ ⋅ ⋅ ⋅ ∪ IK . I With this. Consider then the set { f j (r2 ) j ∈ N}. (∃N ∈ N) (∀m. Note that since we don’t know what the limit is. . We proceed inductively. . we use instead the Cauchy criterion for convergence. Why? Because for every the diagonal sequence is eventually a ( ) subsequence of f j . .14 Chapter 1 Proof : Let (rk ) be a sequence that contains all the rational numbers in I. this set is bounded. t1 − t2 < δ)(∀ f ∈ F ) ( f (t1 ) − f (t2 ) < ε). and m. and for every k derive a sequence f j converges for every i ≤ j. t2 . we can prove Cauchy’s theorem: . This means that (∀r ∈ I ∩ Q)(∀ε > 0)(∃N ∈ N) (∀m. Since F is uniformly bounded. which concludes the proof. (∃δ > 0) (∀t1 . therefore there exists a subsequence f j f j (r2 ) converges (and also f j (r1 ) converges). and therefore has an accumu(1) (1) lation point: there exists a sequence f j ∈ F for which f j (r1 ) converges. In every subinterval Ik select a rational number rk . Partition the interval I into subintervals.
3 (CauchyPeano) Let f (t. We have no a priori knowledge on how long this interval is. The sequence (ϕk ) is uniformly bounded on I because (∀t ∈ I)(∀k ∈ N) ( ϕk (t) − y0 < b). It remains to show that y is a solution to the diﬀerential equation. Then the diﬀerential equation y′ (t) = f (t. It only guarantees the existence of a solution on some interval. t0 + α] satisfying the initial condition y(t0 ) = y0 . and (∀ t1 − t2 < δ)(∀k ∈ N) ( ϕk (t2 ) − ϕk (t1 ) ≤ ε). This theorem is known as a local existence theorem. and satisﬁes the initial conditions. and therefore. At this stage not a word about uniqueness. t0 + α] to the diﬀerential equation. Proof : Let (εk ) be a sequence of positive numbers that tends to zero. given ε > 0 we set δ = ε M.Existence and uniqueness 15 Theorem 1. denote the limit by y. Comments: Note that our only conditions on f is that it be continuous. t2 ∈ I)(∀k ∈ N) ( ϕk (t2 ) − ϕk (t1 ) ≤ M t2 − t1 ). It is also equicontinuous because (∀t1 . . It follows from the ArzelaAscoli theorem that (ϕk ) has a subsequence (which we do not relabel) that converges uniformly on I. y) and the rectangle R0 be deﬁned as above. We proved that for every k there exists a function ϕk I → Rn that is an εk approximation on I = [t0 . y(t)) has a solution on the interval [t0 .
Show that the equation y′ (t) = f (t.2 Let D ⊂ R × Rn be an open domain. and its derivative satisﬁes the fundamental theorem of calculus. f (yk ) − f (xk ) < ε. 2 y(0) = 0. yk ) that satisﬁes ∑k (yk − xk ) < δ. y0 ) ∈ D. k Using the fact that ϕk is an εk approximation to the diﬀerential equation.4 Uniqueness Example: Consider the initial value problem. we can diﬀerentiate this equation and obtain y′ (t) = f (t. ϕk (s))] ds. we get that y(t) = y0 + t t0 f (s. t t0 [ϕ′ (s) − f (s. Since y ∈ C 1 (I). y(t)). y(t)). . The functions ϕk are absolutely continuous ( ). k A function is absolutely continuous if and only if it is almost everywhere diﬀerentiable. and (t0 . y′ (t) = y1 3 (t). k Letting k → ∞ and using the fact that the convergence of ϕk to y is uniform. k The fact that ϕ′ is not deﬁned at a ﬁnite number of points is immaterial. y(s)) ds. which means that there exists for every ε > 0 a δ > 0 such that for every ﬁnite sequence of disjoint intervals (xk . y(t0 ) = y0 has a solution on some open interval I that contains t0 . I Exercise 1. ϕk (s))] ds ≤ εk t − t0 .16 Chapter 1 For each k ∈ N it follows from the fundamental theorem of calculus that2 ϕk (t) = y0 + = y0 + t t0 t t0 ϕ′ (s) ds k f (s. 1. ϕk (s)) ds + t t0 [ϕ′ (s) − f (s. f ∈ C(D).
ϕ1 (t0 ) − ϕ2 (t0 ) < δ.5 2 2. 2t 3 2 . does not guarantee the uniqueness of the solution.5 1 1.4 Another solution is 0 2(t−c) 3 3 2 0≤t≤c t > c.2 0 0 0.Existence and uniqueness 17 By the CauchyPeano theorem there exists an α > 0 such that this equation has a solution in the interval [0. It turns out that in order to guarantee uniqueness we need more stringent requirements on the function f . Proposition 1.6 1. y(t)) contained in D on the interval I. 1. however.8 0. Let ϕ1 and ϕ2 be ε1 . there are inﬁnitely many solutions of the form y(t) = for every c > 0. . y(t) = 3 In fact.1 Let f R × Rn → Rn by continuous in t and Lipschitz continuous in y with constant L in the domain D.and ε2 approximations to the equation y′ (t) = f (t. 1.6 0. The CauchyPeano theorem.2 1 0.4 0. and moreover for a certain t0 ∈ I. A trivial solution to this equation is y(t) = 0.5 3 (What would have happened had we approximated the solution with Euler’s method?) The above example shows that existence of solutions does not necessarily go hand in hand with uniqueness. α].
ϕ2 (s))] ds ≤ (ε1 +ε2 )(t−t0 ). For i = 1. for all t ∈ I. ϕ2 (s)) ≤ ε2 . y) − f (t. ϕ1 (s)) ≤ ε1 1 ′ ϕ2 (s) − f (s. ϕ1 (s)) − f (s. 2. without loss of generality that t > t0 . By the deﬁnition of εapproximations. for every t0 ≤ s ≤ t. ϕ1 (t) − ϕ2 (t) < δeL t−t0 + ε1 + ε2 L t−t0 e −1 . L Comment: By Lipschitz continuity in y we mean that for every (t. ϕ′ (s) − f (s. i f (s. Proof : Suppose. f (s. ϕi (s)) ds + t t t0 [ϕ′ (s) − f (s. y) and (t. z) ≤ L y − z . ϕi (t) = ϕi (t0 ) + hence ϕi (t) − ϕi (t0 ) − t0 t t0 f (s. f (t. . a − b ≤ a + b . [ϕ1 (t) − ϕ2 (t)] − [ϕ1 (t0 ) − ϕ2 (t0 )] − It further follows that ϕ1 (t) − ϕ2 (t) ≤ ϕ1 (t0 ) − ϕ2 (t0 ) + t t0 t t0 [ f (s. z) in D. 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. ϕi (s))] ds. ϕ1 (s)) − f (s.18 Chapter 1 Then. ϕ2 (s)) ds + (ε1 + ε2 )(t − t0 ). Using the triangle inequality. ϕi (s)) ds ≤ εi (t − t0 ). and (ii) a deviation due to the fact that both functions are only approximate solutions to the same equation.
Next. Proof : This is an immediate corollary of the previous proposition as δ = ε1 = ε2 = 0. Our goal is to get an explicit bound for r(t). which is a diﬀerential inequality. Then ϕ1 = ϕ2 . y(t0 ) = y0 . and it only remains to integrate the right hand side explicitly3 and substitute back into (1.2 Various Gronwall inequalities.1) This is an integral inequality. we may integrate this inequality from t0 to t and get eL(t−t0 ) R(t) − R(0) ≤ 0 t t0 ′ eL(s−t0 ) (δ + (ε1 + ε2 )(s − t0 )) ds. Then. R(t0 ) = 0.1). Since integration preserves order. y(t)). using the Lipschitz continuity of f and the bound on the initial deviation. I 3 t 0 t 1 se−Ls ds = − e−Lt − 2 e−Lt − 1 .Existence and uniqueness 19 Deﬁne now r(t) = ϕ1 (t) − ϕ2 (t) . We proceed as follows: deﬁne R(t) = so that t r(s) ds. L L . where f is continuous in t and Lipschitz continuous in y. for every t0 ≤ s ≤ t: eL(s−t0 ) R(s) = eL(s−t0 ) [R′ (s) − LR(s)] ≤ eL(s−t0 ) (δ + (ε1 + ε2 )(s − t0 )) . Theorem 1.4 (Uniqueness) Let ϕ1 and ϕ2 be two solutions of the initial value problem y′ (t) = f (t. I TA material 1. (1. t0 R′ (t) ≤ δ + LR(t) + (ε1 + ε2 )(t − t0 ). r(t) ≤ δ + L t t0 r(s) ds + (ε1 + ε2 )(t − t0 ).
and it is of sufﬁcient interest to justify a second proof. Deﬁne the function y(0) (t) ≡ y0 . y(s)) ds. and further deﬁne y(1) (t) = y0 + t t0 f (s. contradicting the fact that τ is the supremum. y(t) = y0 + t t0 f (s. ∆(n) = y(n+1) − y(n) . We then deﬁne inductively for every n ∈ N: y(n+1) (t) = y0 + t t0 f (s. Recall that a solution to the initial value problem y′ (t) = f (t. ϕ(t) − y(t) < L (5 hrs. Deﬁne. y(n+1) (t)) remains in R0 for all t ∈ I. y(t0 ) = y0 is also a solution to the integral equation.5 The PicardLindl¨ f existence proof o There is another standard way to prove the existence of solutions.20 Chapter 1 Comment: We have obtained also an estimate for the error of Cauchy’s approximation. For what values of t is it deﬁned? For the same [t0 . The diﬀerence with the proof based on Euler’s method is that we need to assume the Lipschitz continuity of f in R0 . . It y is the solution and ϕ is an εapproximation with exact initial data. y(t)). and show as above that (t. We are now going to show that the sequence y(n) converges uniformly to a solution to the initial value problem. y(1) (t)) remains in R0 as follows: let τ = sup{t ∈ I If τ < t0 + α then y(1) (t) − y0 ≤ b}. then ε L t−t0 e −1 . t0 + α] as in the previous sections. y(n) (s)) ds. y(0) (s)) ds. ( )) 1. We can show that (t. y(1) (τ) − y0 ≤ M(t − t0 ) < b.
∆(n+1) (t) = and ∆(n+1) (t) ≤ L with ∆(1) (t) ≤ t0 t t0 t0 21 t f (s. I purposely did it here “by hand” to show how straightforward the proof is. Note also that we can estimate the error of the successive approximations. We have y(n) − y(m) = n−1 ∆(k) . L n! n−1 k=0 Now we write the nth function as follows: y(n) = y(n) − y(n−1) + y(n−1) − y(n−2) + ⋅ ⋅ ⋅ + y0 = y0 + ∆(k) .Existence and uniqueness then. we have an integral inequality. All that remains is to let n → ∞ in the equation y(n+1) (t) = y0 + to obtain y(t) = y0 + t0 t t0 t f (s. ∆(n) (t) ≤ t t0 t t0 ∆(1) (s) ds ≤ LM L2 M 2 t t0 t t0 (s − t0 ) ds ≤ LM t − t0 2 . y(n−1) (s)) ds. 2 L2 M t − t0 3 . y0 ) ds ≤ M t − t0 . This method is known as the method of successive approximations. Comment: Usually one puts this proof in the context of the Picard’s contractive mapping theorem. 3! ∆(2) (s) ds ≤ (s − t0 )2 ds ≤ M (L t − t0 )n . y(n) (s)) ds. f (s. y(n) (s)) − f (s. k=m . f (s. By the Weierstaß Mtest this series converges uniformly to a limit y. Once again. but we may proceed diﬀerently: ∆(2) (t) ≤ L ∆(3) (t) ≤ L and generally. t ∆(n) (s) ds. y(s)) ds.
y) = y2 ). This solution however diverges as t → 1. τ]}. 1−t is a solution on the interval [0. indicating that it can’t be further continued. t→a y(0) = 1. = Lm! ∆(k) (t) ≤ n−1 n−1 The right hand side does not depend on n. y) ∈ R × Rn (as is f (t. Suppose that f (t. y) is continuous in t and Lipschitz continuous in y for all (t.22 hence y(n) (t) − y(m) (t) ≤ Chapter 1 Lk−1 M (t − t0 )k k! k=m k=m ∞ ∞ Lm+k−1 M Lk−1 M (t − t0 )k = (t − t0 )m+k ≤ k! (m + k)! k=0 k=m ∞ Lm+k−1 M Lm−1 M(t − t0 )m ∞ Lk ≤ (t − t0 )m+k = (t − t0 )k m!k! m! k! k=0 k=0 m M[L(t − t0 )] L(t−t0 ) e . . It is easy to check that 1 . y(t) = This situation is typical to equations that do not have global solutions. so we can let n → ∞ to get an estimate for y − y(m) . 1. 1). and we know therefore that it is unique. and deﬁne a = sup{τ > t0 there exists a solution on [t0 . If a is ﬁnite and lim y(t) = b < ∞. t0 + α]. 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) = y2 (t).6 Continuation of solutions The existence and uniqueness proof shows that there exists an α such that a unique solution exists in the interval [t0 .
ξ) = ξ. ∂t ϕ(τ. we can viewed the solution of y′ (t) = f (t.Existence and uniqueness then consider the initial value problem y′ (t) = f (t. τ. τ. as if it were. This observation has an important implication: whenever you want to show that a certain equation has global solutions. ( )) 1. for example. but it can also be viewed as a function of the initial data. (6 hrs.τ. τ. violating the deﬁnition of a as a supremum. y(t)). y(a) = b. τ. We know that ϕ is diﬀerentiable in t. This is an important question. if a is ﬁnite then lim y(t) = ∞.8 Generalized solutions Continuity of solutions with respect to initial conditions The solution of an ivp is of course a function of “time” t. In this . namely. ξ). discontinuous with respect to the initial data. but to all ivps together. as it satisﬁes the equations. ∂ϕ (t.2) The question is how regular it is with respect to the initial data τ and ξ.e. y(t)). the only way a solution can fail to exist indeﬁnitely is if it diverges. i. you can try to show that as long as a solution exists it is bounded uniformly.. y(τ) = ξ as a function of t.7 1. t→a namely. for now we denote this solution as ϕ(t. This is a diﬀerent point of view in which we have a function that represents the solution not to a single ivp. a unique solution exists on [t0 . (1. 23 By the existence and uniqueness theorems there exists a γ such that a unique solution exists on [a. ξ)). and ξ. Thus. a + γ]. it would not be an appropriate model for any real life problem. a + γ]. ϕ(t. ξ) = f (t.
Theorem 1. uniqueness. and continuity on initial data (and possibly other parameters in the equation). ξ) τ ∈ I. b] and let ψ I → Rn be a solution to the diﬀerential equation. τ. such that {(t. .24 Chapter 1 context. Moreover. ξ) = ξ. Let D ⊂ R × Rn be a domain in which f is continuous in t and Lipschitz in y. ξ − ψ(τ) < δ} There exists a solution ϕ to (1. one speaks of the wellposedness of an equation.2) on I with ϕ(τ.5 There exists a δ > 0 such that for every (τ. ϕ is continuous on I × Uδ . ψ(t)) t ∈ I} ⊂ D. Let I = [a. ξ) ∈ Uδ = {(τ. which includes existence.
Ax ≤ A x .1 Preliminaries Recall that for every normed space (X . We will deal with the vector space Rn endowed with the Euclidean norm. X ) with an operator norm. This norm has the following (additional) properties: 1.Chapter 2 Linear equations 2. . ⋅ ) we can endow the vector space of linear operators L(X . Exercise 2. For every x ∈ X and A ∈ L(X . X ). A = sup x≠0 Ax . For every A. 2. 3. X ). One can ask then what is the explicit form of the corresponding operator norm for matrices A Rn×n . AB ≤ A B . B ∈ L(X .1 Find the explicit form of the operator norm corresponding to the Euclidean norm on Rn . It is always the case that id = 1. x It can be shown that this is indeed a norm.
. This is done as follows. i. which are realvalued functions. y(t) = ξ + t τ A(s)y(s) ds. .1) a jk (t)yk (t). y) = A(t)y is continuous in t and Lipschitz in y.2. y(τ) = ξ.26 Chapter 2 2. y′ (t) = A(t)y(t) is called a linear homogeneous system.1 Linear homogeneous systems The space of solutions Let I ⊂ R be a closed interval and let A I → Rn×n be a continuous nbyn matrixvalued function of I. there exists a unique function y I → Rn such that y′ (t) = A(t)y(t). hence a local solution exists for some interval containing τ.e. in fact this remains true also if I = R. We ﬁrst write the corresponding integral equation. that it cannot diverge. Example: A linear homogeneous system for n = 2: y1 0 1 (t) = 1 y2 −1 + 2 sin t 0 ′ y1 (t). We denote its entries. Comment: The main claim is that a (unique) solution exists on the whole of I. In component notation. by ai j .1 For every τ ∈ I and ξ ∈ Rn . The ﬁrstorder system of diﬀerential equations.2 2. y′j (t) = n k=1 (2. y2 Proposition 2. We need to show that the solution can be continued on the entire of I. Proof : The function f (t. which holds as long as a solution exists.
ξ e−M(t−τ) − 1 . which means that there exists a uniform bound A(s) ≤ M. e−M(t−τ) Y(t) ≤ and further ξ e M(t−τ) − 1 . Deﬁning Y(t) = ∫τ y(s) ds we get Y ′ (t) ≤ ξ + MY(t). Y(τ) = 0. t This is an integral inequality that we can solve. −M Integrating from τ to t. This solution is called the trivial solution. so is A . y(t) ≤ ξ + t τ 27 A(s)y(s) ds ≤ ξ + t τ A(s) y(s) ds. ∀t ∈ I. I The trivial solution One characteristic of linear homogeneous systems is that y ≡ 0 is always a solution. M Substituting back into the inequality for y(t) . which proves that a solution to the ivp exists on I. Y(t) ≤ y(t) ≤ ξ + M ξ e M(t−τ) − 1 = ξ e M(t−τ) .2 Let X be the space of functions y I → Rn satisfying the linear . Proposition 2. Then.1). e−M(t−τ) [Y ′ (t) − MY(t)] ≤ ξ e−M(t−τ) . homogeneous system (2. Because A is continuous on I. Then X is a vector space over the complex ﬁeld C with respect to pointwise addition and scalar multiplication.Linear equations Taking norms. M This can never diverge in I. This implies that y(t) ≤ ξ + M t τ y(s) ds.
. not all of them zero. z ∈ X . . such that n+1 k=1 ck ϕk = 0. . ϕn ∈ X . . we denote the kth component of ξ j by ξk j . then dim X = n. Moreover. such that n k=1 ck ϕk = 0 implies that ck = 0 for all k. i. and select n independent vectors ξ1 . the ivp ϕ′j (t) = A(t)ϕ j (t). β ∈ C.. but every (n + 1) vectors are linearly dependent.3 Let X be deﬁned as above. Recall that the dimension of a vector space is deﬁned as the maximal number n for which there exists a set of n linearly independent vectors. Proof : Take τ ∈ I. .28 Proof : Let y. Any linear combination of solutions is also a solution. For each ξ j . for every set of n + 1 function ϕk ∈ X there exist scalars ck . . y′ (t) = A(t)y(t) For every α. . that is there exists a set of functions ϕ1 . I Comment: The elements of X are functions I → Rn . Chapter 2 and z′ (t) = A(t)z(t). ξn ∈ Rn .e. Proposition 2. (αy + βz)′ (t) = αA(t)y(t) + βA(t)z(t) = A(t)(αy + βz)(t). which proves that αy + βz ∈ X . ϕ j (τ) = ξ j . Comment: This is what physicists call the principle of superposition. .
for if they were linearly dependent.Linear equations 29 has a unique solution. we would have ξn = but then ψ= n−1 k=1 n−1 k=1 ck ξk . Set ξ = ϕ(τ). suppose that n k=1 ck ϕk ≡ 0. but because the ξk are independent. and by uniqueness ψ = ϕn . . which contradicts the linear independence of the ϕk . n k=1 ck ξk = 0. . n k=1 n k=1 ck ξk . The ξn are linearly independent. ϕ= which proves that the ϕk form a basis. . Let ϕ1 . Take any τ ∈ I and let ξ j = ϕ(τ). ϕn be linear independent solutions. . I . ck ϕk is a solution to the diﬀerential system satisfying ψ(τ) = ξn . ck ϕk . Indeed. Suppose now that ϕ is another solution of the diﬀerential system. the ck are all zero. there exist coeﬃcients ck such that ξ= By the same argument as above. Because the ξ j span Rn . then in particular for t = τ. We claim that these solution are independent.
it has to satisfy the matrix equation (2.2) whose columns are linearly independent.2) and its columns have to be independent for some τ ∈ I. ϕi j is the ith component of the jth basis function. This matrix satisﬁes the matrixvalued diﬀerential equation. Φ(t) = ϕ11 ϕn1 (t). y = Φc. which in components reads ϕ′ j (t) = aik (t)ϕk j (t). any solution y I → Rn can be represented as yi = n j=1 c j ϕi j . Φ′ (t) = A(t)Φ(t). that is. and in vector form. In particular.1). .1). .4 For a matrixvalued function to be a fundamental matrix of (2. i Φ(t) is called a fundamental matrix ( ) for the linear homogeneous system (2.1). Because the columns of a fundamental matrix span the space of solutions to (2. Proposition 2.2 Fundamental matrices I → Rn×n be the matrixvalued function ϕ1n ϕnn Let now ϕ1 . . where the c j ’s are constant. .30 Chapter 2 2. ϕn be a basis.1) any solution X I → Rn×n of the matrix equation X ′ (t) = A(t)X(t). .2. Let Φ whose columns are the vectors ϕ j . we call a fundamental matrix for (2. (2. if this holds then its columns are independent for all t ∈ I.
ϕ′ 11 ϕn1 ∑ j a1 j ϕ j1 ϕn1 ϕ′ 1n ϕnn + + = ∑ j a1 j ϕ jn ϕnn ϕ11 + ∑ j an j ϕ j1 ϕ1n ∑ j an j ϕ′jn .5 Let Φ I → Rn×n be a fundamental matrix. ( )) Here is another proof that the columns of the matrix remain independent. hence from basic linear algebra. . Then det Φ(t) = det Φ(t0 ) exp t t0 Tr A(s) ds . Fix a τ ∈ I.1) can be represented as y = Φc. This is a linear equation for c. In particular. I (8 hrs. Since the same considerations hold for every t ∈ I.Linear equations 31 Proof : Let Φ I → Rn×n be a fundamental matrix. we know that it is a solution of (2. det Φ(t) ≠ 0 for all t ∈ I. Proof : Since Φ= we diﬀerentiate to obtain Φ = ′ ϕ11 ϕn1 ϕ11 ϕ′ n1 + ϕ1n ϕnn ϕ1n ϕ′ nn .2) and that any solution y I → Rn of (2. if det Φ(t0 ) ≠ 0 then det Φ(t) ≠ 0 for all t ∈ I. we know that it has a unique solution (because the columns of Φ span the space of solutions). The vector c is determined by the equation y(τ) = Φ(τ)c. but it comes with a quantitative estimate: Proposition 2. det Φ(τ) ≠ 0.
a13 times the third line. . 0 0 This matrix is singular. What does it imply? That this matrixvalued function cannot be a fundamental matrix of a linear homogeneous system. if Ψ is a fundamental matrix then Ψ = ΦC for some nonsingular complex matrix C. I Example: Consider the matrix Φ(t) = t t2 . Because it is nonsingular for all t. We do the same with the second determinant and remains with a22 det Φ. Then ΦC is also a fundamental matrix. Φ′ = (Tr A) Φ. and 0 = (ΨΨ−1 )′ = Ψ(Ψ−1 )′ + Ψ′ Ψ−1 .1 Let Φ be a fundamental matrix for (2. We subtract from the ﬁrst line a12 times the second line. 0 Theorem 2. and so on. Proof : By deﬁnition Φ′ = AΦ and we know that det Φ(t) ≠ 0 for all t. Then for every nonsingular complex matrix C.1) and let C ∈ Cn×n be a non singular complex (constant) matrix. Ψ−1 exists. and it remains to integrate it. etc. Let Ψ be a fundamental matrix. This is a scalar linear ode. Thus. (ΦC)′ = AΦC and det Φ(t)C = det Φ(t) det C ≠ 0 for all t ∈ I. there exists no matrix A(t) such that t 0 ′ t = A(t) 0 and t2 0 ′ = A(t) t2 . Then we remain with a11 det Φ.32 Chapter 2 Consider the ﬁrst determinant. Moreover. but its columns are linearly independent.
e. Conversely.1) then Ψ is a fundamental matrix for the adjoint system if and only if Ψ∗ Φ = C. This equation is known as the equation adjoint ( ) to (2. (Ψ∗ )′ = −Ψ∗ A. and since Φ is nonsingular (Ψ∗ )′ Φ = −Ψ∗ AΦ.. where C is a nonsingular constant matrix. (Ψ−1 Φ)′ = (Ψ−1 )′ Φ + Ψ−1 Φ′ = −Ψ−1 AΦ + Ψ−1 AΦ = 0.1). and it remains to take the Hermitian adjoint of both sides. if Ψ is a fundamental solution of the adjoint equation then (Ψ∗ Φ)′ = 0. (9 hrs. 33 It follows that Ψ−1 Φ = C. ( )) Proof : Let Ψ satisfy the above equation then (Ψ∗ )′ Φ + Ψ∗ Φ′ = 0. I Comment: The matrix (Ψ∗ )−1 satisﬁes the equation.Linear equations from which we get that (Ψ−1 )′ = −Ψ−1 Ψ′ Ψ−1 = −Ψ−1 AΨΨ−1 = −Ψ−1 A.6 If Φ is a fundamental matrix for (2. [(Ψ∗ )−1 ]′ = −A∗ (Ψ∗ )−1 . I . which concludes the proof. or Φ = ΨC. i. Now. Proposition 2.
In particular.. The matrix Φ∗ (t)Φ(t) determines the magnitudes of the columns of Φ(t) and the angles between pairs of vectors.. − sin t 0 1 y(t).1 Let A be antihermitian. −1 0 A fundamental matrix is therefore Φ(t) = It follows that. i. and Φ∗ (t)Φ(t) = C. then the linear homogeneous system and its adjoint coincide. ϕ j ) = αi α jCi j . Φ∗ (t)Φ(t) = sin t cos t cos t − sin t sin t cos t 1 0 = . j=1 n j=1 α j ϕ j (t). cos t − sin t 0 1 sin t cos t . Example: Consider the linear system y′ (t) = Two independent solutions are ϕ1 (t) = sin t cos t and ϕ2 (t) = cos t . n i. if Φ is a fundamental matrix for (2. cos t − sin t . j=1 = αi α j (ϕi . i. since any solution is of the form y(t) = then y(t) 2 n i.e.e. A(t) = −A∗ (t). the norm of any solution is independent of t.1) then it is also a fundamental matrix for its adjoint.34 Chapter 2 Corollary 2. In particular.
then the solution is of the form y(t) = Φ(t)c.. and the linear transformation between the solution at two diﬀerent times is the matrix R(t. ≡R(t. We call the matrix R(t. but it doesn’t. c = Φ−1 (t0 )y(t0 ). cos t − sin t = α2 sin t cos t + β2 cos t − sin t sin t + 2αβ sin t cos t cos t cos t = α2 + β2 . t0 ). and then Ψ(t)Ψ−1 (t0 ) = Φ(t)CC −1 Φ−1 (t0 ) = Φ(t)Φ−1 (t0 ). y(t0 ) = Φ(t0 )c. . y(t0 ) = given. i. It doesn’t because the solution is unique.3 The solution operator y′ (t) = A(t)y(t). The vector c is found by substituting the initial data. Consider the initial value problem.2. and the solution is therefore.Linear equations Any solution of this linear system is of the form y(t) = α Now. We can also see it as any other fundamental matrix Ψ(t) can be represented as Ψ(t) = Φ(t)C. t0 ) the solution operator ( ) between times t0 and t. y(t) = Φ(t)Φ−1 (t0 ) y(t0 ).e. If Φ(t) is a fundamental matrix. The solution operator seems to depend on the choice of the fundamental matrix. − sin t cos t − sin t 2. y(t) 2 35 sin t cos t +β .t0 ) Comments: It follows that the solution at time t depends linearly on the solution at time t0 .
t0 ) is nonsingular and [R(t. t0 )]−1 = R(t0 . t1 )R(t1 . Example: Consider the diﬀerential system in some interval that does not include the origin: y′ (t) = A solution to this system is ϕ1 (t) = 1 t . t0 ).2. t0 )y(t0 ). t0 )y(t0 ). Chapter 2 R(t2 .36 For every t0 . ϕ m . but also. This gives an explicit solution for any initial value problem. t) = I for all t. Why? Because for every y(t0 ). . . and suppose that we have m linearly independent solutions ϕ1 . y(t2 ) = R(t2 . with m < n. t). 3 2t2 −1 2t . 2. t1 . −1 t2 0 1 y(t). t0 ) = R(t2 . The catch is that in most cases we will not be able to ﬁnd n independent solutions. we need to ﬁnd n independent solutions. R(t. t2 . . y(t2 ) = R(t2 . R(t. . Let us recapitulate: given a homogeneous linear system. t1 )R(t1 .4 Order reduction Consider an ndimensional linear homogeneous system. Then we can construct a fundamental matrix and the associated solution operator. t1 )y(t1 ) = R(t2 .
−1 t2 All we need in order to be able to solve an initial value problem is to ﬁnd another independent solution. a scalar equation). t0 ) = Φ(t)Φ−1 (t0 ). Consider now a system of the form where A I → Rn×n and b I → Rn are continuous. then the solution for the nonhomogeneous system is y(t) = R(t.2 Let Φ(t) be a fundamental matrix for the homogenous system. t0 R′ (t. s)b(s) ds + R(t. s)b(s) ds. It turns out that if a fundamental matrix for the corresponding homogeneous system is known. 2. Proof : Note that the solution operator also satisﬁes the homogeneous system. t0 )y(t0 ) + A(t) = A(t)y(t) + b(t). R′ (t. t0 ) = A(t)R(t. s)b(s) ds + b(t) . then we can solve the nonhomogeneous system in explicitly form.Linear equations as ϕ′ (t) = 1 37 −1 t2 0 1 = 3 2 −1 2t 2 t 3 2t 1 t . Theorem 2. t)b(t) t t0 = A(t)R(t. t0 )y(t0 ) + where as above R(t. y′ (t) = R′ (t. t0 )y(t0 ) + t t0 t R(t.3 Nonhomogeneous linear systems y′ (t) = A(t)y(t) + b(t). t0 ) Diﬀerentiating the above solution. R(t. Such a system is called a linear nonhomogeneous system. The knowledge of one solution enables us to get for the other solution an equation of lower order (in this case.
(y − yin )′ = A(t)(y − yin ).38 By uniqueness. ( )) Chapter 2 I Look at the structure of the solution. and substituting the initial data we get that C1 = Φ−1 (t0 )y(t0 ). It is a sum of the solution of the solution to the homogeneous problem and a term that “accumulates” the nonhomogeneous term. A(t)Φ(t) hence Φ(t)C ′ (t) = b(t) C(t) = t t0 or C ′ (t) = Φ−1 (t)b(t). The space of solutions is an aﬃne space Another point of practical interest is the following. which implies that any solution to the nonhomogeneous system can be expressed as yin (t) plus a solution to the homogeneous system. y(t) = t t0 Integrating we ﬁnd a solution to the nonhomogeneous equation. Φ(t)Φ−1 (s)b(s) ds.e. We then look for a solution where C is a function of time. Diﬀerentiating we get y′ (t) = Φ′ (t) C(t) + Φ(t)C ′ (t) = A(t)y(t) + b(t) = A(t)Φ(t)C(t) + b(t). .2 is the method of variation of constants. Suppose that yin (t) is a solution to the nonhomogeneous equation and let y(t) by any other solution. Then. this is the solution. The solution to the homogeneous system is of the form y(t) = Φ(t)C. y(t) = Φ(t)C(t). Φ−1 (s)b(s) ds i. The method of variation of constants Another way to derive Theorem 2. (10 hrs. This is not the most general equation: we can always add a solution to the homogeneous system: y(t) = Φ(t)C1 + t t0 Φ(t)Φ−1 (s)b(s) ds.
−1 0 sin ωt where ω is a frequency. It is a general approach for expressing the solution of a linear nonhomogeneous system in terms of the solution operator of the homogeneous system. cos t − sin t Substituting into Theorem 2. This system describes a unit mass attached to a spring with unit force constant and forced externally with frequency ω. sin ωs cos(t − s) 0 ds sin ωs Assume that t0 = 0. y(0) + 2 ω(cos t − cos ωt) − sin t cos t ω −1 The solution splits into a part that depends on the initial condition plus a part that depends on the forcing.Linear equations 39 Comment: The method shown here is a special case of something known as Duhammel’s principle.2 we get y(t) = = cos(t − t0 ) sin(t − t0 ) y(t0 ) + − sin(t − t0 ) cos(t − t0 ) cos(t − t0 ) sin(t − t0 ) y(t0 ) + − sin(t − t0 ) cos(t − t0 ) t t0 t t0 cos(t − s) sin(t − s) − sin(t − s) cos(t − s) sin ωs sin(t − s) ds. y′ (t) = 0 1 0 y(t) + . Then y(t) = 1 ω sin t − sin ωt cos t sin t . cos t0 − sin t0 − sin(t − t0 ) cos(t − t0 ) sin t cos t . The entries of y(t) are the displacement and the momentum. The expression for the nonhomogeneous term holds . A fundamental matrix for the homogeneous system is Φ(t) = Since Φ−1 (t) = then the solution operator is R(t. Example: Consider a forced linear oscillator. t0 ) = sin t cos t cos t − sin t sin t0 cos t0 cos(t − t0 ) sin(t − t0 ) = . cos t − sin t sin t cos t .
the solution diverges linearly in time. Recall that any matrix is similar to a matrix in Jordan form: that is. n=0 n! . a phenomenon known as resonance.40 Chapter 2 only of ω ≠ 1. it commutes with its adjoint. .2 The exponential of a matrix ∞ Let A be a (constant) nbyn matrix. there exists a nonsingular complex matrix P such that A = PJP−1 . then the “natural frequency” coincides with the ”forcing frequency. s. (Recall that a matrix is diagonalizable if it is normal.4 2.. .e.e. If ω = 1.) 2. and for i = 1. and J is of the form J0 J= where λ1 J0 = λ2 λq λq+i 1 0 λq+i ⋅ Ji = ⋅ 0 0 0 0 0 1 ⋅ 0 0 0 0 ⋅ 0 0 0 0 ⋅ .4. We deﬁne eA = An . y(0) + t sin t − sin t cos t 2 i.4. i. If A is diagonalizable then it only has the diagonal part J0 . 1 λq+i J1 Js . .1 Systems with constant coeﬃcients The Jordan canonical form Let A be a (constant) nbyn matrix.. . 2. If ω = 1 then we get y(t) = 1 sin t − t cos t cos t sin t .
n! n! n=0 n=0 ∞ which means that it is suﬃcient to learn to exponentiate matrices in Jordan canonical form. What can we say about the exponential of a matrix? In general. then eA = (PJP−1 )n ∞ PJ n P−1 = = Pe J P−1 . k>n and by Cauchy’s criterion the series converges absolutely. J= ⋅ ⋅ ⋅ ⋅ 0 0 0 0 1 0 0 0 0 λ The exponential of a diagonal matrix λ1 J0 = eλ1 λ2 λq is e J0 = eλ2 eλq . . eA+B ≠ eA eB . because. . e J0 eJ = e J1 e Js so that it is suﬃcient to learn to exponentiate diagonal matrices and matrices of the form λ 1 0 0 0 0 λ 1 0 0 ⋅ . unless A and B commute. ∞ A n Ak A k ∞ A k+n A k+n ≤ = ≤ ≤ e k! n! k≥n k! k=0 (k + n)! k=0 n! k! A n→∞ 41 → 0. where J is the Jordan canonical form. Note that for a blockdiagonal matrix J. If A = PJP−1 .Linear equations Is it welldeﬁned? Yes.
1 1 For later use. Suppose for example that ri = 4. 1 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 Ei2 = 0 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0 Ei3 = 0 0 0 0 0 0 0 0 0 0 1 0 0 0 Ei4 = 0. −t 0 .42 Chapter 2 The nondiagonal blocks have the following structure: Ji = λq+i Iri + Eri . The matrix Ei is nilpotent. 1 0 eEi = 0 0 1 1 2! 1 1 0 1 0 0 1 3! 1 2! 1 0 ⋅ 0 0 0 1 ⋅ 0 0 0 0 ⋅ 0 0 0 0 ⋅ . we note that if we multiply J j by a scalar t. where ri is the size of the ith block and 0 0 Ei = ⋅ 0 0 Because Iri and Ei commute then e Ji = eλq+i I eEi = eλq+i eEi . then 1 0 0 0 t t 2! 1 t 0 1 0 0 2 etJi = etλq+i t3 3! t2 2! t 1 Example: Let A= 0 t . then 0 0 Ei = 0 0 Then.
4. A3 = −t3 A A4 = t4 I. − sin t cos t 2. etA = cos t sin t . y′ (t) = Ay(t). − sin t cos t t3 t4 t5 t6 t2 I− A+ I+ A− I− 2! 3! 4! 5! 6! = cos t I + sin t A.e. 1 1 1 A= √ 2 i −i and then 1 etA = √ 2 1 1 = 2 i = = 1 1 i −i 1 −i eit 0 0 e−it 1 1 −i √ 2 1 i it 0 0 −it 1 1 −i √ 2 1 i . we diagonalize the matrix.Linear equations One way to calculate etA is to calculate powers of A: A2 = −t2 I hence etA = I + tA − i..3 Linear system with constant coeﬃcients We now consider homogeneous linear systems in which the coeﬃcients are constant. solving this system amounts to ﬁnding a fundamental matrix. As we know. . where A is a (constant) nbyn matrix. 43 Alternatively. eit −ieit e−it ie−it 1 eit + e−it −ieit + ie−it 2 ieit − ie−it eit + e−it cos t sin t .
dt It follows that the solution operator is R(t. which is what prevents us from proceeding. t0 ) = etA e−t0 A = e(t−t0 )A . one may wonder whether the exponential is only a fundamental solution for constant coeﬃcients. t Φ(t) = e∫0 A(s) ds ? Let’s try to diﬀerentiate: Φ(t + h) − Φ(t) e∫0 = h However. . These matrices commute either if A is constant.44 Chapter 2 Proposition 2. h A(s) ds t A(s) ds ≠ e∫0 A(s) ds e∫t t t+h . It follows that the solution of the inhomogenous equation y′ (t) = Ay(t)+b(t) is t e(t−s)A b(s) ds. y(t) = e(t−t0 )A y(t0 ) + t0 Inspired by the case of n = 1. lim ehA − I tA e(t+h)A − etA = lim e = A etA . Comments: Since A and etA commute it is also the case that d tA e = etA A. Isn’t the solution to y′ (t) = A(t)y(t). s. Proof : From the deﬁnition of the derivative. in general e∫0 t+h t+h A(s) ds − e∫0 A(s) ds . or if it is diagonal (in which case we have n independent equations). where we used the fact that etA and e sA commute for every t.7 The matrix etA is a fundamental matrix for the system with constant coeﬃcients A. h→0 h→0 h h I hence y(t) = etA is a fundamental solution.
.) A (scalar) diﬀerential equation of the form Ln y = 0 is called a homogeneous linear equation of order n. We deﬁne the linear diﬀerential operator. etA P = PetJ is also a fundamental solution. Y ′ (t) = A(t)Y(t). 2. . Ln C 1 (I) C 0 (I). an I → C be continuous complexvalued functions deﬁned on some interval. dt + an−1 f ′ + an f.1 Linear diﬀerential equations of order n The Wronskian Let a0 . (It will be assume that a0 (t) ≠ 0 for all t ∈ I. where λ is an eigenvalue of A and m is between zero is one minus the multiplicity of λ. a0 (t) a0 (t) a0 (t) a0 (t) As usual.5 2. Ln ( f ) = a0 f (n) + a1 f (n−1) + or in another notation. a1 . . . . It can also be written in the form y(n) (t) = − a1 (t) (n−1) a2 (t) (n−2) an−1 (t) ′ an (t) y (t) − y (t) − ⋅ ⋅ ⋅ − y (t) − y(t).5.Linear equations If A = PJP−1 then 45 etA = PetJ P−1 . and by the properties of fundamental matrices. Ln = a0 dn−1 dn + a1 n−1 + dtn dt + an−1 d + an . we can convert an nth order equation into a ﬁrstorder system of n equations. This means that the t dependence of all the components of the fundamental matrix is a linear combinations of terms of the form tm eλt .
. ϕn )(t) = det Φ(t). .46 where 0 0 ⋅ 0 1 0 ⋅ 0 Chapter 2 A= 0 1 ⋅ 0 0 0 ⋅ 0 0 0 ⋅ 1 −a1 a0 . . . as the matrix A has most of its elements zero. . . It is however a very special system. ϕn )(t0 ) exp − t Tr A(s) ds t0 t t0 a1 (s) ds . . ϕn . . −an a0 −an−1 a0 −an−2 a0 −an−3 a0 This is a homogeneous linear system. every solution of the linear system has the following structure: ϕ ϕ′ ϕ′′ . n solutions ϕ j (t) are linearly independent if the Wronskian (at any point t) is not zero. a fundamental matrix is determined by n independent scalar functions ϕ j : ϕ1 ϕ2 ϕn ′ ′ ϕ1 ϕ2 ϕ′ n . . . . . .5 W(ϕ1 . . ϕn )(t0 ) exp = W(ϕ1 . In fact. . . . hence we know a lot about the structure of its solutions. a0 (s) As we know. Φ= ⋅ ⋅ ⋅ (n−1) (n−1) (n−1) ϕ1 ϕ2 ϕn The determinant of this matrix is called the Wronskian of the system Ln y = 0 with respect to the solutions ϕ1 . By Proposition 2. We denote it as follows: W(ϕ1 . ϕ(n−1) In particular. . . . ϕn )(t) = W(ϕ1 . .
Linear equations The fact that for any n linearly independent solutions ϕ j , W(ϕ1 , . . . , ϕn )(t) = c exp − is known as Abel’s theorem.
t
47
a1 (s) ds a0 (s)
TA material 2.1 Use Abel’s 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 nth order linear operator Ln on I and n solutions ϕ j I → R satisfying Ln ϕ j = 0 for which W(ϕ1 , . . . , ϕn )(t) ≠ 0, then every function y I → R satisfying Ln y = 0 is a linear combination of the ϕ j ’s. It turns out that the opposite is also true. To every n linearly independent functions corresponds a linear diﬀerential operator of order n for which they form a basis:
Theorem 2.3 Let ϕ1 , . . . , ϕn be ntimes continuously diﬀerentiable functions on an
interval I such that ∀t ∈ I W(ϕ1 , . . . , ϕn )(t) ≠ 0. Then there exists a unique linear diﬀerential operator of order n (assuming that the coeﬃcient of the nth derivative is one) for which these functions form a basis.
Proof : We start by showing the existence of such an equation. Consider the equation W(y, ϕ1 , . . . , ϕn )(t) = 0. (−1)n W(ϕ1 , . . . , ϕn )(t) This is an nth order linear diﬀerential equation for y. Also the coeﬃcient of y(n) is 1. Clearly, W(ϕ j , ϕ1 , . . . , ϕn )(t) = 0 for all j’s, which means that the ϕ j ’s form a basis for this equation.
48
Chapter 2
It remains to prove the uniqueness of this equation. The solutions ϕ j provide us with a fundamental matrix Φ for the associated ﬁrstorder 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) a0 (t). I
2.5.2
The adjoint equation
In the context of ﬁrstorder 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 nth order operator system is 0 0 −1 0 0 −1 −A∗ = ⋅ ⋅ 0 0 0 0 Ln , and assume that a0 (t) = 1. The adjoint 0 0 0 ⋅ 0 0 0 0 0 ⋅ 0 0 an 0 0 an−1 0 an−2 ⋅ ⋅ 0 a2 −1 a1
Let us write down this equation in components: y′ = an yn 1 y′ = −y1 + an−1 yn 2 y′ = −y2 + an−2 yn 3 y′ = −yn−2 + a2 yn n−1 y′ = −yn−1 + a1 yn . n Diﬀerentiate the last equation (n − 1) times, yn = −yn−1 + (a1 yn )(n−1) .
(n) (n−1)
Linear equations Diﬀerentiating (n − 2) times the equation for y′ and substituting: n−1 yn = yn−2 − (a2 yn )(n−2) + (a1 yn )(n−1) . We proceed, and eventually get an nth order scalar equation for yn ,
∗ Ln y = 0, ∗ where the adjoint operator Ln is given by ∗ Ln = (−1)n (n) (n−2)
49
dn dn−2 dn−1 + (−1)n−1 n−1 a1 + (−1)n−2 n−2 a2 − dtn dt dt
−
d an−1 + an . dt
Theorem 2.4 (Lagrange identity) Suppose that the coeﬃcients a j in the operator
Ln are diﬀerentiable suﬃciently many times. Then for every pair of functions u, v that are ntimes diﬀerentiable,
∗ v Ln u − u Ln v = [uv]′ ,
where [uv] is deﬁned as follows: [uv] =
n
(−1) j u(k) (an−m v)( j) .
m=1 j+k=m−1
Proof : We will prove the case n = 2. The general proof is left as an exercise. For n = 2 (which is what we’ll end up using later anyway): L2 = and
∗ L2 =
d2 d + a1 + a2 2 dt dt d2 d − a1 + a2 . 2 dt dt
Thus,
∗ v Ln u − u Ln v = v(u′′ + a1 u′ + a2 u) − u(v′′ − (a1 v)′ + a2 v) = (u′′ v + a1 u′ v + a2 uv) − (u(v)′′ − u(a1 v)′ + ua2 v) = (u′ v − uv′ + u a1 v)′ .
I
(Ln y)(t) = b(t). We may assume without loss of generality that a0 (t) ≡ 1.2 (Green’s formula) For every t1 . In this section we derive an expression for the solution of an nonhomogeneous nth order linear equation. . Proof : Obvious I 2. where A(t) is as above and 0 0 B(t) = 0 b(t) The diﬀerences between this problem and the general inhomogenous system is: The vector B(t) has all entries but one zero. . The corresponding ﬁrstorder system is Y ′ (t) = A(t)Y(t) + B(t).5. t2 : t2 t1 ∗ v Ln u − u Ln v dt = [uv](t2 ) − [uv](t1 ).50 Chapter 2 Corollary 2. We are only looking for the 1st component of the solution.3 Nonhomogeneous equation We know how to express the solution of the nonhomogeneous linear equation in terms of the solution operator of the homogeneous system for a general ﬁrstorder system.
Since we are only interested in the ﬁrst component and by the structure of B(t). ϕn )(s) b(s) ds. Wk (ϕ1 . y(t) = yh (t) = n j=1 t0 t Φ1 j (t)Φ−1 (s)b(s) ds jn ϕ j (t) Thus. I TA material 2. then the solution of the initialvalue problem (Ln y)(t) = b(t) is y(t) = yh (t) + n k=1 Y(t0 ) = Y0 . W(ϕ1 . . . and ϕ1 ϕ′ 1 ⋅ ϕk−1 0 ϕk+1 0 ϕ′ ϕ′ k+1 k−1 ⋅ ⋅ (n−1) (n−1) ϕk−1 1 ϕk+1 ϕn ϕ′ n ⋅ Wk (vp1 .5 Let ϕ1 . .2 Apply this formula for n = 2. ϕn ) . . . ϕn be n independent solution of the homogeneous equation Ln y = 0. ϕn ) but this follows from basic linear algebra (formula of the inverse by means of cofactors). it only remains to show that Φ−1 = jn Wk (ϕ1 . . . ϕn )(s) ϕk (t) t t0 where yh (t) is the solution to the homogeneous initialvalue problem.Linear equations 51 Theorem 2. ϕn ) = ϕ1 (n−1) ϕn (n−1) Proof : We know that for we can express the nonhomogeneous part of the solution (in vector form) as t t0 Φ(t)Φ−1 (s)B(s) ds. W(ϕ1 . . . .
Proof : We prove it inductively on n. where 0 1 0 0 0 0 1 0 ⋅ ⋅ ⋅ A= ⋅ 0 0 0 0 −an −an−1 −an−2 −an−4 0 0 ⋅ . 1 −a1 + an y. λI − A = 0 λ a3 a2 λ + a1 p3 (λ) = λp2 (λ) + a3 . For n = 2 λI − A = and so p2 (A) = λ(λ + a1 ) + a2 = λ2 + a1 λ + a2 . . λ −1 0 −1 .5. It is easy to see that pn (λ) = λ pn−1 (λ) + an . Y ′ (t) = AY(t).52 Chapter 2 2.4 Constant coeﬃcients We now consider the case of an nth order equation with constant coeﬃcients: Ln y = y(n) + a1 y(n+1) + where the a j ’s are now constant. Lemma 2. For n = 3. In matrix form.1 The characteristic polynomial of A is p(λ) = det(λI − A) = λn + a1 λn−1 + + an−1 λ + an . a2 λ + a1 Hence. I λ −1 .
. Dividing by eλ1 t . Ln tk eλt = Ln dk dk dk λt e = k Ln eλt = k pn (λ)eλt . dλk dλ dλ (k) If λ has multiplicity m and k < m. s j=1 m j = n. Suppose that they were dependent. t2 eλ j t . It remains to prove that these solutions are linearly independent. Then there would have existed constants c jk . Ln tk eλt = 0. . First. . tm−1 eλ j t . Then. Then the space of solutions is spanned by the functions eλ j t . . Let λ be a root of the characteristic polynomial of multiplicity m. Proof : We need to show that these n functions are indeed solutions and that they are linearly independent. note that for every λ (not necessarily a root of the characteristic polynomial): Ln eλt = pn (λ) eλt . . s. . . t eλ j t . λs be distinct roots of the characteristic polynomial of A such that λ j has multiplicity m j .6 Let λ1 . .Linear equations 53 Theorem 2. P1 (t) + e(λ2 −λ1 )t P2 (t) + + e(λs −λ1 )t P s (t) = 0. . j = 1. then pn (λ) = p′ (λ) = n which proves that = pn (λ) = 0. and let 0 ≤ k ≤ m − 1. such that s m j −1 j=1 k=0 P j (t) c jk tk eλ j t = 0. .
and te−it . The characteristic polynomial is (λ2 + 1)2 = 0 and it has two roots of multiplicity 2: λ = ±i..e. Thus. eit and e−it . Diﬀerentiate suﬃciently many times until P1 disappears. Example: Consider the equation y′′′′ − y = 0. eit − e−it eit + e−it and sin t = 2 2i we may also change basis and use instead the realvalued functions cos t = et . This means that the linear space of solutions is spanned by et . I + e(λs −λ1 )t Q s (t) = 0. t cos t. We will get an equation of the form: e(λ2 −λ1 )t Q2 (t) + We then divide by e(λ2 −λ1 )t . ±i. Since Example: Consider the equation y′′′′ + 2y′′ + y = 0. teit . We proceed this way until we get that Q s (t) = 0. the space of solution is spanned by the basis eit . . e−t .54 Chapter 2 The P j are polynomials. e−it . e−t . cos t and sin t. sin t. cos t. all the c j ’s were zero. to after a change in bases. i. and t sin t. The characteristic polynomial is λ4 − 1 = 0. It follows then backwards that all the Q j ’s (or the P j ’s) vanish. get Q2 (t) + e(λ3 −λ2 )t Q3 (t) + + e(λs −λ2 )t Q s (t) = 0. and its roots are λ = ±1.
In this case we set y(t) = z(t) + εη(t). . that is T periodic. An important question is whether this periodic solution is stable.6. We will assume that T is the smallest number for which this holds.1 Linear systems with periodic coeﬃcients: Floquet theory Motivation Consider a linear homogeneous system of the form y′ (t) = A(t)y(t). The periodic solution is said to be stable if solutions with “slightly perturbed” initial data. Such equations are of interest in systems in which the dynamics have some inherent periodicity (e.6 2. where η 1 do not diverge “too fast” from z(t).. Note that A is automatically kT periodic for every integer k. Substituting into the equation we get z′ (t) + ε η′ (t) = f (z(t) + ε η(t)). z(t). Another important application of linear systems with periodic coeﬃcients is the stability of periodic solutions. and that this solution passes through the point z0 .Linear equations 55 2. such a system is called autonomous. Suppose that this system has a particular solution. where for every t. A(t + T ) = A(t). we can arbitrarily assume that z(0) = z0 . Since the system is autonomous. dynamics of some phenomenon on earth that is aﬀected by the position of the moon). Note that f does not depend explicitly on t. Consider a (generally nonlinear) diﬀerential system: y′ (t) = f (y(t)). y(0) = z0 + εη0 .g. (2.3) 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.
which is 2πperiodic. This is a linear equation with T periodic coeﬃcients. and moreover.56 Chapter 2 We then use Taylor’s theorem (assuming that f is twice diﬀerentiable): z′ (t) + ε η′ (t) = f (z(t)) + ε ∇ f (z(t))η(t) + O(ε2 ).2 General theory A natural question is whether the periodicity of the equation implies the periodicity of the solution. Since z′ (t) = f (z(t)) we remain with η′ (t) = ∇ f (z(t))η(t) + O(ε). Might be interesting to read in this context: the van der Pol oscillator. it diverges as t → ∞). See: http://www.scholarpedia.6.7 Let Φ(t) be a fundamental matrix for the periodic equation. The rationale of this approach is that if η(t) remains bounded in time then for small enough ε. Otherwise. . then Φ(t + T ) is also a fundamental solution. Example: Consider the scalar equation: y′ (t) = (1 + sin t) y(t). One can check by direct substitution that the general solution is y(t) = c exp (t − cos t) . there exists a constant matrix B.org/article/Van_der_Pol_oscillator 2. such that Φ(t + T ) = Φ(t)B. the perturbed solution diverges away from the periodic solution. and det B = exp 0 T Tr A(s) ds . which is not periodic (in fact. the perturbed solution will remain in an ε neighborhood of z(t). Theorem 2. The above is a very rough analysis intended to motivate the study of linear systems with periodic coeﬃcients.
if both Φ(t) and Ψ(t) are fundamental matrices. which proves the ﬁrst part. . it follow that B = Φ−1 (0)Φ(T ). I Lemma 2. Note that in the last step we used the following property: T 0 Tr A(s) ds .2 Let f be periodic with period T . we use Proposition 2. Furthermore. 57 As we know. This means that the knowledge of the fundamental matrix in the interval [0. Since this relation holds for every t. then for every t. we may always choose the fundamental matrix Φ(t) such that Φ(0) = I. we may as well integrate from 0 to T . The fact that Φ(t + T ) = Φ(t)B implies that the fundamental matrix is in general not periodic.Linear equations Proof : Let Φ(t) be a fundamental matrix and set Ψ(t) = Φ(t + T ).. and since A(t) is periodic. Ψ′ (t) = Φ′ (t + T ) = A(t + T )Φ(t + T ) = A(t)Ψ(t). Φ(t + T ) = Φ(t)Φ(T ). such that Ψ(t) = Φ(t + T ) = Φ(t)B. i. for every t.e. Then. T ] determines the fundamental matrix everywhere. Finally. then there exists a constant matrix B. in which case B = Φ(T ).5 to get that det Φ(t + T ) = det Φ(t) t+T t Tr A(s) ds . det Φ(t) det B = det Φ(t) which concludes the proof. t+T t f (s) ds = T 0 f (s) ds.
where t mod T = t − t T T. y(t + T ) = ρy(t).58 Chapter 2 Moreover. The constants µi deﬁned by ρi = eµi T are called the characteristic exponents ( ) of the periodic linear system. y(t + T ) = eµt p(t). Thus.8 Let ρ and µ be a pair of characteristic multiplier/exponent of the periodic linear system. (Note that the characteristic exponents are only deﬁned modulo 2πik T . y(t) is a solution of (2. The forthcoming analysis will make this notion more precise. where p(t) is T periodic. There there exists a solution y(t) such that for every t. there exists a solution y(t) such that for every t. By the deﬁnition of a fundamental matrix. and deﬁne y(t) = Φ(t)u. which proves the ﬁrst part. Proof : Let Bu = ρu. Φ(t) = Φ(t mod T ) B t T periodic T . for arbitrary t.1 The eigenvalues of B = Φ(T ).3) and y(t + T ) = Φ(t + T )u = Φ(t)Bu = ρΦ(t)u = ρy(t). it is easy to see by induction that for k ∈ N: Φ(t + kT ) = Φ(t)Bk . . Also. Deﬁnition 2. ρi are called the characteristic multipliers ( ¯) of the periodic system.) Theorem 2. It follows that the behaviour of solution for large t is dominated by powers of the matrix B.
µn t e fundamental matrix for some constant vector C. every solution is of the form eµ1 t pn (t) y(t) = p1 (t) C. µi )..e. and the general solution to the periodic linear system is n y(t) = αk eµk t pk (t). like the stability of a limit cycle. An immediate corollary is that the system will have periodic solutions if one of the characteristic exponents of B is zero. has a complete set of eigenvectors). a pendulum in which the point of support oscillates vertically). 1 59 which concludes the proof.e.. In other words. Comment: At this stage the theory we developed is not useful in ﬁnding the fundamental matrix. Such equations arise in mechanics when then force on the particle is proportional to its location and the coeﬃcient of proportionality is periodic in time (i. We will now exploit the results of the previous section to study a periodic (scalar) second order equation of the form. 2.3 Hill’s equation I ended up not teaching this section. k=1 where the functions pk (t) are T periodic. makes sense only with a good problem to study. Its importance is so far theoretical in that it reveals the structure of the solution as a combination of exponential functions times periodic functions.Linear equations Next deﬁne p(t) = e−µt y(t). Then there are n pairs (ρi . I Suppose now that B is diagonalizable (i. where a(t + T ) = a(t). y′′ (t) + a(t)y(t) = 0.6. . then p(t + T ) = e−µ(t+T ) y(t + T ) = e−µT ρ e−µt y(t) = p(t).
ϕ′ (t) ϕ′ (t) 11 12 ϕ11 (T ) ϕ12 (T ) . B= ϕ11 (t) ϕ12 (t) . 1 ϕ11 (t) ϕ12 (t) .. We are looking for a fundament matrix Φ(t) = with Φ(0) = I. ϕ′ (T ) ϕ′ (T ) 11 12 Moreover. For any 2by2 matrix B= a b c d the eigenvalues are satisfy the characteristic equation (a − ρ)(d − ρ) − bc = ρ2 − (Tr B)ρ + det B = 0. −a(t) 0 Note that Tr A(t) = 0.2 = γ ± γ2 − 1. Since. . ϕ21 (t) ϕ22 (t) ϕ′ j (t) = ϕ2 j (t). where A(t) = 0 1 . 12 11 The characteristic multiplier are the eigenvalues of B. it follows that ϕ11 (t)ϕ′ (t) − ϕ12 (t)ϕ′ (t) = 1. As usual. it follows that Φ(t) = Thus. we start by rewriting this equation as a ﬁrst order system. ρ1 ρ2 = 1.60 Chapter 2 Note that there is no general explicit solution to a linear secondorder equation with nonconstant coeﬃcients.e. which implies that det B = 1. W(t) = det B = 1. since the Wronskian is constant. y′ (t) = A(t)y(t). and since det B = 1 we get that ρ1. i.
Case 2: γ < −1 In this case both ρ1. hence µ1 + µ2 = 0. and furthermore.2 = eµ1. µ1 = eµ1 T = ρ1 e(µ1 −ıπ T )T = ρ1 . T . where pi (t) are periodic. Case 1: γ > 1 In this case ρ1. Note that 12 2 61 ρ1 + ρ2 = 2γ and ρ1 ρ2 = 1. which implies that e(µ1 +µ2 )T = 1. Even though we do not have a general solution for this type of problems.Linear equations 1 where γ = 2 Tr B = 1 [ϕ11 (T ) + ϕ′ (T )]. and 2γ = eµ1 T + eµ2 T = 2 cosh µ1 T.e. ρ2 < −1 < ρ1 < 0. The general solution is of the form y(t) = c1 eµ1 t p1 (t) + c2 e−µ1 t p2 (t).2 are real and negative.. Thus. in which case implies that i.2 are both real and positive. we are able to express the characteristic coeﬃcients in terms of the single parameter γ: cosh µ1 T and µ1 + µ2 = 0.2 T . iπ + log ρ1 . 0 < ρ2 < 1 < ρ1 . µ1 and µ2 = −µ1 are both realvalued. It follows that there are no periodic solutions and that in general the solution diverges as t → ∞. it turns out that the range of possible solutions can be determined by the single parameter γ. Thus. The characteristic exponents are ρ1.
ρ1. In this case the solutions remain bounded. Case 4: γ = 1 Case 5: γ = −1 . however they will not be periodic. Case 3: −1 < γ < 1 Now ρ1. period 2T period 2T Note the doubling of the period. Once again there are no periodic solutions and that in general the solution diverges as t → ∞.62 Chapter 2 Thus.2 = e±iσT (this relation deﬁnes σ) and thus µ1. unless it happens that σT = 2π m for some integer m. the general solution is of the form y(t) = c1 elog ρ1 t eiπt T p1 (t) +c2 e− log ρ1 t e−iπt T p2 (t) .2 are both complex and located on the unit circle.2 = ±iσ.2 (t) periodic. In this case the general solution is of the form y(t) = c1 eiσt p1 (t) + c2 e−iσt p2 (t). with p1.
one has to specify initial conditions. ∂2 y ∂2 y = c2 2 . 0) = f (x) (It is a initialboundary problem. y(x. Consider a ﬁnite string. Boundary value problems arise also in the study of partial diﬀerential equations. x ∈ [0.Chapter 3 Boundary value problems 3. t) = y(L. The constant c is the wave speed that depends on the mass density and tension of the string. 0) = g(x).) and ∂y (x. Example: The equation that describes the motion of an elastic string is known as the wave equation ( ). L] tied at its two ends. ∂t2 ∂x where the unknown is a function y(x.1 Motivation Many problems in science involve diﬀerential equations with conditions that are speciﬁed at more than one point. t) of space x and time t. ∂t . such that (∀t) y(0. In addition. t) = 0.
there exists a constant k. but only if k = πm L for some integer m. 1]: Ly = λy y(0) = y(1) = 0. Consider the equation for u(x). t) and y2 (x. so that if y1 (x. such that u′′ (x) = ku(x) and v′′ (t) = k v(t). we have a boundary value problem: u′′ (x) = ku(x) u(0) = u(L) = 0. Substituting into the wave equation: u′′ (x)v(t) = c2 u(x)v′′ (t). this time on the interval [0. u(x) = √ √ sin( kx). Are there more solutions? Such questions are going to be addressed in the course of this chapter. and any linear combination of these solutions is also a solution to the wave equation. Are there other solutions? yes. Note that this equation is linear. namely. c2 Note that we may have many such solutions for various values of k. Let’s get back to the same boundary value problem. and further.64 Chapter 3 This problem is analytically solvable. where L=− d2 . we will denote it by x rather than t. u ≡ 0. u(x) v(t) Since this equation holds for all x and t. Because the independent variable in boundary value problem is often space rather than time. t) are solutions to the wave equation. One method of solution is to look for solutions that have a separation of variables: y(x. t) = u(x)v(t). dx2 . v′′ (t) u′′ (x) = c2 . then so is any linear combination. it must be the case that both sides are equal to the same constant. Of course. Because of the boundary conditions. Does it have a solution? Yes. a notation is nothing but a notation.
1] with an inner product: ( f. We will provide all the needed information.e. The class of functions that are (well. In case you didn’t take this course. the values of λ for which the boundary value problem has a solution are called eigenvalues. . 1 0 f (x) 2 dx exists is known as L2 [0. Note that this is an orthonormal set. For a function f ∈ L2 [0. the normalized eigenfunctions are √ ψm (x) = 2 sin(mπx). . this looks like an eigenvalue problem! Indeed. no worry.Boundary value problems 65 The parameter λ can be in general complex (the reason for the minus sign in L will be made clear later). (ψm .. for m = 1. Then. What we see is that the equation Ly = λy with boundary values has a solution only for selected values of λ. which is an object that you learn about in Advanced Calculus 2. We know that the general solution to the equation Ly = λy is y(x) = A sin λ1 2 x + B cos λ1 2 x. Since we further require y(1) = 0. Does it remind us something? L is a linear operator. (m = 0 is the trivial solution and we ignore it as of now). we endow the space of functions on [0.. it follows that λ = π 2 m2 . ψk ) = 2 1 0 sin(mπx) sin(kπx) dx = δmk .. λ is a scalar. . . Lebesgue. i. 1]. The corresponding “eigenvectors” are rather called eigenfunctions ( ). 1]. yes. fˆ(k) = ( f. its innerproduct with the elements of the orthogonal set ψk are called its Fourier coeﬃcients. ψk ) = √ 2 0 1 f (x) sin(kπx) dx.) square integrable. Since we require y(0) = 0 then B = 0. Like for vectors. g) = 1 0 f (x)g(x) dx. 2. It is a Hilbert space..
N. the norm of f and its Fourier coeﬃcients satisfy a relation known as the Parseval identity: ∞ f 2= fˆ(k) 2 . N − 1. Talk about harmonics. . . 1] norm to f . We get a linear equation for the discrete vector (yi ).1 Solve the wave equation initialvalue problem. TA material 3. Comment: An interesting connection between these inﬁnitedimensional boundary value problems and “standard” eigenvalue problems can be made by considering discretizations. n→∞ lim f − n k=1 fˆ(k)ψk = 0. Finally. . i = 1.66 The series Chapter 3 ∞ k=1 fˆ(k)ψk (x) converges in the L2 [0. 2∆x2 where ∆x = 1 N is the mesh size. namely. . . these results hold in a much wider scope. i . Suppose that we want to solve the boundary value problem −y′′ (x) = λ y(x). In addition we set y0 = yN = 0. N and look for yi = y(xi ). − 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 y1 y2 yN−2 yN−1 y1 y2 yN−2 yN−1 =λ . xi = yi−1 + yi+1 − 2yi . k=1 We have stated all these facts without any proof or any other form of justiﬁcation. We can approximate the solution by deﬁning a mesh. . by approximating the diﬀerential equation by a ﬁnitediﬀerence approximation. Talk about what happens when you touch the center of a guitar string. y(0) = y(1) = 0. . . i = 0. . As we will see.
with n = 2.2. n dx dx dx where p j ∈ C n− j (a.1 Consider a linear nth order diﬀerential operator. b). p1 (x) = p2 (x) = 0. k = 1. . L = p0 (x) dn−1 d dn + p2 (x) n−1 + ⋅ ⋅ ⋅ + pn−1 (x) + pn (x). Example: The example we studied in the previous section is such an instance. b) is in general complexvalued.2 Selfadjoint eigenvalue problems on a ﬁnite interval Deﬁnitions 3. Uk y = Mk1 y(a) + Mk2 y′ (a) + + Nk1 y(b) + Nk2 y′ (b) + The equation Ly = λy. we assume that p0 (x) ≠ 0 for all x ∈ (a. for which we expect N eigenvalues. We also deﬁne a linear operator U C n (a. b) → Rn of the form Uk y = n j=1 Mk j y( j−1) (a) + Nk j y( j−1) (b) . . Uy = 0 + Mkn y(n−1) (a) + Nkn y(n−1) (b).Boundary value problems 67 This is a “standard” eigenvalue problem. . p0 (x) = −1. 3. Compare your results with the solution of the continuous equation. is called an eigenvalue problem.1 Find the eigenvalues and eigenvectors of this discretized system. That is. . where the M jk and N jk are constant. . n. As we reﬁne the mesh and take N → ∞. the number of eigenvalues is expected to tend to inﬁnity. Exercise 3. and verify their dependence on N.
This corresponds to periodic boundary conditions. L∗ v) = [uv](b) − [uv](a). .e. and then (Lu. and other M jk . In the ﬁrst case Uu = 0 implies that u(a) = u(b) = 0.e. v) = − b a u′′ (x)v(x) dx = b a u′ (x)v′ (x) dx = − b a u(x)v′′ (x) dx = (u. the corresponding functions are called eigenfunctions. U1 y = M11 y(a) + N11 y(b) + M12 y′ (a) + N12 y′ (b) = y(a) − y(b) U2 y = M21 y(a) + N21 y(b) + M22 y′ (a) + N22 y′ (b) = y′ (a) − y′ (b). Thus. v ∈ C n (a. if L∗ = L and Uu = Uv = 0 implies that [uv](b) = [uv](a)..68 and Chapter 3 M11 = N21 = 1. M22 = −N22 = 1. b) that satisfy the boundary conditions Uu = Uv = 0. A selfadjoint problem always has a trivial solution for any value of λ. Lv) for every u. Values of λ for which a nontrivial solution exists are called eigenvalues. Recall the adjoint operator L∗ derived in the previous chapter: dn dn−1 d p0 (x) + (−1)n−1 n−1 p1 (x) + − pn−1 (x) + pn (x). and other M jk . Deﬁnition 3. then the problem is selfadjoint. N jk are zero. Lv). U1 y = M11 y(a) + N11 y(b) + M12 y′ (a) + N12 y′ (b) = y(a) U2 y = M21 y(a) + N21 y(b) + M22 y′ (a) + N22 y′ (b) = y(b). N jk are zero. i. Example: Another class of examples for n = 2: M11 = −N11 = 1. v) − (u.. n dx dx dx The Green identity proved in the previous chapter showed that L∗ = (−1)n (Lu. v) = (u. ) if Example: The two above examples are selfadjoint. i.1 The eigenvalue problem is said to be selfadjoint ( (Lu.
ϕ). Comment: The last part of the proof relies on complex analysis. i.. ψ). λ). λ(ψ. Comment: At this stage there is no guarantee that eigenvalues exist. λ) = δ jk . j = 1. Lψ) = λ(ψ. Then. Lψ = λψ.Boundary value problems 69 3. Eigenfunctions that correspond to distinct eigenvalues are orthogonal. . Let λ ≠ µ be distinct eigenvalues. n. ψ) = 0. Lϕ = µϕ.e. which proves that ψ is real. λ(ψ. ϕ) = 0. Lϕ) = µ(ψ. . ψ) = (Lψ. µ) = (Lψ.1 Consider a selfadjoint boundary value problem. ψ) = (ψ. which implies that ψ and ϕ are orthogonal.2 Properties of the eigenvalues Proposition 3. The eigenvalues form at most a countable set that has no (ﬁnitevalued) accumulation point. (λ − µ)(ψ. . (k−1) ϕ j (c. . i. ϕ) = (ψ.2. Fasten your seat belts. . (λ − λ)(ψ. The eigenvalues are real. Then. Let now ϕ j (x. so as of now this theorem can hold in a vacuous sense.. Proof : Let λ be an eigenvalue with eigenfunction ψ. Then.e. be solutions of the equation Lϕ j = λϕ j with initial conditions.
. for ﬁxed x the function λ ϕ(x. the functions ϕ j are continuous both in x and λ. It is wellknown (especially if you learn it. hence every solution of the equation Ly = λy is of the form y(x) = For λ to be an eigenvalue we need U y= i.e. j=1 k=1 This is a homogeneous linear system for the coeﬃcients c j . ϕ1 (c) = 1 ϕ′ (c) = ϕ′′ (c) = ϕ′′′ (c) = 1 1 1 ϕ′ (c) = 1 ϕ2 (c) = ϕ′′ (c) = ϕ′′′ (c) = 2 2 2 ϕ′′ (c) = 1 3 ϕ3 (c) = ϕ′ (c) = ϕ′′′ (c) = 3 3 etc. Also. n. λ) is analytic. . the ϕ j (x.) that the zeros of an entire analytic function that it not identically zero do not have accumulation points (except at inﬁnity). . . λ) + N k ϕ j A j (λ) (k−1) (b.e. For it to have a nontrivial solution the determinant of the matrix A j (λ) must vanish. . b). n n n j=1 n j=1 c j ϕ j (x. In fact. = 1. det A j (λ) is analytic in the entire complex plane. . λ) c j = 0. That is. c j U ϕ j (⋅. . = 1. That is. we are back in initial values problems.. 1 The functions ϕ j are linearly independent. λ) = 0. there no issue of existence nor nonuniqueness. M kϕ j (k−1) (a. n. . i. . I 1 A concept learned in complex analysis. the eigenvalues λ are identiﬁed with the roots of det A j (λ). It is not identically zero because it can only vanish on the real line. . λ) are deﬁned for every (complex) λ..70 Chapter 3 where c is some point in (a.. λ). I will say a few words about it. = ϕ1 (n−1) (n−1) (n−1) (c) = 0 (c) = 0 (c) = 0 = ϕ2 = ϕ3 Note.
for example. For n = 1 we get d2 ϕ = 4πρ dx2 ϕ(0) = ϕ(L) = 0. λ) be as in the previous section solutions to Lϕ j (⋅.Boundary value problems 71 3. p0 (s) Note that p0 and p1 are the same for L and L − λ id. A (unique) solution exists if the matrix A − λI is not singular.2. Uy = 0.3 Nonhomogeneous boundary value problem We turn now to study the nonhomogeneous equation Ly = λy + f. The Poisson equation for the electric potential in a domain Ω ⊂ Rn with a metallic boundary is ∆ϕ = 4πρ. . ϕn )(x) = exp − x c (k−1) (c. Example: Such boundary value problems may arise. . we will see that the nonhomogeneous boundary value problem has a solution for values of λ that are not eigenvalues of the homogeneous equation. Let ϕ j (x. .. . . p1 (s) ds . . Similarly. ϕ ∂Ω = 0. in electrostat ics. subject to the boundary conditions ϕj Note that W(ϕ1 . where f ∈ C(a. .e. What to expect? Think of the linearalgebraic analog. ϕn )(c) = det I = 1. hence W(ϕ1 . . Ax = λx + b. λ) = λϕ j (⋅. λ). if λ is not an eigenvalue of A. λ) = δ jk . b). . i. where ∆ is the Laplacian and ρ(x) is the charge density. which is of the same type as above with λ = 0.
.72 Chapter 3 We now deﬁne the following function: ϕ1 (ξ. . λ) Lϕn (x. λ) = 0. λ) ϕ′ (ξ. . λ) ϕ′ (ξ. ξ. . ξ. λ) = b a x a ϕn (ξ. λ) ϕ′ (ξ. dx 0 . ϕn )(ξ) (n−2) ϕ1 (ξ. λ) ϕn (x. b]. ξ. ϕ1 (ξ. λ) f (x). The (n − 1)st and nth derivatives are continuous in every subdomain. ξ. . λ) = p0 (ξ)W(ϕ1 . 1 dn−1 K(ξ+ . ξ). ξ. u′ (x. λ) 1 1 ⋅ K(x. λ) n ⋅ = λK(x. λ) = d K(ξ+ . in particular K is continuous for n ≥ 2. λ) = x a d K(x. . λ) f (ξ) dξ. both taking values in [a. ξ. λ) K(x. . λ) 1 1 ⋅ LK(x. . ξ. λ) ϕ1 (x. ξ. ξ. Deﬁne now the function u(x. λ) f (ξ) dξ + K(x. λ) This relation holds trivially for x < ξ. It is apparently discontinuous. however note that K(ξ+ . for x > ξ. λ) f (ξ) dξ = K(x. n−1 dx p0 (ξ) Finally. λ) = . ϕn )(ξ) (n−2) ϕ1 (ξ. x− . λ) = p0 (ξ)W(ϕ1 . λ). n−2 dx hence K and its ﬁrst n − 2 derivatives with respect to x are continuous over all (x. λ) ϕ′ (ξ. λ) n ⋅ (n−2) ϕn (ξ. λ) ξ≤x x < ξ. We calculate its ﬁrst n derivatives with respect to n. λ) = dx = dn−2 K(ξ+ . λ) 0 ϕn (ξ. λ) Lϕ1 (x. (n−2) ϕn (ξ. ξ. Also. This function depends on x and ξ.
e. b]: UkG(⋅. i. and set c j (ξ) such that for every k = 1. x− . λ) + f (x). λ) f (x). . dxn−1 dx 0 Finally. Set then u(x. This linear system has a unique solution if det Uk ϕ j (⋅. λ) = = b a b a G(x. ξ. λ) f (ξ) dξ + K(x. λ) f (x) . λ) is a solution to the nonhomogeneous equation..Boundary value problems u′′ (x. i. x− . λ) = x a x a 73 d d2 K(x. 2 dx dx 0 dn−1 dn−2 K(x. ξ. Deﬁne G(x. u(n) (x. .. dxn dx f (x) p0 (x) It follows that Lu(x. λ) f (ξ) dξ + n−2 K(x. λ) = 0. We are looking for a solution of the nonhomogeneous equation that satisﬁes the boundary conditions. λ) f (ξ) dξ + n j=1 ϕ j (x) b a c j (ξ) f (ξ) dξ. x− . if λ is not an eigenvalue of the homogeneous boundary value problem. Uk y = 0. λ) = x a dn−1 dn K(x. λ) = and so until u(n−1) (x. . λ) = K(x. u(x. ξ. λ) f (x) . . λ) + n j=1 c j (ξ)ϕ j (x. ξ. n and every ξ ∈ [a. ξ. λ) = x a LK(x. ξ. ξ. . λ) f (ξ) dξ + f (x) = λu(x. λ) ≠ 0. ξ. λ) = Uk K(⋅. λ) f (ξ) dξ + n−1 K(x. ξ.e. λ) + n j=1 c j (ξ)Uk ϕ j (⋅. ξ. λ). λ) f (ξ) dξ K(x.
λ) + n j=1 c j (ξ)Uk ϕ(⋅. ξ. 2 ϕ2 (x) = λ−1 2 sin λ1 2 x. cos λ1 2 ξ cos λ1 2 x sin λ1 2 ξ = sin λ1 2 (ξ − x) sin λ1 2 x ξ≤x x<ξ . ϕ1 (0) = 1 ϕ′ (0) = 0 1 G(x. cos λ1 2 x and −ϕ′′ (x) = λϕ1 (x). ξ. ξ.2 have to satisfy. . λ) for this particular example. ξ. λ) = λu(x. ϕ2 )(x) = Next. λ) = − 0 Now. The coeﬃcients c1. and G(1. λ) and G(x. λ) = b a Uk K(⋅. 2 namely. we found a solution to the nonhomogeneous boundary value problem for every value of λ that is not an eigenvalue of the homogeneous boundary value problem. and Uk u(⋅. λ) f (ξ) dξ = 0. ξ. Solution 3. λ) − c1 (ξ) cos λ1 2 x − c2 (ξ)λ−1 2 sin λ1 2 x. W(ϕ1 .2: We start by constructing the two independent solution.2 Solve the boundary value problem. Exercise 3. y(0) = y(1) = 0. λ) = sin λ1 2 (ξ − 1) − c1 (ξ) cos λ1 2 − c2 (ξ)λ−1 2 sin λ1 2 = 0. cos λ1 2 x −λ1 2 sin λ1 2 x λ1−0 2 sin λ1 2 x = 1.74 Then. λ) = K(x. λ). K(x. ξ. For what values of λ solutions exist? Repeat the above analysis (including the calculation of K(x. Thus. Chapter 3 Lu(x. G(0. since p0 (ξ) = −1. −ϕ′′ (x) = λϕ2 (x). λ) = 0 − c1 (ξ) = 0. −y′′ (x) = λy(x) + f (x). 1 ϕ2 (0) = 0 ϕ′ (0) = 1. ξ. ξ. ϕ1 (x) = cos λ1 2 x Thus.
ξ. GLu = u for all u ∈ C n (a. λ) = x 0 sin λ1 2 (ξ − x) f (ξ) dξ − sin λ1 2 x sin λ1 2 1 0 sin λ1 2 (ξ − 1) f (ξ) dξ. b).Boundary value problems from which we get that G(x. ξ. Suppose now that zero is not an eigenvalue of the homogeneous boundary value problem. (Note that L and G are almost inverse to each other. Note that by deﬁnition LG(x. ξ). b) satisfying Uu = 0. (G f. λ) = K(x. ξ. g) = ( f. We will denote G(x. ξ) = G(ξ. It is easy to see that the boundary conditions are indeed satisﬁed. ξ) = 0 at all points where x ≠ ξ. λ) − sin λ1 2 (ξ − 1) sin λ1 2 x sin λ1 2 75 Thus. x). 0) simply by G(x. and consider the boundary value problem Ly = f Uy = 0. Deﬁne the linear integral operator: G f b a G(⋅. Proposition 3. ξ) f (ξ) dξ. You may now check by direct diﬀerentiation that the diﬀerential equation is indeed satisﬁed.2 The integral operator G (and the Green function G) satisﬁes the following properties: LG f = f and UG f = 0 for all f ∈ C(a. the solution to the boundary value problem is u(x.) . G(x. Gg). 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 λ0 that is not en eigenvalue and then L − λ0 I will be used as an operator for which zero is not en eigenvalue. which maps the nonhomogeneous term into the solution of the boundary value problem.
g). I 3.76 Chapter 3 Proof : The ﬁrst statement follows from the fact that G maps f into the solution of the boundary value problem. x) f (ξ)g(x) dxdξ = 0. which proves the second statement. ξ)g(ξ) dξ dx.3 The existence of eigenvalues Ly = λy. The ﬁrst step will be to show that eigenfunctions of L coincide with eigenfunctions of G: . Changing the order of integration and interchanging x and ξ in the second integral: b a a b G(x. LGg). Uy = 0. (LG f. It follows from the second statement that b a a b G(x. Assuming that zero is not an eigenvalue we deﬁned the Green function G(x. LGg) = (u. ξ) and the operator G. Since this holds for every f and g the third statement follows. So far we haven’t shown that the eigenvalue problem (3. Gg) = (G f. ξ) f (ξ) dξ g(x) dx = b a f (x) b a G(x. We only know properties of solutions if such exist. hence by the ﬁrst statement ( f. Finally. which implies that GLu = u. g). b) and u ∈ C n (a. b) satisfying Uu = 0. for every g ∈ C(a.1) has solutions. ξ) − G(ξ. By the selfadjointness of L and the fact that UG f = UGg = 0. g) = (u. Gg) = (G f. In this section we will show that selfadjoint eigenvalue problems always have a countable set of eigenvalues. (GLu.
I Thus. Then. Then.2) Proof : Let (λ. 1).3 ϕ is an eigenfunction of (3. λ (3.1) has eigenvalues we can rather show that G has eigenvalues. suppose that (3. . namely 1 Gϕ = ϕ.1) with eigenvalue λ if and only if it is also an eigenfunction of G with eigenfunction 1 λ.2) holds. in order to show that (3. λ Conversely. Lemma 3.1). where we used the fact that the range of G is functions that satisfy the boundary conditions.2 that 1 Gϕ = ϕ. it follows from the fourth statement of Proposition 3. u ≤ 1} ⋅ is the L2  is bounded (in the supnorm) and equicontinuous (but the norm norm). 1 Lϕ = λL ϕ = λLGϕ = λϕ. λ and Uϕ = λUGϕ = 0.Boundary value problems 77 Proposition 3. λ λ and since Uϕ = 0. ϕ) be eigensolutions of (3. 1 1 Gϕ = Gλϕ = GLϕ.1 The set of functions X = {Gu u ∈ C(0.
ξ)−G(x2 . for all u ∈ X and x ∈ [a. b]) (∀x1 .1 that G < ∞ (i. ξ) ∈ [a.. Then. u =1 (u. it is also bounded. x1 − x2 < δ) b G(x1 . G is a bounded linear operator). (∀ε > 0)(∃δ > 0) (∀u ∈ X ) (∀x1 . b) → L2 (a. x2 it holds that Gu(x1 ) − Gu(x2 ) ≤ b a x1 −x2 < δ) G(x1 . Gu)1 2 . For n ≥ 2 the function G(x. x2 Thus. I which proves the uniform boundedness of X . b]2 . the case of n = 1 requires a slightly diﬀerent treatment. ξ) < ε. ξ) is continuous on the square (x. b) to avoid measuretheoretic subtleties. b). It follows from the ArzelaAscoli theorem: Corollary 3. which proves the equicontinuity of X . Gu(x) ≤ b a G(x. 2 We therefore deﬁne its operator norm: G = sup Gu = sup (Gu. ξ) is continuous. ξ) u(ξ) dξ ≤ ε u(ξ) dξ. Since G(x. (∀ε > 0)(∃δ > 0) (∀ξ ∈ [a. ξ) u(ξ) dξ ≤ K(b − a)1 2 u . Gu(x1 ) − Gu(x2 ) ≤ b a 1 2 u(ξ) 2 dξ a b 1 2 dξ = ε(b − a)1 2 u . b) to C n (a. The map G is actually a map from L2 (a. ξ) − G(x2 . and in particular.u)=1 It follows from Lemma 3. b].e.78 Chapter 3 Proof : We will consider the case of n ≥ 2. hence it is uniformly continuous. a Using the CauchySchwarz inequality. .1 Every sequence in X has a subsequence that converges uniformly on [a. let K be a bound. 2 We deﬁned it as a map from C(a. b].
(G(u + v). Substitute now v = Gu Gu . w) w =1 ≤ 2 sup (Gw.4 G = sup (Gu. v) ≤ sup (Gw. for every u and v. u) + (Gv. w) w =1 u 2 + Gu Gu 2 . u) + (Gv. u) ≤ Gu u ≤ G u i. u + v) = (Gu. v) − (Gu. u =1 2 = G . v) − (Gv. u) = (Gu. u) = (Gu. w =1 Subtracting the second inequality from the ﬁrst.e. u) .Boundary value problems 79 Proposition 3. u) ≤ G . Gu or. sup (Gu. u 2 +1 . u − v) = (Gu. v) ≤ 2 sup (Gw. 4R(Gu. Gu 4R Gu.3) By the bilinearity of the innerproduct and the selfadjointness of G. u) + (Gv. w =1 and (G(u − v).. v) + (Gv. u) + (Gv. Then. v) + 2R(Gu. w) u + v 2 . . it follows from the CauchySchwarz inequality and the deﬁnition of the norm that (Gu. v) − 2R(Gu. u =1 Proof : For every normalized u. v) + (Gu. 4 Gu ≤ 2 sup (Gw. w) w =1 u 2 + v 2 . w) u − v 2 . (3. v) ≥ − sup (Gw.
the second case is treated similarly. u =1 w =1 (3. Proof : Denote G = µ0 . then either µ0 = sup (Gu. Comments: This proves the existence of an eigenvalue to the boundary value problem. n→∞ . there exists a sequence of functions (un ) such that un = 1 lim (Gun . w) . or µ0 = − sup (Gu. w =1 and since this holds for every such u. u). G = sup Gu ≤ sup (Gw. u =1 Consider the ﬁrst case. (3.3) and (3. thus the sup in the deﬁnition of G is attained.4) together give the desired result.4) I Eqs. By the deﬁnition of the supremum. Theorem 3. un ) = µ0 . w) . Everything we have been doing holds in the ﬁnitedimensional case. If A is an hermitian matrix then either A or − A is an eigenvalue.80 For u = 1 we get Chapter 3 Gu ≤ sup (Gw. Since there exists an eigenfunction ϕ such that Gϕ = G ϕ it follows that Gϕ = G ϕ . u) u =1 or Gϕ = − G ϕ.1 Either G or − G is an eigenvalue of G.
Gun − µ0 un Moreover. Gun − ϕ0 ) = Gun 2 + ϕ0 2 − 2R(Gun .2 G has an inﬁnite number of eigenfunctions. un ) → 0. I Theorem 3. 0 lim Gun − µ0 un = 0. ϕ0 ) → Gun 2 − ϕ0 2 . Finally. we get the convergence of the norm.) Now. b]. which concludes the proof. Since 0 ← (Gun − ϕ0 . n→∞ lim Gun − ϕ0 = 0. (It is always true that convergence in norm implies the convergence of the norm. Letting n → ∞ we get that Gϕ0 = µ0 ϕ0 .Boundary value problems 81 By Corollary 3. ≤ 2µ2 − 2µ0 (Gun . n→∞ lim Gun = ϕ0 . 0 from which follows that ϕ0 ≠ 0 (we need this to claim that ϕ0 is nontrivial). . denote its limit by ϕ0 . it follows that 0 ≤ Gun − µ0 un from which follows that n→∞ 2 2 = Gun 2 + µ2 − 2µ0 (Gun . un ) → ϕ0 0 2 − µ2 . uniform convergence implies L2 convergence.1 the sequence (Gun ) has a subsequence (not relabeled) that converges uniformly on [a. by the triangle inequality Gϕ0 − µ0 ϕ0 ≤ Gϕ0 − G(Gun ) + G(Gun ) − µ0 Gun + µ0 Gun − µ0 ϕ0 ≤ G ϕ0 − Gun + G Gun − µ0 un + µ0 Gun − ϕ0 . In particular.
G1 v). ϕ0 and ϕ1 are orthogonal. i. Let. then for every f ∈ C(a. ξ) − µ1 ϕ1 (x)ϕ1 (ξ) and ﬁnd a third orthonormal eigenfunction of G. G1 ϕ1 = µ1 ϕ1 Note that G1 ϕ0 = 0. Hence. ϕ j ) = 0. which means that ϕ1 is also an eigenfunction of G. Gϕ1 = G1 ϕ1 + µ0 ϕ1 (ϕ1 . By the maximality of µ0 . ϕ0 ) = G1 ϕ1 = µ1 ϕ1 .e. µ1 ≤ µ0 . ξ) − µ0 ϕ0 (x)ϕ0 (ξ). unless G1 = 0 (the zero operator). v) = (Gu. G1 f (x) = This operator is also selfadjoint. ϕ0 (x)ϕ0 (ξ)u(ξ)v(x) dξdx = (u. whose absolute value is equal to G1 . Gm f (x) = G f (x) − m−1 j=1 µ j ϕ j (x)( f. ξ) f (ξ) dξ. and the corresponding operator. The only way this process can stop is if there is a Gm = 0 for some m. ϕ0 = 1. ϕ0 ) = 1 1 (G1 ϕ1 .. . ξ) = G(x. (G1 u. v) + µ0 b a b a G1 (x. it has an eigenvalue µ1 . G1 ϕ0 ) = 0. deﬁning next G2 (x. Suppose this were the case.82 Chapter 3 Proof : Let ϕ0 and µ0 be as above and assume that ϕ0 has been normalized. We continue this way. Then. µ1 µ1 ϕ1 = 1. Deﬁne a new kernel G1 (x. b). ϕ0 ) = (ϕ1 . hence (ϕ1 . ξ) = G1 (x.
2 (Bessel’s inequality) Let f ∈ L2 (a. j f (x) = m−1 j=1 83 ϕ j (x)( f. ϕ j ). This would imply that every continuous function is in the span of m diﬀerentiable functions. Proof : For every m ∈ N. b). which is wrong. ϕ j ) 2 . 0≤ f − m j=0 2 ( f. using the fact that LG f = f and Lϕ j = µ−1 ϕ j . .Boundary value problems Applying L on both sides. Lemma 3. ϕ j ) 2 converges and ∞ j=0 ( f.4 Boundary value problems and complete orthonormal systems We have thus proved the existence of a countable orthonormal system. I which proves both statements. ϕ j )ϕ j = f 2 − m j=0 ( f. with corresponding eigenvalues µ0 ≥ µ1 ≥ µ2 ≥ . I 3. Then the sequence ∞ j=0 ( f. ϕ j ) 2 ≤ f 2 . ϕ j .
x)ϕ j (ξ) dξ = µ j ϕ j (x). ϕ j )ϕ j . ξ) = G(ξ. x). ∞ j=0 ( f. b].84 Chapter 3 Theorem 3. . ξ) 2 dξdx ≤ K 2 (b − a)2 . (Gϕ j )(x) = b a G(x. m j=0 µj 2 ≤ b a G(x. m j=0 b a G(ξ. (G(x. Integrating over x. ξ) It follows at once that ∑∞ µ j 2 converges and in j=0 particular µ j → 0. we get f 2 = ∞ ( f. By Bessel’s inequality applied to G(x. b]. ξ) 2 dξ. Corollary 3. Multiplying this equation by f¯ and integrating over [a. ⋅) 2 = b a G(x. ϕ j ) 2 j=0 Proof : For every j ∈ M and x ∈ [a. then f= where the convergence is uniform. where K is a bound on G(x. ⋅). interchanging the order of integration and summation.3 Let f ∈ C n (a. ⋅). m j=0 µ j 2 ≤ G(x. ⋅). ϕ j ) 2 = m j=0 µ j 2 ϕ j (x) 2 = m j=0 µ j 2. (G(x.2 An immediate consequence is Parseval’s identity. ξ)ϕ j (ξ) dξ = µ j ϕ j (x). b) and U f = 0. ϕ j ) = Squaring and summing over j. Taking the complex conjugate and using the fact that G(x.
ϕ j )ϕ j . j=p where we used the CauchySchwarz inequality. For q > p. Then. ϕ j )ϕ j = G q (u. ϕ j )ϕ j converges unij=p formly. By the way we constructed the sequence of eigenfunctions. ϕ j )ϕ j = 0. b). . b). it follows that ∞ j=0 µ j (u. Gm u = Gu − Letting m → ∞. where the convergence is uniform. ϕ j )ϕ j = Gu. ϕ j )ϕ j = K(b − a) q j=p 1 2 1 2 (u. j=p Hence. Since it converges to Gu in L2 and the latter is continuous. ϕ j )ϕ j ≤ K(b − a) q 1 2 (u. q j=p µ j (u. Since the right hand side vanishes as p. and ∞ j=0 µ j (L f. Thus. q j=p µ j (u. ϕ j ) 2 . U f = 0.Boundary value problems For m ∈ N consider Gm (x. We will show that it j=0 converges in fact uniformly. ξ) = G(x. ϕ j )ϕ j converges to Gu in L2 . Let now f ∈ C n (a. m→∞ m−1 j=0 µ j (u. This means that ∑∞ µ j (u. we know that Gm = µm . for every u ∈ C(a. ξ) − m−1 j=0 85 µ j ϕ j (x)ϕ j (ξ). it follows by Cauchy’s criterion that ∑q µ j (u. L f is continuous. q → ∞. lim Gu − m−1 j=0 µ j (u. ϕ j )ϕ j ≤ µm u . ϕ j )ϕ j = GL f.
I .86 Since GL f = f and Lϕ j = µ−1 ϕ j . ϕ j )ϕ j = f. j ∞ j=0 Chapter 3 ( f.
. 0) y = (2. Consider the following nonlinear system: then ψ(t) = ϕ(t − t0 ) is a solution to the initial data problem y′ (t) = f (y(t)). in a system y′ (t) = f (t. Fixed points For the above system. four points in the yplane that are of special interest: y = (0. 2). y(t0 ) = y0 . i. y(t)).Chapter 4 Stability of solutions 4. 2 This system does not depend explicitly on t. such systems are called autonomous. ⋅) is called autonomous if it does not depend on t.e. 1) and y = (−1. 2) y = (0.1 Deﬁnitions y′ = 1 (y1 − y2 )(1 − y1 − y2 ) 1 3 y′ = y1 (2 − y2 ). f Rn → Rn . More generally. . y′ (t) = f (y(t)). y(0) = y0 . the family of vector ﬁelds f (t. Note that if ϕ is a solution to the autonomous initial data problem.
We will be concerned in this chapter with the global behaviour of solutions. the global behaviour of solutions can be inferred by knowing the behaviour of solutions in the vicinity of ﬁxed points. Chaotic. Yet. . Asymptotic convergence to a limit cycle. y(t) − u(t) < ε ∀t > t0 . t→∞ Comment: The stability of ﬁxed points is a particular case of stability of solutions as deﬁned above. The above system is nonlinear and cannot be solved by analytical means for arbitrary initial values. If any of these points is prescribed as an initial value then the solution will be constant. a ﬁxed 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 ﬁxed point. here is a list of possible behaviors: Fixed point. Roughly speaking. Asymptotic convergence to a ﬁxed point. It is called asymptotically stable if there exists a δ such that y(t0 ) − u(t0 ) < δ implies that lim y(t) − u(t) = 0. such that for every solution y(t) satisfying y(t0 ) − u(t0 ) < δ. or equilibrium points ( ). Limit cycle (periodic solution). or stationary points. Tends asymptotically to inﬁnity.88 Chapter 4 These are ﬁxed points ( ). Each such point corresponds to a solution to the system that does not change in time. Quasiperiodic. in many cases. The notion of stability can be made more precise and more general: Deﬁnition 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 ε > 0 there exists a δ > 0.
Since we are interested in understanding the behaviour of solutions in the vicinity of u(t).2). Show that the stability/instability of the zero solution of (4.) .2) hence the stability of u(t) in the nonlinear equation (4. Substituting in the equation we get u′ (t) + z′ (t) = f (u(t) + z(t)). we set y(t) = u(t) + z(t). It is sensible to think that if z(t0 ) is “very small”. a solution u(t) of (4. (4.1). then as long as z(t) is in some compact set U ⊂ Rn . that linear stability implies nonlinear stability. then at least as long as it remains so.1) for which the zero solution of the linearized equations is stable is called linearly stable ( ).2) is stable/unstable. hence the devision of the solution from u(t) is governed by the linearized equation.2) implies the stability/instability of the solution u(t) of the nonlinear equation (4. Since u′ (t) = f (u(t)).1) can be determined by the stability of of the solution y = 0 of the linear equation (4. the quadratic correction to the Talor expansion is negligible.Stability of solutions 89 4. we get a diﬀerential 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 diﬀerentiable. y′ (t) = D f (u(t))y(t).1) and a particular solution u(t). The task is therefore 2fold: Show that the zero solution of the linearized equation (4.2 Linearization y′ (t) = f (y(t)). (That is. z′ (t) = D f (u(t))z(t) + O( z(t) 2 ). Consider an autonomous system (4.
y(t) = eD f (u) t y(0). 1 where δ ≥ 0.90 Chapter 4 The ﬁrst task is generally diﬃcult. we know the solution to the linearized equation. This describes a particle that is repelled from the origin when close it. (0. and (ii) if an eigenvalue has a zero realpart then it is of multiplicity one. In this case.1) is linearly stable if and only if (i) all the eigenvalues of D f (u) have nonpositive real parts. it is called a source ( ) all the eigenvalues have positive real parts. Moreover.. i. since we do not have a general way of solving linear equations with nonconstant coeﬃcients. y′ = y2 1 ′ y2 = y1 − y3 − δy2 . we will also prove it using a new technique. 0) and (±1. This system has three ﬁxed points. but attracted to it when far away from it. . The ﬁrst task is however easy if u(t) is a constant solution. and can therefore determine the following: Theorem 4. D f (u) has real part equal to zero. It is called a sink ( ) if all the eigenvalues have negative real parts. if u is a ﬁxed point.1 A ﬁxed point u of (4. we know the form of the exponential eD f (u) t . it is called a saddle ( ¯ ) if certain eigenvalues have positive real parts while other have negative real parts. The second task is in fact not always possible. 0). one can ﬁnd situations in which a solution is linearly stable but nonlinearly unstable. it is called a center ( ¯ ) if all the eigenvalues are imaginary. While we already know this is true.e. Deﬁnition 4.2 A ﬁxed point u is called hyperbolic if none of the eigenvalues of Example: Consider the unforced Duﬃng oscillator.
for deﬁne z(t) = y(t) − u. which implies that these ﬁxed points are stable.e. . u2 ) be the ﬁxed point. 0) there is a positive and a negative eigenvalue. Suppose that all the eigenvalues of D f (u) have negative real parts. 1 0 1 . which is stable if and only if u is a stable ﬁxed point of f . 4. Comment: Without loss of generality we may assume that u = 0. 2 1 − 3u1 −δ 91 For the ﬁxed point (0. Then. which implies that it is unstable..Stability of solutions Let u = (u1 . 0) both eigenvalues have negative real parts. Zero is a ﬁxed point of F.3 Lyapunov functions Theorem 4. Comment: This is a situation in which linear stability implies nonlinear stability.2 Let f Rn → Rn be twice diﬀerentiable. z′ (t) = f (z + u) ≡ F(z). let u be a ﬁxed point of f . D f (u) = The eigenvalues satisfy λ(λ + δ) − (1 − 3u2 ) = 0. For the ﬁxed point (±1. 1 i. Then u is asymptotically stable. furthermore. The linearized equation is y′ = y2 1 ′ y2 = y1 − 3u2 y1 − δy2 . DF(0) = D f (u). 1 which implies that λ= 1 −δ ± 2 δ2 + 4(1 − 3u2 ) .
solutions are “trapped” within those level sets. D f (0)y) < −α y 2 . such that ∀y ∈ U f (y) − D f (0)y ≤ C y 2 . Moreover. f (y(t)) − D f (0)y(t)) . By the chain rule. Close enough to the ﬁxed point. there exists a constant α > 0 such that (y. There exists a neighborhood U of the origin and a constant C > 0. .. Then. f (y)) ≤ 0.92 Chapter 4 Proof : The idea of the proof is based on a socalled Lyapunov function. Suppose we construct a function V Rn → R that satisﬁes the following conditions: V(y) ≥ 0 for all y. g′ (t) = (y(t). That is. then R(t) ≤ y(t) ⋅ C y(t) 2 g′ (t) < −2αg(t) + R(t). Back to the proof. i. there exists a neighborhood U of the origin. if y(t) ∈ U. V(y) = 0 iﬀ y = 0. ≡R(t) 1 2 y .e. Deﬁne the Lyapunov function V(y) = and set g(t) = V(y(t)). the vector ﬁeld f points “inward” on the level sets of V. such that ∀y ∈ U (DV(y). f (y(t))) = (y(t). D f (0)y(t)) + (y(t). The sets Kc = {y ∈ Rn V(y) ≤ c} are compact. 2 Since all the eigenvalues of D f (0) have negative real part. connected and contain the origin as internal point. ≤ C1 g3 2 (t).
d αt (e g(t)) < 0. and as long as y(t) ∈ U ′ g′ (t) < −αg(t). such that ∀x ∈ Rn (Ay.1 Show that if A is a real matrix whose eigenvalues have all negative real parts.e. Consider the following system. 0) is a center and we have no current way to determine its (nonlinear) stability. Consider then the Lyapunov function 1 V(y) = (y2 + y2 ). y′ = y2 1 ′ y2 = −y1 − εy2 y2 . the solution is asymptotically stable. Example: An important use of Lyapunov functions is in cases where linear stability cannot conclusively determine the (nonlinear) stability of a ﬁxed point.Stability of solutions 93 In particular. y) < −β y 2 . i. which implies that t→∞ eαt g(t) < g(0). I Exercise 4. y2 ) y2 = −εy2 y2 1 2 −y1 − εy2 y2 1 . there exists a constant k > 0 such that Kc ⊂ U. (DV(y). then there exists a constant β > 0. lim g(t) = 0. 1 The point (0. 0) is a ﬁxed point. and in turn this implies that y(t) → 0. 2 1 2 Then. If follows that if g(0) < c then g(t) < c for all t > 0. dt and further. −1 0 and the eigenvalues are ±i. This means that (0. In addition.. The linearized equation about the origin is y′ = 0 1 y. f (y)) = (y1 .
the origin is a center. Since V(y(t)) is monotonically decreasing it must converge to a limit δ > 0. The origin is a ﬁxed point.3 Let V be a Lyapunov function.e. Consider the Lyapunov function 1 V(y) = (y2 + 3y2 ). Example: Consider the system y′ = −3y2 1 y′ = y1 + α(2y3 − y2 ). Theorem 4. i. 2 2 where α > 0. and D f (0) = 0 −3 1 0 √ The eigenvalues are ± 3i..e. . dt then the origin is asymptotically stable. g′ (t) = y1 (−3y2 ) + 3y2 y1 + α(2y3 − y2 ) = −α(3y2 − 6y4 ). That is. also the level sets of V are closed. then g′ (t) ≤ 0 for all t.. Proof : Suppose that y(t) does not converge to zero. 2 2 2 1 It follows that g′ (t) ≤ 0 as long as y2 < 2 . if 2 g(0) < 3. 2 2 1 Setting g(t) = V(y(t)).e.. V(y) ≥ 0 with V(y) = 0 iﬀ y = 0. i. i. ∀t > 0 δ ≤ V(y(t)) ≤ V(y0 ).94 Chapter 4 The origin is therefore stable for ε > 0. If d V(y(t)) < 0.
I Exercise 4.3 Consider the system x′ = −x3 − y2 y′ = xy − y3 . . Prove that the origin is stable using the Lyapunov function V(x. which leads to a contradiction for large t. where g is continuous and satisﬁes xg(x) > 0 for x ≠ 0 and k > 0. Thus. max(DV(y). d V(y(t)) = (DV(y(t)). y) = 1 + 2 x g(s)ds. (DV(y). 0 Exercise 4.Stability of solutions The set C = {y δ ≤ V(y) ≤ V(y0 )} is closed and bounded and hence compact.2 Consider the nonlinear system x′ = y y′ = −ky − g(x). dt and therefore V(y(t)) ≤ V(y0 ) − αt.) Show that the origin is stable by using the Lyapunov function V(x. f (y)) ≡ −α < 0. y∈C 95 Then. y) = −x − log(1 − x) − y − log(1 − y). f (y(t)) ≤ −α. (This might represent the equation of a mass and nonlinear spring subject to friction. f (y)) must attain its maximum on C.
O(y) = {Φt (y) t ∈ R} is an invariant set. Example: Any trajectory. Deﬁnition 4. if y1 is the solution at time t of the diﬀerential system. Φt Rn → Rn Deﬁnition 4. A set S ⊂ Rn is That is. y(0) = y0 . ) if It is said to be negatively invariant ( Φt (y) ∈ S for all y ∈ S and for all t < 0. namely y1 = Φt (y0 ). This vector ﬁeld induces a ﬂow map ( via the trajectories of the solutions.4 Invariant manifolds ). It is said to be positively invariant ( ) if Φt (y) ∈ S for all y ∈ S and for all t > 0.3 Let f Rn → Rn be an autonomous vector ﬁeld. Any union of trajectories.4 Let f Rn → Rn be an autonomous vector ﬁeld. y′ = f (y). O(y) y∈A . Another way to write it is d Φt (y) = f (Φt (y)) dt said to be invariant if Φt (y) ∈ S for all y ∈ S and for all t ∈ R.96 Chapter 4 4. a trajectory that starts in S has always been in S and remains in S . Φ0 (y) = y.
where A = D f (u). . Consider again an autonomous system. y′ = Ay. e s } This is of course a nonrigorous deﬁnition. we may decompose Rn as follows: Rn = E s ⊕ Eu ⊕ Ec . and conversely. for a rigorous deﬁnition you need to take a course in diﬀerential geometry. and any backward trajectory.. . i. y′ = f (y).e. Any forward trajectory. The solution of the linearized system is y(t) = eAt y0 .5 An invariant set S is said to be an invariant manifold ( ) if “looks” locally like a smooth surface embedded in Rn . and let u be a ﬁxed point. . any invariant set is a union of trajectories. Deﬁnition 4. 97 Example: A ﬁxed point is an invariant set (no matter if it is stable or not). O+ (y) = {Φt (y) t ≤ 0} is a negativelyinvariant set. As we know from linear algebra. . where E s = Span{e1 . O+ (y) = {Φt (y) t ≥ 0} is a positivelyinvariant set. 1 1 .Stability of solutions is an invariant set. The behavior of the solution near the ﬁxed point is studied by means of the linearized equation. the ﬂow map of the linearized system is given by the linear map Φt = eAt .
. λ2 < 0 and λ3 > 0. These three linear subspaces are examples of invariant manifolds (of the linearized system!). e3 be the corresponding normalized eigenvectors. The characterization of the unstable manifold Eu . Then. eλ1 t 0 0 At λ2 t 0 T −1 y0 . e s+u } is the span of (generalized) eigenvectors corresponding to eigenvalues with positive real part. is that t→−∞ lim Φt (y) = 0 ∀y ∈ Eu . e2 } and Eu = Span{e3 } are invariant manifolds. Eu = Span{e s+1 . e2 . The characterization of the stable manifold E s . . Example: Let n = 3 and suppose that A has three distinct real eigenvalues. . e s+u+c } is the span of (generalized) eigenvectors corresponding to eigenvalues with zero real part. A = e1 e2 e3 As we know. The geometry of the trajectories is depicted in the following ﬁgure: λ1 0 0 0 λ2 0 0 0 λ3 e1 e2 e3 −1 ≡ T ΛT −1 . . . λ1 . is that t→∞ lim Φt (y) = 0 ∀y ∈ E s . and Ec = Span{e s+u+1 . . . y(t) = e y0 = T 0 e 0 0 eλ3 t In this case E s = Span{e1 . . .98 Chapter 4 is the span of (generalized) eigenvectors corresponding to eigenvalues with negative real part. Let e1 .
in which ρ −ω 0 Λ= ω ρ 0 . e1 + e2 e1 + e2 e1 − e2 A √ =ρ √ −ω √ 2 2 2i e1 − e2 e1 − e2 e1 + e2 A √ =ρ √ +ω √ . . the corresponding eigenvectors are complex.Stability of solutions 99 Eu! Es! Example: Suppose now that λ1. 2i 2i 2 Thus. however A(e1 + e2 ) = (ρ + iω)e1 + (ρ − iω)e2 = ρ(e1 + e2 ) + iω(e1 − e2 ) A(e1 − e2 ) = (ρ + iω)e1 − (ρ − iω)e2 = ρ(e1 − e2 ) + iω(e1 + e2 ). 0 0 λ3 The solution to this system is eρt cos ωt −eρt sin ωt 0 y(t) = T eρt sin ωt eρt cos ωt 0 T −1 y0 .2 = ρ ± iω and λ3 > 0. there exists a transformation of variables A = T ΛT −1 . where ρ < 0. Then. 0 0 eλ3 t or.
100 Chapter 4 Here again there is a twodimensional stable subspace and a onedimensional unstable subspace. This is a deﬁcient system. 0 0 λ3 eλ1 t teλ1 t 0 y(t) = T 0 eλ1 t 0 T −1 y0 .4 Consider the following diﬀerential systems: y′ = y′ = y′ = y′ = y′ = y′ = λ 0 y 0 µ λ 0 y 0 µ λ −ω y ω λ 0 0 y 0 λ 0 λ y 0 0 0 0 y. ω > 0. A = T ΛT −1 with Λ= and the solution is λ1 1 0 0 λ1 0 . SHOW FIGURE. λ > 0. Describe the stable and unstable manifolds of the origin. . Then. Example: Consider now the case where λ1 < 0 is an eigenvalue of multiplicity two and λ3 > 0. µ > 0. 0 0 λ < 0. e2 } and Eu = Span{e3 }. λ < 0. SHOW FIGURE. Exercise 4. hence has only one direction “entering” the ﬁxed point. λ < 0. µ < 0. In each case calculate all the trajectories and illustrate them on the plane. λ < 0. 0 0 eλ3 t In this transformed system E s = Span{e1 .
. Moreover. . W u and W c. ∀y ∈ W s and ∀y ∈ W u t→−∞ t→∞ lim Φt (y) = 0. Then there exists a neighborhood of the origin in which there are C k invariant manifold. Can you infer from this analysis the nonlinear stability? The question is what happens in the nonlinear case. Proof : Not in this course. I Example: Consider again the Duﬃng oscillator.Stability of solutions 101 Exercise 4.4 Suppose that the vector ﬁeld f is ktimes diﬀerentiable and let the origin be a ﬁxed point. How does the structure of the linearized equation aﬀect the properties of the solutions in the vicinity of the ﬁxed point? Theorem 4. x′ = y y′ = x − x3 − δy. namely. the nonlinear stable and unstable manifolds retain the asymptotic properties of the linearized manifolds. that intersect at the origin are are tangent at the origin to the invariant manifolds of the linearized invariant manifolds. lim Φt (y) = 0.5 Consider the nonlinear systems: y′ = y2 1 y′ = −δy2 − µy1 − y2 2 1 y′ = −y1 + y3 1 1 y′ = y1 + y2 2 . W s. Find all the ﬁxed points and determine their linear stability.
y i. i. The origin is a ﬁxed point and the linearized system is x y ′ = 1 0 0 −1 x . The origin is a saddle.2 = y = λ1. Wu# Ws# Example: Consider the following artiﬁcial system.2 x. 2 The corresponding (linear) stable and unstable manifolds are λ1. it has a onedimensional stable manifold and a one dimensional unstable manifold.e. ... 1 −δ 1 √ 2 ± δ +4−δ . E s = Span{e2 } and Eu = Span{e1 } are the stable and unstable manifolds of origin in the linearized system. x′ = x y′ = −y + x2 .102 The linearized vector ﬁeld is Chapter 4 y′ = with eigenvalues 0 1 y.e.
6 A solution through a point y is said to be periodic with period Φt (y) = y.5 T > 0 if Periodic solutions Deﬁnition 4. . we get a onedimensional manifold which is invariant (since it is a trajectory) and tangent to the unstable manifold of 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. x 3 For C = 0. y′ (x) = ′ x (t) x This is a linear system whose solution is y(x) = C x2 + . y y′ (t) = − + x. We will now concentrate on periodic solutions of autonomous systems in the plane. We look for trajectories y(x). These satisfy the diﬀerential equation. Ws# Wu# 4. The stable manifold of the nonlinear system is the yaxis. x = 0.
S Γ where n is the unit normal to Γ.5 (Bendixon’s criterion) Consider the planar system. f (x. Let Γ be a closed curve in the plane. y) . φ(x.6 Index theory In this section we are concerned with autonomous diﬀerential systems in the plane: x′ = f (x. then at any point along the curve the vector ﬁeld makes an angle φ with the x axis. I Example: For the Duﬃng oscillator. hence this system cannot have any closed orbit. then there are no closed orbits lying entirely in D. as assumed div f is neither identically zero nor changes sign. n) d = 0. and ( f. If in a simply connected domain D div f = ∂ f1 ∂ f2 + ∂y1 ∂y2 is not identically zero and does not change sign. Proof : Suppose there was a closed orbit Γ that encloses a domain S ⊂ D.104 Chapter 4 Theorem 4. y) = tan−1 g(x. y) y′ = g(x. y). y′ = f (y) where f is at least C 1 . If. div f = −δ. n) = 0 because f is tangent to Γ. By Green’s theorem. then this is impossible. div f dy = ( f. y) . 4. If the curve does not intersect a ﬁxed point.
or center is 1. or a center. The index of a periodic orbit is 1.Stability of solutions 105 When we complete a cycle along Γ the angle φ must have changed by a multiple of 2π. source. Sink% Source% Saddle% Periodic% Corollary 4. We deﬁne the index of the curve by i(Γ) = 1 2π γ dφ. The index of a curve is equal to the sum of the indexes of the ﬁxed points that it encloses. then there must be 2n + 1 of those with n saddles and n + 1 sources/sinks/centers.1 Every periodic trajectory encloses at least one ﬁxed point. The index of a curve that encloses a single sink. as long as it does not pass through a ﬁxed point. The index of a saddle is −1. . The index of a curve that does not enclose any ﬁxed point is zero. It turns out that the index of a curve satisﬁes the following properties: The index of a curve does not change when it is deformed continuously. If it encloses a single ﬁxed point then it is either a sink. a source. If all the enclosed ﬁxed points are hyperbolic.
A point y belongs to the ωlimit of y. ˜ y ∈ ω(y).7 Let y ∈ Rn . Deﬁnition 4. ˜ if there exists a sequence tn → ∞ such that n→∞ lim Φtn (y) = y. ˜ .7 Asymptotic limits We are back to autonomous systems in Rn .106 Chapter 4 Example: According to index theory. we denote by Φt Rn → Rn the ﬂow map. possible limit cycles in the Duﬃng oscillator are as follows: 4. The trajectory through a point y is the set: O(y) = {Φt (y) t ∈ R} We also have the positive and negative trajectories: O+ (y) = {Φt (y) t ≥ 0} O− (y) = {Φt (y) t ≤ 0}.
1 Let M be a compact positively invariant set (i. 4. As before we denote the ﬂow function by Φt (x. then α(z) = {u}. ω(y) = {u}. ˜ Thus the ωset ( ) ω(y) denotes the collection of all “future partial limits” of trajectories that start at y. y). Proposition 4. We consider again twodimensional systems. Example: Let y R → Rn be a periodic solution. ˜ if there exists a sequence tn → −∞ such that n→∞ 107 lim Φtn (y) = y. If z belongs to the unstable manifold of u. Then for every t0 . Then.8 The Poincar´ Bendixon theorem e y′ = f (y).. Let p ∈ M. and α(y) denotes the collection of all “past partial limits” of trajectories that pass through y.e.Stability of solutions It belongs to the αlimit y. . Example: Let u be a ﬁxed point and let y belong to the stable manifold of u. y ∈ α(y). trajectories that start in M remain in M at all future times). ω(y(t0 )) = α(y(t0 )) = O(y(t0 )). Then.
Since Φtn (p) ∈ M and M is compact this sequence has a converging subsequence. which is therefore nonempty. Let q ∈ ω(p). and hence ω(p) is invariant. there exists a neighborhood U of q and a time T > 0 such that Φt (p) ∈ U ∀t > T.108 Chapter 4 ω(p) ≠ . i. lim Φtn (p) = q.e. ω(p) is invariant. Then. i. and Φt (p) ∈ V which implies that z ∈ ω(p).. every point z ∈ U has a neighborhood V ⊂ U. n→∞ n→∞ which proves that Φτ (q) ∈ ω(p) for all τ ∈ R. By deﬁnition. By deﬁnition q ∈ ω(p). ω(p) is connected. k→∞ lim Φtnk (p) = q ∈ M. (ω(p))c is open. hence ω(p) is closed. . Let q ∈ ω(p).e. it is a union of trajectories. suchthat n→∞ ∀t > T. Since U is open.. n→∞ lim Φtn +τ (p) = lim Φτ (Φtn (p)) = Φτ lim Φtn (p) = Φτ (q). Let τ ∈ R be arbitrary. Proof : Take any sequence tn → ∞. there exists a sequence tn → ∞. By the deﬁnition of the ωset. ω(p) is closed.
V1# V2# K# I Deﬁnition 4.Stability of solutions 109 Suppose by contradiction that ω(p) was not connected. It follows that the trajectory O(p) must pass from V1 to V2 and back inﬁnitely often. V2 such that ω(p) ⊂ V1 ∪ V2 ω(p) ∩ V1 ≠ and ω(p) ∩ V2 ≠ . namely.e. f (Σ(t)) ≠ Σ′ (t) f (Σ(t) . Then there exist disjoint open sets V1 .8 Let Σ be a continuous and connected arc in R2 (i. f" ∑" . hence must have there an accumulation point. The trajectory must pass inﬁnitely often in the compact set K V1 .. Consider a compact set K that encloses V1 and not intersect V2 . (Σ′ (t). a map I → R2 ). in contradiction to ω(p) being contained in V1 ∪ V2 . It is called transversal to the ﬂow ﬁeld f neither vanishes on Σ nor is tangent to Σ.
Proof : Proof by sketch: pi# pi+1# Posi)vely#invariant#set# I Corollary 4. for suppose that p1 .110 Chapter 4 arc in M. that is. Proof : The existence of two intersection points would violate the previous lemma. the positive trajectory O+ (p) intersects Σ monotonically. then the sequence si is monotonic. or that it has to intersect it more than once. Lemma 4.2 Let M ⊂ R2 be a positively invariant set and let Σ be a transversal arc in M. Then for every p ∈ M. if pi = Σ(si ) is the ith intersection of O+ (p) and Σ. p2 ∈ Σ ∩ ω(p). ω(p) intersects Σ at at most one point. Then for every p ∈ M. # Σ ∩ ω(p) ≤ 1. .1 Let M ⊂ R2 be a positively invariant set and let Σ be a transversal Comment: The lemma does not say that O+ (p) must intersect Σ at all.
If ω(p) does not contain any ﬁxed point. Since O(q) ⊂ ω(p) (an ωlimit set is invariant) and ω(p) is closed. then all the qn must coincide with x. but since qn ∈ ω(p). Proof : Let q ∈ ω(p) and let x ∈ ω(q) (we know that both are nonempty sets). Let Σ now be a transversal arc at q.2 Let M be a positively invariant set and let p ∈ M. hence it is periodic. O+ (p) would intersect Σ at inﬁnitely many points in disjoint neighborhoods of p1 and p2 . we can trace through it a transversal arc Σ. it follows that ω(p) = O(q). It remains to show that O(q) coincides with ω(p). O(r) ⊂ ω(p). Note that this point is somewhat subtle. Any trajectory visiting this neighborhood intersects U ∩ Σ. Since ω(p) is connected. Take any neighborhood U of p1 and consider U ∩ Σ. then ω(p) is a periodic trajectory. By transversality. x is not a ﬁxed point. then ω(q) ⊂ ω(p). and there would be such an r for which O(r) intersects Σ. . but does it mean that it has to intersect Σ each time? Yes. The trajectory O+ (q) will intersect Σ monotonically at points qn → x.Stability of solutions 111 Then. for if every neighborhood of O(q) contained points in r ∈ ω(p). We claim that O(q) has a neighborhood in which there are no other points that belong to ω(p). i. {Φt (U ∩ Σ) −ε < t < e} is an open neighborhood of p1 in R2 .e. Since x is not a ﬁxed point. It follows that ω(q) does not contain ﬁxed points. I Proposition 4. It follows that O(q) intersects x more than once.. If p1 ∈ Σ ∩ ω(p) we are guaranteed that the trajectory will visit inﬁnitely often any neighborhood of p1 . which is an open neighborhood of p1 in Σ.
There exists at most one trajectory. then it is called a homoclinic orbit.9 A heteroclinic orbit is a trajectory that connects two ﬁxed points. such that α(O(y)) = {p1 } and ω(O(y)) = {p2 }. γ ⊂ ω(p). where p1 and p2 are ﬁxed points.e. such that {p1 } = α(γ) and {p2 } = ω(γ). a trajectory O(y).. Let p1 . If p1 = p2 (a saddle). Proof : Proof by sketch: . p2 ∈ ω(p) for some p ∈ M. Proposition 4. i. Since ω(p) is connected and invariant (a union of trajectories). There exists a heteroclinic orbit that connects p1 and p2 (or vice versa). p2 be ﬁxed points and p1 .3 Let M be a positively invariant set.112 Chapter 4 p" q" A"zone"in"which"there"" are"no"point"in"w(p)" I Deﬁnition 4.
ω(p) is a limit cycle. If ω(p) does not contain any ﬁxed point then it is a limit cycle (we proved it). Then. one and only one of the following occurs: ω(p) is a single ﬁxed point. on a trajectory that connects between two ﬁxed points in ω(p). . Let p ∈ M. Similarly. ω(p) is a union of ﬁxed points pi and heteroclinic orbits γ j that connect them. as claimed. hence q is on a trajectory that converges asymptotically to a ﬁxed point in ω(p). let q ∈ ω(p) and consider the set ω(q). α(q) contains a single ﬁxed point. We will show that it contains a single ﬁxed point.Stability of solutions 113 p1# p# p2# I Proposition 4. so that q lies.4 (Poincar´Bendixon) Let M be a positively invariant set that e contains a ﬁnite number of ﬁxed points. Remains the case where ω(p) contains ﬁxed points and points that are not ﬁxed points. Proof : If ω(p) contains only ﬁxed points then it only contains one due to connectedness.
then we would trace a transversal arc Σ through r. but then we know that it can only contain a single ﬁxed point. it must intersect Σ at a single point. It follows that ω(q) contains only ﬁxed points. I . which implies that O(q) is a periodic orbit. but this is impossible because ω(p) is connected. The trajectory O(q) must intersect this arc inﬁnitely often.114 Chapter 4 So let x ∈ ω(q). Suppose that x was not a ﬁxed point. but since O(q) ∈ ω(p).