Ordinary Differential Equations

Modern Perspective
Ordinary Differential Equations
Modern Perspective
Mohan C Joshi
IIT Bombay
Preface
This book has culminated into its title as an outgrowth of series of undergradu-
ate and graduate level courses on ordinary differential equations (ODE), taught
by the author for the past several years. It is our aim to present a unified and
comprehensive treatment to ODE based on widely used abstract formulations in
the underlying spaces. The emphasis is more on qualitative analysis of solvabil-
ity of ODE than the actual methodology of computing the closed form solution.
In our presentation we presume only the basic knowledge of linear algebra and
advanced calculus of the undergraduate level. Hence the text is suitable for
graduate students of mathematics and other branches of science and engineer-
ing requiring a broad perspective of ODE. The contents have been class tested
for several years in a regular course on ODE in the Department of Mathematics
at IIT Bombay.
With this approach in mind, we have divided the text into eight chapters. Chap-
ter 1 introduces ODE with several examples drawn from applied areas. These
examples are continuously tracked at appropriate places in later chapters for
illustrating the theory.
Chapter 2 is devoted to the development of mathematical structure required to
deal with linear and nonlinear operators in function spaces. These operators
will arise while discussing various facets of ODE.
Chapter 3 gives basic existence and uniqueness theory needed for initial value
problems, the starting point for any discussion of ODE. The subsequent analysis
rests heavily on the concept of transition matrix. It is, therefore, dealt with in
detail in Chapter 4. While in Chapter 5, we use the properties of transition
matrix in the study of stability theory of both linear and nonlinear systems.
Chapter 6 lays emphasis on the clasical theory of series solution for ODE. We
focus on Legendre, Hermite and Bessel equations. Boundary value problems
are investigated in Chapter 7 through the methods of Green’s function and
eigenfunction expansion. We give rather an elongated description on Green’s
function in view of its inherent importance.
The closing Chapter 8 is a fitting finale, giving interesting glimpse of vast ap-
plicability potential of ODE to control theory - a fertile inter-disciplinary area
involving of science and engineering.
Acknowledgements
This book could not have been written without direct and indirect assistance
from various institutional bodies within IIT Bombay. Firstly, we would like to
acknowledge the financial support received from the Curriculum Development
Program of the Institute. Secondly, we thank the Department of Mathematics
and Industrial Mathematics Group for providing the necessary secretarial and
computational help throughout the preparation of the manuscript while it went
through the number of iterations. Several colleagues and students helped us
from time to time in many ways giving correct references, proof-reading the
manuscript etc. In this connection, we would like to acknowlege the help rendred
by Amiya K Pani, M K Ramaprasad and Soumya Mohanty. Our special thanks
to Asmit Pany for type setting the manuscript in LATEX.
Finally, this author would like to thank his wife for patiently supporting him in
implementing the project.
Mohan C. Joshi
Powai, Mumbai
Contents
1 Introduction To Ordinary Differential Equations 1
1.1 First Order Ordinary Differential Equations . . . . . . . . . . . . 1
1.2 Classification of ODE . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Higher Order ODE . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Difference Equations . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.6 Techniques for Solving First Order Equations . . . . . . . . . . . 26
1.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2 Mathematical Preliminaries 37
2.1 Finite and Infinite Dimensional Spaces . . . . . . . . . . . . . . . 37
2.2 Linear and Nonlinear Operators . . . . . . . . . . . . . . . . . . . 47
2.3 Lipschitz Continuity and Differentiability . . . . . . . . . . . . . 57
2.4 Fixed Point Theorems and Monotone Operators . . . . . . . . . 63
2.5 Dirac-Delta Function . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3 Initial Value Problems 79
3.1 Existence and Uniqueness Theorems . . . . . . . . . . . . . . . . 79
3.2 The Maximal Interval of Existence . . . . . . . . . . . . . . . . . 89
3.3 Dependence on Initial Condition . . . . . . . . . . . . . . . . . 93
3.4 Cauchy-Peano Existence Theorem . . . . . . . . . . . . . . . . . 96
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4 Linear System And Transition Matrix 101
4.1 Linear System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.2 Transition Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.3 Periodic Linear System . . . . . . . . . . . . . . . . . . . . . . . . 114
4.4 Computation of Transition Matrix . . . . . . . . . . . . . . . . . 118
4.5 Euler’s Linear System . . . . . . . . . . . . . . . . . . . . . . . . 131
4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
CONTENTS
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5 Stability Analysis 137
5.1 Asymptotic Behaviour of Linear System . . . . . . . . . . . . . . 137
5.2 Stability Analysis - Formal Approach . . . . . . . . . . . . . . . . 149
5.3 Phase Portrait Analysis . . . . . . . . . . . . . . . . . . . . . . . 154
5.4 Lyapunov Stability . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
6 Series Solution 175
6.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
6.2 Linear System with Analytic Coefficients . . . . . . . . . . . . . . 181
6.3 Linear System with Regular Singularities . . . . . . . . . . . . . 188
6.4 Properties of Legendre, Hermite and Bessel Functions . . . . . . 196
6.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
7 Boundary Value Problems 207
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
7.2 Adjoint of a Differential Operator . . . . . . . . . . . . . . . . . . 209
7.3 Green’s Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
7.4 Construction of Green’s Function and Modified Green’s Function 221
7.5 Eigenfunction Expansion - Sturm-Liouville System . . . . . . . . 233
7.6 Nonlinear Boundary Value Problem . . . . . . . . . . . . . . . . 237
7.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
8 Control Theory Of Linear Systems 243
8.1 Controllability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
8.2 Observability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
8.3 Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
8.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
Chapter 1
Introduction To Ordinary
Differential Equations
In this chapter we begin by introducing the concept of solvability of initial value
problems (IVP) corresponding to first order differential equations (ODE) in '
n
.
We also give, along with appropriate examples of practical importance, all re-
lated notions like orbit or trajectory, direction field, isocline and solution curve
of IVP.
The reader is also familiarized with difference equations and numerical solutions
of ODE through the process of discretization. An interesting concept of neural
solution of ODE is also touched upon. Towards the end, we recall standard
methods of solving some classes of first order ODE.
1.1 First Order Ordinary Differential Equations
Let Ω ⊂ '
n
be open and let f : Ω →'
n
be a mapping ( not necessarily linear)
given by
f(x
1
, x
2
, x
n
) = (f
1
(x
1
, x
2
, . . . , x
n
), f
2
(x
1
, x
2
, . . . x
n
), . . . , f
n
(x
1
, x
2
, . . . , x
n
))
We shall refer to f as vector field on Ω. A vector field may also depend on an
additional parameter t, that is, f = f(t, x).
The main thrust of our study is the following differential equation
d¯ x
dt
= f(t, x) (1.1.1)
where f : I Ω →'
n
, I a subinterval of '.
The value x(t) ∈ Ω ⊂ '
n
is sometimes referred as the state of the system
described by the ordinary differential equation (ODE) - Eq.(1.1.1). The set Ω
is then called the state space or phase space.
1
2CHAPTER1. INTRODUCTIONTOORDINARYDIFFERENTIAL EQUATIONS
Definition 1.1.1 The Initial Value Problem (IVP) corresponding to Eq. (1.1.1)
is given by
dx
dt
= f(t, x(t)) (1.1.1(a))
x(t
0
) = x
0
(1.1.1(b))
Definition 1.1.2 A function x(t) is said to be a solution of the IVP - Eq.
(1.1.1), if there exists an interval J ⊂ I, containing t
0
such that x(t) is differ-
entiable on J with x(t) ∈ Ω for all t ∈ J and ¯ x(t) satisfies Eq. (1.1.1).
As x(t) needs to satisfy Eq. (1.1.1) only on J, this solution is sometimes referred
as a local solution. If t →f(t, .) is defined on the entire line ' and ¯ x(t) satisfies
Eq. (1.1.1) on ', then ¯ x(t) is said to be a global solution of the IVP. In Chapter
3, we shall discuss the local and global solvability of IVP in detail.
At times, it is desirable to indicate the dependence of the solution x(t) on
its initial value x
0
. Then we use the notation x(t, t
0
, x
0
) for x(t).
Definition 1.1.3 The orbit or the trajectory of the ODE- Eq.(1.1.1), is the set
¦x(t, t
0
, ¯ x
0
) : t ∈ J¦ in the state space Ω. Whereas the solution curve is the set
¦(t, x(t, t
0
, ¯ x
0
)) : t ∈ J¦ ⊂ I Ω.
Definition 1.1.4 The direction field of ODE - Eq. (1.1.1), is the vector field
(1, f(t, x)).
It is clear from the above definitions that the orbit of the ODE - Eq. (1.1.1) is
tangent to the vector field whereas the solution curve is tangent to the direction
field at any point.
The following proposition gives the equivalence of solvability of IVP - Eq.
(1.1.1) with the solvability of the corresponding integral equation.
Proposition 1.1.1 Assume that f : I Ω → '
n
is continuous. ¯ x(t) is a
solution the IVP - Eq. (1.1.1) iff ¯ x(t) is the solution of the integral equation
¯ x(t) = ¯ x
0
+

t
t0
f(s, ¯ x(s))ds, t ∈ J (1.1.2)
Proof : If ¯ x(t) is a solution of Eq.(1.1.1), by direct integration on the interval
[t
0
, t] for a fixed t, we get

t
t0

d¯ x
ds

ds =

t
t0
f(s, ¯ x(s))ds
This gives
¯ x(t) − ¯ x(t
0
) =

t
t0
f(s, ¯ x(s))ds
and hence we have Eq. (1.1.2).
1.1. FIRST ORDER ORDINARY DIFFERENTIAL EQUATIONS 3
Conversely, if Eq.(1.1.2) holds, ¯ x(t) given by Eq. (1.1.2) is differentiable and
differentiating this equation we get
d¯ x
dt
=
d
dt
¸
t
t0
f(s, ¯ x(s))ds

= f(t, x(t))
Also we have ¯ x(t
0
) = ¯ x
0
.
Definition 1.1.5 The set of points in Ω where f(t, ¯ x) = 0, is called the set of
equilibrium points of ODE - Eq. (1.1.1). It is clear that at these points the orbit
or trajectory is a singleton.
Example 1.1.1 (Population Dynamics Model)
We investigate the variation of the population N(t) of a species in a fixed time
span. An important index of such investigation is growth rate per unit time,
denoted by
R(t) =
¸
1
N(t)

dN
dt
(1.1.3)
If we only assume that the population of species changes due to birth and death,
then growth rate is constant and is given by,
R(t) = R(t
0
) = b −d
where b and d are birth and death rates, respectively.
If the initial population N(t
0
) = N
0
, we get the following initial value prob-
lem
dN
dt
= R(t
0
)N (1.1.3(a))
N(t
0
) = N
0
(1.1.3(b))
The solution of the above IVP is given by
N(t) = N
0
exp(R(t −t
0
)), t ∈ '
Although the above model may predict the population N(t) in the initial stages,
we realise that no population can grow exponentially for ever. In fact, as pop-
ulation grows sufficiently large, it begins to interact with its environment and
also other species and consequently growth rate R(t) diminishes. If we assume
the form of R(N) as equal to a−cN(a, c positive constant), we get the following
IVP
dN
dt
= N(a − cN) (1.1.4a)
N(t
0
) = N
0
(1.1.4b)
This is known as logistic equation with a as the growth rate without environ-
mental influences and c representing the effect of increase population density.
4CHAPTER1. INTRODUCTIONTOORDINARYDIFFERENTIAL EQUATIONS
The equilibrium points are 0 and a/c, C = a/c is called the carrying capacity of
the environment.
One can easily solve the above ODE by the method of separation of variables
and get
N(t) =
a
c

1 +

a
c
−N
0

N0
exp(−a(t −t
0
))
¸
¸
It follows from the above representation that lim
t→∞
N(t) = a/c = C (carrying
capacity). So the above solution has the representation
N(t) =
C
¸
1 +
(C −N
0
)
N
0
exp(−a(t −t
0
))
(1.1.5)
The solution curve of the above population model is as under.
2 4 6 8 10
t
2
4
6
8
No
C=a\c
Nt
Figure 1.1.1: Growth of population N(t)
The parameters a and C are usually not known but can be estimated by min-
imising the mean square deviation φ(a, C) between the known discrete population
data N
d
(t
m
) and theoretical data N(t
m
) :
φ(a, C) =
M
¸
m=1
[N(t
m
) −N
d
(t
m
)]
2
(1.1.6)
By using one of the algorithms of the unconstrained optimization techniques
(refer Algorith 2.3.1), we can actually carry out the computation of a and C
from the given data.
1.1. FIRST ORDER ORDINARY DIFFERENTIAL EQUATIONS 5
Example 1.1.2 Consider the following ODE
dx
dt
= −tx (1.1.7)
A general solution of the above ODE is given by
x(t) = ae

1
2
t
2
, a is any constant
The solution passing through (t
0
, x
0
) in (t, x) plane is given by
x(t) = x
0
e

1
2
(t
2
−t
2
0
)
(1.1.8)
The following graphics demonstrates the solution curve and the direction field
of the above ODE.
-2 -1 1 2
t
0.5
1
1.5
2
2.5
3
3.5
xt
Figure 1.1.2: Solution curves x(t)
6CHAPTER1. INTRODUCTIONTOORDINARYDIFFERENTIAL EQUATIONS
Figure 1.1.3: Direction field of the ODE
If IVP corresponding to Eq. (1.1.1) has a unique solution x(t, t
0
, x
0
) pass-
ing through (t
0
, x
0
), no two solution curves of the IVP can intersect. Thus
the uniqueness theorem, to be dealt with in the Chapter 3, will be of utmost
importance to us.
As we have seen through Example 1.1.2, the direction field for a given differential
equation can be an effective way to obtain qualitative information about the
behaviour of the system and hence it would be more appropriate to expand on
this concept.
To sketch the direction field for a given ODE in 1-dimension, a device called
isocline can be used.
Definition 1.1.6 Isocline is a curve in t − x plane through the direction field
along which p = f(t, x) is constant. The family of isoclines is then the family
of curves f(t, x) = p in t −x plane.
Example 1.1.3 We wish to find the direction field of the following ODE
1 +
dx
dt
1 −
dx
dt
= 2

t +x
Solving for
dx
dt
, we get
dx
dt
(1 + 2

t +x) = 2

t +x −1
which gives
dx
dt
=
2

t +x −1
2

t +x + 1
= p
1.2. CLASSIFICATION OF ODE 7
So iscolines are given by
2

t +x −1
2

t +x + 1
= p
That is
t +x =
1
2
¸
1 +p
1 −p

2
, −1 ≤ p < 1 (1.1.9)
Thus isoclines are all lines of slope -1 with −1 ≤ p < 1. So we get the following
graphics. The solution of the ODE passing through (t
0
, x
0
) is given by
(x −t) +

t +x = (x
0
−t
0
) +

t
0
+x
0
(1.1.10)
1 2 3 4
t
-3
-2
-1
1
2
3
4
xt
p=-1
p=0
p=0.2
p=0.33
p=0.5
Figure 1.1.4: Isoclines with solution curves
1.2 Classification of ODE
The ODE
dx
dt
= f(t, x) (1.2.1)
is called linear if f has the form
f(t, x) = A(t)x +b(t) (1.2.2)
where A(t) ∈ '
n·n
and b(t) ∈ '
n
for all t. A linear ODE is called homogeneous
if b(t) ≡ 0 and if A(t) = A (constant matrix), then we call it a linear ODE with
constant coefficients. Linear ODE will be discussed in detail in Chapter 4.
8CHAPTER1. INTRODUCTIONTOORDINARYDIFFERENTIAL EQUATIONS
The ODE - Eq. (1.2.1) is called autonomous if f does not depend on t. That is
ODE has the form
dx
dt
= f(x) (1.2.3)
For autonomous ODE we have the following proposition (refer Mattheij and
Molenaar [9]).
Proposition 1.2.1 If x(t) is a solution of Eq. (1.2.3) on the interval I =
(a, b), then for any s ∈ ', x(t + s) is a solution of Eq. (1.2.3) on the interval
(a − s, b − s). Hence the trajectory x(t) of Eq. (1.2.3) satisfies the following
property
x(t, t
0
, x
0
) = x(t −t
0
, 0, x
0
) for all t
0
∈ I
This implies that the solution is completely determined by the initial state ¯ x
0
at
t = 0.
A non-autonomous ODE - (1.2.1) is called periodic if
f(t +T, x) = f(t, x) for some T > 0 (1.2.4)
The smallest of such T is called its period. From Eq. (1.2.4), it follows that
the form of the solution is not affected if we shift t
0
to t
0
± nT, n ∈ ℵ (set of
natural numbers). That is
x (t ±nT, t
0
±nT, x
0
) = x(t, t
0
, x
0
)
It is not necessary that a solution of periodic ODE is periodic as we see in the
following example.
Example 1.2.1
dx
dt
=
¸
0 1
−1 0

x +
¸
0
2 cos t

f(t, x) = Ax + b(t), b(t) = (0, 2 cos t). f is periodic with period 2π but its
solution (to be discussed in Chapter 4) is given by
x(t) = (t −sin t, sint +t cos t)
However, we have the following theorem regarding the periodic solution of the
periodic ODE.
Proposition 1.2.2 If a solution of a periodic ODE is periodic. Then it has the
same period as the vector field.
Proof : Assume that f(t, x) has a period T and the solution x(t) has period
S with S = T. Because ¯ x is periodic, its derivative
˙
x(t) = f(t, x) will also be
periodic with period S as well T. That is
f(t +S, x) = f(t, x)
1.2. CLASSIFICATION OF ODE 9
Also, T periodicity of f implies that
f(t +S −nT, x) = f(t, x)
We choose n such that 0 < S − nT < T. It implies that f is periodic with a
period smaller than T, a contradiction. Hence T = S and the solution ¯ x(t) has
the same period as the vector field.
Example 1.2.2 (Predator-Prey Model)
Predator-prey model represents the interaction between two species in an envi-
ronment. For example, we shall focus on sharks and small fish in sea.
If the food resource of sharks (fish) is non-existent, sharks population exponen-
tially decays and is increased with the existence of fish. So, the growth rate of
sharks

1
S
dS
dt

is modelled as −k without the fish population and −k+λF with
fish population F. Thus
dS
dt
= S(−k +λF)
For the growth rate of fish, we note that it will decrease with the existence of
sharks and will flourish on small plankton (floating organism in sea) without
sharks. Thus
dF
dt
= F(a −cS)
Thus, we have, what is called the famous Lotka - Volterra model for the predator-
prey system
dF
dt
= F(a −cS) (1.2.4(a))
dS
dt
= S(−k +λF) (1.2.4(b))
This reduces to the following autonomous system of ODE in '
2
d¯ x
dt
= f(¯ x)
where
x = (F, S), f(¯ x) = (F(a −cS), S(−k +λF))
The equilibrium points are given by
ˆ
F = k/λ,
ˆ
S = a/c;
ˆ
F = 0,
ˆ
S = 0.
We shall have the discussion of the phase-plane analysis in the Chapter 5. How-
ever, if we linearize the system given by Eq.(1.2.4) around the equilibrium point
10CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
ˆ
F = k/λ,
ˆ
S = a/c, we get the following linear system (refer Section 2.3 for
linearization)
dx
dt
= Ax(t)
x = (F −
ˆ
F, S −
ˆ
S)
A =

0 −
kc
λ

c
0
¸
¸
¸ (1.2.5)
The solution of the IVP corresponding to Eq. (1.2.5) is given by
S =
ˆ
S +S
0
cos wt +

cw
F
0
sin wt (1.2.6a)
F =
ˆ
F +F
0
sinwt −
cw
a
S
0
cos wt (1.2.6b)
(Refer Section 4.3 for solution analysis of this linear system).
Thus, the solution is periodic with period

w
= 2π(ak)
−1/2
. The following graph-
ics depict the fish and shark population in t −F/S plane.
2 4 6 8 10 12
t
9.5
10
10.5
11
St
Ft
Figure 1.2.1: Fish and shark population with time
The average population of the general predator-prey system given by Eq.
(1.2.4) is obtained as follows.
We have
1
S
dS
dt
= −k +λF
This gives
ln
¸
S(t)
S(t
0

= −k(t −t
0
) +λ

t
t0
F(τ)dτ (1.2.7)
1.3. HIGHER ORDER ODE 11
In view of the information obtained from the linearized model, we can assume
that both S and F are periodic of period T = t
f
− f
0
. That is, S(t
f
) = S(t
0
)
and F(t
f
) = F(t
0
) and hence Eq. (1.2.7) gives
0 = −kT +λ

t0+T
t0
F(τ)dτ
This implies that
1
T
¸

t0+T
t0
F(τ)dτ
¸
= k/λ (1.2.8(a))
and similarly
1
T
¸

t0+T
t0
S(τ)dτ
¸
= a/c (1.2.8(b))
So in predator-prey system, no matter what the solution trajectory is, the average
population remains around the equilibrium point.
It is also interesting to analyse the men’s influence on the above ecosystem if we
fish both predator and prey. Then, we have
dF
dt
= F(a −cS) −σ
1
F
dS
dt
= S(−k +λF) −σ
2
S
This is equivalent to
dF
dt
= F(a
t
−cS)
dS
dt
= S(−k
t
+λF)
a
t
= a −σ
1
, k
t
= k +σ
2
Hence it follows that the average population of predator is
a
t
c
=
a −σ
1
c
(de-
creases) and that of prey is
k
t
c
=
k +σ
2
c
(increases).
1.3 Higher Order ODE
Assume that x(t) is a scalar valued n−times continuously differentiable function
on an interval I ⊂ '. Let us denote the derivative
d
k
x
dt
k
by x
(k)
(t).
We shall be interested in the following higher order ordinary differential equation
x
(n)
(t) = g(t, x(t), x
(1)
(t), . . . , x
(n−1)
(t)) (1.3.1)
where g : I ''. . . ' →' is a given mapping. We define a vector x(t)
with components x
i
(t), i = 1, . . . n by
x
i
(t) = x
(i−1)
(t), x
1
(t) = x(t); 2 ≤ i ≤ n (1.3.2)
12CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
and the vector field f(t, ¯ x) by
f(t, ¯ x) = (x
2
, x
3
, . . . , g(t, x
1
, . . . , x
n
))
Then the higher order ODE - Eq. (1.3.1) is equivalent to the following first
order ODE
dx
dt
= f(t, x)
A linear higher order ODE has the form
x
(n)
(t) +a
n−1
(t)x
(n−1)
(t) +. . . +a
0
(t)x(t) = b(t) (1.3.3)
where a
i
(t)(0 ≤ i ≤ n − 1) and b(t) are given functions. Then in view of Eq.
(1.3.2) - Eq. (1.3.3) we have
dx
1
dt
= x
2
(t)
dx
2
dt
= x
3
(t)
.
.
.
dx
n
dt
= b(t) −a
0
(t)x
1
(t) −a
1
(t)x
2
(t) − −a
n−1
(t)x
n
(t)
This is equivalent to the following first order system
dx
dt
= A(t)x(t) +b(t) (1.3.4)
where
A(t) =

0 1 0 . . . 0
0 0 1 . . . 0
.
.
.
0 0 0 . . . 1
−a
0
−a
1
−a
2
. . . −a
n−1
¸
¸
¸
¸
¸
¸
¸
b(t) = (0, 0, . . . , b(t))
The matrix A(t) is called the companion matrix.
Example 1.3.1 (Mechanical Oscillations)
A particle of mass m is attached by a spring to a fixed point as given in the
following diagram.
1.3. HIGHER ORDER ODE 13
a
m
F(t)
x
Figure 1.3.1: Mechanical oscillations
We assume that spring obeys Hook’s law (tension is proportional to its exten-
sion) and resistance (damping) is proportional to the particle speed. The external
force applied to the particle is F(t). By equilibrium of forces we have
mF = m
d
2
x
dt
2
+T +mk
dx
dt
By Hook’s Law, we get T =
λx
a
and hence
F =
d
2
x
dt
2
+
λ
ma
x +k
dx
dt
(1.3.5)
Equivalently, we get the following second order equation modelling the spring
problem
d
2
x
dt
2
+k
dx
dt

2
x = F(t), ω
2
=
λ
ma
(1.3.6)
Case 1: No resistance and no external force (Harmonic Oscillator)
d
2
x
dt
2

2
x = 0, ω
2
=
λ
ma
(1.3.7)
Eq. (1.3.7) is equivalent to the following system of first order differential equa-
tions
dx
dt
=
¸
0 1

λ
ma
0
¸
x (1.3.8)
It is easy to see the solution of the above system is given by (refer Section 4.4)
x(t) = x
0
cos ω(t −t
0
) +
x
0
ω
sin ω(t −t
0
). (1.3.9)
˙ x(t) = −x
0
ω sinω(t −t
0
) +x
0
cos ω(t −t
0
) (1.3.10)
14CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
Equivalently, we have
x(t) = αcos(ωt +β)
This is a simple harmonic motion with period

ω
and amplitude α. ω is called
its frequency.
We can easily draw plots of phase-space (normalized ω = 1) and also solution
curve of the ODE - Eq. (1.3.8).
-1 -0.5 0.5 1
-2
-1
1
2
Figure 1.3.2: Phase space of the linear system
-5
0
5
t
-1
-0.5
0
0.5
1
xt
-2
-1
0
1
2
x’t
-5
0
5
t
-2
-1
0
1
2
x’t
Figure 1.3.3: Solution curve of the linear system
Case 2: Solution with resistance
dx
dt
=
¸
0 1
−ω
2
−k

x(t) (1.3.11)
(i) k
2
−4ω
2
< 0
1.3. HIGHER ORDER ODE 15
The solution curve x(t) is a damped oscillation given by (refer Section 4.4)
x(t) = exp(
−kt
2
)(Acos bt +Bsin bt)
b =


2
−k
2
2
5 10 15 20 25
t
-1
-0.5
0.5
1
xt
Figure 1.3.4: Solution curve for Case2(i)
The oscillation curves for the other cases are as under.
(ii) (k
2
−4ω
2
) = 0
x(t) = Aexp(
−kt
2
) +Bt exp(
−kt
2
)
10 20 30 40 50 60
t
-1.5
-1
-0.5
0.5
1
xt
Figure 1.3.5: Solution surve for case2(ii)
16CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
(iii) k
2
−4ω
2
> 0
x(t) = exp(
−kt
2
)

A
1
e
ct
+A
2
e
−ct

c =

k
2
−4ω
2
1 2 3 4 5 6
t
5
10
15
20
25
30
35
xt
Figure 1.3.6: Solution curve for case2(iii)
Case 3: Effect of resistance with external force
dx(t)
dt
=
¸
0 1
−ω
2
−k

x +F(t), F(t) = (0, F(t))
The solution x(t) is given by
x = x
p
+x
c
where x
c
is the solution of the homogeneous equation with F(t) = 0. We have
seen that x
c
→ 0 as t → ∞. A particular solution (response function) x
p
(t) of
the above system is given by
x
p
(t) =
F
0
D
[cos (βt −φ)]
corresponding to the external force (input function) F(t) = F
0
cos βt.
The constant D is given by
D =


2
−β
2
)
2
+ k
2
β
2

1/2
and
sinφ =

D
, cos φ =
ω
2
−β
2
D
Thus the forced oscillations have the same time period as the applied force but
with a phase change φ and modified amplitude
F
0
D
=
F
0
[(ω
2
−β
2
)
2
+k
2
β
2
]
1/2
1.3. HIGHER ORDER ODE 17
Amplitude modification depends not only on the natural frequency and forcing
frequency but also on the damping coefficient k.
As k →0 we have
x
p

F
0

2
−β
2
)
→∞ as ω →β.
For k = 0, the response is given by
x
p
=
F
0

t sin wt
which implies that the system resonates with the same frequency but with
rapidly increasing magnitude, as we see in the following graph.
2 4 6 8 10 12 14
t
-2
2
4
Input
Response
Figure 1.3.7: Input-Response curve
Example 1.3.2 (Satellite Problem)
A satellite can be thought of a mass orbiting around the earth under inverse
square law field.
r
θ
u
m
u
2
1
Figure 1.3.8: Satellite in orbit around the Earth
18CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
We assume that m = 1 and also the satellite has thrusting capacity with radial
thrust u
1
and tangential thrust u
2
.
Equating forces in normal and tangential direction on the orbiting satellite, we
get
¸
d
2
r
dt
2
−r(t)(

dt
2
)
2

= −
k
r
2
(t)
+u
1
(t)
¸
r
d
2
θ
dt
2
+ 2

dt
dr
dt

= u
2
(t)
This gives a pair of second order differential equations
d
2
r
dt
2
= r(t)(

dt
)
2

k
r
2
(t)
+u
1
(t) (1.3.12a)
d
2
θ
dt
2
= −
2
r(t)

dt
dr
dt
+
u
2
(t)
r(t)
(1.3.12b)
] If u
1
= 0 = u
2
, then one can show that Eq. (1.3.12) has a solution given by
r(t) = σ, θ(t) = ωt (σ, ω are constant and σ
3
ω
2
= k). Make the following
change of variables:
x
1
= r −σ, x
2
= ˙ r, x
3
= σ(θ −ωt), x
4
= σ(
˙
θ −ω)
This gives
r = x
1
+σ, ˙ r = x
2
, θ =
x
3
σ
+ωt,
˙
θ =
x
4
σ

So Eq. (1.3.12) reduces to
dx
1
dt
= x
2
dx
2
dt
= (x
1
+σ)(
x
4
σ
+ω)
2

k
(x
1
+σ)
2
+u
1
(1.3.13)
dx
3
dt
= x
4
dx
4
dt
= −2σ(
x
4
σ
+ω)
x
2
(x
1
+σ)
+
u
2
σ
(x
1
+σ)
Eq. (1.3.13) is a system of nonlinear ordinary diferential equations involving
the forcing functions ( controls) u
1
and u
2
and can be written in the compact
form as
d¯ x
dt
= f(¯ x, ¯ u), x(t) ∈ '
4
, u(t) ∈ '
2
(1.3.14)
1.4. DIFFERENCE EQUATIONS 19
Here f is a vector function with components f
1
, f
2
, f
3
, f
4
given by
f
1
(x
1
, x
2
, x
3
, x
4
; u
1
, u
2
) = x
2
f
1
(x
1
, x
2
, x
3
, x
4
; u
1
, u
2
) = (x
1
+σ)(
x
4
σ
+ω)
2

k
(x
1
+σ)
2
+u
1
f
1
(x
1
, x
2
, x
3
, x
4
; u
1
, u
2
) = x
4
f
1
(x
1
, x
2
, x
3
, x
4
; u
1
, u
2
) = −2σ(
x
4
σ
+ω)
x
2
(x
1
+σ)
+
u
2
σ
(x
1
+σ)
We shall be interested in the solution, x(t) ∈ '
4
of linearized equation corre-
sponding to Eq. (1.3.14) in terms of the control vector u(t) ∈ '
2
.
1.4 Difference Equations
In some models (as we shall see subsequently), the state vector x(t) may not
depend on t continuously. Rather, x(t) takes values at discrete set of points
¦t
1
, t
2
, . . . , t
k
, ¦
In such a situation we use difference quotients instead of differential quotients
and that leads to difference equations. Suppose there exists a sequence of vector
fields f
i
(x) : Ω → '
n
, Ω ⊆ '
n
. Then the first order difference equation has the
form
x
i+1
= f
i
(x
i
), ¯ x
i
∈ Ω, i = 1, 2 . . . (1.4.1(a))
If in addition
x
0
= z
0
(a given vector in '
n
) (1.4.1(b))
then Eq. (1.4.1(a)) - Eq. (1.4.1(b)) is called the IVP corresponding to a differ-
ence equation.
As for ODE, we can define orbit or trajectory and solution curve as discrete
subsets of '
n
and '
n+1
, respectively in an analogous way. A stationary point
or equilibrium point of the difference equation Eq. (1.4.1) is a constant solution
x such that
x = f
i
(x), i ∈ ℵ
Eq. (1.4.1) is called linear if it has the form
x
i+1
(t) = A
i
x
i
+b
i
, i ∈ ℵ
where A
i
∈ '
n·n
and b
i
∈ '
n
.
For the scalar case n = 1, we have the linear difference equation
x
i+1
= a
i
x
i
+b
i
, i ∈ ℵ
Its solution is given by
x
i
= (Π
i−1
j=0
(a
j
))x
0
+
i−1
¸
j=0

i
l=j+1
a
l
)b
j
20CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
A k
th
order difference equation in 1-dimension is given by
x
i+1
= g
i
(x
i
, x
i+1
, . . . , x
i+1−k
), i = k −1, k, . . .
A linear k
th
order difference equation is given by
x
i+1
=
k
¸
j=1
a
ij
x
i−j+1
+ b
i
.
This can be written as the first order system
x
i+1
= A
i
x
i
+b
i
, i ≥ 0 (1.4.2)
where
¯ x
i
=

x
i
.
.
.
x
i+k−1
¸
¸
¸,
¯
b
i
=

0
.
.
.
0
b
i+k
¸
¸
¸
¸
¸
A
i
=

0 1
.
.
. 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 1
a
i+k−1,k
a
i+k−1,k−1
a
i+k−1,1
¸
¸
¸
¸
¸
¸
¸
¸
¸
For the homogeneous case b
i
= 0, with A
i
= A for all i, we get
x
i+1
= Ax
i
(1.4.3)
The solution of this system is of the form
x
i
= r
i
c
where scalar r and vector c are to be determined.
Plugging this representation in Eq. (1.4.3) we get
(A−rI)c = 0 (1.4.4)
This is an eigenvalue problem. If A has distinct eigenvalues λ
1
, λ
2
, . . . , λ
n
with
linearly independent eigenvectors c
1
, . . . c
n
, then the general solution of Eq.
(1.4.3) is given by
x
m
=
n
¸
i=1
d
i
λ
m
i
c
i
, m = 1, 2 . . . (1.4.5)
If the initial value is x
0
then d = [d
1
, . . . , d
n
] is given by d = C
−1
(x
0
). Here
C = [ ¯ c
1
, ¯ c
2
, , ¯ c
n
]. So the solution to the IVP corresponding to Eq. (1.4.3) is
completely given by Eq. (1.4.5).
1.5. DISCRETIZATION 21
Example 1.4.1 (Discrete One Species Population Dynamics Model)
We wish to measure population changes in a species with one year age distri-
bution. Let N
i
(t) denote the number of i
th
year old species and let b
i
, d
i
the
corresponding birth and death rates, respectively (0 ≤ i ≤ M). Then we have
N
0
(t +´t) = b
0
N
0
(t) +b
1
N
1
(t)
+. . . +b
M
N
M
(t), (´t = 1) (1.4.6(a))
N
i+1
(t +´t) = (1 −d
i
)N
i
(t), 0 ≤ i ≤ M −1 (1.4.6(b))
Let N(t) = (N
0
(t), N
1
(t), . . . , N
M
(t)). Then Eq. (1.4.6) becomes
N(t +´t) = AN(t) (1.4.7)
where
A =

b
0
b
1
b
M
1 −d
0
0 0
0 1 −d
1

.
.
.
.
.
.
0 0 1 −d
M−1
¸
¸
¸
¸
¸
¸
¸
If we denote N
m
= N(m´t), then Eq. (1.4.7) can be viewed as a difference
equation of the form
N
m+1
= AN
m
(1.4.8)
This is of the type given by Eq. (1.4.3) and hence the solution of the above
equation is of the form
N
m
=
M
¸
i=1
a
i
λ
m
i
φ
i
where ¦λ
i
¦
M
i=1
and ¦φ
i
¦
M
i=1
are eigenvalues (distinct) and linearly independent
eigenvectors of A, respectively. a = [a
1
, . . . , a
m
] is a constant vector. If the
initial population N
0
is known, the constant vector a is given by a = CN
0
where the nonsingular matrix C is given by
C = [φ
1
, φ
2
, . . . , φ
M
]
It is to be noticed that the population of each age group grows and decays de-
pending upon the sign of the eigenvalues λ
1
, λ
2
, . . . , λ
M
.
1.5 Discretization
In this section we shall describe some well-known one-step numerical methods,
which are used to solve ODE. However, we shall not be concerned with the
stability and convergence aspects of these methods.
We also introduce to reader the concept of neural solution of ODE. This is also
based on discretization but uses neural network approach (refer Haykin [6]).
22CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
It is interesting to observe that sometimes neural solution is a better way of
approximating the solution of ODE.
Let us begin with scalar ordinary differential equation
dx
dt
= f(t, x) (1.5.1)
corresponding to the scalar field f(t, x) : I ' →'.
Let t
0
, t
1
, . . . , t
N
be a set of time points (called grid-points) wherein we would
like to approximate the solution values x(t
i
), 0 ≤ i ≤ N.
Integrating the above equation on [t
i
, t
i
+ 1] we get
x(t
i+1
) = x(t
i
) +

ti+1
ti
f(s, x(s))ds
If f is approximated by its values at t
i
, we get the Euler’s forward method
x(t
i+1
) = x(t
i
) +hf(t
i
, x(t
i
)), h = x
i+1
−x
i
(1.5.2)
which is the so-called explicit, one-step method.
On the other hand, if f is approximated by its values at t
i+1
we get Euler’s
backward method
x(t
i+1
) = x(t
i
) +hf(t
i+1
, x(t
i+1
)) (1.5.3)
which has to be solved implicitly for x
ti+1
.
If the integral

ti+1
ti
f(s, x(s))ds is approximated by the trapezoidal rule, we
get
x(t
i+1
) = x(t
i
) +
1
2
h[f(t
i
, x(t
i
)) +f(t
i+1
, x(t
i+1
))] (1.5.4)
which is again implicit.
Combining Eq. (1.5.2) and Eq. (1.5.4) to eliminate x(t
i+1
), we get the Heun’s
method (refer Mattheij and Molenaar[9])
x(t
i+1
) = x(t
i
) +
1
2
h[f(t
i
, x(t
i
)) +f(t
i+1
, x(t
i
) +hf(t
i
, x(t
i
)))] (1.5.5)
An important class of one-step numerical methods is Runge-Kutta method.
Consider the integral equation
x(T) = x(t) +

T
t
f(s, x(s))ds (1.5.6)
Approximating the integral

ti+1
ti
f(s, x(s))ds by a general quadrature formula
m
¸
j=1
β
j
f(t
ij
, x(t
ij
)), we get
x(t
i+1
) = x(t
i
) +h
m
¸
j=1
β
j
f(t
ij
, x(t
ij
)) (1.5.7)
1.5. DISCRETIZATION 23
where h = t
i+1
− t
i
and t
ij
= t
i
+ ρ
j
h(0 ≤ ρ
j
≤ 1) are nodes on the interval
[t
i
, t
i+1
].
As x(t
ij
) are unknowns, to find them we apply another quadrature formula for
the integral

tij
ti
f(s, x(s))ds to get
x(t
ij
) = x(t
i
) +h
m
¸
l=1
r
jl
f(t
il
, x(t
il
)) (1.5.8)
Combining Eq. (1.5.7) and Eq. (1.5.8), we get the Runge-Kutta formula
x(t
i+1
) = x(t
i
) +h
m
¸
j=1
β
j
k
j
(1.5.9)
k
j
= f(t
i

j
h, x(t
i
) +h
m
¸
l=1
r
jl
k
l
)
Eq. (1.5.9) is explicit if r
jl
= 0, l ≥ j. If ρ
1
= 0, ρ
2
= 1, r
11
= 0 = r
12
, r
21
=
1, r
22
= 0 and β
1
= 1/2 = β
2
, we get the Heun’s formula given by Eq. (1.5.5).
Also, it is easy to get the following classical Runge-Kutta method from Eq.
(1.5.9)
x(t
i+1
) = x(t
i
) +h
¸
1
6
k
1
+
1
3
k
2
+
1
3
k
3
+
1
6
k
4

(1.5.10)
k
1
= f(t
i
, x(t
i
))
k
2
= f(t
i
+
1
2
h, x(t
i
) +
1
2
hk
1
)
k
3
= f(t
i
+
1
2
h, x(t
i
) +
1
2
hk
2
)
k
4
= f(t
i
+h, x(t
i
) +hk
3
)
In case we have f(t, x) as a vector field, we obtain a similar Runge - Kutta
method, with scalars x(t
i
), x(t
i
+ 1) being replaced by vectors x(t
i
), x(t
i+1
),
and scalars k
i
being replaced by vectors k
i
(1 ≤ i ≤ 4).
We now briefly discuss the concept of neural solution of the IVP associated
with Eq. (1.5.1). A neural solution (refer Logaris et al [8]) x
N
(t) can be written
as
x
N
(t) = x
0
+tN(t, W) (1.5.11)
where N(t, W) is the output of a feed forward neural network (refer Haykin [
6 ]) with one input unit t and network weight vector W. For a given input t,
output of the network is N =
¸
k
i=1
v
i
σ(z
i
) where z
i
= w
i
t −u
i
, w
i
denotes the
weight from the input unit t to the hidden unit i, v
i
denotes the weight from
24CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
x
N
u
1
w
2
w
3
w
k
v
v
v
1
2
3
k
Figure 1.5.1: Neural network with one input
the hidden unit i to the output and u
i
denotes the bias of the unit i, σ(z) is
the sigmoid transfer function
σ(z) =
1
1 +e
−z
and k is the number of hidden neurons.
We first determine the network parameters w
i
, v
i
and u
i
in such a manner
that x
N
(t) satisfies Eq. (1.5.1) in some sense. For this we discretize the interval
[t
0
, t
f
] as t
0
< t
i
< . . . < t
m
= t
f
.
As x
N
(t) is assumed to be a trial solution, it may not be exactly equal to
f(t
j
, x
N
(t
j
)) at point t
j
, 0 ≤ j ≤ m. Hence we train the parameter vector W
in such a way that it minimizes the error function φ(W) given as
φ(W) =
m
¸
j=0

dx
N
dt
[
t=tj
−f(t
j
, x
N
(t
j
))

2
One can use any unconstrained optimization algorithms (refer Joshi and Kannan
[ 7 ]) to minimize φ w.r.t. the neural network parameters to obtain their optimal
values u

i
, v

i
, w

i
to get the neutral solution x
N
(t) given by Eq. (1.5.11)
This methodology will be clear from the following example, wherein we com-
pute numerical solutions by Euler’s and Runge-Kutta methods and also neural
solution. Further, we compare these solutions with the exact solution.
Example 1.5.1 Solve the IVP
dx
dt
= x −x
2
, t ∈ [0, 1] (1.5.12)
x(0) =
1
2
The exact solution of this ODE (Bernoulli’s differential equation) is given by
x(t) =
1
1 +e
−t
(1.5.13)
(Refer Section 6 of this chapter).
1.5. DISCRETIZATION 25
For neural solution we take three data points t
j
= 0, 0.5, 1 and three hidden
neurons with a single hidden layer network. The neural solution is of the form
x
N
(t) =
1
2
+t
3
¸
i=1
v
i

1 +e
(−wit+ui)
(1.5.14)
The weight vector W = (w
1
, w
2
, w
3
, v
1
, v
2
, v
3
, u
1
, u
2
, u
3
) is computed in such a
way that we minimize φ(W) given by
φ(W) =
3
¸
j=1
¸
dx
N
(t)
dt
[
t=tj
−f(t
j
, x
N
(t
j
)

2
=
3
¸
j=1
¸
1
2
+
v
1
1 +e
u1−w1tj
+
v
2
1 +e
u2−w2tj
+
v
3
1 +e
u3−w3tj

v
1
1 +e
u1−w1tj
+
v
2
1 +e
u2−w2tj
+
v
3
1 + e
u3−w3tj

t
j
+

e
u1−w1tj
v
1
w
1
(1 +e
u1−w1tj
)
2
+
e
u2−w2tj
v
2
w
2
(1 +e
u2−w2tj
)
2
+
e
u3−w3tj
v
3
w
3
(1 +e
u3−w3tj
)
2

t
j
+

1
2
+

v
1
1 +e
u1−w1tj
+
v
2
1 +e
u2−w2tj
+
v
3
1 + e
u3−v3tj

t
j

2
¸
2
We have used the steepest descent algorithm (refer to Algorithm 2.3.1 ) to
compute the optimal weight vector W . The neural solution is given by Eq.
(1.5.14). We note that the neural solution is a continuous one. We can compare
the values of this solution at a discrete set of points with the ones obtained by
Euler and Runge-Kutta method ( refer Conte and deBoor [3]). The following
tables give the distinction between the three types of approximate solutions.
Table 1.5.1: Euler solution and actual solution
Euler’s method Actual solution Absolute difference
0.525, 0.52498 0.00002
0.54994 0.54983 0.00010
0.57469 0.57444 0.00025
0.59913 0.59869 0.00044
0.62315 0.62246 0.00069
0.64663 0.64566 0.00097
0.66948 0.66819 0.00129
0.69161 0.68997 0.00163
0.71294 0.71095 0.00199
0.73340 0.73106 0.00234
26CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
Table 1.5.2: Runge-Kutta solution and actual solution
Runge Kutta method Actual solution Absolute difference
0.52498 0.52498 1.30339
−9
0.54983 0.54983 2.64162
−9
0.57444 0.57444 4.05919
−9
0.59869 0.59869 5.59648
−9
0.62246 0.62246 7.28708
−9
0.64566 0.64566 9.15537
−9
0.66819 0.66819 1.12147
−8
0.68997 0.68997 1.34665
−8
0.71095 0.71095 1.58997
−8
0.73106 0.73106 1.84916
−8
Table 1.5.3: Neural solution and actual solution
Neural Network sol Actual sol Absolute Difference
0.52511 0.52498 0.00013
0.54973 0.54983 0.00011
0.57383 0.57444 0.00062
0.59741 0.59869 0.00127
0.62048 0.62246 0.00198
0.64302 0.64566 0.00264
0.66503 0.66819 0.00316
0.68650 0.68997 0.00348
0.70743 0.71095 0.00352
0.72783 0.73106 0.003231
1.6 Techniques for Solving First Order Equa-
tions
The general form of a first order linear differential equation is
a(t)
dx
dt
+b(t)x +c(t) = 0 (1.6.1)
where a(t), b(t) and c(t) are continuous functions in a given interval with a(t) =
0. Dividing by a(t), we get the equivalent equation in normal form
dx
dt
+P(t)x = Q(t) (1.6.2)
1.6. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS 27
This equation is solved by multiplying Eq. (1.6.2) by e
R
P(t)dt
and integrating,
to give us the solution
x(t) = e

R
P(t)dt
¸
e
R
P(t)dt
Q(t)dt +c

(1.6.3)
Exact Equations
The differential equation
M(t, x)dt + N(t, x)dx = 0 (1.6.4)
or
M(t, x) +N(t, x)
dx
dt
= 0 = M(t, x)
dt
dx
+N(t, x)
is called exact if there is a differentiable function u(t, x) such that
∂u
∂t
= M,
∂u
∂x
= N (1.6.5)
If Eq. (1.6.5) holds, we have
M(t, x)dt + N(t, x)dx =
∂u
∂t
dt +
∂u
∂x
dx
= d(u(t, x))
Hence, intergrating the above wquation we get the solution
u(t, x) = C
of the exact equation given by Eq. (1.6.4).
The following theorem gives a test for the exactness of a differential equation.
Theorem 1.6.1 If M and N are continuously differentiable function of (t, x) ∈
D ⊆ '
2
( with no hole in D), then the differential equation
M(t, x)dt +N(t, x)dx = 0 is exact iff
∂M
∂x
=
∂N
∂t
in D (1.6.6)
Proof : If the differential equation is exact, we have
∂M
∂x
=

∂x
¸
∂u
∂t

and
∂N
∂t
=

∂t
¸
∂u
∂x

. Since M and N are continuously differentiable in D, it follows
that

2
u
∂x∂t
=

2
u
∂t∂x
and hence
∂M
∂x
=
∂N
∂t
in D.
Conversly, let Eq. (1.6.6) holds.
28CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
(t,x)
τ
y
∆ (t+ t ,x)
0
(t ,x )
0
Figure 1.6.1: A curve in τ −y plane
We explicitily define the function u(t, x) as the solution of the initial value
problem
∂u
∂t
= M(t, x),
∂u
∂x
= N(t, x)
u(t
0
, x
0
) = u
0
Then, u(t, x) is given by
u(t, x) = u
0
+

(t,x)
(t0,x0)
[M(τ, y)dτ +N(τ, y)dy]
where the integral on the RHS is a line integral along a path joining (t
0
, x
0
)
and (t, x). It can be shown that this line integral is independent of path if Eq.
(1.6.6) holds.
This gives
u(t +´t, x) −u(t, x) =

(t+Zt,x)
(t,x)
[M(τ, y)dτ +N(τ, y)dy]
=

t+Zt
t
M(τ, y)dτ (as y = x and dy = 0)
= M(τ
1
, x)´t, t < τ
1
< t +´t
(by meanvalue theorem for integrals).
Hence
∂u
∂t
= lim
Zt→0
¸
u(t +´t, x) −u(t, x)
´t

= lim
Zt→0
[M(τ
1
, x)] = M(t, x)
1.6. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS 29
Similarly, we have
∂u
∂x
= N(t, x)
It may happen that the equation M(t, x)dt +N(t, x)dx = 0 is not exact but
after multiplying both sides of this equation by a function µ(t, x), it becomes
exact. Such a function is called an integrating factor. For example, xdy −ydx =
(x
2
+y
2
)dx is not exact. But dividing by (x
2
+y
2
), we get
xdy −ydx
x
2
+y
2
= dx
which is equivalent to
x
x
2
dy −
y
x
2
dx
1 +
y
2
x
2
= dx
This is now an exact differential equation with solution
tan
−1
(
y
x
) = x +c
Change of Variables
There are a few common types of equations, wherein substitutions suggest them-
selves, as we see below
[A]
dx
dt
= f(at +bx), a, b ∈ ' (1.6.7)
We introduce the change of variable X = at +bx, which gives
x =
1
b
(X −at),
dx
dt
=
1
b
(
dX
dt
−a)
and hence Eq. (1.6.7) becomes
1
b
(
dX
dt
−a) = f(X)
or
dX
dt
= a +bf(X)
which can be easily solved.
[B]
dx
dt
= f(
x
t
) (1.6.8)
30CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
in which RHS depends only on the ratio
x
t
.
Introduce the change of variable u =
x
t
. Thus
x = ut,
dx
dt
= u +t
du
dt
Then Eq. (1.6.8) becomes
u +t
du
dt
= f(u)
or
du
dt
=
f(u) −u
t
which can be easily solved.
[C] Consider the equation
dx
dt
= f

at +bx +p
ct +dx +q

, a, b, c, d, p, q, ∈ ' (1.6.9)
in which RHS depends only on the ratio of the linear expression.
We substitute T = t −h, X = x −k.
Then
dx
dt
=
dX
dt
where we choose h and k such that
ah +bk +p = 0 = ch + dk +q (1.6.10)
Then
dX
dT
= f

at +bX
ct +dX

which is of type [B].
Eq. (1.6.10) can always be solved for h, k except when ad − bc = 0. In
that case we have cx +dy = m(ax +by) and hence Eq. (1.6.9) becomes
dx
dt
= f
¸
at +bx +p
m(at +bx) +q

which is of the type [A].
[D] The equation
dx
dt
+P(t)x = Q(t)x
m
(1.6.11)
in which the exponent is not necessarily an integer, is called Bernoulli’s
equation. Assume that m = 1. Then introduce the change of variable
1.6. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS 31
X = x
1−m
. Then x = X
(
1
1−m
)
,
dx
dt
=
1
1 −m
X
(
m
1−m
)
dX
dt
. Hence Eq.
(1.6.11) becomes
1
1 −m
X
(
m
1−m
)
dX
dt
+P(t)X
(
1
1−m
)
= Q(t)X
(
m
1−m
)
Equivalently
dX
dt
+ (1 −m)P(t)X = (1 −m)Q(t)
which is a linear differential equation in X.
Example 1.6.1 Solve
(t +x
2
)dx + (x −t
2
)dt = 0
M(t, x) = x −t
2
, N(t, x) = t +x
2
∂M
∂x
= 1 =
∂N
∂t
and hence this equation is exact.
So we write the solution u(t, x) as
u(t, x) = u
0
+

(t,x)
(t0,x0)

(τ −y
2
)dτ + (τ +y
2
)dy

(0,0)
C
2
C
1
(t, ) x
τ
y
Figure 1.6.2: Path from (0, 0) to (t, x)
In the integrand on RHS, we go along the path C
1
, C
2
. Hence the integral
become

t
0
(−τ
2
)dτ +

x
0
(t +y
2
)dy = −
t
3
3
+tx +
x
3
3
32CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
Hence, the solution is
−t
2
3
+tx +
x
3
3
= C
Example 1.6.2 Solve
dx
dt
=
t +x −1
t + 4x + 2
Solve h+k−1 = 0 = h+4k+2. This gives h = 2, k = −1. Hence t = T +2, x =
X −1 and
dX
dT
=
T +X
T + 4X
.
We now make substitution X = TU, to get
T
dU
dT
=
1 −4U
2
1 + 4U
which is equivalent to
1 + 4U
1 −4U
2
dU =
dT
T
This yields
(1 + 2U)(1 −2U)
3
T
4
= C
or
(T + 2X)(T −2X)
3
= C
or
(t + 2x)(t −2x −4)
3
= C
Example 1.6.3 Solve
dx
dt
+tx = t
3
x
4
This is a Bernoulli’s equation. We make the change of variable X = x
−3
and
get an equivalent linear differential equation
dx
dt
−3tX = −3t
3
The general solution of this equation is given by
X(t) = e
3
R
tdt
¸
(3t
3
)e

R
3tdt
+ C

= e
3
2
t
2
¸
(−3t
3
)e

3
2
t
2
dt +c

1.6. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS 33
To compute I =

(−3t
3
)e

3
2
t
2
dt +c, put t
2
= u, 2tdt = du. This gives
I =

e

3
2
u
¸

3
2
udu

= −
3
2
¸

2
3
e

3
2
u
u +
2
3

e

3
2
u
du

= −
3
2
¸

2
3
e

3
2
u
u −
2
3
2
3
e

3
2
u

= e

3
2
u
u +
2
3
e

3
2
u
= e

3
2
t
2
t
2
+
2
3
e

3
2
t
2
and hence
X(t) = e

3t
2
2
¸
e

3
2
t
2
t
2
+
2
3
e

3
2
t
2
+c

=
¸
t
2
+
2
3
+ce
3
2
t
2

Equivalently
x =
1
X
3
=
1

t
2
+
2
3
+ce
3
2
t
2

3
Example 1.6.4 Solve the IVP
dx
dt
= x −x
2
, x(0) =
1
2
We make the change of variable X =
1
x
and get an equivalent linear equation
dX
dt
+X = 1, X(0) = 2
The unique solution X(t) is given by
X(t) = e
−t
¸
e
t
dt +t

= e
−t
+ 1
This gives
x(t) =
1
1 +e
−t
For more on techniques for solving first order differential equation refer Golomb
and Shanks[4].
For more on real life problems giving rise to mathematical models generated by
ordinary differential equations, refer to Braun et al[1], Burghe and Borrie[2] and
Haberman[5].
34CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
1.7 Exercises
1. Sketch the direction field for the following differential equations.
(a)

dx
dt

2
= x
2
+t
2
(b) t
dx
dt
−x =

dx
dt

2
−1
(c) x
2
−(1 −t
2
)

dx
dt

2
= 0
Sketch the solution x(t), wherever possible.
2. Consider the differential equation
dx
dt
= f(t, x(t))
f(t, x) =

2tx
t
2
+x
2
, (t, x) = (0, 0)
0, (t, x) = (0, 0)
(a) Sketch the direction field of the above equation.
(b) Solve the above equation.
(c) Find the number of solutions passing through (0, 0) and (x
0
, y
0
) =
(0, 0).
3. Consider the differential equation
dx
dt
= (x
2
−1)t
p
, p ≥ 0, t ≥ 0
(a) Determine the equilibrium points.
(b) Sketch the vector field.
(c) Find solution passing through (0, −2), (0, 0) and (0, 2).
4. Consider the linear system in '
2
given by
dx
1
dt
= (x
2
+ 1) cos at
dx
2
dt
= x
1
+x
2
(a) Find the stationary points.
(b) Sketch the vector field in '
2
.
5. Solve following initial value problems.
1.7. EXERCISES 35
(a) t
2
dx
dt
+x(t + 1) = t, x(−1) = −2.
(b) (1 +t
2
)

tan
−1
t

dx
dt
+x = t, x(−1) = −1
(c) 2(1 −t
2
)
dx
dt
−2tx = a(t
3
−t), x(0) =
a
3
6. Test if the following equations are exact. If so, solve them.
(a) sin t sin
2
xdt −(cos t cos 2xtan x + cos t tan x) dx = 0.
(b)
¸
xlog(t
2
+x
2
) +
2t
2
x
t
2
+x
2
+x

dx +
¸
t +log(t
2
+x
2
) +
2t
2
x
t
2
+x
2
+t

dx = 0
(c)
¸
(t +x)
2
1 + (t +x)
2

dt −
¸
1
1 + (t +x)
2
−1

dx = 0
7. Using the line integrals, solve the following exact equations which pass
through the given point.
(a) [2t +exp(t) sin x] dt +exp(t) cos xdx = 0, x(0) =
π
2
(b) (cos t cos x) dx −(sin t sinx) dx = 0, x(
π
2
) =
π
3
(c)
¸
(t +x)
t
2
+x
2

dt −
¸
t −x
t
2
+x
2

dx = 0, x(1) = 1
8. Solve the following Bernoulli equations.
(a)
dx
dt
+x = tx
3/2
(b) 2tx
3
dx
dt
+t
4
−x
4
= 0
(c) 3
dx
dt
+t sec x = x
4
(1 −sint)
2
9. Find all curves with the property that the angle of inclination of its tangent
line at any point is three times the inclination of the radius vector to that
point from the orgin.
10. Find the fimily of curves orthogonal to the family of cardiods r = λ(1 +
cos θ) in polar cordinates (r, θ).
11. In a model of epidemics, a single infected individual is introduced in to a
community containing n individuals susceptible to the desease. Let x(t)
denote the number of uninfected individuals in the population at time t.
If we assume that the infection spreads to all those susceptible, then x(t)
decreases from x(0) = n to x(0) = 0. Give a mathematical model for the
above problem and solve it.
36CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
References
[1 ] Martin Braun, Courtney S Coleman and Donald Drew, Differential Equa-
tions Model, Ellis Horwood Ltd., 1981.
[2 ] D. N. Burghe and M S Borrie, Modelling with Differential Equation,
Springer Verlag, 1983.
[3 ] S. D. Conte and C deBoor, Elementary Numerical Analysis- An Algo-
rithmic Approach, McGraw Hill Book Co., 1981.
[4 ] Michael Golomb and Merrill Shanks, Elements of Ordinary Differneial
Equations, McGraw Hill Book Company, 1965.
[5 ] Richard Haberman, Mathematical Models: Mechanical Vibrations, Pop-
ulation Dynamics and Traffic Flow, Prentice Hall, 1977.
[6 ] S. Haykin, Neural Networks, Macmillan College Publishing Company,
Boston, 1994.
[7 ] Mohan C Joshi and Kannan S Moudgalya, Optimization - Theory and
Practice, Narosa Publishing House Pvt. Ltd, 2004.
[8 ] I. E. Logaris, A. Likas and D. I. Fotiadis, Artificial Neural Networks for
Solving Ordinary and Partial Differential Equations, IEEE Trans. Neural
Networks, 9(5), 987-1000, 1998.
[9 ] R.M.M Mattheij and J Molenaar, Ordinary Differential Equations in
Theory and Practice, John Wiley and Sons, 1996.
Chapter 2
Mathematical Preliminaries
This chapter is meant for laying the foundation for abstract concepts needed for
our subsequent discussion of ODE in function spaces.
We start with vector spaces and go on to define normed and inner product
spaces and then dwell on notions like Schwartz inequality, Bessel’s inequality,
Parseval’s equality, orthormal system and Fourier series.
Bounded linear and nonlinear operators in function spaces along with their
much needed properties - Lipscntiz continuity, monotonicity, differentiability
and compactness - are given a fair treatment.
We close by defining Dirac-delta function as a distribution function. This ap-
proach is essential for introducing Green’s function in Chapter 7.
2.1 Finite and Infinite Dimensional Spaces
Definition 2.1.1 Let X be a nonempty set which is closed with respect to binary
operation x + y (x, y ∈ X) and scalar multiplication αx (α ∈ ', x ∈ X). X is
said to be a vector space if the following holds.
1. For any x, y, z ∈ X
(a) (x +y) +z = x + (y +z)
(b) (x +y) = (y +x)
(c) ∃ an element 0 ∈ X such that x + 0 = x
(d) ∃ −x ∈ X such that x + (−x) = 0
2. For any α, β ∈ ' and x, y ∈ X
(a) α(x +y) = αx +αy
(b) (α +β)x = αx +βx
(c) α(βx) = (αβ)x
(d) 1x = x
37
38 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Definition 2.1.2 An inner product (x, y) in a vector space X is a function on
X X with values in ' such that the following holds.
1. (x, x) ≥ 0 forall x ∈ X and equality holds iff x =
¯
0
2. (x, y) = (y, x) for all x, y ∈ X
3. (αx +βy, z) = α(x, z) +β(y, z); α, β ∈ ', x, y, z ∈ X
The vector space X with an inner product defined is called inner product
space.
In an inner product space X, x is said to be orthogonal to y if (x, y) = 0.
This is denoted by x ⊥ y.
Definition 2.1.3 A vector space X is said to be a normed space if there exists
a function |x| from X to ' such that the following properties hold.
1. |x| ≥ 0 for all x ∈ X and |x| = 0 iff x = 0
2. |αx| = [α[|x| for all x ∈ X and α ∈ '
3. |x +y| ≤ |x| +|y| for all x, y ∈ X
In an inner product space, the induced norm is defined by
|x|
2
= (x, x), x ∈ X
In a normed space X, the induced metric d(x, y) (distance between two vectors)
is defined as
d(x, y) = |x −y| ∀ x, y ∈ X
In view of Definition 2.1.3, this metric d(x, y) satisfies the following properties.
[P
1
] d(x, y) ≥ 0 for all x, y ∈ X and d(x, y) = 0 iff x = y
[P
2
] d(x, y) = d(y, x) for all x, y ∈ X
[P
3
] d(x, z) ≤ d(x, y) +d(y, z) for all x, y, z ∈ X
It is possible to define a metric d(x, y) satisfying properties [P
1
] −[P
3
] in any
set without having a vector space structure. Such a set is called a metric space.
Definition 2.1.4 A metric space X is a non-empty set with a metrix d : X
X →' defined on it, which satisfies the properties [P
1
] −[P
3
].
In a metric space, without the vector space structure, we can easily define
the concepts of convergence, Cauchy convergence, completeness etc., as we see
below.
Definition 2.1.5 A sequence ¦x
k
¦ in a metrix space (X, d) is said to be con-
vergent to x ∈ X if d(x
k
, x) → 0 as k → ∞. ¦x
k
¦ is said to be Cauchy if
d(x
k
, x

) → as k, l →∞.
2.1. FINITE AND INFINITE DIMENSIONAL SPACES 39
Definition 2.1.6 A metrix space (X, d) is said to be complete if everyCauchy
sequence in (X, d) converges.
A complete normed space is called a Banach space, whereas a complete inner
product space is called a Hilbert space.
In a normed space X, it is possible to define infinite sum
¸

i=1
x
i
. We say
that s =
¸

i=1
x
i
iff the sequence of partial sums s
n
=
¸
n
i=1
x
i
∈ X, converges
to s ∈ X.
Definition 2.1.7 A set S of vectors ¦¯ x
1
, ¯ x
2
, , ¯ x
n
¦ in a vector space X is
said to be linearly dependent (l.d.) if ∃ scalars α
i
∈ ' (not all zero) such that
α
1
x
1

2
x
2
+. . . +α
n
x
n
= 0.
Otherwise, this set is called linearly independent (.i.).
If the set S consists of infinite elements of the spaces X. Then S is said to
be .i. if every finite subset of S is .i.
Definition 2.1.8 A set of vectors S = ¦x
1
, x
2
, . . . , x
n
¦ in a vector space X is
said to be a basis for X if
1. S is .i. and
2. L[S] = ¦α
1
x
1
+ α
2
x
2
+ + α
n
x
n
; α
i
∈ '¦, called the linear span of S,
equals X.
The number of elements in a basis is unique (refer Bapat[1]).
Definition 2.1.9 A vector space X is said to be finite dimensional if it has
a basis consisting of finite number of elements and this number is called its
dimension and is denoted by dimX.
A vector space X which is not finite dimensional, is called infinite dimensional
space.
Example 2.1.1 In the space '
n
of n-tuples, the Euclidean norm and inner
product are defined as
|x|
2
= x
2
1
+x
2
2
+. . . +x
2
n
, x = (x
1
, x
2
, . . . , x
n
) (2.1.1(a))
(x, y) = x
1
y
1
+x
2
y
2
+. . . x
n
y
n
, x = (x
1
, x
2
, . . . , x
n
), y = (y
1
, y
2
, . . . , y
n
)
(2.1.1(b))
In terms of the matrix notation we can represent (x, y) as
(x, y) = y
¯
x = x
¯
y
if we treat x, y as column vectors and x
¯
, y
¯
as row vectors.
One can show that '
n
is complete with respect to the norm induced by the inner
product defined by Eq. (2.1.1(b)) and hence it is a Hilbert space.
40 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Remark 2.1.1 It is possible to define other norms in the space '
n
, as we see
below
|x|
1
= [x
1
[ +[x
2
[ +. . . +[x
n
[
|x|

= max
1≤i≤n
¦[x
i

Definition 2.1.10 Two different norms |x|
a
and |x|
b
in a normed space X
are said to be equivalent if ∃ positive constants α, β such that
α|x|
a
≤ |x|
b
≤ β|x|
a
One can show that all norms in '
n
are equivalent (refer Limaye[8]) and hence
'
n
is also complete w.r.t. |x|
1
and |x|

, defined earlier.
Example 2.1.2 In the space C[0, 1] of all real valued continuous functions, one
can define norm and inner product as under
|f|
2
2
=

1
0
f
2
(t)dt (2.1.2(a))
(f, g) =

1
0
f(t)g(t)dt (2.1.2(b))
One can show that the space C[0, 1] w.r.t. the above norm is not complete and
hence it is not a Banach space. To see this consider the sequence ¦f
n
(t)¦ of
continuous functions on [0, 1] defined as
f
n
(t) = t
n
, t ∈ [0, 1]
0.2 0.4 0.6 0.8 1
0.2
0.4
0.6
0.8
1
f1
f2
f3
f4
Figure 2.1.1: A sequence which is Cauchy but not complete
|f
n
−f
m
|
2
2
=

1
0
(t
n
−t
m
)
2
dt →0 as n, m →∞
2.1. FINITE AND INFINITE DIMENSIONAL SPACES 41
and hence ¦f
n
¦ is Cauchy but f
n
(t) converges to f(t) defined as
f(t) =

1, t = 1
0, t = 1
which is not a continuous function and hence the space is C[0, 1] is not complete
w.r.t. the above norm.
However, if |f|

is defined as |f|

= sup
t∈[0,1]
[f(t)[, then C[0, 1] is complete
with respect to |f|

norm and hence it is a Banach space.
Example 2.1.3 Let L
2
[a, b] denote the space of all square integrable functions.
L
2
[a, b] = ¦f : [a, b] →',

b
a
f
2
(t)dt < ∞¦
Let the norm and inner product be defined as in Eq. (2.1.2). Then L
2
[a, b]
is a Hilbert space.
Remark 2.1.2 '
n
is finite dimensional, where as C[a, b) and L
2
[a, b] are infi-
nite dimensional spaces
In an inner product space, we have the following identities.
1. Pythagorus identity:
x ⊥ y iff |x +y|
2
= |x|
2
+|y|
2
(2.1.3)
2. Parallelogram law:
|x +y|
2
+|x −y|
2
= 2[|x|
2
+|y|
2
] (2.1.4)
Parallelogram law is relevant in view of the following theorem.
Theorem 2.1.1 A normed space X is an inner product space iff the norm in
X satisfies the parallelogram law given by Eq. (2.1.4).
Refer Goffman and Pedrick [5] for the proof of the above Theorem.
Example 2.1.4 X = C[0, 1] with
|f| = sup
t∈(0,1]
[f(t)[, f ∈ X t ∈ [0, 1]
Take f = t, g = 1. One can check that parallegram law is not satisfied and hence
the Banach space C[0, 1] with respect to the above norm can never be made a
Hilbert space.
Theorem 2.1.2 (Schwartz inequality)
Let X be an inner product space. Then for all x, y, ∈ X, we have the following
inequality
[(x, y)[ ≤ |x||y| (2.1.5)
Equality holds iff x and ¯ y are l.d.
42 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Proof : If any of the elements x, y is zero, we are through. So, assume that
x, y = 0. By normalizing y as y/|y| = ¯ e, the above inequality reduces to
[(x, e)[ ≤ |x| for all x ∈ X (2.1.6)
So, it suffices to show Eq. (2.1.6). We have
0 ≤ (x −(x, e)e, x −(x, e)e)
= (x, x) −(x, e)
2
This gives Eq. (2.1.5). Also, if equality holds in Eq. (2.1.5), we get
(x −(x, y)y, x −(x, y)y) = 0
This implies that
x −(x, y)y = 0
and hence x, y are l.d.
On the other hand if ¯ x, ¯ y are l.d., then Eq. (2.1.5) is true. This proves the
theorem.
Let M be a closed subspace of a Hilbert space X. Given an element x ∈ X,
we wish to obtain z ∈ M which is closest to M. We have the following theorem
in this direction.
Theorem 2.1.3 Suppose x ∈ X and M a closed subspace of X. Then ∃ a
unique element z ∈ M such that
|x −z| = inf
y∈M
|x −y|
Proof : Let d = inf
y∈M
|x −y|. Then ∃ a sequence ¦¯ y
n
¦ ∈ M such that
|x −y
n
| →d
By parallelogram law, we have
|y
n
−y
m
| = |(¯ y
n
−x) −(y
m
−x)|
2
= [2|y
n
−x|
2
+ 2|y
m
−x|
2
+| −2x +y
n
+y
m
|
2
]
≤ [2|y
n
−x|
2
+ 2|y
m
−x|
2
−4d
2
]
→ 2d
2
+ 2d
2
−4d
2
= 0
That is, ¦y
n
¦ is Cauchy and since M is a closed subspace of X, it is complete
and hence y
n
→z ∈ M. It is clear that
|x −z| = lim
n→∞
|x −y
n
| = d
Uniqueness of z is easy to prove.
This gives us the following theorem. In this theorem M

represent the
spaces of all elements orthogonal to M.
2.1. FINITE AND INFINITE DIMENSIONAL SPACES 43
Theorem 2.1.4 Let M be a closed subspace of X. Then z ∈ M is the point of
minimum distance from x ∈ X iff x −z ∈ M

.
This, in turn, implies that every x ∈ X has a unique representation
x = z +w, z ∈ M, w ∈ M

This representation is written as
X = M ⊕M

X is called the direct sum of M and M

.
Definition 2.1.11 A set S ⊂ X is called an orthonormal set if e
α
, e
β
∈ S ⇒
(e
α
, e
β
) = 0, α = β and |e
α
| = 1 ∀ e
α
∈ S.
We have the following property for orthonormal sets in X.
Theorem 2.1.5 Let ¦e
α
¦
α∈A
be a collection of orthonormal elements in X.
Then, for all x ∈ X, we have
¸
α∈A
[(x, e
α
)[
2
≤ |x|
2
(2.1.7)
and further

¯ x −
¸
α∈A
(x, e
α
)e
α

⊥ e
β
∀ β ∈ A.
The inequality Eq. (2.1.7) is referred as Bessel’s inequality.
Definition 2.1.12 Let S be an orthonormal set in a Hilbert space X. Then S is
called a basis for X (or a complete orthonormal system) if no other orthonormal
set contains S as a proper set.
The following theorem presents the most important property of a complete
orthonormal set.
Theorem 2.1.6 Let X be a Hilbert space and let S = ¦e
α
¦
α∈A
be an orthonor-
mal set in X. Then the following are equivalent in X.
(1) S = ¦e
α
¦
α∈A
is complete.
(2) 0 is only vector which is orthogonal to every e
α
∈ S. That is, x ⊥ e
α
for
every α ⇒x = 0.
(3) Every vector x ∈ X has Fourier series expansion x =
¸
α∈A
(x, e
α
)e
α
.
(4) Every vector x ∈ X satisfies the Parseval’s equality
|x|
2
=
¸
α∈A
|(x, e
α
)[
2
(2.1.8)
44 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Proof : (1) ⇒ (2)
Suppose (2) is not true, then ∃ x = 0 such that x ⊥ e
α
∀α ∈ A. Defining
e = x/|x|, we get an orthonormal set S ∪ ¦e¦ which properly contains S. This
contradicts the completeness of S. (2) ⇒ (3)
x −
¸
α∈A
(x, e
α
)e
α
is orthogonal to e
β
for every β. But by (2), it follows that it
must be the zero vector 0 and hence
x =
¸
α∈A
(x, e
α
)e
α
(3) ⇒ (4)
We have
|x|
2
= (x, x) = (
¸
α∈A
(x, e
α
)e
α
,
¸
β∈A
(x, e
β
)e
β
)
=
¸
α∈A
[(x, e
α
)[
2
(4) ⇒ (1)
If S = ¦e
α
¦
α∈A
is not complete, it is properly contained in an orthonormal set
ˆ
S and let e ∈
ˆ
S −S. then Parseval’s equation Eq. (2.1.8) gives
|e|
2
=
¸
α∈A
[(e, e
α
)[
2
= 0
contradicting the fact that e is a unit vector.
Example 2.1.5 In X = L
2
[−π, π], the collection of functions
S =

1


,
cos t

π
,
cos 2t

π
, . . .
sin t

π
,
sin 2t

π
, . . .

is a complete orthonormal set and every f ∈ L
2
[−π, π] has a Fourier series
expansion
f(t) =
a
0
2
+

¸
k=1
[a
k
cos kt +b
k
sin kt] (2.1.9)
where
a
k
=
1
π

π
−π
f(t) cos ktdt and b
k
=
1
π

π
−π
f(t) sin ktdt, k = 0, 1, 2, . . .
The fact that the above set S forms an orthonormal basis in L
2
[−π, π] will be
proved in Chapter 7. We also note that for any f ∈ L
2
[−π, π] we have the
Parseval’s relation

π
π
f
2
(t)dt =
(f, 1)
2

+

¸
n=1
1
π
[(f, cos nt)
2
+ (f, sin nt)
2
]
2.1. FINITE AND INFINITE DIMENSIONAL SPACES 45
The Fourier series representation given by Eq. (2.1.9) will be used while dis-
cussing the solution of boundary value problems.
For a piecewise continuous function f(t) defined on [−π, π],
f(t) =

1, t ≥ 0
−1, t ≤ 0
we have the following Fourier series representation along with its graphics (Fig-
ure 2.1.2)
f(t) =
4
π

¸
n=0
sin(2n + 1)t
2n + 1
-3 -2 -1 1 2 3
t
-1
-0.5
0.5
1
ft
Figure 2.1.2: Sketch of first few terms of the Fourier series
We now give the following Gram-Schmidt procedure which helps in producing an
orthonormal collection ¦e
1
, e
1
, . . . , e
n
¦ out of the .i. collection ¦¯ x
1
, ¯ x
2
, . . . , ¯ x
n
¦.
We first obtain orthogonal vectors ¦¯ y
1
, ¯ y
2
, . . . , ¯ y
n
¦, inductively
¯ y
1
= ¯ x
1
.
.
.
¯ y
i
= ¯ x
i
−α
i,1
¯ y
1
−α
i,2
¯ y
2
, . . . , α
i,i−1
¯ y
i−1
where
α
i,j
=
(¯ x
i
, ¯ y
j
)
(¯ y
j
, ¯ y
j
)
, i ≤ j ≤ i −1
This is continued inductively for i + 1, i + 2, , . . . , n. It is clear that
L[¯ x
1
, ¯ x
2
, . . . , ¯ x
n
] = L[¯ y
1
, ¯ y
2
, . . . , ¯ y
n
]
46 CHAPTER 2. MATHEMATICAL PRELIMINARIES
for each n. Normalizing ¯ y
i
, we get the orthonormal collection ¦¯ e, ¯ e
2
, . . . , ¯ e
n
¦,
¯ e
i
= ¯ y
i
/|¯ y
i
|.
Example 2.1.6 In the space L
2
[−1, 1], the set of all polynomials
¦P
0
(t), P
1
(t), . . .¦, called Legendre polynomials, which are obtained by Gram-
Schmidt orthonormalisation procedure, is an orthonormal basis for L
2
[−1, 1].
Computation for p
0
, p
1
, p
2
, p
3
, p
4
is given below. The higher degree Legendre
polynomials are inductively obtained.
p
0
(t) = 1, p
1
(t) = t
p
2
(x) = t
2
−α
0
p
0
(t) −α
1
p
1
(t), α
0
=
(t
2
, 1)
(1, 1)
=
1
3
,
α
1
=
(t
3
, t)
(t, t)
= 0
and hence
p
2
(t) = t
2

1
3
p
3
(t) = t
3
−α
0
p
0
(t) −α
1
p
1
(t) −α
2
p
2
(t), α
0
=
(t
3
, 1)
(1, 1)
= 0,
α
1
=
(t
3
, t)
(t, t)
=
3
5
, α
2
= 0
and hence
p
3
(t) = t
3

3
5
t
p
4
(t) = t
4
−α
0
p
0
(t) −α
1
p
1
(t) −α
2
p
2
(t) −α
3
p
3
(t),
α
0
=
(t
4
, 1)
(1, 1)
=
1
5
, α
1
= 0,
α
2
=
(t
4
, p
2
(t))
(p
2
(t), p
2
(t))
=
6
7
, α
3
= 0
and hence
p
4
(t) = 0 = t
4

6
7
t
2

1
5
Normalising these polynomials we get the Legendre polynomials P
i
(t), 1 ≤ i ≤ 4,
given by P
0
(t) = 1, P
1
(t) = t, P
2
(t) =
1
2
(3t
2
−1), P
3
(t) =
1
2
(5t
3
−3t),
P
4
(t) =
1
8
(35t
4
−30t
2
+ 3).
2.2. LINEAR AND NONLINEAR OPERATORS 47
For f ∈ L
2
[−1, 1], we have the Legendre series representation
f(t) =

¸
i=0
a
i
P
i
(t), a
i
= (f, P
i
) =

1
−1
f(t)P
i
(t)dt, i ≥ 0
For f(t) = sinπt we have the following graphical representation of first few
terms of the Legendre series.
-3 -2 -1 1 2 3
t
-4
-2
2
4
ft
Figure 2.1.3: Sketch of First few of terms of Legendre series for sin πt
Legendre polynomials will be used while discussing series solution of Legendre
differential equation, to be discussed in Chapter 6.
2.2 Linear and Nonlinear Operators
In this section, the spaces under consideration will be normed spaces, unless
otherwise stated.
Definition 2.2.1 T : X →Y is said to be linear if
T(α
1
x
1

2
x
2
) = α
1
Tx
1

2
Tx
2
, ∀α
1
, α
2
∈ ' ∀ x
1
, x
2
∈ X
Otherwise, T is called nonlinear.
Definition 2.2.2 T : X →Y is said to be continuous at x ∈ X if
x
n
→x in X ⇒Tx
n
→Tx in Y
T is said to be continuous on X if T is continuous at every x ∈ X.
48 CHAPTER 2. MATHEMATICAL PRELIMINARIES
One can show that a linear operator T is continuous on X iff ∃ a constant
k > 0 such that
|Tx| ≤ k|x| ∀ x ∈ X (2.2.1)
A linear T satisfying Eq. (2.2.1) is said to be a bounded operator. B(X, Y ) will
denote the space of all bounded linear operators from X to Y. One can show that
B(X, Y ) is also normed space with |T| defined by
|T| = sup¦
|Tx|
|x|
, x =
¯
0¦ (2.2.2)
B(X) will denote the space of bounded linear operator from X into itself.
Example 2.2.1 Let X = '
n
and T : '
n
→'
n
be defined by a matrix (α
ij
)
n·n
[Tx]
i
=
n
¸
j=1
α
ij
x
j
, x = (x
1
, . . . , x
n
) ∈ '
n
T is obviously linear. Also
|Tx|
1

¸
max
i
n
¸
j=1

ij
[
¸

|x|
1
Thus T is a bounded linear operator on '
n
, equipped with 1-norm. But on '
n
,
all norms are equivalent and hence T is also a bounded linear operator on '
n
,
equipped with the Euclidean norm |x|
2
=

n
¸
i=1
x
2
i

1/2
.
Example 2.2.2 For a discussion on boundary value problems, we shall be en-
countering a linear operator of the form
[Tf](t) =

1
0
k(t, s)f(s)ds, f ∈ L
2
[0, 1]
Assume that k : [0, 1] [0, 1] →' is square integrable :

1
0

1
0
k
2
(t, s)dtds < ∞.
We have
[Tf(t)[
2

1
0
[k(t, s)f(s)[
2
ds

1
0
[k(t, s)[
2
ds

1
0
[f
2
(s)[ds
(by the application of Schwartz inequality to f(s) and k(., s) ∈ L
2
[0, 1]). This
inequality implies that

1
0
[Tf(t)[
2
dt ≤ (

1
0
[k(t, s)[
2
dtds)(

1
0
[f
2
(s)[ds)
2.2. LINEAR AND NONLINEAR OPERATORS 49
Denote by k
2
=

1
0

1
0
k(t, s)[
2
dtds. Then we get
|Tf| ≤ k|f| ∀ f ∈ L
2
[0, 1]
That is, T ∈ B(L
2
[0, 1]).
The normed space B(X, ') of all continuous linear operators fromX to ' will
play a significant role throughout the solvability analysis of ordinary differential
equations. This space is called the conjugate space or dual space of X and is
denoted by X

. Elements of this space are called continuous linear functionals
or simply functionals.
As claimed before, X

is a normed space with norm of a functional f ∈ X

defined as
|f| = sup
¦x¦≤1
[f(x)[
We also have the second conjugate X
∗∗
= (X

)

defined in a similar way.
One can show that X ⊆ X
∗∗
. However, if X = X
∗∗
, then X is said to be
reflexive.
Definition 2.2.3 A one to one and onto linear mapping from X to Y is called
isomorphism. The spaces X and Y are then called isomorphic.
If X and Y are finite dimensional spaces of the same dimension, then one can
define an isomorphism between X and Y by defining a linear operator which
maps basis to basis. With this notion, it follows that every finite dimensional
space of dimension n is isomorphic to '
n
.
Example 2.2.3 Let X = '
n
. Then X

= '
n
(up to an isomorphism). To see
this we proceed as follows.
Let S = ¦e
1
, e
2
, . . . , e
n
¦ be an orthonormal basis for '
n
. Define a set S

=
¦f
1
, f
2
, . . . , f
n
¦ ⊂ X

as follows:
f
j
(e
i
) = δ
ij
=

0, i = j
1, i = j
It is clear that ¦f
j
¦
n
j=1
are .i. Also, if f ∈ X

, one can write
f = f(¯ e
1
)
¯
f
1
+f(¯ e
2
)f
2
+. . . + f(¯ e
n
)f
n
Thus S

is a basis for X

and S

consists of n .i. vectors. So X

is equal
'
n
(up to an isomorphism).
Example 2.2.4 We denote by L
p
[0, 1] (lp) the space of p
th
power absolutely
integrable functions (p
th
power absolutely summable sequences), 1 < p < ∞.
One can show that L

p
[0, 1](lp

) = L
q
[0, 1](lq),
1
p
+
1
q
= 1. (Refer Royden [10]).
From this result it follows that L
p
(lp) spaces are reflexive spaces. In particular,
we have L

2
[0, 1] = L
2
[0, 1](l

2
= l
2
).
50 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Definition 2.2.4 Let T ∈ B(X, Y ). We define the conjugate operator T

:
Y

→X

, of T as follows:
[T

f][x] = f(Tx), f ∈ X

, x ∈ X
One can show that T

is also bounded and |T

| = |T|.
However, if X is a Hilbert space, then corresponding to T ∈ B(X, Y ) we define
T

∈ B(Y, X) as :
(x, T

y)
X
= (Tx, y)
Y
∀ x ∈ X, y ∈ Y. (2.2.3)
We say that T ∈ B(X) is self-adjoint if T

= T.
Definition 2.2.5 Let X be Hilbert space and T ∈ B(X) be a self adjoint oper-
ator. T is said to be positive semi-definite if
(T ¯ x, ¯ x) ≥ 0 for all ¯ x ∈ X
T is called positive definite if the above inequality holds and is strict for ¯ x =
¯
0.
For T ∈ B(X, Y ), the subspaces N(T) and R(T) are defined as follows
N(T) = ¦x ∈ X : Tx = 0¦, R(T) = ¦y ∈ Y : y = Tx, x ∈ X¦
We have the following relationship between N(T

) and R(T) and vice-versa.
Theorem 2.2.1 Let X and Y be Hilbert spaces and let T ∈ B(X, Y ). Then
N(T

) = R(T)

and R(T

)

= N(T).
Proof : It is sufficient to prove the first one. Assume that x ∈ N(T

), then
T

¯ x = 0 and hence (T

x, y) = 0 for all y ∈ X. This implies that (x, Ty) = 0 for
all y ∈ X. That is, x ∈ R(T)

and hence N(T

) ⊂ R(T)

. Reversing the above
argument, one can show that R(T)

⊂ N(T

) and hence the first part of the
result.
This theorem can be extended to get the following result.
Theorem 2.2.2 Let X and Y be Hilbert spaces. If T ∈ B(X, Y ), then R(T) =
N(T

)

and R(T

) = N(T)

.
Remark 2.2.1 If R(T) is closed, we get
R(T) = N(T

)

, R(T

) = N(T)

For operators T ∈ B(X) with closed range in a Hilbert space, we get the or-
thogonal projector P from X to M = R(T), if T

T is invertible. This is done
as follows by making use of Theorem 2.1.4.
Let M = R(T). By Theorem 2.1.4 we can write z ∈ X as follows:
z = u +v, u ∈ R(T) = M, v ∈ R(T)

= N(T

)
2.2. LINEAR AND NONLINEAR OPERATORS 51
v ∈ N(T

) implies that T

(z −Tx) = 0 and hence T

z = T

Tx. This gives
x = (T

T)
−1
T

z (2.2.4)
The operator P : X → R(T) = M is called the orthogonal projector and is
given by
Pz = u = Tx = [T(T

T)
−1
T

]z
This element
u = T(T

T)
−1
T

z ∈ R(T) = M (2.2.5)
minimizes the functional f(x) = |z −Tx| in R(T) = M.
Example 2.2.5 Let X = '
n
and A be the matrix (α
ij
). Then A

= (β
ij
) where
β
ij
= α
ji
. That is A

, reduces to the transpose A
¯
of the matrix A. We have
(Ax, y) =
n
¸
i=1
n
¸
j=1
α
ij
x
j
y
i
=
n
¸
j=1
n
¸
i=1
α
ji
x
i
y
j
=
n
¸
i=1
(
n
¸
j=1
α
ji
y
j
)x
i
As β
ij
= α
ji
, we get
(Ax, y) =
n
¸
i=1
(
n
¸
j=1
β
ij
y
j
)x
i
= (x, By), B = (β
ij
)
This implies that A

= B = (β
ij
) = (α
ji
) = A
¯
Example 2.2.6 If X = '
n
and A = (a
ij
) is a mn matrix with (A
¯
A)
m·m
invertible, then Eq. (2.2.5) gives the element u ∈ R(A) nearest to R(A) from
¯ z ∈ '
n
. This formulation can be used in linear data fitting model leading to the
least square approximation, as we see below.
Data fitting model is used in measuring relationship between two physical
quantities. Let the first quantity be represented by variables x
1
, x
2
, . . . , x
n
, giving
the vector x ∈ '
n
, the second quantity by y and the functional relationship be
given by
y = f(x), x = (x
1
, . . . , x
n
)
We assume that the observed data f
k
approximate the value f(x
(k)
). The
problem of data fitting is to recover the function f(x) from the given data
f
k
, k = 1, 2, . . . , m. We try to fit the data by a linear function F(x) given
by
F(¯ x) = a
0
+a
1
x
1
+ +a
n
x
n
(2.2.6)
52 CHAPTER 2. MATHEMATICAL PRELIMINARIES
We shall compute parameters a
k
so that the deviations
d
k
= f
k
−F(x
(k)
; a
0
, a
1
, . . . , a
n
), k = 1, . . . , m
are as small as possible.
Hence we minimize the error d
k
w.r.t. the 2-norm
|
¯
d|
2
=
m
¸
k=1
d
2
k
,
¯
d = (d
1
, d
2
, . . . , d
m
)
From the linear model given by Eq. (2.2.6), we have
|
¯
d|
2
=
m
¸
k=1
(f
k
−F(¯ x
(k)
))
2
=
m
¸
k=1
(f
k

n
¸
i=0
a
i
x
(k)
i
)
2
Let ¯ a = (a
0
, a
1
, . . . , a
n
), ¯ z = (f
1
, f
2
, . . . , f
n
)
A =

1 x
(1)
x
(2)
x
(n)
1 x
(1)
2
x
(2)
2
x
(n)
2
.
.
.
.
.
.
.
.
.
.
.
.
1 x
(1)
m
x
(2)
m
x
(n)
m
¸
¸
¸
¸
¸
¸
.
Then the least square problem (data fitting problem) reduces to finding ¯ u

= Aa

such that
|z −u

| = inf
a∈J
n
|z −Aa|
By using Eq. (2.2.5), the vector u

is given by
¯ u

= A(A
¯
A)
−1
A
¯
¯ z
if (A
¯
A) is invertible.
Remark 2.2.2 From the point of view of applications to differential equations,
it is not possible to define the linear operator T =
d
dt
on the whole space L
2
[0, 1].
In such a case we first define the domain D(T) as follows:
D(T) = ¦f ∈ L
2
[0, 1] :
˙
f =
df
dt
exists a.e. with
˙
f ∈ L
2
[0, 1]¦
D(T) is called the space AC[0, 1] of all absolutely continuous functions on [0, 1].
It can be shown that D(T) is dense in L
2
[0, 1]. However, it is not a bounded
operator as it maps the sequence f
n
(t) =

sin nπt
n

which converges to zero in
L
2
[0, 1] to a sequence Tf
n
(t) = π cos nπt which is not convergent.
2.2. LINEAR AND NONLINEAR OPERATORS 53
Definition 2.2.6 Let X be a Hilbert space and T : D(T) → X be such that
D(T) is dense in X. Then we define D(T

) to consist of y ∈ X for which the
linear functional
f(x) = (Tx, y), x ∈ D(T) ⊂ X
is continuous in D(T). By Riesz representation theorem (refer Bose and Joshi
[2]), there exists a unique y

∈ X such that
(Tx, y) = (x, y

) ∀ x ∈ D(T)
The adjoint T

of T is defined on D(T

) by T

y = y

. In other words the adjoint
T

of T for densely defined operators is given by
(Tx, y) = (x, T

y), ∀ x ∈ D(T), y ∈ D(T

)
T is said to be self-adjoint iff D(T

) = D(T) and T

= T.
Example 2.2.7 Let X = L
2
[0, 1] and Tf(t) = −
¨
f(t) = −
d
2
f
dt
2
, where D(T) is
defined as under
D(T) = ¦f ∈ L
2
[0, 1]¦ :
˙
f,
¨
f ∈ L
2
[0, 1] with f(0) = f(1) = 0¦
One can show that D(T) is dense in L
2
[0, 1] and
(Tf, g) = −

1
0
¨
f(t)g(t)dt = −
˙
f(t)g(t)

1
0
+

1
0
˙
f(t) ˙ g(t)dt
= −
˙
f(1)g(1) −
˙
f(1)g(0) +f(t) ˙ g(t)

1
0

1
0
f(t)¨ g(t)dt
= −
˙
f(1)g(1) +
˙
f(1)g(0) +f(1) ˙ g(1) −f(0) ˙ g(0) −

1
0
f(t)¨ g(t)dt, f ∈ D(T)
= −
˙
f(1)g(1) +
˙
f(1)g(0) −

1
0
f(t)¨ g(t)dt
= −

1
0
f(t)¨ g(t)dt
if g(0) = g(1) = 0.
Hence
(Tf, g) = (f, Tg)
where T

g = −¨ g and D(T

) = ¦g ∈ X : ˙ g, ¨ g ∈ L
2
, g(0) = g(1) = 0¦.
Thus T

= T and D(T) = D(T

). So, T is a self-adjoint operator and is densely
defined. Also one can show that T is closed. That is if x
n
→x in X = L
2
[0, 1]
and Tx
n
→y in X. Then x ∈ D(T) and y = Tx.
54 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Definition 2.2.7 T ∈ B(X, Y ) is said to be compact if it maps bounded sets
B ⊂ X onto relatively compact set T(B) (T(B) is compact). Equivalently, T
is compact iff for every bounded sequence ¦¯ x
n
¦ ⊂ X, ¦T ¯ x
n
¦ has a convergent
subsequence in Y .
Definition 2.2.8 T ∈ B(X, Y ) is called finite rank operator if the range of T
is finite dimensional.
Theorem 2.2.3 Every finite rank operator T ∈ B(X, Y ) is compact.
Proof : As R(T) is finite dimensional, we have R(T) = L[y
1
, y
2
, . . . , y
N
].
Consider now a bounded sequence ¦x
n
¦ ⊂ X, then Tx
n
=
¸
N
j=1
α
(n)
j
y
j
for
some scalars α
(n)
j
.
Since T is bounded ¦α
(n)
j
¦ is bounded for each j and hence ∃ subse-
quence ¦α
(n
k
)
j
¦ such that α
(n
k
)
j
→ α
j
, 1 ≤ j ≤ N. It now follows that
Tx
nk
=
¸
N
j=1
α
(n
k
)
j
y
j
→ y = α
1
y
1
+ α
2
y
2
+ . . . + α
N
y
N
. This proves the
compactness of T.
The following theorem gives the compactness property of the adjoint of a
compact operator and also the limit of compact operators. X and Y are assumed
to be Hilbert spaces.
Theorem 2.2.4 (1) The composition of a compact operator with a bounded
operator is compact.
(2) Limit of a sequence of compact operator is compact.
(3) T ∈ B(X, Y ) is compact iff T

∈ B(Y, X) is compact.
Refer Bose and Joshi [2] for the proof of this theorem.
Example 2.2.8 Consider the operator T : L
2
[0, 1] → L2[0, 1], as defined in
Example 2.2.2:
[Tf](t) =

1
0
k(t, s)f(s)ds, f ∈ L
2
[0, 1]
with

1
0

1

[k(t, s)]
2
dtds < ∞.
Let ¦φ
i
(t)¦ be a complete orthonormal set in L
2
[0, 1], then the set
¦φ
i
(t)φ
j
(s)¦

i,j=1
is a complete orthonormal set in L
2
([0, 1] [0, 1]). Hence
k(t, s) =

¸
i=1

¸
j=1
a
ij
φ
i
(t)φ
j
(s)
where a
ij
=

1
0

1
0
k(t, s)φ
i
(t)φ
j
(s)dtds. By Parseval’s equality, we have

1
0

1
0
[k(t, s)]
2
dtds =

¸
i=1

¸
j=1
(a
ij
)
2
2.2. LINEAR AND NONLINEAR OPERATORS 55
Further, we have

1
0

1
0

k(t, s) −
n
¸
i=1
n
¸
j=1
a
ij
φ
i
(t)φ
j
(s)
¸
¸
2
dtds
=

1
0

1
0


¸
i,j=n+1
[a
ij
φ
i
(t)φ
j
(s)]
¸
¸
2
dtds
=

¸
i,j=n+1
[a
ij
[
2
→0
since

¸
i=1

¸
j=1
[a
ij
[
2
< ∞.
Now define K
n
(t, s) =
n
¸
i,j
a
ij
φ
i
(t)φ
j
(s).
Then T
n
: X →X defined by
(T
n
f)(t) =

1
0
K
n
(t, s)f(s)ds
is a finite rank operator and hence it is compact. Also T
n
→ T in B(X) and
hence by the previous theorem T is compact.
We have the following theorem, called Fredholm Alternative, giving the solv-
ability of the operator equation involving compact operators.
Consider three related equations in a Hilbert space X (over the field of complex
numbers)
Tx −λx = 0 (H)
T

z −λz = 0 (H

)
Ty −λy = f (NH)
Theorem 2.2.5 1. Either both (H) and (H

) have only trivial solutions or both
have nontrivial solutions.
2. The necessary and sufficient condition for (NH) to have a nontrivial
solution is that f be orthogonal to every solution of (H

).
In particular, if λ is not an eigenvalue of T then λ is not an eigenvalue of T

and (NH) has one and only one solution for each f ∈ X.
We now state the following theorem which gives orthonormal basis for a Hilbert
space X arising out of eigenvectors of T ∈ B(X), refer Bose and Joshi[2] for the
proof.
Theorem 2.2.6 Let T be a self-adjoint compact operator on a Hilbert space.
Then the eigenvectors of T form a complete orthonormal basis for X.
56 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Corollary 2.2.1 Let T be a compact and self-adjoint operator in a Hilbert space
X. Then T is positive semidefinite (positive definite) iff all eigenvalues of T are
nonegative (positive).
Example 2.2.9 Let X = '
n
and A : '
n
→ '
n
be given by a symmetric
matrix (a
ij
)
n·n
. Then A ∈ B('
n
) is compact self-adjoint operator and hence
the eigenvectors of A form a basis.
Let

i
= λ
i
φ
i
, 1 ≤ i ≤ n
where ¦φ
i
¦
n
i=1
are orthonormal.
Let P = [φ
i
, φ
2
, . . . , φ
n
]. Then P is an orthogonal matrix (P
¯
P = I) and
AP = [Aφ
i
, . . . , Aφ
n
] = [λ
1
φ
1
, . . . λ
n
φ
n
] = PD
where D = diag[λ
1
, . . . , λ
n
].
This gives us P
−1
AP = D. Thus A is similar to a diagonal matrix. This is
called diagonalisation procedure and will be used in Chapter 4 while computing
transition matrix of a given system of differential equations. However, it may be
pointed out A can be diagonalised even if A is not self-adjoint provided it has
distinct eigenvalues. Assume that the eigenvectors corresponding to distinct
eigenvalues are linearly independent and hence defining P as before, we get
P
−1
AP = D. But P is no longer an orthogonal matrix, it is merely nonsingular.
We now examine the properties of a nonlinear operator F : X → Y, X and Y
are normed spaces.
Definition 2.2.9 F : X → Y is said to be continuous if x
n
→ x in X ⇒
Fx
n
→Fx ∀ x ∈ X. F is said to be bounded if it maps bounded sets in X into
bounded sets in Y.
Compact nonlinear operators are defined in the same way as compact linear
operators.
Remark 2.2.3 For nonlinear operators continuity is not equivalent to bound-
edness, refer Joshi and Bose [6].
Example 2.2.10 Consider the following ODE
dx
dt
= f(t, x(t)), t ∈ I = [t
0
, t
f
] (2.2.7)
with f : I ' →' continuous and bounded.
Then by Proposition 1.1.1, the solvability of Eq. (2.2.7) is equivalent to the
solvability of the equation
x(t) = x(t
0
) +

t
t0
f(s, x(s))ds, t ∈ I = [t
0
, t
f
] (2.2.8)
2.3. LIPSCHITZ CONTINUITY AND DIFFERENTIABILITY 57
The RHS of Eq. (2.2.8) gives rise to a nonlinear operator F on C[t
0
, t
f
] defined
as follows
[Fx](t) = x(t
0
) +

t
t0
f(s, x(s))ds
We have
[[Fx](t
1
) −[Fx](t
2
)[ ≤

t2
t1
[f(s, x(s))[ds
≤ m[t
1
−t
2
[(m = sup
(s,x)∈I·J
[f(s, x)[)
For this inequality it follows that Fx is continuous and hence Fx ∈ C[t
0
, t
f
] if
x ∈ C[t
0
, t
f
]. Further
[Fx
k
(t) −Fx(t)[ ≤

t
t0
[f(s, x
k
(s)) −f(s, x(s)[ds
Let x
k
→x in C[t
0
, t
f
]. This implies that x
k
(s) → x(s) in ' for all s ∈ [t
0
, t
1
].
Since (s, x) →f(s, x) is continuous, it follows by bounded convergence theorem,
that
[Fx
k
(t) −Fx(t)[ ≤

t
t0
[f(s, x
k
(s)) −f(s, x, s)[ds →0
for all t ∈ [t
0
, t
f
]. One can also show that this convergence is uniform and hence
|Fx
k
−Fx| = sup
t∈[t0,t
f
]
[Fx
k
(t) −Fx(t)[ →0
Hence F is a continuous nonlinear operator from X = C[t
0
, t
f
] in to itself.
Similarly, one can show that F is bounded on X.
The above result also holds true if f(t, x) is defined on I Ω → Ω where Ω is
an open subset of '
n
. One then denotes by C[t
0
, t
f
] the space of all '
n
− valued
continuous functions and the norm of the elements x ∈ C[t
0
, t
f
] is defined as
|x| = sup
t∈[t0,t
f
]
|x(t)|
J
n
2.3 Lipschitz Continuity and Differentiability
In this section too the spaces under consideration will be normed space.
Definition 2.3.1 F : X → Y is said to be Lipschitz continuous if ∃ constant
k > 0 such that
|Fx −Fy| ≤ k|x −y| ∀ x, y ∈ X (2.3.1)
58 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Definition 2.3.2 F : X → Y is said to be differentiable at x ∈ X if ∃ A ∈
B(X, Y ) such that
F(x +h) −F(x) = Ah +w(x, h) (2.3.2)
where lim
¦
¯
h¦→0
|w(x, h)|
|h|
= 0
A is called the derivative of F at x and is denoted by F
t
(x). Its value at h
will be denoted by F
t
(x)h. One can show that F is differentiable at x ∈ X with
derivative A iff
lim
t→0
F(x +th) −F(x)
t
= Ah ∀ h ∈ X
and ¯ x →F
t
(¯ x) is continuous.
Equivalently, writing φ(t) = F(x + th) for x, h ∈ X, we see that F has
derivative F
t
(¯ x) iff
d
dt
[φ(t)]
t=0
= F
t
(x)h, for all
¯
h ∈ X
We have the following variant of the meanvalue theorem for F : X → Y. Refer
Joshi and Bose[ 6 ] for the proof of this theorem.
Theorem 2.3.1 Let F : X → Y be differentiable at every ¯ x in X. Then for
points ¯ x, ¯ x +
¯
h ∈ X and ¯ e ∈ Y

, there exists a constant τ, 0 < τ < 1, such that

¯ e, F(¯ x +
¯
h) −F(x)

=

¯ e, F
t
(¯ x +τ
¯
h)
¯
h

This immediately gives the following Corollary.
Corollary 2.3.1 If |F
t
(¯ x)| ≤ K ∀ ¯ x ∈ X then F is Lipschitz continuous.
Example 2.3.1 Let F : '
n
→'
n
be given by
F(x
1
, x
2
, . . . , x
n
) = (f
1
(x
1
, x
2
, . . . , x
n
), . . . , f
m
(x
1
, x
2
, . . . x
n
)). Let A = (a
ij
) be
the mn (matrix, which we want to compute) such that F
t
(x) = A.
We choose
¯
h ∈ '
n
as j
th
coordinate vector e
j
= (0, . . . , 1, . . . , 0) Then
lim
t→0
|
1
t
F(x +th) −F(x) −tAh| = 0
⇔ lim
t→0
[
1
t
[f
i
(x +e
j
) −f
i
(x) −ta
ij
][ = 0, 1 ≤ i ≤ m, 1 ≤ j ≤ n
This shows that F(x) is differentiable at x ∈ '
n
iff f
i
(x) has continuous partial
derivatives at x ∈ '
n
and F
t
(x) = A =

∂f
i
∂x
j
(x)

, if partial derivative are
2.3. LIPSCHITZ CONTINUITY AND DIFFERENTIABILITY 59
continuous. So we have
F
t
(x) =


1
f
1
(¯ x) ∂
n
f
1
(¯ x)

1
f
2
(¯ x) ∂
n
f
2
(¯ x)
.
.
.
.
.
.

1
f
m
(¯ x) ∂
n
f
m
(¯ x)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
In case f is a functional, that is, f is a mapping from '
n
to ', then f
t
(x) =
∇f(x), called the gradient of f at x.
Example 2.3.2 Let X be a Hilbert space and let f : X →' be given by
f(x) =
1
2
(Ax, x), x ∈ X
We assume that A is a bounded self-adjoint linear operator. The derivative of
f is computed as under.
f
t
(x)h = lim
t→0
¸
f(x +th) −f(x)
t

= lim
t→0
1
2
¸
(A(x +th), (x +th)) −(Ax, x)
t

= lim
t→0
1
2
¸
t

(Ax, h) + (x, Ah)

+t
2
(Ah, h)
t
¸
= lim
t→0
1
2

(Ax, h) + (A

x, h) +t(Ah, h)

=
1
2

A+A

2

x, h

= (Ax, h)
as A

= A.
This implies that f
t
(x) = Ax ∀ x ∈ X. This derivative f
t
of a functional f is
also denoted as ∇f.
From this representation it also follows that f

(x) = A.
Remark 2.3.1 In dealing with minimization problem involving several vari-
ables we shall encounter convex functionals as in Example 2.3.2. A function
f : X →' is said to be convex if
f(αx + (1 −α)y) ≤ αf(x) + (1 −α)f(y), 0 < α < 1
for all x, y ∈ X. X is any normed space.
60 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Remark 2.3.2 One can show that f is strictly convex on a Hilbert space if
f

(x) is positive definite (refer problem number 8 in sectoin 2.6). Hence the
functional of Example 2.3.2 is strictly convex, if A is positive definite.
Example 2.3.3 (Least Square Approximation - Data Fitting) In Exam-
ple 2.2.5 involving problems of least square approximations, we minimize |d|
2
given by
|d|
2
= |z −Aa|
2
=
m
¸
k=1
(f
k

n
¸
i=0
a
i
x
(k)
i
)
2
where z and A are as defined in the example. Let φ(a) be defined as follows
φ(a) = (z −Aa, z −Aa)
= (z, z) −2(Aa, z) + (Aa, Aa)
= (z, z) −2(Aa, z) + (A
¯
Aa, a), ¯ a ∈ '
n
Proceeding as in Example 2.3.2, we get
∇φ(a) = −2A
¯
z + 2A
¯
Aa
It is clear that A
¯
A is positive semi-definite. If A
¯
A is invertible then A
¯
A
is positive definite and hence f is strictly convex and its critical point is the
minimizer of φ. Also a

is the critical point of φ iff ∇φ(a

) = 0
This implies that A
¯
z = A
¯
Aa

and hence a

= (A
¯
A)
−1
A
¯
z.
This is the same result, which we obtained earlier.
Remark 2.3.3 While solving differential equations, we shall also be estimating
a set of parameters which will arise in the defining model. Estimating parameters
requires the minimization of a functional involving these parameters.
Such functionals f may not be convex and hence there is a need to generate
algorithms which will computationally capture the minimizer (local / global).
We give below one such algorithm - called the steepest descent algorithm - which
captures the local minimizer of a function f defined on the entire space '
n
.
Steepest Descent Algorithm 2.3.1
Step 1 : Start with x
(k)
(k = 0)
Step 2 : Set d
(k)
= −g
(k)
, g
(k)
= ∇f(x
(k)
)
Step 3 : Minimize 1-dimensional function φ(α) = f(d
(k)
+αd
(k)
) to get
α
(k)
Step 4 : Set x
(k+1)
= x
(k)

(k)
d
(k)
Step 5 : Use any stopping criterion and set x

= x
(k+1)
. Otherwise set
k = k + 1 or GO TO STEP 2
Step 6 : Compute f(x

), x

is the point of optimum.
2.3. LIPSCHITZ CONTINUITY AND DIFFERENTIABILITY 61
In Example 2.3.2, the iteration procedure for the functional
f(x) =
1
2
(Ax, x), x ∈ '
n
is given by
x
(k+1)
= x
(k)

g
(k)
, g
(k)

Ag
(k)
, g
(k)
g
(k)
(2.3.3)
Example 2.3.4 Let F : X = C[0, 1] →C[0, 1] be defined by
[Fx](t) =

1
0
k(t, s)f(s, x(s))ds
where k and f satisfy the following properties
(i) k(t, s) is continuous on [0, 1] [0, 1].
(ii) x → f(t, x) is differentiable mapping from ' to ' for all t ∈ [0, 1] such
that
[
∂f
∂x
(t, x)[ ≤ M for all (t, x) ∈ [0, 1] '
As in Example 2.2.11, we can prove that F is a continuous mapping of C[0, 1]
to itself. We now compute the derivative of F.
[F
t
(x)h] =
d

[F(x +αh)]
α=0
= lim
α→0
[
F(x +αh) −F(x)
α
]
= lim
α→0

1
0
k(t, s)
[f(s, (x +αh)(s)) −f(s, ¯ x(s))]
α
=

1
0
k(t, s)

∂x
f(s, x(s))h(s)ds
by Lebesgue dominated convergence theorem. Thus F is differentiable at x with
F
t
(x) as a linear integral operator generated by the kernel
g(t, s) = k(t, s)
∂f
∂x
(s, x(s)) and is given by
F
t
(x)h(t) =

1
0
g(t, s)h(s)ds.
Also, as g(t, s) is a bounded on [0, 1] [0, 1], it follows that |F
t
(x)| ≤ M
and hence by Corollary 2.3.1 it follows that F is Lipschitz continuous on X =
L
2
[0, 1].
62 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Let F be differentiable at x
0
∈ X, then it follows that we have the following
expansion for F(x) with F
t
(¯ x
0
) ∈ B(X).
F(x) = F(x
0
) +F
t
(x
0
)(x −x
0
) +w(x, h) (2.3.4)
where
|w(x, h|
|h|
→0 as |h| →0.
Definition 2.3.3 If F is differentiable at x
0
then the function
ˆ
F(x) = F(x
0
) +F
t
(x
0
)(x −x
0
) (2.3.5)
is called the linearization of F at x
0
.
Remark 2.3.4 The linearization process can be used to find the zeros of the
nonlinear function F(x) in a neighbourhood of the operating point x
0
∈ X if
[F

(¯ x
0
)] is invertible.
By linearization, it suffices to compute the zeros
ˆ
F(x) and hence
0 =
ˆ
F(x) = F(x
0
) +F
t
(x
0
)(x −x
0
).
This gives
x = x
0
+ [F
t
(x
0
)]
−1
F(x
0
)
Hence we get the Newton’s algorithm to find the zeros of F(x), starting with an
initial guess ¯ x
0
.
Algorithm 2.3.2 - Newton’s Algorithm
Step 1 Set k = 0 and input initial guess x
k
and the tolerance > 0
Step 2 x
(k+1)
= x
(k)
−[F
t
(x
(k)
)]
−1
F(x
(k)
)
Step 3 If |x
(k+1)
− x
(k)
| <∈, STOP and set x

= x
(k)
as the computed zero of
F(¯ x). Otherwise increment k by 1.
Step 4 GO To Step 2
Example 2.3.5 We have shown in Example 1.3.2 that the motion of a satellite
orbiting around the earth is given by an ordinary differential equation (ODE) in
'
4
:
d¯ x
dt
= f(¯ x, ¯ u), ¯ x ∈ '
4
, ¯ u ∈ '
2
(2.3.6)
¯ x(t) = (x
1
(t), x
2
(t), x
3
(t), x
4
(t)) represents the state vector and
¯ u(t) = (u
1
(t), u
2
(t)) is the control vector (thrusts on the satellite), f is given by
f(¯ x, ¯ u) = (f
1
(¯ x, ¯ u), f
2
(¯ x, ¯ u), f
3
(¯ x, ¯ u), f
4
(¯ x, ¯ u))
where
f
1
= x
2
, f
2
= (x
1
+σ)(
x
4
σ
+w)
2

k
(x
1
+σ)
2
+u
1
,
f
3
= x
4
, f
4
= −2σ(
x
4
σ
+w)
x
2
(x
1
+σ)
+
u
2
σ
(x
1
+σ)
2.4. FIXED POINT THEOREMS AND MONOTONE OPERATORS 63
Using linearization procedure for f(¯ x, ¯ u), described by Eq. (2.3.5), around
¯ x = 0, ¯ u = 0 we get
ˆ
f(¯ x, ¯ u) = f
t
x
(
¯
0,
¯
0)¯ x +f
t
u
(
¯
0,
¯
0)¯ u
A = f
t
x
(
¯
0,
¯
0) =

∂f
1
∂x
1
∂f
1
∂x
2
∂f
1
∂x
3
∂f
1
∂x
4
∂f
2
∂x
1
∂f
2
∂x
2
∂f
2
∂x
3
∂f
2
∂x
4
∂f
3
∂x
2
∂f
3
∂x
2
∂f
3
∂x
3
∂f
3
∂x
4
∂f
4
∂x
2
∂f
4
∂x
2
∂f
4
∂x
3
∂f
4
∂x
4
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
(¯ o,¯ o)
=

0 1 0 0
3w
2
0 0 2w
0 0 0 1
0 −2w 0 0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
(2.3.7)
B = f
t
u
(0, 0) =

∂f
1
∂u
1
∂f
1
∂u
2
∂f
2
∂u
1
∂f
2
∂u
2
∂f
3
∂u
1
∂f
3
∂u
2
∂f
4
∂u
1
∂f
4
∂u
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
(¯ o,¯ o)
=

0 0
1 0
0 0
0 1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
(2.3.8)
Here σ is normalized to 1. So the linearized ODE, corresponding to Eq. (2.3.6)
is
d¯ x
dt
= A¯ x +B¯ u
where A and B are given by Eq. (2.3.7) and Eq. (2.3.8), respectively.
2.4 Fixed Point Theorems and Monotone Oper-
ators
Definition 2.4.1 Let X be a metric space and F : X →X. An element x ∈ X
is said to be a fixed point of the mapping F if Fx = x.
We begin with the following constructive fixed point theorem by Banach,
known as Banach contraction mapping principle. This theorem will be one of the
important instruments in obtaining existence theorems for ordinary differential
equations.
Definition 2.4.2 Let (X, d) be a metric space. F : X → X is said to be
contraction if F is Lipschitz continuous with Lipschitz constant strictly less than
64 CHAPTER 2. MATHEMATICAL PRELIMINARIES
1. That is, there exists a positive constant α < 1 such that
d(Fx, Fy) ≤ αd(x, y) for all x, y ∈ X (2.4.1)
If X is a normed space, then (2.4.1) is equivalent to the following inequality
|F ¯ x −F ¯ y| ≤ α|¯ x − ¯ y| for all ¯ x, ¯ y ∈ X (2.4.2)
Theorem 2.4.1 Let F be a contraction mapping of a complete metric space X
into itself. Then F has a unique fixed point x

∈ X. Moreover if x
0
∈ X is any
arbitrary point, and ¦x
n
¦ a sequence defined by
x
n+1
= Fx
n
(n = 0, 1, 2, . . .)
then lim
n→∞
x
n
= x

and d(x
n
, x

) ≤
α
n
1 −α
d(x
1
, x
0
).
Proof : We have
d(x
k
, x
k−1
) = d(Fx
k−1
, Fx
k−2
)
≤ αd(x
k−1
, x
k−2
)
≤ αd(x
1
, x
0
)
This gives
d(x

, x
m
) ≤ d(x

, x
−1
) +. . . +d(x
m+1
, x
m
)
≤ [α
m

m+1
+. . . +α
l−1
]d(x
1
, x
0
)
≤ α
m
[1 +α +α
2
+. . .]d(x
1
, x
0
)
=
¸
α
m
1 −α

d(x
1
, x
0
), l > m
→0 as m →∞
This implies that ¦x
n
¦ is Cauchy in X and as X is complete, it converges to
x

∈ X. As x
n
→ x

, x
n+1
= Fx
n
and F is continuous, it follows that
x

= Fx

.
The uniqueness of the fixed point x

follows from the fact that for two fixed
point x

and y

, we have
d(x

, y

) = d(Fx

, Fy

)
≤ αd(x

, y

)
< d(x

, y

)
and hence d(x

, y

) = 0.
The error estimate d(x
n
, x

) is given by
d(x

n
, x

) ≤ d(x
m
, x

) +d(x
m
, x
n
)
≤ d(x
m
, x

) +
α
n
1 −α
d(x
1
, x
0
), m > k
2.4. FIXED POINT THEOREMS AND MONOTONE OPERATORS 65
Letting m →∞, we get
d(x
n
, x

) ≤
α
n
1 −α
d(x
1
, x
0
)
As a Corollary of this theorem, we obtain the solvability of the linear equa-
tion
x −Ax = y (2.4.3)
in a Banach space X. Here, A ∈ B(X) with |A| < 1 and ¯ y ∈ X.
Corollary 2.4.1 Let A ∈ B(X) be such that |A| < 1. Then (I −A) is bounded
and invertible with
|(I −A)
−1
| ≤
1
1 −|A|
(2.4.4)
Proof : We first show that
x −Ax = y, y ∈ X
has a unique solution. Let Fx = Ax +y. Then
|Fx
1
−Fx
2
| = |Ax
1
−Ax
2
| ≤ |A||x
1
−x
2
| for all x
1
, x
2
∈ X
As α = |A| < 1, it follows that F is a contraction and hence it has a unique
fixed point x in the Banach space X (complete metric space). That is, Eq.
(2.4.3) has a unique solution x. Further, let
x = (I −A)
−1
y
This implies that x −Ax = y, y ∈ X arbitrary and hence
|x| ≤ |Ax| +|y|
≤ |A||x| + |y|
This gives
|¯ x| ≤
¸
1
1 −|A|

|¯ y|
Hence
|(I −A)
−1
¯ y| ≤
¸
1
1 −|A|

|¯ y|
for all ¯ y ∈ X.
This implies that
|(I −A)
−1
| ≤
1
1 −|A|
66 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Example 2.4.1 Solve the linear system
A¯ x =
¯
b (2.4.5)
in '
n
, A = (α
ij
)
n·n
. Expanding Eq. (2.4.5) in terms of its component we get
α
11
x
1

12
x
2
+. . . +α
1n
xn = b
1
α
21
x
1

22
x
2
+. . . +α
2n
xn = b
2
.
.
.
α
n1
x
1

n2
x
2
+. . . +α
nn
x
n
= b
n
(2.4.6)
Assume that α
ii
> 0 and the matrix A is diagonally dominant:
n
¸
i=1
i,=j

ij
[ < α
ii
, 1 ≤ i ≤ n.
Then Eq. (2.4.6) is equivalent to
x
1
=
b
1
α
11

α
12
x
2
α
11
. . . −
α
1n
x
n
α
11
x
2
=
b
2
α
22

α
21
x
1
α
22
−. . . −
α
2n
x
n
α
22
.
.
.
x
n
=
b
n
α
nn

α
n
, x
1
α
nn
−. . . −
α
n,n−1
x
n−1
α
nn
(2.4.7)
Define
B = −

¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
0
α
12
α
11

α
1n
α
11
α
21
α
22
0
α
2n
α
22
.
.
.
α
n1
α
nn
α
n2
α
nn
0
¸

, c =

¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
b
1
α
11
b
2
α
22
.
.
.
b
n
α
nn
¸

Then the solvability of Eq. (2.4.7) is equivalent to the solvability of the matrix
equation
¯ x −B¯ x = ¯ c, |B| < 1.
Hence by the above corollary, this has a unique solution ¯ x

and the iteration
scheme
¯ x
(k+1)
= ¯ c +B¯ x
(k)
2.4. FIXED POINT THEOREMS AND MONOTONE OPERATORS 67
converges to ¯ x

.
This gives us the well known Jacobi’s iteration scheme (refer Conte and deBoor
[4]).
x
(k+1)
1
=
b
1
α
11

α
12
α
11
x
(k)
2

α
2n
α
11
x
(k)
n
x
(k+1)
2
=
b
n
α
22

α
21
α
22
x
(k)
1

α
2,n
α
22
x
(k)
n
.
.
. =
.
.
. −
.
.
. −
.
.
.
x
(k+1)
n
=
b
n
α
nn

α
n1
α
nn
x
(k)
1

α
n,n−1
α
nn
x
(k)
n
For the diagonally dominant matrix
A =

¸
¸
¸
24 3 9 7
1 12 −3 −4
0 1 22 −3
7 −4 −3 24
¸

,
¯
b = (1, 1, 2, −4)
we get the following iterated solution ¯ x
(n)
= (x
(n)
1
, x
(n)
2
, x
(n)
3
, x
(n)
4
) of A¯ x =
¯
b.
Table 2.4.1: Iterated solution of the matrix equation
n x
(n)
1
x
(n)
2
x
(n)
3
x
(n)
4
|x −x
(n)
|
0 1 0 -1 2
1 -0.16667 0.41667 0.36364 -0.58333 5.53030
2 0.02336 -0.00631 -0.00758 -0.00316 1.56439
3 0.04622 0.07844 0.09077 -0.17548 0.37828
4 0.04901 0.04368 0.06341 -0.15573 0.08465
5 0.05785 0.04319 0.06769 -0.16575 0.02363
6 0.05923 0.04018 0.06634 -0.16788 0.00786
7 0.06073 0.03902 0.06619 -0.16895 0.00389
8 0.06125 0.03850 0.06610 -0.16960 0.00178
9 0.06154 0.03822 0.06603 -0.16986 0.00089
It may happen that F : (X, d) →(X, d) is not a contraction but some power
of it is a contraction. We still get a unique fixed point, as we see in the following
corollary.
Corollary 2.4.2 Let (X, d) be a complete metric space. F : (X, d) →(X, d) be
such that F
N
is a contraction for some positive integer N. Then F has a unique
fixed point.
68 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Proof : By the contraction mapping principle, F
N
has a unique fixed point
x

. That is, ∃ x

∈ X such that F
N
x

= x

. Applying F to both sides, we get
F
N+1
(x

) = F(x

) and hence F
N
(F(x

)) = F(x

). This implies that F(x

) is
also a fixed point of F
N
and uniqueness implies that Fx

= x

. Also every fixed
point of F is also a fixed point of F
N
and hence fixed point of F is unique.
Example 2.4.2 Let F : C[a, b] →C[a, b] be defined by
[Fx](t) =

t
a
x(s)ds
We have
[F
k
x](t) =
1
k −1!

t
a
(t −s)
k−1
x(s)ds
This gives
[(F
k
x −F
k
y)(t)[ ≤
1
k −1!

t
a
(t −s)
k−1
[x(s) −y(s)[ds

(b −a)
k
k!
sup
s∈[a,b]
[x(s) −y(s)[
=
(b −a)
k
k!
|x −y|
This implies that
|F
k
x −F
k
y| ≤
(b −a)
k
k!
|x −y| ∀ x, y ∈ X = C[a, b]
As
(b −a)
k
k!
→0 as k →∞, it follows that ∃ positive integer N such that
(b −a)
N
N!
< 1. Thus there exist a positive integer N such that α =
(b −a)
N
N!
< 1
and |F
N
x − F
N
y| ≤ α|x − y| ∀ x, y ∈ X = C[a, b]. That is F
N
is a
contraction. However, F is not a contraction if (b −a) > 1.
At times, we shall encounter mappings in a normed space, which are not
contractions but have compactness property. For such operators, we have the
following Schauder’s theorem, giving a fixed point of F.
Theorem 2.4.2 Let K be a nonempty closed convex subset of a normed space
X. Let F be a continuous mapping of K into a compact subset of K. Then F
has a fixed point in K.
Another important notion which will be used in solvability analysis of differ-
ential equation is the concept of monotone operators in Hilbert spaces. We shall
now define this concept and obtain some abstract results which subsequently
will be used in Chapter 3 while dealing with existence and uniqueness theorems.
In the following, the space X is assumed to be a Hilbert space.
2.4. FIXED POINT THEOREMS AND MONOTONE OPERATORS 69
Definition 2.4.3 F : X →X is said to be monotone if
(Fx
1
−Fx
2
, x
1
−x
2
) ≥ 0 for all x
1
, x
2
∈ X (2.4.8)
F is called strictly monotone if the above inequality holds and is strict for x
1
=
x
2
. F is called strongly monotone if ∃ constant c > 0 such that
(Fx
1
−Fx
2
, x
1
−x
2
) ≥ c|x
1
−x
2
|
2
for all x
1
, x
2
∈ X (2.4.9)
If F is not defined on the whole space, then we verify Eq. (2.4.8) - Eq. (2.4.9)
on D(F), the domain of F.
Definition 2.4.4 A multivalued mapping F : X →2
X
is called monotone if
(¯ y
1
− ¯ y
2
, ¯ x
1
, ¯ x
2
) ≥ 0 ∀ ¯ x
1
, ¯ x
2
∈ D(F) and ¯ y
1
∈ F ¯ x
1
, ¯ y
2
∈ F ¯ x
2
A monotone mapping F is called maximal monotone if it has no proper mono-
tone extensions.
Example 2.4.3 A : L
2
[0, 1] be the differential operator −
d
2
dt
2
with D(A) defined
as
D(A) = ¦x ∈ L
2
[0, 1] : ˙ x(t), ¨ x(t) ∈ L
2
[0, 1]; x(0) = x(1) = 0¦
A has dense domain and it is self adjoint
(A¯ x, ¯ y) =

d
2
x
dt
2

, y(t)

=

x, −

d
2
y
dt
2

; x, y ∈ D(A)
A is monotone as
(Ax, x) = −

1
a
d
2
x
dt
2
x(t)dt
=

1
0
(
dx
dt
)
2
dt, ¯ x ∈ D(A)
≥ 0
Example 2.4.4 Let f(s, x) be a function defined on [0, 1] ' to ' be such that
f(s, x) is measurable w.r.t. s for all x ∈ ' and continuous w.r.t. x for almost
all s ∈ [0, 1]. Let F : X = L
2
[0, 1] →L
2
[0, 1] be defined by
[Fx](s) = f(s, x(s))
If we assume a growth condition of the type
[f(s, x)[ ≤ a(s) +b[x[, a ∈ L
2
[0, 1], b > 0
on f, then one can show that F maps X into itself and is continuous and bounded
(refer Joshi and Bose[6]). Further, assume that the mapping x → f(s, x) is
monotone increasing for almost all s ∈ [0, 1]. Then we have
(F ¯ x
1
−F ¯ x
2
, ¯ x
1
− ¯ x
2
) =

1
0
(f(s, x
1
(s)) −f(s, x
2
(s)))(x
1
(s) −x
2
(s))ds

(2.4.10)
As the integrand in the above equation is nonnegative for all x
1
, x
2
∈ X, it
follows that F is monotone.
70 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Example 2.4.5 Let F : X → X be a contraction. Then (I − F) is strongly
monotone.
((I −F)¯ x
1
−(I −F)¯ x
2
, ¯ x
1
− ¯ x
2
) = (¯ x
1
− ¯ x
2
, ¯ x
1
− ¯ x
2
) −(F ¯ x
1
−F ¯ x
2
, ¯ x
1
− ¯ x
2
)
≥ |¯ x
1
− ¯ x
2
|
2
−|F ¯ x
1
−F ¯ x
2
| |x
1
−x
2
|
≥ |¯ x
1
− ¯ x
2
|
2
−α|¯ x
1
− ¯ x
2
|
2
= (1 −α)|¯ x
1
− ¯ x
2
|
2
∀ ¯ x
1
, ¯ x
2
∈ X
We have the following solvability result due to Browder[3], concerning strongly
monotone continuous operators on X.
Theorem 2.4.3 Let F : X → X be a strongly monotone continuous operator.
Then F ¯ x = ¯ y has a unique solution for any y ∈ X. Moreover F
−1
is Lipschitz
continuous and monotone.
In view of Example 2.4.5, we get the following Corollary.
Corollary 2.4.3 Let F : X →X be a contraction on the Hilbert space X. Then
¯ x = F ¯ x + ¯ y has a unique solution ¯ x for every ¯ y ∈ X and this solution ¯ x varies
continuously w.r.t. ¯ y.
Definition 2.4.5 F : X →X is called coercive if lim
¦¯ x¦→∞
(F ¯ x, ¯ x)
|¯ x|
= ∞.
For functions f defined on ', this corresponds to the condition lim
x→∞
f(x) = ∞
and lim
x→−∞
f(x) = −∞.
We have the following result due to Minty [9].
Theorem 2.4.4 Let F : X →X be continuous, monotone and coercive. Then
R(F) = X.
While dealing with differential equations, we shall encounter operators, which
are only densely defined. For example, F = L + F
0
where L is densely defined
linear maximal monotone operator and F
0
is monotone operator defined on the
whole space. Then the following theorem gives the surjectivity of F = L +F
0
.
Refer Joshi and Bose [6] for the proof.
Theorem 2.4.5 Let F
1
be a maximal monotone operator with dense domain
containing 0 and F
2
a single valued operator which is defined on the whole space
and is continuous, bounded and coercive. Then R(F) = X where F = F
1
+F
2
.
2.5. DIRAC-DELTA FUNCTION 71
2.5 Dirac-Delta Function
During the course of our discussion on boundary value problems, we shall en-
counter Dirac-delta function δ(x), which is a symbolic function representing the
charge density corresponding to a unit charge at the origin. To specify it, we
take a continuously distributed charge on the real axis with charge density
ρ

(t) =
1
π
¸

t
2
+
2

> 0 and ρ

(t) is small for small but for a peak height
1
π
at the origin.
The cumulative charge distribution γ

(t) is given by
γ

(t) =

t
−∞
ρ

(x)dx =
1
2
+
1
π
tan
−1
(t/)
Hence
lim
t→∞
γ

(t) =


−∞
ρ

(x)dx = 1 for all
So, the total charge for the densily distribution ρ

(t) on the line is 1, whereas
lim
→0
ρ

(t) =

0, t = 0
∞, t = 0
This limiting charge is denoted by δ(t). That is
δ(t) = lim
→0
ρ

(t)
So, vaguely the Dirac-delta function δ(t) can be defined as
δ(t) =

0, t = 0
∞, t = 0
such that


−∞
δ(t)dt = 1
Obviously, this is not a function.
72 CHAPTER 2. MATHEMATICAL PRELIMINARIES
-0.4 -0.2 0.2 0.4
1
2
3
4
5
6
Figure 2.5.1: Charge density
-0.4 -0.2 0.2 0.4
-10
-5
5
10
Figure 2.5.2: Cumulative charge density
The description of a point charge by considering it as a limit of the sequence
ρ

(t) is not quite satisfactory, mathematically. For, one can take various types
of sequences to get the limit δ(t), yet these sequences can behave differently at
t = 0.
To be more mathematically rigorous, we define what we call the action of
ρ

(t) on a class of test functions which consists of functions continuous at the
origin and well behaved at infinity.
Definition 2.5.1 The support of a function f(t) is the closure of the set of all
points where f(t) is different from zero, that is,
sup f = ¦t : f(t) = 0¦
2.5. DIRAC-DELTA FUNCTION 73
Definition 2.5.2 The space T(') of test functions consists of all real valued
functions φ(t) which are infinitely times differentiable and which have compact
support. This space is a vector space with respect to pointwise addition and
scalar multiplication.
Example 2.5.1
φ(t) =

0, t ≥ 1
exp

1
t
2
−1

, [t[ < 1
(2.5.1)
φ(t) is infinite times differentiable and derivatives of all order vanish for [t[ ≥ 1.
Definition 2.5.3 A sequence of function ¦φ
n
(t)¦ in T(') is said to converge
to zero if
(i) all φ
n
(t) have the same compact support,
(ii) sequence ¦φ
n
(t)¦ converges to zero uniformly and for any positive integer
k, the derivatives ¦φ
(k)
n
(t)¦ also converges uniformly to zero.
Example 2.5.2 The sequence

φ(t)
n


n=1
with φ(t) defined by Eq. (2.5.1) con-
verges to zero in T('). Whereas, the sequence

φ

t
n

does not converge to
zero in T(') as these functions do not have the same compact support.
Definition 2.5.4 A continuous linear function on the space T(') is called a
distribution. The space of all distributions is denoted by T
t
('). The action of
f ∈ T
t
(') is denoted by (f, φ).
Definition 2.5.5 A sequence of functions ¦f
n
¦

n=1
in T
t
(') is said to converge
if for every φ ∈ T(') the sequence of scalars ¦(f
n
, φ)¦ converges to the limit
¦(f, φ)¦. One can show that f is also a distribution.
Example 2.5.3 Define a functional δ on T(') by 'δ, φ` = φ(0), ∀φ ∈ T(')
It is easy to see that δ is linear and continuous and hence δ ∈ T
t
('). The
distribution functional δ is defined as above is called the Dirac-delta function.
Example 2.5.4 Dipole distribution
˜
δ is defined as (
˜
δ, φ) = φ
t
(0) ∀ φ ∈ T(')
Definition 2.5.6 A locally integrable function f generates a distribution via
the definition
(f, φ) =


−∞
f(t)φ(t) ∀ φ ∈ T(')
This is called a regular distribution. A distribution which is not a regular
distribution is called a singular distribution.
74 CHAPTER 2. MATHEMATICAL PRELIMINARIES
Theorem 2.5.1 The Dirac-delta distribution is a singular distribution.
For the proof of this fact refer Bose and Joshi [2]
We can define translation, scale expansion, derivative of a distribution as follows.
Definition 2.5.7 The translation f(t − a) and scale expansion f(at) of a dis-
tribution is given by
(f(t −a), φ(t)) = (f(t), φ(t +a)) ∀ φ ∈ T(')
(f(at), φ(t)) =
1
[a[
(f(t), φ(t/a)), a = 0 ∀ φ ∈ T(')
Example 2.5.5 The translation δ(t − a) and scale expansion δ(−t) are given
by
(δ(t −a), φ(t)) = (δ, φ(t +a)) = φ(a)
and
(δ(−t), φ(t)) = (δ, φ(−1)) = φ(0)
= (δ, φ(t)) ∀ φ ∈ T(')
Definition 2.5.8 For any distribution f, its derivative f
t
is defined as
(f
t
, φ) = (f, −φ
t
) = −(f, φ
t
) ∀ φ ∈ T(')
By repeated application of the above definition, we get
(f
(n)
, φ) = (−1)
n
(f, φ
(n)
) ∀ φ ∈ T(')
Here f
(n)
=
d
n
f
dt
n
.
Example 2.5.6 Let H(t) be the Heaviside function
H(t) =

0, t < 0
1/2, t = 0
1, t > 0
H(t) generates a distribution given by
(H(t), φ(t)) =


−∞
H(t)φ(t)dt
=


0
φ(t)dt
Its derivative H
t
(t) is given by
(H
t
(t), φ(t)) = −(H, φ
t
)
= −


0
φ
t
(t)dt = φ(0)
= (δ, φ(t))
Hence we have H
t
= δ.
2.5. DIRAC-DELTA FUNCTION 75
Example 2.5.7 Let f(t) be differentiable except at points a
1
, . . . , a
n
where f(t)
jumps by the respective amount ´f
1
, ´f
2
, . . . , ´f
n
. We assume that f
t
is locally
integrable wherever it exists. Then the distributional derivative f
t
is computed
as follows
(f
t
, φ) = −(f, φ
t
) = −


−∞
f(t)φ
t
(t)dt
= −


−∞
f(t)φ
t
(t)dt −

a2
a1
f(t)φ
t
(t)dt . . .


an
f(t)φ
t
(t)dt
Integrating by parts and noting that f(a
+
k
) −f(a

k
) = ´f
k
, we get
(f
t
, φ) =


−∞
df
dt
φ(t)dt +

a2
a1
df
dt
φ(t)dt +. . .
+


an
df
dt
φ(t)dt +
n
¸
k=1
´f
k
φ(a
k
)
Therefore
(f
t
, φ) =


−∞
[f
t
]φ(t)dt +
n
¸
k=1
´f
k
φ(a
k
)
This gives f
t
= [f
t
] +
¸
n
k=1
´f
k
δ(t −a
k
).
Here [f
t
] indicates the usual derivative of f, wherever it exists.
Example 2.5.8 If g(t) is infinitely differentiable, then g(t)f
(n)
(t) can be found
as follows
(gf
(n)
(t), φ) = (f
(n)
, gφ)
= (−1)
n
(f,
d
n
dt
n
(gφ))
Example 2.5.9 Let L be a linear differential operator given by
L = a
0
(t)
d
n
dt
n
+a
1
(t)
d
n−1
dt
n−1
+. . . +a
n−1
(t)
d
dt
+a
n
=
n
¸
k=0
a
n−k
(t)
d
k
dt
k
with coefficients a
k
(t) infinite times differentiable. Then for the distribution f,
we have
(Lf, φ) = (
n
¸
k=0
a
n−k
(t)f
(k)
, φ)
= (−1)
k

f,
d
k
dt
k
(a
n−k
(t)φ(t))

= (f, L

φ)
76 CHAPTER 2. MATHEMATICAL PRELIMINARIES
where L

φ =
n
¸
k=0
(−1)
k
d
k
dt
k
(a
n−k
(t)φ(t)).
The n
th
order differential operator L

is known as the formal adjoint of L.
Example 2.5.10 Consider the sequence ¦sin nt¦

n=1
. As it is locally integrable,
it generates a sequence of distributions. This sequence does not converge point
wise except at t = 0. However, we have
[(sin nt, φ(t))[ = [


−∞
sin ntφ(t)dt[

1
n


−∞
[ cos ntφ
t
(t)[dt

1
n


−∞

t
(t)[dt
→0 as n →∞
So ¦sinnt¦ converges in T
t
(') to zero distribution.
Example 2.5.11 Let f
n
(t) =

0, [t[ ≥
1
n
n
2
, [t[ <
1
n
(f
n
, φ) =


−∞
f
n
(t)φ(t)dt =
1
n

1
n
n
2
φ(t)dt
=
n
2
1
n

1
n
φ(t)dt
This implies that
lim
n→∞
(f
n
, φ) = lim
n→∞
n
2

1/n
−1/n
φ(t)dt
= φ(0) = (δ, φ)
Thus the sequence of regular distribution ¦f
n
¦ converges in T
t
(') to the delta
distribution whereas the corresponding functions converge almost everywhere to
zero.
2.6 Exercises
1. Test the convergence of the following sequences of the functions in the
space L
2
[0, 1]. Compute the limit function, if it exists.
(a) x
n
(t) = sin nπt
2.6. EXERCISES 77
(b) x
n
(t) = t
n
(c) x
n
(t) = t sinnπt
(d) x
n
(t) =
sinnπt
n
2. Let A ∈ '
n·n
and let N(A) and R(A) denote the null space and range
space of A, respectively. Show that
'
n
= N(A) ⊕R(A)
3. Let M ⊂ L
2
[−π, π] be the linear span of
¸
1, t
2
, cos t
¸
. Find an element
in M closest to cos 2t ∈ L
2
[−π, π].
4. Show that
[Lx](t) =

t
0
x(s)ds
is a bounded linear operator in L
2
[0, 1]. Also, find its adjoint L

and show
that L is a singular operator.
5. Let L : L
2
[0, 1] →L
2
[0, 1] be defined by
[Lx](t) =

1
0
(s + sin t)
2
x(s)ds
From the first principle, show that L is compact.
6. Linearize the following nonlinear operators around the given points.
(a) F(x
1
, x
2
) =

exp(x
1
x
2
) + cos x
1
sin x
2
, x
2
2
x
1
exp(x
1
, x
2
)

(x
1
, x
2
) = (0, 0)
(b) G(x
1
, x
2
) = (x
1
+ sin x
1
cos x
2
, cos x
1
sinx
2
+x
1
x
2
)
(x
1
, x
2
) = (0, 0)
7. Determine if the function F : '
3
→'
3
defined as
F(x
1
, x
2
, x
3
) =
1
10

sin
2
x
1
+ cos x
2
cos x
3
, sin x
2
cos x
3
,
cos x
1
cos x
2
cos x
3
)
is a contraction in '
3
w.r.t 1-norm in '
3
.
8. Let f : X → ' (X, a Hilbert space ) be differentiable. Show that f is
convex iff ¯ x →f(¯ x) is monotone for all ¯ x ∈ X.
9. Show that
Ψ(t) =

exp(−(t
−2
))exp(−(t −a)
−2
), 0 ≤ t ≤ a
0, t ≤ 0, or t ≥ a
is a test function.
78 CHAPTER 2. MATHEMATICAL PRELIMINARIES
10. Show that
(a) g(t)δ
t
(t) = g(0)δ
t
(t) −g
t
(t)δ(t), g ∈ C

(b) δ
t
(b −a) =

δ
t
(b −a)δ(b −a)dt
(c)
d
dt
(sgnt) = 2δ(t)
References
[1 ] R B Bapat, Linear Algebra and Linear Models, Hindustan Book Agency,
1993.
[2 ] R. K. Bose and M. C. Joshi, Methods of Mathematical Physics, Tata
McGraw Hill Publishing Co. Ltd., 1984.
[3 ] F. E. Browder, The Solvability of Nonlinear Functional Equations, Duke-
Math Journal 30, 557-566, 1963.
[4 ] S. D. Conte and C. deBoor, Elemenatry Numerical Analysis - An Algo-
rithmic Approach, McGraw Hill Book Co., 1981.
[5 ] Casper Goffman and George Pedrick, First Course in Functional Anal-
ysis, Prentice Hall of India, 1974.
[6 ] Mohan C Joshi and Ramendra K Bose, Some Topics in Nonlinear Func-
tional Analysis, Wiley Eastern Ltd, 1985
[7 ] M. A. Krasnoselskii, Topological Methods in the Theory of Nonlinear
Integral Equation, Pergaman Press, 1964.
[8 ] Balmohan V Limaye, Functional Analysis, New Age International, 1996.
[9 ] G. Minty, Monotone (Nonlinear) Operator in a Hilbert Space, Duke Math
Journal, 29, 341-346, 1962.
[10 ] H. L. Royden, Real Analysis, Macmillan Publications, 1988.
Chapter 3
Initial Value Problems
This chapter primarily deals with the existence and uniqueness theory of initial
value problems. The notions of - Picard iterates, maximal interval of existence,
continuation of solution and continuity of solution with respect to initial condi-
tion - are the prime targets of our investigation.
3.1 Existence and Uniqueness Theorems
Let I be an interval in ' and Ω an open set containing (t
0
, x
0
) ∈ I Ω. Let
f : I Ω →'
n
be a function, not necessarily linear. As defined earlier (Defini-
tion 1.1.1), by Initial Value Problem (IVP), we mean the following differential
equation with initial condition
dx
dt
= f(t, x(t)) (3.1.1a)
x(t
0
) = x
0
(3.1.1b)
Also, let us recall what we mean by the solution of the IVP - Eq. (3.1.1). x(t)
is a solution of Eq. (3.1.1), if ∃ an interval J ⊂ I, containing ‘t
0
’ such that x(t)
is differentiable on J with x(t) ∈ Ω, ∀ t ∈ J and x(t) satisfies Eq. (3.1.1).
The following theorem gives the existence of unique solution of Eq. (3.1.1),
requiring Lipschitz continuity on f(t, x) with respect to x.
Theorem 3.1.1 Let E = I Ω ⊆ ' '
n
be an open set containing (t
0
, x
0
)
and let f : E →'
n
be continuous. Further assume that x →f(t, x) is Lipschitz
continuous :
∃ a constant M such that
[[f(t, x) − f(t, y)[[ ≤ M[[x −y[[ ∀ t ∈ I and ∀ x, y ∈ Ω (3.1.2)
Then the IVP given by Eq. (3.1.1) has a unique solution.
79
80 CHAPTER 3. INITIAL VALUE PROBLEMS
Proof : As shown in Proposition 1.1.1, solvability of Eq. (3.1.1) is equivalent
to the solvability of the following integral equation
x(t) = x
0
+

t
t0
f(s, x(s))ds (3.1.3)
in some appropriate subinterval J ⊂ I.
Let B denote a closed ball with centre at (t
0
, ¯ x
0
) and contained in E. Let
m = sup
(t,x)∈B
[f(t, ¯ x)[. ‘m’ is finite as f is continuous on a compact set B ⊆ I'
n
.
Choose two positive numbers ‘δ’ and ‘τ’ such that Mτ < 1, mτ < δ and
¦ (t, ¯ x) ∈ I '
n
: [t −t
0
[ ≤ τ, [[x −x
0
[[ ≤ δ ¦ ⊆ B.
Denote by J = [t
0
− τ, t
0
+ τ] and Ω
0
= ¦ x ∈ Ω : [[x −x
0
[[ ≤ δ ¦. Define
X ⊂ C(J) to be set of all continuous functions on J with values in Ω
0
⊆ Ω.
Then X is a closed subset of C(J) and hence a complete metric space with
metric induced by the sup-norm of C(J). We define a nonlinear operator F on
X as follows
[Fx](t) = x
0
+

t
t0
f(s, x(s))ds, t ∈ J (3.1.4)
It is clear that Fx is a continuous function defined on J. Also,
|[Fx](t) −x
0
|
J
n = |

t
t0
f(s, x(s))ds|
J
n

t
t0
|f(s, x(s))|
J
nds
≤ m

t
t0
ds
≤ mτ
≤ δ
This implies that [Fx](t) ∈ Ω
0
∀ t ∈ J and hence F maps X into itself.
Further, ∀¯ x, ¯ y ∈ X, we have,
[[Fx(t) −Fy(t)[[
J
n ≤

t
t0
[[f(s, x(s)) −f(s, y(s))[[
J
nds
≤ M

t
t0
[[x(s) −y(s)[[
J
n
ds
≤ M [[x −y[[ (t −t
0
)
≤ τM [[x −y[[
and hence
[[Fx −Fy[[ ≤ τM[[x −y[[ ∀ x, y ∈ X
3.1. EXISTENCE AND UNIQUENESS THEOREMS 81
As α = τM < 1, it follows that F is a contraction on the complete metric space
X. Invoking Theorem 2.4.1, it follows that F has a unique fixed point x

∈ X.
That is, ∃ ¯ x

∈ X = C(J, Ω
0
) such that
x

(t) = Fx

(t) = x
0
+

t
t0
f(s, x

(s))ds
and x

is unique.
This proves that x

(t) is unique solution of Eq. (3.1.3) and hence that of Eq.
(3.1.1).
Corollary 3.1.1 The iterations x
(k)
(t), defined as
x
(k)
(t) = x
0
+

t
t0
f(s, x
(k−1)
(s))ds (3.1.5)
with x
(0)
(t) = x
0
, converge to the unique solution x

(t) of IVP given by Eq.
(3.1.1).
The iterates defined by Eq. (3.1.5) are called Picard iterates.
Proof : This follows from Theorem 2.4.1, wherein the converging iterates are
defined as
x
(k)
= Fx
(k−1)
, x
(0)
∈ X, (arbitrary)
Here F is defined by (3.1.4). We note that x
(0)
(t) = x
0
∈ X = C(J, Ω
0
).
We now consider an n
th
order differential equation
d
n
x
dt
n
+a
n−1
(t)
d
n−1
x
dt
n−1
+ +a
1
(t)
dx
dt
= f(t, x(t)) (3.1.6a)
with initial values
x(0) = x
(0)
1
, x
(1)
(0) = x
(0)
2
, , x
(n−1)
(0) = x
(0)
n
(3.1.6b)
Here x
(k)
(t) refer to the k
th
derivative of x(t).
Corollary 3.1.2 Let Ω ⊆ ' be open and f : I Ω → ' be continuous with
x →f( . , x ) Lipschitz continuous with Lipschitz constant M. Further, assume
that a
i
(t), 1 ≤ i ≤ n −1 are continuous functions on I. Then n
th
order initial
value problem given by Eq. (3.1.6) has a unique solution.
Proof : We shall reduce Eq. (3.1.6) in to an initial value problem in '
n
. For
this, set
x
1
= x, x
2
=
.
x
1
=
.
x, x
3
=
.
x
2
=
..
x
1
=
..
x, , x
n
=
.
x
n−1
Proceeding as in section (1.3), the IVP given by Eq. (3.1.6) reduces to the
following IVP on I Ω Ω Ω ⊆ I '
n
dx
dt
= F(t, x(t)) (3.1.7a)
x(0) = x
0
(3.1.7b)
82 CHAPTER 3. INITIAL VALUE PROBLEMS
Here ¯ x = (x
1
, x
2
, , x
n
), ¯ x
0
=

x
(0)
1
, x
(0)
2
, , x
(0)
n

∈ '
n
and
F(t, x) = (x
2
, x
3
, . . . , x
n
, f(t, x
1
) −a
1
(t)x
2
− −a
n−1
(t)x
n
) . It is clear that
F is continuous on I Ω
n
and further x → F(., x) is Lipschitz continuous as
we see below.
[[F(t, x) −F(t, y)[[
2
J
n
≤ [x
2
−y
2
[
2
+[x
3
−y
3
[
2
+ +[x
n
−y
n
[
2
+ n

[f(t, x
1
) −f(t, y
1
)[
2
+[a
1
(t)[
2
[x
2
−y
2
[
2
+ +[a
n−1
[
2
[x
n
−y
n
[
2

≤ nM
2
[x
1
− y
1
[
2
+ (1 + n[a
1
(t)[
2
)[x
2
−y
2
[
2
+ +

1 + n[a
n−1
(t)[
2

[x
n
−y
n
[
2
≤ α
2

[x
1
−y
1
[
2
+[x
2
−y
2
[
2
+ +[x
n
−y
n
[
2

.
where α
2
= max

nM
2
, 1 + na
2
1
, 1 +na
2
2
, , 1 +na
2
n−1

,
with a
i
= sup
t∈I
[a
i
(t)[, 1 ≤ i ≤ n.
Hence we have,
[[F(t, x) −F(t, y)[[
2
J
n
≤ α
2
[[x −y[[
2
J
n
Thus the function F appearing in the IVP given by Eq. (3.1.7), satisfies all
conditions of Theorem 3.1.1 and hence it has a unique solution. This in turn
implies that the n
th
order differential equation given by Eq. (3.1.6) has a unique
solution.
Example 3.1.1
dx
dt
= x
2
, x(0) = 1
Clearly, the function f(x) = x
2
is Lipschitz continuous on a closed, bounded
subset of '. Hence by Theorem 3.1.1 this IVP has a unique solution. The
Picard’s iterates x
(k)
(t) are given by
x
(k)
(t) = x
0
+

t
0
f(s, x
(k−1)
(s))ds
= 1 +

t
0

x
(k−1)
(s)

2
ds
So, we get the following Picard iterates for the above problem.
x
(1)
(t) = 1 +t
x
(2)
(t) = (1 +t +t
2
) +
t
3
3
.
.
.
x
(n)
(t) = (1 +t +t
2
+ +t
k
) +o(t
k+1
)
3.1. EXISTENCE AND UNIQUENESS THEOREMS 83
We can actually compute the solution of the above IVP by the method of
separation of variables and obtain x(t) =
1
1 −t
. The largest interval on which
this solution is defined in (−∞, 1). For small
t
t
t
, the function x(t) has the series
expansion given by
x(t) = 1 +t
2
+t
3
+ +t
n
+ .
(refer Chapter 6).
Thus the Picard iterates x
(k)
(t) do converge to the actual solution x(t) = 1+t
2
+
t
3
+ in a small interval [t[ < 1 around t = 0, as envisaged in the Corollary
3.1.1.
The following graph gives first three Picard iterates.
The actual solution x(t) =
1
1 −t
is plotted in Figure 3.1.2.
-1 -0.75 -0.5 -0.25 0.25 0.5 0.75
t
1
2
3
4
x1
x2
x3
Figure 3.1.1: First three Picard iterates
84 CHAPTER 3. INITIAL VALUE PROBLEMS
-1 -0.75 -0.5 -0.25 0.25 0.5 0.75
t
2
4
6
8
10
xt
Figure 3.1.2: Actual solution x(t) =
1
1 −t
Example 3.1.2
dx
dt
= f(x), x(0) = 0
where f(x) is given by
f(x) =

2

x, x > 0
0, x ≤ 0
As
f(x)
x
=
2

x
, x > 0,
f(x)
x
→∞as x →0
+
, and hence f is not Lipschitz con-
tinuous at ‘0’. Hence Theorem 3.1.1 is not applicable. However, by inspection
one can easily compute the following solutions of above IVP :
x
1
(t) =

t
2
, t > 0
0, t ≤ 0
x
2
(t) = 0 ∀ t ∈ '
That is, the above IVP has at least two distinct solutions or in other words, the
solution is not unique.
Remark 3.1.1 This example concludes that there may exist a solution of IVP
of the type Eq. (3.1.1) without Lipschitz continuity assumptions on the nonlinear
function f(t, ¯ x). However it may not be unique. This point will discussed in
detail in section 3.4.
3.1. EXISTENCE AND UNIQUENESS THEOREMS 85
Example 3.1.3
dx
dt
= f(t, x)
f(t, x) =

2tx
t
2
+x
2
(t, x) = (0, 0)
0 (t, x) = (0, 0)
The solution of the above differential equation passing through (0, 0) is given by
the following curve in (t, x)-plane
(x −c)
2
−t
2
= c
2
, c is arbitrary constant (3.1.8)
The solution curve is plotted as under.
-6 -4 -2 2 4 6
t
-4
-2
2
4
6
8
10
xt
Figure 3.1.3: Solution curve for the above example
It may noted that the above IVP has a unique solution passing through the
point away form (0, 0). It can be easily seen that (t, x) → f(t, x) is not even
continuous at (0, 0). Yet, there are infinitely many solutions passing through
(0, 0), of the above differential equation.
86 CHAPTER 3. INITIAL VALUE PROBLEMS
Example 3.1.4
d
2
x
dt
2
= cos x, x(0) = 0, ˙ x(0) = 0
As f(x) = cos x is Lipschitz continuous, it follows by Corollary 3.1.2 that the
above second order initial value problem has a unique solution. However, a close
form solution of the above problem is difficult to obtain. One can, therefore,
resorts to Picard iterative procedure to obtain iterates x
(k)
(t) converging to the
unique solution of the above second order IVP.
x
1
(t) = x(t), x
2
(t) =
.
x; ¯ x(t) = (x
1
(t), x
2
(t))
d¯ x
dt
= F(¯ x(t)), F(x) = (x
2
, cos x
1
), x(0) =
¯
0
We have
x
(1)
1
(t) =

t
0
x
(0)
2
(s)ds = 0
x
(1)
2
(t) =

t
0
cos x
(0)
1
(s)ds =

t
0
ds = t
x
(2)
1
(t) =

t
0
x
(1)
2
(s)ds =

t
0
sds =
t
2
2
x
(2)
2
(t) =

t
0
cos x
(1)
1
(s)ds =

t
0
ds = t
x
(3)
1
(t) =

t
0
x
(2)
2
(s)ds =

t
0
sds =
t
2
2
x
(3)
2
(t) =

t
0
cos x
(2)
1
(s)ds =

t
0
cos
s
2
2
ds
One can approximate cos (
s
2
2
) by series expansion and obtain x
(3)
2
(t) and

x
(4)
1
(t), x
(4)
2
(t)

,
up to some accuracy.
x
(4)
1
(t) =
t
2
2

t
6
240
+
t
10
34560
x
(4)
2
(t) = t −
t
5
40
+
t
9
3456
3.1. EXISTENCE AND UNIQUENESS THEOREMS 87
-3 -2 -1 1 2 3
t
1
2
3
4
xt
x12
x14
Figure 3.1.4: Graph of iterates x
(2)
1
(t), x
(4)
1
(t)
-3 -2 -1 1 2 3
t
-3
-2
-1
1
2
3
xt
x22
x24
Figure 3.1.5: Graph of iterates x
(2)
2
(t), x
(4)
2
(t)
There do exist examples where the nonlinear function f(t, ¯ x) under consider-
ation is not Lipschitz continuous but satisfies other nice property like monotonic-
ity with respect to the variable ¯ x. We shall now state and prove the existence
and uniquness theorem for IVP of the type given by Eq. (3.1.1) under the
monotonicity assumption.
Using the notion of monotonicity of Chapter 2 (Definition 2.4.3), we shall say
that ¯ x →f(t, ¯ x) is monotone in '
n
if
(f(t, ¯ x
1
) −f(t, ¯ x
2
), x
1
− ¯ x
2
)
J
n ≥ 0 ∀ ¯ x
1
, ¯ x
2
∈ '
n
(3.1.9)
It is said to be strong monotone if ∃ c > 0 such that
(f(t, ¯ x
1
) −f(t, ¯ x
2
), x
1
− ¯ x
2
)
J
n
≥ c |¯ x
1
− ¯ x
2
|
2
J
n ∀ ¯ x
1
, ¯ x
2
∈ '
n
(3.1.10)
88 CHAPTER 3. INITIAL VALUE PROBLEMS
In the IVP given by Eq. (3.1.1), we assume that t
0
= 0 and ¯ x
0
= 0.
Definition 3.1.1 A function ¯ x(t) ∈ L
2
[0, t
f
] is said to satisfy Eq. (3.1.1) in
almost everywhere (a.e) sense if ¯ x(t) is differentiable for all t ∈ [0, t
f
] = I and
satisfies Eq. (3.1.1) for all t ∈ I except on a set of measure zero in I.
Theorem 3.1.2 Let f(t, x) : [0, t
f
] '
n
→ '
n
be continuous and x → f(t, ¯ x)
be strongly monotone and further it satisfies a growth condition of the form
|f(t, x)|
J
n ≤ a(t) +b|x|
J
n, a ∈ L
2
[0, t
f
], b > 0 (3.1.11)
Then there exist unique ¯ x(t) ∈ L
2
[0, t
f
] satisfying Eq. (3.1.1) in almost every-
where sense.
Proof : We are interested in the solvability of the IVP of the form
dx
dt
+f(t, x) = 0 (3.1.12)
x(0) = 0
in the interval I = [0, t
f
], in almost every where sense. Set X = L
2
[0, t
f
]. Here
L
2
[0, t
f
] is defined to be the set of all square integrable functions with values in
'
n
. Define L as L¯ x =
d¯ x
dt
with D(L) = ¦x ∈ X :
d¯ x
dt
∈ X and x(0) = 0¦. Then
D(L) is dense in X and L is monotone as
(L¯ x, ¯ x) = (Lx, ¯ x)
L2[0,t
f
]
=

t
f
0

d¯ x
dt
, x(t)

J
n
dt
= [[x(t
f
)[[
2
J
n −

t
f
0

x(t),
d¯ x
dt

J
n
dt
= [[¯ x(t
f
)[[
2
J
n −(L¯ x, ¯ x)
This implies that
(L¯ x, ¯ x) =
|x(t
f
)|
2
≥ 0 ∀ x ∈ L
2
[0, t
f
]
Further, one can show that L is maximal monotone as R(L+λI) = X for λ > 0
(refer Pazy[6]).
We now define the nonlinear operator N on X as in Example 2.4.4:
[Nx] = f(t, ¯ x(t)) (3.1.13)
N is a bounded, continuous operator on X = L
2
[0, t
f
] (refer Joshi and Bose
3.2. THE MAXIMAL INTERVAL OF EXISTENCE 89
[4]). Further, it is strongly monotone as we have
(N¯ x
1
−N¯ x
2
, ¯ x
1
− ¯ x
2
) =

t
f
0
(f(t, ¯ x
1
(t)) −f(t, ¯ x
2
(t)), (¯ x
1
(t) − ¯ x
2
(t))) dt
≥ c

t
f
0
[[x
1
(t) −x
2
(t)[[
2
dt
= c[[x
1
−x
2
[[
2
for all ¯ x
1
, ¯ x
2
∈ X.
Thus the solvability of equation Eq. (3.1.12) reduces to the solvability of the
operator equation
Lx +Nx = 0 (3.1.14)
in the space X = L
2
[0, t
f
]. L : D(L) → X is a linear maximal monotone
operator with dense domain. N is a bounded, continuous strongly monotone
operator defined on the whole space X. Hence by Theorem 2.4.5, it follows
that Eq. (3.1.14) has a unique solution and hence the unique solvability of Eq.
(3.1.12), in almost everywhere sense in X.
Example 3.1.5
dx
dt
+αx + cos x = 0
x(0) = 0
f(x) = αx + cos x is strongly monotone if α > 1.
Hence it is follows by Theorem 3.1.2 that the above initial value problem has
a unique solution x(t), (a.e. for t) in a finite subinterval of [0, ∞). However,
Theorem 3.1.1 is not applicable.
3.2 The Maximal Interval of Existence
Theorem 3.1.1 and Theorem 3.1.2 establish the fact that under suitable continu-
ity and Lipschitz continuity amd monotoncity assumptions on f(t, x), the IVP
of the form Eq. (3.1.1) has a unique solution x(t) on some interval surrounding
t
0
. In this section we shall show that the solution x(t) is defined on the maximal
interval of existence (α, β).
Assumption 1: Let Ω ⊆ '
n
be an open set containing x
0
. Let f : 'Ω →'
n
satisfies the assumptions of Theorem 3.1.1
Theorem 3.2.1 Let f satisfies Assumption 1. If x(t), y(t) are solutions of the
IVP given by Eq. (3.1.1) on the intervals I
1
, I
2
. Then t
0
∈ I
1
∩ I
2
and if I is
any open interval around ‘t
0
’ contained in I
1
∩ I
2
, then x(t) = y(t) ∀ t ∈ I.
90 CHAPTER 3. INITIAL VALUE PROBLEMS
Proof : It is clear that t
0
∈ I
1
∩ I
2
. If t
0
∈ I ⊂ I
1
∩ I
2
, then by Theorem
3.1.1, ∃ some open interval I
a
= (a − t
0
, a + t
0
) ⊂ I such that Eq. (3.1.1) has
a unique solution. In view of the uniqueness of solutions on I
a
, x(t) = y(t) on
I
a
⊂ I ⊂ I
1
∩ I
2
.
Define I

= ∪I
a
. Then I

is the largest open interval contained in I on which
x(t) = y(t). If I

is properly contained in I, then one of the end points t

of I

is contained in I ⊂ I
1
∩I
2
. As both x(t) and y(t) are continuous on I, it follows
that,
lim
t→t

x(t) = lim
t→t

y(t) = ¯ x

Now examine the initial value problem
dx
dt
= f(t, x(t)), x(t

) = ¯ x

(3.2.1)
Eq. (3.2.1) has a unique solution on some interval I
0
= (t

− a

, t

+ a

) ⊂ I.
Again by uniqueness on I
0
⊂ I ⊂ I
1
∩I
2
, it follows that, x(t) = y(t) for t ∈ I

∪I
0
and I

is proper subinterval of I

∪I
0
and hence a contradiction of the fact that
I

is the largest open interval contained in I on which x(t) = y(t). Hence I

= I
and thereby implying that x(t) = y(t) ∀ t ∈ I.
Theorem 3.2.2 Let f satisfies Assumption 1. Then for each x
0
∈ Ω, ∃
a maximal interval I

on which the IVP given by Eq. (3.1.1) has a unique
solution. Furthermore, the maximal interval J is open, that is I

= (α, β).
Proof : As in Theorem 3.2.1, we define I

= ∪I
a
where I
a
is an open interval
surrounding t
0
, on which Eq. (3.1.1) has a unique solution x
a
. Define x : I

→Ω
as follows,
x(t) = x
a
(t), t ∈ I
a
This is well defined. for, if t ∈ I
1
∩I
2
, where I
1
and I
2
are two intervals on which
Eq. (3.1.1) has solutions x
1
(t) and x
2
(t). Then by Theorem 3.2.1 it follows that
x
1
(t) = x
2
(t). Also x(t) is a unique solution of Eq. (3.1.1) as x
a
(t) is a unique
solution of Eq. (3.1.1) for t ∈ I
a
.
By definition, I

is maximal. Also let the maximal interval of existence be not
open say of the form (α, β] then we can extend the solution of Eq. (3.1.1) on
(α, β +a] with a > 0 as in the proof of Theorem 3.2.1.
If (α, β) is the maximal interval of existence for IVP given by Eq. (3.1.1), then
t
0
∈ (α, β). The intervals [t
0
, β) and (α, t
0
] are called the right and the left
maximal intervals of existence respectively. Without loss of generality we can
concentrate on the right maximal interval of existence. We wish to find out if
the right maximum interval of existence [t
0
, β) can be extended to [t
0
, ∞). This
will then lead to global solvability of IVP given by Eq. (3.1.1). The following
theorem is an important one in that direction.
3.2. THE MAXIMAL INTERVAL OF EXISTENCE 91
Theorem 3.2.3 Let Ω and f be as defined in Theorem 3.2.2 with x
0
∈ Ω. Let
(α, β) be the maximal interval of existence of the solution x(t) of the IVP given
by Eq. (3.1.1). Assume that β < ∞. Then given any compact set K ⊂ Ω,
∃ t ∈ (α, β) such that x(t) / ∈ K.
Proof : Let K be compact set in '
n
. Since f is continuous on (α, β) K. It
follows that
m = sup
(t,x)∈(α, β)·K
|f(t, x)| < ∞
If possible, let the solution x(t) of Eq. (3.1.1) be such that x(t) ∈ K ∀ t ∈ (α, β)
with β < ∞. For α < t
1
< t
2
< β, we have
|x(t
1
) −x(t
2
)| ≤

t2
t1
|f(s, x(s))|ds
≤ m|t
2
−t
1
|
As t
1
, t
2
→ β, it follows that |x(t
1
) − x(t
2
)| → 0 which by Cauchy’s criterion
of convergence in '
n
implies that lim
t→β−
x(t) exists. Let x
1
be this limit. Since
x(t) ∈ K and K is compact, it follows that x
1
∈ K ⊂ Ω. Now define y(t) on
(α, β] by
y(t) =

x(t), t ∈ (α, β)
x
1
, t = β
y(t) is differentiable on (α, β] with
˙
¯ y(β) = f(β, y(β)), as y(t) satisfies the
integral equation
y(t) = x
0
+

t
t0
f(s, y(s))ds
Thus y is a continuation of the solution x(t) on (α, β].
We now examine the IVP
dx
dt
= f(t, x(t))
¯ x(β) = ¯ x
1

(3.2.2)
By Theorem 3.1.1 the equation (3.2.2) has a unique solution z(t) on some open
interval (β −a, β +a) and hence z(t) = y(t) on (β −a, β) and at the end point
β we have z(β) = x
1
= y(β).
Now we define a function v(t) as
v(t) =

y(t), t ∈ (α, β]
z(t), t ∈ [β, β +a)
Then v(t) is a solution of the IVP given by Eq. (3.1.1) on (α, β + a), contra-
dicting the assumption that (α, β) is the maximal interval of existence for given
IVP. Hence, if β < ∞, ∃ a t ∈ (α, β) such that x(t) / ∈ K.
As a Corollary of the above result, we immediately get the following theorem,
called continuation theorem. The process of continuation is already described
in the proof of Theorem 3.2.3.
92 CHAPTER 3. INITIAL VALUE PROBLEMS
Theorem 3.2.4 Let Ω, f be as defined in Theorem 3.2.3 with x
0
∈ Ω. Let
[t
0
, β) be the right maximal interval of existence of solution x(t) of the IVP
given by Eq. (3.1.1).
Assume that ∃ a compact set K ⊂ Ω such that
¸
y ∈ '
n

y = x(t) for t ∈ [ t
0
, β)
¸
⊂ K.
Then β = ∞. That is,IVP given by Eq. (3.1.1) has a solution x(t) on [t
0
, ∞).
That is, x(t) is a global solution.
Corollary 3.2.1 Suppose Ω = '
n
and satisfies the following condition:
∃ ρ > 0 such that
[[f(t, x)[[
J
n ≤ [a(t)[ [[x[[ ∀ [[x[[
J
n > ρ
with a(t) as a continuous function.Then the initial value problem given by Eq.
(3.1.1) has a global solution.
Proof : We have
dx
dt
= f(t, x(t))
Taking scalar product with x(t) we get

x(t),
dx
dt

=

x(t),
dx
dt

J
n
= (f(t, x(t)), x(t) )
≤ [[f(t, x(t))[[ [[x(t)[[
≤ [a(t)[ [[x(t)[[
2
for [[x(t)[[ ≥ ρ
This implies that
d
dt

[[x(t)[[
2


[a(t)[
2
[[x(t)[[
2
and hence
[[x(t)[[
2
≤ [[x(0)[[
2
e
R
t
t
0
|a(t)|
2
dt
for [[x(t)[[ ≥ ρ
If [[x(t)[[ ≤ ρ for all t, then we apply the previous theorem to K = ¦x : [[x[[ ≤ ρ¦
and get the result. If [[x(t)[[ > ρ, we again apply the same theorem to
K =

x : [[x[[
2
≤ [[x(0)[[
2
e
R
β
t
0
|a(t)|
2
dt
¸
.
Corollary 3.2.2 Let x → f( ., x) be Lipschitz continuous for all x(t) ∈ '
n
.
Then Eq. (3.1.1) has a global solution.
Proof : This follows from the previous corollary.
3.3. DEPENDENCE ON INITIAL CONDITION 93
Corollary 3.2.3 Let f(t, x) : I '
n
→ '
n
be continuous with ¯ x → −f(t, ¯ x)
strongly monotone and f(t, 0) = 0. Then the IVP given by Eq. (3.1.1) has a
global solution.
Proof : We have a solution x(t) of Eq. (3.1.1) satisfying
dx
dt
−f(t, x(t)) = 0
Taking scalar product with x(t), we get

dx
dt
, x(t)

−(f(t, x(t)), x(t))) = 0
As −f strongly monotone, this gives
d
dt

[[x(t)[[
2

= 2 (f(t, x(t)), x(t))
≤ −2c[[x(t)[[
2
This implies that
[[x(t)[[
2
≤ [[x(0)[[
2
e
−2ct
≤ [[x(0)[[
2
∀ t
This gives that x(t) ∈ K = ¦x ∈ '
n
: [[x[[ ≤ [[x(0)[[¦. Hence Theorem 3.2.4,
implies that x(t) is a global solution of Eq. (3.1.1).
For more details on this section, refer Perko[ 7 ].
3.3 Dependence on Initial Condition
In this section we investigate the dependence of the solution of the differential
equation
d¯ x
dt
= f(t, ¯ x(t)) (3.3.1)
on the initial condition ¯ x(t
0
) ∈ '
n
. We first state and prove the following lemma
called Gronwall’s lemma.
Lemma 3.3.1 Suppose that g(t) is a continuous real valued function satisfying
the conditions
(i) g(t) ≥ 0 (ii) g(t) ≤ c +k

t
t0
g(s)ds ∀t ∈ [t
0
, t
1
]
(c, k are positive constants.)
Then, we have
g(t) ≤ c exp(kt) ∀ t ∈ [t
0
, t
1
]
94 CHAPTER 3. INITIAL VALUE PROBLEMS
Proof : Let G(t) = c +k

t1
t0
g(s)ds, t ∈ [t
0
, t
1
].
Then, G(t) ≥ g(t) and G(t) > 0 on [t
0
, t
1
]. By second fundamental theorem of
calculus
˙
G(t) = kg(t)
and hence
˙
G(t)
G(t)
= k
g(t)
G(t)
≤ k ∀ t ∈ [t
0
, t
1
]
This gives
d
dt
(log G(t)) ≤ k
Equivalently
G(t) ≤ G(t
0
) exp (kt) ∀t ∈ [t
0
, t
1
]
which is the required result.
Theorem 3.3.1 Let f(t, ¯ x) satisfy all conditions of Theorem 3.1.1. Then the
solution ¯ x(t) of the differential equation (3.1.1) depends continuously on the
initial condition ¯ x
0
.
Proof : Let ¯ x(t), ¯ y(t) be unique solutions of Eq. (3.1.1), with respect to the
initial conditions ¯ x
0
and ¯ y
0
, respectively. Denote by J the interval of existence
of solutions.
Using the corresponding integral equation analog we have
¯ x(t) = ¯ x
0
+

t
t0
f(s, ¯ x(s)) ds t ∈ J
¯ y(t) = ¯ y
0
+

t
t0
f(s, ¯ y(s)) ds t ∈ J
Substracting the above equations and using triangle inequality we get
[[¯ x(t) − ¯ y(t)[[
J
n ≤ [[¯ x
0
− ¯ y
0
[[
J
n +

t
t0
[[f(s, ¯ x(s)) −f(s, ¯ y(s))[[
J
n ds
Borrowing our earlier notation of Theorem 3.1.1, we have ¯ x, ¯ y ∈ X =
C(J, Ω
0
), Ω
0
⊂ Ω. Using Lipschitz continuity of f and Gronwall inequality,
we get
[[¯ x(t) − ¯ y(t)[[
J
n
≤ [[¯ x
0
− ¯ y
0
[[
J
n
+M

t
t0
[[¯ x(s) − ¯ y(s)[[
J
n
ds (3.3.2)
and
[[¯ x(t) − ¯ y(t)[[
J
n
≤ [[¯ x
0
− ¯ y
0
[[
J
n
exp(Mt) ∀t ∈ J = [t
0
, t
f
] (3.3.3)
3.3. DEPENDENCE ON INITIAL CONDITION 95
This gives that
[[¯ x − ¯ y[[ = sup
t∈J
[[¯ x(t) − ¯ y(t)[[
J
n
≤ [[¯ x
0
− ¯ y
0
[[
J
n
exp (Mt
f
) (3.3.4)
This implies that the mapping ¯ x
0
→ ¯ x of '
n
into X = C(J, Ω
0
) is continuous.
We shall give an alternate proof of the above theorem by using the following
lemma, using fixed points, rather than Gronwall’s lemma.
Lemma 3.3.2 Let (X, d) be a complete metric space and let F, F
k
be contrac-
tion mappings on X with the same contraction constant α < 1 and with the
fixed points x, x
k
respectively. Suppose that lim
k→∞
F
k
y = Fy ∀y ∈ X. Then
lim
k→∞
x
k
= x.
Proof : Using the error estimate given in Theorem 2.4.1 for F
k
, we get
d(x
k
, F
l
k
y) ≤
α
l
1 −α
d(F
k
y, y), for any arbitrary y ∈ X and for any l.
Set l = 0 and y = x. We have
d(x
k
, x) ≤
1
1 − α
d(F
k
x, x) →0, as k →∞
This implies that x
k
→x.
Alternate Proof (of Theorem 3.3.1): Let the space X and constants M, m, δ,
τ be as defined in Theorem 3.1.1 .
Let ¯ x
k
(t), ¯ x(t) be solutions of Eq. (3.3.1) with initial conditions ¯ x
(0)
k
and ¯ x
(0)
respectively and let ¯ x
(0)
k
→ ¯ x
(0)
.
Define F
k
, F : X →X as under
[F
k
¯ x] (t) = ¯ x
(0)
k
+

t
t0
f(s, ¯ x(s)) ds, ¯ x ∈ X
[F ¯ x] (t) = ¯ x
(0)
+

t
t0
f(s, ¯ x(s)) ds, ¯ x ∈ X
We have
[[F
k
¯ x(t) − ¯ x
(0)
[[ ≤ [[¯ x
(0)
k
− ¯ x
(0)
[[ +mτ
< δ for k large ( as ¯ x
(0)
k
→ ¯ x
(0)
)
Thus F
k
maps X to itself and so is the case for F (already proved in section 1).
Also ¯ x
k
, ¯ x are fixed points of F
k
and F, respectively with the same contraction
constant α = τm < 1. Also, F
k
¯ x → F ¯ x for ¯ x ∈ X, as ¯ x
(0)
k
→ ¯ x
(0)
in X. Using
Lemma 3.3.2, we get that the fixed points ¯ x
k
of F
k
and ¯ x of F also converge.
This proves the theorem.
96 CHAPTER 3. INITIAL VALUE PROBLEMS
3.4 Cauchy-Peano Existence Theorem
We shall prove an existence theorem for IVP given by Eq. (3.1.1) with the
assumptions of only continuity and boundedness of f.
For this purpose, we need Schauder’s fixed point theorem - Theorem 2.4.2.
We also need the following theorem, called Ascoli-Arzela theorem ( refer Rudin [
8 ] ), which gives the relative compactness for a class T of continuous functions.
Definition 3.4.1 A class T of continuous functions is said to be equicontinous,
if given > 0, ∃ δ > 0 such that
[[f(t) −f(t
t
)[[ < whenever [t −t
t
[ < δ for all f ∈ T
( δ is independent of f ∈ T ).
Definition 3.4.2 T is said to be uniformly bounded if ∃ M independent of
f ∈ T such that
[[f(t)[[ ≤ M, ∀ t and ∀ f ∈ T
Theorem 3.4.1 (Ascoli-Arzela theorem) Let T be a class of continuous
functions defined over some interval J. Then T is relatively compact iff T is
equicontinous and uniformly bounded.
Theorem 3.4.2 (Cauchy-Peano theorem) Let E be as defined in Theorem
3.1.1 and (t
0
, x
0
) ∈ E. Assume that f : E →'
n
is continuous. Then the initial
value problem given by Eq. (3.1.1) has a solution.
Proof : Let the space X and constants m, τ, δ, be as defined in Theorem
3.1.1
Recall that X is a complete metric space of all continuous mappings from
J = [t
0
− τ, t
0
+ τ] into Ω
0
= ¦x ∈ Ω : [[¯ x − ¯ x
0
[[ ≤ δ¦ and hence is a closed,
convex bounded subset of the space C(J).
As before, define F : X →X as
[F ¯ x(t)] = ¯ x
0
+

t
t0
f(s, ¯ x(s)) ds, t ∈ J
Define T = ¦F ¯ x : ¯ x ∈ X¦
T is a collection of continuous functions defined on J. T is equicontinous, since
[[F ¯ x(t
1
) −F ¯ x(t
2
)[[
J
n

t2
t1
[[f(s, ¯ x(s)) [[
J
n
ds
≤ m[t
1
−t
2
[ ∀ ¯ x ∈ X
Further, T is uniformly bounded as
[[F ¯ x(t)[[
J
n ≤ [[¯ x
0
[[
J
n +

t
t0
[[f(s, ¯ x(s))[[
J
n ds
≤ [[¯ x
0
[[
J
n
+mτ ∀ ¯ x ∈ X
3.5. EXERCISES 97
Hence by Ascoli-Arzela theorem, T is relatively compact.
To apply Schauder’s fixed point theorem, it remains to be shown that F : X →
X is continuous.We first observe the following.
f : J Ω
0
→ '
n
is continuous on a compact set and hence it is uniformly
continuous. This implies that given > 0, ∃ δ() > 0 such that
[[f(s, ¯ x(s)) −f(s, ¯ y(s))[[
J
n <

τ
whenever [[x(s) −y(s)[[
J
n < δ
and this δ depends only on .
We have
[[F ¯ x(t) −F ¯ y(t)[[
J
n ≤

t
t0
[[f(s, ¯ x(s)) −f(s, ¯ y(s))[[
J
n ds
This gives
[[¯ x − ¯ y[[ < δ ⇒[[F ¯ x −F ¯ y[[ ≤

t0+τ
t0
[[f(s, ¯ x(s)) −f(s, ¯ y(s))[[
J
nds


τ
τ =
That is, F : X → X is continuous. Thus F : X → X satisfies all assumptions
and hence has a fixed point ¯ x ∈ X. That is
¯ x(t) = ¯ x
0
+

t
t0
f(s, ¯ x(s))ds, t ∈ J
This proves that Eq. (3.1.1) has a solution.
However, this solution need not be unique. In Example 1.2.1, the existence of
a solution of the initial value problem
dx
dt
= f(x), ¯ x(0) = 0
f(x) =

2

x, x ≥ 0
0, x ≤ 0
is guaranteed by Cauchy-Peano theorem. This solution is not unique, as already
seen before.
For more on existence and uniqueness of IVP refer Amann [1], Arnold [2], Hart-
man [3] and Mattheij and Molenaar [5].
3.5 Exercises
1. Find the regions (in '
n
'
n
) of existence and uniquness of solutions of
the IVP associated with the following differential equations.
98 CHAPTER 3. INITIAL VALUE PROBLEMS
(a)
dx
dt
=
3

xt
(b)
dx
dt
=

[x[
(c)
dx
dt
=


x x ≥ 0


−x x ≤ 0
(d)
dx
dt
=

2x
t
t = 0
0 x = 0
2. (a) Let x(t) be a solution of the differential equation
dx
dt
= (t
2
−x
2
) sin x +x
2
cos x
which vanishes for t = t
0
. Show that x(t) ≡ 0.
(b) Find the solution of the IVP
dx
dt
= x[x[
x(t
0
) = 0
3. Show that the IVP
d
2
x
dt
2
= f(t, x(t))
x(t
0
) = x
01
, ˙ x(t
0
) = x
02
is equivalent to the integral equation
x(t) = x
01
+ (t −t
0
)x
02
+

t1
t0
(t −s)f(s, x(s))ds, t ∈ J
where J is the interval of existence of solution.
4. Let x(t) and y(t) be two solutions of the nonhomogeneous linear differen-
tial equation
d
n
x
dt
n
+a
n−1
(t)
d
n−1
x
dt
n−1
+ +a
1
(t)
dx
dt
+a
0
(t)x = b(t)
in the interval J around the initial point t
0
. Show that
u(t
0
)exp(−2k[t −t
0
[) ≤ u(t) ≤ u(t
0
)exp(k[t −t
0
[)
where u(t) is given by
u(t) =
n−1
¸
j=0

x
(j)
(t) −y
(j)
(t)

3.5. EXERCISES 99
x
(j)
(t) =
d
j
x
dt
j
, y
(j)
=
d
j
x
dt
j
and k = 1 +
n−1
¸
j=1
sup
t∈J
[a
j
(t)[.
5. (a) Show that 0 < x
0
< x(t) < π if 0 < x
0
< π, where x(t) is the solution
of the IVP
dx
dt
= t sin t
x(0) = x
0
(b) Show that x
0
< x(t) < 1 for t ≥ 0 and 0 < x(t) < x
1
for t ≤ 0, if
x(t) is the solution of the IVP
dx
dt
= x −x
2
x(0) = x
0
, 0 < x
0
< 1
6. Consider the IVP
dx
dt
= x
p
x(0) = 1
Find the soultion of this problem for different values of p. Find how the
maximum interval I
max
of existence depends on p.
7. Let x(t) be the solution of Eq. (3.1.1), where x → f(t, x) is Lipschitz
continuous from ' to ' for all t ∈ '.
Let x
1
(t), x
2
(t) be two continuous function on ', such that
dx
1
dt
≤ f(t, x
1
(t))
x
1
(t
0
) ≤ x
0
and
dx
2
dt
≥ f(t, x
2
(t))
x
2
(t
0
) ≥ x
0
Show that
x
1
(t) ≤ x(t) ≤ x
2
(t) ∀t ≥ t
0
8. Find the maximal interval of existence of solution for the IVP
dx
1
dt
= −
x
2
x
3
,
dx
2
dt
= −
x
1
x
3
,
dx
3
dt
= 1
¯ x(
1
π
) = (0, 1,
1
π
)
100 CHAPTER 3. INITIAL VALUE PROBLEMS
9. Show that for the IVP
dx
dt
= ax, x(0) = 1
the Picard iterates x
n
(t) converge to the unique solution e
at
, for more
than one initial guess x
0
(t).
10. Consider the following perturbed IVP (corresponding to the above prob-
lem)
dy
dt
= ax, y(0) = 1 +
Show that [x(t) −y(t)[ ≤ [[e
]a]t
Graphically, show that the above estimate is very accurate for a > 0, but
very inaccurate for a < 0.
References
[1 ] Herbet Amann, Ordinary Differential Equations : Introduction to Non-
linear Analysis, Watter de Gruyter, 1990.
[2 ] V. I. Arnold, Geometrical Methods in the Theory of Ordinary Differential
Equations, Springer-Verlag, 1983.
[3 ] Philip Hartman, Ordinary Differential Equations, Birkauser, 1982.
[4 ] Mohan C Joshi and R. K. Bose, Some Topics in Nonlinear Functional
Analysis, Halsted Press, 1985.
[5 ] R. M. M. Mattheij and J. Molenaar, Ordinary Differential Equations in
Theory and Practice, John Wiley, 1996.
[6 ] A. Pazy, Semigroup of Linear Operators and Applications to Partial Dif-
ferential Equations, Springer - Verlag, 1983.
[7 ] Lawrence Perko, Differential Equations and Dynamical Systems, Springer
- Verlag, 1991.
[8 ] Walter Rudin, Real and Complex Analysis, Tata McGraw Hill Publica-
tions, 1974.
Chapter 4
Linear System And
Transition Matrix
The focus of this chapter is on the linear system of differential equations, wherein
the concept of transition matrix plays a crucial role.
We give the definition of transition matrix, its properties and the method of
computation. Some of the illustrative examples, introduced earlier in Chapter
1, are recalled for demonstrating the methodology involved in computation.
4.1 Linear System
In this chapter we shall be concerned with the properties of the linear system
of the form
dx
dt
= A(t)x(t) +b(t) (4.1.1)
This is a special case of Eq. (3.1.1) of Chapter 3.
Here A(t) is an nn matrix and b(t) is a vector in '
n
, with entries a
ij
(t) and
b
i
(t) respectively, as continuous functions of t. The existence and uniqueness of
global solution of Eq. (4.1.1) with the initial condition
x(t
0
) = x
0
(4.1.1(a))
follows from Theorem 3.1.1.
To study the solution of Eq. (4.1.1), it helps to study the homogeneous system
dx
dt
= A(t)x(t) (4.1.2)
We shall first prove the existence of n linearly independent solutions of Eq.
(4.1.2).
101
102 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
Definition 4.1.1 Let x
(1)
(t), x
(2)
(t), , x
(n)
(t) be n-vector valued functions.
The Wronskian W(t), is defined by
W(t) =

x
(1)
1
(t) x
(2)
1
(t) x
(n)
1
(t)
x
(1)
2
(t) x
(2)
2
(t) x
(n)
2
(t)
.
.
.
.
.
.
.
.
.
x
(1)
n
(t) x
(2)
n
(t) x
(n)
n
(t)

Here x
(i)
=

x
(i)
1
, x
(i)
2
, , x
(i)
n

, 1 ≤ i ≤ n.
Theorem 4.1.1 Let W(t
0
) = 0 for some t
0
∈ J ⊆ '. Then x
(1)
(t), x
(2)
(t),
, x
(n)
(t) are linearly independent in the interval J.
Proof : If possible, let
¸
x
(1)
(t), , x
(n)
(t)
¸
be linearly dependent. This
implies that there exist n constants c
1
, c
2
, , c
n
(not all zero) such that
c
1
x
(1)
(t) +c
2
x
2
(t) + +c
n
x
(n)
(t) =
¯
0 (4.1.3)
Eq. (4.1.3) implies that
c
1
x
(1)
1
(t
0
) + c
2
x
(2)
1
(t
0
) + + c
n
x
(n)
1
(t
0
) = 0
c
1
x
(1)
2
(t
0
) + c
2
x
(2)
2
(t
0
) + + c
n
x
(n)
2
(t
0
) = 0
.
.
. + +
.
.
. +
.
.
.
.
.
.
c
1
x
(1)
n
(t
0
) + c
2
x
(2)
n
(t
0
) + + c
n
x
(n)
n
(t
0
) = 0
(4.1.4)
Eq. (4.1.4) is a linear system of equations in n-unknowns c
1
, c
2
, , c
n
. As
the determinant W(t
0
) of this system is nonzero, it follows that c
i
= 0 for
1 ≤ i ≤ n (as RHS of Eq. (4.1.4) is zero). This is a contradiction. Hence the
set
¸
x
(1)
(t), x
(2)
(t), , x
(n)
(t)
¸
is linearly independent.
This gives us the following theorem.
Theorem 4.1.2 There exists n linearly independent solutions of Eq. (4.1.2) in
J.
Proof : We consider n initial value problems
dx
dt
= A(t)x(t), t ∈ J
x(t
0
) = e
(i)
, 1 ≤ i ≤ n
where e
(i)
are the unit vectors (0, , 1 , , 0) .
4.1. LINEAR SYSTEM 103
By Theorem 3.1.1, the above set of n initial value problems has n unique solu-
tions x
(1)
(t), x
(2)
(t), , x
(n)
(t), with initial conditions x
(i)
(t
0
) = e
(i)
, (1 ≤
i ≤ n). Let
W(t) =

x
(1)
1
(t) x
(2)
1
(t) x
(n)
1
(t)
x
(1)
2
(t) x
(2)
2
(t) x
(n)
2
(t)
.
.
.
.
.
.
.
.
.
.
.
.
x
(1)
n
(t) x
(2)
n
(t) x
(n)
n
(t)

Then
W(t
0
) =

1 0 0
0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 1

= 1
By Theorem 4.1.1, x
(1)
(t), , x
(n)
(t) are linearly independent on J.
Let X be the vector space of solutions of Eq. (4.1.2). The following theorem
gives the dimension of X.
Theorem 4.1.3 The dimension of the solution space X of the linear homoge-
neous system given by Eq. (4.1.2) is n.
Proof : We have already proved in Theorem 4.1.2 that Eq. (4.1.2) has n
linearly independent solutions x
(1)
(t), , x
(n)
(t) with x
(i)
(t
0
) = e
(i)
.
Let x(t) be any general solution of Eq. (4.1.2). Let us assume that x(t
0
) =
x
(0)
, with x
(0)
=

x
(0)
1
, x
(0)
2
, , x
(0)
n

. Consider the function y(t) defined as
y(t) = x
(0)
1
x
(1)
(t) +x
(0)
2
x
(2)
(t) + +x
(0)
n
x
(n)
(t). It is clear that
dy
dt
= A(t)y(t)
Also,
y(t
0
) = x
(0)
1
x
(1)
(t
0
) +x
(0)
2
x
(2)
(t
0
) + +x
(0)
n
x
(n)
(t
0
)
= x
(0)
1
e
(1)
+x
(0)
2
e
(2)
+ +x
(0)
n
e
(n)
=

x
(0)
1
, , .x
(0)
n

= x
(0)
Thus y(t) is a solution of the initial value problem
dx
dt
= A(t)x(t), x(t
0
) = x
(0)
104 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
By uniqueness theorem of IVP, it follows that
x(t) = y(t) =
n
¸
i=1
x
(0)
i
x
(i)
(t)
That is, ¯ x(t) is spanned by n linearly independent solutions ¯ x
(1)
(t), ¯ x
(2)
(t),
, ¯ x
(n)
(t) of Eq. (4.1.2) and hence X has dimension n.
At times, it helps to know the relation between W(t) and the matrix A(t).
The following theorem gives the relation.
Theorem 4.1.4 Let x
(1)
(t), , x
(n)
(t) be solutions of Eq. (4.1.2) and let
t
0
∈ J. Then we have the Abel’s formula
W(t) = W(t
0
) exp

t
t0
Tr [A(s)] ds

(4.1.5)
where Tr [A(t)] is defined as
Tr[A(t)] =
n
¸
i=1
a
ii
(t)
Proof : We have
W(t) =

x
(1)
1
(t) x
(2)
1
(t) x
(n)
1
(t)
x
(1)
2
(t) x
(2)
2
(t) x
(n)
2
(t)
.
.
.
.
.
.
.
.
.
.
.
.
x
(1)
n
(t) x
(2)
n
(t) x
(n)
n
(t)

This gives
dW(t)
dt
=
n
¸
i=1

x
(1)
1
(t) x
(n)
1
(t)
x
(1)
1
(t) x
(n)
1
(t)
.
.
.
.
.
.
.
.
.
x
(1)
i−1
(t) x
(n)
i−1
(t)
.
x
(1)
i
(t)
.
x
(n)
i
(t)
x
(1)
i+1
(t) x
(n)
i+1
(t)
.
.
.
.
.
.
.
.
.
x
(1)
n
(t) x
(n)
n
(t)

(4.1.6)
4.1. LINEAR SYSTEM 105
In view of Eq. (4.1.2), we have
.
x
(k)
i
(t) =
dx
(k)
i
(t)
dt
=
n
¸
j=1
a
ij
(t)x
(k)
j
(t), 1 ≤ k ≤ n
and hence Eq. (4.1.6) gives
dW(t)
dt
=
n
¸
i=1

x
(1)
1
(t) x
(n)
1
(t)
.
.
.
.
.
.
.
.
.
¸
n
j=1
a
ij
(t)x
(1)
j
(t)
¸
n
j=1
a
ij
(t)x
(n)
j
(t)
.
.
.
.
.
.
.
.
.
x
(1)
n
(t) x
(n)
n
(t)

Multiply the first row by a
i1
(t), the second by a
i2
(t) and so on except the i
th
row and substract their sum from the i
th
row. We get
dW(t)
dt
=
n
¸
i=1

x
(1)
1
(t) x
(2)
1
(t) x
(n)
1
(t)
.
.
.
.
.
.
.
.
.
.
.
.
a
ii
(t)x
(1)
j
(t) a
ii
(t)x
(2)
j
(t) a
ii
(t)x
(n)
j
(t)
.
.
.
.
.
.
.
.
.
.
.
.
x
(1)
n
(t) x
(2)
n
(t) x
(n)
n
(t)

That is
dW
dt
=
n
¸
i=1
a
ii
(t)W(t) = W(t)Tr[A(t)]
which gives
W(t) = W(t
0
) exp

t
t0
Tr[A(t)]dt

Corollary 4.1.1 If ¯ x
(1)
(t), , ¯ x
(n)
(t) are linearly independent solutions of
Eq. (4.1.2), then W(t) = 0 on J.
Example 4.1.1 Cosider the n
th
order linear differential equation in J ⊆ '
d
n
x(t)
dt
n
+a
n−1
(t)
d
n−1
x(t)
dt
+ +a
1
(t)
dx(t)
dt
+a
0
x(t) = 0 (4.1.7)
with a
i
(t), 0 ≤ i ≤ n −1 as continuous functions of t ∈ J.
106 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
Proceeding as in Corollary 3.1.2, one can show that Eq. (4.1.7) is equivalent to
the linear first order homogeneous system
dx
dt
= A(t)x(t) (4.1.8)
where
A(t) =

0 1 0
0 0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1
−a
0
−a
1
−a
2
−a
n−1
¸
¸
¸
¸
¸
¸
¸
, x(t) =

x(t)
dx
dt
.
.
.
d
n−1
x(t)
dt
n−1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
Recall that A(t) is the companion matrix of Eq. (4.1.7).
It is now clear from Theorem 4.1.3 that Eq. (4.1.8) has n linearly independent
solutions ¯ x
1
(t), ¯ x
2
(t), , ¯ x
n
(t) and its Wronskian is defined as
W(t) =

x
1
(t) x
2
(t) x
n
(t)
x
(1)
1
(t) x
(1)
2
(t) x
(1)
n
(t)
x
(2)
1
(t) x
(2)
2
(t) x
(n)
n
(t)
.
.
.
.
.
.
.
.
.
x
(n)
1
(t) x
(n)
2
(t) x
(n)
n
(t)

Here x
(n)
k
(t) denotes
d
n
x
k
(t)
dt
n
.
Remark 4.1.1 The converse of Theorem 4.1.1 is not true, as we see from the
following example.
x
(1)
(t) =
¸
t
1

, x
(2)
(t) =
¸
t
2
t

are linearly independent functions defined
on ' but W(t) ≡ 0.
4.2 Transition Matrix
In the previous section we proved that the homogeneous linear system
dx
dt
= A(t)x(t) (4.2.1)
4.2. TRANSITION MATRIX 107
has n linearly independent solutions x
(1)
, x
(2)
, , x
(n)
.
Define a nonsingular matrix Ψ(t) as
Ψ(t) =

x
(1)
(t), x
(2)
(t), , x
(n)
(t)

(4.2.2)
Then, we have
dΨ(t)
dt
=
¸
dx
(1)
(t)
dt
,
dx
(2)
(t)
dt
, ,
dx
(n)
(t)
dt
¸
=

A(t)x
(1)
(t), A(t)x
(2)
(t), , A(t)x
(n)
(t)

That is

dt
= A(t)Ψ(t) (4.2.3)
Definition 4.2.1 A nonsingular matrix Ψ(t) is called a fundamental matrix if
it satisfies the matrix differential equation given by Eq. (4.2.3).
Definition 4.2.2 A nonsingular matrix Φ(t, t
0
) is called principal fundamental
matrix or transition matrix or evolution matrix if it satisfies the matrix differ-
ential equation given by Eq. (4.2.3) with the initial condition Φ(t
0
, t
0
) = I.
That is, transition matrix Φ(t, t
0
), is the solution of the matrix initial value
problem
dΦ(t)
dt
= A(t)Φ(t), Φ(t
o
) = I (4.2.4)
Here the solution Φ(t) of the above differential equation lies in the space ´
n·n
of nn matrices. Note that ´
n·n
is a Hilbert space ( finite dimensional) with
inner product and norm defined as
(A, B)
'n×n
=
n
¸
i=1

¯ a
(i)
,
¯
b
(i)

, |A|
2
'n×n
=
n
¸
i=1
|¯ a
(i)
|
2
where A =

¯ a
(1)
, ¯ a
(2)
, , ¯ a
(n)

and B =

¯
b
(1)
,
¯
b
(2)
, ,
¯
b
(n)

.
Proceeding as before, the solvability of Eq. (4.2.4) is equivalent to the solvability
of the Volterra integral equation
φ(t) = I +

t
t0
A(s)φ(s)ds, t ∈ J = (t
0
, t
f
) ⊂ (t
0
, ∞) (4.2.5)
in the space C[J, ´
n·n
]. Its unique solution φ(t) is given by
φ(t) = I +

¸
n=1

t
t0
A
n
(s)ds
108 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
where A
n
(t) =

t
t0

σ1
t0

σn−1
t0
[A(σ
1
)A(σ
2
) A(σ
n
) dσ
n

1
]. Equiva-
lentely, we have
φ(t) = I +

t
t0
A(σ
1
)dσ
1
+

t
t0

σ1
t0
A(σ
1
)A(σ
2
) dσ
2

1
+ (4.2.6)
Eq. (4.2.5) or Eq. (4.2.6) is called the Peano-Baker series for the transition ma-
trix Φ(t) of the linear differential equation given by Eq. (4.2.1) ( refer Brockett
[2]).
Example 4.2.1 Let A(t) be a constant matrix A. Then the transition matrix
of the homogeneous system given by Eq. (4.2.1) is of the form
Φ(t, t
0
) = I +

t
t0
Ads +

t
t0

s
t0
A
2
dτds +
= I +A(t −t
0
) +A
2
(t −t
0
)
2
2!
+
= exp (A(t −t
0
))
Example 4.2.2 Let A =
¸
0 1
−1 0

A
2
= −I, A
3
= −A, A
4
= I, A
5
= A, , A
2k
= (−1)
k
I, A
2k+1
= (−1)
k
A.
This gives
exp(A(t −t
0
)) = I +A(t −t
0
) −I
(t −t
0
)
2
2!

A(t −t
0
)
3
3!
+
(t −t
0
)
4
4!
+A
(t −t
0
)
5
5!
+
= I
¸
1 −
(t −t
0
)
2
2!
+
(t −t
0
)
4
4!
+

+A
¸
(t −t
0
) −
(t −t
0
)
3
3!
+

=
¸
cos (t −t
0
) sin(t −t
0
)
−sin(t −t
0
) cos (t −t
0
)

The following theorem implies that two different fundamental matrices corre-
sponding to the same system differ by a multiple of a constant nonsingular
matrix.
Theorem 4.2.1 Let Ψ
1
(t) be a fundamental matrix of the system given by Eq.
(4.2.1). Then Ψ
2
(t) is a fundamental matrix of Eq. (4.2.1) iff there exists a
nonsingular constant matrix C such that Ψ
2
(t) = Ψ
1
(t)C.
4.2. TRANSITION MATRIX 109
Proof : Let Ψ
2
(t) = Ψ
1
(t)C. Then

2
(t)
dt
=

1
(t)
dt
C = AΨ
1
(t) C = AΨ
2
(t)
This implies that Ψ
2
(t) is a fundamental matrix of Eq. (4.2.1). Conversely, let
Ψ
2
(t) be another fundamental matrix of Eq. (4.2.1).
Set
C(t) = Ψ
−1
1
(t)Ψ
2
(t)
This gives
Ψ
2
(t) = Ψ
1
(t)C(t)
and

2
(t)
dt
=

1
(t)
dt
C(t) + Ψ
1
(t)
dC(t)
dt
Hence
A(t)Ψ
2
(t) = A(t)Ψ
1
(t)C(t) + Ψ
1
(t)
dC(t)
dt
= A(t)Ψ
2
(t) + Ψ
1
(t)
dC(t)
dt
This implies that
Ψ
1
(t)
dC(t)
dt
= 0
and hence
dC(t)
dt
= 0
Thus C is a constant matrix. Also C is invertible as Ψ
1
(t), Ψ
2
(t) are invertible.
Corollary 4.2.1 If Ψ(t) is a fundamental matrix of Eq. (4.2.1), the transition
matrix Φ(t, t
0
) of Eq. (4.2.1) is given by
Φ(t, t
0
) = Ψ(t)Ψ
−1
(t
0
)
Corollary 4.2.2 The initial value problem
dx
dt
= A(t)x(t), x(t
0
) = x
0
has a solution x(t), given by
¯ x(t) = Ψ(t)Ψ
−1
(t
0
)¯ x
0
where Ψ(t) is a fundamental matrix of Eq. (4.2.1). That is
x(t) = Φ(t, t
0
)x
0
110 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
Proof : We have
dΨ(t)
dt
= A(t)Ψ(t)
This gives
¸
dx
(1)
(t)
dt
,
dx
(2)
(t)
dt

dx
(n)
(t)
dt
¸
=

Ax
(1)
, Ax
(2)
, . . . Ax
(n)

where
Ψ(t) =

x
(1)
, x
(2)
, . . . x
(n)

with ¦x
(i)
¦
n
i=1
linearly independent.
This implies that
dx
i
dt
= Ax
i
, 1 ≤ i ≤ n
Thus we have n linearly independent solutions of Eq. (4.2.1) and as its solution
space is of dimension n, it follows that every solution x(t) of Eq. (4.2.1) is given
by
x(t) = c
1
x
(1)
(t) + +c
n
x
(n)
(t)
That is
x(t) = Ψ(t)c
where ¯ c = (c
1
, c
2
, , c
n
).
Since x(t
0
) = ¯ x
0
, we get
c = Ψ
−1
(t
0
)x
0
This gives
x(t) = Ψ(t)Ψ
−1
(t
0
)x
0
= Φ(t, t
0
)x
0
We also note the following facts.
Remark 4.2.1 If Ψ
1
(t) and Ψ
2
(t) are two fundamental matrices of the systems
dx
dt
= A
1
(t)x(t) and
dx
dt
= A
2
(t)x(t), respectively (A
1
= A
2
). Then, Ψ
1
(t) =
Ψ
2
(t).
This follows from the fact that A
1
(t) =

1
(t)
dt
Ψ
−1
1
(t) and
A
2
(t) =

2
(t)
dt
Ψ
−1
2
(t).
Remark 4.2.2 Φ(t, t
0
) satisfies the following properties
(i) Φ(t, t
0
) = Φ(t, t
1
)Φ(t
1
, t
0
),
4.2. TRANSITION MATRIX 111
(ii) Φ
−1
(t, t
0
) = Φ(t
0
, t)
(iii) Φ(t, t) = I
for all t, t
0
, t
1
∈ J
To realise this, just observe that Φ(t, t
0
) = Ψ(t)Ψ
−1
(t
0
).
We shall say a linear homogeneous differential equation in ¯ p(t) is adjoint equa-
tion associated with the given equation
d¯ x
dt
= A(t)¯ x(t), provided that for any
initial data, the scalar product (x(t), p(t)) is constant for all t.
This gives the following theorem.
Theorem 4.2.2 The adjoint equation associated with
d¯ x
dt
= A(t)¯ x(t)
is
d¯ p
dt
= −A(t)
¯
¯ p(t) (4.2.7)
Proof : We have
d
dt
[(¯ x(t), p(t))] =

d¯ x
dt
, ¯ p(t)

+

¯ x(t),
d¯ p(t)
dt

= (A(t)¯ x(t), ¯ p(t)) +

¯ x(t), −A(t)
¯
¯ p(t)

= ([A(t) −A(t)] ¯ x(t), ¯ p(t))
= 0
This gives
(¯ x(t), ¯ p(t)) = constant
That is Eq. (4.2.7) is the adjoint equation associated with Eq. (4.2.1).
Theorem 4.2.3 If Ψ(t) is a fundamental matrix of the system given by Eq.
(4.2.1), then

Ψ
−1
(t)

¯
is a fundamental matrix of the adjoint system.
Proof : We have Ψ(t)Ψ
−1
(t) = I. Differentiating this equality, we get
dΨ(t)
dt
Ψ
−1
(t) + Ψ(t)

−1
(t)
dt
= 0
and hence
A(t) + Ψ(t)

−1
(t)
dt
= 0
This gives

−1
(t)
dt
= −Ψ
−1
(t)A(t)
112 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
and hence
¸

−1
(t)
dt

¯
= −A(t)
¯

Ψ
−1
(t)

¯
Theorem 4.2.1 gives us the following corollary.
Corollary 4.2.3 If Ψ(t) is a fundamental matrix of Eq. (4.2.1) then χ(t) is
fundamental matrix of its adjoint system given by Eq. (4.2.7) iff
χ(t)
¯
Ψ(t) = C
where C is a constant nonsingular matrix.
Remark 4.2.3 If A(t) is skew symmetric, that is, −[A(t)]
¯
= A(t), then the
transition matrix is orthogonal

[Φ(t, t
0
)]

Φ
¯
(t, t
0
)

= I

. This follows from the
fact that
(i)
d
dt

φ
¯
(t, t
0
)φ(t, t
0
)

= φ
¯
(t, t
0
)

A
¯
(t) +A(t)

φ(t, t
0
) = 0 and
(ii) φ(t
0
, t
0
) = I.
We shall now give the method of variation of parameter to solve the non-
homogeneous system
dx
dt
= A(t)x(t) +b(t) (4.2.8)
We set
x(t) = Φ(t, t
0
)v(t) (4.2.9)
which satisfies Eq. (4.2.8), v(t) to be determined.
Plugging this representation of ¯ x(t) in Eq. (4.2.8), we get
¸
d
dt
Φ(t, t
0
)

v(t) + Φ(t, t
0
)
dv(t)
dt
= A(t) Φ(t, t
0
) v(t) +b(t)
This gives
A(t)Φ(t, t
0
)¯ v(t) + Φ(t, t
0
)
dv(t)
dt
= A(t)Φ(t, t
0
)¯ v(t) +b(t)
and hence
dv(t)
dt
= Φ
−1
(t, t
0
)b(t)
= Φ(t
0
, t)b(t)
4.2. TRANSITION MATRIX 113
Integrating this equation, we get
v(t) = v(t
0
) +

t
t0
Φ(t
0
, s)b(s)ds
As x(t) = Φ(t, t
0
)v(t), it follows x(t
0
) = v(t
0
) and
x(t) = Φ(t, t
0
)v(t
0
) +

t
t0
Φ(t, t
0
)Φ(t
0
, s)b(s)ds
This gives us the variation of parameter formula
x(t) = Φ(t, t
0
)x
0
+

t
t0
Φ(t, s)b(s)ds
Example 4.2.3 Solve
dx
dt
= Ax(t) +b(t)
A =
¸
0 1
−1 0

, b(t) =
¸
t
0

The transition matrix for this system is given by
Φ(t, t
0
) =
¸
cos (t −t
0
) sin (t −t
0
)
−sin(t −t
0
) cos (t −t
0
)

By the method of variation of parameter we have
x(t) = x
(0)
+

t
t0
Φ(t, s)b(s)ds
where x
(0)
= x(t
0
).

t
t0
Φ(t, s)b(s)ds =

t
t0
s cos (t −s)ds

t
t0
−s sin(t −s)ds
¸
¸
¸
=

t
0
sin (t −t
0
) + 1 −cos (t −t
0
)
t
0
cos (t −t
0
) − 1 + sin (t −t
0
)
¸
¸
This gives the solution of the non-homogeneous system as

x
1
(t)
x
2
(t)
¸
¸
=

x
(0)
1
x
(0)
2
¸
¸
¸ +

t
0
sin (t −t
0
) + 1 −cos (t −t
0
)
t
0
cos (t −t
0
) −1 + sin (t −t
0
)
¸
¸
For more on this section refer Agarwal and Gupta [1].
114 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
4.3 Periodic Linear System
Let us first consider a homogenous periodic system
d¯ x
dt
= A(t)¯ x(t), t ∈ ' (4.3.1a)
A(t +T) = A(t), t ∈ ' (4.3.1b)
We shall make use of the following lemma, which we shall state without proof
(refer Coddington and Levinson [3] for further details).
Lemma 4.3.1 Let C be any nonsingular matrix, then ∃ a matrix R such that
C = e
R
.
We observe that Eq. (4.3.1) is a non-autonomous system, yet it is possible
to get a simpler representition of the transition matrix φ(t, t
0
) in view of the
periodicity of A(t).
noindent The Peano-Baker series representation of the transition matrix
φ(t, t
0
) gives
φ(t, t
0
) = I +

t
t0
A(σ
1
) dσ
1
+

t
t0
A(σ
1
)

σ1
t0
A(σ
2
)dσ
2

1
+
As

t+T
t0+T
A(σ) dσ =

t
t0
A(σ) dσ, it follows that φ(t, t
0
) is also periodic of period
T :
φ(t +T, t
0
+T) = φ(t, t
0
).
Theorem 4.3.1 (Floquet-Lyapunov). If A(t + T) = A(t), the associated tran-
sition matrix φ(t, t
0
) of the system given by Eq. (4.3.1) has the form
φ(t, t
0
) = P
−1
(t)e
R(t−t0)
P(t
0
)
where P(t) is a nonsingular periodic matrix function and R is a fixed matrix.
Proof : As φ(T, 0) is a nonsingular matrix, by Lemma 4.3.1, there exists a
matrix R such that φ(T, 0) = e
RT
(redefine R in Lemma 4.3.1).
We now define a nonsingular P(t), through its inverse P
−1
(t) as
P
−1
(t) = φ(t, 0)e
−Rt
Then P
−1
(t +T) is satisfies the relationship
P
−1
(t +T) = φ(t +T, 0)e
−R(t+T)
= φ(t +T, T)φ(T, 0)e
−RT
e
−Rt
= φ(t +T, T)e
−Rt
= P
−1
(t)
4.3. PERIODIC LINEAR SYSTEM 115
That is P(t) is periodic.
Also we have
φ(t, 0) = P
−1
(t)e
Rt
, φ(0, t) = e
−Rt
P(t) (4.3.2)
Eq. (4.3.2) imply that
φ(t, t
0
) = φ(t, 0)φ(0, t
0
)
= P
−1
(t)e
Rt
e
−Rt0
P(t
0
)
= P
−1
(t)e
R(t−t0)
P(t
0
)
This proves the theorem.
It is a pleasant surprise to observe that if we make the transformation
¯ x(t) = P
−1
(t)¯ z(t) (4.3.3)
then ¯ z(t) satisfies the autonomous system
d¯ z
dt
= Rz(t) (4.3.4)
This can be seen as follows. Since

dt
= A(t)φ(t), we have
d
dt

P
−1
(t)e
R(t−t0)
P(t
0
)

= A(t)

P
−1
(t)e
R(t−t0)
P(t
0
)

Equivalently,
¸
d
dt
P
−1
(t)

e
R(t−t0)
P(t
0
) +P
−1
(t) Re
R(t−t0)
P(t
0
)
= A(t)P
−1
(t) e
R(t−t0)
P(t
0
)
which gives
P
−1
(t)R = A(t)P
−1
(t) −
d
dt
P
−1
(t) (4.3.5)
Operating Eq. (4.3.5) on ¯ z(t) we get
P
−1
(t)R¯ z(t) = A(t)P
−1
(t)¯ z(t) −
¸
d
dt
P
−1
(t)

¯ z(t)
= A(t)¯ x(t) −
¸
d
dt
P
−1
(t)

¯ z(t) (4.3.6)
Using transformation given by Eq. (4.3.3) we get
d¯ x
dt
=
¸
d
dt
P
−1
(t)

¯ z(t) +P
−1
(t)
d¯ z
dt
116 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
and hence Eq. (4.3.6) reduces to
P
−1
(t)R¯ z(t) = A(t)¯ x(t) −
d¯ x
dt
+P
−1
(t)
d¯ z
dt
= P
−1
(t)
d¯ z
dt
This gives
d¯ z
dt
= R¯ z(t)
which is the desired result.
Example 4.3.1 Consider the following periodic system
d¯ x
dt
= A(t)¯ x(t)
A(t) =
¸
−2 cos
2
t −1 −sin 2t
1 −sin 2t −2 sin
2
t

A(t + 2π) = A(t)
One can verify that
φ(t) =

e
−2t
cos t −sint
e
−2t
sin t cos t
¸
¸
is a fundamental matrix of this system.
One observes that
φ(t, 0) =
¸
cos t −sint
sin t cos t
¸
e
−2t
0
0 1

= P
−1
(t)e
Rt
where P(t) = φ(t, 0) =
¸
cos t −sin t
sin t cos t

is periodic of period 2π and R =
¸
−2 0
0 0

Let us now examine the non-homogeneous initial value periodic system
d¯ x
dt
= A(t)¯ x(t) +
¯
b(t) (4.3.7a)
¯ x(t
0
) = ¯ x
0
(4.3.7b)
where A(t +T) = A(t),
¯
b(t +T) =
¯
b(t).
We state the following theorem regarding the solvability of Eq. (4.3.7),
refer Brocket[2] for a proof of this theorem.
4.3. PERIODIC LINEAR SYSTEM 117
Theorem 4.3.2 Let φ(t, t
0
) be the transition matrix generated by A(t). Then
the solution of the non-homogeneous system given by Eq. (4.3.7) can be written
as
¯ x(t) = ¯ x
p
(t) +φ(t, t
0
) [ ¯ x
0
− ¯ x
p
(t
0
)] (4.3.8)
iff
¯
b(t) is orthogonal to the solution ¯ p(t) of the adjoint equation
d¯ p
dt
= −A
¯
(t)¯ p(t) (4.3.9)
(
¯
b is orthogonal to ¯ p iff

t0+T
t0

¯ p(σ),
¯
b(σ)

dσ = 0.)
Example 4.3.2 Consider the harmonic oscillator problem with periodic exter-
nal force b(t) (refer Example 1.3.1)
d
2
x
dt
2

2
x(t) = b(t), ω
2
=
λ
ma
This is equivalent to

dx
1
dt
dx
2
dt
¸
¸
¸
¸
¸
=
¸
0 1
−ω
2
0
¸
x
1
x
2

=
¸
0
b(t)

⇐⇒
d¯ x
dt
= A¯ x(t) +
¯
b(t) (4.3.10)
with x
1
(t) = x(t), x
2
(t) = ˙ x(t).
The eigenvalues of A are ±ωi and hence the transition matrix corresponding to
A is given by
φ(t, t
0
) =
¸
1 0
0 ω

cos ω(t −t
0
) sin ω(t −t
0
)
−sinω(t −t
0
) cos ω(t −t
0
)
¸
¸

1 0
0
1
ω
¸
¸
¸
=

cos ω(t −t
0
)
1
ω
sin ω(t −t
0
)
−ω sin ω(t −t
0
) cos ω(t −t
0
)
¸
¸
¸
Hence the solution of the homogeneous equation corresponding to Eq. (4.3.10)
is given by
x
H
(t) =

x
(1)
0
cos ω(t −t
0
) +
x
(2)
0
ω
sin ω(t −t
0
) ,
−x
(1)
0
ω sin ω(t −t
0
) +x
(2)
0
cos ω(t −t
0
)

118 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
This gives
x
H
(t) = x
H
(0) cos ω(t −t
0
) +
˙ x
H
(0)
ω
sin ω(t −t
0
) (4.3.11)
The adjoint system corresponding to Eq. (4.3.10) is given by
d¯ p
dt
=
¸
0 ω
2
−1 0

¯ p(t)
The transition matrix of the adjoint system is

cos ω(t −t
0
) ω sin ω(t −t
0
)

1
ω
sinω(t −t
0
) cos ω(t −t
0
)
¸
¸
Hence the non-homogeneous equation will have a solution iff ¯ p(t) is orthogonal
to
¯
b(t). That is, b(t) satisfies the orthogonality condition

t0+T
t0

cos ω(t −t
0
) ω sinω(t −t
0
)

1
ω
sinω(t −t
0
) cos ω(t −t
0
)
¸
¸

0
b(t)
¸
¸
dt = 0
This implies that b(t) satisfies the conditon

t0+T
t0
sin ω(t −t
0
) b(t)d(t) =

t0+T
t0
cos ω(t −t
0
) b(t)d(t) = 0
where T is the time period

ω
.
4.4 Computation of Transition Matrix
We shall discuss some basic properties of the linear transformation exp Ain order
to facilitate its computaton. We note that exp (A(t −t
0
)) is the representation
of the transition matrix of the system
dx
dt
= Ax(t) (4.4.1)
where A is independent of t.
We note the following properties.
[1] If P and Q are linear transformation on '
n
and S = PQP
−1
, then
exp (S) = P exp(Q)P
−1
.
This follows from the fact
exp S = lim
n→∞
n
¸
k=0
(PQP
−1
)
k
k!
= P

lim
n→∞
n
¸
k=0
Q
k
k!

P
−1
= P exp (Q)P
−1
4.4. COMPUTATION OF TRANSITION MATRIX 119
This implies that if P
−1
AP = diag[λ
j
], then
Φ(t, t
0
) = exp (A(t −t
0
)) = P diag[ exp(λ
j
(t −t
0
)) ] P
−1
Here, diag[λ
j
] represents the diagonal matrix with entries λ
j
.
[2] If P and Q are linear transformation on '
n
which commute then
exp (P +Q) = exp (P) exp (Q)
Example 4.4.1 If A =
¸
a b
0 a

, then
exp (A(t −t
0
)) = exp (a(t −t
0
))
¸
1 b(t −t
0
)
0 1

This follows from the fact that
A = aI +B, B =
¸
0 b
0 0

Also, B commutes with I and B
2
= 0. Using property[2], we get the required
representation as follows,
exp(A(t −t
0
) ) = exp (a(t −t
0
)I ) expB(t −t
0
)
= exp (a(t −t
0
) )[I +B(t −t
0
)]
= exp (a(t −t
0
) )
¸
1 b(t −t
0
)
0 1

[3] If A =
¸
a −b
b a

, then exp(A) = exp (a)
¸
cos b −sinb
sin b cos b

Denote λ = a +ib. By induction, we have
¸
a −b
b a

k
=
¸
Re(λ
k
) −Im(λ
k
)
Im(λ
k
) Re(λ
k
))

This gives us exp(A) as
exp(A) =

¸
k=0

Re

λ
k
k!

−Im

λ
k
k!

Im

λ
k
k!

Re

λ
k
k!

¸
¸
=
¸
Re [exp (λ)] −Im[exp (λ)]
Im[exp (λ)] Re [exp (λ)]

= exp (a)
¸
cos b −sinb
sin b cos b

120 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
In a more general setting, we have the following theorem.
Theorem 4.4.1 If the set of eigenvalues ¦λ
1
, λ
2
, , λ
n
¦ of an nn matrix
A is real and distinct, then any set of corresponding eigenvectors
¸
u
1
, , u
n
¸
forms a basis for '
n
. The matrix P =

u
(1)
, , u
(n)

is invertible and P
−1
AP =
diag [λ
1
, λ
2
, . . . , λ
n
] and hence the transition matrix φ(t, t
0
) is given by
φ(t, t
0
) = P

exp (λ
1
(t −t
0
)) 0 0
0 exp (λ
2
(t −t
0
)) 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 exp λ
n
(t −t
0
)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
P
−1
Proof : We have
A¯ u
(i)
= λ
i
¯ u
(i)
, ¯ u
(i)
=
¯
0, λ
i
= λ
j
We claim that ¯ u
(i)
, , ¯ u
(n)
are linearly independent. We proceed by induction.
Obviously, the result is true for k = 1 as ¯ u
(1)
= 0 is linearly independent. So let
the result be true for k ≤ n −1. We shall show that it is also true for all k = n.
Let
α
1
¯ u
(1)

2
¯ u
(2)
+ +α
n
¯ u
(n)
=
¯
0, α
i
∈ ' (4.4.2)
Operating by A we get
α
1
λ
1
¯ u
(1)

2
λ
2
¯ u
(2)
+ +α
n
λ
n
¯ u
(n)
=
¯
0 (4.4.3)
Multiply Eq. (4.4.2) by λ
n
and subtract from Eq. (4.4.3), we get
α
1

1
−λ
n
)¯ u
(1)

2

1
−λ
n
)¯ u
(2)
+ +α
n−1

n−1
−λ
n
)¯ u
(n−1)
=
¯
0
By induction ¯ u
(1)
, , ¯ u
(n−1)
are linear independent and hence α
i
= 0, 1 ≤ i ≤
n −1.
Putting these values in Eq. (4.4.2) we get α
n
λ
n
¯ u
(n)
= 0 which gives α
n
= 0 as
λ
n
¯ u
(n)
= 0.
Thus ¯ u
(1)
¯ u
(n)
are linear independent by induction. Hence the matrix P =

¯ u
(1)
, , ¯ u
(n)

is nonsingular. Further, we have
AP =

A¯ u
(1)
, , A¯ u
(n)

=

λ
1
¯ u
(1)
, , λ
n
¯ u
(n)

= P

λ
1
.
.
.
λ
n
¸
¸
¸ = P diag [λ
j
]
4.4. COMPUTATION OF TRANSITION MATRIX 121
This implies that P
−1
AP = diag [λ
j
]
Proceeding as before, the transition matrix φ(t, t
0
) is given by
e
A(t−t0)
= P exp ( diag [λ
j
(t −t
0
)]) P
−1
= P

e
λ1(t−t0)
0 0
0 e
λ2(t−t0)
0
.
.
.
.
.
.
.
.
.
0 0 e
λn(t−t0)
¸
¸
¸
¸
¸
P
−1
Example 4.4.2 A =
¸
−1 −3
0 2

λ
1
= −1, λ
2
= 2. A pair of corresponding eigenvectors is (1, 0), (−1, 1).
P =
¸
1 −1
0 1

, P
−1
=
¸
1 1
0 1

This gives
Φ(t −t
0
) = exp (A(t −t
0
))
= P

exp (−(t −t
0
)) 0
0 exp (−2(t −t
0
))
¸
¸
P
−1
=

exp (−(t −t
0
)) exp (−(t −t
0
)) −exp (2(t −t
0
))
0 exp(2(t −t
0
))
¸
¸
If eigenvalues are complex, we state the following theorem concerning the com-
putation of the transition matrix.
Theorem 4.4.2 Let A be a 2n 2n matrix with 2n distinct complex eigen-
values λ
j
= a
j
+ ib
j
, λ
j
= a
j
− ib
j
with corresponding eigenvectors w
(j)
=
u
(j)
± i v
(j)
, 1 ≤ j ≤ n. Then
¸
u
(1)
, v
(1)
, , u
(n)
, v
(n)
¸
is a basis for '
2n
and the matrix P =

u
(1)
, v
(1)
, , u
(n)
, ¯ v
(n)

is invertible and
P
−1
AP = diag
¸
a
j
b
j
−b
j
a
j

with 2 2 blocks along the diagonal . The tran-
siton matrix φ(t, t
0
) has the representation.
φ(t, t
0
) = P exp

diag
¸
a
j
b
j
−b
j
a
j

P
−1
= P diag

¸
(exp(a
j
(t −t
0
)))

cos (b
j
(t −t
0
)) sin (b
j
(t −t
0
))
−sin(b
j
(t −t
0
)) cos (b
j
(t −t
0
))
¸
¸
¸

P
−1
122 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
Example 4.4.3 A =

1 −1 0 0
1 1 0 0
0 0 3 −2
0 0 1 1
¸
¸
¸
¸
A has eigenvalues λ
1,2
= 1±i, λ
3,4
= 2±i. A corresponding pair of eigenvectors
is
w
(1)
= (i, 1, 0, 0) , w
(2)
= (0, 0, 1 +i, 1)
P =

¯ v
(1)
¯ u
(1)
¯ v
(2)
¯ u
(2)

=

1 0 0 0
0 1 0 0
0 0 1 1
0 0 0 1
¸
¸
¸
¸
P
−1
=

1 0 0 0
0 1 0 0
0 0 1 −1
0 0 0 1
¸
¸
¸
¸
P
−1
AP =

1 −1 0 0
1 1 0 0
0 0 2 −1
0 0 1 2
¸
¸
¸
¸
Φ(t, t
0
) =
P

exp (t) cos t −exp (t) sin t 0 0
exp (t) sin t exp (t) cos t 0 0
0 0 exp(2t) cos t −exp(2t) sin t
0 0 exp (2t) sin t exp (2t) cos t
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
P
−1
=

exp(t) cos t −exp(t) sin t 0 0
exp (t) sin t exp (t) cos t 0 0
0 0 exp (2t) (cos t + sin t) −2 exp(2t) sin t
0 0 2 exp(2t) sin t exp (2t) (cos t −sin t)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
For multiple eigenvalues of A we proceed as follows.
4.4. COMPUTATION OF TRANSITION MATRIX 123
Definition 4.4.1 let λ be an eigenvalue of nn matrix A of multiplicity m ≤ n.
Then for k = 1 . . . m, any non zero solution v of
[(A−λI)
k
]v = 0
is called a generalized eigenvector of A.
Definition 4.4.2 A matrix N is said to be nilpotent of order k if N
k−1
=
0 and N
k
= 0.
Theorem 4.4.3 Let A be a real n n matrix with eigenvalues λ
1
, . . . , λ
n
re-
peated according to their multiplicity. Then there exist a basis of generalized
eigenvectors v
(1)
, v
(2)
, . . . , v
(n)
with P = [v
(1)
, v
(2)
, . . . , v
(n)
] invertible and A =
S+N, where P
−1
SP = diag [λ
j
]. Further, the matrix N = A−S is nilpotent of
order k ≤ n and S and N commute. The transition matrix φ(t, t
0
) is computed
as follows
Φ(t, t
0
) =

Pdiag [exp (λ
j
(t −t
0
))] P
−1

¸
1 +N(t −t
0
) + +N
k−1
(t −t
0
)
k−1
(k −1)!

If λ is of multiplicity n, then S = diag[λ].
For proofs of Theorem 4.4.2 and Theorem 4.4.3, refer Perko[4].
Example 4.4.4 A =

0 −2 −1 −1
1 2 1 1
0 1 1 0
0 0 0 1
¸
¸
¸
¸
λ = 1 is of multiplicity 4. S = [I]
4·4
and N = A−S.
N =

−1 −2 −1 −1
1 1 1 1
1 2 1 1
0 1 0 0
¸
¸
¸
¸
N
2
=

−1 −1 −1 −1
0 0 0 0
1 1 1 1
0 0 0 0
¸
¸
¸
¸
N
3
= 0
Φ(t, t
0
) = exp (t −t
0
)
¸
I + N(t −t
0
) +
N
2
(t −t
0
)
2
2!

124 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
=
exp (t −t
0
)

1 −[(t −t
0
) −[2(t −t
0
) −[(t −t
0
) −[(t −t
0
)
+ + + +
(t −t
0
)
2
2
]
(t −t
0
)
2
2
]
(t −t
0
)
2
2
]
(t −t
0
)
2
2
]
(t −t
0
) 1 + (t −t
0
) (t −t
0
) (t −t
0
)
(t −t
0
)
2
2
(t −t
0
) +
(t −t
0
)
2
2
1 +
(t −t
0
)
2
2
(t −t
0
)
2
2
0 0 0 1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
Example 4.4.5 A =

1 0 0
−1 2 0
1 1 2
¸
¸
Here λ
1
= 1, λ
2
= λ
3
= 2
v
(1)
= (1, 1, −2 ), v
(2)
= (0, 0, 1 ), v
(3)
, obtained by solving (A − 2I)
2
v = 0,
is given by
0 =

1 0 0
1 0 0
−2 0 0
¸
¸
v
That is v
(3)
= (0, 1, 0)
So, we get
P =

1 0 0
1 0 1
−2 1 0
¸
¸
P
−1
=

1 0 0
2 0 1
−1 1 0
¸
¸
N =

0 0 0
0 0 0
−1 1 0
¸
¸
Φ(t, 0) = P

exp (t) 0 0
0 exp(2t) 0
0 0 exp (2t)
¸
¸
P
−1
[1 +Nt]
=

exp(t) 0 0
exp (t) −exp (2t) exp (2t) 0
−exp(2t) + (2 −t) exp(2t) t exp(2t) exp (2t)
¸
¸
¸
¸
¸
¸
4.4. COMPUTATION OF TRANSITION MATRIX 125
Example 4.4.6 Let us consider the problem of mechanical oscillations dis-
cussed in Example 1.3.1, which is modelled by the second order differential
equaiton
d
2
x
dt
2
+k
dx
dt

2
x = F(t) (4.4.4)
k is the resistance and ω
2
=
λ
ma
.
This is equivalent to the following non-homogenous linear system
d¯ x
dt
=
¸
0 1
−ω
2
−k

¯ x +
¸
0
F(t)

= A¯ x +
¯
b(t)
A =
¸
0 1
−ω
2
−k

,
¯
b =
¸
0
F(t)

(i) Eigenvalues of A are complex if k
2
−4ω
2
< 0 and are given by

k
2
±
i
2


2
−k
2
= a ±bi

a = −
k
2
and b =

ω
2
−(k
2
/4)

.
Hence the transition matrix is given by
φ(t, t
0
)
=
e
a(t−t0)

1 0

k
2
b
¸
¸
¸

cos b(t −t
0
) sin b(t −t
0
)
−sin b(t −t
0
) cos b(t −t
0
)
¸
¸

1 0
k
2b
1
b
¸
¸
¸
=
e
a(t−t0)

cos b(t −t
0
) +
k
2b
sin b(t −t
0
)
1
b
sin b(t −t
0
)

ω
2
b
sin b(t −t
0
) cos b(t −t
0
) −
k
2b
sinb(t −t
0
)
¸
¸
¸
¸
¸
This gives the solution of the homogeneous equation corresponding to Eq.
(4.4.4), passing through the initial point x(t
0
) = x
(1)
0
, ˙ x(t
0
) = x
(2)
2
as
x(t) = e
a(t−t0)
¸
x(t
0
)

cos b(t −t
0
) +
k
2b
sin b(t −t
0
)

+
˙ x(t
0
)
b
sin b(t −t
0
)

(ii) Eigenvalues of A are real and distinct if k
2
− 4ω
2
> 0 and are given by

k
2
±
1
2

k
2
−4ω
2
=
−k ±c
2
, c =

k
2
−4ω
2
.
126 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
So, the transition matrix corresponding to A is given by
φ(t, t
0
)
=

1 1
−k +c
2
−k −c
2
¸
¸
¸

e

−k +c
2

(t −t
0
)
0
0 e

−k −c
2

(t −t
0
)
¸
¸
¸
¸
¸
¸
¸

k +c
2c
1
c
−k +c
2c

1
c
¸
¸
¸
¸
¸
=

e

−k +c
2

(t −t
0
)
e

−k −c
2

(t −t
0
)

−k +c
2

e

−k +c
2

(t −t
0
)

−k −c
2

e

−k −c
2

(t −t
0
)
¸
¸
¸
¸
¸
¸
¸
¸

k +c
2c
1
c
−k +c
2c

1
c
¸
¸
¸
¸
¸
=

k +c
2c
e

−k +c
2

(t −t
0
)

1
c
e

−k +c
2

(t −t
0
)
+ −
−k +c
2c
e

−k −c
2

(t −t
0
)
¸
¸
¸
1
c
e

−k −c
2

(t −t
0
)
¸
¸
¸

c
2
−k
2
4c
e

−k +c
2

(t −t
0
)

−k +c
2c
e

−k + c
2

(t −t
0
)
+ +
k
2
−c
2
4c
e

−k −c
2

(t −t
0
)
¸
¸
¸
k +c
c
e

−k −c
2

(t −t
0
)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
4.4. COMPUTATION OF TRANSITION MATRIX 127
Hence the solution of the homogeneous equation passing through the initial
point is given by
x(t) =
x(t
0
)

k +c
2c
e

−k +c
2

(t −t
0
)
+
−k + c
2c
e

−k −c
2

(t −t
0
)
¸
¸
¸
+ ˙ x(t
0
)

1
c
e

−k +c
2

(t −t
0
)

1
c
e

−k −c
2

(t −t
0
)
¸
¸
¸
=
e

k
2
(t −t
0
)

x(t
0
)

k +c
2c
e

c
2

(t − t
0
)
+
−k +c
2c
e

−c
2

(t −t
0
)

¸
¸
¸
+ e

k
2
(t −t
0
)
¸
˙ x(t
0
)

1
c
e
c
2
(t −t
0
)

1
c
e

c
2
(t −t
0
)
¸¸
= e

k
2
(t −t
0
)
¸
e
c
2
(t −t
0
)

k +c
2c
x(t
0
) +
1
c
˙ x(t
0
)

¸
+e

k
2
(t −t
0
)

e
−c
2
(t −t
0
)

−k +c
2c
x(t
0
) −
1
c
˙ x(t
0
)

¸
¸
= e

k
2
(t −t
0
)

αe
c
2
(t −t
0
)
+βe
−c
2
(t −t
0
)
¸
¸
,
α =
k +c
2c
x(t
0
) +
˙ x(t
0
)
c
, β =
−k +c
2c
x(t
0
) −
˙ x(t
0
)
c
(iii) If k
2
− 4ω
2
= 0, then the eigenvalues of A are repeated
−k
2
,
−k
2
. So, the
transition matrix is given by
φ(t −t
0
) =

e

−k
2

(t −t
0
)
0
0 e

−k
2

(t −t
0
)
¸
¸
¸
¸
¸
[I +N(t −t
0
)]
128 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
where N is the nilpotent matrix given by
N = A−


k
2
0
0 −
k
2
¸
¸
¸
¸
¸
=

k
2
1
−ω
2

k
2
¸
¸
¸
¸
¸
This gives
φ(t, t
0
) =

e

−k
2

(t −t
0
)
0
0 e

−k
2

(t −t
0
)
¸
¸
¸
¸
¸
¸
¸

1 +
k
2
(t −t
0
) (t −t
0
)
−ω
2
(t −t
0
) 1 −
k
2
(t −t
0
)
¸
¸
¸
¸
¸
=

¸
1 +
k
2
(t −t
0
)

e

−k
2

(t −t
0
)
(t −t
0
)e

−k
2

(t −t
0
)
−ω
2
(t −t
0
)e

−k
2

(t −t
0
)
¸
1 −
k
2
(t −t
0
)

e

−k
2

(t −t
0
)
¸
¸
¸
¸
¸
¸
¸
¸
¸
Hence, the solution of the homogeneous system is given by
x(t) = x(t
0
)

1 +
k
2
(t −t
0
)

e

−k
2

(t −t
0
)
+˙ x(t
0
)e

−k
2

(t −t
0
)
(t −t
0
)
4.4. COMPUTATION OF TRANSITION MATRIX 129
= e

−k
2

(t − t
0
)
x(t
0
) + (t −t
0
)e

−k
2

(t −t
0
)

k
2
x(t
0
) + ˙ x(t
0
)

= αe

−k
2

(t −t
0
)
+β(t −t
0
)e

−k
2

(t −t
0
)
where
α = x(t
0
), β =
k
2
x(t
0
) + ˙ x(t
0
)
Example 4.4.7 Let us consider the linearized satellite problem, discussed ear-
lier in Chapters 1 and 2,
d¯ x
dt
= A¯ x(t) +B¯ u(t)
where ¯ x = (x
1
, x
2
, x
3
, x
4
), ¯ u = (u
1
, u
2
),
A =

0 1 0 0

2
0 0 2ω
0 0 0 1
0 −2ω 0 0
¸
¸
¸
¸
, B =

0 0
1 0
0 0
0 1
¸
¸
¸
¸
To compute the transition matrix of the above problem, we note that A has
eigenvalues 0, 0, ±ωi.
We follow the procedure described in Theorem 4.4.3.
The generalized eigenvectors corresponding to the eigenvalue λ = 0
are (1, 0, 0, −
3
2
ω), (0, 0, 1, 0) and the eigenvectors corresponding to the complex
eigenvalues λ = ±iω are (1, 0, 0, 2ω), (0, ω, 2, 0).
This gives
P =

1 0 1 0
0 0 0 ω
0 1 0 2

3
2
ω 0 −2ω 0
¸
¸
¸
¸
, P
−1
=

4 0 0
2
ω
0 −
2
ω
1 0
−3 0 0 −
2
ω
0
1
ω
0 0
¸
¸
¸
¸
and hence the nilpotent matrix N = A−S, where
S = P diag[B
j
]P
−1
=

1 0 1 0
0 0 0 ω
0 1 0 2

3
2
ω 0 2ω 0
¸
¸
¸
¸

0 0 0 0
0 0 0 0
0 0 0 ω
0 0 −ω 0
¸
¸
¸
¸

4 0 0
2
ω
0 −
2
ω
1 0
−3 0 0 −
2
ω
0
1
ω
0 0
¸
¸
¸
¸
=

0 1 0 0

2
0 0 2ω
6ω 0 0 4
0 −2ω 0 0
¸
¸
¸
¸
130 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
That is
N = A−S =

0 0 0 0
0 0 0 0
−6ω 0 0 −3
0 0 0 0
¸
¸
¸
¸
N is nilpotent matrix of order 2. So, the transition matrix φ(t, t
0
) is given by
φ(t, t
0
)
=

P

1 0 0 0
0 1 0 0
0 0 cos ω(t −t
0
) sinω(t −t
0
)
0 0 −sinω(t −t
0
) cos ω(t −t
0
)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
P
−1

[I +N(t −t
0
)]
=

4 −3 cos ω(t −t
0
)
sin ω(t −t
0
)
ω
0
2
ω
(1 −cos ω(t −t
0
))
3ω sin ω(t −t
0
) cos ω(t −t
0
) 0 2 sinω(t −t
0
)
6 sinω(t −t
0
)
−2
ω
(1 −cos ω(t −t
0
)) 1
4
ω
sin ω(t −t
0
)
6ω(−1 + cos ω(t −t
0
)) −2 sinω(t −t
0
) 0 −3 + 4 cos ω(t −t
0
)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

1 0 0 0
0 1 0 0
−6ω(t −t
0
) 0 1 −3(t −t
0
)
0 0 0 1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
4.5. EULER’S LINEAR SYSTEM 131
=

4 −3 cos ω(t −t
0
)
sinω(t −t
0
)
ω
0
2(1 −cos ω(t −t
0
))
ω
3ω sin ω(t −t
0
) cos ω(t −t
0
) 0 2 sinω(t −t
0
)
[6(−ω)(t −t
0
) −
¸
2
ω
1 −[3(t −t
0
)
+ − −
6 sinω(t −t
0
))]
2
ω
cos ω(t −t
0
)

4
ω
sin ω(t −t
0
)]
[−6ω −2 sinω(t −t
0
) 0 −[3
+ −
6ω cos ω(t −t
0
)] 4 cos ω(t − t
0
)]
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
4.5 Euler’s Linear System
The equation of the form
d¯ x
dt
=
1
t
A¯ x (4.5.1)
where A is an autonomous matrix, is called Euler’s system.
We make a change variable of the form t = e
s
in the independent variable, so
that Eq. (4.5.1) reduces to a linear sytem of the form
d¯ x
dt
= A¯ x (4.5.2)
One can now make use of the linear system theory to compute the solution of
Eq. (4.5.2) and hence that of Eq. (4.5.1)
Example 4.5.1 Solve
t
2
d
2
x
dt
2
−2t
dx
dt
+ 2x = 0 (4.5.3)
Using the notation x
1
= x and x
2
= t
dx
dt
, we get
dx
1
dt
=
1
t
x
2
dx
2
dt
= −
2
t
x
1
+
3
t
x
2
So, we get the Euler’s system
d¯ x
dt
=
1
t

0 1
−2 3

¯ x (4.5.4)
132 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
Making the change of variable t = e
s
we get
d¯ x
ds
= A¯ x(s) (4.5.5)
where A =

0 1
−2 3

has eigenvalues 1,2 with eigenvectors

1
1

,

1
2

.
Hence a fundamental matrix of the system given by Eq. (4.5.5) is
φ(s) =

2 −1
−1 1

e
s
0
0 e
2s

1 1
1 2

=

2e
s
−e
2s
2e
s
−2e
2s
−e
s
+e
2s
−e
s
+ 2e
2s

Hence the solution of Eq. (4.5.4) is given by
¯ x(t) =

¸
c
0
(2t −t
2
) +c
1
(2t − t
2
)
c
0
(−t +t
2
) +c
1
(−t +t
2
)
¸

Example 4.5.2 Solve
t
3
d
3
x
dt
3
+t
dx
dt
−x = 0 (4.5.6)
Using the notation x
1
= x, x
2
= t
dx
dt
, x
3
= t
2
d
2
x
dt
2
, we get that Eq. (4.5.6) is
equivalent to the Euler’s system
d¯ x
dt
=
1
t

¸
0 1 0
0 1 1
1 −1 2
¸

¯ x (4.5.7)
A =

¸
0 1 0
0 1 1
1 −1 2
¸

has 1 as eigenvalues of multiplicity 3.
We use Theorem 4.4.3 to compute a fundamental matrix for the system
d¯ x
dt
= A¯ x(s), A =

¸
0 1 0
0 1 1
1 −1 2
¸

(4.5.8)
S = I, N = A−S =

¸
−1 1 0
0 0 1
1 −1 1
¸

4.6. EXERCISES 133
N is nilpotent of order 3 and hence
φ(s) = exp(s)

I + Ns + N
2
s
2

= exp(s)

¸
1 0 0
0 1 0
0 0 1
¸

+s

¸
−1 1 0
0 0 1
1 −1 1
¸

+ s
2

¸
1 −1 1
1 −1 1
0 0 0
¸

¸
¸
= exp(s)

¸
¸
¸
¸
¸
1 −s +s
2
s −s
2
s
2
s
2
1 −s
2
s + s
2
s −s 1 +s
¸

¸
¸
¸
¸
¸
¸
Hence the solution of the Euler’s system Eq. (4.5.7) is given by
¯ x(t) = t

1 −lnt + (lnt)
2
lnt −(lnt)
2
(lnt)
2
(lnt)
2
1 −(lnt)
2
lnt + (lnt)
2
lnt −lnt 1 +lnt
¸
¸
¸
¸
¸
¸

c
1
c
2
c
3
¸
¸
¸
¸
¸
¸
This gives the solution of Eq. (4.5.6) as
x(t) = c
1
t

1 −lnt + (lnt)
2

+c
2

lnt −(lnt)
2

+c
3
t(lnt)
2
= c
1
t +tlnt(c
2
−c
1
) +t(lnt)
2
(c
1
−c
2
+c
3
)
or
x(t) = d
1
t +d
2
t lnt +d
3
t (lnt)
2
4.6 Exercises
1. If A =
¸
0 1
−1 −2δ

, show that
e
At
=

e
−δt

cos ωt +
δ
ω
sin ωt

1
ω
e
−δt
sin ωt

1
ω
e
−δt
sin ωt e
−δt

cos ωt −
δ
ω
sin ωt

¸
¸
¸
¸
¸
where ω =

1 −δ
2
, δ ≤ 1.
134 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
2. Let A and B be two n n matrices given by
A =

λ 1 0 0
0 λ 1 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1
0 0 0 λ
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
, B =

0 1 0 0
0 0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1
0 0 0 0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
Show that
(a) B
n
= 0.
(b) (λI)B = B(λ)I.
(c) e
At
= e
λt
¸
I +tB +
t
2
2!
B
2
+ +
1
n!
t
n−1
B
n−1

3. Find two linearly independent solutions of the system
d¯ x
dt
= A(t)¯ x(t)
where
A(t) =

0 1

1
t
2

1
t
¸
¸
¸, t = 0
Also, compute the Wronskian of the linearly independent solutions.
4. Compute the transition matrix of the system
d¯ x
dt
=

0 e
t
0 0
¸
¸
¯ x(t)
by using Peano-Baker series.
5. Verify that
φ(t) =

e
−2t
cos t −sin t
e
−2t
sin t cos t
¸
¸
is a fundamental matrix of the system
d¯ x
dt
=

−2 cos
2
t −1 −sin 2t
1 −sin 2t −2 sint
¸
¸
¯ x(t)
4.6. EXERCISES 135
6. Solve the non-homogeneous system
d¯ x
dt
= A(t)¯ x(t) +
¯
b(t), ¯ x(0) = 0
where a(t) is an given in Problem 5 and
¯
b(t) = (1, e
−2t
).
7. Compute the transition matrix of the following system
d¯ x
dt
=

2 1 0
1 3 1
0 1 2
¸
¸
¸
¸
¸
¸
¯ x(t)
8. Solve the initial value problem
d¯ x
dt
= A¯ x(t)
¯ x(0) = (1, t)
where
(a) A =

0 −1 0 0
1 0 0 0
0 0 0 −1
2 0 1 0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
(b) A =

−1 −1 0 0
0 −1 0 0
0 0 0 −2
0 0 1 2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
9. Show that the transition matrices corresponding to autonomous systems
described by two similar matrices are necessarily similar.
10. Verify the following properties of the transition matrix φ(t, t
0
) of the sys-
tem described by Problem 7.
(a) [φ(t, s)]
−1
= φ(s, t)
(b) [φ(t, t
0
)φ(t
0
, s)] = φ(t, s)
(c) φ(t, t) = I
136 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
11. (a) Find two distinct fundamental matrices of the system
d¯ x
dt
=

1 1
0 2

¯ x(t)
(b) Find a fundamental matrix of the adjoint system of the above.
References
[1 ] Ravi P Agarwal and Ramesh C Gupta, Essentials of Ordinary Differen-
tial Equatoins,McGraw Hill Book Company, 1991
[2 ] Roger W Brockett, Finite Dimensional Linear System, John Wiley and
Sons, Inc, 1970
[3 ] Earl A Coddington and Norman Levinson, Theory of Ordinary Differ-
ential Equations, McGraw Hill, 1955
[4 ] Lawrence Perko, Differential Equations and Dynamical Systems, Springer
Verlag, 1991
Chapter 5
Stability Analysis
In this chapter we utilize the theory of transition matrix discussed in Chapter 4,
to study boundedness property of solutions of both linear and nonlinear systems
along with the stability of equilibrium points.
Various notions of stability including the one due to Lyapunov are introduced
and examined leading to some interesting theorems.
Phase portraits, around the equilibrium points in two dimensions, are discussed
in detail. This gives rise to the notion of saddle point, stable and unstable node,
focus and centre. We also sketch phase portraits, vector fields and solution
curves of some illustrative examples.
5.1 Asymptotic Behaviour of Linear System
Theorem 4.4.1 - 4.4.3 can be completely described in terms of the following
theorem, known as Jordan canonical form of a matrix, which can be used in
investigating the behaviour of solution of a linear system of the form given by
Eq. (4.1.1) for large values of t.
Theorem 5.1.1 ( Jordan canonical form)
Let A be a real matrix with real eigenvalues λ
j
, 1 ≤ j ≤ k, and complex eigen-
values λ
j
= a
j
±ib
j
, k + 1 ≤ j ≤ n.
Then there exists a basis ¦¯ v
1
, ..., ¯ v
k
, ¯ v
k+1
, ¯ u
k+1
, ..., ¯ v
n
, ¯ u
n
¦ of '
2n−k
where
¯ v
j
are generalized eigenvectors of A corresponding to real eigenvalues λ
j
(1 ≤
j ≤ k) and ¯ w
j
are generalized eigenvectors of A corresponding to complex
eigenvalues λ
j
(k + 1 ≤ j ≤ n) with ¯ u
j
= Re( ¯ w
j
) and ¯ v
j
= Im( ¯ w
j
). Let
P = [¯ v
1
, ..., ¯ v
k
, ¯ v
k+1
, ¯ u
k+1
, ..., ¯ v
n
, ¯ u
n
].
Then P is invertible and
P
−1
AP =

B
1
B
2
.
.
.
B
r
¸
¸
¸
¸
¸
(5.1.1)
137
138 CHAPTER 5. STABILITY ANALYSIS
The elementary Jordan blocks B = B
j
, j = 1, 2, ..., r are either of the form
B =

λ 1 0
λ 1
.
.
.
.
.
. 1
0 λ
¸
¸
¸
¸
¸
(∗)
for λ as one of the real eigenvalues of A or of the form
B =

D I
2
0
D I
2
.
.
. I
2
0 D
¸
¸
¸
¸
¸
(∗∗)
with
D =
¸
a b
−b a

, I
2
=
¸
1 0
0 1

for λ = a ±ib as a pair of complex eigenvalues of A.
This gives us the solvability of the linear system
d¯ x
dt
= A¯ x(t) (5.1.2)
¯ x(t
0
) = ¯ x
0
Theorem 5.1.2 The solution ¯ x(t) of the system given by Eq. (5.1.2) is of the
form
x(t) = φ(t, t
0
)¯ x
0
where φ(t, t
0
) = φ(t−t
0
) = P diag [exp B
j
(t −t
0
)] P
−1
is the transition matrix
of the system.
(a) If B
j
= B is an mm matrix of the form (*), then B = λI +N and
exp B(t −t
0
) = expλ(t − t
0
) exp N(t −t
0
)
= expλ(t − t
0
)

1 t − t
0
(t−t0)
2
2!

(t−t0)
m−1
(m−1)!
0 1 t −t
0

(t−t0)
m−2
(m−2)!
.
.
.
.
.
.
0 0 (t −t
0
)
0 0 1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
5.1. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM 139
(b) If B
j
= B is 2m2m matrix of the form (**) and λ = a±ib is a complex
pair of eigenvalues, then
exp B(t −t
0
) = exp a(t −t
0
)

R R(t −t
0
) R
(t−t0)
2
2!
R
(t−t0)
m−1
(m−1)!
0 R R(t −t
0
) R
(t−t0)
m−2
(m−2)!
.
.
.
.
.
.
0 0 R R(t −t
0
)
0 0 0 R
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
where R =

cos b(t −t
0
) sinb(t −t
0
)
−sinb(t −t
0
) cos b(t −t
0
)
¸
¸
Corollary 5.1.1 Each component of the solution of the linear system Eq. (5.1.2)
is a linear combination of functions of the form
(t −t
0
)
l
exp a(t −t
0
) cos b(t −t
0
) or (t −t
0
)
l
exp a(t −t
0
) sin b(t −t
0
),
where λ = a ± ib is a pair of eigenvalues of the matrix A and l is smaller than
the dimension of the space.
Let multiplicity of each eigenvalues λ
j
be p
j
where λ
j
= a
j
+ib
j
. Let
a = max
1≤j≤k
a
j
and p = max
1 ≤j ≤l
p
j
Then by Corollary (5.1.1), it follows that ∃ t
1
≥ 0 such that
|φ(t)| = | exp(At)| ≤ c exp(at) t
p
∀ t ≥ t
1
where c is some suitable constant.
Corollary 5.1.2 Every solution of the linear system given by Eq. (5.1.2) tends
to zero as t →∞iff the real parts of the eigenvalues of the matrix A are negative.
Corollary 5.1.3 Every solution of Eq. (5.1.2) is bounded iff the real parts of
the multiple eigenvalues of the matrix A are negative and real parts of the simple
eigenvalues of A are non positive.
Theorem5.1.1 gives us the following well-known theorem called Cayley-Hamilton
theorem.
Theorem 5.1.3 Let p(λ) = det(λI −A), A ∈ '
n
. Then p(A) = 0. That is, if
det(λI −A) = λ
n
+p
n−1
λ
n−1
+ +p
0
, then A
n
+p
n−1
A
n−1
+ +p
0
I = 0.
140 CHAPTER 5. STABILITY ANALYSIS
For the proof of the Theorem 5.1.1 and Theorem 5.1.3, refer Gantmacher[5]
We shall now investigate the following perturbed linear system
d¯ x
dt
= A¯ x +B(t)¯ x (5.1.3)
where B(t) is a continuous matrix on the interval [ t
0
, ∞) . We wish to examine
the relationship between the boundedness of solutions of Eq. (5.1.2) and Eq.
(5.1.3).
Let φ(t, t
0
) = exp(A(t −t
0
)) be the transition matrix of the the linear system
given by Eq. (5.1.2). By variation of parameter formula, the solvability of Eq.
(5.1.3) is given by
¯ x(t) = φ(t, t
0
)¯ x
0
+

t
t0
φ(t, s)B(s)¯ x(s)ds (5.1.4)
We first prove the following variation of Gronwall’s lemma.
Lemma 5.1.1 (Generalized Gronwall’s inequality)
Let u(t), p(t), q(t) be nonnegative continuous functions in the interval I =
[t
0
, t
f
] and let
u(t) ≤ p(t) +

t
t0
q(s)u(s)ds, t ∈ I (5.1.5)
Then we have
u(t) ≤ p(t) +

t
t0
p(τ)q(τ) exp

t
τ
q(s)ds

dτ, t ∈ I (5.1.6)
Proof : Let
r(t) =

t
t0
q(s)u(s)ds
Then r(t
0
) = 0 and ˙ r(t) = q(t)u(t).
By assumption, we have u(t) ≤ p(t) +r(t) and hence
˙ r(t) = q(t)u(t) ≤ p(t)q(t) +r(t)q(t)
This gives
˙ r(t) −r(t)q(t) ≤ p(t)q(t)
and hence multiplying both side by exp

t
t0
q(s)ds

, we get
d
dt
¸
r(t) exp

t
t0
q(s)ds

≤ exp

t
t0
q(s)ds

p(t)q(t)
which implies that
exp

t
t0
q(s)ds

r(t) ≤

t
t0
p(τ)q(τ) exp

τ
t0
q(s)ds


5.1. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM 141
This gives
r(t) ≤

t
t0
p(τ)q(τ)
¸
exp

t0
τ
q(s)ds +

t
t0
q(s)ds


=

t
t0
q(τ)p(τ) exp

t
τ
q(s)ds


and hence
u(t) ≤ p(t) +

t
t0
q(τ)p(τ) exp

t
τ
q(s)ds


Corollary 5.1.4 If p(t) is nondecreasing on I, then
u(t) ≤ p(t) exp

t
t0
q(s)ds

(5.1.7)
Proof : As p(t) is nondecreasing, (5.1.6) gives
u(t) ≤ p(t)
¸
1 +

t
t0
q(τ) exp

t
τ
q(s)ds

= p(t)
¸
1 −

t
t0
d

exp

t
τ
q(s)ds

= p(t)
¸
exp

t
t0
q(s)ds

We are now ready to state and prove the boundedness of solutions of the
perturbed system given by Eq. (5.1.3).
Theorem 5.1.4 Let all solutions of Eq. (5.1.2) be bounded on [t
0
, ∞). Then
all solutions of Eq. (5.1.3) are also bounded provided


t0
|B(t)| ≤ ∞ (5.1.8)
Proof : As all solutions of Eq. (5.1.2) are bounded, it follows that ∃ c such
that
| exp(At)| ≤ c ∀ t ∈ I = [t
0
, ∞).
From Eq. (5.1.4), we have
|¯ x(t)| ≤ | exp(A(t −t
0
))| |¯ x
0
| +

t
t0
|B(s) exp (A(t −s))| |¯ x(s)|ds
≤ c|¯ x
0
| +c

t
t0
|B(s)| |¯ x(s)| ds (5.1.9)
142 CHAPTER 5. STABILITY ANALYSIS
Applying the generalized Gronwall’s inequality given by Eq. (5.1.7) to the above
inequality we have
|¯ x(t)| ≤ c|¯ x
0
|
¸
exp

c

t
t0
|B(s)|ds

≤ c|¯ x
0
| exp(M) ∀ t ∈ I
where
M = exp

c


t0
|B(s)|ds

< ∞
This implies that Eq. (5.1.3) has bounded solutions.
Remark 5.1.1 We examine the boundedness of solutions of the non-homogeneous
system
d¯ x
dt
= A¯ x(t) +
¯
b(t) (5.1.10)
under the assumption that
(a) Eq. (5.1.2) has bounded solutions and
(b)


t0
|
¯
b(s)|ds < ∞.
One can show that the solutions of Eq. (5.1.10) remains bounded. The proof
is similar to that of Theorem 5.1.3. The inequality given by Eq. (5.1.9) in the
proof of Theorem 5.1.3 will be replaced by the following inequality
|¯ x(t)| ≤ c| ¯ x
0
| +c

t
t0
|
¯
b(s)|ds
The following theorem gives asymptotic convergence of solutions of Eq.
(5.1.3) vis-a-vis asymptotic convergence of solutions of Eq. (5.1.2).
Theorem 5.1.5 Let all solutions of Eq. (5.1.2) be such that they tend to zero
as t →∞. Then all solutions of Eq. (5.1.3) are also asymptotically convergent
to zero, provided
|B(t)| →0 as t →∞ (5.1.11)
Proof : By Corollary 5.1.2, it follows that all eigenvalues of A have negative
real parts and hence there exist constants c and δ > 0 such that
| exp(At)| ≤ c exp (−δt) ∀ t ≥ t
0
Using this inequality in Eq. (5.1.4) we get,
|¯ x(t)| ≤ c exp (−δ(t −t
0
)) |¯ x
0
| +c

t
t0
exp (−δ(t −s)) |B(s)| |¯ x(s)|ds
5.1. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM 143
In view of Eq. (5.1.11), given any c
1
> 0, ∃ t
1
≥ t
0
> 0 such that
|B(t)| ≤ c
1
for t ≥ t
1
and hence we get
|¯ x(t)| exp(δt) ≤ c exp(δt
0
) |¯ x
0
| +c

t1
t0
exp (δs) |B(s)| |¯ x(s)| ds
+

t
t1
|B(s)| exp(δs) |¯ x(s)| ds
≤ c exp(δt
0
) |¯ x
0
| +c

t1
t0
exp (δs) |B(s)| |¯ x(s)| ds
+ c

t
t1
c
1
exp(δs) |¯ x(s)| ds.
Let u(t) = exp(δt) [[¯ x(t)[[ and C denote the first two terms of RHS of the above
inequality:
C = c exp (δt
0
) [[¯ x
0
[[ +c

t1
t0
exp (δs) [[B(s)[[ [[¯ x(s)[[ds
Then we get,
u(t) ≤ C +c c
1

t
t1
u(s)ds , t ∈ I
By Gronwall’s lemma we get
u(t) ≤ C exp (

t
t1
c c
1
ds)
= C exp(c c
1
(t −t
1
)) ∀ t ∈ I
This gives
[[¯ x(t)[[ ≤ exp(−δt)C exp(c c
1
(t −t
1
))
= C exp(t(c c
1
− δ) − c c
1
t
1
), ∀ t ∈ I
As c
1
is arbitrary, we choose c
1
such that cc
1
< δ and hence exp (t(c c
1
− δ)) →
0 as t →∞. This gives that |¯ x(t)| →0 as t →∞.
This proves the theorem.
Remark 5.1.2 If Eq. (5.1.2) has solutions asymptoticaly convergent to zero,
then the solutions of non-homogeneous system given by Eq. (5.1.10) satisfies
the same property if |
¯
b(t)| →0 as t →∞.
144 CHAPTER 5. STABILITY ANALYSIS
We now consider the boundedness and stability of solution of non-autonomous
linear system
d¯ x
dt
= A(t)¯ x(t) (5.1.12)
Observe that,
[[¯ x(t)[[
2
= (¯ x(t), ¯ x(t))
Differentiating the above relation, we get
d
dt

[[¯ x(t)[[
2

=

d¯ x(t)
dt
, ¯ x(t)

+

¯ x(t),
d¯ x(t)
dt

= (A(t)¯ x(t), ¯ x(t)) + (¯ x(t), A(t)¯ x(t))
=

(A(t) + [A(t)]
¯
)¯ x(t), ¯ x(t)

A(t) + [A(t)]
¯
is a symmetric matrix and hence its eigenvalues are real. Let
M(t) be the largest eigenvalue of A(t) +[A(t)]
¯
. We have the following theorem
Theorem 5.1.6 (a) If


t0
M(s)ds is bounded then Eq. (5.1.12) has bounded
solutions.
(b) If


t0
M(s)ds = −∞, then Eq. (5.1.12) has solutions asymptotically con-
verging to zero.
Proof : We have
d
dt

[[¯ x(t)[[
2

(A(t) + [A(t)]
¯
¯ x(t), ¯ x(t)

≤ M(t) [[¯ x(t)[[
2
This implies that
[[¯ x(t)[[
2
≤ [[¯ x(0)[[
2
exp

t
t0
M(s)ds

, t ∈ I
If


t0
M(s)ds < ∞, then we get (a). If
lim
t→∞

t
t0
M(s)ds

= −∞ then we get
(b).
Example 5.1.1
d¯ x
dt
= [A+B(t)] ¯ x(t)
A =
¸
0 1
−1 0

and B(t) =
¸
0 0
0
2a
at +b
¸
5.1. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM 145
One can compute that the fundamental matrix of the unperturbed system is
¸
cos t sint
−sint cos t

which is bounded.
The perturbed system has the fundamental matrix as
¸
a sint −(at +b) cos t a cos t + (at +b) sin t
(at +b) sint (at +b) cos t

which is unbounded. We note that

t
0
[[B(s)[[ds →∞ as t →∞.
Example 5.1.2
A(t) =

1
(1 +t)
2
t
2
−t
2
−1
¸
¸
and A(t) +A(t)
¯
=

2
(1 +t)
2
0
0 −2
¸
¸
M(t) =
2
(1 +t)
2
and this gives


0
M(s)ds = 2.
Hence by Theorem 5.1.5 the non-autonomous system
d¯ x
dt
= A(t)¯ x(t)
has bounded solutions.
We shall now consider boundedness and asymptotic stability of the perturbed
system of the form
d¯ x
dt
= [A(t) +B(t)] ¯ x(t) (5.1.13)
corresponding to the non-autonomous system given by Eq. (5.1.12).
Assumptions 1 The matrix A(t) corresponding to the system given by Eq.
(5.1.12) is such that
(a)
lim
t→∞
inf

t
t0
TrA(s)ds > −∞ or TrA(s) = 0 (5.1.14)
(b) The perturbation matrix B(t) satisfies the condition


t0
[[B(t)[[dt < ∞
Theorem 5.1.7 Under Assumptions 1, if all solution of Eq. (5.1.12) are
bounded, then all solutions of Eq. (5.1.13) are also bounded.
146 CHAPTER 5. STABILITY ANALYSIS
Proof : Let ψ(t) be a fundamental matrix of the system Eq. (5.1.12). Then,
the boundedness of solutions implies that ψ(t) is bounded. Also, by Abel’s
lemma (Theorem 4.1.4), we have
det[ψ(t)] = det[ψ(t
0
)]exp(

t
t0
TrA(s)ds)
and hence
ψ
−1
(t) =
adjψ(t)
detψ(t)
=
adjψ(t)
det[ψ(t
0
)]exp(

t
t0
TrA(s)ds)
(5.1.15)
In view of Assumptions 1(a), it follows that [[ψ
−1
(t)[[ is bounded.
We now apply the variation of parameter formula to the IVP corresponding to
Eq. (5.1.13), to get the following integral represntation of the solution ¯ x(t).
¯ x(t) = ψ(t)ψ
−1
(t
0
) ¯ x
0
+

t
t0
ψ(t)ψ
−1
(s)B(s)x(s)ds (5.1.16)
Define
c = max
¸
sup
t≥t0
[[ψ(t)[[, sup
t≥t0
[[ψ
−1
(t)[[
¸
, c
0
= c[[ ¯ x
0
[[.
Then Eq. (5.1.16) gives
[[¯ x(t)[[ ≤ c
0
+c
2

t
t0
[[B(s)[[[[¯ x(s)[[ds
and hence Gronwall’s inequality gives
[[¯ x(t)[[ ≤ c
0
exp

c
2

t
t0
[[B(s)[[ds

< ∞
This proves the theorem.
Theorem 5.1.8 Under Assumptions 1 , if all solution of system Eq. (5.1.12)
are asymptotically convergent to zero, so is the case for all soluiton for Eq.
(5.1.13).
Proof : Asymptotic convergence of solution of Eq. (5.1.12) will imply that
[[ψ(t)[[ −→0 as t −→∞. Using this fact in Eq. (5.1.15), we get that [[ψ
−1
(t)[[ →
0 as t −→∞.
Let c be defined as before. Estimating Eq. (5.1.16), we get
[[¯ x(t)[[ ≤ [[ψ(t)[[
¸
c
1
+c

t
t0
[[B(s)[[[[¯ x(s)[[ds

5.1. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM 147
where c
1
= [[ψ
−1
(t
0
) ¯ x
0
[[. The above inequality gives
[[¯ x(t)[[
[[ψ(t)[[
≤ c
1
+c

t
t0
[[B(s)[[[[ψ(s)[[

[[x(s)[[
[[ψ(s)[[

ds
Put [[y(t)[[ =
[[¯ x(t)[[
[[ψ(t)[[
and apply Gronwall’s inequality to get
[[y(t)[[ ≤ c
1
¸
exp

t
t0
c
2
[[B(s)[[ds

thereby implying that
[[¯ x(t)[[ ≤ c
1
[[ψ(t)[[
¸
exp

t
t0
c
2
[[B(s)[[ds

−→0 as t −→∞
These theorems immediately lead to the following result corresponding to
the asymptotic stability of the non-homogeneous system
d¯ x
dt
= A(t)¯ x(t) +
¯
b(t) (5.1.17)
Assumptions 2
(a) Same as Assumption 1(a)
(b) The function
¯
b(t) is such that

t
t0
[[
¯
b(s)[[ds < ∞
Theorem 5.1.9 Let Assumptions 2 hold. Then
(a) boundedness of solutions of Eq. (5.1.12) will imply boundedness of solu-
tions of Eq. (5.1.17) and
(b) asymptotic convergence of solutions of Eq. (5.1.12) to zero will imply
asymptotic convergence of solutions to zero of the non-homogeneous equa-
tion Eq. (5.1.17).
If the non-autonomous system under consideration is periodic, we can dispatch
with the Assumption 1(a). The boundedness and asymptotic convergence of the
perturbed system Eq. (5.1.13) is preserved under Assumption 1(b).
Theorem 5.1.10 Let A(t) be periodic. Let Assumption 1(b) hold. Then
(a) boundedness of solutions of Eq. (5.1.12) implies boundedness of solutions
of Eq. (5.1.13) and
148 CHAPTER 5. STABILITY ANALYSIS
(b) asymptotic convergence of solutions of Eq. (5.1.12) implies asymptotic
convergence of solution of Eq. (5.1.13).
Proof : By Floquet-Lyapunov theorem - Theorem 4.3.1, the transition matrix
of the periodic system Eq. (5.1.12) is given by
φ(t, t
0
) = P
−1
(t)e
R(t−t0)
P(t
0
)
where P is periodic nonsingular matrix and R is a fixed martix.
By variation of parameter formula applied to the IVP corresponding to Eq.
(5.1.13) we get
¯ x(t) = P
−1
(t)e
R(t−t0)
P(t
0
) ¯ x
0
+

t
t0
P
−1
(t)e
Rt
e
−Rs
P(s)B(s)¯ x(s)ds
This gives
|¯ x(t)| ≤ |P
−1
(t)||e
Rt
||e
−Rt0
P(t
0
)¯ x
0
|
+

t
t0
|P
−1
(t)||e
R(t−s)
||P(s)||B(s)||¯ x(s)|ds (5.1.18)
P(t) is periodic and nonsingular and hence det P(t) = 0 and hence both [[P(t)[[
and [[P
−1
(t)[[ are bounded on [t
0
, ∞). Let
d = max

sup
t≥t0
[[P(t)[[, sup
t≥t0
[[P
−1
(t)[[

Eq. (5.1.18) gives
|¯ x(t)| ≤ d
1
|e
Rt
| +d
2

t
t0
|e
R(t−s)
||B(s)||x(s)|ds (5.1.19)
where d
1
= d[[e
−Rt0
P(t
0
) ¯ x
0
[[.
Boundedness of solutions of Eq. (5.1.12) implies that [[e
Rt
[[ ≤ d
2
and hence Eq.
(5.1.19) gives
[[¯ x(t)[[ ≤ d
1
d
2
+d
2
d
2

t
t0
[[B(s)[[[[¯ x(s)[[ds
and hence
[[¯ x(t)[[ ≤ d
1
d
2
¸
exp

d
2
d
2

t
t0
[[B(s)[[ds

This proves the boundedness of solutions of Eq. (5.1.13).
If solutions of Eq. (5.1.12) are asymptotically convergent to zero, we must
have [[e
Rt
[[ ≤ d
3
e
−αt
for some positive constant d
3
and α. This inequality
immediately gives
[[¯ x(t)[[ ≤ d
1
d
3
e
−αt
+d
2
d
3

t
t0
e
−α(t−s)
[[B(s)[[[[¯ x(s)[[ds
5.2. STABILITY ANALYSIS - FORMAL APPROACH 149
and hence
[[¯ x(t)[[ ≤ d
1
d
3
e
−αt
¸
exp

d
2
d
3

t
t0
[[B(s)[[ds

−→0 as t −→∞
This proves the theorem.
5.2 Stability Analysis - Formal Approach
We shall formally define various concepts of stability of solution of the initial
value problem
d¯ x
dt
= f(t, ¯ x), ¯ x(t
0
) = ¯ x
0
(5.2.1)
We assume that Eq. (5.2.1) has a global solution on [t
0
, ∞).
Definition 5.2.1 Let ¯ x(t, t
0
, ¯ x
0
) denote the solution of the initial value problem
Eq. (5.2.1), indicating its dependence on t and also the initial point t
0
and initial
value ¯ x
0
. This solution is said to be stable if for a given > 0, ∃ δ > 0 such
that
[[∆¯ x
0
[[ < δ ⇒ [[¯ x(t, t
0
, ¯ x
0
+ ∆¯ x
0
) − ¯ x(t, t
0
, ¯ x
0
)[[ <
Definition 5.2.2 ¯ x(t, t
0
, ¯ x
0
) is said to be asymptotically stable if it is stable
and ∃ δ
0
> 0 such that
[[∆¯ x
0
[[ < δ
0
⇒[[¯ x(t, t
0
, ¯ x
0
+ ∆¯ x
0
) − ¯ x(t, t
0
, ¯ x
0
)[[ →0 as t → ∞ (5.2.2)
To be more explicit, Eq. (5.2.2) can be restated as, given µ > 0, there exist δ
and T(µ) such that
[[∆¯ x
0
[[ < δ ⇒[[¯ x(t, t
0
, ¯ x
0
+ ∆¯ x
0
) − ¯ x(t, t
0
, ¯ x
0
)[[ < µ ∀ t ≥ t
0
+T(µ)
Definition 5.2.3 ¯ x(t, t
0
, ¯ x
0
) is said to be uniformly stable if for any > 0,
∃ δ = δ() > 0 such that for any solution ¯ x
(1)
(t, t
0
, ¯ x
1
) of
d¯ x
dt
= f(t, ¯ x), ¯ x(t
0
) = ¯ x
1
[[¯ x(t
1
) − ¯ x
(1)
(t
1
)[[ < δ ⇒[[¯ x(t) − ¯ x
(1)
(t)[[ < ∀ t ≥ t
1
≥ t
0
Theorem 5.2.1 For the linear system
d¯ x
dt
= A(t)x(t) (5.2.3)
stability is equivalent to the boundedness of the solution.
150 CHAPTER 5. STABILITY ANALYSIS
Proof : Let φ(t, t
0
) be the transition matrix. Then,
[[¯ x(t, t
0
, ¯ x
0
+ ∆¯ x
0
) − ¯ x(t, t
0
, ¯ x
0
)[[ = [[φ(t, t
0
)∆¯ x
0
[[ ≤ [[φ(t, t
0
)[[ [[∆¯ x
0
[[
If φ(t, t
0
) is bounded, it follows that the solution ¯ x(t, t
0
, ¯ x
0
) is stable.
Conversely, let the solution of the linear system be stable. Then, given >
0, ∃ δ > 0 such that
[[¯ x(t, t
0
, ¯ x
0
+ ∆¯ x
0
) − ¯ x(t, t
0
, ¯ x
0
)[[ = [[φ(t, t
0
)∆¯ x
0
[[ <
whenever [[∆¯ x
0
[[ < δ.
Choose ∆¯ x
0
=
δ
2
¯ e
(j)
. Then
[[φ(t, t
0
)∆¯ x
0
)[[ =
δ
2
[[φ(t, t
0
) ¯ e
(j)
[[ <
But φ(t, t
0
) ¯ e
(j)
is the j
th
column vector φ
(j)
of the matrix φ(t, t
0
). Hence we
have
[[φ(t, t
0
)[[ = max
1 ≤ j ≤ n
[[φ
(j)
[[ = max
1≤j≤n
[[φ(t, t
0
)¯ e
(j)
[[ ≤
2
δ
This gives
[[¯ x(t, t
0
, ¯ x
0
)[[ = [[φ(t, t
0
)¯ x
0
[[ ≤
2
δ
[[¯ x
0
[[
That is, the solution of the linear system given by Eq. (5.2.3) is bounded.
Corollary 5.2.1 Let φ(t) be a fundamental matrix of the system given by Eq.
(5.2.3). Then Eq. (5.2.3) has asymptotically stable solutions iff [[φ(t)[[ → 0
as t → ∞.
For the uniform stability of the linear system we have the following theorem
Theorem 5.2.2 Let φ(t) be a fundamental matrix of Eq. (5.2.3). Then all
solutions of Eq. (5.2.3) are uniformly stable iff there exists a constant c > 0
such that
[[φ(t)φ
−1
(t
1
)[[ ≤ c ∀ t
0
≤ t
1
≤ t < ∞ (5.2.4)
Proof : Let ¯ x(t) = ¯ x(t, t
0
, ¯ x
0
) be a solution of Eq. (5.2.3). Then for any
t
1
≥ t
0
we have ¯ x(t) = φ(t)φ
−1
(t
1
)¯ x(t
1
). Let ¯ x
(1)
(t) = φ(t)φ
−1
(t
1
)¯ x
(1)
(t
1
) be
any other solution. As Eq. (5.2.4) is satisfied, we have
[[¯ x
(1)
(t) − ¯ x(t)[[ ≤ [[φ(t)φ
−1
(t
1
)[[[[¯ x
(1)
(t
1
) − ¯ x(t
1
)[[
≤ c[[¯ x
(1)
(t
1
) − ¯ x(t
1
)[[ ∀ t
0
≤ t
1
≤ t < ∞
Let > 0 be given.
[[¯ x
(1)
(t
1
) − ¯ x(t
1
)[[ ≤

c
= δ() > 0 imply that [[¯ x
(1)
(t) − ¯ x(t)[[ <
5.2. STABILITY ANALYSIS - FORMAL APPROACH 151
and hence ¯ x(t) is uniformly stable.
Conversley, let Eq. (5.2.3) uniformly stable and hence zero solution is uniformly
stable. Hence given > 0, ∃ δ() such that t
1
≥ t
0
and
[[¯ x
(1)
(t)[[ < δ ⇒[[¯ x
(1)
(t)[[ < ∀ t ≥ t
1
.
Thus we have [[φ(t)φ
−1
(t
1
)¯ x
(1)
(t
1
)[[ < ∀ t ≥ t
1
.
As in Theorem 5.2.1, we choose ¯ x
(1)
(t
1
) =
δ
2
¯ e
(j)
. Then
[[φ(t)φ
−1
(t
1
)¯ x
(1)
(t
1
)[[ = [[φ
(j)
[[
δ
2
< .
Here φ
(j)
(t) is the j
th
column of φ(t)φ
−1
(t
1
). Hence, it follows that
[[φ(t)φ
−1
(t
1
)[[ = max
1≤j≤n
[[φ
(j)
(t)[[ <
2
δ
, t ≥ t
1
This proves Eq. (5.2.4) and the theorem.
However, for nonlinear systems, boundedness and stability are distinct concepts.
Example 5.2.1 Consider,
dx
dt
= t
This has a solution of the form x(t) = x(t
0
) +
t
2
2
. This is a stable solution but
not bounded. Whereas for
dx
dt
= 0, all solutions are of the form x(t) = x(t
0
)
which are uniformly stable but not asymptotically stable.
Example 5.2.2 Every solution of the system
dx
dt
= p(t)x(t)
is of the form
x(t) = x(t
0
) exp

t
t0
p(s)ds

Hence x(t) ≡ 0 is asymptotically stable iff

t
t0
p(s)ds →−∞ as t →∞.
For the nonlinear initial value problem given by Eq. (5.2.1), it is difficult to
give the stability criterion. Instead, we examine a special form of Eq. (5.2.1),
which is of the type
dx
dt
= Ax(t) +g(t, x(t)) (5.2.5)
152 CHAPTER 5. STABILITY ANALYSIS
g(t, x) satisfies the following condition
|g(t, x)|
|x|
→0 as |x| →0 (5.2.6)
In view of Eq. (5.2.6), it follows that g(t, 0) = 0 and hence 0 is a trivial solution
of Eq. (5.2.5) with 0 initial condition.
We have the following stability theorem for such a kind of system.
Theorem 5.2.3 Suppose the real part of eigenvalues of the matrix A are neg-
ative and the function g(t, x) satisfies Eq. (5.2.6). Then the trivial solution of
Eq. (5.2.5) is asympotically stable.
Proof : The solvability of Eq. (5.2.5) with x(t
0
) = x
0
is equivalent to the
solvability of the following equation
x(t) = exp(A(t − t
0
)) x
0
+

t
t0
exp(A(t −s)) g(s, x(s))ds (5.2.7)
Since real(λ) < 0, ( λ is any eigenvalue of A), it follows that there exists c, δ > 0
such that [[ exp(At)[[ ≤ c exp (−δt) ∀ t ≥ 0. In view of this fact, Eq. (5.2.7)
gives
|x(t)| ≤ c exp (−δ(t −t
0
)) |x
0
| + c

t
t0
exp (−δ(t −s)) |g(s, x(s))| ds(5.2.8)
Also the condition given by Eq. (5.2.6) on g(t, x) implies that for given m >
0, ∃ d > 0 such that
[[g(t, v)[[ ≤ m[[v[[ whenever [[v[[ ≤ d (5.2.9)
We assume that the initial condition x(t
0
) = x
0
is such that [[¯ x
0
[[ < d. Then in
view of continuity of the function x(t), there exists t
1
> t
0
such that [[x(t)[[ < d
for all t ∈ [t
0
, t
1
] = I.
Hence t ∈ I, implies that [[x(t)[[ < d and Eq. (5.2.9) gives that [[g(t, x(t))[[ ≤
m[[x(t))[[ for all t ∈ I. Using this inequality in Eq. (5.2.8), we get
|x(t)| exp (δt) ≤ c exp (δt
0
)|x
0
| +c m

t
t0
exp(δs) [[x(s)[[ds, t ∈ I
By Gronwall’s lemma, we get that
[[x(t)[[ exp (δt) ≤ c [[x
0
[[ exp (δt
0
) exp(c m(t −t
0
)), t ∈ I
That is
[[x(t)[[ ≤ c [[x
0
[[ exp ((c m−δ)(t −t
0
)), t ∈ I (5.2.10)
Since x
0
and m are at our disposal, we choose m and ¯ x
0
in such a way that
c m < δ and |x
0
| ≤ min

d,

c

. Then Eq. (5.2.10) implies that [[x(t)[[ ≤ d
5.2. STABILITY ANALYSIS - FORMAL APPROACH 153
for t ∈ I and this bound d does not depend on t and hence by continuation
theorem (Theorem 3.2.4) the solution x(t) can be extended for t ∈ [t
0
, ∞).
Further Eq. (5.2.10) implies that given > 0, ∃ η = min

d,

c

such that
|x
0
| < η ⇒ |x(t, t
0
, x
0
)| < ∀ t ∈ [t
0
, ∞)
and also [[x(t)[[ →0 as t →∞.
This gives us the asymptotic stability of the zero solution of Eq. (5.2.5).
Example 5.2.3 The motion of a simple pendulum with damping is given by
¨
θ +
k
m
˙
θ +
g
l
sin θ = 0
Approximating sin θ by θ −
θ
3
3!
+. . ., we get
sin θ = θ +f(θ),
f(θ)
θ
→0 as θ →0
This gives
¨
θ +
k
m
˙
θ +
g
l
θ +
g
l
f(θ) = 0 (5.2.11)
where
f(θ)
θ
→0 as θ →0.
Eq. (5.2.11) is equivalent to
dx
dt
= Ax +F(x)
where
x = (θ,
˙
θ), A =
¸
0 1

g
l

k
m
¸
and F(x) = F(θ,
˙
θ) =

0, −
g
l
f(θ)

A has eigenvalues −
k
2m
±

k
2
4m
2

g
l

1
2
both of which have negative real parts
if k, m, g and l are positive and
|Fx| =
g
l

θ
3
3!
−. . .

≤ M[θ[
3
→0 as |x| →0
Thus Theorem 5.2.3 is applicable and hence the main pendulum problem given
by Eq. (5.2.11) is asymptotically stable.
154 CHAPTER 5. STABILITY ANALYSIS
In a similar way we have the following theorem for the unstability of zero solution
of Eq. (5.2.5).
Theorem 5.2.4 Assume that A possesses at least one eigenvalue with positive
real part and g(t, x) satisfies Eq. (5.2.6). Then the trivial solution of Eq. (5.2.5)
is unstable.
For details analysis concerning this section refer Agarwal and Gupta[1].
5.3 Phase Portrait Analysis
For the phase portrait, we shall concentrate on the autonomous system
dx
dt
= f(x(t)) (5.3.1)
Definition 5.3.1 The set of solution curves in the phase space '
n
of the system
given by Eq. (5.3.1) is called the phase portrait of the system.
We shall be interested in the stability of the system around equilibrium points
- those points where f(x) = 0. Also, phase space will be assumed to be two
dimensional so that its geometry can be studied clearly.
If ¯ x
0
is an isolated equilibrium point of f(¯ x) and if f is differentiable at ¯ x
0
, then
we have (by Eq. (2.3.4))
f(¯ x) = f( ¯ x
0
) +f
t
(x
0
)(¯ x − ¯ x
0
) +g(¯ x)
where
|g(¯ x)|
|¯ x − ¯ x
0
|
→0 as |¯ x − ¯ x
0
| →0.
As f( ¯ x
0
) = 0, we have f(¯ x) = f
t
(x
0
)(¯ x − ¯ x
0
) + g(¯ x). Translating the origin to
¯ x
0
, we get that Eq. (5.3.1) is equivalent to the system
d¯ x
dt
= A¯ x +g(¯ x) (5.3.2)
with
|g(¯ x)|
|¯ x|
→0 as |¯ x| →0 and A = f
t
(x
0
).
We shall now study Eq. (5.3.2) through the linear system
d¯ x
dt
= A¯ x (5.3.3)
Theorem 5.2.3 and 5.2.4 give the relationship between the stability of the triv-
ial solution of the nonlinear system of the form Eq. (5.3.2) with that of the
linear system given by Eq. (5.3.3). We restate this relationship in terms of the
following theorem.
Theorem 5.3.1 (a) If the equilibrium point zero of the linear system given
by Eq. (5.3.3) is asymptotically stable, so is the equilibrium point zero of
the nonlinear system given by Eq. (5.3.2).
5.3. PHASE PORTRAIT ANALYSIS 155
(b) If the equilibrium point zero of the Eq. (5.3.3) is unstable, so is the un-
stability of the equilibrium point zero of Eq. (5.3.2).
(c) If the equilibrium point zero Eq. (5.3.3) is stable then the zero solution of
Eq. (5.3.2) can be stable, asymptotically stable or unstable.
Example 5.3.1 We investigate the phase portrait of the linear system given by
Eq. (5.3.3) around the origin with A =
¸
−1 0
0 2

. The solution of the linear
system is given by
x(t) =
¸
exp(−t) 0
0 exp (2t)

c, c = x
0
= x(t
0
).
The solution curves of the above system are of the form
x
2
=
k
x
1
2
, k = c
1
2
c
2
(c = (c
1
, c
2
)).
The solution x(t) defines a motion along these curves, that is each point c ∈ '
2
moves to the point x(t) ∈ '
2
after a time t. Hence the phase portrait in '
2
is
as under.
-3 -2 -1 1 2 3
x1
-2
-4
2
4
x2
Figure 5.3.1: Phase portrait of Example 5.3.1
The function f(x) = Ax defines a mapping on '
2
. Recall that this is called
a vector field. Since
dx
dt
= f(x) = Ax, at each point x ∈ '
2
, the solution curve
x(t) is tangent to the vectors in the vector field f(x) = Ax. So the vector field
is as given below.
156 CHAPTER 5. STABILITY ANALYSIS
Figure 5.3.2: Vector field of Example 5.3.1
Example 5.3.2 Consider now the coupled system
dx
dt
=

−1 −3
0 2
¸
¸
x = Ax
The eigenvalues of the matrix A are −1 and 2 and the pair of corresponding
eigenvectors is (1, 0), (−1, 1).
P =

1 −1
0 1
¸
¸
, P
−1
=

1 1
0 1
¸
¸
, P
−1
AP =

−1 0
0 2
¸
¸
= D
So, under the transformation y = P
−1
¯ x, we obtain the decoupled system
dy
dt
= Dy
The phase portraits in y
1
−y
2
plane and x
1
−x
2
plane are as under.
5.3. PHASE PORTRAIT ANALYSIS 157
-3 -2 -1 1 2 3
y1
-2
-4
2
4
y2
Figure 5.3.3: Phase portrait in y
1
−y
2
plane
-4 -2 2 4
x1
-3
-2
-1
1
2
3
x2
Figure 5.3.4: Phase portrait in x
1
−x
2
plane
In general, the phase portrait of a linear system given by Eq. (5.3.3) is best
studied by an equivalent linear system
dx
dt
= Bx (5.3.4)
where the matrix B = P
−1
AP has one of the following form ( by appealing to
Theorem 5.1.1)
B =
¸
λ 0
0 µ

, B =
¸
λ 1
0 λ

, or B =
¸
a −b
b a

158 CHAPTER 5. STABILITY ANALYSIS
The solution of the initial value problem given by Eq. (5.3.4) with x(0) = x
0
is
x(t) =
¸
exp (λt) 0
0 exp (µt)

x
0
, x(t) = exp (λt)
¸
1 t
0 1

x
0
or
x(t) = exp(at)
¸
cos bt −sinbt
sin bt cos bt

x
0
Different cases will have different phase portraits.
Case 1: B =
¸
λ 0
0 µ

, λ < 0 < µ
The phase portrait is of the form given in Figure 5.3.3.
In this case the system given by Eq. (5.3.4) is said to have a saddle point at
0. If λ > 0 > µ, then the arrows in the vector field are reversed as we see in the
following Figure.
Figure 5.3.5: Vector field with arrows reversed
Case 2: B =
¸
λ 0
0 µ

, λ ≤ µ < 0 or B =
¸
λ 1
0 λ

, λ < 0
In this case the origin is referred to as a stable node ( unstable if λ ≥ µ > 0
). The phase portraits are as under:
5.3. PHASE PORTRAIT ANALYSIS 159
-3 -2 -1 1 2 3
x1
-6
-4
-2
2
4
6
x2
Figure 5.3.6: Phase portrait with λ = µ < 0 (stable node)
-3 -2 -1 1 2 3
x1
-1.5
-1
-0.5
0.5
1
1.5
x2
Figure 5.3.7: Phase portrait with λ < µ < 0 (stable node)
When B =
¸
λ 1
0 λ

, λ < 0, the phase portrait is as under.
160 CHAPTER 5. STABILITY ANALYSIS
-4 -2 2 4
x1
-2
-1
1
2
x2
Figure 5.3.8: Origin as stable node
Case 3: B =
¸
a −b
b a

If a < 0, the origin is a stable focus with trajectories as spiral converging
to zero. If a > 0, we have unstable focus.
-10 -5 5
x1
-10
-5
5
10
x2
Figure 5.3.9: Stable focus
5.3. PHASE PORTRAIT ANALYSIS 161
-10 -7.5 -5 -2.5 2.5 5 7.5
x1
-10
-5
5
10
x2
Figure 5.3.10: Unstable focus
Case 4: B =
¸
0 −b
b 0

In this case the system given by Eq. (5.3.4) is said to have a centre at the
origin. The phase portrait is as under.
-1 -0.5 0.5 1
x1
-1
-0.5
0.5
1
x2
Figure 5.3.11: Origin as centre
Example 5.3.3
d¯ x
dt
= A¯ x, A =
¸
0 −4
1 0

162 CHAPTER 5. STABILITY ANALYSIS
A has eigenvalues λ = ±2i.
P =
¸
2 0
0 1

, P
−1
=
¸
1
2
0
0 1

, B = P
−1
AP =
¸
0 −2
2 0

¯ x(t) = P
¸
cos 2t −sin2t
sin 2t cos 2t

P
−1
¯ c
=
¸
cos 2t −2 sin2t
1
2
sin 2t cos 2t

¯ c
¯ c = ¯ x(0).
It can be seen that x
1
2
+ 4x
2
2
= c
1
2
+ 4c
2
2
.
-4 -2 2 4
x1
-2
-1
1
2
x2
Figure 5.3.12: Phase portrait with origin as centre
The above discussion can be described in the following theorem, giving the
saddle point, node, focus or centre at the origin of the linear system. We assume
that det A = 0, so that
¯
0 is the only equilibrium point.
Theorem 5.3.2 Let δ = detA and τ = traceA and consider the linear system
given by Eq. (5.3.3).
(a) If δ < 0, then Eq. (5.3.3) has a saddle point at the origin.
(b) If δ > 0 and τ
2
− 4δ ≥ 0, then Eq. (5.3.3) has a node at the origin. it is
stable if τ < 0 and unstable if τ > 0.
(c) If δ > 0 and τ
2
− 4δ < 0 and τ = 0, then Eq. (5.3.3) has a focus at the
origin. It is stable if τ < 0 and unstable if τ > 0.
(d) If δ > 0 and τ = 0 then Eq. (5.3.3) has a centre at the origin.
5.4. LYAPUNOV STABILITY 163
Example 5.3.4 Consider the following linearized Predator-prey model (linearized
around the equilibrium point (
k
λ
,
a
c
)), given by Eq. (1.2.5)
d¯ x
dt
= A¯ x(t), A =

0 −
kc
λ

c
0
¸
¸
¸
¸
¸
where ¯ x = (F, S), with F representing fish population and S shark population.
Using the notation of the above theorem (Theorem 5.3.2), we get δ = det A =
ak > 0 and τ = trace A = 0. As δ > 0 and τ = 0, it implies that the equilibrium
point (
k
λ
,
a
c
) is a centre.
Example 5.3.5 Consider the problem of mechanical oscillations discussed in
Example 1.3.1
d
2
x
dt
2
+k
dx
dt
+w
2
x = 0
k is the resistance and w
2
=
λ
ma
.
This is equivalent to the following system
d¯ x
dt
= A¯ x
where ¯ x =

x,
dx
dt

and A =

0 1
−w
2
−k
¸
¸
.
δ = detA = w
2
> 0 and τ = traceA = −k < 0.
(i) If τ
2
−4δ = k
2
− 4w
2
< 0, we have origin as a stable focus.
(ii) If k
2
−4w
2
≥ 0, we have origin as a stable node.
For more on phase portrait analysis refer Perko [9].
5.4 Lyapunov Stability
The principle idea in the Lyapunov method is the following physical reasoning.
If the rate of change
dE(¯ x)
dt
of the energy of a physical system is negative
for every possible state ¯ x except for a single equilibrium state ¯ x
e
, then the
energy will continually decrease until it finally assumes its minimum. From
mathematical perspective, we look for a scalar valued Lyapunov function V (¯ x)
of the state which is (a) positive (b) with
˙
V (¯ x) < 0(¯ x = ¯ x
e
) and (c) V (¯ x) =
˙
V (¯ x) = 0 for ¯ x = ¯ x
e
. One of the attractions of the Lyapunov method is its
appeal to geometric intution as we see in the following two examples.
164 CHAPTER 5. STABILITY ANALYSIS
Example 5.4.1 Let us recall Example 1.3.1, case 1 of a harmonic oscillator
given by a second order differential equation.
d
2
x
dt
2

2
x = 0, ω
2
=
λ
ma
In terms of a first order system we have
d¯ x
dt
=
¸
0 1
−ω
2
0

¯ x, ¯ x = (x
1
, x
2
)
The trajectories of the above system in the phase-plane are ellipses about the
origin. The energy of the system is given by
E(¯ x) = E(x
1
, x
2
) = ω
2
x
2
1
+x
2
2
= V (¯ x)
The derivative
˙
V and V along the trajectory of the above system is given by
˙
V (¯ x) =
dV ¯ x(t)
dt
= (¯ x(t), V (¯ x))
= ˙ x
1
∂V
∂x
1
+ ˙ x
2
∂V
∂x
2
= 2ω
2
x
1
x
2
+ 2x
2
˙ x
2
= 0
So energy remains constant and hence the system is conservative.
Example 5.4.2 Consider a slight modification of the earlier example
˙ x
1
= x
2
−ax
1
˙ x
2
= −x
1
−ax
2
with a > 0.
We define V (¯ x) = (x
2
1
+x
2
2
). Then
˙
V (¯ x(t)) = ( ˙ x(t), V (¯ x))
= −2a(x
2
1
+x
2
2
)
2
= −2aV (¯ x)
< 0 if ¯ x = 0
= 0 if ¯ x = 0
This implies that V (¯ x) is constantly decreasing along any solution of the above
system. The trajectory of the above system is given by
R(t) = R
0
e
−(t−t0)
, R(t) = x
2
1
(t) +x
2
2
(t)
5.4. LYAPUNOV STABILITY 165
x
x
2
1
V(x)=constant
Figure 5.4.1: Trajectory along V (¯ x) = constant
Hence the trajectory of the above system cross the boundary of every region
V (¯ x) = constant from the outside towards inside. This leads to the following
intersting geometric interpretation of the Lyapunov function. Let V (¯ x) be a
measure of the distance of the state ¯ x from the orgin in the state space resulting
in to V (¯ x) > 0 for ¯ x = 0 and V (0) = 0. Suppose this distance is continuously
decreasing as t →∞, that is,
˙
V (¯ x(t)) < 0. Then ¯ x(t) →0.
We now investigate the role of Lyapunov function to ensure the stability of the
equilibrium state of the non-autonomous nonlinear system.
d¯ x
dt
= f(t, ¯ x(t)), (t, ¯ x) ∈ I '
n
(5.4.1)
where f : I '
n
→'
n
with origin as an equilibrium point (f(t, 0) = 0).
Assumptions 1: There exists a scalar valued function V (t, ¯ x) : I '
n
→'
n
,
called Lyapunov function, such that it is continuous and has continuous partial
derivatives w.r.t ¯ x and t with V (t, 0) = 0 and
(i) V (t, ¯ x) is positive definite, that is, there exist a continuous, non-decreasing
function α such that α(0) = 0 and for all t and all ¯ x = 0, we have
0 < α([[¯ x[[) ≤ V (t, ¯ x) (5.4.2)
(ii) there exists a continuous scalar valued function r such that r(0) = 0 and
the derivative
˙
V (¯ x) of V (¯ x) along the motion ¯ x(t) satisfies the following
˙
V (t, ¯ x(t)) ≤ −r([[¯ x[[) < 0 ∀ (t, ¯ x) ∈ I '
n
(5.4.3)
166 CHAPTER 5. STABILITY ANALYSIS
(ii) (a)
˙
V (t, ¯ x(t)) ≤ 0 ∀ (t, ¯ x) ∈ I '
n
(5.4.4)
(iii) there exists a continuous, non-decreasing function β such that β(0) = 0
and
V (t, ¯ x(t)) ≤ β([[¯ x[[) ∀ t ∈ I (5.4.5)
Theorem 5.4.1 Let V (t, ¯ x) be a Lyapunov function of the system (5.4.1).
Then the equilibrium state
¯
0 is
(a) stable if Assumptions (i) and (ii) hold
(b) uniformly stable if Assumptions (i),(ii)(a) and (iii) hold
(c) asymptotically stable if (i),(ii) and (iii) hold.
Proof :
(a) Let β

(t, [[¯ x[[) be the maximum of V (t, ¯ y) for [[¯ y[[ ≤ [[¯ x[[. Let > 0 be
arbitary. As β

(t
0
, [[¯ x[[) is a continuous function of ¯ x, we can choose
δ(t
0
, ) such that
β

(t
0
, [[ ¯ x
0
[[) < α() if [[ ¯ x
0
[[ < δ (5.4.6)
Using (5.4.2), (5.4.4) and (5.4.6), we get
α([[¯ x[[) ≤ V (t, ¯ x(t, t
0
, ¯ x
0
)) ≤ V (t
0
, ¯ x
0
) ≤ β

(t
0
, ¯ x
0
) < α()
forall t ≥ t
0
and [[ ¯ x
0
[[ < δ.
Since α is non-decreasing and positive, the above inequality implies that
[[¯ x(t, t
0
, ¯ x
0
)[[ < whenever [[ ¯ x
0
[[ < δ for t ≥ t
0
.
This proves the stability of the equilibrium state
¯
0.
(b) Let (i), (ii)(a) and (iii) hold and assume that > 0 is arbitary. We get
δ() (independent of t
0
) such that β(δ) < α(). Let [[ ¯ x
0
[[ < δ. Then
proceeding as in part(a), we have
α() > β(δ) ≥ V (t
0
, ¯ x
0
) ≥ V (t, ¯ x(t, t
0
, ¯ x
0
)) ≥ α([[¯ x(t, t
0
, ¯ x
0
)[[)
Here, for the second inequality, we have used Assumption 1(iii).
The above inequality gives
[[¯ x(t, t
0
, ¯ x
0
)[[ < for t ≥ t
0
and [[ ¯ x
0
[[ < δ
where δ is independent of t
0
.
This implies the uniform stability of the zero solution the equilibriun point.
5.4. LYAPUNOV STABILITY 167
(c) For asymptotic stability of the zero solution, we only need to show that
∃ r > 0 such that
[[ ¯ x
0
[[ ≤ r ⇒[[¯ x(t, t
0
, ¯ x
0
)[[ →0 uniformly with t →∞.
Equivallently, given µ > 0, ∃ T(µ, r) such that
[[ ¯ x
0
[[ ≤ r ⇒[[¯ x(t, t
0
, ¯ x
0
)[[ < µ for t ≥ t
0
+T
Let functions α(µ) and β(µ) be as defined in Eq. (5.4.3) and Eq. (5.4.5).
Take any positive constant c
1
and find r > 0 such that β(r) < α(c
1
). Let
[[ ¯ x
0
[[ ≤ r. Then by part (b), we have
[[ ¯ x
0
[[ ≤ r ⇒[[¯ x(t, t
0
, ¯ x
0
)[[ < c
1
∀ t ≥ t
0
Let 0 < µ ≤ [[ ¯ x
0
[[. As β(µ) is continuous at 0, we find λ(µ) > 0 such that
β(λ) < α(µ). Also r([[¯ x[[) is a continuous function and hence it assumes
its minimum on a compact set λ ≤ [[¯ x[[ ≤ c
1
(r) and let this minimum be
c
2
(µ, r).
Define T(µ, r) =
β(r)
c
2
(µ, r)
> 0.
We first claim that [[¯ x(t, t
0
, ¯ x
0
)[[ = λ for some t = t
2
in [t
0
, t
1
], t
1
=
t
0
+T. If not, then [[¯ x(t, t
0
, ¯ x
0
)[[ > λ on [t
0
, t
1
]. Hence Eq. (5.4.2) gives
0 < λ ≤ V (t
1
, ¯ x(t, t
0
, ¯ x
0
))
= V (t
0
, x
0
) +

t1
t0
˙
V (¯ x(τ, t
0
, ¯ x
0
))dτ
≤ V (t
0
, x
0
) −

t1
t0
r([[¯ x(τ)[[)dτ
≤ V (t
0
, x
0
) −(t
1
, t
0
)c
2
≤ β(r) −Tc
2
= 0
which is a contradiction.
Hence, we have
[[ ¯ x
2
[[ = [[x(t
2
, t
0
, ¯ x
0
)[[ = λ
168 CHAPTER 5. STABILITY ANALYSIS
Therefore
α([[¯ x(t, t
0
, ¯ x
0
)[[) ≤ V (t, ¯ x(t, t
2
, x
2
))
≤ V (t
2
, x
2
)
≤ β(λ)
< α(µ) ∀ t ≥ t
0
This implies that
[[¯ x(t, t
0
, ¯ x
0
)[[ < µ ∀ t ≥ t
0
+T(µ, r) ≥ 0
The real weakness of Lyapunov method is that no general technique is known for
the construction of the Lyapunov function V (t, ¯ x). However, we shall describe
the method of construction of Lyapunov function for asymptotically stable linear
autonomuous system
d¯ x
dt
= A¯ x
We assume that A has k distinct real eigenvalues λ
1
, λ
2
, , λ
k
and the re-
maining are complex eigenvalues λ
j
= a
j
± ib
j
, k + 1 ≤ j ≤ n. Then by
Theorem 5.1.1, there exists a basis ¦ ¯ v
1
, ¯ v
2
, , ¯ v
k
, ¯ v
k+1
, ¯ v
k+2
, , ¯ v
n
, ¯ u
n
¦ of
'
2n−k
where ¯ v
j
(1 ≤ j ≤ k) are eigenvectors corresponding real eigenvalues
and u
j
±iv
j
(k + 1 ≤ j ≤ n) are the complex eigenvectors.
If P = [¯ v
1
, ¯ v
2
, , ¯ v
k
, ¯ v
k+1
, ¯ v
k+2
, , ¯ v
n
, ¯ u
n
], then P is nonsingular and P
−1
AP
= D, where D is block diagonal matrix of the form
D =

λ
1
λ
2
.
.
.
λ
k
B
k+1
.
.
.
B
n
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
with B
i
=
¸
a
i
−b
i
b
i
a
i

as a Jordan block corresponding to the eigenvalue a
i
±
ib
i
. Using the transformation ¯ x = P ¯ y, we get the reduced system
d¯ y
dt
= P
−1
AP ¯ y = D¯ y
5.4. LYAPUNOV STABILITY 169
For asymptotic stability, we require all diagonal elements of D to be negative.
We seek Lyapunov function V of the forrm
V (¯ y) = (B¯ y, ¯ y) (5.4.7)
where B is a suitable positive definite matrix. Differentiating Eq. (5.4.7), we
get
dV
dt
=

B
d¯ y
dt
, ¯ y(t)

+

B¯ y(t),
dy
dt

= (BD¯ y(t), ¯ y(t)) + (B¯ y(t), D¯ y(t))
= (BD¯ y(t), ¯ y(t)) +

D
¯
B¯ y(t), ¯ y(t)

=

BD +D
¯
B

¯ y(t), ¯ y(t)

(5.4.8)
In order to ensure that
dV
dt
is negative definite, we shall require that
dV
dt
= −(Q¯ y(t), ¯ y(t)) (5.4.9)
where Q is a given positive definite matrix. Eq. (5.4.9) will hold, if we assume
that
BD +D
¯
B = −Q (5.4.10)
Thus, to get a Lyapunov function which is positive definite with
dV
dt
as negative
definite, we need to solve the matrix equation given by Eq. (5.4.10) for B with
a given positive definite matrix Q (say Q = I).
Example 5.4.3 Construct a Lyapunov function for the system
d¯ x
dt
=

−1 −1 0 0
0 −1 0 0
0 0 −3 −2
0 0 1 −1
¸
¸
¸
¸
¯ x
The eigenvalues of this system are −1 ±i, −2 ±2i.
One gets
P =

1 0 0 0
0 1 0 0
0 0 −1 −1
0 0 1 0
¸
¸
¸
¸
with P
−1
=

1 0 0 0
0 1 0 0
0 0 0 1
0 0 −1 −1
¸
¸
¸
¸
Then
P
−1
AP = D =

−1 1 0 0
−1 −1 0 0
0 0 −2 −2
0 0 −2 −2
¸
¸
¸
¸
170 CHAPTER 5. STABILITY ANALYSIS
We make the transformation ¯ x = P ¯ y to get the reduced system
d¯ y
dt
= D¯ y
Take Q = I and solve the matrix equation
BD +D
T
B = −I
to get
B =

¸
¸
¸
¸
¸
¸
¸
¸
1
2
0 0 0
0
1
2
0 0
0 0
1
4
0
0 0 0
1
4
¸

This gives the Lyapunov function V (¯ y) as
V (¯ y) =
1
2
y
2
1
+
1
2
y
2
2
+
1
4
y
2
3
+
1
4
y
2
4
where
¯ y = P
−1
¯ x =

¸
¸
¸
1 0 0 0
0 1 0 0
0 0 0 1
0 0 −1 −1
¸

We have
y
1
= x
1
, y
2
= x
2
, y
3
= x
4
, y
4
= −x
3
−x
4
Hence
V (¯ x) =
x
2
1
2
+
x
2
2
4
+
(x
3
+x
4
)
2
4
+
x
2
4
4
Example 5.4.4 Consider the nonlinear system of the form
d¯ x
dt
= A¯ x +g(¯ x) (5.4.11)
A has eigenvalues with negative real part and g(¯ x) satisfies the growth condition
of the form
[[g(¯ x)[[ ≤ k[[¯ x[[ ∀ ¯ x
with small k.
We construct a Lyapunov function for Eq. (5.4.11) in the following way. V (¯ x)
= (B¯ x, ¯ x) , B a positive definite matrix. Solve B to satisfy the matrix equation
A
¯
B +B
¯
A = −I
5.5. EXERCISES 171
We have
dV
dt
=
˙
¯ x
¯
B¯ x +
˙
¯ x
¯
B
˙
¯ x
=

¯ x
¯
A
¯
+ [g(¯ x)]
¯

B¯ x + ¯ x
¯
B(A¯ x +g(¯ x))
= ¯ x
¯
(A
¯
B +BA)¯ x +

[g(¯ x)]
¯
B¯ x + ¯ x
¯
B[g(¯ x)]

= −¯ x
¯
¯ x +

[g(x)]
¯
B¯ x + ¯ x
¯
B[g(¯ x)]

Estimating the second term within the square bracket, we get
[[[g(¯ x)]
¯
B¯ x + ¯ x
¯
B[g(¯ x)][[ ≤ 2[[g(¯ x)[[[[B[[[[¯ x[[
≤ 2k[[B[[[[¯ x[[
2
Hence
dV
dt
≤ −[[¯ x[[
2
[1−2k[[B[[] = −α([[x[[) where α(r) = αr
2
, α−2k[[B[[ > 0.
Thus the Lyapunov function V (¯ x) is such that V is positive definite with
dV
dt
negative definite and hence the system Eq. (5.4.11) is asymptotically stable.
For more on stability results refer Aggarwal and Vidyasagar [2], Amann [3],
Borckett [4], Lakshmikantham et al[6,7] and Mattheij and Molenaar [8].
5.5 Exercises
1. Show that all the solution of the linear system
d¯ x
dt
= A(t)¯ x
where A(t) =

−t 0 0
0 −t
2
0
0 0 −t
2
¸
¸
¸
¸
¸
¸
,

e
−t
−1 −cos t
1 −e
−2t
t
2
cos t −t
2
e
−3t
¸
¸
¸
¸
¸
¸
tend to zero as t →∞.
2. By using the Cayley-Hamilton theorem (Theorem 5.1.3), show that for
A ∈ '
n·n
, there exist functions α
k
(t), 1 ≤ k ≤ n − 1 such that e
At
=
n−1
¸
k=0
α
k
(t)A
k
.
3. Show that all solutions of the following differential equations are bounded
in [ t
0
, ∞)
172 CHAPTER 5. STABILITY ANALYSIS
(i)
d
2
x
dt
2
+c
dx
dt
+
¸
1 +
1
1 +t
2

x = 0, c > 0.
(ii)
d
2
x
dt
2
+c
dx
dt
+
¸
1 +
1
1 +t
4

x = 0, c > 0.
4. Test the stability, asymptotic stability or unstability for the trivial solution
of the following linear systems.
(i)
d¯ x
dt
=

0 1
−1 0
¸
¸
¯ x
(ii)
d¯ x
dt
=

0 1 0
0 0 1
−1 −6 −5
¸
¸
¸
¸
¸
¸
¯ x
(iii)
d¯ x
dt
=

1 2 0
0 1 1
1 3 1
¸
¸
¸
¸
¸
¸
¯ x
(iv)
d¯ x
dt
=

1 −1 −1
1 1 −3
1 −5 −3
¸
¸
¸
¸
¸
¸
¯ x
5. Test the stability, asymptotic stability or unstability of trivial solution of
each of the following system
(i)
dx
1
dt
= −2x
1
+x
2
+ 3x
3
+ 8x
2
1
+x
3
2
dx
2
dt
= −6x
2
−5x
3
+ 7x
4
3
dx
3
dt
= −2x
3
+x
4
1
+x
2
2
+x
3
3
5.5. EXERCISES 173
(ii)
dx
1
dt
= 2x
1
+x
2
−x
2
1
−x
2
2
dx
2
dt
= x
1
+ 3x
2
−x
3
1
sin x
3
dx
3
dt
= x
2
+ 2x
3
+x
2
1
+x
2
2
6. Discuss the stability or unstability of the origin of the following linear
sytem and sketch the phase portraits.
(i)
dx
1
dt
= −2x
1
+x
2
dx
2
dt
= −5x
1
−6x
2
(ii)
dx
1
dt
= 4x
1
+x
2
dx
2
dt
= 3x
1
+ 6x
2
(iii)
dx
1
dt
= x
2
dx
2
dt
= 2x
1
−x
2
(iv)
dx
1
dt
= x
1
+x
2
dx
2
dt
= 3x
1
−x
2
7. Consider the nonlinear system
dx
dt
= a(t)x
3
Find its solution.
(i) Show that origin is uniformly asymptotically stable if a(t) < 0.
(ii) If a(t) = −
1
t + 1
, show that origin is asymptotically stable but not
uniformly stable.
(iii) If a(t) = −
1
(t + 1)
2
, show that the solution of the corresponding IVP
approaches a constant value as t →∞ but origin is uniformly stable.
8. Consider the nonlinear system
dx
dt
= −x
p
, p > 1 and p is integer
Show that
174 CHAPTER 5. STABILITY ANALYSIS
(i) If p is odd, then origin is uniformly stable and
(ii) If p is even, then origin is unstable.
9. Find Lyapunov functions around the origin for the following system.
(i)
dx
dt
= −x
(ii)
dx
dt
= −x(1 −x)
(iii)
d¯ x
dt
=
¸
0 1
−1 −a

¯ x, a ≥ 0
10. Show that the origin is unstable equilibrium point of the system
d
2
x
dt
2

dx
dt

2
sgn( ˙ x) +x = 0
References
[1 ] Ravi Agarwal and Ramesh C Gupta, Essentials of Ordinary Differential
Equations, McGraw Hill Book Company, 1991
[2 ] J K Aggarwal and M. Vidyasagar, (Editors) Nonlinear Systems, Stability
Analysis, Dowden, Hutchinson and Ross, Inc., 1977
[3 ] Herbert Aman, Ordinary Differential Equations: Introduction to Non-
linear Analysis, Walter de Gruyter, 1990
[4 ] Roser W Brockett, Finite Dimensional Linear Systems, John Wiley and
Sons, Inc., 1970
[5 ] F. R. Gantmacher, The Theory of Matrices Vols. I and II, Chelsea
Publishing Company, 1960
[6 ] V. Lakshimantham and S. G. Deo, Method of Variation of Parameters
for Dynamic Sytems, Gorden and Breach Science Pub. , 1998
[7 ] V. Lakshmikantham, S. Leela and A.A Martynyuk, Stability Analysis of
Nonlinear Systems, Marcel Dekker, 1989
[8 ] R M M Mattheij and J Molenaar, Ordinary Differential Equations in
Theory and Practice, John Wiley, 1996
[9 ] Lawrence Perko, Differential Equations and Dynamical Systems, Springer-
Verlag, 1991
Chapter 6
Series Solution
This chapter lays emphasis on the classical theory of series solution for ODE by
using power series expansion around ordinary and singular points.
We mainly focus on Legendre, Hermite and Bessel differential equations. We
also deduce various interesting properties of Legendre and Hermite polynomials
as well as Bessel functions.
6.1 Preliminaries
A power series in powers of (t −t
0
) is an infinite series of the form

¸
k=0
c
k
(t −t
0
)
k
= c
0
+c
1
(t −t
0
) +c
2
(t −t
0
)
2
+ (6.1.1)
where t is a variable and c
0
, c
1
, c
2
, are constants. Recall that a series

¸
k=0
c
k
(t −t
0
)
k
is convergent at a point t if the limit of the partial sums s
n
(t) =
n
¸
k=0
c
k
(t −t
0
)
k
exists. This limit f(t) is denoted as the sum of the series at the
point t. A series which does not converge is said to be a divergent series.
Example 6.1.1 Consider the geometric power series

¸
k=0
t
k
= 1 + t + t
2
+
The partial sums s
n
(t) = 1 +t + +t
n
satisfy the relation
ts
n
(t) = t +t
2
+ +t
n
+t
n+1
and hence s
n
(t) =
1 −t
n+1
1 −t
.
This gives f(t) = lims
n
(t) =
1
1 −t
for [t[ < 1.
175
176 CHAPTER 6. SERIES SOLUTION
Definition 6.1.1 A power series

¸
k=0
c
k
(t −t
0
)
k
is said to converge absolutely
at a point t if the series

¸
k=0
[c
k
(t −t
0
)
k
[ converges.
One can show that if the series converges absolutely, then it also converges.
However, the converse is not true.
We have the following tests for checking the convergence or divergence of a
series of real numbers.
(i) Comparision test
(a) Let a series

¸
k=0
a
k
of real numbers be given and let there exists a
convergent series

¸
k=0
b
k
of nonnegative real numbers such that
[a
k
[ ≤ b
k
, k ≥ 1
Then the orginal series

¸
k=0
a
k
converges.
(b) Let a series

¸
k=0
a
k
of real numbers be given and let there exists a
divergent series

¸
k=0
d
k
of nonnegative real numbers such that
[a
k
[ ≥ d
k
, k ≥ 1
Then the orginal series

¸
k=0
a
k
diverges.
(ii) Ratio test
Let the series

¸
k=0
a
k
( with a
n
= 0) be such that lim
k→∞

a
k+1
a
k

= L.
(a) The series

¸
k=0
a
k
converges absolutely if L < 1.
(b) The series

¸
k=0
a
k
diverges if L > 1.
(c) No conclusion if L = 1.
6.1. PRELIMINARIES 177
(iii) Root test
If the series

¸
k=0
a
k
is such that lim
k→∞

([a
k
[)
1
k

= L.
(a) The series

¸
k=0
a
k
converges absolutely of L < 1.
(b) The series

¸
k=0
a
k
diverges if L > 1.
(c) No conclusion if L = 1.
Theorem 6.1.1 If the power series given by Eq. (6.1.1) converges at a point
t = t
1
, then it converges absolutely for every t for which [t −t
0
[ < [t
1
−t
0
[.
Proof : Since the series given by Eq. (6.1.1) converges for t = t
1
, it follows
that the partial sums s
n
(t
1
) converge and hence s
n
(t
1
) is Cauchy. This implies
that
s
n+1
(t
1
) −s
n
(t
1
) →0 as n →∞
This, in turn, implies that c
k
(t
1
− t
0
)
k
−→ 0 as k → ∞. Hence the elements
c
k
(t
1
−t
0
)
k
of the series given by Eq. (6.1.1) are bounded at t = t
1
. That is,
[c
k
(t
1
−t
0
)
k
[ ≤ M for all k ≥ 1 (6.1.2)
Eq. (6.1.2) implies that
[c
k
(t −t
0
)
k
[ =

c
k
(t
1
−t
0
)
k

t −t
0
t
1
−t
0

k

≤ M

t −t
0
t
1
−t
0

k
(6.1.3)
If [t−t
0
[ < [t
1
−t
0
[, the series

¸
k=0

t −t
0
t
1
−t
0

k
converges and hence by comparision
test the series

¸
k=0
c
k
(t −t
0
)
k
converges absolutely.
Definition 6.1.2 If the series

¸
k=0
c
k
(t −t
0
)
k
converges absolutely for [t −t
0
[ <
r and diverges for [t −t
0
[ > r, then r is called the radius of convergence.
The power series

¸
k=1
t
k
k
converges for [t[ < 1 and diverges for [t[ > 1. At t = 1,
it diverges and t = −1, it converges. Thus, the radius of convergence of this
series is 1.
178 CHAPTER 6. SERIES SOLUTION
The radius of convergence of the power series given by Eq. (6.1.1) may be
determined from the coefficients of the series as follows.
Theorem 6.1.2 (Radius of convergence)
(i) Suppose that the sequence ¦[c
k
[
1
k
¦ converges . Let L denotes its limit.
Then, if L = 0, the radius of convergence r of the power series is
1
L
.
(ii) If L = 0, then r = ∞ and hence the series Eq. (6.1.1) converges for all t.
(iii) If ¦[c
k
[
1
k
¦ does not converge, but it is bounded, then r =
1
l
where l =
sup¦[c
k
[
1
k
¦. If this sequence is not bounded, then r = 0 and hence the
series is convergent only for t = t
0
.
Let

¸
k=0
c
k
(t −t
0
)
k
be a power series with non zero radius of convergence r.
Then the sum of the series is a function f(t) of t and we write
f(t) =

¸
k=0
c
k
(t −t
0
)
k
(6.1.4)
One can easily show the following.
(i) The function f(t) in Eq. (6.1.4) is continuous at t = t
0
.
(ii) The same function f(t) can not be represented by two different power
series with the same centre. That is, if
f(t) =

¸
k=0
c
k
(t −t
0
)
k
=

¸
k=0
d
k
(t −t
0
)
k
in a disk: [t −t
0
[ < r, then c
k
= d
k
for all k ≥ 0.
We can carry out the standard operations on power series with ease - addition
and subtraction, multiplication, term by term differentiation and integration.
(i) Addition and Subtraction
Two power series
¸

k=0
c
k
(t −t
0
)
k
and
¸

k=0
d
k
(t −t
0
)
k
can be added and
subtracted in the common radius of convergence.
If f(t) =

¸
k=0
c
k
(t −t
0
)
k
in [t −t
0
[ < r
1
and
g(t) =

¸
k=0
d
k
(t −t
0
)
k
in [t −t
0
[ < r
2
Then f(t) ±g(t) =

¸
k=0
(c
k
±d
k
)(t −t
0
)
k
in [t −t
0
[ < r = min(r
1
, r
2
).
6.1. PRELIMINARIES 179
(ii) Multiplication
If f(t) =

¸
k=0
c
k
(t −t
0
)
k
in [t −t
0
[ < r
1
and
g(t) =

¸
k=0
d
k
(t −t
0
)
k
in [t −t
0
[ < r
2
Then h(t) = f(t)g(t) defined within the radius of convergence of each
series and
h(t) =

¸
k=0
c
k
(t −t
0
)
k
where c
k
=
k
¸
m=0
c
m
d
k−m
. That is
h(t) = c
0
d
0
+ (c
0
d
1
+c
1
d
0
) t + (c
0
d
2
+c
1
d
1
+c
2
d
0
) t
2
+
The series converges absolutely within the radius of convergence of each
series.
(iii) Differentiation
Let f(t) =

¸
k=0
c
k
(t −t
0
)
k
in [t −t
0
[ < r
Then
df
dt
=
˙
f =

¸
k=1
c
k
k(t −t
0
)
k−1
in [t −t
0
[ < r
(iv) Integration
Let f(t) =

¸
k=0
c
k
(t −t
0
)
k
in [t −t
0
[ < r
Then

f(t)dt =

¸
k=0
c
k
k + 1
(t −t
0
)
k+1
in [t −t
0
[ < r
Definition 6.1.3 We shall say that a function f(t) is analytic at t = t
0
if it
can be expanded as a sum of a power series of the form

¸
k=0
c
k
(t −t
0
)
k
with
a radius of convergence r. It is clear that if f(t) is analytic at t
0
, then c
k
=
f
(k)
(t
0
)
k!
, k = 0, 1, 2, ( f
k
(t
0
) denotes the k
th
derivative of f at t
0
).
180 CHAPTER 6. SERIES SOLUTION
Example 6.1.2 We have the binomial expansion for (1 −t)
−k
for a fixed pos-
itive number k to give us
(1 −t)
−k
= 1 +kt +
k(k + 1)
2!
t
2
+
k(k + 1)...(k +r −1)
r!
t
r
+ for [t[ < 1
We denote by
u
r
=
k(k + 1)...(k +r −1)
r!
Then
(1 −t)
−k
=

¸
r=0
u
r
t
r
, [t[ < 1
As a power series can be differentiated term by term in its interval of conver-
gence, we have
k(1 −t)
−k−1
=

¸
r=1
ru
r
t
r−1
, [t[ < 1
This gives
k(1 − t)
−k
1
(1 −t)
=

¸
r=1
ru
r
t
r−1
, [t[ < 1
Again, using the power series expansion for
1
1 −t
= 1 +t +t
2
+ , we get
k


¸
r=0
t
r


¸
r=0
u
r
t
r

=

¸
r=1
ru
r
t
r−1
Using the product formula for LHS, we get
k

¸
r=0
t
r
r
¸
l=0
u
l
=

¸
r=0
(r + 1)u
r+1
t
r
Uniqueness of power series representation gives
(r + 1)u
r+1
= k
r
¸
l=0
u
l
(6.1.5a)
where
u
r
=
k(k + 1).....(k +r −1)
r!
, k is a fixed integer (6.1.5b)
We now make an attempt to define the concept of uniform convergence. To
define this concept, we consider the following series whose terms are functions
¦f
k
(t)¦

k=0

¸
k=0
f
k
(t) = f
0
(t) +f
1
(t) +f
2
(t) + (6.1.6)
Note that for f
k
(t) = c
k
(t −t
0
)
k
, we get the power series.
6.2. LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS 181
Definition 6.1.4 We shall say that the series given by Eq. (6.1.6) with sum
f(t) in an interval I ⊂ ' is uniformly convergent if for every > 0, we can find
N = N(t), not depending on t, such that
[f(t) −s
n
(t)[ < for all n ≥ N()
where s
n
(t) = f
0
(t) +f
1
(t) + +f
n
(t).
Theorem 6.1.3 A power series

¸
k=0
c
k
(t −t
0
)
k
with nonzero radius of conver-
gennt is uniformly convergent in every closed interval [t − t
0
[ ≤ p of radius
p < r.
Proof : For [t −t
0
[ ≤ p and any positive integers n and l we have

c
n+1
(t −t
0
)
n+1
+ +c
n+l
(t −t
0
)
n+l

≤ [c
n+1
[ p
n+1
+ +[c
n+l
[ p
n+l
(6.1.7)
The series

¸
k=0
c
k
(t −t
0
)
k
converges absolutely if [t −t
0
[ ≤ p < r ( by Theorem
6.1.1) and hence by Cauchy convergence, given > 0, ∃ N() such that
[c
n+1
[ p
n+1
+ +[c
n+p
[ p
n+l
< for n ≥ N(), l = 1, 2, ..
Eq. (6.1.7) gives

c
n+1
(t −t
0
)
n+1
+ +c
n+p
(t −t
0
)
n+l

≤ [c
n+1
[ p
n+1
+ +[c
n+p
[ p
n+l
< for n ≥ N(), [t −t
0
[ ≤ p < r
This implies that, given > 0, ∃ N() such that [f(t) −s
n
(t)[ < for n ≥ N()
and all t where s
n
(t) =
n
¸
k=0
c
k
(t −t
0
)
k
. This proves the uniform convergence of
the power series inside the interval [t −t
0
[ ≤ p < r.
Example 6.1.3 The geometric series 1 + t + t
2
+ is uniformly convergent
in the interval [t[ ≤ p < 1. It is not uniformly convergent in the whole interval
[t[ < 1.
6.2 Linear System with Analytic Coefficients
We now revisit the non-autonomous system in '
n
d¯ x
dt
= A(t)¯ x +g(t) (6.2.1)
¯ x(0) = ¯ x
0
182 CHAPTER 6. SERIES SOLUTION
where the matrix A(t) is analytic at t = 0 and hence has the power series
representation
A(t) =

¸
k=0
A
k
t
k
(6.2.2)
in its interval of convergence [t[ < r. Here each A
k
is n n matrix.
So, the homogeneous system corresponding to Eq. (6.2.1) is given by
d¯ x
dt
=

¸
k=0
A
k
t
k
¯ x (6.2.3)
¯ x(0) = ¯ x
0
We shall look for analytic solution of Eq. (6.2.3), which is of the form ¯ x(t) =

¸
k=0
¯ c
k
t
k
. The vector coefficients ¯ c
k
are to be determined. The point t = 0 is
called ordinary point of the above system.
The following theorem gives the analytic solution of the system given by Eq.
(6.2.3)
Theorem 6.2.1 The homogeneous system given by Eq. (6.2.3) has analytic
solution ¯ x(t) =

¸
k=0
¯ c
k
t
k
in the interval of convergence [t[ < r. This solution ¯ x(t)
is uniquely determined by the initial vector ¯ x
0
.
Proof : Let ¯ x(t) =

¸
k=0
¯ c
k
t
k
, where the vector coefficient ¯ c
k
are yet to deter-
mined. Plugging this representation in Eq. (6.2.3) we get

¸
k=1
k¯ c
k
t
k−1
=


¸
k=0
A
k
t
k


¸
k=0
¯ c
k
t
k

=

¸
k=0

k
¸
l=0
A
k−l
¯ c
l

t
k
Equivalently, we have

¸
k=0
(k + 1)¯ c
k+1
t
k
=

¸
k=0

k
¸
l=0
A
k−l
¯ c
l

t
k
(6.2.4)
Uniqeness fo power series in the interval [t[ < r, gives
(k + 1)¯ c
k+1
=
k
¸
l=0
A
k−l
¯ c
l
(6.2.5)
6.2. LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS 183
and hence
(k + 1) |¯ c
k+1
| ≤
k
¸
l=0
|A
k−l
| |¯ c
l
| (6.2.6)
By Theorem 6.1.3, the power series

¸
k=0
A
k
t
k
converges absolutely and uniformly
in the interval [t[ ≤ p < r and hence the terms A
k
p
k
must be uniformly bounded.
That is, ∃ M such that
|A
k
|p
k
≤ M, k ≥ 0 (6.2.7)
Using the above inequality in Eq. (6.2.6), we get
(k + 1)|¯ c
k+1
| ≤
k
¸
l=0
M
|¯ c
l
|
p
k−l
Put d
l
= p
l
|¯ c
l
|, then the above inequality becomes
(k + 1)d
k+1
≤ Mp
k
¸
l=0
d
l
(6.2.8)
Using Eq. (6.2.8) inductively, we get
d
1
≤ Mpd
0
d
2

Mp
2
(d
0
+d
1
)

1
2

Mp +M
2
p
2

d
0
.
.
.
d
k

Mp(Mp + 1)(Mp + 2) + + (Mp +k −1)
k!
d
0
(6.2.9)
To claim that Eq. (6.2.8) holds for all k, we need to prove this inequality by
induction. So, let this inequality be true for all r < k. By Eq. (6.2.8), we have
(r + 1)d
r+1
≤ Mp|¯ c
0
|
r
¸
l=0
d
l
≤ Mp|¯ c
0
|
r
¸
l=0
Mp(Mp + 1)(Mp + 2) + + (Mp +l −1)
l!
Using the notation of Example 6.1.2, set
u
l
=
Mp(Mp + 1)(Mp + 2) + + (Mp +l −1)
l!
184 CHAPTER 6. SERIES SOLUTION
Hence using Eq. (6.1.5), we get
(r + 1)d
r+1
≤ Mp|¯ c
0
|
r
¸
l=0
u
l
= |¯ c
0
|(r + 1)u
r+1
This gives
(r + 1)d
r+1
≤ (r + 1)
Mp(Mp + 1)(Mp + 2) + + (Mp +r)
(r + 1)!
|¯ c
0
|
That is
d
r+1

Mp(Mp + 1)(Mp + 2) + + (Mp +r)
(r + 1)!
|¯ c
0
|
This proves the induction.
Hence, it follows that
|¯ c
k
| =
d
k
p
k

Mp(Mp + 1)(Mp + 2) + + (Mp +k −1)
k!p
k
|¯ c
0
|
This gives
|¯ x| =


¸
k=0
¯ c
k
t
k

≤ |¯ c
0
|


¸
k=0
Mp(Mp + 1)(Mp + 2) + + (Mp +k −1)
k!

[t[
p

k

That is
|¯ x| ≤
|¯ c
0
|

1 −
[t[
p

Mp
(6.2.10)
provided [t[ ≤ p < r.
This proves the existence of analytic solution of the system given by Eq. (6.2.3),
This solution is uniquely determined by the initial value ¯ x
0
. For if ¯ x(t) and ¯ y(t)
are two different solutions of the initial value problem given by Eq. (6.2.3), then
¯ z(t) = ¯ x(t) − ¯ y(t) is the solution of the initial value problem
d¯ z
dt
= A(t)¯ z(t)
¯ z(0) =
¯
0
Since ¯ c
0
= 0, it follows that ¯ z =
¯
0, thereby implying that ¯ x = ¯ y.
6.2. LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS 185
Example 6.2.1 ( Legendre Differential Equation)
(1 −t
2
)
d
2
x
dt
2
− 2t
dx
dt
+n(n + 1)x = 0 (6.2.11)
By substituting x
1
= x, x
2
=
dx
dt
, this differential equation is equivalent to the
following linear homogeneous system
d¯ x
dt
= A(t)¯ x(t)
where
A(t) =

¸
¸
0 1
−n(n + 1)
1 −t
2
2t
1 −t
2
¸

As A(t) is analytic in the interval [t[ < 1, it follows by Theorem 6.2.1 that
Eq. (6.2.11) has a unique analytic solution of the form

¸
k=0
c
k
t
k
. We shall now
determine the coefficients c
k
. We substitute
x(t) =

¸
k=0
c
k
t
k
,
dx
dt
=

¸
k=1
kc
k
t
k−1
,
d
2
x
dt
2
=

¸
k=2
k(k −1)c
k
t
k−2
in Eq. (6.2.11) to get
(1 −t
2
)

¸
k=2
k(k −1)c
k
t
k−2
−2t

¸
k=1
kc
k
t
k−1
+n(n + 1)

¸
k=1
c
k
t
k
= 0
Equivalently, we have

¸
k=0
[(k + 2)(k + 1)c
k+2
+c
k
[n(n + 1) −(k −1)k −2k]] t
k
= 0
The uniqueness of series representaion in [t[ < 1 gives
c
k+2
= −
[n(n + 1) −(k −1)k −2k]
(k + 2)(k + 1)
c
k
= −
(n −k)(n +k + 1)
(k + 2)(k + 1)
c
k
, k = 0, 1, 2, (6.2.12)
This gives us
c
2
= −
n(n + 1)
2!
c
0
, c
4
= −
(n + 3)(n −2)
4.3
c
2
=
(n + 3)(n + 1)n(n −2)
4!
c
0
c
3
= −
(n + 2)(n −1)
3!
c
1
, c
5
= −
(n + 4)(n −3)
5.4
c
3
=
(n + 4)(n + 2)(n −3)(n −1)
5!
c
1
186 CHAPTER 6. SERIES SOLUTION
and hence the coefficients c
k
’s are inductively defined.
So, we have
x(t) = c
0
x
1
(t) +c
1
x
2
(t)
x
1
(t) = 1 −
n(n + 1)
2!
t
2
+
(n −2)n(n + 1)(n + 3)
4!
t
4

x
2
(t) = t −
(n −1)(n + 2)
3!
t
3
+
(n −3)(n −1)(n + 2)(n + 4)
5!
t
5

Since RHS of Eq. (6.2.12) is zero for k = n, we have 0 = c
n+2
= c
n+4
=
and hence if n is even, x
1
(t) reduces to a polynomial of degree n and if n is odd,
the same is true for x
2
(t). These polynomials multiplied by some constants, are
called Legendre polynomials.
In Eq. (6.2.12), writing c
k
in terms of c
k+2
, we get
c
k
= −
(k + 2)(k + 1)
(n −k)(n +k + 1)
c
k+2
(k ≤ n −2)
This gives
c
n−2
= −
n(n −1)
2(2n −1)
c
n
We fix c
n
and then compute the lower order coefficients. It is standard to choose
c
n
= 1 when n = 0 and
2n!
2
n
(n!)
2
=
1.3.5...(2n −1)
n!
for n = 0. This defines
c
n−2
= −
n(n −1)
2(2n −1)
(2n)!
2
n
(n!)
2
= −
n(n −1)2n(2n −1)(2n −2)!
2(2n −1)2
n
(n −1)!n(n −1)(n −2)!
= −
(2n −2)!
2
n
(n −1)!(n −2)!
In general
c
n−2m
=
(−1)
m
(2n −2m)!
2
n
m!(n −m)!(n −2m)!
So we get a solution of Legendre differential equation given by Eq. (6.2.11) as
P
n
(t) =
M
¸
m=0
(−1)
m
(2n −2m)!
2
n
m!(n −m)!(n −2m)!
t
n−2m
where M =
n
2
or
n −1
2
whichever is an integer. Some of the Legendre polyno-
mials are given as below
6.2. LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS 187
P
0
(t) = 1, P
1
(t) = t, P
2
(t) =
1
2
(3t
2
−1), P
3
(t) =
1
2
(5t
3
−3t)
P
4
(t) =
1
8
(35t
4
−30t
2
+ 3), P
5
(t) =
1
8
(63t
5
−70t
3
+ 15t)
We shall discuss the properties of Legendre differential equation in section 4.
Example 6.2.2 (Hermite Differential Equation)
In the study of the linear harmonic oscillator in quantum mechanics, one en-
counters the following deifferential equation
d
2
u
dt
2
+

λ −t
2

u(t) = 0 (6.2.13)
over the interval (−∞, ∞). As we look for the solution u ∈ L
2
(−∞, ∞), we
put u(t) = e

−t
2
2

x(t). So Eq. (6.2.13) becomes
d
2
x
dt
2
−2t
dx
dt
+ (λ −1) x = 0
We take λ = 2n + 1 and get the equation
d
2
x
dt
2
−2t
dx
dt
+ 2nx = 0 (6.2.14)
Eq. (6.3.13) (or alternatively Eq. (6.2.14)) is called Hermite differential equa-
tion.
As the coefficient of the differential equation given by Eq. (6.2.14) are analytic
in (−∞, ∞), we get a series solution of the form
x(t) =

¸
k=0
c
k
t
k
for − ∞< t < ∞ (6.2.15)
Plugging the representation for x(t),
dx
dt
and
dx
2
dt
2
in Eq. (6.2.14) we see that
c
k
satisfy the recurrence relation
c
k+2
=
2k + 1 −(2n + 1)
(k + 2)(k + 1)
c
k
=
2(k −n)
(k + 2)(k + 1)
c
k
(6.2.16)
It is clear that c
n+2
= 0 and hence the solution given by Eq. (6.2.15) is a
polynomial. As in the previous example, we normalize c
n
and put it equal to
2
n
and then compute the coefficients in descending order. This gives us the
solution, which is denoted by H
n
(t), the Hermite polynomial. It is given by
H
n
(t) =
M
¸
m=0
(−1)
m
n!
m!(n −2m)!
(2t)
n−2m
We shall discuss the properties of the Hermite polynomials in section 4.
188 CHAPTER 6. SERIES SOLUTION
6.3 Linear System with Regular Singularities
A system of the form
d¯ x
dt
=
1
t −t
0
A(t)¯ x (6.3.1)
where A(t) is analytic at t
0
, is said to have a regular singularity at t
0
.
As before, we can assume that t
0
= 0 and A(t) has series representation of the
form
A(t) =

¸
k=0
A
k
t
k
(6.3.2)
in the interval of the convergence [t[ < r.
Eq. (6.3.1) then reduces to the equation of the form
d¯ x
dt
=
1
t


¸
k=0
A
k
t
k

¯ x (6.3.3)
We prove the following theorem, giving the existence of a series solution of the
equation given by Eq. (6.3.3)
Theorem 6.3.1 The differential equation given by Eq. (6.3.3) has a series
solution of the form ¯ x(t) = t
µ

¸
k=0
¯ c
k
t
k
in the interval [t[ < r, provided µ is an
eigenvalue of A
0
and no other eigenvalue of the form µ+n exists for A
0
, where
n is a positive integer.
Proof : We introduce a new dependent variable ¯ y(t) = t
µ
¯ x(t). This gives
d¯ y
dt
=
1
t
[A(t) −µI] ¯ y(t)
=
1
t
¸

¸
k=0
A
k
t
k
−µI
¸
¯ y(t)
Assuming that ¯ y(t) =

¸
k=0
¯ c
k
t
k
, we get
6.3. LINEAR SYSTEM WITH REGULAR SINGULARITIES 189

¸
k=1
¯ c
k
kt
k
=

¸
k=0
¸
k
¸
l=0
A
k−l
¯ c
l

−µ¯ c
k
¸
t
k
. This gives us the recurrence relations
(A
0
−µI)¯ c
0
= 0
(A
0
−(µ +I)) ¯ c
1
= −A
1
¯ c
0
(6.3.4)
.
.
.
(A
0
−(µ +n)I)¯ c
n
= −
n−1
¸
l=0
A
n−l
¯ c
l
.
.
.
Since [(A
0
−(µ+n)I)¯ c
n
[ = 0 for all n ≥ 1, the above relations iteratively define
¯ c
0
, ¯ c
1
, , ¯ c
n
, We shall now show that the series x(t) = t
µ

¸
k=0
¯ c
k
t
k
converges.
As in Section 6.2, for some positive numbers M, and p < r we have
|A
k
|p
k
≤ M, k = 0, 1, 2, (6.3.5)
We first observe the following
(i) [A
0
−(µ +k)I[ = 0 for all positive integer k
(ii) lim
k→∞

A
0

(µ +k)
k
I

= lim
k→∞

(A
0
−µI)
k

k
k
I

= n
In view of property (ii), it follows that there exists a positive quantity > 0
such that
|A
0
−(µ +k)I|
k
> , k = 1, 2, 3,
So, the recurrence relation given by Eq. (6.3.4) provides
(k + 1)¯ c
k+1
= [A
0
−(µ +k + 1)I]
−1
(k + 1)
k
¸
l=0
A
k+1−l
¯ c
l
Using Eq. (6.3.5) and (ii) we get
(k + 1) |¯ c
k+1
| ≤
M(k + 1)

k
¸
l=0
|¯ c
l
|
p
k+1−l
Equivalently,
(k + 1) |¯ c
k+1
| p
k+1

M

(k + 1)
k
¸
l=0
|¯ c
l
|p
l
190 CHAPTER 6. SERIES SOLUTION
Proceeding as in Theorem 6.2.1, we have the relation
|¯ c
k
| ≤

M

M

+ 1

M

+k −1

k!p
k
|¯ c
0
|
for all k ≥ 1.
Applying this estimate to the series expansion for ¯ x(t), we get
|¯ x(t)| ≤

t
µ

¸
k=0
¯ c
k
t
k

≤ [t[
µ
|¯ c
0
|

¸
k=0

M

M

+ 1

M

+k −1

k!

[t[
p

k
=
[t[
µ
|¯ c
0
|

1 −
[t[
p

M

in the interval of uniform convergence [t[ ≤ p < r.
This proves the theorem.
Example 6.3.1 (Bessel’s Equation)
t
2
d
2
x
dt
2
+t
dx
dt
+ (t
2
−µ
2
)x(t) = 0 (6.3.6)
⇐⇒
d
2
x
dt
2
+
1
t
dx
dt
+ (1 −
µ
2
t
2
)x(t) = 0
We use the substitution x
1
= x(t), x
2
= t ˙ x(t) to get
dx
1
dt
=
1
t
x
2
dx
2
dt
=
dx
1
dt
+t
d
2
x
1
dt
2
=
dx
1
dt

dx
1
dt
−(t −
µ
2
t
)x
1
(t)
= −
1
t
(t
2
−µ
2
)x
1
(t)
This reduces Eq. (6.3.6) to
d¯ x
dt
=
1
t

¸
0 1
µ
2
−t
2
0
¸

¯ x
=
1
t

A
0
+A
2
t
2

¯ x
6.3. LINEAR SYSTEM WITH REGULAR SINGULARITIES 191
where A
0
=

0 1
µ
2
0

, A
2
=

0 0
−1 0

. It is clear that A is analytic in
(−∞, ∞) and A
0
has eigenvalues ±µ.
Case 1: µ is not an integer
Then Eq. (6.3.6) has solution of the form x(t) = t
µ
¸

k=0
¯ c
k
t
k
in (−∞, ∞).
For the eigenvalue µ > 0, plugging the representation
x(t) =

¸
m=0
c
m
x
m+µ
,
dx
dt
=

¸
m=0
c
m
(m+µ)t
m+µ−1
d
2
x
dt
2
=

¸
m=0
c
m
(m+µ)(m +µ −1)t
m+µ−2
in Eq. (6.3.6) we get the recurrence relation
c
2m
= −
1
2
2
m(µ +m)
c
2m−2
, m = 1, 2
c
2m+1
= 0, m = 0, 1, 2,
This gives
c
2
= −
c
0
2
2
(µ + 1)
c
4
= −
c
2
2
2
2(µ + 2)
=
c
0
2
4
2!(µ + 1)(µ + 2)
.
.
.
c
2m
=
(−1)
m
c
0
2
2m
m!(µ + 1)(µ + 2) (µ +m)
, (6.3.7)
m = 1, 2,
c
2m+1
= 0, m = 0, 1, 2, 3,
Now define the gamma function Γ(µ) as
Γ(µ) =


0
e
−t
t
µ−1
dt, µ > 0
which is a convergent integral. Γ(µ) satisfies the property that Γ(µ+1) = µΓ(µ)
and Γ(n + 1) = n! If we put c
0
=
1
2
µ
Γ(µ + 1), then c
2m
is given by
c
2m
=
(−1)
m
2
2m+µ
m!(µ + 1)(µ + 2) (µ +m)Γ(µ + 1)
=
(−1)
m
2
2m+µ
m!Γ(µ +m+ 1)
192 CHAPTER 6. SERIES SOLUTION
This gives us the first solution J
µ
(t), called Bessl’s function of order µ of the
differential equation given by Eq. (6.3.6), as
J
µ
(t) = t
µ

¸
m=0
(−1)
m
t
2m
2
2m+µ
m!Γ(µ +m+ 1)
Extending gamma function for µ < 0, we get the representation for the second
solution J
−µ
(t) as
J
−µ
(t) = t
−µ

¸
m=0
(−1)
m
t
2m
2
2m−µ
m!Γ(−µ +m + 1)
J
µ
(t) and J
µ
(t) are linearly independent.
Case 2: µ = n integer
We have
J
n
(t) = t
n

¸
m=0
(−1)
m
t
2m
2
2m+n
m!(n +m)
One can easily derive that
J
−n
(t) = (−1)
n
J
n
(t)
This gives
J
0
(t) =

¸
m=0
(−1)
m
t
2m
2
2m
(m!)
2
= 1 −
t
2
2
2
(1!)
2
+
t
4
2
4
(2!)
2

t
6
2
6
(3!)
2
+
J
1
(t) =

¸
m=0
(−1)
m
t
2m+1
2
2m+1
(k!)(k + 1)!
= t −
t
3
2
3
(1!)(2!)
+
t
5
2
5
(2!)(3!)

t
7
2
7
(3!)(4!)
+
J
0
looks similar to cosine function and J
1
looks similar to sine function. The
zeros of these functions are not completely regularly spaced and also oscillations
are damped as we see in the following graphics.
6.3. LINEAR SYSTEM WITH REGULAR SINGULARITIES 193
5 10 15 20
t
-0.4
-0.2
0.2
0.4
0.6
0.8
1
J0
J1
Figure 6.3.1: Sketch of J
0
and J
1
Example 6.3.2 (Bessel’s equation of order zero)
Consider the equation
x
d
2
x
dt
2
+
dx
dt
+x = 0
This is equivalent to the first order system of the form
d¯ x
dt
= −
1
t

0 1
−t
2
0

¯ x (6.3.8)
where x
1
= x, x
2
= t
dx
dt
.
As we have seen before, a solution of this system is given by
x
1
(t) =

¸
k=0
(−1)
k
t
2k
4
k
(k!)
2
(6.3.9)
x
2
(t) =

¸
k=0
(−1)
k
t
2k
4
k
(k!)
2
(6.3.10)
To get another solution ¯ y(t) linearly independent of ¯ x(t) and satisfying Eq.
(6.3.8), we set
¯ y(t) = ψ(t)¯ z(t) where ψ(t) =

¸
1 x
1
0 x
2
¸

.
194 CHAPTER 6. SERIES SOLUTION
This gives
d¯ y
dt
=

dt
¯ z(t) +ψ(t)
d¯ z
dt
= A(t)ψ(t)¯ x(t)
That is

¸
¸
0
1
t
x
2
0 −tx
1
¸

¯ z +

¸
1 x
1
0 x
2
¸

dz
dt
=

¸
¸
0
1
t
x
2
−t −tx
1
¸

¯ z(t)
Solving for
dz
dt
we get

¸
1 x
1
0 x
2
¸

dz
dt
=

¸
0 0
−t 0
¸

¯ z(t)
and hence
dz
dt
=

¸
¸
¸
t
x
1
x
2
0

t
x
2
0
¸

¯ z(t)
So, we need to solve
dz
1
dt
= t
x
1
x
2
z
1
= −
˙ x
2
x
2
z
1
(6.3.11a)
dz
2
dt
= −
t
x
2
z
1
(6.3.11b)
Integrating Eq. (6.3.11(a)) - Eq. (6.3.11(b)) we get
z
1
=
c
1
x
2
, z
2
= −c
1

t
x
2
2
(t)
dt
Hence ¯ y(t) = ψ(t)¯ z(t) is given by
¯ y(t) =

¸
¸
¸
¸
c
1
x
2
−c
1
x
1

tdt
x
2
2
(t)
−c
1
x
2

t
x
2
2
(t)
dt
¸

That is
y
1
(t) =
c
1
x
2
−c
1
x
1
¸
− ˙ x
2
/x
1
x
2
2

dt
= −c
1
x
1

˙ x
2
x
2
x
2
1
dt
= −c
1
x
1

1
tx
2
1
dt
6.3. LINEAR SYSTEM WITH REGULAR SINGULARITIES 195
y
2
(t) = −c
1
x
2

t
x
2
2
dt
Using the represntations of Eq. (6.3.9) - Eq. (6.3.10), we get
x
1
(t) = 1 −
t
2
4
+
x
2
(t) = −
t
2
2
+
t
4
16
+
This gives
x
1

dt
tx
2
1
= x
1

1 +
t
2
2
+
t
dt
= x
1
lnt +x
1
(
t
2
4
+ )
and
x
2

t
x
2
2
dt = 4x
2

t

1 +
t
2
4
+

t
4
dt
=
−2x
2
t
2
+x
2
lnt +
Hence, it follows that the second solution ¯ y(t), linearly independent of x(t), is
given by
y(t) = x(t)lnt +x(t)

1 +
t
2
4
+

To be more precise, one can show that
y(t) = J
0
(t)lnt +

¸
m=1
(−1)
m
h
m
t
2m
2
2m
(m!)
2
where h
m
=

1 +
1
2
+ +
1
m

.
The function Y
0
(t) defined as
Y
0
(t) = ay(t) +bJ
0
(t), a =
2
π
, b = r −ln2
where r is the Euler’s constant lim
n→∞

1 +
1
2
+ +
1
m
−lnn

is called the
Bessel’s function of the second kind (order zero). Hence, it is given by
Y
0
(t) =
2
π
¸
J
0
(t)

ln
t
2
+r

+

¸
m=1
(−1)
m
h
m
t
2m
2
2m
(m!)
2
¸
196 CHAPTER 6. SERIES SOLUTION
6.4 Properties of Legendre, Hermite and Bessel
Functions
Legendre polynomial P
n
(t) is a solution of the differential equation
(1 −t
2
)
d
2
x
dt
2
−2t
dx
dt
+n(n + 1)x(t) = 0 (6.4.1)
In terms of the notation of Chapter 6, this is eigenvalue problem of the form
Lx = λx (6.4.2)
where L is the second order differential operator
Lx = (1 −t
2
)
d
2
x
dt
2
−2t
dx
dt
=
d
dt

(1 −t
2
)
dx
dt

(6.4.3)
We can think of L as an operator defined on L
2
[−1, 1] with D(L) as the set of
all functions having second order derivatives.
Then Green’s formula (to be done in Section 7.1) gives

1
−1
(yLx −xL

y) dt = J(x, y)

1
−1
where J(x, y)[
1
−1
= a
0

y
dx
dt
−x
dy
dt

1
−1
with a
0
(t) = (1 −t
2
).
As a
0
(1) = a
0
(−1) = 0, it follows that J(x, y)[
1
−1
= 0 and hence

1
−1
(yLx −xL

y) dt = 0 ∀x, y ∈ D(L) = D(L

)
That is L

= L and D(L

) = D(L).
We appeal to Theorem 7.5.1, of Chapter 7 to claim that the eigenfunctions
of L form a complete orthonormal set. We have already proved in Example
6.2.1 that Eq. (6.4.1) has sequence of eigenvalues ¦n(n + 1)¦ with ¦P
n
(t)¦
corresponding eigenfunctions and hence they are orthogonal. However, for the
sake of completeness we prove this result directly.
Theorem 6.4.1 The set of Legendre polymonials ¦P
n
(t)¦ satisfying Eq. (6.4.1)
are orthogonal set of polynomials in L
2
[−1, 1].
Proof : We have
d
dt

(1 −t
2
)
dP
n
dt

= −n(n + 1)P
n
(t) (6.4.4)
d
dt

(1 −t
2
)
dP
m
dt

= −m(m+ 1)P
m
(t) (6.4.5)
6.4. PROPERTIES OF LEGENDRE, HERMITE ANDBESSEL FUNCTIONS197
Multiplying Eq. (6.4.4) by P
m
(t) and Eq. (6.4.5) by P
n
(t) and substracting the
equations and integrating, we get

1
−1
[m(m + 1) −n(n + 1)] P
m
(t)P
n
(t)dt
=

1
−1
¸
d
dt

(1 −t
2
)
dP
n
dt

P
m
(t) −
d
dt

(1 −t
2
)
dP
m
dt
P
n
(t)

dt
=

(1 −t
2
)

˙
P
n
(t)P
m
(t) −
˙
P
m
(t)P
n
(t)

1
−1

1
−1

(1 −t
2
)

˙
P
n
(t)
˙
P
m
(t) −
˙
P
m
(t)
˙
P
n
(t)

= 0
Hence (P
m
, P
n
) =

1
−1
P
m
(t)P
n
(t)dt = 0. That is P
m
⊥ P
n
.
Theorem 6.4.2 The Legendre polynomials P
n
(t) satisfy the Rodrigues formula
P
n
(t) =
1
2
n
n!
d
n
dt
n

(t
2
−1)
n

, n = 0, 1, 2, (6.4.6)
Proof : We have
P
n
(t) =
1
2
n
n!
M
¸
k=0
(−1)
k
n!
k!(n −k)!
(2n −2k)!
(n −2k)!
t
n−2k
Since
d
n
dt
n

t
2n−2k

=
(2n −2k)!
(n −2k)!
t
n−2k
, it follows that
P
n
(t) =
1
2
n
n!
d
n
dt
n
¸
M
¸
k=0
(−1)
k
n!
k!(n −k)!
t
2n−2k
¸
It is now clear that the sum
¸
¸
M
k=0
(−1)
k
n!
k!(n −k)!
t
2n−2k

is the binomial expansion
of (t
2
−1)
n
and hence
P
n
(t) =
1
2
n
n!
d
n
dt
n
(t
2
−1)
n
Definition 6.4.1 Let ¦f
n
(t)¦ be a sequence of functions in an interval I. A
function F(t, u) is said to be a generating function of this sequence if
F(t, u) =

¸
n=0
f
n
(t)u
n
(6.4.7)
198 CHAPTER 6. SERIES SOLUTION
We have the following theorem giving the generating function for the sequence
¦P
n
(t)¦ of Legendre polynomials.
Theorem 6.4.3 For Legendre polynomials ¦P
n
(t)¦, we have
(1 −2tu +u
2
)

1
2
=

¸
n=0
P
n
(t)u
n
(6.4.8)
Proof : Let [t[ ≤ r (arbitary positive number) and [u[ < (1 +r
2
)
1
2
−r. Then
[2tu −u
2
[ ≤ 2[t[[u[ +[u
2
[
≤ 2r(1 +r
2
)
1
2
− 2r
2
+ (1 +r
2
) +r
2
−2r(1 +r
2
)
1
2
= 1
So expanding (1 −2tu +u
2
)

1
2
in a binomial series, we get
[1 − u(2t −u)]

1
2
= 1 +
1
2
u(2t −u) +
1
2
3
4
u
2
(2t −u)
2
+
+
1.3......(2n −1)
1.2.....(2n)
u
n
(2t −u)
n
+
The coefficient of u
n
in this expression is
1.3......(2n −1)
2.4....(2n)
(2t)
n

1.3......(2n −3)
2.4....(2n −2)
(2t)
n−2
+
=
1.3......(2n −1)
n!
¸
t
n

n(n −1)
(2n −1)
2
t
n−2
+
n(n −1)(n −2)(n −3)
(2n −1)(2n −3)2.4
t
n−4
+

= P
n
(t)
Theorem 6.4.4 The Legendre polynomials satisfy the following recurrence re-
lation
(n + 1)P
n+1
(t) = (2n + 1)tP
n
(t) −nP
n−1
(t), n = 1, 2, ... (6.4.9)
Proof : Differentiating the relation given by Eq. (6.4.8) with respect to u, we
get
(t −u)(1 −2tu +u
2
)

3
2
=

¸
n=1
nP
n
(t)u
n−1
This gives
(t −u)(1 −2tu +u
2
)

1
2
= (1 −2tu +u
2
)

¸
n=0
nP
n
(t)u
n−1
6.4. PROPERTIES OF LEGENDRE, HERMITE ANDBESSEL FUNCTIONS199
That is
(t −u)

¸
n=1
P
n
(t)u
n
= (1 −2tu +u
2
)

¸
n=0
nP
n
(t)u
n−1
Equivalently,

¸
n=0
tP
n
(t)u
n


¸
n=0
P
n
(t)u
n+1
=

¸
n=1
nP
n
(t)u
n−1
−2

¸
n=1
ntP
n
(t)u
n
+

¸
n=0
nP
n
(t)u
n+1
Rewriting, we get

¸
n=1
tP
n
(t)u
n


¸
n=1
P
n−1
(t)u
n
=

¸
n=1
(n + 1)P
n+1
(t)u
n
−2

¸
n=1
ntP
n
(t)u
n
+

¸
n=1
(n −1)P
n−1
(t)u
n
Comparing the coefficient of u
n
, we get
(n + 1)P
n+1
(t) = (2n + 1)tP
n
(t) −nP
n−1
(t)
Corollary 6.4.1
|P
n
|
2
=

1
−1
P
2
n
(t)dt =
2
2n + 1
(6.4.10)
Proof : Recurrence relation gives
(2n + 1)tP
n
(t) = (n + 1)P
n+1
(t) + nP
n−1
(t)
(2n −1)tP
n−1
(t) = nP
n
(t) + (n −1)P
n−2
(t)
These two equations give
0 =

1
−1
[tP
n
(t)P
n−1
(t) −tP
n−1
(t)P
n
(t)] dt
=
(n + 1)
(2n + 1)

1
−1
P
n+1
(t)P
n−1
(t) +
n
(2n + 1)

1
−1
P
2
n−1
(t)dt

(n −1)
(2n −1)

1
−1
P
n−2
(t)P
n
(t) −
n
(2n −1)

1
−1
P
2
n
(t)dt
200 CHAPTER 6. SERIES SOLUTION
This implies that

1
−1
P
2
n
(t)dt =
(2n −1)
(2n + 1)

1
−1
P
2
n−1
(t)dt
and hence

1
−1
P
2
n
(t)dt =
(2n −1)
(2n + 1)
(2n −3)
(2n −1)

2
1

1
−1
P
2
0
(t)dt
=
2
(2n + 1)
We now examine the properties of the Hermite polynomials. Recall that Hermite
polynomial H
n
(t) satisfy the differential equation
d
2
x
dt
2
−2t
dx
dt
+ 2nx = 0
and is given by
H
n
(t) =

¸
k=0
(−1)
n
n!
k!(n −2k)!
(2t)
n−2k
We have the following theorem concerning the orthogonality of ¦H
n
(t)¦ in
L
2
(−∞, ∞) with respect to the weight function e
−t
2
.
Theorem 6.4.5 The Hermite Polynomials H
n
(t) are orthogonal set of polyno-
mials in the space L
2
(−∞, ∞) with respect to the weight function e
−t
2
.
Theorem 6.4.6 For Hermite polynomials H
n
(t), we have the following formu-
lae.
(i) Rodgriues formula
H
n
(t) = (−1)e
t
2 d
n
dt
n
e
−t
2
(ii) Generating function
e
2tu−u
2
=

¸
n=0
H
n
(t)
u
n
n!
(iii) Recurrence relation
H
n+1
(t) = 2tH
n
(t) −2nH
n−1
(t), H
0
= 1, H
1
= 2t
We now enunciate the properties of the Bessel’s function J
n
(t).
6.4. PROPERTIES OF LEGENDRE, HERMITE ANDBESSEL FUNCTIONS201
Theorem 6.4.7 For each fixed nonnegative integer n, the sequence of Bessel
functions J
n
(k
m
t), where k
m
are the zeros of J
n
(k), form an orthogonal set in
L
2
[0, 1] with respect to the weight function t. That is

1
0
tJ
n
(k
l
t)J
n
(k
m
t)dt = 0, l = m (6.4.11)
Proof : The Bessel function J
n
(t) satisfies the differential equation (refer Eq.
(6.3.6))
t
2
¨
J
n
(t) +t
˙
J
n
(t) + (t
2
−n
2
)J
n
(t) = 0
Set t = ks, then the above equation reduces to the differential equation
d
ds

s
˙
J
n
(ks)

+


n
2
s
+k
2
s

J
n
(ks) = 0 (6.4.12)
Let k
m
be the zeros of the Bessel function J
n
(k), then we have
d
ds

s
˙
J
n
(k
l
s)

+


n
2
s
+k
2
l
s

J
n
(k
l
s) = 0 (6.4.13)
and
d
ds

s
˙
J
n
(k
m
s)

+


n
2
s
+k
2
m
s

J
n
(k
m
s) = 0 (6.4.14)
Eq. (6.4.13) - Eq. (6.4.14) give
(k
2
l
−k
2
m
)

1
0
sJ
n
(k
l
s)J
n
(k
m
s)ds
=

1
0
¸
d
ds

s
˙
J
n
(k
l
s)

J
n
(k
m
s) −
d
ds

s
˙
J
n
(k
m
s)

J
n
(k
l
s)

ds
= 0, k
l
= k
m
using the fact that J
n
(k
l
) = 0 = J
n
(k
m
).
Theorem 6.4.8 The generating function for the sequence ¦J
n
(t)¦ of Bessel
functions is given by
exp

1
2
t

u −
1
u

=

¸
n=−∞
J
n
(t)u
n
(6.4.15)
and hence J
−n
(t) = (−1)
n
J
n
(t).
202 CHAPTER 6. SERIES SOLUTION
Proof : Expanding exp

1
2
t

u −
1
u

in powers of u, we get
exp

1
2
t

u −
1
u

=

exp
1
2
tu

exp
1
2

−t
u

=


¸
k=0
(tu)
k
2
k
k!


¸
l=0
(−t)
l
2
l
l! u
l

=

¸
−∞
c
n
(t)u
n
The coefficient c
n
(t) of u
n
in the above expression is

¸
k=0
(−1)
k

t
2

2k+n
k!(k +n)!
= J
n
(t)
and hence we get the generating function relation given by Eq. (6.4.15) for
J
n
(t).
If we replace u by −
1
v
in Eq. (6.4.15), we get
exp(
1
2
(v −
1
v
)t) =

¸
n=−∞
J
n
(t)(−1)
n
v
−n
=

¸
n=−∞
J
−n
(t)v
n
Hence, it follows that J
−n
(t) = (−1)
n
J
n
(t)
Theorem 6.4.9 The Bessel functions J
µ
(t) satisfy the following recurrence re-
lations
J
µ−1
(t) +J
µ+1
(t) =

t
J
µ
(t) (6.4.16a)
J
µ−1
(t) −J
µ+1
(t) = 2J
µ
(t) (6.4.16b)
Proof : We have
J
µ
(t) = t
µ


¸
k=0
(−1)
k
t
2k
2
2k+µ
(k!)Γ(µ +k + 1)

Multiplying J
µ
(t) by t
µ
and pulling t

under the summation, we have
t
µ
J
µ
(t) =

¸
k=0
(−1)
k
t
2k+2µ
2
2k+µ
(k!)Γ(µ +k + 1)
6.4. PROPERTIES OF LEGENDRE, HERMITE ANDBESSEL FUNCTIONS203
Differentiating the above expression we have
d
dt
(t
µ
J
µ
(t)) =

¸
k=0
(−1)
k
2(k +µ)t
2k+2µ−1
2
2k+µ
(k!)Γ(µ +k + 1)
= t
µ
t
µ−1

¸
k=0
(−1)
k
t
2k
2
2k+µ−1
(k!)Γ(µ +k)
= t
µ
J
µ−1
(t)
Thus, we have
d
dt
(t
µ
J
µ
(t)) = t
µ
J
µ−1
(t)
and
d
dt

t
−µ
J
µ
(t)

= −t
−µ
J
µ+1
(t)
Expanding the LHS of the above expressions, we have
µt
µ−1
J
µ
+t
µ
˙
J
µ
= t
µ
J
µ−1
−µt
µ−1
J
µ
+t
µ
˙
J
µ
= −t
µ
J
µ+1
Adding and substracting the above relations, we get
J
µ−1
+J
µ+1
=

t
J
µ
(t)
J
µ−1
−J
µ+1
= 2
˙
J
µ
(t)
Corollary 6.4.2 From the definition of J
µ
(t), we have
J1
2
(t) =

2
πt
sin t, J

1
2
(t) =

2
πt
cos t
and hence the above recurrence relations give
J3
2
(t) =

2
πt

sin t
t
−cos t

J

3
2
(t) =

2
πt


cos t
t
−sint

We have the following graphics for the above functions.
204 CHAPTER 6. SERIES SOLUTION
5 10 15 20
t
-0.25
0.25
0.5
0.75
1
J-1\2
J1\2
Figure 6.4.2: Sketch of J
−1/2
(t) and J
1/2
(t)
5 10 15 20
t
-1
-0.75
-0.5
-0.25
0.25
0.5
J3\2
J-3\2
Figure 6.4.3: Sketch of J
3/2
(t) and J
−3/2
(t)
For more details on varioius topics in this chapter, refer Agarwal and Gupta
[1], Brown and Churchill [2], Hochstadt [3] and Kreyzig [4].
6.5. EXERCISES 205
6.5 Exercises
1. Let the functions f, g : ' →' be defined as follows
f(t) =

sin t
t
, t = 0
1, t = 0
, g(t) =

1 −cos t
t
2
, t = 0
1
2
, t = 0
Show that f and g are analytic at t = 0.
2. Find the series solution of the following IVPs
(i) ¨ x +t ˙ x −2x = 0, x(0) = 1, ˙ x(0) = 0
(ii) t(2 −t)¨ x −6(t −1) ˙ x −4x = 0, x(1) = 1, ˙ x(1) = 0
(iii) ¨ x +e
t
˙ x + (1 +t
2
)x = 0, x(0) = 1, ˙ x(0) = 0
(iv) ¨ x −(sin t)x = 0, x(π) = 1, ˙ x(π) = 0
3. Show that the general solution of the Chebyshev differential equation
(1 −t
2
)¨ x −t ˙ x +a
2
x = 0
is given by
x(t) = c
0
¸
1 +

¸
n=1
(−a)
2
(2
2
−a
2
)......((2n −2)
2
−a
2
)
(2n)!
t
2n
¸
+ c
1
¸
t +

¸
n=1
(1 −a
2
)(3
2
−a
2
)......((2n −1)
2
−a
2
)
(2n −1)!
t
2n+1
¸
4. Show that the two linearly independent solutions of the following differ-
ential equation
t
2
¨ x +t(t −
1
2
) ˙ x +
1
2
x = 0
are given by
x
1
(t) = [t[

¸
n=0
(−1)
n
(2t)
n
(2n + 1)(2n −1)......3.1
x
2
(t) = [t[
1
2

¸
n=0
(−1)
n
t
n
n!
5. Show that the two linearly independent solutions of the differential equa-
tion
t(1 −t)¨ x + (1 −t) ˙ x −x = 0
206 CHAPTER 6. SERIES SOLUTION
are given by
x
1
(t) = 1 +

¸
n=1
1.2.5......((n −1)
2
+ 1)
(n!)
2
t
n
x
2
(t) = x
1
(t)ln[t[ + 2


¸
n=1
1.2.5.....((n −1)
2
+ 1)
(n!)
2

n
¸
k=1
k −2
k((k −1)
2
+ 1)

t
n
6. Using Rolle’s theorem show that between two consucative zeros of J
n
(t),
there exists precisely one zero of J
n+1
(t).
7. Prove the following identities for Bessel functions J
µ
(t), µ ∈ '.
(i)

t
µ
J
µ−1
(t)dt = t
µ
J
µ
(t) +c
(ii)

t
−µ
J
µ+1
(t)dt = −t
−µ
J
µ
(t) +c
(iii)

J
µ+1
(t)dt =

J
µ−1
(t)dt −2J
µ
(t)
8. Show that
|J
n
(k
lm
t)|
2
=

1
0
tJ
2
n
(k
lm
t)dt =
1
2
J
2
n+1
(k
lm
)
9. Using Rodrigues formula, show that
P
n
(0) =

0, n is odd
(−1)
n/2
1.3....n −1
2.4.....n
, n is even
10. Let p(t) be a polynomial of degree n. Show that p(t) is orthogonal to all
Legendre polynomials of degree strictly less than n.
References
[1 ] Ravi P. Agarwal and Ramesh C. Gupta, Essentials of Ordinary differen-
tial Equations, McGraw Hill Book Company, 1993
[2 ] James Ward Brown and Rulev Churchill, Fourier Series and Boundary
Value Problems, McGraw Hill, Inc., 1993
[3 ] Harry Hochstadt, Differential Equations, A Modern Approach, Dover
Publications, Inc., 1964
[4 ] Erwin Kreyszig, Advanced Engineering Mathematics, John Wiley and
Sons , Inc., 1999
Chapter 7
Boundary Value Problems
In this chapter we mainly deal with boundary value problems involving sec-
ond order differential operators. The main tool for solvability analysis of such
problems is the concept of Green’s function. This transforms the differential
equation in to integral equation, which, at times, is more informative.
Hence there is a detailed discussion involving the definition of Green’s func-
tion through Dirac-delta function, its properties and construction. Numerous
examples are given for illustration.
Towards the end, we give eigenfunction expansion technique for computing the
solution of boundary value problems. Nonlinear boundary value problem is also
the topic of our investigation. We use both fixed point theory and monotone
operators for this analysis.
7.1 Introduction
We shall be concerned with the solution of the ordinary differential equation
Lx = f (7.1.1)
over the interval a ≤ t ≤ b, subject to certain boundary conditions. Here L is a
second order linear differential operator of the form
L = a
0
(t)
d
2
dt
2
+a
1
(t)
d
dt
+a
2
(t), a
0
(t) = 0 on [a, b]
Since L is of second order, there will be two boundary conditions of the form
B
i
(x) = C
i
, i = 1, 2 (7.1.2)
where C
t
i
s are given constants, B
t
i
s are functions of the unknown variable x.
Throughout this chapter, we shall limit ourselves to those B
t
i
s which are linear
combinations of x and its derivative. So the most general form of the boundary
207
208 CHAPTER 7. BOUNDARY VALUE PROBLEMS
condition Eq. (7.1.2) will be written as
B
1
(x) ≡ α
11
x(a) +α
12
˙ x(a) +β
11
x(b) +β
12
˙ x(b) = C
1
B
2
(x) ≡ α
21
x(a) +α
22
˙ x(a) +β
21
x(b) +β
22
˙ x(b) = C
2
(7.1.3)
where the coefficients α
ij
, β
ij
are given real numbers. To ensure that Eq. (7.1.3)
gives two distinct boundary conditions we assume that the row vectors

11
, α
12
, β
11
, β
12
) and (α
21
, α
22
, β
21
, β
22
) are linearly independent .
If C
1
= C
2
= 0 , we say that boundary conditions are homogeneous,
If β
11
= β
12
= α
21
= α
22
= 0, the boundary conditions are called unmixed.
Eq. (7.1.3) then reduces to
α
11
x(a) +α
12
˙ x(a) = C
1
β
21
x(b) +β
22
˙ x(b) = C
2
(7.1.4)
If α
12
= β
11
= β
12
= α
21
= β
21
= β
22
= 0, we have initial conditions
x(a) = C
1
˙ x(a) = C
2
The boundary conditions are periodic if they are of the form
x(a) = x(b), ˙ x(a) = ˙ x(b) (7.1.5)
Example 7.1.1
d
2
x
dt
2
= f(t)
x(0) = x(1) = 0 (7.1.6)
This denotes the deflection of a string stretched under a unit tension at fixed
end points 0 and 1 and subjected to force distribution f(t). Writing Eq. (7.1.6)
in the form Eq. (7.1.1), we have
Lx = f, B
1
(x) = 0 = B
2
(x)
where L ≡
d
2
dt
2
, B
1
(x) ≡ x(0), B
2
(x) ≡ x(1).
It is clear that we have homogeneous unmixed boundary conditions.
Example 7.1.2 Consider a rod of length unity and unit cross-section. We
assume that the rod is homogeneous and the flow of heat is in the direction of
t only. Suppose there is a heat source of density f(t) in the interior, then the
temperature distribution x(t) satisfies

d
dt

K
dx
dt

= f(t)
7.2. ADJOINT OF A DIFFERENTIAL OPERATOR 209
Here, K is the thermal conductivity of the rod, which is a function of t.
If the end points of the rod are kept at zero temperature we have the boundary
value problem
Lx = f, B
1
(x) = 0 = B
2
(x)
where L ≡ −
d
dt

K
d
dt

, B
1
(x) ≡ x(0), B
2
(x) ≡ x(1).
Here again, we have homogeneous unmixed boundary conditions.
Example 7.1.3 Consider the motion of a particle along a straight line. The
force on the particle is along the line and the particle starts from rest at the
origin. Then the displacement x(t) satisfies Newton’s law
m
d
2
x
dt
2
= f(t)
with initial conditions x(0) = 0,
dx
dt
(0) = 0.
We can write this in our standard form as
Lx = f, B
1
(x) = 0 = B
2
(x)
where L ≡ m
d
2
dt
2
, B
1
(x) ≡ x(0), B
2
(x) ≡ ˙ x(0).
This is an initial-value problem.
7.2 Adjoint of a Differential Operator
As before, we shall denote by L
2
[a, b] the Hillbert space of all square integrable
real valued functions on [a, b].
Definition 7.2.1 Let L : L
2
[a, b] →L
2
[a, b] be a second order differential opera-
tor. By domain D(L), we mean a subspace of L
2
[a, b] which consists of functions
which have piecewise-continuous derivatives of second order and which satisfy
homogeneous boundary conditions
B
1
(x) = 0 = B
2
(x)
Definition 7.2.2 Let L be the differential operator acting on the Hilbert space
L
2
[a, b] with domain D. Then, using Definition 2.2.6 the adjoint operator L

:
L
2
[a, b] →L
2
[a, b] is defined implicitly by the equation
(y, Lx) = (L

y, x); x ∈ D(L) and y ∈ D(L

) (7.2.1)
D(L

) will be obtained in the following steps:
(y, Lx) =

b
a
yLxdt
=

b
a
y(a
0
¨ x +a
1
˙ x +a
2
x) dt
210 CHAPTER 7. BOUNDARY VALUE PROBLEMS
On the RHS, integrating by parts we get
(y, Lx) = [a
0
(y ˙ x −x ˙ y) +xy(a
1
− ˙ a
0
)]
b
a
+

b
a
x
¸
d
2
dt
2
(a
0
y) −
d
dt
(a
1
y) +a
2
y

dt
Applying the definition of L

we see that
(L

y, x) =

b
a
x
¸
d
2
dt
2
(a
0
y) −
d
dt
(a
1
y) +a
2
y

dt
+ [a
0
(y ˙ x −x ˙ y) +xy(a
1
− ˙ a
0
)]
b
a
(7.2.2)
Thus we see that L

consists of two parts, a second order differential operator
d
2
dt
2
(a
0
y) −
d
dt
(a
1
y) +a
2
y (7.2.3)
appearing in the integrand and some boundary terms.
Definition 7.2.3 The operator given by Eq. (7.2.3) is called the formal adjoint
of L and is denoted by L

. So the formal adjoint of the operator
L = a
0
d
2
dt
2
+a
1
d
dt
+a
2
is
L

= a
0
d
2
dt
2
+ (2˙ a
0
−a
1
)
d
dt
+ (¨ a
0
− ˙ a
1
+a
2
)
Definition 7.2.4 The expression J(x, y) = a
0
(y ˙ x −x ˙ y) +(a
1
− ˙ a
0
)xy is called
the bilinear concomitant of x and y. From Eq. (7.2.2) we get the Green’s formula

b
a
(yLx −xL

y)dt = J(x, y)

b
a
Definition 7.2.5 L is said to be formally self-adjoint if L

= L.
It is clear that L is formally self-adjoint if ˙ a
0
(t) = a
1
(t). So, a formally
self-adjoint operator L can be written in the form
L =
d
dt

a
0
(t)
d
dt

+ a
2
(t)
J(x, y) is then given by
a
0
(y ˙ x − ˙ yx)
and the Green’s formula reduces to

b
a
(yLx −xLy) dt = J(x, y) = a
0
(y ˙ x − ˙ yx)

b
a
7.2. ADJOINT OF A DIFFERENTIAL OPERATOR 211
We say that y ∈ D

(domain of L

) if it is twice differentiable and if J(x, y)[
b
a
= 0
for every x ∈ D. From the definition of J(x, y), it can be seen that J(x, y)[
b
a
= 0
for all x ∈ D only if y satisfies two homogeneous boundary conditions B

1
(y) =
B

2
(y) = 0. These are called adjoint boundary conditions.
Definition 7.2.6 Consider the boundary value problem
Lx = f; a ≤ t ≤ b
B
1
(x) = 0 = B
2
(x) (7.2.4)
Its adjoint boundary value problem is
L

x = f
B

1
(x) = 0 = B

2
(x) (7.2.5)
Boundary value problem given by Eq. (7.2.4) is said to be self-adjoint, if
L = L

and D = D

Remark 7.2.1 If we consider L

as an operator on the Hilbert space L
2
[a, b]
with domain D

, one can show that L is self-adjoint as an unbounded operator
if the boundary value problem given by Eq. (7.2.4) is self-adjoint in the sense
of the above definition (Refer Dunford and Schwartz [3]).
Example 7.2.1 L =
d
2
dt
2
with boundary conditions
B
1
(x) ≡ x(0) = 0
B
2
(x) ≡ ˙ x(0) = 0
L is formally self-adjoint and so L

=
d
2
dt
2
.
To determine adjoint boundary conditions, we need to find J(x, y)[
1
0
.
J(x, y) = (y ˙ x − ˙ yx) [
1
0
= ˙ x(1)y(1) − ˙ y(1)x(1) − ˙ x(0)y(0) + ˙ y(0)x(0)
= ˙ x(1)y(1) −x(1) ˙ y(1)
So J(x, y)[
1
0
= 0 for all x ∈ D iff B

1
(y) ≡ y(1) = 0, B

2
(y) ≡ ˙ y(1) = 0.
Thus we see that though the operator is formally self-adjoint. D

= D.
Hence the boundary value problem is not self-adjoint.
Example 7.2.2 Consider the second order differential operator
L = a
0
d
2
dt
2
+a
1
(t)
d
dt
+a
2
(t)
with boundary conditions B
1
(x) ≡ x(0) = 0, B
2
(x) ≡ ˙ x(1) = 0. We know that
L

= a
0
d
2
dt
2
+ (2˙ a
0
−a
1
)
d
dt
+ (¨ a
0
− ˙ a
1
+a
2
)
212 CHAPTER 7. BOUNDARY VALUE PROBLEMS
L will be formally self-adjoint iff ˙ a
0
= a
1
.
To determine the adjoint boundary conditions we first determine J(x, y)[
1
0
.
J(x, y)[
1
0
= a
0
(y ˙ x −x ˙ y) + (a
1
− ˙ a
0
)xy[
1
0
= a
0
(1) [y(1) ˙ x(1) −x(1) ˙ y(1)] + [a
1
(1) − ˙ a
0
(1)] x(1)y(1)
−a
0
(0) [y(0) ˙ x(0) −x(0) ˙ y(0)] + [a
1
(0) − ˙ a
0
(0)]x(0)y(0)
= [a
1
(1) − ˙ a
0
(1)]x(1)y(1) −a
0
(1) ˙ y(1)x(1) −a
0
(0)y(0) ˙ x(0)
J(x, y)[
1
0
= 0 for all x ∈ D iff
B

1
≡ y(0) = 0
B

2
≡ ˙ y(1) −
[a
1
(1) − ˙ a
0
(1)]y(1)
a
0
(1)
= 0
Thus, we see that the adjoint boundary conditions differ from the original one
and D

= D. However, if L is formally delf-adjoint, then D

= D. That is, the
boundary value problem is self-adjoint.
7.3 Green’s Function
Let the boundary value problem be given by
Lx = f
B
1
(x) = 0 = B
2
(x) (7.3.1)
We want to find L
−1
, the inverse of L, and use it to solve the above equation.
As we will see later, this inverse operator is an integral operator. The kernel of
this operator will turn out to be the Green’s function of the system Eq. (7.3.1),
which we now define.
First we define fundamental solution for the operator L.
Definition 7.3.1 Any solution of the equation
LT = δ(t −s) (7.3.2)
is said to be a fundamental solution for the operator L.
Since RHS of Eq. (7.3.2) is a distribution, the solution T of Eq. (7.3.2) must
be interpreted in the sense of a distribution. But we will show that the solution
of Eq. (7.3.2) is a weak solution.
7.3. GREEN’S FUNCTION 213
We prove this assertion by writing Eq. (7.3.2) formally as
LT = 0, t = s
Let x(t) be an arbitrary solution of the homogeneous equation and let y
s
(t) be
any other solution which is to be determined.
We get a particular solution of Eq. (7.3.2) in the following way:
T =

y
s
(t), t > s
x(t), t < s
We will determine y
s
(t) by appropriate matching conditions at t = s. Assume
that T is continuous at t = s, then integrating Eq.(7.3.2) we get
lim
→0

s+
s−
a
0
(t)
d
2
T
dt
2
dt = 1,
where a
0
is the coefficient of
d
2
T
dt
2
occuring on the expression for LT. That is,
a
0
(s)
¸
dT
dt

t=s+

dT
dt

t=s−
¸
= 1
or ˙ y
s
(s) = ˙ x(s) +
1
a
0
(s)
Thus a solution T of (7.3.2) is given by
T(t, s) =

y
s
(t), t > s
x(t), t < s
(7.3.3)
where x(t) is an arbitrary solution of the homogeneous equation and y
s
(t) is
another solution of the homogeneous equation satisfying
y
s
(s) = x(s)
˙ y
s
(s) = ˙ x(s) +
1
a
0
(s)
By the theory of differential equaitons, such a solution y
s
(t) exists and is deter-
mined by x(t) and a
0
(t).
We now show that T(t, s) defined by Eq. (7.3.3) is a weak solution of Eq.
(7.3.2).
That is (T, L

φ) = (δ, φ) = φ(s) for every test function φ.
By definition
L

φ =
2
¸
k=0
(−1)
k
d
k
dt
k
(a
n−k
φ)
214 CHAPTER 7. BOUNDARY VALUE PROBLEMS
(T, L

φ) =


−∞
T(t)L

φ(t)dt
Integrating by parts twice and using boundary conditions, we get
(T, L

φ) = φ(s) +

s−
−∞
φLxdt +


s+
φLy
s
dt
Since both x and y
s
satisfy the homogeneous equation LT = 0, we get
(T, L

φ) = φ(s) = (δ(t −s), φ)
That is
LT = δ(t −s)
We are now in a position to define the Green’s function of Eq. (7.3.1).
Definition 7.3.2 The Green’s function g(t, s) of Eq.(7.3.1) is a fundamental
solution of
Lg = δ(t −s)
which satisfies additional boundary condition B
1
(g) = 0 = B
2
(g). In view of
what we have just proved, alternatively we can define the Green’s function g(t, s)
as a solution of the boundary value problem
Lg = 0, a ≤ t < s, s < t ≤ b
B
1
(g) = 0 = B
2
(g) (7.3.4)
g is continuous at t = s and
dg
dt

t=s+

dg
dt

t=s−
=
1
a
0
(s)
.
Example 7.3.1 Consider the string problem
d
2
x
dt
2
= f(t)
x(0) = x(1) = 0
Green’s function g(t, s) for the above sustem is the solution of the auxiliary
problem
d
2
g
dt
2
= δ(t −s)
g(0, s) = 0 = g(1, s)
That is
d
2
g
dt
2
= 0, 0 ≤ t < s, s < t ≤ 1
g(0, s) = 0 = g(1, s)
7.3. GREEN’S FUNCTION 215
g(t, s) is continuous at t = s,
dg
dt

t=s+

dg
dt

t=s−
= 1.
Integrating the above equation and imposing boundary conditions at 0 and 1,
we get
g(t, s) =

At, 0 ≤ t < s
C(t −1), s < t ≤ 1
The continuity of g(t, s) at t = s gives
g(t, s) =

At, 0 ≤ t ≤ s
A(1 −t)s
(1 −s)
, s ≤ t ≤ 1
Finally, the jump condition on the derivatives
dg
dt
at t = s gives
As
s −1
−A = 1, A = s −1
So
g(t, s) =

t(s −1), 0 ≤ t ≤ s
s(t −1), s ≤ t ≤ 1
Example 7.3.2 Consider the initial-value problem representing motion of a
particle in a straight line
m
d
2
x
dt
2
= f(t), 0 ≤ t < ∞
x(0) = ˙ x(0) = 0
Green’s function for the above system is given by
m
d
2
g
dt
2
= 0, 0 ≤ t < s, s < t < ∞
g(0, s) = 0 = ˙ g(0, s)
g is continuous at t = s and
dg
dt

t=s+

dg
dt

=s−
=
1
m
.
Integrating and using the boundary conditions we get
g(t, s) =

0, t ≤ s
At +B, s ≤ t
216 CHAPTER 7. BOUNDARY VALUE PROBLEMS
Using continuity of g at t = s we get B = −As.
So
g(t, s) =

0, t ≤ s
A(t −s), s ≤ t
Finally the jump condition for the derivative gives A =
1
m
.
Hence
g(t, s) =

0, t ≤ s
(t −s)
m
, s ≤ t
Definition 7.3.3 The adjoint Green’s function h(t, s) is defined as the solution
of the system
L

h = δ(t −s)
B

1
(h) = 0 = B

2
(h) (7.3.5)
Theorem 7.3.1 Theadjoint Green’s function h(t, s) and the Green’s function
g(t, s) satisfy the relation h(t, s) = g(s, t). In particular, if the original boundary
value problem given by Eq. (7.3.1) is self-adjoint, then the Green’s function
g(t, s) is symmetric.
Proof : We have
Lg = δ(t −s) (7.3.6)
B
1
(g) = 0 = B
2
(g)
and
L

h = δ(t −η) (7.3.7)
B

1
(h) = 0 = B

2
(h)
Multiply Eq. (7.3.6) by h(t, η), Eq. (7.3.7) by g(t, s), subtract and integrate
from t = a to t = b. Then we get

b
a
(hLg −gL

h)dt =

b
a
h(t, η)δ(t −s)dt −

b
a
g(t, s)δ(t −η)dt
Using Green’s formula

b
a
(hLg −gL

h) = J(g, h)

b
a
and substituting this in the
left hand side of the above equation we get
J(g, h)[
b
a
= h(s, η) −g(η, s)
Since h satisfies the adjoint boundary condition J(g, h)[
b
a
= 0 and hence
h(s, η) = g(η, s) for all s, η. If the system is self-adjoint, then we know that
h = g and hence
g(s, η) = g(η, s) for all s, η
7.3. GREEN’S FUNCTION 217
That is, g is symmetric.
Now we shall show how Green’s function can be used to obtain the solution
of the boundary value problem by constructing the inverse of the corresponding
differential operator.
Theorem 7.3.2 Suppose g(t, s) is a Green’s function for the boundary value
problem Lx = f, a ≤ t ≤ b, B
1
(x) = 0 = B
2
(x). Then it has a solution given
by
x(t) =

b
a
g(t, s)f(s)ds (7.3.8)
We shall use the following lemma which is stated without proof.
Lemma 7.3.1 (Leibnitz) Let f(t, η) be a continuous function of two variables
in a region α(t) ≤ η ≤ β(t), c ≤ t ≤ d, where α(t) and β(t) are continuous
functions. Then the integral
g(t) =

β(t)
α(t)
f(t, η)dη; c ≤ t ≤ d (7.3.9)
represents a continuous function of t. If the derivatives ˙ α(t) and
˙
β(t) exist, and
∂f
∂t
as well as f are continuous, then the derivative of the integral in Eq. (7.3.9)
is given by the generalized Leibnitz formula
˙ g(t) =

β(t)
α(t)

∂t
f(t, η)dη +f(t, β(t))
˙
β(t) −f(t, α(t)) ˙ α(t)
Proof (Proof of the main theorem): Let g(t, s) be the Green’s function for
the boundary value problem given by Eq. (7.3.1). We can write Eq. (7.3.8) as
x(t) =

b
a
g(t, s)f(s)ds =

t−
a
g(t, s)f(s)ds +

b
t+
g(t, s)f(s)ds
Then, using the above mentioned lemma, we have
˙ x(t) =

t−
a
g
t
(t, s)f(s)ds +

b
t+
g
t
(t, s)f(s)ds
+ [g(t, t−)f(t) −g(t, t+)f(t)]
Due to the continuity of g(t, s), the last bracketed expression vanishes. Also
¨ x(t) =

t−
a
g
tt
(t, s)f(s)ds +

b
t+
g
tt
(t, s)f(s)ds
+ g
t
(t, t−)f(t) −g
t
(t, t+)f(t)
We note that g
t
(t, t+) = g
t
(t−, t) and g
t
(t, t−) = g
t
(t+, t) due to the continuity
of g(t, s) in the triangular regions ABC and ABD respectively (Refer Fig. 7.2.1).
218 CHAPTER 7. BOUNDARY VALUE PROBLEMS
A
B
D
C
S
t
t=s
Figure 7.3.1: Integration in a triangular region
So, substituting the expressioins for ¨ x(t) and ˙ x(t) in Lx ≡ a
0
(t)¨ x +a
1
(t) ˙ x +
a
2
(t)x, we have
Lx =

t−
a
[a
0
(t)g
tt
(t, s) +a
1
(t)g
t
(t, s) +a
2
(t)g(t, s)] ds
+

b
t+
[a
0
(t)g
tt
(t, s) +a
1
(t)g
t
(t, s) +a
2
(t)g(t, s)] ds
+ f(t) [g
t
(t, t−) −g
t
(t, t+)] a
0
(t)
= 0 + 0 +f(t)
¸
1
a
0
(t)
a
0
(t)

= f(t)
Hence x(t) =

b
a
g(t, s)f(s)ds satisfies the given differential equation. In exactly
the same manner one can show that x given by Eq. (7.3.8) satisfies B
1
(x) =
0 = B
2
(x). Thus, x(t) given by Eq. (7.3.8) is the solution of the boundary value
problem given by Eq. (7.3.1).
Theorem 7.3.3 The operator L and the integral operator G defined as
Gf =

b
a
g(t, s)f(s)ds
are inverses of each other.
7.3. GREEN’S FUNCTION 219
Proof : It is clear from the definition of L and G that LGf = f and GLx = x.
Hence the result.
Green’s function also enables us to get a solution of the boundary value
problem with non-homogeneous boundary conditions:
Lx = f
B
1
(x) = α, B
2
(x) = β (7.3.10)
Theorem 7.3.4 If g(t, s) is a Green’s function of Eq. (7.3.10), then it has a
solution x(t) given by
x(t) =

b
a
g(t, s)f(s)ds −J(x(s), g(t, s))

b
a
Proof : We recall that the adjoint Green’s function h(t, s) satisfies the system
L

h = δ(t −s) (7.3.11)
B

1
(h) = 0 = B

2
(h)
Multiply Eq. (7.3.10) by h(t, s), Eq. (7.3.11) by x(t) , subtract and integrate
from t = a to t = b. to get

b
a
(hLx −xL

h)dt =

b
a
f(t)h(t, s)dt −

b
a
δ(t −s)x(t)dt
Using Green’s formula in the LHS, we have
J(x, h)[
b
a
=

b
a
h(t, s)f(t)dt −x(s)
That is
x(s) =

b
a
h(t, s)f(t)dt −J(x, h)

b
a
Using the fact that h(t, s) = g(s, t) we obtain
x(s) =

b
a
g(s, t)f(t)dt −J(x(t), g(s, t))

b
a
or
x(t) =

b
a
g(t, s)f(s)ds −J(x(s), g(t, s))

b
a
J(x(s), g(t, s))[
b
a
can be calculated explicitly by the given boundary conditions.
An alternative method of solution of Eq. (7.3.10) is as follows.
220 CHAPTER 7. BOUNDARY VALUE PROBLEMS
Let x
1
(t) and x
2
(t) be non-trivial solutions of the homogeneous system, cor-
responding to Eq. (7.3.10) satisfying B
1
(x
1
) = 0 and B
2
(x
2
) = 0, respectively.
Since a completely homogeneous system has only the trivial solution, we must
have B
1
(x
2
) = 0. We write the solution of the system as
x(t) =

b
a
g(t, s)f(s)ds +c
1
x
1
(t) +c
2
x
2
(t)
g(t, s) is the Green’s function for the system with homogeneous boundary con-
ditions and c
1
and c
2
are constants to be determined. Applying boundary
conditions, we have
c
2
B
1
(x
2
) +c
1
B
1
(x
1
) +

b
a
B
1
(g(t, s))f(s)ds = α
c
1
B
2
(x
1
) +c
2
B
2
(x
2
) +

b
a
B
2
(g(t, s))f(s)ds = β
Therefore,
x(t) =

b
a
g(t, s)f(s)ds +
β
B
2
(x
1
)
x
1
(t) +
α
B
1
(x
2
)
x
2
(t)
It is clear that this solution depends continuously on f, α and β and is unique.
Example 7.3.3 Let us compute the solution of the non-homogeneous boundary
value problem
d
2
x
dt
2
= f(t)
x(0) = α, x(1) = β
As in Example 7.3.1, its Green’s function g(t, s) is given by
g(t, s) =

t(s −1), t ≤ s
s(t −1), t ≥ s
It is a symmetric function and satisfies the boundary conditions g(t, s)[
t=0
=
0 = g(t, s)[
t=1
.
From the above
˙ g(t, s) =

s −1, t ≤ s
s, t ≥ s
˙ g(t, s)[
t=0
= (s −1), ˙ g(t, s)[
t=1
= s
7.4. CONSTRUCTIONOF GREEN’S FUNCTIONAND MODIFIEDGREEN’S FUNCTION221
Using the formula for J(x, g) we get
J(x, g)[
1
0
= [ ˙ xg −x˙ g]
1
0
= [ ˙ x(1)g(1) −x(1) ˙ g(1)] −[ ˙ x(0)g(0) −x(0) ˙ g(0)]
= x(0) ˙ g(0) −x(1) ˙ g(1)
= α˙ g(0) −β ˙ g(1)
= α(s −1) −βs
Hence, by Theorem 7.3.4, the solution x(s) of the above boundary value problem
is given by
x(s) =

1
0
g(t, s)f(t)dt −J(x, g)[
1
0
=

1
0
g(t, s)f(t)dt −α(s −1) +βs
=

s
0
t(s −1)f(t)dt +

1
s
s(t −1)f(t)dt +α(1 −s) + βs
or
x(t) =

t
0
s(t −1)f(s)ds +

1
t
t(s −1)f(s)ds +α(1 −t) +βt
7.4 Construction of Green’s Function and Mod-
ified Green’s Function
We shall consider the boundary value problem
Lx = f
B
1
(x) = α, B
2
(x) = β
We prove the existence of Green’s function of the above system where the bound-
ary conditions are of unmixed and initial type only, by construction.
Theorem 7.4.1 For the restricted case of unmixed and initial boundary condi-
tions of the completely homogeneous system
Lx = 0, B
1
(x) = 0 = B
2
(x) (7.4.1)
which has only the trivial solution, the Green’s function for the boundary value
problem exists and is unique.
Proof : Uniqueness of the solution is immediate. For, if there are two Green’s
functions g
1
and g
2
then (g
1
−g
2
) is a non-trivial solution of the completely ho-
mogeneous equation Eq. (7.4.1) which is a contradiction to the hypothesis. We
need to prove the existence in two parts, first for unmixed boundary conditions
and then for initial conditions.
222 CHAPTER 7. BOUNDARY VALUE PROBLEMS
Case 1 Unmixed boundary conditions are given by
B
1
(x) = α
11
x(a) +α
21
˙ x(a) = 0
B
2
(x) = β
21
x(b) +β
22
˙ x(b) = 0
Let x
1
(t) be a non-trivial solution of Lx = 0 satisfying boundary condition
at t = a and x
2
(t) be a non-trivial solution satisfying boundary condition
at t = b. Such a solution exists, for one can take x
1
(t) to be the unique
solution of Lx = 0, x(a) = α
12
, ˙ x(a) = −α
11
. This follows from the theory
of differential equation (Refer Coddington and Levinson[1].) Similarly, we
have the existence of x
2
(t). x
1
and x
2
are linearly independent, otherwise
the completely homogeneous system will have a non-trivial solution. The
Green’s function is then of the form

Ax
1
(t), a ≤ t < s
Bx
2
(t), s < t ≤ b
Continuity and jump conditions of g(t, s) give
Ax
1
(s) −Bx
2
(s) = 0
−A˙ x
1
(s) −B˙ x
2
(s) =
1
a
0
(s)
Solving the above of equations for A and B we get
A =
x
2
(s)
a
0
(s)W(x
1
, x
2
, s)
B =
x
1
(s)
a
0
(s)W(x
1
, x
2
, s)
where W(x
1
, x
2
, s) is the Wronskian of x
1
and x
2
and is non-zero since x
1
and x
2
are linearly independent.
So
g(t, s) =

x
1
(t)x
2
(s)
a
0
(s)W(x
1
, x
2
, s)
, a ≤ t < s
x
1
(s)x
2
(t)
a
0
(s)W(x
1
, x
2
, s)
, s < t ≤ b
This proves the existence of Green’s function for the boundary value prob-
lem.
Case 2 Boundary conditions are of initial type
x(a) = 0 = ˙ x(a)
7.4. CONSTRUCTIONOF GREEN’S FUNCTIONAND MODIFIEDGREEN’S FUNCTION223
Green’s function is of the form
g(t, s) =

0, a ≤ t < s
Ax
1
(t) +Bx
2
(t), t > s
where x
1
, x
2
are two linearly independent solutions of Lx = 0.
Continuity and jump conditions on g(t, s) give
Ax
1
(s) +Bx
2
(s) = 0
A˙ x
1
(s) +B ˙ x
2
(s) =
1
a
0
(s)
Solving for A and B, we get
A =
x
2
(s)
a
0
(s)W(x
1
, x
2
, s)
B =
x
1
(s)
a
0
(s)W(x
1
, x
2
, s)
This gives
g(t, s) =

0, a ≤ t ≤ s
x
2
(t)x
1
(s) −x
1
(t)x
2
s
a
0
(s)W(x
1
, x
2
, s)
, t ≥ s
Example 7.4.1 Let us compute the Green’s function for the string problem
d
2
x
dt
2
= f(t)
x(0) = 0 = x(1)
We take x
1
(t) = t x
2
(t) = 1 −t. Then
W(x
1
, x
2
, s) =

s 1 −s
1 −1

= −1
=

x
1
(s) x
2
(s)
˙ x
1
(s) ˙ x
2
(s)

A =
x
2
(s)
a
0
(s)W(x
1
, x
2
, s)
=
1 −s
−1
B =
x
1
(s)
a
0
(s)W(x
1
, x
2
, s)
=
s
−1
224 CHAPTER 7. BOUNDARY VALUE PROBLEMS
So
g(t, s) =

t(s −1), t ≤ s
s(t −1), t ≥ s
Example 7.4.2 Consider the initial-value problem describing the motion of a
particle in a straight line
m
d
2
x
dt
2
= f(t)
x(0) = 0 = ˙ x(0)
We take x
1
(t) = 1, x
2
(t) = t as two linearly independent solutions of the
homogeneous differential equation
d
2
x
dt
2
= 0.
Then
W(x
1
, x
2
, s) =

1 s
0 1

= 1
Hence
g(t, s) =

0, t ≤ s
(t −s)
m
, t ≥ s
Example 7.4.3 Consider the problem of heat distribution in a rod of unit length
and unit cross-section and with thermal conductivity k(t). If the heat source has
density f(t), then the governing equation is

d
dt

k
dx
dt

= f(t)
Assume that the ends of the rod are kept at 0

temperature
x(0) = 0 = x(1)
We look for two linearly independent solutions x
1
(t) and x
2
(t) of the homoge-
neous equation

d
dt

k
dx
dt

= 0
which satisfy the boundary conditions at 0 and 1 respectively. So we write
x
1
(t) =

t
0
ds
k(s)
, x
2
(t) =

t
1
dx
k(s)
W(x
1
, x
2
, s) =

x
1
(s) x
2
(s)
1
k(s)

1
k(s)

7.4. CONSTRUCTIONOF GREEN’S FUNCTIONAND MODIFIEDGREEN’S FUNCTION225
= −
x
1
(s) +x
2
(s)
k(s)
= −

1
0
k
−1
(s)ds
k(s)
Hence
g(t, s) =

x
1
(t)x
2
(s)
k(s)
¸
k(s)

1
0
k
−1
(t)dt
¸
, t ≤ s
x
1
(s)x
2
(t)
k(s)
¸
k(s)

1
0
k
−1
(t)dt
¸
, t ≥ s
=
¸
1
0
k
−1
(t)dt
−1

x
1
(t)x
2
(s), t ≤ s
x
2
(t)x
1
(s), t ≥ s
Example 7.4.4 In the study of heat flow in a cylindrical shell, one comes across
the Bessel’s equation of the type
d
dt

t
dg
dt

+tg = δ(t −s); a ≤ t ≤ b
with boundary conditions
g = 0 at t = a and t = b
As we know, the homogeneous Bessel’s equation
d
dt

t
dg
dt

+tg = 0
has two linearly independent solutions J
0
(t) and Y
0
(t)
J
0
(t) =

¸
n=0
(−1)
n
t
2n
2
2n
(n!)
2
and
Y
0
(t) =
2
π
¸
γ +
lnt
2

J
0
(t) +
2
π
(t/2)
2
(1!)
2

2
π
(t/2)
4
(2!)
2
(1 + 1/2)
+
2
π
(t/2)
6
(3!)
2
(1 + 1/2 + 1/3) −
where γ

= 0.5772 is the so called Euler constant which is the limit of 1 +
1
2
+
+
1
5
− lns as s approaches infinity. But J
0
(t) and Y
0
(t) fail to satisfy the
boundary conditions at a and b. So we define two other linearly independent
solutions
Z
1
0
(t) = J
0
(t)Y
0
(a) −Y
0
(t)J
0
(a)
Z
2
0
(t) = J
0
(t)Y
0
(b) −Y
0
(t)J
0
(b)
226 CHAPTER 7. BOUNDARY VALUE PROBLEMS
Obviously now Z
1
0
(a) = 0 = Z
2
0
(b).
It can be shown that
W(Z
1
0
, Z
2
0
; t) = W[J
0
, Y
0
; t][J
0
(a)Y
0
(b) −Y
0
(a)J
0
(b)]
=
2
πt
[J
0
(a)Y
0
(b) −Y
0
(a)J
0
(b)]
So
g(t, s) =

Z
1
0
(t)Z
2
0
(s)
C
, t ≤ s
Z
1
0
(s)Z
2
0
(t)
C
, t ≥ s
where C =
2
π
[J
0
(a)Y
0
(b) −Y
0
(a)J
0
(b).
Example 7.4.5 Consider the boundary value problem

d
2
x
dt
2
−λx = f(t)
x(0) = 0 = x(1)
To find the Green’s function for this problem we first look for two linearly inde-
pendent solutions of the completely homogeneous problem −
d
2
x
dt
2
−λx = 0, which
satisfy the boundary condition at 0 and 1, respectively.
x
1
(t) = sin

λt, x
2
(t) = sin

λ(1 −t)
W(x
1
, x
2
; t) =

sin

λt sin

λ(1 −t)

λcos

λt −

λ cos

λ(1 −t)

=

λsin

λ(1 −t) cos

λt
+sin

λt cos

λ(1 −t)
=

λsin

λ(1 −t +t)
=

λsin

λ
So
g(t, s) =

sin

λt sin

λ(1 −s)

λsin

λ
, t ≤ s
sin

λ(1 −t) sin

λs

λsin

λ
, t ≥ s
7.4. CONSTRUCTIONOF GREEN’S FUNCTIONAND MODIFIEDGREEN’S FUNCTION227
In general, where the boundary conditions are mixed, the Green’s function may
be found in a straignt forward way.
Let x
1
(t) and x
2
(t) be any two linearly independent solutions of Lx = 0.
Then write
g(t, s) = Ax
1
(t) +Bx
2
(t) +
x
1
(t)x
2
(s)H(s −t) +x
2
(t)x
1
(s)H(t −s)
W(x
1
, x
2
; s)a
0
(s)
where A and B are constants which will be determined so that g(t, s) satisfies
the given boundary conditions.
It is clear that g(t, s) is continuous and its derivative has a jump of magni-
tude
1
a
0
(s)
. Now consider the boundary conditions. Since they are linear and
homogeneous, we have
B
1
(g) ≡ AB
1
(x
1
) +BB
1
(x
2
) +B
1
(r) = 0 (7.4.2)
B
1
(g) ≡ AB
1
(x
1
) +BB
1
(x
2
) +B
1
(r) = 0 (7.4.3)
where r(x) stands for
x
1
(t)x
2
(s)H(s −t) +x
2
(t)x
1
(s)H(t −s)
W(x
1
, x
2
; s)a
0
(s)
The boundary conditions give us two linear equations which will determine
the values of A and B. These equations will have solutions if the determinant

B
1
(x
1
) B
1
(x
2
)
B
2
(x
1
) B
2
(x
2
)
¸
¸
does not vanish.
If the determinant vanishes, either B
1
(x
1
) = B
2
(x
1
) = 0, or B
1
(x
2
) =
B
2
(x
2
) = 0 which implies that either x
1
(t) or x
2
(t) is a non-trivial solution of the
completely homogeneous system which is a contradiction. The other possiblility
for the determinant to vanish is that there exists a constant c such that
B
1
(x
2
) +cB
1
(x
1
) = 0
B
2
(x
2
) +cB
2
(x
1
) = 0
These equations imply that B
1
(x
2
+ cx
1
) = B
2
(x
2
+ cx
1
) = 0, which is not
possible as discussed earlier. So by solving Eq. (7.4.2)- Eq. (7.4.3) for A and
B, we get the desired Green’s function.
So far, we have proved the existence of Green’s function for the system
Lx = f, B
1
(x) = 0 = B
2
(x)
under the assumption that the completely homogeneous system Lx = 0, B
1
(x)
= 0 = B
2
(x) has only the trivial solution. What happens when the com-
pletely homogeneous system has a non-trivial solution ?
228 CHAPTER 7. BOUNDARY VALUE PROBLEMS
We state the following Alternative theorem which answers the question raised
above
Consider the related systems:
(i) The compeletely homogeneous system
Lx = 0; a ≤ t ≤ b; B
1
(x) = 0 = B
2
(x)
(ii) The inhomogeneous system
Ly = f; a ≤ t ≤ b; B
1
(y) = 0 = B
2
(y)
(iii) The adjoint homogeneous system
L

z = 0; a ≤ t ≤ b; B

1
(z) = 0 = B

2
(z)
Theorem 7.4.2 (a) If (i) has only the trivial solution x(t) = 0 in a ≤ t ≤ b,
then (iii) has also the trivial solution and (ii) has one and only one solution.
(b) If (i) has a non-trivial solution, (iii) will also have non-trivial solution
and (ii) has solution iff

b
a
f(t)z(t) = 0 for every z which is a solution of (iii).
For a proof of this, refer Stakgold [5].
Example 7.4.6 Consider the self-adjoint system
d
2
g
dt
2
= δ(t −s) (7.4.4)
g(0) = 0, ˙ g(1) = g(1)
The completely homogeneous system
d
2
g
dt
2
= 0, g(0) = 0, ˙ g(1) = g(1)
has a non-trivial solution g = t.
So it follows by alternative theorem that Eq. (7.4.4) will have a solution iff

1
0
δ(t −s)dt = 0. Since

1
0
δ(t −s)dt = 0, Eq. (7.4.4) will have no solution.
That is, the ordinary Green’s function does not exist.
Example 7.4.7 Consider the self-adjoint system
d
2
g
dt
2
= δ(t −s), 0 ≤ t ≤ 1 (7.4.5)
g(0) = 0 = ˙ g(1)
7.4. CONSTRUCTIONOF GREEN’S FUNCTIONAND MODIFIEDGREEN’S FUNCTION229
The completely homogeneous system
d
2
g
dt
2
= 0, ˙ g(0) = 0 = ˙ g(1)
has a non-trivial solution g = const. Hence Eq. (7.4.5) has no solution, for

1
0
δ(t −s)ds = 0.
That is, the ordinary Green’s function can not be constructed for Eq. (7.4.5).
Modified Green’s Function
Assume that the self-adjoint system
Lx = f; a ≤ t ≤ b; B
1
(x) = 0 = B
2
(x)
is such that the completely homogeneous system
Lx = 0; a ≤ t ≤ b; B
1
(x) = 0 = B
2
(x)
has non-trivial solution of the form x
1
(t), where x
1
(t) is normalized

b
a
x
2
1
(t)dt = 1

.
It now follows that the Green’s function which is a solution of the system
Lg = δ(t −s); B
1
(g) = 0 = B
2
(g)
will not exist for

b
a
x
1
(t)δ(t −s)dt = 0. However, in order to compute the
Green’s function one has to resort to modified Green’s function by the addition
of new source density of strength x
1
(s).
Definition 7.4.1 Modified Green’s function is the solution of the system
Lg
M
(t, s) = δ(t −s) −x
1
(t)x
1
(s)
B
1
(g
M
) = 0 = B
2
(g
M
)
It is clear that the solution of the above system exists for

b
a
[δ(t −s) −x
1
(t)x
1
(s)] x
1
(t)dt = 0
The method of construction of g
M
is now similar to that of ordinary Green’s
function which we have already discussed.
Theorem 7.4.3 The modified Green’s function g
M
(t, s) is symmetric if

b
a
g
M
(t, s)x
1
(t)dt = 0 for every s
230 CHAPTER 7. BOUNDARY VALUE PROBLEMS
Proof : Consider g
M
(t, s
1
) and g
M
(t, s
2
) wich satisfy the systems:
Lg
M
(t, s
1
) = δ(t −s
1
) −x
1
(t)x
1
(s
1
) (7.4.6)
B
1
(g
M
(t, s
1
)) = 0 = B
2
(g
M
(t, s
1
))
and
Lg
M
(t, s
2
) = δ(t −s
2
) −x
1
(t)x
1
(s
2
) (7.4.7)
B
1
(g
M
(t, s
2
)) = 0 = B
2
(g
M
(t, s
2
))
Multiply Eq. (7.4.6) by g
M
(t, s
2
) and Eq. (7.4.7) by g
M
(t, s
1
) and subtract and
integrate from a to b. Then we get

b
a
[g
M
(t, s
2
)(Lg
M
(t, s
1
)) −g
M
(t, s
1
)(Lg
M
(t, s
2
))] dt
=

b
a
g
M
(t, s
2
)δ(t −s
1
)dt −

b
a
x
1
(t)g
M
(t, s
2
)dt

x
1
(s
1
)

b
a
g
M
(t, s
1
)δ(t −s
2
)dt +

b
a
x
1
(t)g
M
(t, s
1
)dt

x
1
(s
2
)
Using Green’s formula and boundary conditions for g
M
(t, s
1
), g
M
(t, s
2
), we have
0 = g
M
(s
1
, s
2
) −g
M
(s
2
, s
1
) −x
1
(s
1
)

b
a
x
1
(t)g
M
(t, s
2
)dt

+x
1
(s
2
)

b
a
x
1
(t)g
M
(t, s
1
)dt

But

b
a
x
1
(t)g
M
(t, s)dt = 0 for every s. Hence
g
M
(s
1
, s
2
) = g
M
(s
2
, s
1
)
On imposing the conditions of symmetry on g
M
(t, s) we can alternatively
define the modified Green’s function as the solution of the system:
Lg
M
(t, s) = δ(t −s) −x
1
(t)x
1
(s); a ≤ t, s ≤ b
satisfying
(i) B
1
(g
M
(t, s)) = 0 = B
2
(g
M
(t, s))
(ii) g
M
(t, s) is continuous at t = s
7.4. CONSTRUCTIONOF GREEN’S FUNCTIONAND MODIFIEDGREEN’S FUNCTION231
(iii)
dg
M
dt
[
t=s+

dg
M
dt
[
t=s−
=
1
a
0
(s)
and
(iv)

b
a
x
1
(t)g
M
(t, s)dt = 0 for every s
For computational purposes, the above defining property of g
M
(t; s) will be used.
We finally have the following theorem which gives a solution of the self-adjoint
system by using the modified Green’s function.
Theorem 7.4.4 Let g
M
(t, s) be the modified Green’s function of the self-adjoint
system
Lx = f; a ≤ t ≤ b; B
1
(x) = 0 = B
2
(x) (7.4.8)
having x
1
(t) as the only normalized solution of the completely homogeneous
system.
If

b
a
f(t)x
1
(t)dt = 0, then Eq. (7.4.8) has a solution x(t) given by
x(t) =

b
a
g
M
(t, s)f(s)ds +Cx
1
(t).
Proof : The existence of a solution of Eq. (7.4.8) is clear by alternative
theorem, for

b
a
f(t)x
1
(t)dt = 0.
We have
Lx = f; B
1
(x) = 0 = B
2
(x)
and
Lg
M
= δ(t −s) −x
1
(t)x
1
(s)
B
1
(g
M
) = 0 = B
2
(g
M
)
By our standard procedure we get

b
a
(g
M
Lx −xLg
M
) =

b
a
g
M
(t, s)f(t)dt −

b
a
x(t)δ(t − s)dt
+x
1
(s)

b
a
x
1
(t)x(t)dt
Using Green’s formula, it reduces to
J(x, g
M
) =

b
a
g
M
(t, s)f(t)dt −x(s) +Cx
1
(s)
The LHS vanishes, fo both x and g
M
satisfy zero boundary conditions and there
fore we have
x(s) =

b
a
g
M
(t, s)f(t)dt +Cx
1
(s)
232 CHAPTER 7. BOUNDARY VALUE PROBLEMS
or
x(t) =

b
a
g
M
(t, s)f(s)ds +Cx
1
(t)
Example 7.4.8 Consider the system
d
2
x
dt
2
= sin 2πt; 0 ≤ t ≤ 1
(7.4.9)
˙ x(0) = 0 = ˙ x(1)
The completely homogeneous system has a non-trivial solution x = const.
Since

1
0
sin 2πtdt = 0, by Theorem 7.4.4, the above system has a solution x(t)
given by
x(t) =

b
a
g
M
(t, s) sin 2πsds +C (7.4.10)
To compute g
M
(t, s) we solve the system
d
2
g
M
dt
2
= δ(t −s) −1 (7.4.11)
˙ g
M
(0) = 0 = ˙ g
M
(1)
g
M
(t, s) is continuous at t = s
dg
M
dt
[
t=s+

dg
M
dt
[
t=s−
= 1

1
0
g
M
(t, s)dt = 0 for every s
Integrating Eq.(7.4.11) and using boundary conditions, we get
g
M
(t, s) =

A−
t
2
2
, t < s
C +t −
t
2
2
, t > s
Continuity at t = s implies that
A−
s
2
2
= C +s −
s
2
2
or
C = A−s
7.5. EIGENFUNCTION EXPANSION - STURM-LIOUVILLE SYSTEM 233
So
g
M
(t, s) =

A−
t
2
2
, t ≤ s
A−s +t −
t
2
2
, t ≥ s
The last defining condition gives

s
0
(A−
t
2
2
)dt +

1
s
(A−s +t −
t
2
2
)dt = 0
That is, As −
s
6
+

A−s +
1
2

1
6

(A−s)s +
s
2
2

s
3
3

= 0, which gives
A =
¸
s −
s
2
2

1
3

Hence, we have the following symmetric modified Green’s function
g
M
(t, s) =

s −
s
2
2

1
3


t
2
2
, t ≤ s

s −
s
2
2

1
3

−s +t −
t
2
2
, t ≥ s
=

s −
1
3

s
2
+t
2
2
, t ≤ s
t −
1
3

s
2
+t
2
2
, t ≥ s
7.5 Eigenfunction Expansion - Sturm-Liouville
System
Consider the self-adjoint system
Lx −λx = f, 0 ≤ t ≤ b
(7.5.1)
B
1
(x) = 0 = B
2
(x)
where a and b are finite constants and f ∈ L
2
[a, b]. Such a self-adjoint system
is called Sturm-Liouville system.
We assume that the operator L acts on L
2
[a, b] with domain D define as
D = ¦x ∈ L
2
[a, b] : ˙ x, ¨ x ∈ L
2
[a, b], B
1
(x) = 0 = B
2
(x)¦
We consider the related eigenvalue problem
Lφ = λφ, B
1
φ = 0 = B
2
φ (7.5.2)
with the assumption that λ = 0 is not an eigenvalue of L. We get the following
important theorem, giving eigenfunctions of L as a basis for L
2
[a, b].
234 CHAPTER 7. BOUNDARY VALUE PROBLEMS
Theorem 7.5.1 The self-adjoint system Eq.(7.5.2) has discrete eigenvalues
¦λ
1
, λ
2
, , λ
n
, ¦ and the corresponding eigenfunctions ¦φ
1
, φ
2
, , φ
n
, ¦
form a compelete orthonormal set in L
2
[a, b].
Proof : As zero is not the eigenvalue of L, there exists Green’s function g such
that the solvability of Eq. (7.5.2) is equivalent to the solvability of the integral
equation
x(t) = λ

b
a
g(t, s)x(s)ds (7.5.3)
in the space L
2
[a, b]. Let K be the integral operator generated by g(t, s)
[Kx](t) =

b
a
g(t, s)x(s)ds
Then Eq. (7.5.3) is equivalent to the following operator equation
[Kx] =
1
λ
x = µx, µ =
1
λ
(7.5.4)
in the space L
2
[a, b].
We note that µ = 0 is not an eigenvalue of K. For, if x = 0 exists satisfying

b
a
g(t, s)x(s)ds = 0, then the boundary value problem
Lw = x, B
1
(w) = 0 = B
2
(w)
has a unique solution w =

b
a
g(t, s)x(s)ds = 0. This implies that x = Lw = 0,
a contradiction.
We observe that

b
a

b
a
g
2
(t, s)dtds < ∞ and hence the operator K generated
by g(t, s) is a compact operator on L
2
[a, b]. K is also self-adjoint as the system
Eq. (7.5.2) is self-adjoint. So, it follows from Theorem 2.2.6 that eigenvalues of
K are discrete and the eigenfunctions of K form a basis for L
2
[a, b]. As K and
L are inverses of each other, it follows that eigenvalues of L are discrete and the
corresponding eigenfunctions ¦φ
1
, φ
2
, , φ
n
, ¦ form a basis for L
2
[a, b].
Theorem 7.5.2 If λ is not any one of the eigenvalues ¦λ
1
, λ
2
, ¦ of L, then
the self-adjoint system Eq. (7.5.1) has one and only one solution x(t) ∈ L
2
[a, b]
given by
x(t) =

¸
n=1
f
n
φ
n
(t)
λ
n
−λ
where f
n
is given by f
n
= (f, φ
n
). Here (., .) is the innerproduct in L
2
(a, b].
7.5. EIGENFUNCTION EXPANSION - STURM-LIOUVILLE SYSTEM 235
Proof : Since λ is not any one of the eigenvalues, uniqueness of the solution
of Eq. (7.5.1) follows by Theorem 7.4.2. ¦φ
1
, φ
2
, ¦ is a complete orthonormal
set in L
2
[a, b] and hence we can write
f(t) =

¸
n=1
f
n
φ
n
; f
n
= (f, φ
n
) =

b
a
f(t)φ
n
(t) dt (7.5.5)
and
x(t) =

¸
n=1
x
n
φ
n
; x
n
= (x, φ
n
) =

b
a
x(t)φ
n
(t) dt (7.5.6)
Operating on Eq. (7.5.6) by (L −λI) we get
(L −λI)x(t) =

¸
n=1
x
n
(L −λI)φ
n
(t) =

¸
n=1
x
n

n
−λ)φ
n
(t)
Since (L −λI)x = f, we get

¸
n=1
x
n

n
−λ)φ
n
(t) =

¸
n=1
f
n
φ
n
(t)
That is

¸
n=1
[f
n
−x
n

n
−λ)] φ
n
(t) = 0. Completeness of ¦φ
1
, φ
2
, ¦ implies
that f
n
−x
n

n
−λ) = 0 and hence x
n
= f
n
/(λ
n
−λ).
Plugging in the value of x
n
in Eq. (7.5.6) we get
x(t) =

¸
n=1
f
n

n
−λ)
φ
n
(t)
Remark 7.5.1 If λ = λ
j
for some j, Eq. (7.5.1) has infinitely many solutions
if f
j
= 0 and no solution if f
j
= 0. This remark follows by Theorem 7.4.2.
Corollary 7.5.1 If zero is not one of the eigenvalues of the boundary value
problem
Lx = f; B
1
(x) = 0 = B
2
(x) (7.5.7)
then the above system has a unique solution x(t) given by
x(t) =

¸
n=1
f
n
λ
n
φ
n
(t) =

¸
n=1

b
a
f(y)φ
n
(y)dy

φ
n
(t)
λ
n
236 CHAPTER 7. BOUNDARY VALUE PROBLEMS
Example 7.5.1 Consider the self adjoint-system

d
2
x
dt
2
= f(t); x(0) = 0, ˙ x(1) =
1
2
x(1)
The associated eigenvalue problem is

d
2
φ
dt
2
= λφ; φ(0) = 0,
˙
φ(1) =
1
2
φ(1)
We take up different cases, depending upon the algebraic sign of λ. It can be
easily seen that there are no negative eigenvalues. For λ > 0, φ(t) = c
1
sin

λt
+c
2
cos

λt. Boundary condition φ(0) = 0 gives c
2
= 0 and hence φ(t) =
c
1
sin

λt.
The second boundary condition implies that λ satisfies the equation
tan

λ = 2

λ
Let λ = k
2
, then the above equation becomes tank = 2k. By considering the
curves y = tan x and y = 2x, we see that the above equation has infinitely many
solutions. So eigenvalues are given by λ = k
2
n
where k
n
is a non-zero solution of
tan k = 2k. The corresponding normalized eigenfunctions are −
2
cos 2k
n
sin k
n
t.
These eigenfunctions are orthogonal and form a complete set. Hence, by Theo-
rem 7.5.2, the solution x(t) of the above system is given by
x(t) =

¸
n=1
4 sink
n
t
k
2
n
(cos 2k
n
)
2
¸
1
0
f(y) sink
n
y dy

Example 7.5.2 Consider the boundary value problem

1
t
d
dt

t
dt
dt

= f(t)
lim
t→0
t
dx
dt
= 0; x(1) = 0
We first make the above system self-adjoint by defining a new inner product
(x, y) =

1
0
tx(t)y(t) dt and hence its eigenvalues are real. The associated eigen-
value problem is

1
t
d
dt

t

dt

+λφ = 0
That is,

dt
+
1
t

dt
+λφ = 0
This is a Bessel’s equation of zero order in the variable

λt and has the solution
φ(t) = AJ
0
(

λt) + BY
0
(

λt). But the first boundary condition gives B = 0.
7.6. NONLINEAR BOUNDARY VALUE PROBLEM 237
Hence, φ(t) = AJ
0
(

λt). The second boundary condition gives J
0
(

λ) = 0.
Thus, the zeros of the Bessel’s function J
0
(t) gives the eigenvalues of the above
boundary value problem. The corresponding eigenfunctions are φ
n
= J
0
(

λ
n
t).
¦J
0
(

λ
n
t)¦ forms a complete set and hence by Theorem 7.5.2
x(t) =

¸
n=1
f
n
φ
n
(t)
λ
n
=

¸
n=1
J
0
(

λ
n
t)
λ
n
|J
0
(

λ
n
t)|
2
¸
1
0
f(y)J
0
(

λ
n
y) dy

Theorem 7.5.3 Let ¦λ
1
, λ
2
, ¦ be a discrete set of eigenvlaues of L with
eigenfunctions ¦φ
1
, φ
2
, ¦ forming a complete set in L
2
[a, b]. Then the Green’s
function for Eq. (7.5.1) is given by g(t, s, λ) =

¸
n=1
φ
n
(t)φ
n
(s)
λ
n
−λ
provided λ is not
one of the eigenvalues.
Proof : Green’s function g(t, s, λ) is the solution of the system Lg − λg =
δ(t − s), B
1
(g) = 0 = B
2
(g). Applying Theorem 7.5.2 for f(t) = δ(t − s)
(formally) we get
g(t, s, λ) =

¸
n=1
δ
n
φ
n
(t)
λ
n
−λ
, where δ
n
=

b
a
δ(t −s)φ
n
(t) dt = φ
n
(t)
That is
g(t, s, λ) =

¸
n=1
φ
n
(t)φ
n
(s)
λ
n
−λ
For more on this section refer Bose and Joshi [2].
7.6 Nonlinear Boundary Value Problem
In this section we shall be concerned with two point boundary value problem of
the form
Lx = f(t, x(t)), t ∈ [a, b] (7.6.1)
B
1
(x) = α, B
2
(x) = β (7.6.2)
where L is the second order differential operator as defined before and B
1
(x), B
2
(x)
are boundary conditions on x.
Here f : [a, b] ' →' is nonlinear function satisfying Caratheodory condi-
tons:
(i) x →f(t, x) is continuous for almost all t ∈ [a, b].
(ii) t →f(t, x) is measurable for all values of x ∈ '.
238 CHAPTER 7. BOUNDARY VALUE PROBLEMS
Let g(t, s) be the Green’s function, corresponding to the operator L with given
zero boundary conditions. Then the solvability of the nonlinear boundary value
problem given by Eq. (7.6.1) is equivalent to the solvability of the integral
equation
x(t) =

b
a
g(t, s)f(s, x(s))ds +y(t) (7.6.3)
where y(t) is a known function given by y(t) = −J(x(s), g(t, s))[
b
a
, J being the
bilinear concomitant of x and g.
We shall prove the existence of solution of Eq. (7.6.3) in the space X =
L
2
[a, b] by using operator theoretic approach in terms oprators K and N, defined
as below
[Kx](t) =

b
a
g(t, s)x(s)ds (7.6.4)
[Nx] (t) = f(t, x(t)) (7.6.5)
Assumptions I
(i) g(t, s) is a Hilbert Schmidt kernel . That is
k
2
=

b
a

b
a
g
2
(t, s)dtds < ∞ (7.6.6)
(ii) f(t, x) satisfies growth condition of the form
[f(t, x)[ ≤ a(t) +b[x[, a ∈ L
2
(a, b), b > 0 (7.6.7)
(iii) x →f(t, x) is Lipschitz continuous - there exists constant γ > 0 such that
[f(t, x
1
) −f(t, x
2
)[ ≤ γ[x
1
−x
2
[ ∀t ∈ [a, b], ∀x
1
, x
2
∈ ' (7.6.8)
Theorem 7.6.1 Let Assumptions I hold and let kγ < 1. Then Eq. (7.6.1)
has a unique solution x

(t) which is obtained as a limit of the Picard iterates
x
n
(t), defined iteratively as
x
n
(t) = Tx
n−1
(t) =

1
0
g(t, s)f(s, x
n−1
(s))ds +y(s) (7.6.9)
with x
0
(t) = y(t).
Proof : Solvability of Eq. (7.6.1) or Eq. (7.6.3) is equivalent to the solvability
of the operator equation
x = KNx +y (7.6.10)
where K, N are as defined by Eq. (7.6.4), Eq. (7.6.5), respectively.
7.6. NONLINEAR BOUNDARY VALUE PROBLEM 239
We have already proved in Example 2.2.8 that K is a compact operator on
X. Also, inview of Assumption I(ii), it follows that N is a bounded continuous
operator on X ( refer Joshi and Bose [4] ). Further Assumption I(iii) implies
that N is Lipschitz continuous, that is,
|Nx
1
−Nx
2
|
X
≤ γ|x
1
−x
2
|
X
∀x
1
, x
2
∈ X
Let
Tx = KNx +y, ∀x ∈ X (y ∈ X fixed)
Then
|Tx
1
−Tx
2
| = |K(Nx
1
−Nx
2
)|
≤ |K| |Nx
1
−Nx
2
|
≤ kγ|x
1
−x
2
| ∀x
1
, x
2
∈ X
If α = kγ < 1, it follows that T is a contraction on the space X and hence
it has a fixed point x

and this point is approximated by the iterates x
n
=
Tx
n−1
, x
0
= y. This proves that x

(t) is a solution of Eq. (7.6.1) where
x

(t) = lim
n→∞
x
n
(t). Thus, we have
x
n
(t) =
¸
y(t) +

b
a
g(t, s)f(s, x
n−1
(s))ds
¸
and
x
0
(t) = y(t)
This proves the theorem.
Example 7.6.1 Let us consider the nonlinear boundary value problem
d
2
x
dt
2
= γ sinx(t) (7.6.11)
x(0) = α, x(1) = β (7.6.12)
Following Example 7.3.1, we find that the Green’s function g(t, s) for the above
problem is given by
g(t, s) =

t(s −1), t ≤ s
s(t −1), t ≥ s
This gives

1
0

1
0
g
2
(t, s)dtds =
1
90
and hence k = |K| =
1
3(

10)
.
240 CHAPTER 7. BOUNDARY VALUE PROBLEMS
Solvability of Eq. (7.6.11) is equivalent to the solvability of the integral equa-
tion
x(t) = γ

1
0
g(t, s) sin x(s)ds +y(t) (7.6.13)
where y(t) = α(1 −t) +βt.
It is clear that f(x) = γ sin x which satisfies Assumptions I(i) and (ii).
If we assume that γ

1
3(

10)

< 1, it follows that Eq. (7.6.11) has a unique
solution x

(t), which is the limit of x
n
(t), defined, iteratively as
x
n
(t) = α(1 −t) +βt +γ

1
0
g(t, s) sin x
n−1
(s)ds
with x
0
(t) = α(1 −t) +βt.
If f is monotone increasing, we may resort to the concept of monotone
operators (see Section 2.4) to get a solvability results for Eq. (7.6.1).
Assumptions II
(i) [g(t, s)] is symmetric positive semidefinite ( K is monotone) and Hilbert -
Schmidt.
(ii) Assumption I(ii) holds.
(iii) x →f(t, x) is strongly monotone: ∃ c > 0 such that
(f(t, x
1
) −f(t, x
2
)) (x
1
−x
2
) ≥ c(x
1
−x
2
)
2
x
1
, x
2
∈ '
Theorem 7.6.2 Under Assumptions II, Eq. (7.6.1) has a unique solution.
Refer Joshi and Bose [4] for the proof of this theorem.
Example 7.6.2 Consider the following boundary value problem:
d
2
x
dt
2
+γx + sin x = 0 (7.6.14)
x(0) = α, x(1) = β (7.6.15)
Assume that γ > 1.
Solvability of Eq. (7.6.14) is equivalent to the solvability of the integral equa-
tion
x(t) +

1
0
g(t, s)[γx(s) + sinx(s)]ds = α(1 −t) +βt
Observe that [Kx](t) =

1
0
g(t, s)x(s)ds is positive semi definite as K is the
inverse of the differential operator Lx = −
d
2
x
dt
2
which is positive semi definite
(monotone).
7.7. EXERCISES 241
As f(t, x) = γx +sinx, we get
∂f(t, x)
∂x
= γ +cos x ≥ (γ −1) > 0 for all x ∈ '.
Hence, x →f(t, x) is a strongly monotone function:
(f(t, x
1
) −f(t, x
2
)) (x
1
−x
2
) ≥ (α −1)(x
1
−x
2
)
2
∀x
1
, x
2
∈ '
As g and f satisfy all Assumptions II, it follows that Eq. (7.6.14) has a
unique solution.
Remark 7.6.1 If γ is such that γ >

3(

10) −1

> 1 then Eq. (7.6.14) is not
solvable through Theorem 7.6.1 whereas Theorem 7.6.2 is applicable.
7.7 Exercises
1. Consider the operator L with boundary conditions ˙ x(0) = x(1), ˙ x(1) = 0.
Find L

and adjoint boundary conditions.
2. Find the Green’s function for the operator L given by
(a) L = (t + 1)
d
2
dt
2
with boundary conditions x(0) = 0 = x(1).
(b) L =
d
2
dt
2
with boundary conditions ˙ x(0) = x(1), ˙ x(1) = 0.
3. Find the modified symmetric Green’s function for the system L =
d
2
dt
2
,
−1 ≤ t ≤ 1 with boundary conditions x(−1) = x(1) and ˙ x(−1) = ˙ x(1).
4. Show that the differential operator
Lx = −
1
w(t)
(a
0
(t) ˙ x) +a
2
(t)x, w(t) > 0
is formally self adjoint if the inner product is defined by
(x, y) =

1
0
w(t)x(t)y(t)dt
5. Prove that the solution x(t) =

1
0
g(t, s)f(s)ds of the system
Lx = f, B
1
(x) = 0 = B
2
(x)
depends continuously on f.
6. Using Green’s function technique, show that the solution x(t) of the
boundary value problem
d
2
x
dt
2
+tx = 1, x(0) = ˙ x(0) = 0
242 CHAPTER 7. BOUNDARY VALUE PROBLEMS
satisfies the Volterra integral equation
x(t) =

t
0
s(s −t)x(s)ds +
1

2
Prove that the converse of the above is also true. Compare this approach
with that of the initial value problem.
7. Obtain the solution of the following boundary value problems through
eigenfunction expansion.
(a)
d
2
x
dt
2
+x = t, x(0) = 0, ˙ x(1) = 0
(b)
d
2
x
dt
2
+x = sint, x(1) = 2, ˙ x(0) = 1
(c)
d
2
x
dt
2
+x = e
t
, x(0) = x(1), ˙ x(0) = ˙ x(1)
8. Compute the first few Picard iterates for the following nonlinear boundary
value problems
(a)
d
2
x
dt
2
= x +x
3
, x(0) = 0 = x(1)
(b)
d
2
x
dt
2
= x + sinx, ˙ x(0) = 0 = ˙ x(1)
References
[1 ] E. A. Coddington and N. Levinson, Theory of Ordinary Differential
Equations, McGraw Hill Book Co., 1955
[2 ] R. K. Bose and M. C. Joshi, Methods of Mathematical Physics, Tata
McGraw Hill Publ., co., Ltd., 1984
[3 ] N. Dunford and J.T. Schwartz, Linear Operators, Parts I and II, Inter-
scinece Publ., Inc., 1966
[4 ] M. C. Joshi and R. K. Bose, Some Topic in Nonlinear Functional Anal-
ysis, Halsted Press, 1968
[5 ] I. Stakgold, Boundary Vlaue Problems of Mathematical Physics, Vol I
and II, The Macmillan Co. 1968
Chapter 8
Control Theory Of Linear
Systems
Control theory deals with the maneuver of the state or trajectory of the sys-
tem, modelled by ODE. Some of the important properties of such systems are
controllability, observability and optimality.
In this chapter we show how operator theoretic approach in function spaces is
useful to discuss the above mentioned properties. We substantiate our theory
by concrete examples.
8.1 Controllability
We shall be interested in the linear system of the form
d¯ x
dt
= A(t)¯ x(t) +B(t)¯ u(t) (8.1.1a)
¯ x(t) = ¯ x
0
(8.1.1b)
where A(t) is nn matrix, B(t) is nm matrix, ¯ u(t) ∈ '
m
and ¯ x(t) ∈ '
n
. ¯ u(t)
is called control or input vector and ¯ x(t) the corresponding trajectory or state
of the system.
The typical controllability problem involves the determination of the control
vector ¯ u(t) such that the state vector ¯ x(t) has the desired properties. We assume
that the entries of the matrices A(t), B(t) are continuous so that the above
system has a unique solution ¯ x(t) for a given input ¯ u(t) .
Definition 8.1.1 The linear system given by Eq. (8.1.1) is said to be control-
lable if given any initial state ¯ x
0
and any final state ¯ x
f
in '
n
, there exist a
control ¯ u(t) so that the correspondoing trajectory ¯ x(t) of Eq. (8.1.1) satisfies
the condition
¯ x(t
0
) = ¯ x
0
, ¯ x(t
f
) = ¯ x
f
(8.1.2)
243
244 CHAPTER 8. CONTROL THEORY OF LINEAR SYSTEMS
The control ¯ u(t) said to steer the trajectory from the initial state ¯ x
0
to the final
state ¯ x
f
.
Assume that φ(t, t
0
) is the transition matrix of the above system. Then by
the variation of parameter formula, Eq. (8.1.1) is equivalent to the following
integral equation
¯ x(t) = φ(t, t
0
)¯ x
0
+

t
t0
φ(t, τ)B(τ)¯ u(τ)dτ (8.1.3)
As (t, τ) → φ(t, τ) is continuous, it follows that |φ(t, τ)| ≤ M for all t, τ ∈
[t
0
, t
f
].
So controllability of Eq. (8.1.1) is equivalent to finding ¯ u(t) such that
¯ x
f
= ¯ x(t
f
) = φ(t
f
, t
0
)¯ x
0
+

t
f
t0
φ(t
f
, τ)B(τ)¯ u(τ)dτ
Equivalently,
¯ x(t
f
) −φ(t
f
, t
0
)¯ x
0
=

t
f
t0
φ(t
f
, τ)B(τ)¯ u(τ)dτ (8.1.4)
Let us define a linear operator L : X = L
2
([t
0
, t
f
], '
m
) by
[L¯ u] =

t
f
t0
φ(t
f
, τ)B(τ)¯ u(τ)dτ (8.1.5)
Then, in view of (8.1.4), the controllability problem reduces to showing that
operator L is surjective.
Theorem 8.1.1 The system given by Eq. (8.1.1) is controllable iff the control-
lability Grammian
W(t
0
, t
f
) =

t
f
t0

φ(t
f
, τ)B(τ)B
¯
(τ)φ
¯
(t, t
f
)

dτ (8.1.6)
is nonsingular. A control ¯ u(t) steering the system from the initial state ¯ x
0
to
the final state ¯ x
f
is given by
¯ u(t) = B
¯
(t)φ
¯
(t
f
, t) [W(t
0
, t
f
)]
−1
[¯ x
f
−φ(t
f
, t
0
)¯ x
0
]
Proof : Let (¯ x
0
, ¯ x
f
) be any pair of vectors in '
n
'
n
. The controllability
problem of Eq. (8.1.1) reduces to the determination of ¯ u ∈ X such that

t
f
t0
φ(t
f
, τ)B(τ)¯ u(τ)dτ = ¯ z
f
= ¯ x
f
−φ(t
f
, t
0
)¯ x
0
This is equivalent to solving
L¯ u = ¯ z
f
8.1. CONTROLLABILITY 245
in the space X for a given ¯ z
f
, where L is given by Eq. (8.1.5).
We first note that L is a bounded linear operator as φ(t, τ) is bounded. Let
L

: '
n
→X be its adjoint. This is determined as follows.
For ¯ u ∈ X and ¯ α ∈ '
n
, we have
(L¯ u, ¯ α)
J
n
=

¯ α,

t
f
t0
φ(t
f
, τ)B(τ)¯ u(τ)dτ

J
n
=

t
f
t0
(¯ α, φ(t
f
, τ)B(τ)¯ u(τ))
J
n

=

t
f
t0

B
¯
(τ)φ
¯
(t
f
, τ)¯ α, ¯ u(τ)

J
n

= (¯ u, L

¯ α)
X
This implies that
[L

¯ α](t) = B
¯
(t)φ
¯
(t
f
, t)¯ α (8.1.7)
To claim that L is onto, it is sufficient to prove that LL

: '
n
→ '
n
is onto
(equivalently nonsingular). That is, given any ¯ z
f
∈ '
n
, find ¯ α ∈ '
n
such that
LL

¯ α = ¯ z
f
(8.1.8)
If Eq. (8.1.8) is solvable, then a control ¯ u which solves L¯ u = ¯ z
f
is given by
¯ u = L

¯ α.
Computing LL

we get
[LL

](¯ α) = L

B
¯
(t)φ
¯
(t
f
, t)¯ α

=
¸
t
f
t0

φ(t
f
, t)B(t)B
¯
(t)φ
¯
(t
f
, t)

dt

¯ α
= W(t
0
, t
f
)¯ α
If the controllability Grammian W(t
0
, t
f
) is nonsingular, we have a control ¯ u(t)
given by
¯ u(t) = L

α = L

(LL

)
−1
¯ z
f
= L

[W(t
0
, t
f
)]
−1
¯ z
f
= B
¯
(t)φ
¯
(t
f
, t) [W(t
0
, t
f
)]
−1
[¯ x
f
−φ(t
f
, t
0
)¯ x
0
]
which will steer the state from the initial state ¯ x
0
to the final state ¯ x
f
.
Conversely, let the Eq. (8.1.1) be controllable. That is, L is onto. Equiva-
lently, R(L) = '
n
. Then, by Theorem 2.2.1, N(L

) = [R(L)]

= ['
n
]

= ¦0¦.
Hence L

is 1-1. This implies LL

is also 1-1, which is proved as under. Assume
that LL

¯ α = 0. This gives 0 = (LL

¯ α, ¯ α) = (L

¯ α, L

α) = |L

α|
2
and hence
L

¯ α =
¯
0. L

1-1 implies that ¯ α = 0. Thus N(LL

) = ¦
¯
0¦.
246 CHAPTER 8. CONTROL THEORY OF LINEAR SYSTEMS
Thus W(t
0
, t
f
) = LL

: '
n
→'
n
is 1-1 and hence nonsingular . This proves
the result.
The controllability Grammian W(t
0
, t
f
) has some intersting properties, which
we now describe.
Theorem 8.1.2 (i) W(t
0
, t
f
) is symmetric and positive semidefinite.
(ii) W(t
0
, t
f
) satisfies the linear differential equation
d
dt
[W(t
0
, t)] = A(t)W(t
0
, t) +W(t
0
, t)A
¯
(t) +B(t)B
¯
(t)
W(t
0
, t
0
) = 0
(iii) W(t
0
, t
f
) satisfies functional equation
W(t
0
, t
f
) = W(t
0
, t) +φ(t
f
, t)W(t, t
f

¯
(t
f
, t)
Proof : We have
W(t
0
, t
f
) =

t
f
t0

φ(t
f
, τ)B(τ)B
¯
(τ)φ
¯
(t
f
, τ)


(i) It is clear that W
¯
(t
0
, t
f
) = W(t
0
, t
f
). Further,
(W(t
0
, t
f
)¯ α, ¯ α)
J
n
=

t
f
t0

φ(t
f
, τ)B(τ)B
¯
(τ)φ
¯
(t
f
, τ)¯ α, ¯ α

J
n

=

t
f
t0
|B
¯
(τ)φ
¯
(t
f
, τ)¯ α|
2
≥ 0
for all ¯ α ∈ '
n
.
(ii) Using Leibnitz rule, we get
d
dt
(W(t
0
, t)) =
d
dt
¸
t
t0
φ(t
f
, τ)B(τ)B
¯
(τ)φ(t, τ)dτ

= B(t)B
¯
(t) +
¸
t
t0
d
dt

φ(t
f
, τ)B(τ)B
¯
(τ)φ
¯
(t, τ)

= B(t)B
¯
(t) +
¸
t
t0
A(t)φ(t
f
, τ)B(τ)B
¯
(τ)φ(t, τ)dτ

+
¸
t
t0
φ(t
f
, τ)B(τ)B
¯
(τ)φ
¯
(t, τ)A
¯
(t)dτ

= A(t)W(t
0
, t) +W(t
0
, t)A
¯
(t) +B(t)B
¯
(t)
Bounadry condition W(t
0
, t
0
) = 0 follows from the definition.
8.1. CONTROLLABILITY 247
(iii) To get the functional equation we express W(t
0
, t
f
) as follows
W(t
0
, t
f
) =

t
t0
φ(t
f
, τ)B(τ)B
¯
(τ)φ
¯
(t
f
, τ)dτ
+

t
f
t
φ(t
f
, τ)B(τ)B
¯
(τ)φ
¯
(t
f
, τ)dτ
= W(t
0
, t) +φ(t
f
, t)

t
f
t
φ(t, τ)B(τ)B
¯
(τ)φ
¯
(t, τ)dτ

φ
¯
(t
f
, t)
= W(t
0
, t) +φ(t
f
, t)W(t, t
f

¯
(t
f
, t)
Remark 8.1.1 Theorem 8.1.1 implies that to check the controllability of the
linear system given by Eq. (8.1.1), one needs to verify the invertibility of the
Grammian matrix W(t
0
, t
f
). This is a very tedious task. However, if A(t) and
B(t) are time invariant matrics A, B, then the controllability of the linear system
is obtained in terms of the rank of the following controllability matrix
C =

B, AB, , A
n−1
B

(8.1.9)
The next theorem, in this direction, is due to Kalman [4].
Theorem 8.1.3 The linear autonomous system
d¯ x
dt
= A¯ x(t) +B¯ u(t) (8.1.10a)
¯ x(t
0
) = ¯ x
0
(8.1.10b)
is controllable iff the controllability matrix C given by Eq. (8.1.9) has rank n.
Proof : Assume the rank of C = n. It follows that N(C
¯
) = ¦
¯
0¦. We shall
prove that N(W(t
0
, t
f
)) = ¦
¯
0¦. Let ¯ α ∈ '
n
be such that W(t
0
, t
f
)¯ α = 0. This
implies that
0 = |W(t
0
, t
f
)¯ α|
2
=

t
f
t0
|B
¯
φ
¯
(t
f
, τ)¯ α|
2

=

t
f
t0
|B
¯
e
A

(t
f
−τ)
¯ α|
2

This implies that
B
¯
e
A

(t
f
−τ)
¯ α = 0 for all t ∈ [t
0
, t
f
] (8.1.11)
Expanding the LHS of Eq. (8.1.11) in a Taylor series arround t = t
f
, we get
that
0 = B
¯
¯ α = B
¯
(A
¯
)¯ α = B
¯
(A
¯
)
2
¯ α = B
¯
(A
¯
)
n−1
¯ α
248 CHAPTER 8. CONTROL THEORY OF LINEAR SYSTEMS
This implies that ¯ α ∈ N(C
¯
) = ¦
¯
0¦.
Conversely, let N(W(t
0
, t
f
)) = ¦
¯
0¦. We shall show that N(C
¯
) = ¦
¯
0¦.
Assume that ¯ α ∈ N(C
¯
). This implies that
B
¯
(A
¯
)
k
¯ α = 0, 1 ≤ k ≤ n −1 (8.1.12)
We have
W(t
0
, t
f
)¯ α =

t
f
t0

φ(t
f
, τ)BB
¯
φ
¯
(t, τ)¯ α


As φ(t
f
, τ) = e
A(t
f
−τ)
, it follows by Cayley-Hamilton theorem (refer prob-
lem number 4 in Section 5.5), that there exist functions α
k
(t
f
− τ) such that
φ(t
f
, τ) =
n−1
¸
k=0
α
k
(t
f
−τ)A
k
and hence W(t
0
, t
f
)¯ α is given by
W(t
0
, t
f
)¯ α =

t
f
t0
φ(t
f
, τ)B(τ)
¸
n−1
¸
k=0
α
k
(t
f
−τ)B
¯
(A
T
)
k
¯ α
¸
dτ (8.1.13)
In view of Eq. (8.1.12), we have
n−1
¸
k=0
α
k
(t
f
− τ)B
¯
(A
¯
)
k
¯ α = 0. This implies
that ¯ α ∈ N(W(t
0
, t
f
)) = ¦
¯
0¦.
This proves the theorem.
Example 8.1.1 Let us recall Example 2.3.5 which gives the linearized motion
of a satellite of unit mass orbiting around the earth. This is given by the control
system of the form given by Eq. (8.1.10), where
A =

0 1 0 0
3w
2
0 0 2w
0 0 0 1
0 −2w 0 0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
B =

0 0
1 0
0 0
0 1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
8.1. CONTROLLABILITY 249
For this system, one can easily compute the controllability matrix,
C =

B, AB, A
2
B, A
3
B

. It is given by
C =

0 0 1 0 0 2w −w
2
0
1 0 0 2w −w
2
0 0 −2w
3
0 0 0 1 −2w 0 0 −4w
2
0 1 −2w 0 0 −4w
2
2w
3
0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
One can verify that rank of C is 4 and hence the linearized motion of the satellite
is controllable. It is intersting to ask the following question.
What happens when one of the controls or thrusts becomes inoperative ?
For this purpose set u
2
= 0 and hence B reduces to B
1
= [0 1 0 0]
¯
.
So, the controllability matrix C
1
= [B
1
, AB
1
, A
2
B
1
, A
3
B
1
], is given by
C
1
=

0 1 0 −w
2
1 0 −w
2
0
0 0 −2w 0
0 −2w 0 2w
3
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
C
1
has rank 3.
On the other hand u
1
= 0 reduces B to B
2
= [0 0 0 1]
¯
and this gives control-
lability matrix C
2
= [B
2
, AB
2
, A
2
B
2
, A
3
B
2
], as
C
2
=

0 0 2w 0
0 2w 0 −2w
3
0 1 0 −4w
2
1 0 −4w
2
0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
C
2
has rank 4.
Since u
1
was radial thrust and u
2
was trangential thrust, we see that the loss
of radial thrust does not destroy controllability where as loss of tangential thrust
does. In terms of practical importance of satellite in motion, the above analysis
means that we can maneuver the satellite just with radial rocket thrust.
250 CHAPTER 8. CONTROL THEORY OF LINEAR SYSTEMS
8.2 Observability
Consider the input-output system of the form
d¯ x
dt
= A(t)¯ x(t) +B(t)¯ x(t) (8.2.1a)
¯ y(t) = C(t)¯ x(t) (8.2.1b)
The questions concerning observability relate to the problem of determining the
values of the state vector ¯ x(t), knowing only the output vector ¯ y(t) over some
interval I = [t
0
, t
f
] of time.
As in the previous section, A(t) ∈ '
n·n
, B(t) ∈ '
n·m
are assumed to be
continuous functions of t. Let φ(t, t
0
) be the transition matrix of the above
system. Then the output vector ¯ y(t) can be expresssed as
¯ y(t) = C(t)φ(t, t
0
)¯ x
(
t
0
) + ¯ y
1
(t) (8.2.2)
where ¯ y
1
(t) is known quantity of the form
y
1
(t) =

t
f
t0
C(t)φ(t, τ)B(τ)u(τ)dτ
Thus, from Eq. (8.2.2) it follows that if we are concerned about the deter-
mination of ¯ x(t
0
), based on the output ¯ y(t), we need only the homogeneous
system
d¯ x
dt
= A¯ x(t) (8.2.3a)
¯ y(t) = C(t)¯ x(t) (8.2.3b)
in place of Eq. (8.2.1).
Definition 8.2.1 We shall say that Eq. (8.2.3) is observable on I = [t
0
, t
f
] if
¯ y(t) = 0 on I implies that ¯ x(t) = 0 on I.
We define a linear operator L : '
n
→X = L
2
([t
0
, t
f
], '
n
) as
[L¯ x
0
](t) = C(t)φ(t, t
0
)¯ x
0
= H(t)¯ x
0
(8.2.4)
where H : t →H(t) is a matrix function which is continuous in t. Observability
of the system Eq. (8.2.3), then, reduces to proving that L is one-one.
The following theorem gives the observability of the system given by Eq. (8.2.3),
in terms of the nonsingularity of the matrix M(t
0
, t
f
).
Theorem 8.2.1 For the homogeneous system given by Eq. (8.2.3), it is possible
to determine the initial state ¯ x(t
0
) within an additive constant vector which lies
with null space of M(t
0
, t
f
) which is defined by
M(t
0
, t
f
) =

t
f
t0
φ
¯
(t, t
0
)C
¯
(t)C(t)φ(t, t
0
)dt (8.2.5)
8.2. OBSERVABILITY 251
Hence ¯ x(t
0
) is uniquely determined if M(t
0
, t
f
) is nonsingular. That is, Eq.
(8.2.3) is observable ifff M(t
0
, t
f
) is nonsingular.
Proof : If the output vector ¯ y(t) is known, solving ¯ y(t) = C(t)φ(t, t
0
)¯ x(t
0
) for
¯ x(t
0
) is equivalent to solving the operator equation
L¯ x(t
0
) = ¯ y (8.2.6)
where L is defined by Eq. (8.2.4). Premultiplying Eq. (8.2.6) by L

, we get
[L

L] ¯ x(t
0
) = L

¯ y (8.2.7)
where L

: X →'
n
is given by
[L

u] =

t
f
t0
φ
¯
(t, t
0
)C
¯
(t)u(t)dt (8.2.8)
Observe that M(t
0
, t
f
) = L

L. If RHS of Eq. (8.2.7) is known, ¯ x(t
0
) is deter-
mined up to an additive constant lying in the null space of L

L = M(t
0
, t
f
). If
L

L is invertible, then it is clear that L is one-one and hence the system given
by Eq. (8.2.3) is observable.
Conversely, if L is one-one then so is L

L and hence R(L

L) = [N(L

L)]

=
'
n
. This implies that L

L is nonsingular. For such an observable system, the
initial state ¯ x(t
0
) is given by
¯ x(t
0
) = (L

L)
−1
L

y = [M(t
0
, t
f
)]
−1
¸
t
f
t0
φ
¯
(t, t
0
)C
¯
(t)y(t)dt

(8.2.9)
This proves the theorem.
M(t
0
, t
f
) defined by Eq. (8.2.5) is called the observability Grammian.
Theorem 8.2.2 The observability Grammian M(t
0
, t
f
) satisfies the following
properties
(i) M(t
0
, t
f
) is symmetric and positive semidefinite.
(ii) M(t
0
, t
f
) satisfies the following matrix differential equation
d
dt
[M(t, t
1
)] = −A
¯
(t)M(t, t
1
) −M(t, t
1
)A(t) −C
¯
(t)C(t)
M(t
1
, t
1
) = 0
(iii) M(t
0
, t
f
) satisfies the functional equation
M(t
0
, t
f
) = M(t
0
, t) +φ
¯
(t, t
0
)M(t, t
1
)φ(t, t
0
)
252 CHAPTER 8. CONTROL THEORY OF LINEAR SYSTEMS
The proof of the above theorem follows along the lines of the proof of the
Theorem 8.1.2. See Brocket [1] for more details.
We also have the following theorem giving the necessary and sufficient con-
tition which is easily verifiable for observability.
Let us denote by O the observability matrix
O = [C, CA, , CA
n−1
] (8.2.10)
Theorem 8.2.3 The system given by Eq. (8.2.3) is observable iff the observable
matrix O given by Eq. (8.2.10) is of rank n.
Example 8.2.1 As a continuation of Example 8.1.1, consider the satellite or-
biting around the Earth.
We assume that we can measure only the distance r of the satellite from the
centre of the force and the angle θ. So, the defining state-output equation of the
satellite is given by
d¯ x
dt
= A¯ x(t)
¯ y(t) = Cx(t)
where
A =

0 1 0 0
3w
2
0 0 2w
0 0 0 1
0 −2w 0 0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
C =

1 0 0 0
0 0 1 0
¸
¸
¯ x = (x
1
, x
2
, x
3
, x
4
), ¯ y = (y
1
, y
2
); y
1
= r, y
2
= θ being the radial and angle
measuments. So, the observability matrix O is given by
O = [C, CA, CA
2
, CA
3
]
=

1 0 0 0
0 0 1 0
0 1 0 0
0 0 0 1
3w
2
0 0 2w
0 −2w 0 0
0 −w
2
0 0
−6w
3
0 0 −4w
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
Rank of C is 4 and hence the above state -output system is observable.
8.3. OPTIMAL CONTROL 253
To minimize the measurement, we might be tempted to measure y
1
only not y
2
.
This gives
C
1
= [1 0 0 0 0]
¯
and
O
1
= [C
1
, C
1
A, C
1
A
2
, C
1
A
3
] =

1 0 0 0
0 1 0 0
3w
2
0 0 2w
0 −w
2
0 0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
which is of rank 3.
However, if y
1
is not measured, we get
C
2
= [0 0 1 0 0]
¯
and
O
2
= [C
2
, C
2
A, C
2
A
2
, C
2
A
3
]
=

0 0 1 0
0 0 0 1
0 −2w 0 0
−6w
3
0 0 −4w
2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
which is of rank 4.
Thus, the state of the satellite is known from angle measurements alone but this
is not so for radial measurements.
8.3 Optimal Control
In this section we shall be concerned with the optimization problem of two types.
(i) Unconstrained optimal control problem:
Find the control ¯ u

∈ U = L
2
([t
0
, t
f
], '
m
) that minimizes the cost func-
tional J(u) which is given by
J(¯ u) =

t
f
t0
[(¯ u(t), ¯ v(t)) + (R¯ x(t), ¯ x(t))] dt (8.3.1)
where ¯ x ∈ X is the state of the dynamical system
d¯ x
dt
= A(t)¯ x(t) +B(t)¯ u(t) (8.3.2a)
¯ x(t
0
) = ¯ x
0
(8.3.2b)
254 CHAPTER 8. CONTROL THEORY OF LINEAR SYSTEMS
corresponding to the given control ¯ u(t). R is a positive definite symmetric
matrix and X = L
2
([t
0
, t
f
], '
n
).
(ii) Constrained optimal control system:
Find the control ¯ u

that minimizes J(¯ u) over the constrained set U
c
⊆ U,
where J(¯ u) and U
c
are given as follows
J(¯ u) =

t
f
t0
(¯ u(t), ¯ u(t)) dt (8.3.3)
U
c
= ¦¯ u ∈ U : ¯ u steers the state ¯ x(t) of Eq. (8.3.1) from the initial state
¯ x
0
to the final state ¯ x
f
¦.
We shall show that unconstrained problem is a problem of minimization of a
strictly convex functional defined on the Hilbert space U.
Using the transition matrix φ(t, t
0
), we obtain the state ¯ x(t) of the linear
system given by Eq. (8.3.2) through the integral equation.
¯ x(t) = φ(t, t
0
)¯ x
0
+

t
t0
φ(t, τ)B(τ)u(τ)dτ (8.3.4)
We now define a bounded linear operator K from U to X as follows.
[K¯ u](t) =

t
t0
φ(t, τ)B(τ)¯ u(τ)dτ
Set ¯ y
0
(t) = φ(t, t
0
)¯ x
0
, then in abstract notation, ¯ x ∈ X is given by
¯ x = ¯ y
0
+K¯ u (8.3.5)
So, the cost functional J on U, determined by Eq. (8.3.1), is given by
J(¯ u) = (¯ u, ¯ u)
U
+ (T ¯ x, ¯ x)
X
where T is a positive definite operator induced by R on X.
Using the concept of the derivative of a functional (refer section 2.3), we
obtain
J
t
(¯ u) = ¯ u +K

TK¯ u +K

TK¯ y
0
(8.3.6)
and
J
tt
(u) = I +K

TK ∀ ¯ u ∈ U (8.3.7)
Theorem 8.3.1 J is a strictly convex functional on the space U and hence it
has a unique minimizer ¯ u

∈ U of J(u) and is given by
¯ u

= −[I +K

TK]
−1
[K

TK

¯ y
0
] (8.3.8)
8.3. OPTIMAL CONTROL 255
Proof : We have
J
t
(¯ u) = ¯ u +K

TK¯ u +K

TK¯ y
0
J
tt
(¯ u) = I +K

TK
for all ¯ u ∈ U.
Note that J
tt
(u) : U →U is positive definite for all u ∈ U, as
(J
tt
(u)¯ v, ¯ v) = (¯ v, ¯ v)
U
+ (K

TK¯ v, ¯ v)
U
= |¯ v|
2
+ (TK¯ v, K¯ v)
X
> |¯ v|
2
(for ¯ v = 0)
in view of positive definiteness of T. As J
tt
(u) is positive definite, it follows by
Remark 2.3.2 that J is strictly and hence it follows that J has unique minimizer
¯ u

∈ U and is given by the critical point of J (refer Joshi and Kannan [5]
Chapter1). Hence, we have
¯ u

+K

TK¯ u

= −K

TK¯ y
0
Also, the invertibility of [I +K

TK] (refer Joshi and Bose [3]), gives
¯ u

= −[I +K

TK]
−1
(K

TK¯ y
0
)
Remark 8.3.1 One can obtain integro-differential equation analog of Eq. (8.3.8)
which is equivalent to the equation
¯ u

(t) +

t
f
t
φ(τ, t)B(TK¯ u

)(τ)dτ +

t
f
t
(TK)(¯ y
0
(τ))dτ = 0 (8.3.9)
Differentiating Eq. (8.3.9) we get
d¯ u

dt
−BTKu

(t) +

t
f
t
d
dt
[φ(τ, t)] BTKu

(τ)dτ = TK¯ y
0
(t)
As
d
dt
[φ(τ, t)] = −φ(τ, t)A, we get the equivalent integro-differential equation
d¯ u

dt
= −BTKu

(t) −

t
f
t
φ(τ, t)ABTKu

(τ)dτ (8.3.10a)
= TK¯ y
0
(t) (8.3.10b)
with u

(t
f
) = 0.
In the above equation we assume that the system Eq.(8.3.2) is autonomous.
We now address ourself to second problem pertaining to constrained optimal
control. We have the following theorem in this direction.
256 CHAPTER 8. CONTROL THEORY OF LINEAR SYSTEMS
Theorem 8.3.2 Let W(t
0
, t
f
) be the controllability Grammian of Eq. (8.3.2).
If ¯ u

is any control of the form
¯ u

(t) = B
¯
(t)φ
¯
(t
f
, t)¯ α (8.3.11)
where ¯ α satisfies the equation
W(t
0
, t
f
)¯ α = ¯ x
f
−φ(t
f
, t)¯ x
0
(8.3.12)
Then ¯ u

∈ U
c
. Further, if ¯ u is any, controllability vector belonging to U
c
, then

t
f
t0
|u

(t)|
2
dt ≤

t
f
t0
|¯ u(t)|
2
dt (8.3.13)
Moreover, if W(t
0
, t
f
) is nonsingular, then

t
f
t0
|¯ u

(t)|
2
dt =

W
−1
(t
0
, t
f
) [¯ x
f
−φ(t
f
, t
0
)] ¯ x
0
, [¯ x
f
−φ(t
f
, t
0
)] ¯ x
0

Proof : We have
¯ z
f
= ¯ x
f
−φ(t
f
, t
0
)¯ x
0
=

t
f
t0
φ(t
f
, τ)B(τ)¯ u(τ)dτ = L¯ u
¯ z
f
= ¯ x
f
−φ(t
f
, t
0
)¯ x
0
=

t
f
t0
φ(t
f
, τ)B(τ)u

(τ)dτ = L¯ u

where L is defined by Eq. (8.1.5).
Subtracting the above two equations, we have
L¯ u −L¯ u

= 0
This gives
(L¯ u

, ¯ α)
U
= (L¯ u, ¯ α)
U
(8.3.14)
where ¯ α solves Eq. (8.3.12). Solvability of Eq. (8.3.12) is equivalent to the
solvability of the following equation
LL

¯ α = ¯ z
f
Also Eq. (8.3.11) is equivalent to defining ¯ u

= L

¯ α. Hence, Eq. (8.3.14) gives
|¯ u

|
2
U
= (¯ u

, ¯ u

)
U
= (¯ u

, L

¯ α)
U
= (L¯ u, α)
U
= (¯ u, L

α)
U
= (¯ u, ¯ u

)
U
(8.3.15)
Eq. (8.3.15) gives
0 ≤ |¯ u − ¯ u

|
2
U
= (¯ u − ¯ u

, ¯ u − ¯ u

)
U
= |¯ u|
2
U
−2(¯ u, ¯ u

)
U
+|¯ u

|
2
U
= |¯ u|
2
U
−|¯ u

|
2
U
8.3. OPTIMAL CONTROL 257
That is
|¯ u

|
2
U
≤ |¯ u|
2
U
or

t
f
t0
(¯ u

(t), ¯ u

(t))
J
m dt ≤

t
f
t0
(¯ u(t), ¯ u(t))
J
mdt
Further, if LL

= W(t
0
, t
f
) is invertible, we get
|¯ u

|
2
= |L

¯ α|
2
= (L

¯ α, L

¯ α)
= (LL

¯ α, ¯ α) = (W(t
0
, t
f
)¯ α, ¯ α)
= (¯ x
f
−φ(t
f
, t)x
0
, W
−1
(t
0
, t
f
) [¯ x
f
−φ(t
f
, t)¯ x
0
])
Example 8.3.1 Consider the following electrical network, given by Figure 8.3.1.
v i c
r
Figure 8.3.1: Electrical network
Its equation is given by
d
dt
(cv(t)) = i(t), c = Capacitance (8.3.16)
i is the current out of the source. The energy dissipated in the resister in the
time interval [0, t
f
] is

t
f
0
ri
2
(t)dt .
We wish to find the current i(t) such that the voltage v(0) = v
0
and v(t
f
) =
v
1
and the energy disspipated is minimum.
Thus, we want to minimize

t
f
0
i
2
(t)dt subject to the output voltage satisfying
the equation Eq. (8.3.16) with the constraint that v(0) = v
0
and v(t
f
) = v
1
.
258 CHAPTER 8. CONTROL THEORY OF LINEAR SYSTEMS
For this problem, the controllability Grammian W(t
0
, t
f
) =
t
f
c
2
and hence the
optimal current i

(t) =
c
t
f
(v
1
−v
0
) and hence is constant on the interval [0, t
f
].
For more on control theory refer Brockett [1] and Gopal[3].
8.4 Exercises
1. Show that the system
d¯ x
dt
= A(t)¯ x(t) +B¯ u(t)
where A(t) =

1 ¯ e
t
0 −1
¸
¸
, B =

0
1
¸
¸
is controllable. Find the control ¯ u

(t) with minimum L
2
- norm, which
will steer the trajectory from the initial state
¯
0 at t = 0 to the final state
(1, 1) at time t
f
= 1.
2. Show that the linear system given by the pair ¦A,
¯
b¦ is controllable for all
values of α
i
’s arising the definition of A.
A =

0 1 0 0
0 0 1 0
0 0 0 0
.
.
.
.
.
.
0 0 0 1
−α
1
−α
2
−α
3
−α
n
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
,
¯
b =

0
0
0
.
.
.
1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
3. Prove that the systems
d¯ x
dt
= A(t)¯ x(t) +B¯ u(t) (∗)
and
d¯ z
dt
= −A
¯
(t)¯ z(t) (∗∗)
¯ w(t) = B
¯
¯ z(t)
are dual of each other. That is (*) is controllable iff (**) is observable.
8.4. EXERCISES 259
4. Show that, if the time invariant system
d¯ x
dt
= A(t)¯ x(t) +B¯ u(t)
is controllable, then there exists a matrix C such that
d¯ x
dt
= (A+BC)¯ x(t) +
¯
b
i
v(t)
is also controllable where
¯
b
i
is any nonzero column vector of B.
5. Consider the system
d¯ x
dt
= g(t) [A¯ x(t) +B¯ u(t)]
with g continuous and bounded (0 < α ≤ g(t) ≤ β < ∞) .
Show that if C =

B, AB, , A
n−1
B

has rank n, then the above system
is controllable.
6. Obtain controllable conditions analogous to Theorem 8.1.1 for the follow-
ing matrix differential equation
dX(t)
dt
= A(t)X(t) +X(t)B(t) +C(t)U(t)D(t)
7. Show that the differential equation
d
2
y
dt
2
+u(t)
dy
dt
+y(t) = 0
is controllable ( in the sense that given any y
0
, ˙ y
0
and y
1
, ˙ y
1
, there exists
a control u(t) such that y(t
0
) = y
0
, y(t
f
) = y
1
, ˙ y(t
0
) = ˙ y
0
, ˙ y(t
f
) = ˙ y
1
),
provided y
2
0
+ ˙ y
2
0
and y
1
+ ˙ y
2
1
are nonzero.
8. Show that there exists an initial state for the adjoint system
d¯ p
dt
= −A
¯
(t)¯ p(t)
such that the control ¯ u(t), which minimizes the functional
φ(¯ u) =

t
f
t0
(¯ u(t), ¯ u(t)) dt
for steering the initial state ¯ x
0
to the final state ¯ x
f
, is given by
¯ u(t) = −B
¯
(t)¯ p(t)
260 CHAPTER 8. CONTROL THEORY OF LINEAR SYSTEMS
References
[1 ] Roger W. Brockett, Finite Dimensional Linear Systems, John Wiley and
Sons, Inc, 1970
[2 ]M. Gopal, Modern Control System Theory, Wiley Eastern Ltd., 1989
[3 ] Mohan C. Joshi and Ramendra K. Bose, Some topic in Nonlinear Func-
tional Analysis, Halsted Press, 1985
[4 ] R. E. Kalman, Mathematical Description of Linear Dynamical Systems,
SIAM J Control, Vol I, pp152-192, 1963
[5 ] Mohan C. Joshi and Kannan M. Moudgalya, Optimization -Theory and
Practice, Narosa Publishing House, Pvt. Ltd., 2004
Index
Abel’s formula, 104
adjoint Green’s function, 216, 219
adjoint operator, 50, 209, 210
adjoint system, 111, 112, 118, 136,
259
Analytic Coefficients, 181
Ascoli-Arzela theorem, 96, 97
asymptotic convergence, 142, 147,
148
asymptotically stable, 149–151, 153–
155, 166, 168, 171, 173
autonomous system, 9, 115, 154, 247
autonomous systems, 135
Banach space, 39–41, 65
Bessel’s function, 195, 200, 237
Boundary condition, 236
Boundary conditions, 222
boundary conditions, 207–209, 211,
212, 214, 215, 219–221, 224,
225, 227, 230–232, 237, 238,
241
Boundary Value Problem, 237
Boundary value problem, 211
boundary value problem, 211, 212,
214, 216–222, 226, 234–237,
240, 241
boundary value problems, 45, 48, 71,
207, 242
bounded linear operator, 48, 77, 245,
254
bounded solutions, 142, 144, 145
carrying capacity, 4
Cauchy sequence, 39
Cauchy-Peano theorem, 96, 97
Cayley-Hamilton theorem, 139, 171,
248
compact operator, 54, 55, 234, 239
compact set, 80, 97
Comparision test, 176
comparision test, 177
continuous linear operators, 49
contraction mapping principle, 63,
68
Controllability, 243
controllability, 244, 247, 256
controllability Grammian, 244–246,
256, 258
controllability matrix, 247, 249
convex functional, 254
differential equation, 1, 11, 26, 27
differential operator, 69, 75, 76, 207,
209–211, 217, 237, 240
Dirac-Delta Function, 71
Dirac-delta function, 71, 73
direction field, 1, 2, 6, 34
Eigenfunction Expansion, 233
eigenfunction expansion, 242
Eigenvalues, 125
eigenvalues, 20, 21, 56, 120–123, 127,
129, 132, 137–139, 156, 168,
196, 234
eigenvectors, 20, 21, 55, 56, 132, 137,
156, 168
equilibrium points, 3, 4, 9, 34, 154
Euler’s system, 131–133
First Order Equations, 26
fixed point, 12, 63–65, 67, 68, 81,
97, 239
261
262 INDEX
fixed point theorem, 63, 96, 97
Floquet-Lyapunov theorem, 148
formal adjoint, 210
fundamental matrix, 107–109, 111,
112, 132, 134, 145, 146, 150
generalized eigenvector, 123
generalized eigenvectors, 123, 129,
137
Generating function, 200
generating function, 197, 201, 202
gradient, 59
Gram-Schmidt orthonormalisation,
46
Green’s function, 37, 207, 212, 214–
217, 219–223, 226–231, 233,
234, 237–239, 241
Gronwall’s inequality, 140, 142, 146,
147
Gronwall’s lemma, 93, 140, 143
growth rate, 3, 9
Harmonic Oscillator, 13
harmonic oscillator, 117, 164, 187
Hermite Differential Equation, 187
Hermite differential equation, 187
Hilbert Schmidt kernel, 238
Hilbert space, 39, 41–43, 50, 53, 55,
56, 59, 60, 70, 107, 209,
211, 254
Hilbert spaces, 50
infinite series, 175
initial condition, 79, 93, 94, 101, 107,
152
Initial Value Problem, 2, 79
inner product space, 38, 39, 41
integral equation, 2, 22, 80, 91, 94,
98, 107, 234, 238, 240, 242,
244, 254
integral operator, 61, 212, 218, 234
integro-differential equation, 255
Isocline, 6
isomorphism, 49
Jordan canonical form, 137
Legendre Differential Equation, 185
Legendre differential equation, 47,
186, 187
linear ODE, 7
linear system, 10, 66, 102, 106, 138–
140, 144, 149, 150, 154, 157,
162, 254, 258
linearization, 10, 62, 63
linearly dependent, 39, 102
linearly independent, 20, 21, 56, 101–
107, 110, 120, 193, 222–225,
227
Lipschitz continuity, 79, 84
locally integrable, 73, 75, 76
Lyapunov function, 163, 165, 166,
168–171
Lyapunov Stability, 163
maximal monotone, 69, 88
mechanical oscillations, 125
metric space, 38, 63, 67, 80, 81, 95,
96
modified Green’s function, 229, 231
monotone operator, 70
motion of a satellite, 62, 248
neural network, 23, 24
neural solution, 21, 23, 25
Newton’s algorithm, 62
non-autonomous system, 114, 145,
147
nonlinear boundary value problem,
238, 239
nonlinear boundary value problems,
242
normed space, 38–41, 49, 57, 59, 64,
68
numerical methods, 21, 22
observability, 250–252
optimization, 4, 24, 253
orbit, 2, 3, 19
orthogonal matrix, 56
orthonormal set, 43, 44, 54, 196, 234,
235
Parseval’s equality, 43, 54
INDEX 263
Peano-Baker series, 108, 114, 134
Periodic Linear System, 114
perturbed system, 141, 145, 147
phase portrait, 154, 155, 158, 159,
161
phase portraits, 156, 158, 173
phase space, 1, 154
Picard iterates, 81–83, 100, 238, 242
positive definite operator, 254
predator-prey system, 9–11
Ratio test, 176
regular singularity, 188
Rodrigues formula, 197
Root test, 177
Schauder’s theorem, 68
Schwartz inequality, 41, 48
singular distribution, 73, 74
solution curve, 2, 4, 5, 14, 15, 19, 85
stable focus, 160
stable node, 158
Sturm-Liouville system, 233
test functions, 72, 73
trajectory, 2, 3, 8, 11, 164, 165, 243,
244, 258
transition matrix, 56, 101, 107–109,
112–114, 117, 118, 120, 121,
123, 125–127, 129, 130, 134,
135, 138, 140, 148, 150, 244,
250, 254
unstable focus, 160
variation of parameter, 112, 113, 140,
146, 148, 244
vector field, 1, 2, 8, 12, 23, 155, 158
vector space, 37–39, 73
Wronskian, 102, 106, 134, 222

Ordinary Differential Equations
Modern Perspective

Mohan C Joshi IIT Bombay

Preface
This book has culminated into its title as an outgrowth of series of undergraduate and graduate level courses on ordinary differential equations (ODE), taught by the author for the past several years. It is our aim to present a unified and comprehensive treatment to ODE based on widely used abstract formulations in the underlying spaces. The emphasis is more on qualitative analysis of solvability of ODE than the actual methodology of computing the closed form solution. In our presentation we presume only the basic knowledge of linear algebra and advanced calculus of the undergraduate level. Hence the text is suitable for graduate students of mathematics and other branches of science and engineering requiring a broad perspective of ODE. The contents have been class tested for several years in a regular course on ODE in the Department of Mathematics at IIT Bombay. With this approach in mind, we have divided the text into eight chapters. Chapter 1 introduces ODE with several examples drawn from applied areas. These examples are continuously tracked at appropriate places in later chapters for illustrating the theory. Chapter 2 is devoted to the development of mathematical structure required to deal with linear and nonlinear operators in function spaces. These operators will arise while discussing various facets of ODE. Chapter 3 gives basic existence and uniqueness theory needed for initial value problems, the starting point for any discussion of ODE. The subsequent analysis rests heavily on the concept of transition matrix. It is, therefore, dealt with in detail in Chapter 4. While in Chapter 5, we use the properties of transition matrix in the study of stability theory of both linear and nonlinear systems. Chapter 6 lays emphasis on the clasical theory of series solution for ODE. We focus on Legendre, Hermite and Bessel equations. Boundary value problems are investigated in Chapter 7 through the methods of Green’s function and eigenfunction expansion. We give rather an elongated description on Green’s function in view of its inherent importance. The closing Chapter 8 is a fitting finale, giving interesting glimpse of vast applicability potential of ODE to control theory - a fertile inter-disciplinary area involving of science and engineering.

Our special thanks to Asmit Pany for type setting the manuscript in LATEX. In this connection. Mumbai . Several colleagues and students helped us from time to time in many ways giving correct references. Secondly. we would like to acknowledge the financial support received from the Curriculum Development Program of the Institute.Acknowledgements This book could not have been written without direct and indirect assistance from various institutional bodies within IIT Bombay. we would like to acknowlege the help rendred by Amiya K Pani. this author would like to thank his wife for patiently supporting him in implementing the project. Firstly. Finally. Joshi Powai. we thank the Department of Mathematics and Industrial Mathematics Group for providing the necessary secretarial and computational help throughout the preparation of the manuscript while it went through the number of iterations. M K Ramaprasad and Soumya Mohanty. proof-reading the manuscript etc. Mohan C.

. .2 Transition Matrix . . . . . . . . . . .2 Linear and Nonlinear Operators . . . . . . . .2 Classification of ODE . . . . . . . . . . . . . . . . . . . . . 1 1 7 11 19 21 26 34 36 37 37 47 57 63 71 76 78 79 . . .4 Fixed Point Theorems and Monotone Operators 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 . . . . . References . . . . . .5 Euler’s Linear System . . . . . . 1. . . . . . . 79 . . . . . . . . . . . . . . . . 3 Initial Value Problems 3. . . . . . . . . . . . . . . . . . . . . . . . 2. . . .4 Difference Equations . . 3. . . . . 97 . . . . . . . . . . . . . . . . 4. . . . . . . . . . . . . . . . . . .3 Higher Order ODE . . 4. . . . . . . . 1.1 Finite and Infinite Dimensional Spaces . . . . . . . . . . . .4 Cauchy-Peano Existence Theorem . . .3 Periodic Linear System . . . . . . . 96 . . . . . . . . . 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Contents 1 Introduction To Ordinary Differential Equations 1. . . . 4. . 2 Mathematical Preliminaries 2. . . . .5 Exercises . . . . . . . . . . . 1. . . . . . . . . . . . . . . . . . 2.1 Linear System . . . . . . 4. . . . . . .1 Existence and Uniqueness Theorems 3. . . . . . . . . . . . . . . . . . . . .6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Lipschitz Continuity and Differentiability . . . . 4. . . . . . . References . . . . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6 Exercises . . . . . . . . . . 101 101 106 114 118 131 133 4 Linear System And Transition Matrix 4. 1. . . . .3 Dependence on Initial Condition . . . . . . . . 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1. . . . . .6 Techniques for Solving First Order Equations . . . .7 Exercises . . . . . . . . . . . . . . . . . . . .5 Dirac-Delta Function . .4 Computation of Transition Matrix . 2. . . . . . . . . .5 Discretization .1 First Order Ordinary Differential Equations . . . 89 . . . . . . . . . . . . . . . . . . . . . . . . . . 100 . . . . . . . . References . . . . . . . . . . . . . .2 The Maximal Interval of Existence . . . . . .

. . . . . . . . . . . . . . . . . .5 Eigenfunction Expansion . . . . . . . . . . . . . . . . . . . . . .5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Preliminaries . . . 6. . . . . . . . . . . . . . . .5 Exercises . . . . . . . . . . . . . . . .3 Linear System with Regular Singularities . . . 136 5 Stability Analysis 5. . . . . . . . 7. .Sturm-Liouville System . . . . . . . . . . . . . . . . . . 7 Boundary Value Problems 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . .2 Adjoint of a Differential Operator . .1 Controllability . . . . . 8. . .3 Phase Portrait Analysis . . . . . . . . .4 Lyapunov Stability . . . . . . . . .1 Asymptotic Behaviour of Linear System 5. . . . . . . .Formal Approach .7 Exercises . . . . . . . . . . . . .4 Properties of Legendre. . . . . . . . . 7. .2 Stability Analysis .4 Construction of Green’s Function and Modified Green’s Function 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Green’s Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . References . . . . . . . . . . . . . . . . . .3 Optimal Control . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . . . . . . . . . 8 Control Theory Of Linear 8. . . . . . . . . .1 Introduction . . . . . . . . . . .CONTENTS References . . . . 7. . . . . . 8. . . . . . 137 137 149 154 163 171 174 175 175 181 188 196 205 206 207 207 209 212 221 233 237 241 242 243 243 250 253 258 260 6 Series Solution 6. . . . . . . . 7. . 8. 5. . . . . . . 7. . . .4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Linear System with Analytic Coefficients . . . . . . . . . . . . . Hermite and Bessel Functions 6. . . . . . . . . . . 6. . . . . . . Systems . . . . . .2 Observability . . . . . . . . . . . . . . . . . .6 Nonlinear Boundary Value Problem . . . . .

. The main thrust of our study is the following differential equation d¯ x = f (t. .1) where f : I × Ω → n . along with appropriate examples of practical importance. x2 . . 1 . An interesting concept of neural solution of ODE is also touched upon. 1. .1). x).1. xn ). f = f (t. . . The value x(t) ∈ Ω ⊂ n is sometimes referred as the state of the system described by the ordinary differential equation (ODE) . . isocline and solution curve of IVP. .1 Let Ω ⊂ given by First Order Ordinary Differential Equations n be open and let f : Ω → n be a mapping ( not necessarily linear) f (x1 . x2 . x2 . xn ) = (f1 (x1 .Chapter 1 Introduction To Ordinary Differential Equations In this chapter we begin by introducing the concept of solvability of initial value problems (IVP) corresponding to first order differential equations (ODE) in n . Towards the end. we recall standard methods of solving some classes of first order ODE. f2 (x1 .(1. x) dt (1. . A vector field may also depend on an additional parameter t.Eq. . direction field. . xn ). . The reader is also familiarized with difference equations and numerical solutions of ODE through the process of discretization. all related notions like orbit or trajectory. We also give. . . fn (x1 . . The set Ω is then called the state space or phase space. that is.1. xn )) We shall refer to f as vector field on Ω. x2 . I a subinterval of .

x(t. ¯ As x(t) needs to satisfy Eq.1.1).1) on . is the set {x(t.2 A function x(t) is said to be a solution of the IVP . x0 )) : t ∈ J} ⊂ I × Ω.1). we get t t0 d¯ x ds t ds = t0 f (s. In Chapter ¯ 3.1).1.1(b)) Definition 1.1).1. t f (s.1. x(s))ds ¯ This gives x(t) − x(t0 ) = ¯ ¯ and hence we have Eq.1.1.2) Proof : If x(t) is a solution of Eq.1. is the vector field (1. If t → f (t. t0 .Eq.1.1 Assume that f : I × Ω → n is continuous. (1.(1.1) is tangent to the vector field whereas the solution curve is tangent to the direction field at any point.1) with the solvability of the corresponding integral equation.1) is given by dx (1. (1. then x(t) is said to be a global solution of the IVP. this solution is sometimes referred as a local solution.Eq.3 The orbit or the trajectory of the ODE. x(t)) dt x(t0 ) = x0 (1. (1. x0 ) : t ∈ J} in the state space Ω. (1.4 The direction field of ODE .1.1.1.Eq.2).1 The Initial Value Problem (IVP) corresponding to Eq. (1.Eq. (1. if there exists an interval J ⊂ I. (1. Whereas the solution curve is the set ¯ {(t.1) only on J. t ∈ J ¯ (1. t] for a fixed t.1. (1. x)).1(a)) = f (t. (1.1. t0 .1. (1. ¯ Definition 1. Then we use the notation x(t. containing t0 such that x(t) is differentiable on J with x(t) ∈ Ω for all t ∈ J and x(t) satisfies Eq.1.1.Eq. x(s))ds ¯ t0 .) is defined on the entire line and x(t) satisfies ¯ Eq.2CHAPTER 1. x0 ) for x(t). INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Definition 1. we shall discuss the local and global solvability of IVP in detail.1. x(t) is a ¯ solution the IVP .(1. f (t. by direct integration on the interval ¯ [t0 .1.1) iff x(t) is the solution of the integral equation ¯ t x(t) = x0 + ¯ ¯ t0 f (s. it is desirable to indicate the dependence of the solution x(t) on its initial value x0 . Definition 1.1. x(s))ds.1).1. At times. The following proposition gives the equivalence of solvability of IVP .Eq. It is clear from the above definitions that the orbit of the ODE . Proposition 1. t0 . .

is called the set of ¯ equilibrium points of ODE . it begins to interact with its environment and also other species and consequently growth rate R(t) diminishes.1. (1.3(a)) dt N (t0 ) = N0 (1. FIRST ORDER ORDINARY DIFFERENTIAL EQUATIONS 3 Conversely.5 The set of points in Ω where f (t. An important index of such investigation is growth rate per unit time.1. t ∈ Although the above model may predict the population N (t) in the initial stages. x) = 0. if Eq. ¯ ¯ t f (s.1.(1. denoted by dN 1 (1.1. we get the following initial value problem dN = R(t0 )N (1. .1.1. x(s))ds = f (t.4b) This is known as logistic equation with a as the growth rate without environmental influences and c representing the effect of increase population density.Eq.1. It is clear that at these points the orbit or trajectory is a singleton.1.3(b)) The solution of the above IVP is given by N (t) = N0 exp(R(t − t0 )). x(t)) ¯ t0 Definition 1. If we assume the form of R(N ) as equal to a − cN (a.1.4a) (1. Example 1. R(t) = R(t0 ) = b − d where b and d are birth and death rates.2) is differentiable and ¯ differentiating this equation we get d d¯ x = dt dt Also we have x(t0 ) = x0 . c positive constant). then growth rate is constant and is given by. (1. x(t) given by Eq. we realise that no population can grow exponentially for ever.3) R(t) = N (t) dt If we only assume that the population of species changes due to birth and death.1). we get the following IVP dN dt N (t0 ) = N (a − cN ) = N0 (1.1.2) holds.1. respectively. as population grows sufficiently large.1 (Population Dynamics Model) We investigate the variation of the population N (t) of a species in a fixed time span. If the initial population N (t0 ) = N0 . In fact.1.

C) = m=1 [N (tm ) − Nd (tm )]2 (1.1). So the above solution has the representation N (t) = (1.1. C=a\c 8 6 Nt 4 2 No 2 4 6 8 10 t Figure 1. we can actually carry out the computation of a and C from the given data. C) between the known discrete population data Nd (tm ) and theoretical data N (tm ) : M φ(a.6) By using one of the algorithms of the unconstrained optimization techniques (refer Algorith 2. .1. C = a/c is called the carrying capacity of the environment.5) The solution curve of the above population model is as under.1: Growth of population N (t) The parameters a and C are usually not known but can be estimated by minimising the mean square deviation φ(a.3.1. One can easily solve the above ODE by the method of separation of variables and get a c N (t) =   a − N0 1 + c exp(−a(t − t0 )) N0 C (C − N0 ) exp(−a(t − t0 )) 1+ N0 It follows from the above representation that lim t→∞ N (t) = a/c = C (carrying capacity). INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS The equilibrium points are 0 and a/c.4CHAPTER 1.

1.1.1.1.2 Consider the following ODE dx = −tx dt A general solution of the above ODE is given by x(t) = ae− 2 t . a is any constant The solution passing through (t0 .8) The following graphics demonstrates the solution curve and the direction field of the above ODE.5 2 1. x) plane is given by x(t) = x0 e− 2 (t 1 2 1 2 5 (1.5 -2 -1 1 2 t Figure 1.5 1 0.2: Solution curves x(t) .1.7) −t2 ) 0 (1. FIRST ORDER ORDINARY DIFFERENTIAL EQUATIONS Example 1. xt 3. x0 ) in (t.1.5 3 2.

INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Figure 1.1.1. x) is constant. (1. x) = p in t − x plane. Example 1. will be of utmost importance to us. Thus the uniqueness theorem. As we have seen through Example 1. To sketch the direction field for a given ODE in 1-dimension. The family of isoclines is then the family of curves f (t.3: Direction field of the ODE If IVP corresponding to Eq.3 We wish to find the direction field of the following ODE dx dt = 2√t + x dx 1− dt 1+ Solving for dx .1. we get dt √ √ dx (1 + 2 t + x) = 2 t + x − 1 dt which gives √ 2 t+x−1 dx = √ =p dt 2 t+x+1 .1. the direction field for a given differential equation can be an effective way to obtain qualitative information about the behaviour of the system and hence it would be more appropriate to expand on this concept. x0 ) passing through (t0 .6 Isocline is a curve in t − x plane through the direction field along which p = f (t. x0 ). no two solution curves of the IVP can intersect. t0 . to be dealt with in the Chapter 3. a device called isocline can be used.6CHAPTER 1.1.1) has a unique solution x(t.2. Definition 1.

9) Thus isoclines are all lines of slope -1 with −1 ≤ p < 1. So we get the following graphics.4: Isoclines with solution curves 1. .2. x0 ) is given by √ √ (x − t) + t + x = (x0 − t0 ) + t0 + x0 (1. The solution of the ODE passing through (t0 . then we call it a linear ODE with constant coefficients.2.1.10) xt 4 p=0.33 1 p=0.2 1 -1 2 3 4 t -2 -3 p=-1 p=0 Figure 1.1.2) where A(t) ∈ n×n and b(t) ∈ n for all t.1. x) dt f (t. A linear ODE is called homogeneous if b(t) ≡ 0 and if A(t) = A (constant matrix). −1 ≤ p < 1 (1. x) = A(t)x + b(t) (1.1) The ODE is called linear if f has the form (1. CLASSIFICATION OF ODE So iscolines are given by √ 2 t+x−1 √ =p 2 t+x+1 1 1+p 2 1−p 2 7 That is t+x = .5 3 2 p=0.2. Linear ODE will be discussed in detail in Chapter 4.2 Classification of ODE dx = f (t.1.

From Eq. t0 .4). t0 ± nT. x0 ) = x(t.2. t0 . b − s).1) is called autonomous if f does not depend on t. x) = Ax + b(t). x) for some T > 0 (1. That is ODE has the form dx (1. A non-autonomous ODE .2.3) on the interval I = (a. b(t) = (0. 2 cos t). (1.3) satisfies the following property x(t.4) The smallest of such T is called its period.2. x) = f (t. sin t + t cos t) However.2. x) . f is periodic with period 2π but its solution (to be discussed in Chapter 4) is given by x(t) = (t − sin t. x0 ) for all t0 ∈ I This implies that the solution is completely determined by the initial state x 0 at ¯ t = 0.3) on the interval (a − s. Hence the trajectory x(t) of Eq.2. b). it follows that the form of the solution is not affected if we shift t0 to t0 ± nT.1 If x(t) is a solution of Eq.3) = f (x) dt For autonomous ODE we have the following proposition (refer Mattheij and Molenaar [9]). we have the following theorem regarding the periodic solution of the periodic ODE. then for any s ∈ . (1.2 If a solution of a periodic ODE is periodic. x) = f (t.1) is called periodic if f (t + T.2. 0. x0 ) It is not necessary that a solution of periodic ODE is periodic as we see in the following example.(1. Proof : Assume that f (t. Proposition 1.2. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS The ODE . n ∈ ℵ (set of natural numbers). (1. (1. Because x is periodic. x) will also be ¯ periodic with period S as well T. That is x (t ± nT. Proposition 1. x0 ) = x(t − t0 .2. (1. Example 1. That is f (t + S.2.Eq. its derivative x(t) = f (t. x(t + s) is a solution of Eq.8CHAPTER 1. x) has a period T and the solution x(t) has period ˙ S with S = T.2.2. Then it has the same period as the vector field.1 dx = dt 0 1 −1 0 x+ 0 2 cos t f (t.

x) 9 We choose n such that 0 < S − nT < T. For example. F = 0. So.2. However. Thus dS = S(−k + λF ) dt For the growth rate of fish.1.2. If the food resource of sharks (fish) is non-existent.(1. S(−k + λF )) x The equilibrium points are given by ˆ ˆ ˆ ˆ F = k/λ. the growth rate of 1 dS sharks is modelled as −k without the fish population and −k + λF with S dt fish population F.2 (Predator-Prey Model) Predator-prey model represents the interaction between two species in an environment. S = 0. Thus dF = F (a − cS) dt Thus.2.4) around the equilibrium point 2 (1. T periodicity of f implies that f (t + S − nT. x) = f (t. We shall have the discussion of the phase-plane analysis in the Chapter 5. if we linearize the system given by Eq.4(b)) . f (¯) = (F (a − cS). S). we note that it will decrease with the existence of sharks and will flourish on small plankton (floating organism in sea) without sharks.2.2. S = a/c. Example 1.4(a)) dt dS = S(−k + λF ) dt This reduces to the following autonomous system of ODE in x d¯ = f (¯) x dt where x = (F. a contradiction. Hence T = S and the solution x(t) has ¯ the same period as the vector field. we have. what is called the famous Lotka . we shall focus on sharks and small fish in sea. It implies that f is periodic with a period smaller than T. CLASSIFICATION OF ODE Also. sharks population exponentially decays and is increased with the existence of fish.Volterra model for the predatorprey system dF = F (a − cS) (1.

2.3 for solution analysis of this linear system).5 2 4 6 8 10 Ft 12 t St Figure 1.5) The solution of the IVP corresponding to Eq.10CHAPTER 1. We have 1 dS = −k + λF S dt This gives t S(t) ln = −k(t − t0 ) + λ F (τ )dτ (1.2.2. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS ˆ ˆ F = k/λ. The following graphw ics depict the fish and shark population in t − F/S plane.4) is obtained as follows.7) S(t0 t0 . the solution is periodic with period 2π = 2π(ak)−1/2 .2.2. (1. (1.3 for linearization) dx = Ax(t) dt ˆ ˆ x = (F − F . Thus.2.2.6b) 11 10.1: Fish and shark population with time The average population of the general predator-prey system given by Eq. S = a/c. (1.5 10 9.5) is given by aλ ˆ F0 sin wt S = S + S0 cos wt + cw cw ˆ F = F + F0 sin wt − S0 cos wt a (Refer Section 4.6a) (1. we get the following linear system (refer Section 2. S − S)   kc 0 −  λ  A =  aλ  0 c (1.

HIGHER ORDER ODE 11 In view of the information obtained from the linearized model. . we can assume that both S and F are periodic of period T = tf − f0 .3.2) . We define a vector x(t) with components xi (t). × → is a given mapping. .2. c c 1. . . x(t).1) where g : I × × × .3 Higher Order ODE Assume that x(t) is a scalar valued n−times continuously differentiable function dk x on an interval I ⊂ . dt We shall be interested in the following higher order ordinary differential equation x(n) (t) = g(t. x(1) (t). . . .1. S(tf ) = S(t0 ) and F (tf ) = F (t0 ) and hence Eq. i = 1.2. x(n−1) (t)) (1. That is.8(b)) So in predator-prey system. Then. the average population remains around the equilibrium point. (1. k = F (a − cS) − σ1 F = S(−k + λF ) − σ2 S = F (a − cS) = S(−k + λF ) = k + σ2 a a − σ1 = (dec c Hence it follows that the average population of predator is creases) and that of prey is k + σ2 k = (increases). It is also interesting to analyse the men’s influence on the above ecosystem if we fish both predator and prey.8(a)) S(τ )dτ = a/c t0 (1.2.3. 2 ≤ i ≤ n (1. . Let us denote the derivative k by x(k) (t). n by xi (t) = x(i−1) (t). we have dF dt dS dt This is equivalent to dF dt dS dt a = a − σ1 . . x1 (t) = x(t).3. no matter what the solution trajectory is.7) gives t0 +T 0 = −kT + λ This implies that 1 T and similarly 1 T t0 +T t0 +T F (τ )dτ t0 F (τ )dτ t0 = k/λ (1.

.. g(t. Example 1. . 0.1 (Mechanical Oscillations) A particle of mass m is attached by a spring to a fixed point as given in the following diagram..3. . . . x) by ¯ f (t. x1 . .. . 0 −a0 1 0 0 −a1 0 1 0 −a2 . 1 . . .3) we have dx1 = x2 (t) dt dx2 = x3 (t) dt .3. .12CHAPTER 1. xn )) ¯ Then the higher order ODE . (1. (1. + a0 (t)x(t) = b(t) (1. . .. .3.3. Then in view of Eq. .4) A(t) = . −an−1 b(t) = (0. .3.3) where ai (t)(0 ≤ i ≤ n − 1) and b(t) are given functions. ..2) .3. x) dt A linear higher order ODE has the form x(n) (t) + an−1 (t)x(n−1) (t) + .. x3 .Eq. . (1. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS and the vector field f (t. . . . 0 0        (1. . x) = (x2 . dxn = b(t) − a0 (t)x1 (t) − a1 (t)x2 (t) − · · · − an−1 (t)xn (t) dt This is equivalent to the following first order system dx = A(t)x(t) + b(t) dt where        0 0 .1) is equivalent to the following first order ODE dx = f (t. b(t)) The matrix A(t) is called the companion matrix. .Eq.

By equilibrium of forces we have mF = m By Hook’s Law. ω (1. we get the following second order equation modelling the spring problem d2 x dx λ +k + ω 2 x = F (t). The external force applied to the particle is F (t).3.3.9) (1.3.7) is equivalent to the following system of first order differential equations 0 1 dx x (1.6) 2 dt dt ma Case 1: No resistance and no external force (Harmonic Oscillator) d2 x λ + ω 2 x = 0. ω 2 = dt2 ma (1.8) = λ 0 dt − ma It is easy to see the solution of the above system is given by (refer Section 4.10) x(t) = −x0 ω sin ω(t − t0 ) + x0 cos ω(t − t0 ) ˙ .7) Eq.1.3.3.3.1: Mechanical oscillations We assume that spring obeys Hook’s law (tension is proportional to its extension) and resistance (damping) is proportional to the particle speed.3.3.3. HIGHER ORDER ODE 13 F(t) a x m Figure 1.4) x(t) = x0 cos ω(t − t0 ) + x0 sin ω(t − t0 ). we get T = dx d2 x + T + mk dt2 dt λx and hence a λ dx d2 x + x+k dt2 ma dt (1. (1. ω 2 = (1.5) F = Equivalently.

Eq.3: Solution curve of the linear system Case 2: Solution with resistance dx = dt (i) k 2 − 4ω 2 < 0 0 −ω 2 1 −k x(t) (1.14CHAPTER 1.5 1 -2 Figure 1. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Equivalently. we have x(t) = α cos(ωt + β) 2π This is a simple harmonic motion with period and amplitude α.3.3.8).2: Phase space of the linear system -5 2 1 t 0 0 x’t -1 5 -2 1 0.5 -1 0. (1.5 -1 xt Figure 1. ω is called ω its frequency.3. 2 1 -1 -0.5 0 -0.11) .3. We can easily draw plots of phase-space (normalized ω = 1) and also solution curve of the ODE .

3.4: Solution curve for Case2(i) The oscillation curves for the other cases are as under. (ii) (k 2 − 4ω 2 ) = 0 x(t) = A exp( −kt −kt ) + Bt exp( ) 2 2 xt 1 0.5 -1 -1.3.1.5: Solution surve for case2(ii) .5 -1 10 15 20 25 Figure 1.5 t 5 -0.5 20 30 40 50 60 Figure 1.4) x(t) −kt )(A cos bt + B sin bt) = exp( 2 √ 4ω 2 − k 2 b = 2 15 xt 1 0.3. HIGHER ORDER ODE The solution curve x(t) is a damped oscillation given by (refer Section 4.5 t 10 -0.

F (t)) The solution x(t) is given by x = xp + xc where xc is the solution of the homogeneous equation with F (t) = 0. The constant D is given by D = (ω 2 − β 2 )2 + k 2 β 2 and sin φ = 1/2 kβ ω2 − β 2 . F (t) = (0.16CHAPTER 1. cos φ = D D Thus the forced oscillations have the same time period as the applied force but with a phase change φ and modified amplitude F0 F0 = 2 − β 2 )2 + k 2 β 2 ]1/2 D [(ω .6: Solution curve for case2(iii) Case 3: Effect of resistance with external force dx(t) = dt 0 −ω 2 1 −k x + F (t). INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS (iii) k 2 − 4ω 2 > 0 x(t) = exp( −kt ) A1 ect + A2 e−ct 2 k 2 − 4ω 2 c = xt 35 30 25 20 15 10 5 1 2 3 4 5 6 t Figure 1. A particular solution (response function) xp (t) of the above system is given by xp (t) = F0 [cos (βt − φ)] D corresponding to the external force (input function) F (t) = F0 cos βt.3. We have seen that xc → 0 as t → ∞.

3. As k → 0 we have F0 xp → 2 → ∞ as ω → β.3.7: Input-Response curve Example 1. HIGHER ORDER ODE 17 Amplitude modification depends not only on the natural frequency and forcing frequency but also on the damping coefficient k.8: Satellite in orbit around the Earth . 4 2 Input t 14 Response 2 -2 4 6 8 10 12 Figure 1.3.2 (Satellite Problem) A satellite can be thought of a mass orbiting around the earth under inverse square law field. θ r u2 m u1 Figure 1.3.1. as we see in the following graph. (ω − β 2 ) For k = 0. the response is given by xp = F0 t sin wt 2ω which implies that the system resonates with the same frequency but with rapidly increasing magnitude.

3. Make the following change of variables: ˙ x1 = r − σ. θ(t) = ωt (σ. x2 = r.13) is a system of nonlinear ordinary diferential equations involving the forcing functions ( controls) u1 and u2 and can be written in the compact form as d¯ x (1.3.13) Eq. Equating forces in normal and tangential direction on the orbiting satellite.3. we get d2 r dθ − r(t)( 2 )2 dt2 dt d2 θ dθ dr r 2 +2 dt dt dt = − k + u1 (t) r2 (t) = u2 (t) This gives a pair of second order differential equations d2 r dt2 d2 θ dt2 dθ 2 k ) − 2 + u1 (t) dt r (t) 2 dθ dr u2 (t) = − + r(t) dt dt r(t) = r(t)( (1. then one can show that Eq. x3 = σ(θ − ωt). u(t) ∈ 2 x ¯ dt . INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS We assume that m = 1 and also the satellite has thrusting capacity with radial thrust u1 and tangential thrust u2 .18CHAPTER 1.3. (1.12b) ] If u1 = 0 = u2 .12) reduces to dx1 dt dx2 dt dx3 dt dx4 dt x3 ˙ x4 + ω + ωt. ω are constant and σ 3 ω 2 = k). (1.12) has a solution given by r(t) = σ. x(t) ∈ 4 . u). (1. x4 = σ(θ − ω) ˙ This gives r = x1 + σ.14) = f (¯.3. θ = σ σ = x2 = (x1 + σ)( = x4 = −2σ( x4 x2 u2 σ + ω) + σ (x1 + σ) (x1 + σ) x4 k + ω)2 − + u1 σ (x1 + σ)2 (1.12a) (1. θ = ˙ So Eq.3. r = x2 .3.

1(b)) then Eq. x2 . x4 . we can define orbit or trajectory and solution curve as discrete subsets of n and n+1 . i = 1. (1. u1 .4. u1 . . . For the scalar case n = 1.4. tk .4 Difference Equations In some models (as we shall see subsequently). 1. (1. DIFFERENCE EQUATIONS Here f is a vector function with components f1 .4. x4 . the state vector x(t) may not depend on t continuously. x4 .1) is a constant solution x such that x = fi (x). f2 .1(b)) is called the IVP corresponding to a difference equation. u1 . . x2 .Eq. i ∈ ℵ Eq. t2 . respectively in an analogous way. Ω ⊆ n . . (1.4.14) in terms of the control vector u(t) ∈ 2 . f4 given by f1 (x1 . x2 . i ∈ ℵ where Ai ∈ n×n and bi ∈ n . (1. f3 . A stationary point or equilibrium point of the difference equation Eq. Suppose there exists a sequence of vector fields fi (x) : Ω → n .4. xi ∈ Ω. u2 ) = x2 f1 (x1 . · · · } In such a situation we use difference quotients instead of differential quotients and that leads to difference equations. As for ODE.1(a)) If in addition x0 = z 0 (a given vector in n ) (1. x4 . x3 . Then the first order difference equation has the form xi+1 = fi (xi ). x(t) takes values at discrete set of points {t1 . x3 .4. 2 . u2 ) = (x1 + σ)( f1 (x1 . u2 ) = −2σ( x4 x2 u2 σ + ω) + σ (x1 + σ) (x1 + σ) x4 k + ω)2 − + u1 σ (x1 + σ)2 19 We shall be interested in the solution.1) is called linear if it has the form xi+1 (t) = Ai xi + bi .4. x3 .1(a)) . x2 . u2 ) = x4 f1 (x1 . . Rather. x(t) ∈ 4 of linearized equation corresponding to Eq.1. we have the linear difference equation xi+1 = ai xi + bi . u1 . i ∈ ℵ Its solution is given by i−1 xi = (Πi−1 (aj ))x0 + j=0 j=0 (Πi l=j+1 al )bj . ¯ (1. . (1. x3 .3.

. bi =   . . . with Ai = A for all i. λn with linearly independent eigenvectors c1 . Plugging this representation in Eq. . i ≥ 0 where  xi . .3) is given by n xm = i=1 di λm ci . c¯ ]. . (1. . xi+1−k ). . i (1. 1      (1. . . .4. · · · . .1         This is an eigenvalue problem.3)     Ai =       xi =  ¯ 0 . . . . . A linear k th order difference equation is given by k xi+1 = j=1 aij xi−j+1 + bi . . . So the solution to the IVP corresponding to Eq.4. . k. . . . dn ] is given by d = C −1 (x0 ). (1. . .4. This can be written as the first order system xi+1 = Ai xi + bi . . 2 . . . . .k−1 ai+k−1. . .3) is ¯ ¯ n completely given by Eq.k  ai+k−1. ··· ···  0 . .4. . . cn . .4. . .4) (1. . 0    ¯   . m = 1. i = k − 1. . . Here C = [c1 . we get xi+1 = Axi The solution of this system is of the form xi = r i c where scalar r and vector c are to be determined. . then the general solution of Eq. (1. 0 ai+k−1. λ2 .3) we get (A − rI)c = 0 (1. . 0 bi+k 0 . xi+k−1 1 .20CHAPTER 1.4. (1. . .2) For the homogeneous case bi = 0.5) If the initial value is x0 then d = [d1 . .4.5).4. . xi+1 . . If A has distinct eigenvalues λ1 . INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS A k th order difference equation in 1-dimension is given by xi+1 = gi (xi . c2 . .

. ( t = 1) t) = (1 − di )Ni (t).4. . .4.4.4.4. λ2 . . . If the initial population N 0 is known. . This is also based on discretization but uses neural network approach (refer Haykin [6]).1 (Discrete One Species Population Dynamics Model) We wish to measure population changes in a species with one year age distribution.4. which are used to solve ODE. . then Eq. am ] is a constant vector.4. . Then we have N0 (t + Ni+1 (t + t) = b0 N0 (t) + b1 N1 (t) + . (1.3) and hence the solution of the above equation is of the form M    A=   b1 0 1 − d1 0 Nm = i=1 a i λm φ i i and are eigenvalues (distinct) and linearly independent where eigenvectors of A. . . respectively (0 ≤ i ≤ M ). NM (t)). DISCRETIZATION 21 Example 1. . However.8) This is of the type given by Eq.6) becomes N (t + where  b0 1 − d0 0 . . φ2 .5. . we shall not be concerned with the stability and convergence aspects of these methods.6(b)) Let N (t) = (N0 (t). + bM NM (t). 0 ≤ i ≤ M − 1 (1. . Let Ni (t) denote the number of ith year old species and let bi . {λi }M i=1 {φi }M i=1 1. ··· bM 0 ··· ··· 1 − dM −1        (1. respectively. . . . . . λM . di the corresponding birth and death rates. the constant vector a is given by a = CN 0 where the nonsingular matrix C is given by C = [φ1 .5 Discretization In this section we shall describe some well-known one-step numerical methods.7) can be viewed as a difference equation of the form N m+1 = AN m (1. (1. 0 t) = AN (t) ··· ··· . . .6(a)) (1. φM ] It is to be noticed that the population of each age group grows and decays depending upon the sign of the eigenvalues λ1 . We also introduce to reader the concept of neural solution of ODE. . . N1 (t). a = [a1 . (1.7) If we denote N m = N (m t).1.4. Then Eq.

. x(s))ds by a general quadrature formula βj f (tij .5.4) to eliminate x(ti+1 ).5. (1. x) dt (1.5.5. ti + 1] we get ti+1 x(ti+1 ) = x(ti ) + ti f (s. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS It is interesting to observe that sometimes neural solution is a better way of approximating the solution of ODE.5) An important class of one-step numerical methods is Runge-Kutta method. we get j=1 m x(ti+1 ) = x(ti ) + h j=1 βj f (tij . ti+1 If the integral get ti f (s.6) Approximating the integral m ti f (s. x(s))ds (1. x(s))ds If f is approximated by its values at ti . Let us begin with scalar ordinary differential equation dx = f (t. x(ti ) + hf (ti . we get the Euler’s forward method x(ti+1 ) = x(ti ) + hf (ti .2) and Eq.4) 2 which is again implicit.7) . if f is approximated by its values at ti+1 we get Euler’s backward method x(ti+1 ) = x(ti ) + hf (ti+1 . x(ti )))] 2 (1.5.3) which has to be solved implicitly for xti+1 .5. x(ti )) + f (ti+1 .5.22CHAPTER 1. x) : I × → . Integrating the above equation on [ti . .1) corresponding to the scalar field f (t. (1. On the other hand. 0 ≤ i ≤ N. x(tij )) (1. t1 . h = xi+1 − xi (1. .5. one-step method. we get the Heun’s method (refer Mattheij and Molenaar[9]) 1 x(ti+1 ) = x(ti ) + h [f (ti . x(ti )) + f (ti+1 . Let t0 . Combining Eq. x(ti )). x(tij )). Consider the integral equation T x(T ) = x(t) + t ti+1 f (s.2) which is the so-called explicit. x(ti+1 ))] (1. x(s))ds is approximated by the trapezoidal rule. tN be a set of time points (called grid-points) wherein we would like to approximate the solution values x(ti ). x(ti+1 )) (1. . we 1 x(ti+1 ) = x(ti ) + h [f (ti .5.

wi denotes the weight from the input unit t to the hidden unit i.7) and Eq.10) 6 3 3 6 k1 k2 k3 k4 = f (ti . For a given input t.5.9) is explicit if rjl = 0.9) rjl kl ) l=1 kj = f (ti + ρj h.11) where N (t. x(s))ds to get m x(tij ) = x(ti ) + h l=1 rjl f (til .1. it is easy to get the following classical Runge-Kutta method from Eq. vi denotes the weight from .5.5. x(ti+1 ). A neural solution (refer Logaris et al [8]) xN (t) can be written as xN (t) = x0 + tN (t. r11 = 0 = r12 . x(ti ) + hk2 ) 2 2 = f (ti + h. DISCRETIZATION 23 where h = ti+1 − ti and tij = ti + ρj h(0 ≤ ρj ≤ 1) are nodes on the interval [ti . As x(tij ) are unknowns.5. x(ti ) + hk3 ) In case we have f (t. W ) (1. and scalars ki being replaced by vectors k i (1 ≤ i ≤ 4). (1. (1. W ) is the output of a feed forward neural network (refer Haykin [ 6 ]) with one input unit t and network weight vector W . we get the Runge-Kutta formula m x(ti+1 ) = x(ti ) + h j=1 βj k j m (1. r21 = 1. (1. we get the Heun’s formula given by Eq.5. k output of the network is N = i=1 vi σ(zi ) where zi = wi t − ui . ti+1 ].5. x(ti ) + h Eq.Kutta method. with scalars x(ti ).1). (1. we obtain a similar Runge . We now briefly discuss the concept of neural solution of the IVP associated with Eq. x(til )) (1. If ρ1 = 0. ρ2 = 1. x(ti + 1) being replaced by vectors x(ti ).5.5.8).9) 1 1 1 1 x(ti+1 ) = x(ti ) + h k1 + k2 + k3 + k4 (1.5. Also. x(ti )) 1 1 = f (ti + h.5. x) as a vector field. r22 = 0 and β1 = 1/2 = β2 . (1. (1.5. l ≥ j.8) Combining Eq. x(ti ) + hk1 ) 2 2 1 1 = f (ti + h. to find them we apply another quadrature formula for tij the integral ti f (s.5).

wi to get the neutral solution xN (t) given by Eq. (1. (1.11) i This methodology will be clear from the following example. < tm = tf .24CHAPTER 1. vi . it may not be exactly equal to f (tj . For this we discretize the interval [t0 . . we compare these solutions with the exact solution. Example 1. xN (tj )) dt 2 One can use any unconstrained optimization algorithms (refer Joshi and Kannan [ 7 ]) to minimize φ w.5.r. We first determine the network parameters wi .1) in some sense.5.1: Neural network with one input the hidden unit i to the output and ui denotes the bias of the unit i.1 Solve the IVP dx dt x(0) = x − x2 . Hence we train the parameter vector W in such a way that it minimizes the error function φ(W ) given as m φ(W ) = j=0 dxN |t=tj − f (tj .5.t.5. t ∈ [0. tf ] as t0 < ti < . .13) .5.5. 0 ≤ j ≤ m. 1] = 1 2 1 1 + e−t (1. As xN (t) is assumed to be a trial solution. wherein we compute numerical solutions by Euler’s and Runge-Kutta methods and also neural solution. the neural network parameters to obtain their optimal ∗ ∗ values u∗ . xN (tj )) at point tj . (1. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS w1 w2 x w3 uk k v1 v2 v3 N Figure 1. σ(z) is the sigmoid transfer function σ(z) = 1 1 + e−z and k is the number of hidden neurons.12) The exact solution of this ODE (Bernoulli’s differential equation) is given by x(t) = (Refer Section 6 of this chapter). Further. vi and ui in such a manner that xN (t) satisfies Eq.

1 and three hidden neurons with a single hidden layer network.69161 0.3.00002 0. u1 . We can compare the values of this solution at a discrete set of points with the ones obtained by Euler and Runge-Kutta method ( refer Conte and deBoor [3]).525.00069 0.00025 0. w2 .57444 0.73340 0. v1 .00163 0. The neural solution is given by Eq. v2 .66819 0.00129 0.71294 0. The neural solution is of the form xN (t) = 1 vi +t 2 1 + e(−wi t+ui ) i=1 3 (1. The following tables give the distinction between the three types of approximate solutions.00097 0.59913 0.1 ) to compute the optimal weight vector W . xN (tj ) dt 2 = j=1 v1 v2 v3 1 + + + u1 −w1 tj u2 −w2 tj u3 −w3 tj 2 1+e 1+e 1+e v1 v2 v3 + + tj 1 + eu1 −w1 tj 1 + eu2 −w2 tj 1 + eu3 −w3 tj eu2 −w2 tj v2 w2 eu3 −w3 tj v3 w3 eu1 −w1 tj v1 w1 + + u1 −w1 tj )2 u2 −w2 tj )2 (1 + e (1 + e (1 + eu3 −w3 tj )2 1 + 2 v1 1 + eu1 −w1 tj + v2 1 + eu2 −w2 tj v3 + 1 + eu3 −v3 tj tj − + + tj 2 2 We have used the steepest descent algorithm (refer to Algorithm 2.1.73106 0.00199 0.54983 0. w3 .71095 0.62315 0. u3 ) is computed in such a way that we minimize φ(W ) given by 3 φ(W ) = j=1 3 dxN (t) |t=tj − f (tj .5. We note that the neural solution is a continuous one.66948 0.59869 0.5.00010 0.68997 0. Table 1.64663 0.52498 0.00234 .00044 0.14) The weight vector W = (w1 . u2 .5.5.54994 0. v3 .62246 0.14). (1.1: Euler solution and actual solution Euler’s method Actual solution Absolute difference 0. 0. DISCRETIZATION 25 For neural solution we take three data points tj = 0.5.64566 0.57469 0. 0.

6.84916−8 Table 1.71095 0. b(t) and c(t) are continuous functions in a given interval with a(t) = 0.72783 0.59869 0.66819 0.003231 1.57444 0.73106 Actual solution 0.2) .73106 0.71095 0.6.64566 0.66819 0.5.26CHAPTER 1.64302 0.66819 0.59869 0.12147−8 1.15537−9 1.57383 0.59741 0.64566 0.68650 0.6 Techniques for Solving First Order Equations dx + b(t)x + c(t) = 0 dt The general form of a first order linear differential equation is a(t) (1.71095 0.52498 0.70743 0.00316 0.05919−9 5.00352 0.62048 0.73106 Absolute difference 1.68997 0.62246 0.3: Neural solution and actual solution Neural Network sol Actual sol Absolute Difference 0.59869 0.00198 0.00264 0.00062 0.34665−8 1.54983 0. Dividing by a(t).00013 0.66503 0.68997 0.57444 0.00011 0.30339−9 2.68997 0.28708−9 9.52511 0.00348 0.54983 0.1) where a(t).52498 0. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Table 1.2: Runge-Kutta solution and actual solution Runge Kutta method 0.5. we get the equivalent equation in normal form dx + P (t)x = Q(t) dt (1.00127 0.62246 0.64162−9 4.59648−9 7.54973 0.54983 0.52498 0.57444 0.62246 0.64566 0.58997−8 1.

Since M and N are continuously differentiable in D. (1. x)) Hence. x)dx = 0 or M (t. Theorem 1. we have M (t.5) holds.5) = d(u(t.6. e Q(t)dt + c (1.6. it follows ∂t ∂t ∂x ∂ 2u ∂M ∂N ∂2u = and hence = in D. (1. x)dt + N (t. (1.2) by e to give us the solution x(t) = e− R P (t)dt R P (t)dt R P (t)dt 27 and integrating. . x)dx = 0 is exact iff Proof : ∂M ∂N = in D ∂x ∂t (1.4). ∂t If Eq.1.6. x) + N (t.1 If M and N are continuously differentiable function of (t. x) such that ∂u = M.6. let Eq.6) and If the differential equation is exact.6. we have ∂ ∂u ∂M = ∂x ∂x ∂t ∂ ∂u ∂N = .3) Exact Equations The differential equation M (t. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS This equation is solved by multiplying Eq. (1. x)dt + N (t. x)dt + N (t. x) = C of the exact equation given by Eq.4) is called exact if there is a differentiable function u(t. The following theorem gives a test for the exactness of a differential equation.6) holds. x)dx = ∂u ∂u dt + dx ∂t ∂x ∂u =N ∂x (1. that ∂x∂t ∂t∂x ∂x ∂t Conversly. intergrating the above wquation we get the solution u(t.6. x) ∈ D ⊆ 2 ( with no hole in D).6.6. then the differential equation M (t. x) dx dt = 0 = M (t.6. x) + N (t. x) dt dx (1.6.

x) ∂x u(t. u(t.x) (t+ ∆t . This gives (t+ t.1: A curve in τ − y plane We explicitily define the function u(t.x) (t 0 . x).6.6. x0 ) Then. x)] = M (t. x) − u(t. x) = u0 + (t0 .x0 ) [M (τ. x) − u(t. y)dτ + N (τ. x) = lim [M (τ1 . y)dy] M (τ. y)dτ + N (τ. Hence ∂u u(t + = lim t→0 ∂t t t.x) = M (t. x) is given by (t.x) u(t + t. It can be shown that this line integral is independent of path if Eq.28CHAPTER 1.6) holds.x) t+ t [M (τ. x) t→0 t . x) t. t < τ1 < t + (by meanvalue theorem for integrals). x). (1. y)dy] where the integral on the RHS is a line integral along a path joining (t0 .x 0) τ Figure 1. y)dτ (as y = x and dy = 0) t = = M (τ1 . x0 ) and (t. x) = (t. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS y (t. x) as the solution of the initial value problem ∂u ∂t u(t0 . = u0 ∂u = N (t.

= ( − a) b dt b dt (1. b ∈ dt We introduce the change of variable X = at + bx.8) .7) and hence Eq. x). we get xdy − ydx = dx x2 + y 2 which is equivalent to x y dy − 2 dx 2 x x = dx y2 1+ 2 x This is now an exact differential equation with solution y tan−1 ( ) = x + c x Change of Variables There are a few common types of equations.1. wherein substitutions suggest themselves. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS Similarly. x)dt + N (t.6. For example. which gives x= 1 dx 1 dX (X − at). [B] dx x = f( ) dt t (1. Such a function is called an integrating factor. xdy − ydx = (x2 + y 2 )dx is not exact. x) ∂x 29 It may happen that the equation M (t.6.6. as we see below [A] dx = f (at + bx). it becomes exact. But dividing by (x2 + y 2 ).7) becomes 1 dX ( − a) = f (X) b dt dX or = a + bf (X) dt which can be easily solved.6. x)dx = 0 is not exact but after multiplying both sides of this equation by a function µ(t. we have ∂u = N (t. (1. a.

(1. a. In that case we have cx + dy = m(ax + by) and hence Eq.8) becomes u+t or du dt du dt which can be easily solved. [C] Consider the equation dx =f dt at + bx + p ct + dx + q . Then introduce the change of variable .30CHAPTER 1. Then Eq. Assume that m = 1. ∈ (1. Then dX dx = where we choose h and k such that dt dt ah + bk + p = 0 = ch + dk + q Then dX =f dT at + bX ct + dX which is of type [B]. t x .10) can always be solved for h. c. q.11) in which the exponent is not necessarily an integer. (1. is called Bernoulli’s equation.6. p.6. (1. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS in which RHS depends only on the ratio Introduce the change of variable u = x = ut. k except when ad − bc = 0. Thus t dx du =u+t dt dt in which RHS depends only on the ratio of the linear expression.6. b.6.10) Eq.6.6. X = x − k.9) = f (u) = f (u) − u t x . d. [D] The equation dx + P (t)x = Q(t)xm dt (1. (1.9) becomes dx at + bx + p =f dt m(at + bx) + q which is of the type [A]. We substitute T = t − h.

x) = x − t2 .6. x) as (t.1. 0) to (t.x) u(t. x) = u0 + (t0 . TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS X = x1−m . C2 .0) Figure 1. N (t.6.x0 ) (τ − y 2 )dτ + (τ + y 2 )dy y (t.6. x) C2 τ C1 (0. Then x = X ( 1−m ) .11) becomes 1 31 m dx dX 1 = X ( 1−m ) .2: Path from (0.6. (1. Example 1. x) In the integrand on RHS. dt 1−m dt m m 1 1 dX X ( 1−m ) + P (t)X ( 1−m ) = Q(t)X ( 1−m ) 1−m dt Equivalently dX + (1 − m)P (t)X = (1 − m)Q(t) dt which is a linear differential equation in X. x) = t + x2 ∂M ∂N =1= ∂x ∂t and hence this equation is exact. Hence the integral become t x t3 x3 (−τ 2 )dτ + (t + y 2 )dy = − + tx + 3 3 0 0 . Hence Eq.1 Solve (t + x2 )dx + (x − t2 )dt = 0 M (t. we go along the path C1 . So we write the solution u(t.

6.32CHAPTER 1. x = dX T +X X − 1 and = . k = −1. the solution is x3 −t2 + tx + =C 3 3 dx t+x−1 = dt t + 4x + 2 Example 1.3 Solve dx + tx = t3 x4 dt = C = C = C 1 − 4U 2 dU = dT 1 + 4U or or This is a Bernoulli’s equation. We make the change of variable X = x −3 and get an equivalent linear differential equation dx − 3tX = −3t3 dt The general solution of this equation is given by X(t) = e3 R tdt (3t3 )e− R 3tdt +C = e2t 3 2 (−3t3 )e− 2 t dt + c 3 2 .6. This gives h = 2. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Hence. Hence t = T + 2. to get T which is equivalent to 1 + 4U dT dU = 2 1 − 4U T This yields (1 + 2U )(1 − 2U )3 T 4 (T + 2X)(T − 2X)3 (t + 2x)(t − 2x − 4)3 Example 1.2 Solve Solve h + k − 1 = 0 = h + 4k + 2. dT T + 4X We now make substitution X = T U.

1.6. 2tdt = du. This gives I = = = = = and hence X(t) = e− = Equivalently x= 2 3 2 3 2 e− 2 t t2 + e− 2 t +c 3 2 3 2 t2 + + ce 2 t 3 3t2 2 3 2 33 3 3 e− 2 u − udu 2 3 2 −2u 3 u+ − − e 2 3 3 2 3 − − e− 2 u u − 2 3 3 2 3 e − 2 u u + e− 2 u 3 3 2 3 2 − 2 t2 2 e t + e− 2 t 3 3 2 e− 2 u du 3 2 2 −3u e 2 33 1 = X3 1 t2 + 2 3 + ce 2 t 3 2 3 Example 1. Burghe and Borrie[2] and Haberman[5].4 Solve the IVP dx 1 = x − x2 . For more on real life problems giving rise to mathematical models generated by ordinary differential equations. put t2 = u. refer to Braun et al[1].6. X(0) = 2 dt The unique solution X(t) is given by X(t) = e−t = e−t + 1 This gives x(t) = 1 1 + e−t et dt + t For more on techniques for solving first order differential equation refer Golomb and Shanks[4]. . TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS To compute I = (−3t3 )e− 2 t dt + c. x(0) = dt 2 1 We make the change of variable X = and get an equivalent linear equation x dX + X = 1.

INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS 1. x) = (0. 0) and (x0 . x) = (0. wherever possible. t ≥ 0 given by = (x2 + 1) cos at = x1 + x2 (a) Find the stationary points. y0 ) = (0. (t. (b) Solve the above equation. x(t))  2tx  . Consider the differential equation dx = (x2 − 1)tp . 3. 0)  2 t + x2 =   0. Consider the differential equation dx dt f (t. Consider the linear system in dx1 dt dx2 dt (b) Sketch the vector field in 2 p ≥ 0.7 Exercises dx dt 2 1. 0) (a) Sketch the direction field of the above equation. 0). Solve following initial value problems. 2).34CHAPTER 1. . 4. 5. x) = f (t. (t. dt (a) Determine the equilibrium points. 0) and (0. (0. Sketch the direction field for the following differential equations. (c) Find the number of solutions passing through (0. (a) (b) t = x2 + t2 dx dt dx dt 2 dx −x = dt −1 2 (c) x2 − (1 − t2 ) =0 Sketch the solution x(t). 2. (b) Sketch the vector field. 2 . (c) Find solution passing through (0. −2).

x(0) = (c) 2(1 − t2 ) dt 3 (a) t2 6. Using the line integrals. In a model of epidemics. dt dx + x = t. Find all curves with the property that the angle of inclination of its tangent line at any point is three times the inclination of the radius vector to that point from the orgin. EXERCISES dx + x(t + 1) = t. x(−1) = −2. . Let x(t) denote the number of uninfected individuals in the population at time t. x(1) = 1 t2 + x 2 t + x2 8. 10. then x(t) decreases from x(0) = n to x(0) = 0.7. (b) xlog(t2 + x2 ) + 2t2 x + x dx + t2 + x 2 2t2 x t + log(t2 + x2 ) + 2 + t dx = 0 t + x2 35 (c) 1 (t + x)2 dt − − 1 dx = 0 2 1 + (t + x) 1 + (t + x)2 7. π (a) [2t + exp(t) sin x] dt + exp(t) cos xdx = 0. x(−1) = −1 (b) (1 + t2 ) tan−1 t dt a dx − 2tx = a(t3 − t). If we assume that the infection spreads to all those susceptible. x(0) = 2 π π (b) (cos t cos x) dx − (sin t sin x) dx = 0. Solve the following Bernoulli equations. dx + x = tx3/2 dt dx (b) 2tx3 + t4 − x4 = 0 dt dx (c) 3 + t sec x = x4 (1 − sin t)2 dt (a) 9. x( ) = 2 3 (t + x) t−x (c) dt − 2 dx = 0. θ). solve the following exact equations which pass through the given point. a single infected individual is introduced in to a community containing n individuals susceptible to the desease. Give a mathematical model for the above problem and solve it. solve them. (a) sin t sin2 xdt − (cos t cos 2x tan x + cos t tan x) dx = 0. Test if the following equations are exact. Find the fimily of curves orthogonal to the family of cardiods r = λ(1 + cos θ) in polar cordinates (r.1. If so. 11.

Springer Verlag. I. John Wiley and Sons. Elementary Numerical Analysis.M Mattheij and J Molenaar. Differential Equations Model.Theory and Practice. 2004. N. Neural Networks. 1994. 1965. Neural Networks. 1996. Ellis Horwood Ltd. Optimization . Fotiadis. 1981. Burghe and M S Borrie. [8 ] I. Ordinary Differential Equations in Theory and Practice. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS References [1 ] Martin Braun.. 1998. [2 ] D. [3 ] S. [4 ] Michael Golomb and Merrill Shanks. A. 1983. Logaris. Boston. Conte and C deBoor. Courtney S Coleman and Donald Drew. McGraw Hill Book Co. Artificial Neural Networks for Solving Ordinary and Partial Differential Equations. 987-1000. 9(5).An Algorithmic Approach. Population Dynamics and Traffic Flow. McGraw Hill Book Company. 1977. Prentice Hall. Likas and D. Mathematical Models: Mechanical Vibrations. Narosa Publishing House Pvt. D. [5 ] Richard Haberman. [9 ] R.36CHAPTER 1. 1981. E.M. Macmillan College Publishing Company. [6 ] S. Haykin.. IEEE Trans. Elements of Ordinary Differneial Equations. . Modelling with Differential Equation. [7 ] Mohan C Joshi and Kannan S Moudgalya. Ltd.

This approach is essential for introducing Green’s function in Chapter 7. z ∈ X (a) (x + y) + z = x + (y + z) (b) (x + y) = (y + x) (d) ∃ − x ∈ X such that x + (−x) = 0 2. y.Lipscntiz continuity. X is said to be a vector space if the following holds. Bessel’s inequality. monotonicity. differentiability and compactness .1 Finite and Infinite Dimensional Spaces Definition 2. 1. Bounded linear and nonlinear operators in function spaces along with their much needed properties . For any x. y ∈ X (a) α(x + y) = αx + αy (b) (α + β)x = αx + βx (c) α(βx) = (αβ)x (d) 1x = x 37 (c) ∃ an element 0 ∈ X such that x + 0 = x .Chapter 2 Mathematical Preliminaries This chapter is meant for laying the foundation for abstract concepts needed for our subsequent discussion of ODE in function spaces. For any α.1. x ∈ X). 2. We close by defining Dirac-delta function as a distribution function. β ∈ and x. y ∈ X) and scalar multiplication αx (α ∈ . orthormal system and Fourier series.1 Let X be a nonempty set which is closed with respect to binary operation x + y (x. We start with vector spaces and go on to define normed and inner product spaces and then dwell on notions like Schwartz inequality.are given a fair treatment. Parseval’s equality.

1. (x. This is denoted by x ⊥ y. y. Such a set is called a metric space. 1. z). y) satisfies the following properties. y) + d(y. this metric d(x. without the vector space structure. x ≥ 0 for all x ∈ X and x = 0 iff x = 0 αx = |α| x for all x ∈ X and α ∈ x + y ≤ x + y for all x. In a metric space.1. y) = 0 iff x = y [P2 ] d(x. . z) + β(y. z) = α(x.2 An inner product (x. x) for all x. x). [P1 ] d(x.1. Definition 2. Definition 2. Definition 2. y ∈ X 3. x is said to be orthogonal to y if (x. y) = 0. In an inner product space X. {xk } is said to be Cauchy if d(xk . x ) → as k. 3. z) ≤ d(x. (x. which satisfies the properties [P1 ] − [P3 ].38 CHAPTER 2. y) = x − y ∀ x. l → ∞. y) = (y. y) in a vector space X is a function on X × X with values in such that the following holds. x) for all x. the induced norm is defined by = (x. y) = d(y. as we see below..5 A sequence {xk } in a metrix space (X. we can easily define the concepts of convergence. y ∈ X and d(x.3.4 A metric space X is a non-empty set with a metrix d : X × X → defined on it. x) → 0 as k → ∞. MATHEMATICAL PRELIMINARIES Definition 2. the induced metric d(x. x) ≥ 0 forall x ∈ X and equality holds iff x = ¯ 0 2. y ∈ X x 2 In an inner product space. z) for all x. 1. z ∈ X It is possible to define a metric d(x.3 A vector space X is said to be a normed space if there exists a function x from X to such that the following properties hold. Cauchy convergence.1. y ∈ X In view of Definition 2. d) is said to be convergent to x ∈ X if d(xk . x ∈ X In a normed space X. y ∈ X [P3 ] d(x. z ∈ X The vector space X with an inner product defined is called inner product space. completeness etc. y.1. x. y) ≥ 0 for all x. (αx + βy. y) (distance between two vectors) is defined as d(x. 2. α. β ∈ . y) satisfying properties [P1 ] − [P3 ] in any set without having a vector space structure.

y as column vectors and x . Then S is said to be . A complete normed space is called a Banach space.d. If the set S consists of infinite elements of the spaces X. y) = y x = x y if we treat x. Otherwise.7 A set S of vectors {¯1 . L[S] = {α1 x1 + α2 x2 + · · · + αn xn .1(b)) In terms of the matrix notation we can represent (x. and 2.1. We say ∞ n that s = i=1 xi iff the sequence of partial sums sn = i=1 xi ∈ X.1(b)) and hence it is a Hilbert space. .i. . . y) as (x. xn ) 1 2 n (x.8 A set of vectors S = {x1 .6 A metrix space (X.1. Example 2. y) = x1 y1 + x2 y2 + . .1. . y = (y1 . this set is called linearly independent ( . . d) is said to be complete if everyCauchy sequence in (X. One can show that is complete with respect to the norm induced by the inner product defined by Eq.1. FINITE AND INFINITE DIMENSIONAL SPACES 39 Definition 2.i.1.9 A vector space X is said to be finite dimensional if it has a basis consisting of finite number of elements and this number is called its dimension and is denoted by dim X. x2 . y2 . Definition 2. . . x = (x1 . + x2 . Definition 2. x2 .1. . . A vector space X which is not finite dimensional. x2 . .1 In the space product are defined as x 2 n of n-tuples. αi ∈ equals X. . xn } in a vector space X is x ¯ ¯ said to be linearly dependent (l.i. ∞ In a normed space X. converges to s ∈ X. yn ) (2. .1. y n as row vectors. . whereas a complete inner product space is called a Hilbert space. called the linear span of S.i. . xn yn .1. it is possible to define infinite sum i=1 xi .). .) if ∃ scalars αi ∈ (not all zero) such that α1 x1 + α2 x2 + . . . . The number of elements in a basis is unique (refer Bapat[1]). xn } in a vector space X is said to be a basis for X if 1. . x2 . S is . . . is called infinite dimensional space. d) converges. x = (x1 . Definition 2. xn ). (2. . the Euclidean norm and inner (2. + αn xn = 0. if every finite subset of S is . }.1.1(a)) = x2 + x2 + .2. · · · .

1 It is possible to define other norms in the space below x 1 = |x1 | + |x2 | + .2 0.t.4 0.1.2 In the space C[0.1.t.40 CHAPTER 2. as we see = max1≤i≤n {|xi |} Definition 2.2(b)) One can show that the space C[0. MATHEMATICAL PRELIMINARIES n Remark 2. x 1 and x ∞ .8 0.2 f1 f2 f3 f4 0. 1] w.1.10 Two different norms x a and x b in a normed space X are said to be equivalent if ∃ positive constants α. Example 2. To see this consider the sequence {f n (t)} of continuous functions on [0. 1] defined as fn (t) = tn . + |xn | x ∞ . 1] 1 0.1.4 0. m → ∞ . the above norm is not complete and hence it is not a Banach space.1. one can define norm and inner product as under f 2 2 1 = 0 1 f 2 (t)dt (2.8 1 Figure 2.6 0. . 1] of all real valued continuous functions.r. β such that α x a ≤ x b ≤β x a One can show that all norms in n are equivalent (refer Limaye[8]) and hence n is also complete w.1.2(a)) (f. t ∈ [0.6 0. g) = 0 f (t)g(t)dt (2. defined earlier.1: A sequence which is Cauchy but not complete fn − fm 2 2 1 = 0 (tn − tm )2 dt → 0 as n. .r.

5) Equality holds iff x and y are l.1 A normed space X is an inner product space iff the norm in X satisfies the parallelogram law given by Eq.1.r. b) and L2 [a. then C[0. ¯ . b] denote the space of all square integrable functions. Example 2. 1] Take f = t. Example 2.2. y. Refer Goffman and Pedrick [5] for the proof of the above Theorem. 1] is complete with respect to f ∞ norm and hence it is a Banach space.4 X = C[0. (2. y)| ≤ x y (2. Parallelogram law: x+y 2 2 = x 2 + y 2 (2.1. b] → .1.1. b] is a Hilbert space. we have the following identities.1.3 Let L2 [a. b L2 [a. 1] with f = sup |f (t)|.d.1] |f (t)|. t = 1 41 which is not a continuous function and hence the space is C[0.1. FINITE AND INFINITE DIMENSIONAL SPACES and hence {fn } is Cauchy but fn (t) converges to f (t) defined as f (t) = 1. One can check that parallegram law is not satisfied and hence the Banach space C[0. b] = {f : [a.1.1.1. Remark 2. Then for all x. t = 1 0. where as C[a. we have the following inequality |(x.1. (2. b] are infinite dimensional spaces In an inner product space. g = 1.4) Parallelogram law is relevant in view of the following theorem.3) + x−y 2 = 2[ x 2 + y 2] (2. Then L2 [a.2 n is finite dimensional. However.4). 1. Pythagorus identity: x ⊥ y iff x + y 2. Theorem 2. 1] with respect to the above norm can never be made a Hilbert space. a f 2 (t)dt < ∞} Let the norm and inner product be defined as in Eq. if f ∞ is defined as f ∞ = supt∈[0.t.1. Theorem 2. the above norm. f ∈ X t∈(0.2).2 (Schwartz inequality) Let X be an inner product space. ∈ X. 1] is not complete w.1] t ∈ [0.

{yn } is Cauchy and since M is a closed subspace of X. we wish to obtain z ∈ M which is closest to M. MATHEMATICAL PRELIMINARIES Proof : If any of the elements x.d.1. assume that x. n→∞ x − yn = d . y are l. We have 0 ≤ (x − (x. e)e) = (x. Let M be a closed subspace of a Hilbert space X. It is clear that x − z = lim Uniqueness of z is easy to prove. it suffices to show Eq. On the other hand if x. y)y.1. This gives us the following theorem. Theorem 2. (2. (2. Given an element x ∈ X.6). we are through. In this theorem M ⊥ represent the spaces of all elements orthogonal to M. e)e. Also. e)2 This gives Eq. we have yn − y m = (¯n − x) − (y m − x) 2 y = [2 yn − x 2 + 2 ym − x 2 2 2 ≤ [2 yn − x + 2 ym − x → 2d2 + 2d2 − 4d2 = 0 + − 2x + y n + y m 2 ] − 4d2 ] That is.5) is true. So. it is complete and hence y n → z ∈ M. By normalizing y as y/ y = e. Then ∃ a unique element z ∈ M such that x − z = inf y∈M (2.3 Suppose x ∈ X and M a closed subspace of X.5). x − (x.d.1. y = 0. y)y = 0 x−y Proof : Let d = inf y∈M x − y . the above inequality reduces to ¯ |(x. then Eq. x) − (x. (2. e)| ≤ x for all x ∈ X So. x − (x. we get (x − (x.1.5). Then ∃ a sequence {¯n } ∈ M such that y x − yn → d By parallelogram law.6) x − (x. y are l.1.1. if equality holds in Eq. (2.. y is zero.42 CHAPTER 2. We have the following theorem in this direction. y)y) = 0 This implies that and hence x. This proves the ¯ ¯ theorem.

eα )eα ⊥ eβ ∀ β ∈ A. Theorem 2. Definition 2.1. x ⊥ eα for every α ⇒ x = 0. Then. we have α∈A |(x. eβ ) = 0. The inequality Eq. eα )eα .7) and further x − ¯ α∈A (x.6 Let X be a Hilbert space and let S = {eα }α∈A be an orthonormal set in X. That is. w ∈ M ⊥ This representation is written as X = M ⊕ M⊥ X is called the direct sum of M and M ⊥ . α∈A = α∈A (x. α = β and eα = 1 ∀ eα ∈ S. z ∈ M. in turn.4 Let M be a closed subspace of X.1. We have the following property for orthonormal sets in X.1. Then z ∈ M is the point of minimum distance from x ∈ X iff x − z ∈ M ⊥ . The following theorem presents the most important property of a complete orthonormal set. implies that every x ∈ X has a unique representation x = z + w.1.11 A set S ⊂ X is called an orthonormal set if eα . Then the following are equivalent in X. (2.1. FINITE AND INFINITE DIMENSIONAL SPACES 43 Theorem 2. eα )|2 (2. (3) Every vector x ∈ X has Fourier series expansion x = (4) Every vector x ∈ X satisfies the Parseval’s equality x 2 (x.5 Let {eα }α∈A be a collection of orthonormal elements in X. Then S is called a basis for X (or a complete orthonormal system) if no other orthonormal set contains S as a proper set. Definition 2.1.2. (1) S = {eα }α∈A is complete. This.1. eα )|2 ≤ x 2 (2.8) .12 Let S be an orthonormal set in a Hilbert space X. (2) 0 is only vector which is orthogonal to every eα ∈ S. Theorem 2.7) is referred as Bessel’s inequality.1. eβ ∈ S ⇒ (eα . for all x ∈ X.1.

5 In X = L2 [−π. eα )eα is orthogonal to eβ for every β. MATHEMATICAL PRELIMINARIES Proof : (1) ⇒ (2) Suppose (2) is not true. eα )|2 = 0 contradicting the fact that e is a unit vector. π π π π 2π is a complete orthonormal set and every f ∈ L2 [−π. then ∃ x = 0 such that x ⊥ eα ∀α ∈ A. then Parseval’s equation Eq. We also note that for any f ∈ L2 [−π. (2. . √ ... cos nt)2 + (f. π] will be proved in Chapter 7.. π] has a Fourier series expansion ∞ a0 f (t) = + [ak cos kt + bk sin kt] (2.1. we get an orthonormal set S ∪ {e} which properly contains S. eβ )eβ ) β∈A = α∈A (4) ⇒ (1) If S = {eα }α∈A is not complete. sin nt)2 ] 2π π n=1 ∞ . π] we have the Parseval’s relation π π f 2 (t)dt = (f. √ .1. This contradicts the completeness of S. Defining e = x/ x . π]. |(x. 1)2 1 + [(f. eα )eα (3) ⇒ (4) We have x 2 = (x.44 CHAPTER 2. it is properly contained in an orthonormal set ˆ ˆ S and let e ∈ S − S. .8) gives e 2 = α∈A |(e. (2) ⇒ (3) (x. k = 0. the collection of functions S= 1 cos t cos 2t sin t sin 2t √ . √ . −π The fact that the above set S forms an orthonormal basis in L2 [−π.1.9) 2 k=1 where ak = 1 π π f (t) cos ktdt −π and bk = 1 π π f (t) sin ktdt. eα )eα . ... Example 2. √ . it follows that it x− α∈A must be the zero vector 0 and hence x= α∈A (x. 1. eα )|2 (x. But by (2). 2.. x) = ( α∈A (x.

. y2 . x ¯ ¯ We first obtain orthogonal vectors {¯1 . i + 2. .5 1 2 3 t -1 Figure 2. π]. .1 y1 − αi. FINITE AND INFINITE DIMENSIONAL SPACES 45 The Fourier series representation given by Eq.i. e1 .i−1 yi−1 ¯ ¯ ¯ ¯ (¯i . x2 . (2. . yn }. . yj ) y ¯ i≤j ≤i−1 This is continued inductively for i + 1. n. . . . .1. . . . = xi − αi. xn ] = L[¯1 . . .9) will be used while discussing the solution of boundary value problems. . . .2 y2 .2) ∞ 4 sin(2n + 1)t f (t) = π n=0 2n + 1 ft 1 0.1. . . . yn ] x ¯ ¯ y ¯ ¯ .5 -3 -2 -1 -0.2: Sketch of first few terms of the Fourier series We now give the following Gram-Schmidt procedure which helps in producing an orthonormal collection {e1 . collection {¯1 . y2 . αi. . (¯j . .j = = x1 ¯ . . . . . It is clear that L[¯1 . x2 . f (t) = 1. en } out of the . t ≤ 0 we have the following Fourier series representation along with its graphics (Figure 2. . . . . For a piecewise continuous function f (t) defined on [−π. t ≥ 0 −1. xn }.1. yj ) x ¯ .1.2. inductively y ¯ ¯ y1 ¯ yi ¯ where αi. .

p1 (t) = t = t2 − α0 p0 (t) − α1 p1 (t). α1 = 0. t) =0 (t. 1 ≤ i ≤ 4. α2 = 0 (t. MATHEMATICAL PRELIMINARIES for each n. α0 = α1 = and hence p2 (t) p3 (t) = t2 − 1 3 (t3 . α0 = α1 = (t3 . (1.1. . The higher degree Legendre polynomials are inductively obtained. 1) 3 = t3 − α0 p0 (t) − α1 p1 (t) − α2 p2 (t). 8 . P1 (t). (1. en }. p2 . 1]. Normalizing yi . 2 2 1 4 2 P4 (t) = (35t − 30t + 3).46 CHAPTER 2. which are obtained by GramSchmidt orthonormalisation procedure. . p2 (t)) 7 6 1 = 0 = t 4 − t2 − 7 5 and hence p4 (t) Normalising these polynomials we get the Legendre polynomials P i (t). Computation for p0 . ¯ ¯ ¯ Example 2. we get the orthonormal collection {¯. t) (t2 . α3 = 0 α2 = (p2 (t). p3 . 1) 1 = . the set of all polynomials {P0 (t). . p1 . 1) (t3 . is an orthonormal basis for L 2 [−1.}. 1) 1 α0 = = . t) 5 and hence p3 (t) p4 (t) 3 = t3 − t 5 = t4 − α0 p0 (t) − α1 p1 (t) − α2 p2 (t) − α3 p3 (t). 1].6 In the space L2 [−1. 1 1 given by P0 (t) = 1. . P1 (t) = t. called Legendre polynomials. . 1) = 0. (t4 . . p4 is given below. . (1. p2 (t)) = . e2 . t) 3 = . 1) 5 4 6 (t . P3 (t) = (5t3 − 3t). p0 (t) p2 (x) = 1. ¯ e ¯ ¯ ei = y i / y i . P2 (t) = (3t2 − 1).

2. we have the Legendre series representation ∞ 1 47 f (t) = i=0 ai Pi (t). ∀α1 .2 Linear and Nonlinear Operators In this section. 2. the spaces under consideration will be normed spaces. T is called nonlinear. LINEAR AND NONLINEAR OPERATORS For f ∈ L2 [−1. unless otherwise stated. x2 ∈ X .1.2.1 T : X → Y is said to be linear if T (α1 x1 + α2 x2 ) = α1 T x1 + α2 T x2 . 1]. Pi ) = −1 f (t)Pi (t)dt.2 T : X → Y is said to be continuous at x ∈ X if xn → x in X ⇒ T xn → T x in Y T is said to be continuous on X if T is continuous at every x ∈ X. ft 4 2 -3 -2 -1 1 2 3 t -2 -4 Figure 2. ∀ x 1 .3: Sketch of First few of terms of Legendre series for sin πt Legendre polynomials will be used while discussing series solution of Legendre differential equation. i≥0 For f (t) = sin πt we have the following graphical representation of first few terms of the Legendre series. Definition 2. ai = (f.2. Definition 2. to be discussed in Chapter 6.2. α2 ∈ Otherwise.

. .2. Y ) is also normed space with T defined by T = sup{ Tx 0} . s) ∈ L 2 [0.2.x = ¯ x (2. s)f (s)ds. Example 2. . .2. . One can show that B(X. MATHEMATICAL PRELIMINARIES One can show that a linear operator T is continuous on X iff ∃ a constant k > 0 such that (2. xn ) ∈ T is obviously linear. 0 1 0 |k(t.2.1) is said to be a bounded operator. 1] is square integrable : 1 1 1 0 0 Assume that k : [0.48 CHAPTER 2.2) B(X) will denote the space of bounded linear operator from X into itself.1 Let X = [T x]i = j=1 n n and T : n → n be defined by a matrix (αij )n×n n αij xj . equipped with the Euclidean norm x 2 = i=1 x2 i . This inequality implies that 1 0 |T f (t)|2 dt ≤ ( 1 0 |k(t. f ∈ L2 [0. Y ) will denote the space of all bounded linear operators from X to Y. x = (x1 .2. we shall be encountering a linear operator of the form 1 [T f ](t) = 0 k(t. equipped with 1-norm. 1]).2 For a discussion on boundary value problems. (2. Also Tx ≤ max i 1  n j=1 Thus T is a bounded linear operator on n . s)f (s)|2 ds |k(t. 1] × [0.1) Tx ≤ k x ∀ x ∈ X A linear T satisfying Eq. B(X. 1] → We have |T f (t)|2 ≤ ≤ k 2 (t. s)dtds < ∞. But on all norms are equivalent and hence T is also a bounded linear operator on n 1/2 |αij | x  1 n n . s)|2 ds 1 0 |f 2 (s)|ds (by the application of Schwartz inequality to f (s) and k(. s)|2 dtds)( 1 0 |f 2 (s)|ds) . Example 2. .

then X is said to be reflexive. With this notion. Example 2.3 A one to one and onto linear mapping from X to Y is called isomorphism. T ∈ B(L2 [0. 1](lq). . . . en } be an orthonormal basis for n . . one can write j=1 f = f (¯1 )f1 + f(¯2 )f 2 + . To see this we proceed as follows. if f ∈ X ∗ . 1](l2 = l2 ). f n } ⊂ X ∗ as follows: f j (ei ) = δij = 0.2. One can show that X ⊆ X ∗∗ . s)|2 dtds. Elements of this space are called continuous linear functionals or simply functionals.2. + = 1. Then X ∗ = n (up to an isomorphism). 1] That is. 1] (lp) the space of pth power absolutely integrable functions (pth power absolutely summable sequences). i = j 1. 1] = L2 [0. The spaces X and Y are then called isomorphic. f 2 . ∗ we have L∗ [0.2. This space is called the conjugate space or dual space of X and is denoted by X ∗ . 2 . LINEAR AND NONLINEAR OPERATORS Denote by k 2 = 1 1 0 0 49 k(t.2. if X = X ∗∗ . Let S = {e1 . . ) of all continuous linear operators from X to will play a significant role throughout the solvability analysis of ordinary differential equations. it follows that every finite dimensional space of dimension n is isomorphic to n . Define a set S ∗ = {f 1 . The normed space B(X.2.i. As claimed before. Then we get Tf ≤ k f ∀ f ∈ L2 [0. .i.3 Let X = n . . If X and Y are finite dimensional spaces of the same dimension. In particular. 1 1 One can show that L∗ [0. (Refer Royden [10]). . So X ∗ is equal n (up to an isomorphism). 1]). + f(¯n )f n e ¯ e e Thus S ∗ is a basis for X ∗ and S ∗ consists of n . i = j It is clear that {f j }n are . Also. . e2 . X ∗ is a normed space with norm of a functional f ∈ X ∗ defined as f = sup |f (x)| x ≤1 We also have the second conjugate X ∗∗ = (X ∗ )∗ defined in a similar way. Definition 2. 1](lp∗ ) = Lq [0. Example 2. then one can define an isomorphism between X and Y by defining a linear operator which maps basis to basis.4 We denote by Lp [0. vectors. p p q From this result it follows that Lp (lp) spaces are reflexive spaces. . 1 < p < ∞. However.

T is said to be positive semi-definite if (T x. X) as : (x. Y ). y)Y ∀ x ∈ X.2.1 Let X and Y be Hilbert spaces and let T ∈ B(X. Reversing the above argument. MATHEMATICAL PRELIMINARIES Definition 2.1. Y ). Definition 2. T y) = 0 for ¯ all y ∈ X. Y ).4 we can write z ∈ X as follows: z = u + v.2. x ∈ R(T )⊥ and hence N (T ∗ ) ⊂ R(T )⊥ .2. if X is a Hilbert space. x ∈ X} We have the following relationship between N (T ∗ ) and R(T ) and vice-versa. By Theorem 2.1. Then N (T ∗ ) = R(T )⊥ and R(T ∗ )⊥ = N (T ). Y ) we define T ∗ ∈ B(Y. Let M = R(T ). This is done as follows by making use of Theorem 2.2. u ∈ R(T ) = M. Assume that x ∈ N (T ∗ ). That is. However.1 If R(T ) is closed. Theorem 2. x ∈ X One can show that T ∗ is also bounded and T ∗ = T . y) = 0 for all y ∈ X. we get the orthogonal projector P from X to M = R(T ). y ∈ Y. the subspaces N (T ) and R(T ) are defined as follows N (T ) = {x ∈ X : T x = 0}. of T as follows: [T ∗ f ][x] = f (T x). f ∈ X ∗ . This theorem can be extended to get the following result. then R(T ) = N (T ∗ )⊥ and R(T ∗ ) = N (T )⊥ . If T ∈ B(X. then T ∗ x = 0 and hence (T ∗ x. if T ∗ T is invertible. then corresponding to T ∈ B(X.50 CHAPTER 2. we get R(T ) = N (T ∗ )⊥ . v ∈ R(T )⊥ = N (T ∗ ) (2.4. R(T ∗ ) = N (T )⊥ For operators T ∈ B(X) with closed range in a Hilbert space. We say that T ∈ B(X) is self-adjoint if T ∗ = T. ¯ ¯ For T ∈ B(X. Theorem 2. We define the conjugate operator T ∗ : Y ∗ → X ∗ .5 Let X be Hilbert space and T ∈ B(X) be a self adjoint operator. R(T ) = {y ∈ Y : y = T x.4 Let T ∈ B(X. x) ≥ 0 for all x ∈ X ¯ ¯ ¯ T is called positive definite if the above inequality holds and is strict for x = 0. Proof : It is sufficient to prove the first one.3) . Y ). T ∗ y)X = (T x. Remark 2.2.2. one can show that R(T )⊥ ⊂ N (T ∗ ) and hence the first part of the result. This implies that (x.2 Let X and Y be Hilbert spaces.

2. . xn ) We assume that the observed data fk approximate the value f (x(k) ). The problem of data fitting is to recover the function f (x) from the given data fk .2. as we see below. m. we get n n ( i=1 j=1 αji yj )xi (Ax. We try to fit the data by a linear function F (x) given by F (¯) = a0 + a1 x1 + · · · + an xn x (2.6) . . Then A∗ = (βij ) where βij = αji . . .5) (Ax. .2.2. By). . y) = ( i=1 j=1 βij yj )xi = (x. B = (βij ) This implies that A∗ = B = (βij ) = (αji ) = A Example 2. x2 . LINEAR AND NONLINEAR OPERATORS v ∈ N (T ∗ ) implies that T ∗ (z − T x) = 0 and hence T ∗ z = T ∗ T x. (2. 2. . Let the first quantity be represented by variables x1 .5) gives the element u ∈ R(A) nearest to R(A) from z ∈ n . x = (x1 . .5 Let X = n and A be the matrix (αij ). giving the vector x ∈ n .2. reduces to the transpose A of the matrix A. xn . y) = i=1 j=1 n n αij xj yi αji xi yj j=1 i=1 n n = = As βij = αji .2. . then Eq. . This formulation can be used in linear data fitting model leading to the ¯ least square approximation. This gives x = (T ∗ T )−1 T ∗ z 51 (2. k = 1. .6 If X = n and A = (aij ) is a m × n matrix with (A A)m×m invertible. We have n n (2. Data fitting model is used in measuring relationship between two physical quantities. Example 2. That is A∗ . .2.2.4) The operator P : X → R(T ) = M is called the orthogonal projector and is given by P z = u = T x = [T (T ∗T )−1 T ∗ ]z This element u = T (T ∗T )−1 T ∗ z ∈ R(T ) = M minimizes the functional f (x) = z − T x in R(T ) = M. the second quantity by y and the functional relationship be given by y = f (x).

(1) (2) 1 x m xm · · · x(n) (n) x2 .r. a0 . 1]} dt ∗ D(T ) is called the space AC[0.5). . Remark 2. d2 . . f2 .2. Hence we minimize the error dk w. it is not a bounded sin nπt operator as it maps the sequence fn (t) = which converges to zero in n L2 [0. . .2 From the point of view of applications to differential equations. (2. . . . . .t. 1]. 1]. . with f˙ ∈ L2 [0. an ). a1 . fn ) ¯ ¯  1 x(1) x(2) · · ·  (1) (2) x2 ···  1 x2 A= . the 2-norm m ¯ d 2 = k=1 ¯ d2 . an ). 1] to a sequence T fn(t) = π cos nπt which is not convergent.52 CHAPTER 2. d on the whole space L2 [0. k = 1. However. It can be shown that D(T ) is dense in L2 [0. . . MATHEMATICAL PRELIMINARIES We shall compute parameters ak so that the deviations dk = fk − F (x(k) . xm (n)    . the vector u is given by u∗ = A(A A)−1 A z ¯ ¯ if (A A) is invertible. . . dm ) k From the linear model given by Eq.e. we have m ¯ d 2 = k=1 m (fk − F (¯(k) ))2 x n = k=1 (fk − a i xi ) 2 i=0 (k) Then the least square problem (data fitting problem) reduces to finding u ∗ = Aa∗ ¯ such that z − u∗ = infn z − Aa a∈ Let a = (a0 . .2. .  . it is not possible to define the linear operator T = dt In such a case we first define the domain D(T ) as follows: D(T ) = {f ∈ L2 [0. 1] : f˙ = df exists a. . 1].   By using Eq. d = (d1 . . . . . m are as small as possible. . z = (f1 . (2.2. a1 . ···  . . . . . . .6). 1] of all absolutely continuous functions on [0. .

T is a self-adjoint operator and is densely defined. y). y ∈ D(T ∗ ) T is said to be self-adjoint iff D(T ∗ ) = D(T ) and T ∗ = T. 1] with f (0) = f (1) = 0} One can show that D(T ) is dense in L2 [0. Then we define D(T ∗ ) to consist of y ∈ X for which the linear functional f (x) = (T x. T g) where T ∗ g = −¨ and D(T ∗ ) = {g ∈ X : g. 1] and T f (t) = −f(t) = − 2 . f ∈ D(T ) g f (t)¨(t)dt g 0 = − f (t)¨(t)dt g 0 if g(0) = g(1) = 0.2. Also one can show that T is closed. . g) = − 0 ¨ ˙ f(t)g(t)dt = −f(t)g(t) 1 + 0 1 ˙ ˙ f(t)g(t)dt 1 0 ˙ ˙ = −f(1)g(1) − f(1)g(0) + f (t)g(t) ˙ 0 − g f (t)¨(t)dt 0 1 0 ˙ ˙ = −f(1)g(1) + f(1)g(0) + f (1)g(1) − f (0)g(0) − ˙ ˙ ˙ ˙ = −f(1)g(1) + f(1)g(0) − 1 1 f (t)¨(t)dt. 1]} : f. y) = (x. In other words the adjoint T ∗ of T for densely defined operators is given by (T x.2. there exists a unique y ∗ ∈ X such that (T x. 1] and T xn → y in X. Hence (T f. ∀ x ∈ D(T ). So. y∗ ) ∀ x ∈ D(T ) The adjoint T ∗ of T is defined on D(T ∗ ) by T ∗ y = y ∗ . g(0) = g(1) = 0}. T ∗ y). x ∈ D(T ) ⊂ X is continuous in D(T ). where D(T ) is dt defined as under ˙ ¨ D(T ) = {f ∈ L2 [0.2.7 Let X = L2 [0. LINEAR AND NONLINEAR OPERATORS 53 Definition 2.6 Let X be a Hilbert space and T : D(T ) → X be such that D(T ) is dense in X. g) = (f. That is if xn → x in X = L2 [0. Then x ∈ D(T ) and y = T x. By Riesz representation theorem (refer Bose and Joshi [2]). y) = (x. g ˙ ¨ Thus T ∗ = T and D(T ) = D(T ∗ ).2. g ∈ L2 . f ∈ L2 [0. d2 f ¨ Example 2. 1] and 1 1 (T f.

we have R(T ) = L[y 1 . MATHEMATICAL PRELIMINARIES Definition 2. y 2 . {T xn} has a convergent x ¯ subsequence in Y . s)]2 dtds = i=1 j=1 (aij )2 . } such that αj (n ) αj k y j (nk ) → y = α1 y1 + α2 y2 + . then the set {φi (t)φj (s)}∞ i. Equivalently. as defined in Example 2. X) is compact.8 T ∈ B(X. 1 ≤ j ≤ N. 1]. Theorem 2. y N ]. Let {φi (t)} be a complete orthonormal set in L2 [0. Definition 2. quence {αj (nk ) (n) Since T is bounded {αj } is bounded for each j and hence ∃ subseN j=1 (n) T xnk = compactness of T. 1]). . X and Y are assumed to be Hilbert spaces. By Parseval’s equality.4 (1) The composition of a compact operator with a bounded operator is compact. s)f (s)ds. 1] × [0.54 CHAPTER 2. s)φi (t)φj (s)dtds. Refer Bose and Joshi [2] for the proof of this theorem.7 T ∈ B(X. Y ) is compact. Y ) is compact iff T ∗ ∈ B(Y. This proves the → αj . Hence ∞ ∞ k(t.2. we have 1 0 0 1 ∞ ∞ [k(t. Proof : As R(T ) is finite dimensional. s)]2 dtds < ∞.2. Example 2. 1] → L2[0.2. 1]. Y ) is said to be compact if it maps bounded sets B ⊂ X onto relatively compact set T (B) (T (B) is compact).3 Every finite rank operator T ∈ B(X. (2) Limit of a sequence of compact operator is compact. . T is compact iff for every bounded sequence {¯n } ⊂ X. .2. (3) T ∈ B(X. .j=1 is a complete orthonormal set in L2 ([0.2. . (n) N Consider now a bounded sequence {xn } ⊂ X. . f ∈ L2 [0. Y ) is called finite rank operator if the range of T is finite dimensional. 1] 1 with 0 1 ∞ [k(t. It now follows that The following theorem gives the compactness property of the adjoint of a compact operator and also the limit of compact operators. s) = i=1 j=1 aij φi (t)φj (s) where aij = 1 1 0 0 k(t.2. then T xn = j=1 αj y j for some scalars αj .8 Consider the operator T : L2 [0. Theorem 2. + αN yN .2: 1 [T f ](t) = 0 k(t.

2.2. We now state the following theorem which gives orthonormal basis for a Hilbert space X arising out of eigenvectors of T ∈ B(X).2. Then the eigenvectors of T form a complete orthonormal basis for X. Also Tn → T in B(X) and hence by the previous theorem T is compact. we have 1 0 0 1 55  n n k(t. Either both (H) and (H ∗ ) have only trivial solutions or both have nontrivial solutions. giving the solvability of the operator equation involving compact operators.j aij φi (t)φj (s). The necessary and sufficient condition for (N H) to have a nontrivial solution is that f be orthogonal to every solution of (H ∗ ). called Fredholm Alternative. n Now define Kn (t.j=n+1 ∞ ∞ |aij |2 since i=1 j=1 |aij |2 < ∞. . 2. if λ is not an eigenvalue of T then λ is not an eigenvalue of T ∗ and (NH) has one and only one solution for each f ∈ X.6 Let T be a self-adjoint compact operator on a Hilbert space. In particular.2. Then Tn : X → X defined by 1 (Tn f )(t) = 0 Kn (t. Consider three related equations in a Hilbert numbers) T x − λx = 0 T ∗ z − λz = 0 T y − λy = f space X (over the field of complex (H) (H ∗ ) (N H) Theorem 2. We have the following theorem. Theorem 2. s) = i. refer Bose and Joshi[2] for the proof. s)f (s)ds is a finite rank operator and hence it is compact.5 1.j=n+1 [aij φi (t)φj (s)] dtds →0 = i. s) − 1 1 0 0 ∞ i=1 j=1 ∞ =   aij φi (t)φj (s) dtds 2 2 i. LINEAR AND NONLINEAR OPERATORS Further.

x(t)). We now examine the properties of a nonlinear operator F : X → Y.2.2.1 Let T be a compact and self-adjoint operator in a Hilbert space X. Aφn ] = [λ1 φ1 . . . tf ] (2. Remark 2. Example 2. t ∈ I = [t0 .1. MATHEMATICAL PRELIMINARIES Corollary 2. . refer Joshi and Bose [6]. .2. Compact nonlinear operators are defined in the same way as compact linear operators. Let Aφi = λi φi .2. X and Y are normed spaces. .9 Let X = n and A : n → n be given by a symmetric matrix (aij )n×n . Then A ∈ B( n ) is compact self-adjoint operator and hence the eigenvectors of A form a basis. tf ] dt with f : I × → continuous and bounded. . . . However.7) is equivalent to the solvability of the equation t x(t) = x(t0 ) + t0 f (s. φ2 .56 CHAPTER 2.2. Definition 2. x(s))ds.3 For nonlinear operators continuity is not equivalent to boundedness. This is called diagonalisation procedure and will be used in Chapter 4 while computing transition matrix of a given system of differential equations. λn ]. 1 ≤ i ≤ n where {φi }n are orthonormal. . This gives us P −1 AP = D. . (2. the solvability of Eq.10 Consider the following ODE dx = f (t. . t ∈ I = [t0 .2. Assume that the eigenvectors corresponding to distinct eigenvalues are linearly independent and hence defining P as before.8) .2.1. . it may be pointed out A can be diagonalised even if A is not self-adjoint provided it has distinct eigenvalues. it is merely nonsingular. F is said to be bounded if it maps bounded sets in X into bounded sets in Y. . .7) Then by Proposition 1. Then P is an orthogonal matrix (P P = I) and AP = [Aφi . i=1 Let P = [φi .2. (2. . Example 2. λn φn ] = P D where D = diag[λ1 . But P is no longer an orthogonal matrix. Thus A is similar to a diagonal matrix.9 F : X → Y is said to be continuous if xn → x in X ⇒ F xn → F x ∀ x ∈ X. φn ]. we get P −1 AP = D. Then T is positive semidefinite (positive definite) iff all eigenvalues of T are nonegative (positive).

3. This implies that xk (s) → x(s) in for all s ∈ [t0 . x(s))|ds sup (s. x)|) For this inequality it follows that F x is continuous and hence F x ∈ C[t0 . one can show that F is bounded on X. LIPSCHITZ CONTINUITY AND DIFFERENTIABILITY 57 The RHS of Eq.tf ] Hence F is a continuous nonlinear operator from X = C[t0 . Similarly. x) is defined on I × Ω → Ω where Ω is an open subset of n .2. Definition 2.tf ] x(t) n 2. The above result also holds true if f (t. (2. One then denotes by C[t0 . tf ]. x(s))ds We have t2 |[F x](t1 ) − [F x](t2 )| ≤ t1 |f (s. y ∈ X (2. tf ]. tf ] the space of all n − valued continuous functions and the norm of the elements x ∈ C[t0 .2. One can also show that this convergence is uniform and hence F xk − F x = sup |F xk (t) − F x(t)| → 0 t∈[t0 .1 F : X → Y is said to be Lipschitz continuous if ∃ constant k > 0 such that F x − F y ≤ k x − y ∀ x. Further t |F xk (t) − F x(t)| ≤ t0 |f (s.x)∈I× ≤ m|t1 − t2 |(m = |f (s. that t |F xk (t) − F x(t)| ≤ t0 |f (s. it follows by bounded convergence theorem. tf ] in to itself.1) . tf ] if x ∈ C[t0 .8) gives rise to a nonlinear operator F on C[t0 . xk (s)) − f (s. Since (s.3. t1 ]. xk (s)) − f (s.3. x) is continuous. x. s)|ds → 0 for all t ∈ [t0 . tf ]. tf ] is defined as x = sup t∈[t0 . x) → f (s. tf ] defined as follows t [F x](t) = x(t0 ) + t0 f (s. x(s)|ds Let xk → x in C[t0 .3 Lipschitz Continuity and Differentiability In this section too the spaces under consideration will be normed space.

x ¯ Example 2. 1 ≤ i ≤ m. h ∈ X. such that ¯ ¯ ¯ ¯ ¯ ¯ e.2) where lim ¯ h →0 w(x. x2 . . F (¯ + h) − F (x) = e. .3. 1. x + h ∈ X and e ∈ Y ∗ . xn ). One can show that F is differentiable at x ∈ X with derivative A iff F (x + th) − F (x) lim = Ah ∀ h ∈ X t→0 t and x → F (¯) is continuous. . . Its value at h will be denoted by F (x)h. for all h ∈ X dt We have the following variant of the meanvalue theorem for F : X → Y.3. . h) =0 h A is called the derivative of F at x and is denoted by F (x). 0 < τ < 1. fm (x1 . . . .3. h) (2. . ¯ We choose h ∈ n as j th coordinate vector ej = (0. x2 . . . ¯ x Equivalently. . 0) Then lim 1 F (x + th) − F (x) − tAh = 0 t t→0 1 ⇔ lim | [fi (x + ej ) − fi (x) − taij ]| = 0. xn ) = (f1 (x1 . Corollary 2. Theorem 2.1 If F (¯) ≤ K ∀ x ∈ X then F is Lipschitz continuous.2 F : X → Y is said to be differentiable at x ∈ X if ∃ A ∈ B(X. if partial derivative are ∂xj n . F (¯ + τ h)h ¯ x ¯ ¯ x This immediately gives the following Corollary. .1 Let F : X → Y be differentiable at every x in X.1 Let F : n → n be given by F (x1 . .3.58 CHAPTER 2. . Then for ¯ points x. . . Let A = (aij ) be the m × n (matrix. 1 ≤ j ≤ n t→0 t This shows that F (x) is differentiable at x ∈ derivatives at x ∈ n and F (x) = A = iff fi (x) has continuous partial ∂fi (x) . . writing φ(t) = F (x + th) for x. which we want to compute) such that F (x) = A. . . xn )). we see that F has derivative F (¯) iff x d ¯ [φ(t)]t=0 = F (x)h. x2 . . . Refer Joshi and Bose[ 6 ] for the proof of this theorem.3. . MATHEMATICAL PRELIMINARIES Definition 2. there exists a constant τ. Y ) such that F (x + h) − F (x) = Ah + w(x.

h) + (x.3. h) as A∗ = A.2. 0 < α < 1 for all x. f (x)h = = = = = lim f (x + th) − f (x) t 1 (A(x + th). 2 x∈X be given by We assume that A is a bounded self-adjoint linear operator. x) 2 t t→0 t→0 lim 1 t (Ax. . f is a mapping from f (x). h) + (A∗ x.   ∂1 fm (¯) · · · x   ∂n f2 (¯)  x     . This implies that f (x) = Ax ∀ x ∈ X. X is any normed space. (x + th)) − (Ax. A function f : X → is said to be convex if f (αx + (1 − α)y) ≤ αf (x) + (1 − α)f (y). This derivative f of a functional f is also denoted as f . h) 2 A + A∗ x. . that is. The derivative of f is computed as under. So we have  ∂1 f1 (¯) x ··· ∂n f1 (¯) x  59 In case f is a functional. Remark 2. h) t→0 2 t lim lim 1 2 t→0 1 (Ax. Ah) + t2 (Ah.3.1 In dealing with minimization problem involving several variables we shall encounter convex functionals as in Example 2. called the gradient of f at x.3.  . y ∈ X.  . h) + t(Ah.2 Let X be a Hilbert space and let f : X → f (x) = 1 (Ax.2. then f (x) = Example 2. x). .    ∂1 f2 (¯) · · · x   F (x) =   . LIPSCHITZ CONTINUITY AND DIFFERENTIABILITY continuous.   ∂n fm (¯) x n to . h 2 = (Ax.3. From this representation it also follows that f (x) = A.

3. z) + (A Aa.3. z) − 2(Aa.2 is strictly convex.3. Hence the functional of Example 2. which we obtained earlier.3.60 CHAPTER 2.Data Fitting) In Example 2.which captures the local minimizer of a function f defined on the entire space n . = (z.called the steepest descent algorithm . we get φ(a) = −2A z + 2A Aa It is clear that A A is positive semi-definite. a ∈ ¯ n Steepest Descent Algorithm 2. . We give below one such algorithm . If A A is invertible then A A is positive definite and hence f is strictly convex and its critical point is the minimizer of φ. z) + (Aa. z − Aa) = (z. if A is positive definite. a).2. we minimize d 2 given by m n d 2 = z − Aa 2 = k=1 (fk − a i xi ) 2 i=0 (k) where z and A are as defined in the example. Such functionals f may not be convex and hence there is a need to generate algorithms which will computationally capture the minimizer (local / global). Example 2.5 involving problems of least square approximations. Estimating parameters requires the minimization of a functional involving these parameters. Aa) Proceeding as in Example 2. g (k) = f (x(k) ) (k) +αd (k) ) to get Set x(k+1) = x(k) + α(k) d Use any stopping criterion and set x∗ = x(k+1) .2.3 (Least Square Approximation .3. x∗ is the point of optimum. Remark 2.1 Step 1 Step 2 Step 3 Step 4 Step 5 Step 6 : : : : : : Start with x(k) (k = 0) Set d (k) Minimize 1-dimensional function φ(α) = f (d α(k) (k) = −g(k) . Also a∗ is the critical point of φ iff φ(a∗ ) = 0 This implies that A z = A Aa∗ and hence a∗ = (A A)−1 A z. Let φ(a) be defined as follows φ(a) = (z − Aa.6).3.3 While solving differential equations. Otherwise set k = k + 1 or GO TO STEP 2 Compute f (x∗ ). This is the same result. we shall also be estimating a set of parameters which will arise in the defining model.2 One can show that f is strictly convex on a Hilbert space if f (x) is positive definite (refer problem number 8 in sectoin 2. MATHEMATICAL PRELIMINARIES Remark 2. z) − 2(Aa.

x) ∈ [0. 1] → C[0. 1] to itself. 1]. g (k) g (k) 1 (Ax. LIPSCHITZ CONTINUITY AND DIFFERENTIABILITY In Example 2. x ∈ 2 n 61 (2. x(s))] ¯ α k(t. s) (s. x(s))h(s)ds ∂x by Lebesgue dominated convergence theorem. x(s))ds where k and f satisfy the following properties (i) k(t. 1] × [0.3) Example 2. s)f (s. 1]. Thus F is differentiable at x with F (x) as a linear integral operator generated by the kernel ∂f g(t. 1] be defined by 1 [F x](t) = 0 k(t. the iteration procedure for the functional f (x) = is given by x(k+1) = x(k) − g(k) .11. x(s)) and is given by ∂x 1 F (x)h(t) = 0 g(t. s) is a bounded on [0. x). (ii) x → f (t.3. Also.2.3. s)h(s)ds. it follows that F (x) ≤ M and hence by Corollary 2. 1] such that ∂f | (t. 1] × [0. s) 0 [f (s. We now compute the derivative of F. g(k) Ag (k) . as g(t. x) is differentiable mapping from to for all t ∈ [0. we can prove that F is a continuous mapping of C[0.1 it follows that F is Lipschitz continuous on X = L2 [0. 1].3.2.3. s) is continuous on [0. . (x + αh)(s)) − f (s.3.2. 1] × ∂x As in Example 2. s) = k(t.4 Let F : X = C[0. [F (x)h] = = = = 0 d [F (x + αh)]α=0 dα α→0 lim [ F (x + αh) − F (x) ] α 1 α→0 1 lim k(t. x)| ≤ M for all (t. s) ∂ f (s.

x F (x) = F (x0 ) + F (x0 )(x − x0 ) + w(x. x4 (t)) represents the state vector and ¯ u(t) = (u1 (t). u). h) where w(x. f3 (¯.2 that the motion of a satellite orbiting around the earth is given by an ordinary differential equation (ODE) in 4 : d¯ x = f (¯.3. u)) x ¯ x ¯ x ¯ x ¯ x ¯ where x4 k f1 = x2 . σ (x1 + σ)2 x2 u2 σ x4 + f3 = x4 . then it follows that we have the following expansion for F (x) with F (¯0 ) ∈ B(X).3. x ˆ By linearization. MATHEMATICAL PRELIMINARIES Let F be differentiable at x0 ∈ X.3. u2 (t)) is the control vector (thrusts on the satellite).5) F (x0 ) Hence we get the Newton’s algorithm to find the zeros of F (x).3. x3 (t).5 We have shown in Example 1. h ˆ F (x) = F (x0 ) + F (x0 )(x − x0 ) is called the linearization of F at x0 .Newton’s Algorithm Step 1 Set k = 0 and input initial guess xk and the tolerance > 0 Step 2 x(k+1) = x(k) − [F (x(k) )]−1 F (x(k) ) Step 3 If x(k+1) − x(k) <∈.2 .3. f4 = −2σ( + w) σ (x1 + σ) (x1 + σ) .3. This gives x = x0 + [F (x0 )] −1 (2.62 CHAPTER 2.3 If F is differentiable at x0 then the function (2. f4 (¯. starting with an initial guess x0 .3. u). STOP and set x∗ = x(k) as the computed zero of F (¯). u).4 The linearization process can be used to find the zeros of the nonlinear function F (x) in a neighbourhood of the operating point x0 ∈ X if [F (¯0 )] is invertible. f2 (¯. u) = (f1 (¯. h → 0 as h → 0. Remark 2. u ∈ 2 x ¯ ¯ ¯ (2.3. Otherwise increment k by 1. u). it suffices to compute the zeros F (x) and hence ˆ 0 = F (x) = F (x0 ) + F (x0 )(x − x0 ).6) dt x(t) = (x1 (t). x ∈ 4 . f is given by ¯ f (¯.4) Definition 2. f2 = (x1 + σ)( + w)2 − + u1 . x Step 4 GO To Step 2 Example 2. ¯ Algorithm 2. x2 (t).

So the linearized ODE.7) and Eq. 0)        ∂x1 ∂f2 ∂x1 ∂f3 ∂x2 ∂f4 ∂x2 ∂x2 ∂f2 ∂x2 ∂f3 ∂x2 ∂f4 ∂x2 ∂x3 ∂f2 ∂x3 ∂f3 ∂x3 ∂f4 ∂x3 ∂x4    ∂f2   ∂x4     ∂f3   ∂x4    ∂f  4 ∂x4    3w2  =   0   0 (¯. respectively.1 Let X be a metric space and F : X → X. FIXED POINT THEOREMS AND MONOTONE OPERATORS 63 Using linearization procedure for f (¯.2 Let (X. (2.3. (2. known as Banach contraction mapping principle. F : X → X is said to be contraction if F is Lipschitz continuous with Lipschitz constant strictly less than .5). described by Eq. We begin with the following constructive fixed point theorem by Banach.7)  Here σ is normalized to 1.  ∂f 1  ∂u1    ∂f2   ∂u 1  B = fu (0.4. This theorem will be one of the important instruments in obtaining existence theorems for ordinary differential equations. corresponding to Eq.3. u = 0 we get ¯ ¯ ˆx ¯ f (¯.8).4 Fixed Point Theorems and Monotone Operators Definition 2.4. (2.6) is d¯ x = A¯ + B u x ¯ dt where A and B are given by Eq.¯) oo  0 0  (2. Definition 2.4. u) = fx (¯ ¯ x + fu (¯ ¯ u 0. d) be a metric space. 0)¯ 0.¯) oo  0 1 0 0 0 0 −2w   0 2w     0 1    0 0 (2. u). (2. around x ¯ x = 0.3. 0)¯  ∂f  ∂f ∂f ∂f 1 1 1 1        A = f x (¯ ¯ =  0. 0) =    ∂f3   ∂u1    ∂f 4 ∂u1 ∂f1  ∂u2    ∂f2   ∂u2     ∂f3   ∂u2    ∂f  4 ∂u2      1 0     =    0 0      0 1 (¯.3.2.3. An element x ∈ X is said to be a fixed point of the mapping F if F x = x.3.8) 2.

xn+1 = F xn and F is continuous. + d(xm+1 .1) If X is a normed space. . x∗ ) + 1−α . F xk−2 ) ≤ αd(xk−1 . y ∗ ) and hence d(x∗ . y ∈ X ¯ ¯ (2.4. we have d(x∗ . .]d(x1 . x∗ ) is given by d(x∗ . x −1 ) + . l > m 1−α → 0 as m → ∞ This implies that {xn } is Cauchy in X and as X is complete. y ∗ ) = 0. y ∗ ) = d(F x∗ . xn ) αn d(x1 . xk−2 ) ≤ αd(x1 . x∗ ) ≤ n→∞ αn d(x1 . xm ) ≤ d(x . there exists a positive constant α < 1 such that d(F x. and {xn } a sequence defined by xn+1 = F xn (n = 0. x0 ) αm = d(x1 . x0 ) ≤ αm [1 + α + α2 + . .1 Let F be a contraction mapping of a complete metric space X into itself. . . m > k ≤ d(xm . Then F has a unique fixed point x∗ ∈ X. As xn → x∗ . The error estimate d(xn . 2. That is. 1−α Proof : We have d(xk .2) Theorem 2. then (2. .4.64 CHAPTER 2.) then lim xn = x∗ and d(xn . + αl−1 ]d(x1 . y ∈ X for all x. x∗ ) + d(xm . x0 ) This gives d(x . it follows that x∗ = F x ∗ . y ∗ ) < d(x∗ . .4. MATHEMATICAL PRELIMINARIES 1. . F y ∗ ) ≤ αd(x∗ . x0 ). The uniqueness of the fixed point x∗ follows from the fact that for two fixed point x∗ and y ∗ . F y) ≤ αd(x. xm ) ≤ [αm + αm+1 + . Moreover if x0 ∈ X is any arbitrary point.1) is equivalent to the following inequality (2. xk−1 ) = d(F xk−1 . x0 ). x∗ ) n ≤ d(xm . x0 ). it converges to x∗ ∈ X. . y) Fx − Fy ≤ α x − y ¯ ¯ ¯ ¯ for all x. 1.4.

4.4. Eq. Let F x = Ax + y. (2.2.1 Let A ∈ B(X) be such that A < 1. it follows that F is a contraction and hence it has a unique fixed point x in the Banach space X (complete metric space). That is.3) has a unique solution x. ¯ Corollary 2.4. y ∈ X has a unique solution.3) in a Banach space X. Then (I − A) is bounded and invertible with 1 (I − A)−1 ≤ (2. x0 ) 1−α As a Corollary of this theorem.4. y ∈ X arbitrary and hence x ≤ ≤ Ax + y A x + y 1 1− A This gives x ≤ ¯ Hence y ¯ (I − A)−1 y ≤ ¯ for all y ∈ X. FIXED POINT THEOREMS AND MONOTONE OPERATORS Letting m → ∞. let x = (I − A)−1 y This implies that x − Ax = y. we obtain the solvability of the linear equation x − Ax = y (2.4. Then F x1 − F x2 = Ax1 − Ax2 ≤ A x1 − x2 for all x1 . Further. we get d(xn . A ∈ B(X) with A < 1 and y ∈ X. x2 ∈ X As α = A < 1. Here.4) 1− A Proof : We first show that x − Ax = y. x∗ ) ≤ 65 αn d(x1 . ¯ This implies that 1 1− A 1 1− A y ¯ (I − A)−1 ≤ .

+ α n n xn = b n (2. c =  22     . + α2n xn = b2 .   . .4.7) is equivalent to the solvability of the matrix equation x − B x = c. .      bn 0 αnn                 = α n . A = (αij )n×n .66 CHAPTER 2. .4. (2.7) Then the solvability of Eq. this has a unique solution x∗ and the iteration ¯ scheme x(k+1) = c + B x(k) ¯ ¯ ¯ . ¯ ¯ ¯ Hence by the above corollary.4.4. 1 ≤ i ≤ n. x1 αn. − α11 α11 α11 α21 x1 α2n xn b2 − −. . .n−1 xn−1 bn − −.5) in terms of its component we get α11 x1 + α12 x2 + .4. αn2 αnn ··· ··· ···  b1 α1n   α11 α11      b α2n    2   α22   α  . α n 1 x1 + α n 2 x2 + .. . Then Eq. .   . . .− αnn αnn αnn b1 α12 x2 α1n xn − .6) Assume that αii > 0 and the matrix A is diagonally dominant: n i=j i=1 |αij | < αii .6) is equivalent to x1 x2 = = . + α1n xn = b1 α21 x1 + α22 x2 + .. ..1 Solve the linear system A¯ = ¯ x b in n (2.. (2. MATHEMATICAL PRELIMINARIES Example 2.. xn Define    α21   α  22 B = −       α n1 αnn  0 α12 α11 0 . Expanding Eq..4. . B < 1.4.5) . (2.− α22 α22 α22 (2. .

.03902 0. ¯ This gives us the well known Jacobi’s iteration scheme (refer Conte and deBoor [4]). . F : (X.08465 0.15573 -0.04622 0.00316 -0.16986 (n) x − x(n) 5.4. −.36364 -0. = . Corollary 2. x1 (k+1) = = b1 α12 (k) α2n (k) − x ···− x α11 α11 2 α11 n bn α21 (k) α2. −.06154 (n) x2 0 0.2. d) be a complete metric space.37828 0.16960 -0. .04319 0.06341 0.06769 0.4.06619 0.09077 0. We still get a unique fixed point.04368 0.16667 0.1: Iterated solution of the matrix equation n 0 1 2 3 4 5 6 7 8 9 x1 1 -0.56439 0. 2.58333 -0.02363 0.n (k) − x ···− x α22 α22 1 α22 n .00786 0.16895 -0. For the diagonally dominant  24 3  1 12 A=  0 1 7 −4 ¯ = (1.05785 0. d) be such that F N is a contraction for some positive integer N. as we see in the following corollary.03850 0. FIXED POINT THEOREMS AND MONOTONE OPERATORS 67 converges to x∗ .2 Let (X. x2 (k+1) .06610 0.16788 -0. −4) b Table 2.n−1 (k) − x ··· x αnn αnn 1 αnn n matrix  9 7 −3 −4  .17548 -0.00089 It may happen that F : (X.06634 0.02336 0. 1.05923 0.03822 (n) x3 -1 0.07844 0. .04901 0.00631 0.06125 0.41667 -0. . d) → (X.4. .06603 (n) x4 2 -0. . Then F has a unique fixed point. x3 . d) is not a contraction but some power of it is a contraction. d) → (X. x(k+1) n = bn αn1 (k) αn.00389 0.53030 1.06073 0. x4 ) of A¯ = ¯ ¯ x b.16575 -0. . 22 −3  −3 24 (n) (n) (n) (n) we get the following iterated solution x(n) = (x1 .04018 0.00178 0. x2 .00758 0.

giving a fixed point of F. ∃ x∗ ∈ X such that F N x∗ = x∗ . That is. Let F be a continuous mapping of K into a compact subset of K. b] be defined by t [F x](t) = a x(s)ds t a We have [F k x](t) = This gives |(F k x − F k y)(t)| ≤ ≤ = This implies that F kx − F ky ≤ As 1 k − 1! (t − s)k−1 x(s)ds t 1 (t − s)k−1 |x(s) − y(s)|ds k − 1! a (b − a)k sup |x(s) − y(s)| k! s∈[a. Thus there exist a positive integer N such that α = <1 N! N! and F N x − F N y ≤ α x − y ∀ x. y ∈ X = C[a. MATHEMATICAL PRELIMINARIES Proof : By the contraction mapping principle. Also every fixed point of F is also a fixed point of F N and hence fixed point of F is unique. y ∈ X = C[a. In the following. b]. For such operators. which are not contractions but have compactness property.2 Let K be a nonempty closed convex subset of a normed space X. we have the following Schauder’s theorem. . the space X is assumed to be a Hilbert space. we get F N +1 (x∗ ) = F (x∗ ) and hence F N (F (x∗ )) = F (x∗ ). At times. This implies that F (x∗ ) is also a fixed point of F N and uniqueness implies that F x∗ = x∗ . Theorem 2.4. That is F N is a contraction. F N has a unique fixed point x∗ . Then F has a fixed point in K.4. it follows that ∃ positive integer N such that k! N (b − a) (b − a)N < 1.68 CHAPTER 2.2 Let F : C[a. We shall now define this concept and obtain some abstract results which subsequently will be used in Chapter 3 while dealing with existence and uniqueness theorems. Another important notion which will be used in solvability analysis of differential equation is the concept of monotone operators in Hilbert spaces.b] (b − a)k x−y k! (b − a)k x−y k! ∀ x. F is not a contraction if (b − a) > 1. Applying F to both sides. b] → C[a. However. b] (b − a)k → 0 as k → ∞. Example 2. we shall encounter mappings in a normed space.

y(t) = x. F is called strongly monotone if ∃ constant c > 0 such that (F x1 − F x2 .4.t. x1 − x2 ) ≥ 0 for all x1 . x1 .4 Let f (s. x) is monotone increasing for almost all s ∈ [0. − d2 y dt2 .Eq. x2 ∈ X (2.4. b > 0 on f. x2 ) ≥ 0 ∀ x1 .4. y ∈ D(A) (Ax. 1]. x1 − x2 ) = ¯ ¯ ¯ ¯ (2. x2 ∈ D(F ) y ¯ ¯ ¯ ¯ ¯ and y 1 ∈ F x1 . 1]. x for almost all s ∈ [0.3 F : X → X is said to be monotone if (F x1 − F x2 . 1].t.2. x) = − 1 a d2 x x(t)dt dt2 = 0 ( dx 2 ) dt. (2. x(s)) If we assume a growth condition of the type |f (s. x1 − x2 ) ≥ c x1 − x2 2 for all x1 .3 A : L2 [0. x. 1] be the differential operator − 2 with D(A) defined dt as 2 2 D(A) = {x ∈ L [0. 1] → L2 [0. x2 (s)))(x1 (s) − x2 (s))ds . x) be a function defined on [0. x(t) ∈ L [0. (2. x)| ≤ a(s) + b|x|. s for all x ∈ and continuous w. x(0) = x(1) = 0} ˙ ¨ A has dense domain and it is self adjoint (A¯.4.10) As the integrand in the above equation is nonnegative for all x1 .9) If F is not defined on the whole space.8) . x ∈ D(A) ¯ dt ≥ 0 Example 2.9) on D(F ). the domain of F . 1]. Then we have 1 (F x1 − F x2 . then we verify Eq. Further. FIXED POINT THEOREMS AND MONOTONE OPERATORS Definition 2. y ) = x ¯ A is monotone as 1 − d2 x dt2 .4.8) F is called strictly monotone if the above inequality holds and is strict for x 1 = x2 .4. d2 Example 2. x2 ∈ X.4.r. 0 (f (s. Definition 2. it follows that F is monotone. x) is measurable w. 1] : x(t). 1] × to be such that f (s.4 A multivalued mapping F : X → 2X is called monotone if (¯1 − y2 . x1 (s)) − f (s. x2 ∈ X 69 (2. then one can show that F maps X into itself and is continuous and bounded (refer Joshi and Bose[6]). y2 ∈ F x2 ¯ ¯ ¯ ¯ A monotone mapping F is called maximal monotone if it has no proper monotone extensions. assume that the mapping x → f (s.4. 1] be defined by [F x](s) = f (s.4.4. a ∈ L2 [0.r. Let F : X = L2 [0.

r.t. x→−∞ x →∞ ¯ lim (F x. While dealing with differential equations. Then (I − F ) is strongly monotone.3 Let F : X → X be a contraction on the Hilbert space X. Theorem 2. x2 ∈ X ¯ ¯ We have the following solvability result due to Browder[3]. For example. Then F x = y has a unique solution for any y ∈ X. Theorem 2.5 Let F : X → X be a contraction. ((I − F )¯1 − (I − F )¯2 . Refer Joshi and Bose [6] for the proof. F = L + F0 where L is densely defined linear maximal monotone operator and F0 is monotone operator defined on the whole space. y.5 Let F1 be a maximal monotone operator with dense domain containing 0 and F2 a single valued operator which is defined on the whole space and is continuous. In view of Example 2. . Then R(F ) = X where F = F1 + F2 .70 CHAPTER 2.5.4 Let F : X → X be continuous.4. x1 − x2 ) − (F x1 − F x2 . which are only densely defined. we shall encounter operators. this corresponds to the condition lim f (x) = ∞ We have the following result due to Minty [9]. x ¯ x→∞ . concerning strongly monotone continuous operators on X.4. Moreover F −1 is Lipschitz ¯ ¯ continuous and monotone. x1 − x2 ) = (¯1 − x2 . x) ¯ ¯ = ∞. Then R(F ) = X.4.4.4. bounded and coercive.4. Theorem 2. MATHEMATICAL PRELIMINARIES Example 2. Corollary 2. monotone and coercive. Then the following theorem gives the surjectivity of F = L + F0 .4. we get the following Corollary.3 Let F : X → X be a strongly monotone continuous operator. Then x = F x + y has a unique solution x for every y ∈ X and this solution x varies ¯ ¯ ¯ ¯ ¯ ¯ continuously w.5 F : X → X is called coercive if For functions f defined on and lim f (x) = −∞. x1 − x2 ) x x ¯ ¯ x ¯ ¯ ¯ ¯ ¯ ¯ ¯ ≥ ≥ x 1 − x2 ¯ ¯ x 1 − x2 ¯ ¯ 2 − F x 1 − F x2 ¯ ¯ − α x 1 − x2 ¯ ¯ 2 2 x 1 − x2 2 = (1 − α) x1 − x2 ¯ ¯ ∀ x1 . ¯ Definition 2.

which is a symbolic function representing the charge density corresponding to a unit charge at the origin.5 Dirac-Delta Function During the course of our discussion on boundary value problems. t = 0 This limiting charge is denoted by δ(t). the total charge for the densily distribution ρ (t) on the line is 1. t = 0 ∞.2. vaguely the Dirac-delta function δ(t) can be defined as δ(t) = ∞ 0. π γ (t) = −∞ ρ (x)dx = ∞ 1 1 + tan−1 (t/ ) 2 π Hence t→∞ lim γ (t) = −∞ ρ (x)dx = 1 for all So. . t = 0 such that −∞ δ(t)dt = 1 Obviously. this is not a function. t = 0 ∞. we take a continuously distributed charge on the real axis with charge density ρ (t) = > 0 and ρ (t) is small for small 1 π t2 + 2 but for a peak height The cumulative charge distribution γ (t) is given by t 1 at the origin. we shall encounter Dirac-delta function δ(x). To specify it. whereas lim ρ (t) = →0 0.5. That is δ(t) = lim ρ (t) →0 So. DIRAC-DELTA FUNCTION 71 2.

mathematically. yet these sequences can behave differently at t = 0.1 The support of a function f (t) is the closure of the set of all points where f (t) is different from zero.5. To be more mathematically rigorous. sup f = {t : f (t) = 0} . For.2 0. we define what we call the action of ρ (t) on a class of test functions which consists of functions continuous at the origin and well behaved at infinity.5. one can take various types of sequences to get the limit δ(t).2 -5 -10 0.4 -0.72 CHAPTER 2.4 -0. MATHEMATICAL PRELIMINARIES 6 5 4 3 2 1 -0. that is.2: Cumulative charge density The description of a point charge by considering it as a limit of the sequence ρ (t) is not quite satisfactory.5.4 Figure 2.1: Charge density 10 5 -0.2 0. Definition 2.4 Figure 2.2 0.

3 A sequence of function {φn (t)} in D( ) is said to converge to zero if (i) all φn (t) have the same compact support.5. Example 2. The distribution functional δ is defined as above is called the Dirac-delta function. Example 2. 1 2−1 t t≥1 .5.5. The action of f ∈ D ( ) is denoted by (f.2 The sequence φ(t) n ∞  exp with φ(t) defined by Eq. the sequence t does not converge to n zero in D( ) as these functions do not have the same compact support. One can show that f is also a distribution.5. φ Definition 2.1 φ(t) =   0. Whereas.4 A continuous linear function on the space D( ) is called a distribution.5.5.5. φ)}. This space is a vector space with respect to pointwise addition and scalar multiplication. φ).4 Dipole distribution δ is defined as (δ. ∀φ ∈ D( ) It is easy to see that δ is linear and continuous and hence δ ∈ D ( ).5. Definition 2.3 Define a functional δ on D( ) by δ.5 A sequence of functions {fn }∞ in D ( ) is said to converge n=1 if for every φ ∈ D( ) the sequence of scalars {(fn . Definition 2. Example 2.2. The space of all distributions is denoted by D ( ). .5.1) φ(t) is infinite times differentiable and derivatives of all order vanish for |t| ≥ 1. DIRAC-DELTA FUNCTION 73 Definition 2. |t| < 1 (2. ˜ ˜ Example 2. (ii) sequence {φn (t)} converges to zero uniformly and for any positive integer (k) k. φ) = φ (0) ∀ φ ∈ D( ) Definition 2.5.1) conn=1 verges to zero in D( ). φ = φ(0). φ) = −∞ f (t)φ(t) ∀ φ ∈ D( ) This is called a regular distribution. the derivatives {φn (t)} also converges uniformly to zero. A distribution which is not a regular distribution is called a singular distribution.5. (2. φ)} converges to the limit {(f.5.2 The space D( ) of test functions consists of all real valued functions φ(t) which are infinitely times differentiable and which have compact support.6 A locally integrable function f generates a distribution via the definition ∞ (f.

φ(t)) = (f (t).6 Let H(t) be the Heaviside   0. φ ) ∞ = − Hence we have H = δ. φ(t + a)) = φ(a) and (δ(−t). φ(t)) = (δ. φ(t)) = −(H. φ(t)) = |a| Example 2. φ(−1)) = φ(0) = (δ. 1/2.5 The translation δ(t − a) and scale expansion δ(−t) are given by (δ(t − a). φ ) ∀ φ ∈ D( ) By repeated application of the above definition.5. a = 0 ∀ φ ∈ D( ) (f (at). φ(t/a)). H(t) generates a distribution given by (H(t). φ(n) ) ∀ φ ∈ D( ) Here f (n) = dn f . scale expansion. derivative of a distribution as follows.5. φ) = (f.5.7 The translation f (t − a) and scale expansion f (at) of a distribution is given by (f (t − a). Definition 2.8 For any distribution f.1 The Dirac-delta distribution is a singular distribution.5. H(t) =  1. φ(t)) = function t<0 t=0 t>0 ∞ H(t)φ(t)dt −∞ ∞ = 0 φ(t)dt Its derivative H (t) is given by (H (t). For the proof of this fact refer Bose and Joshi [2] We can define translation.74 CHAPTER 2. φ(t)) . φ(t)) ∀ φ ∈ D( ) Definition 2. φ(t)) = (δ. −φ ) = −(f. MATHEMATICAL PRELIMINARIES Theorem 2.5. φ(t + a)) ∀ φ ∈ D( ) 1 (f (t). its derivative f is defined as (f . φ) = (−1)n (f. φ (t)dt = φ(0) 0 = (δ. dtn Example 2. we get (f (n) .

dt + an df φ(t)dt + dt ∞ fk φ(ak ) k=1 n Therefore (f . DIRAC-DELTA FUNCTION 75 Example 2. .2. = (f. . We assume that f is locally integrable wherever it exists.5. a1 Integrating by parts and noting that f (a+ ) − f (a− ) = k k (f .9 Let L be a linear differential operator given by L = a0 (t) n = k=0 dn dn−1 d + a1 (t) n−1 + . an where f (t) jumps by the respective amount f1 . φ ) = − = − − ∞ f (t)φ (t)dt −∞ ∞ −∞ ∞ an a2 f (t)φ (t)dt − f (t)φ (t)dt f (t)φ (t)dt . then g(t)f (n) (t) can be found as follows (gf (n) (t). .5. φ) = −∞ n [f ]φ(t)dt + k=1 fk φ(ak ) This gives f = [f ] + k=1 fk δ(t − ak ). we get df φ(t)dt + dt n a2 a1 df φ(t)dt + .7 Let f (t) be differentiable except at points a1 . . . .8 If g(t) is infinitely differentiable. fn . Then the distributional derivative f is computed as follows ∞ (f . . Example 2. Here [f ] indicates the usual derivative of f. . . φ) = ( k=0 an−k (t)f (k) .5. . φ) = (f (n) . . . + an−1 (t) + an n dt dt dt dk an−k (t) k dt with coefficients ak (t) infinite times differentiable. φ) = −∞ ∞ fk . dn (gφ)) dtn Example 2. we have n (Lf. L∗ φ) .5. . φ) = −(f. Then for the distribution f. f2 . φ) dk (an−k (t)φ(t)) dtk = (−1)k f. wherever it exists. gφ) = (−1)n (f. .

dtk The nth order differential operator L∗ is known as the formal adjoint of L. φ(t))| = | ≤ sin ntφ(t)dt| −∞ ∞ 1 | cos ntφ (t)|dt n −∞ 1 ∞ |φ (t)|dt ≤ n −∞ → 0 as n → ∞ So {sin nt} converges in D ( ) to zero distribution. 1]. MATHEMATICAL PRELIMINARIES (−1)k k=0 where L∗ φ = dk (an−k (t)φ(t)).   0.6 Exercises 1. |t| < 1 2 n ∞ 1 n (fn .10 Consider the sequence {sin nt}n=1 . φ) = −∞ fn (t)φ(t)dt = n 2 1 n 1 −n n φ(t)dt 2 = This implies that φ(t)dt 1 −n n→∞ lim (fn . Example 2. However. This sequence does not converge point wise except at t = 0. Compute the limit function.5.5. we have ∞ ∞ |(sin nt. 2. As it is locally integrable. |t| ≥ 1   n Example 2. φ) Thus the sequence of regular distribution {fn } converges in D ( ) to the delta distribution whereas the corresponding functions converge almost everywhere to zero. φ) = n→∞ lim n 2 1/n φ(t)dt −1/n = φ(0) = (δ. (a) xn (t) = sin nπt . if it exists. it generates a sequence of distributions.11 Let fn (t) =   n  . Test the convergence of the following sequences of the functions in the space L2 [0.76 n CHAPTER 2.

π] be the linear span of 1. (X. x2 . x2 ) = (0. Show that n = N (A) ⊕ R(A) 3. t2 . Also. x3 ) = 3 1 sin2 x1 + cos x2 cos x3 .r. (a) F (x1 . .2. 0) (b) G(x1 . x2 ) = exp(x1 x2 ) + cos x1 sin x2 . Show that Ψ(t) = 3 . 1] be defined by 1 [Lx](t) = 0 (s + sin t)2 x(s)ds From the first principle. x2 x1 exp(x1 . EXERCISES (b) xn (t) = tn (c) xn (t) = t sin nπt sin nπt (d) xn (t) = n 77 2. 10 cos x1 cos x2 cos x3 ) w.t 1-norm in 3 → 3 defined as is a contraction in 8. sin x2 cos x3 . x ¯ 0≤t≤a t ≤ 0. respectively. Let M ⊂ L2 [−π. show that L is compact. Linearize the following nonlinear operators around the given points. x2 ) 2 (x1 . find its adjoint L∗ and show that L is a singular operator. Show that f is f (¯) is monotone for all x ∈ X. π]. Determine if the function F : F (x1 . 4. 6. 1]. Let L : L2 [0. x2 ) = (0.  0. 1] → L2 [0.6. is a test function. x2 ) = (x1 + sin x1 cos x2 . Let f : X → convex iff x → ¯ 9. 5. Show that [Lx](t) = 0 t x(s)ds is a bounded linear operator in L2 [0. a Hilbert space ) be differentiable. or t ≥ a   exp(−(t−2 ))exp(−(t − a)−2 ). cos x1 sin x2 + x1 x2 ) (x1 . 0) 7. cos t . Let A ∈ n×n and let N (A) and R(A) denote the null space and range space of A. Find an element in M closest to cos 2t ∈ L2 [−π.

Duke Math Journal.. Monotone (Nonlinear) Operator in a Hilbert Space. Ltd. Show that CHAPTER 2. Functional Analysis. A. Conte and C. [9 ] G. Joshi. Some Topics in Nonlinear Functional Analysis. Minty. Wiley Eastern Ltd. Browder. 1963. 29.An Algorithmic Approach. g ∈ C ∞ d (sgnt) = 2δ(t) dt δ (b − a)δ(b − a)dt References [1 ] R B Bapat. MATHEMATICAL PRELIMINARIES (b) δ (b − a) = (c) (a) g(t)δ (t) = g(0)δ (t) − g (t)δ(t). Krasnoselskii. 1974. The Solvability of Nonlinear Functional Equations. 1962. Tata McGraw Hill Publishing Co. 341-346. DukeMath Journal 30. [5 ] Casper Goffman and George Pedrick. 1993. McGraw Hill Book Co. 1996. [3 ] F. D. 1988. [6 ] Mohan C Joshi and Ramendra K Bose. Topological Methods in the Theory of Nonlinear Integral Equation. Royden.. Macmillan Publications. Bose and M. 1984. Methods of Mathematical Physics. deBoor. Elemenatry Numerical Analysis . Pergaman Press. [4 ] S. 557-566. . E. [10 ] H. 1981. Linear Algebra and Linear Models.78 10. 1985 [7 ] M. [2 ] R. Real Analysis. Prentice Hall of India. C. Hindustan Book Agency. New Age International. L. K. [8 ] Balmohan V Limaye. First Course in Functional Analysis. 1964.

1).1.1. Let f : I × Ω → n be a function.1b) Also.1) has a unique solution. let us recall what we mean by the solution of the IVP . x) is Lipschitz continuous : ∃ a constant M such that ||f (t. 79 (3. (3.Picard iterates. continuation of solution and continuity of solution with respect to initial condition . x) with respect to x. ∀ t ∈ J and x(t) satisfies Eq.1. we mean the following differential equation with initial condition dx dt x(t0 ) = f (t.Eq.1 Let E = I × Ω ⊆ × n be an open set containing (t0 . x) − f (t. The notions of .1.1.1).1). by Initial Value Problem (IVP).1. y)|| ≤ M ||x − y|| ∀ t ∈ I and ∀ x.2) . containing ‘t0 ’ such that x(t) is differentiable on J with x(t) ∈ Ω.Chapter 3 Initial Value Problems This chapter primarily deals with the existence and uniqueness theory of initial value problems. Further assume that x → f (t. requiring Lipschitz continuity on f (t. y ∈ Ω Then the IVP given by Eq.1.1a) (3.1.1. (3.1). Theorem 3. 3. maximal interval of existence. not necessarily linear. (3.1 Existence and Uniqueness Theorems Let I be an interval in and Ω an open set containing (t0 . x0 ) and let f : E → n be continuous. (3. x(t) is a solution of Eq. (3. x(t)) = x0 (3. if ∃ an interval J ⊂ I. The following theorem gives the existence of unique solution of Eq. As defined earlier (Definition 1.1.are the prime targets of our investigation. x0 ) ∈ I × Ω.1).

Choose two positive numbers ‘δ’ and ‘τ ’ such that M τ < 1.1.x)∈B n . y(s))|| t t0 n ds ≤ M ||x(s) − y(s)|| n ds ≤ M ||x − y|| (t − t0 ) ≤ τ M ||x − y|| and hence ||F x − F y|| ≤ τ M ||x − y|| ∀ x. ‘m’ is finite as f is continuous on a compact set B ⊆ I × (t.1. t ∈ J (3. Define X ⊂ C(J) to be set of all continuous functions on J with values in Ω0 ⊆ Ω. x(s))ds (3.3) in some appropriate subinterval J ⊂ I. Then X is a closed subset of C(J) and hence a complete metric space with metric induced by the sup-norm of C(J). y ∈ X. t [F x](t) − x0 n = t0 t f (s. solvability of Eq. Let B denote a closed ball with centre at (t0 . x ¯ t ||F x(t) − F y(t)|| n ≤ t0 ||f (s. x(s))ds. we have. Denote by J = [t0 − τ. (3. Let ¯ ¯ m = sup |f (t. mτ < δ and { (t. Further. t0 + τ ] and Ω0 = { x ∈ Ω : ||x − x0 || ≤ δ }. x(s))ds f (s.1) is equivalent to the solvability of the following integral equation t x(t) = x0 + t0 f (s. x) ∈ I × ¯ n : |t − t0 | ≤ τ. We define a nonlinear operator F on X as follows t [F x](t) = x0 + t0 f (s. ∀¯. INITIAL VALUE PROBLEMS Proof : As shown in Proposition 1. y ∈ X . x(s)) − f (s.1. Also. x0 ) and contained in E. x)|.1. x(s)) t0 t n n ≤ ds ≤ m ≤ mτ ≤ δ ds t0 This implies that [F x](t) ∈ Ω0 ∀ t ∈ J and hence F maps X into itself.4) It is clear that F x is a continuous function defined on J.80 CHAPTER 3. ||x − x0 || ≤ δ } ⊆ B.1.

Then nth order initial value problem given by Eq.1. x(t)) dt x(0) = x0 (3. x2 =x1 =x. (3. defined as t x(k) (t) = x0 + t0 f (s.6a) (3. x(1) (0) = x2 . .1. ∃ x∗ ∈ X = C(J. This proves that x∗ (t) is unique solution of Eq.1.3) and hence that of Eq.1.1).7b) .1. Invoking Theorem 2. wherein the converging iterates are defined as x(k) = F x(k−1) .1. (3. Corollary 3.5) are called Picard iterates. (3. x(0) ∈ X.6) in to an initial value problem in this. converge to the unique solution x∗ (t) of IVP given by Eq. Ω0 ) such that ¯ x∗ (t) = F x∗ (t) = x0 + t f (s. x3 =x2 =x1 =x. (3.1. assume that ai (t). xn =xn−1 n . set . EXISTENCE AND UNIQUENESS THEOREMS 81 As α = τ M < 1. .1.1. Proof : We shall reduce Eq.1.1).5) with x(0) (t) = x0 . x(n−1) (0) = x(0) n Here x(k) (t) refer to the k th derivative of x(t). the IVP given by Eq.1. .1. (3.1. Corollary 3. 1 ≤ i ≤ n − 1 are continuous functions on I.1. Further. x1 = x. That is.7a) (3.1 The iterations x(k) (t).. For Proceeding as in section (1. x∗ (s))ds t0 and x∗ is unique.1.6) reduces to the following IVP on I × Ω × Ω · · · Ω ⊆ I × n dx = F (t.3). it follows that F is a contraction on the complete metric space X. . (3. Proof : This follows from Theorem 2.4. · · · . The iterates defined by Eq.3.1. x(t)) (3. x ) Lipschitz continuous with Lipschitz constant M . it follows that F has a unique fixed point x ∗ ∈ X. · · · . x(k−1) (s))ds (3.1. . . = f (t.4).4. We note that x(0) (t) = x0 ∈ X = C(J.1.2 Let Ω ⊆ be open and f : I × Ω → be continuous with x → f ( .6) has a unique solution. Ω0 ). (arbitrary) We now consider an nth order differential equation dn−1 x dx dn x + an−1 (t) n−1 + · · · + a1 (t) n dt dt dt with initial values (0) (0) Here F is defined by (3.6b) x(0) = x1 .. (3.

bounded subset of .82 CHAPTER 3. . (n) x (t) = (1 + t + t2 + · · · + tk ) + o(tk+1 ) . · · · . x2 . with ai = sup |ai (t)|. x) is Lipschitz continuous as we see below. x3 .1. 1 ≤ i ≤ n. The Picard’s iterates x(k) (t) are given by x(k) (t) t = x0 + 0 t f (s. x(k−1) (s))ds x(k−1) (s) 2 = 1+ 0 ds So. ||F (t. we get the following Picard iterates for the above problem.1.1 and hence it has a unique solution. ||F (t.1. satisfies all conditions of Theorem 3. . 1 + na2 . x0 = x1 . x1 ) − f (t. INITIAL VALUE PROBLEMS (0) (0) (0) Here x = (x1 . y)||2 n ≤ |x2 − y2 |2 + |x3 − y3 |2 + · · · + |xn − yn |2 + · · · + |an−1 |2 |xn − yn |2 + n |f (t. where α2 = max nM 2 .6) has a unique solution.1. x(0) = 1 dt Clearly.. f (t. the function f (x) = x2 is Lipschitz continuous on a closed.1 ≤ α2 ||x − y||2 n dx = x2 . . y)|| 2 n Thus the function F appearing in the IVP given by Eq. 1 + na2 1 2 n−1 . . x) − F (t. x) = (x2 . It is clear that F is continuous on I × Ωn and further x → F (. (3. (3. Example 3.1. xn . t∈I Hence we have. xn ∈ n and ¯ ¯ F (t. 1 + na2 .7). xn ). x1 ) − a1 (t)x2 − · · · − an−1 (t)xn ) . y1 )|2 + |a1 (t)|2 |x2 − y2 |2 ≤ nM 2 |x1 − y1 |2 + (1 + n|a1 (t)|2 )|x2 − y2 |2 + · · · + 1 + n|an−1 (t)|2 |xn − yn |2 ≤ α2 |x1 − y1 |2 + |x2 − y2 |2 + · · · + |xn − yn |2 . x2 . . · · · . This in turn implies that the nth order differential equation given by Eq.1 this IVP has a unique solution. x(1) (t) = 1 + t x(2) (t) = (1 + t + t2 ) + t3 3 . Hence by Theorem 3. · · · . . x) − F (t.

The actual solution x(t) = 1−t 4 x3 3 x2 x1 1 2 -1 -0.25 0.3.1. For small t .5 0.75 -0. The largest interval on which 1−t this solution is defined in (−∞. as envisaged in the Corollary 3.1: First three Picard iterates .1.25 0.1. 1).1.5 -0. the function x(t) has the series expansion given by x(t) = 1 + t2 + t3 + · · · + tn + · · · . (refer Chapter 6).75 t Figure 3. EXISTENCE AND UNIQUENESS THEOREMS 83 We can actually compute the solution of the above IVP by the method of 1 separation of variables and obtain x(t) = .2. The following graph gives first three Picard iterates. 1 is plotted in Figure 3. Thus the Picard iterates x(k) (t) do converge to the actual solution x(t) = 1+t2 + t3 + · · · in a small interval |t| < 1 around t = 0.1.

1. by inspection one can easily compute the following solutions of above IVP : As x1 (t) = t2 . x(0) = 0 dt where f (x) is given by f (x) = √ 2 x. t>0 t≤0 ∀t∈ x2 (t) = 0 That is. Hence Theorem 3. x). .25 0.1.5 0. However. (3. x > 0 0.1) without Lipschitz continuity assumptions on the nonlinear function f (t. → ∞ as x → 0+ . INITIAL VALUE PROBLEMS xt 10 8 6 4 2 t -1 -0.4.1. x > 0.1. the solution is not unique. and hence f is not Lipschitz conx x x tinuous at ‘0’. the above IVP has at least two distinct solutions or in other words.5 -0. This point will discussed in ¯ detail in section 3.2: Actual solution x(t) = 1 1−t Example 3.84 CHAPTER 3.1.2 dx = f (x).25 0.1 This example concludes that there may exist a solution of IVP of the type Eq.75 -0. Remark 3.1 is not applicable.75 Figure 3. However it may not be unique. 0. x≤0 f (x) 2 f (x) = √ .

Yet. x) =   0 dx dt 85 (t.3. It can be easily seen that (t. x) = (0. 0). c is arbitrary constant The solution curve is plotted as under. 0).8) xt 10 8 6 4 2 -6 -4 -2 -2 2 4 6 t -4 Figure 3. x) → f (t. x) is not even continuous at (0.1. 0) (t. .1. (3. EXISTENCE AND UNIQUENESS THEOREMS Example 3. 0) The solution of the above differential equation passing through (0. there are infinitely many solutions passing through (0. x)-plane (x − c)2 − t2 = c2 . of the above differential equation. 0).3 = f (t.1. x)   2tx  2 t + x2 f (t. 0) is given by the following curve in (t.1.3: Solution curve for the above example It may noted that the above IVP has a unique solution passing through the point away form (0. x) = (0.

F (x) = (x2 .4 CHAPTER 3. However. t2 t6 t10 − + 2 240 34560 t9 t5 (4) + x2 (t) = t − 40 3456 (4) x1 (t) = . therefore. x(0) = 0 ˙ dt2 As f (x) = cos x is Lipschitz continuous.1. One can.1. x2 (t) . x(0) = 0. x(t) = (x1 (t).86 Example 3. x2 (t) =x. = 0 t x2 (s)ds = 0 cos x1 (s)ds = t (0) t (0) = 0 ds = t 0 = 0 t x2 (s)ds = (1) (1) t sds = 0 t t2 2 = 0 t cos x1 (s)ds = x2 (s)ds = t (2) (2) t ds = t 0 = 0 sds = 0 t t2 2 s2 ds 2 x1 (t). (4) (4) = 0 cos x1 (s)ds = cos 0 s2 (3) One can approximate cos ( ) by series expansion and obtain x2 (t) and 2 up to some accuracy. resorts to Picard iterative procedure to obtain iterates x(k) (t) converging to the unique solution of the above second order IVP. x(0) = ¯ x 0 dt We have x1 (t) x2 (t) x1 (t) x2 (t) x1 (t) x2 (t) (3) (3) (2) (2) (1) (1) t . it follows by Corollary 3. cos x1 ).2 that the above second order initial value problem has a unique solution. INITIAL VALUE PROBLEMS d2 x = cos x. x2 (t)) ¯ d¯ x = F (¯(t)). a close form solution of the above problem is difficult to obtain. x1 (t) = x(t).

1.10) . x1 ) − f (t. x2 ∈ ¯ ¯ n (3. we shall say that x → f (t. x2 ∈ ¯ ¯ 2 n (3.4. (3.4: Graph of iterates x1 (t). We shall now state and prove the existence ¯ and uniquness theorem for IVP of the type given by Eq. x1 − x2 ) ¯ ¯ ≥c n (2) (4) ≥ 0 ∀ x 1 . EXISTENCE AND UNIQUENESS THEOREMS xt 87 4 x12 3 x14 2 1 -3 -2 -1 1 2 3 t Figure 3.9) It is said to be strong monotone if ∃ c > 0 such that n x 1 − x2 ¯ ¯ n ∀ x1 . Using the notion of monotonicity of Chapter 2 (Definition 2.1. x) under consider¯ ation is not Lipschitz continuous but satisfies other nice property like monotonicity with respect to the variable x. x1 (t) xt 3 2 1 x22 (2) (4) x24 1 2 3 t -3 -2 -1 -1 -2 -3 Figure 3. x1 − x2 ) ¯ ¯ ¯ ¯ (f (t.1.1. x2 (t) There do exist examples where the nonlinear function f (t.5: Graph of iterates x2 (t). x2 ).1.1. x) is monotone in n if ¯ ¯ (f (t.3. x1 ) − f (t.1) under the monotonicity assumption.3). x2 ).

1) in ¯ almost everywhere (a. x) ¯ be strongly monotone and further it satisfies a growth condition of the form f (t.1. Define L as L¯ = x dt dt D(L) is dense in X and L is monotone as tf (L¯. tf ] is defined to be the set of all square integrable functions with values in d¯ x d¯ x n with D(L) = {x ∈ X : ∈ X and x(0) = 0}. x) n ≤ a(t) + b x n . one can show that L is maximal monotone as R(L + λI) = X for λ > 0 (refer Pazy[6]).1.1. x) = x ¯ x(tf ) ≥ 0 ∀ x ∈ L2 [0. (3.13) N is a bounded.1) in almost every¯ where sense.12) in the interval I = [0. x(t)) ¯ (3. tf ] × n → n be continuous and x → f (t. x) x This implies that (L¯. b > 0 (3. ¯ Definition 3.tf ] x ¯ ¯ = 0 d¯ x . x)L2 [0. tf ] satisfying Eq. 0 d¯ x dt dt n x ¯ = ||¯(tf )||2 n − (L¯. in almost every where sense. tf ].1). tf ] is said to satisfy Eq. Set X = L2 [0.1 A function x(t) ∈ L2 [0. Then . x) : [0.1. Proof : We are interested in the solvability of the IVP of the form dx + f (t.4: [N x] = f (t. Here L2 [0. x) = 0 dt x(0) = 0 (3. tf ]. t f ] = I and ¯ satisfies Eq. x(t) dt tf dt n = ||x(tf )||2 n − x(t). Theorem 3.1.1.1. we assume that t0 = 0 and x0 = 0. (3. (3.88 CHAPTER 3.1) for all t ∈ I except on a set of measure zero in I. a ∈ L2 [0. INITIAL VALUE PROBLEMS In the IVP given by Eq. continuous operator on X = L2 [0. tf ] (refer Joshi and Bose .4.1. x) = (Lx.2 Let f (t. tf ] 2 Further. We now define the nonlinear operator N on X as in Example 2.e) sense if x(t) is differentiable for all t ∈ [0. (3. tf ].1.11) Then there exist unique x(t) ∈ L2 [0.

In this section we shall show that the solution x(t) is defined on the maximal interval of existence (α. the IVP of the form Eq. (3. it is strongly monotone as we have tf 89 (N x1 − N x2 . L : D(L) → X is a linear maximal monotone operator with dense domain.5 dx + αx + cos x = 0 dt x(0) = 0 f (x) = αx + cos x is strongly monotone if α > 1. x2 ∈ X. THE MAXIMAL INTERVAL OF EXISTENCE [4]).2 The Maximal Interval of Existence Theorem 3.1.2 establish the fact that under suitable continuity and Lipschitz continuity amd monotoncity assumptions on f (t.1 and Theorem 3. then x(t) = y(t) ∀ t ∈ I. N is a bounded.2. continuous strongly monotone operator defined on the whole space X. x1 − x2 ) = ¯ ¯ ¯ ¯ ≥ c 0 (f (t. Let f : × Ω → n satisfies the assumptions of Theorem 3. ¯ ¯ Thus the solvability of equation Eq. ∞).4. (a.14) Lx + N x = 0 in the space X = L2 [0. x2 (t)). (3.1. (3.12) reduces to the solvability of the operator equation (3. x1 (t)) − f (t.1.1) has a unique solution x(t) on some interval surrounding t0 .1.1) on the intervals I1 . it follows that Eq. If x(t).1. Further. . I2 .1.e. y(t) are solutions of the IVP given by Eq.1. (¯1 (t) − x2 (t))) dt ¯ ¯ x ¯ tf 0 ||x1 (t) − x2 (t)||2 dt = c||x1 − x2 ||2 for all x1 .1.3. in almost everywhere sense in X.1. Theorem 3. Hence by Theorem 2.1 Theorem 3. 3. for t) in a finite subinterval of [0. Assumption 1: Let Ω ⊆ n be an open set containing x0 .12). x). (3. (3. β). Then t0 ∈ I1 ∩ I2 and if I is any open interval around ‘t0 ’ contained in I1 ∩ I2 .1.1. tf ].5. Hence it is follows by Theorem 3. Example 3.1. However.2 that the above initial value problem has a unique solution x(t).14) has a unique solution and hence the unique solvability of Eq.1 Let f satisfies Assumption 1.1 is not applicable.2.

This will then lead to global solvability of IVP given by Eq. Then for each x0 ∈ Ω. we define I ∗ = ∪Ia where Ia is an open interval surrounding t0 . β) can be extended to [t0 . Then by Theorem 3. Proof : As in Theorem 3. ∃ a maximal interval I ∗ on which the IVP given by Eq.1.2. (3. x(t) = y(t) on Ia ⊂ I ⊂ I 1 ∩ I 2 . (3. (3. β) and (α.90 CHAPTER 3.1.1.1) Eq. it follows that. x(t) = xa (t).1). If I ∗ is properly contained in I. β] then we can extend the solution of Eq.1). β). Hence I ∗ = I and thereby implying that x(t) = y(t) ∀ t ∈ I. (3. t∗ + a∗ ) ⊂ I. If (α.1 it follows that x1 (t) = x2 (t).1) has a unique solution xa . The following theorem is an important one in that direction.2. then t0 ∈ (α.2. (3. Also let the maximal interval of existence be not open say of the form (α. In view of the uniqueness of solutions on Ia . (3. a + t0 ) ⊂ I such that Eq.2. (3.2. The intervals [t0 . for.2 Let f satisfies Assumption 1. Again by uniqueness on I0 ⊂ I ⊂ I1 ∩I2 . then one of the end points t∗ of I ∗ is contained in I ⊂ I1 ∩ I2 . Then I ∗ is the largest open interval contained in I on which x(t) = y(t).1) as x a (t) is a unique solution of Eq. x(t) = y(t) for t ∈ I ∗ ∪I0 and I ∗ is proper subinterval of I ∗ ∪ I0 and hence a contradiction of the fact that I ∗ is the largest open interval contained in I on which x(t) = y(t). if t ∈ I1 ∩I2 . it follows that. Define I ∗ = ∪Ia .1. Define x : I ∗ → Ω as follows.1.1) for t ∈ Ia . By definition.1.1. then by Theorem 3.1) has a unique solution. β).2. As both x(t) and y(t) are continuous on I. We wish to find out if the right maximum interval of existence [t0 . ∞).1. I ∗ is maximal. . Without loss of generality we can concentrate on the right maximal interval of existence. where I1 and I2 are two intervals on which Eq. (3.1) has a unique solution. x(t∗ ) = x∗ ¯ dt (3. that is I ∗ = (α. (3.1) on (α. β + a] with a > 0 as in the proof of Theorem 3.1. Also x(t) is a unique solution of Eq. Furthermore.1.1) has solutions x1 (t) and x2 (t). If t0 ∈ I ⊂ I1 ∩ I2 . β) is the maximal interval of existence for IVP given by Eq. lim∗ x(t) = lim y(t) = x∗ ¯ ∗ t→t t→t Now examine the initial value problem dx = f (t. t0 ] are called the right and the left maximal intervals of existence respectively. x(t)). on which Eq. ∃ some open interval Ia = (a − t0 . the maximal interval J is open. (3. INITIAL VALUE PROBLEMS Proof : It is clear that t0 ∈ I1 ∩ I2 .1) has a unique solution on some interval I0 = (t∗ − a∗ . t ∈ Ia This is well defined.1.1. Theorem 3.1.

β) is the maximal interval of existence for given IVP. t=β ˙ y(t) is differentiable on (α. It follows that m= sup f (t. β + a) Thus y is a continuation of the solution x(t) on (α. x(s)) ds t1 ≤ m t 2 − t1 As t1 . β] t ∈ [β. β) and at the end point β we have z(β) = x1 = y(β). ∃ t ∈ (α.1. β) × K. (3. β] with y (β) = f (β.2 with x0 ∈ Ω. THE MAXIMAL INTERVAL OF EXISTENCE 91 Theorem 3. let the solution x(t) of Eq.3.2. β]. as y(t) satisfies the ¯ integral equation t y(t) = x0 + t0 f (s.1) on (α. called continuation theorem.2) Then v(t) is a solution of the IVP given by Eq. β + a). if β < ∞. x) < ∞ (t. / As a Corollary of the above result. β)×K If possible. β) such that x(t) ∈ K. (3.3 Let Ω and f be as defined in Theorem 3. Now we define a function v(t) as v(t) = y(t). Let x1 be this limit. Assume that β < ∞.1. The process of continuation is already described in the proof of Theorem 3.x)∈(α. y(s))ds By Theorem 3. β) y(t) = x1 .1). .1 the equation (3. ∃ a t ∈ (α. t ∈ (α.2. we have t2 x(t1 ) − x(t2 ) ≤ f (s.2.1) be such that x(t) ∈ K ∀ t ∈ (α.1. Hence. Let (α. contradicting the assumption that (α. β + a) and hence z(t) = y(t) on (β − a. β) be the maximal interval of existence of the solution x(t) of the IVP given by Eq. Since t→β− x(t) ∈ K and K is compact.1. For α < t1 < t2 < β. it follows that x1 ∈ K ⊂ Ω. We now examine the IVP  dx = f (t. Then given any compact set K ⊂ Ω. Now define y(t) on (α. it follows that x(t1 ) − x(t2 ) → 0 which by Cauchy’s criterion of convergence in n implies that lim x(t) exists. y(β)). / Proof : Let K be compact set in n .2. t ∈ (α.2) has a unique solution z(t) on some open interval (β − a.3. Since f is continuous on (α. we immediately get the following theorem. β] by x(t).2. (3. β) such that x(t) ∈ K. z(t).2. x(t))  dt  x(β) = x1 ¯ ¯ (3. t2 → β. β) with β < ∞.

1) has a global solution. β) be the right maximal interval of existence of solution x(t) of the IVP given by Eq. x(t) ) n ≤ ||f (t. Proof : This follows from the previous corollary.2.92 CHAPTER 3.2. (3. Proof : We have dx = f (t. we again apply the same theorem to K = x : ||x||2 ≤ ||x(0)||2 e t0 . then we apply the previous theorem to K = {x : ||x|| ≤ ρ} and get the result. Let [t0 . .1).2.1 Suppose Ω = ∃ ρ > 0 such that ||f (t. n .1.1. x) be Lipschitz continuous for all x(t) ∈ Then Eq.2 Let x → f ( . That is. β) ⊂ K. x(t)) dt Taking scalar product with x(t) we get x(t). (3. x(t))|| ||x(t)|| ≤ |a(t)| ||x(t)||2 for ||x(t)|| ≥ ρ This implies that |a(t)| d ||x(t)||2 ≤ ||x(t)||2 dt 2 and hence ||x(t)||2 ≤ ||x(0)||2 e Rβ |a(t)| dt 2 Rt t0 |a(t)| dt 2 for ||x(t)|| ≥ ρ If ||x(t)|| ≤ ρ for all t.. x)|| n n and satisfies the following condition: >ρ ≤ |a(t)| ||x|| ∀ ||x|| n with a(t) as a continuous function.IVP given by Eq. (3. Corollary 3.4 Let Ω. (3.3 with x0 ∈ Ω.2. dx dt = (f (t. INITIAL VALUE PROBLEMS Theorem 3. If ||x(t)|| > ρ. ∞). dx dt = x(t).1) has a global solution.Then the initial value problem given by Eq.1. Assume that ∃ a compact set K ⊂ Ω such that y∈ n y = x(t) for t ∈ [ t0 . That is.1. f be as defined in Theorem 3. x(t) is a global solution. Corollary 3. x(t)). Then β = ∞.1) has a solution x(t) on [t0 .

3.1 Suppose that g(t) is a continuous real valued function satisfying the conditions t (i) g(t) ≥ 0 (ii) g(t) ≤ c + k t0 g(s)ds ∀t ∈ [t0 .2. Lemma 3. x) : I × n → n be continuous with x → −f (t.4.1. we get dx .3.1). we have g(t) ≤ c exp (kt) ∀ t ∈ [t0 . DEPENDENCE ON INITIAL CONDITION 93 ¯ ¯ Corollary 3. Hence Theorem 3. (3.1. x(t)) = 0 dt Taking scalar product with x(t).1) has a global solution. (3.3 Let f (t. x(t)) ≤ −2c||x(t)||2 3.2. t1 ] (c.3 Dependence on Initial Condition In this section we investigate the dependence of the solution of the differential equation d¯ x = f (t. x(t)).1) dt on the initial condition x(t0 ) ∈ n . Then the IVP given by Eq. x(t))) = 0 dt As −f strongly monotone.3. Proof : We have a solution x(t) of Eq. implies that x(t) is a global solution of Eq. k are positive constants.1.1) satisfying dx − f (t. x) strongly monotone and f (t. this gives d ||x(t)||2 dt This implies that ||x(t)||2 ≤ ||x(0)||2 e−2ct ≤ ||x(0)||2 ∀ t This gives that x(t) ∈ K = {x ∈ n : ||x|| ≤ ||x(0)||}.) Then. refer Perko[ 7 ]. For more details on this section. x(t) − (f (t. (3.3. We first state and prove the following lemma ¯ called Gronwall’s lemma. 0) = 0. x(t)) ¯ (3. x(t)). = 2 (f (t. t1 ] .

we have x. y ∈ X = ¯ ¯ C(J.2) n ≤ ||¯0 − y0 || x ¯ n exp (M t) ∀t ∈ J = [t0 . t1 ]. y(s))|| ¯ ¯ n ds Borrowing our earlier notation of Theorem 3.94 Proof : Let G(t) = c + k t1 t0 CHAPTER 3. x(s)) − f (s.3. By second fundamental theorem of calculus ˙ G(t) = kg(t) and hence ˙ G(t) g(t) =k ≤ k ∀ t ∈ [t0 . ¯ Proof : Let x(t). This gives Theorem 3. Ω0 ⊂ Ω. y(s)) ds ¯ t0 t∈J t∈J = y0 + ¯ Substracting the above equations and using triangle inequality we get t ||¯(t) − y (t)|| x ¯ n ≤ ||¯0 − y0 || x ¯ n + t0 ||f (s. with respect to the ¯ ¯ initial conditions x0 and y0 .1. Then. y (t) be unique solutions of Eq.1) depends continuously on the ¯ initial condition x0 . tf ] (3.1. Denote by J the interval of existence ¯ ¯ of solutions.1). G(t) ≥ g(t) and G(t) > 0 on [t0 . we get t ||¯(t) − y (t)|| x ¯ and ||¯(t) − y (t)|| x ¯ n ≤ ||¯0 − y0 || x ¯ n +M t0 ||¯(s) − y(s)|| x ¯ n ds (3. t ∈ [t0 . Ω0 ). Using Lipschitz continuity of f and Gronwall inequality. Then the ¯ solution x(t) of the differential equation (3.1 Let f (t.1.1. t1 ] which is the required result. x) satisfy all conditions of Theorem 3. respectively.3) .1. x(s)) ds ¯ f (s.1.3. (3. t1 ]. Using the corresponding integral equation analog we have t x(t) ¯ y(t) ¯ = x0 + ¯ t0 t f (s.3. t1 ] G(t) G(t) d (log G(t)) ≤ k dt Equivalently G(t) ≤ G(t0 ) exp (kt) ∀t ∈ [t0 . INITIAL VALUE PROBLEMS g(s)ds.

3.3.

DEPENDENCE ON INITIAL CONDITION

95

This gives that ||¯ − y || = sup ||¯(t) − y(t)|| x ¯ x ¯
t∈J
n

≤ ||¯0 − y0 || x ¯
n

n

exp (M tf )

(3.3.4)

This implies that the mapping x0 → x of ¯ ¯

into X = C(J, Ω0 ) is continuous.

We shall give an alternate proof of the above theorem by using the following lemma, using fixed points, rather than Gronwall’s lemma. Lemma 3.3.2 Let (X, d) be a complete metric space and let F, Fk be contraction mappings on X with the same contraction constant α < 1 and with the fixed points x, xk respectively. Suppose that lim Fk y = F y ∀y ∈ X. Then
k→∞

lim xk = x.

k→∞

Proof : Using the error estimate given in Theorem 2.4.1 for Fk , we get
l d(xk , Fk y) ≤

αl d(Fk y, y), 1−α

for any arbitrary y ∈ X and for any l.

Set l = 0 and y = x. We have d(xk , x) ≤ This implies that xk → x. Alternate Proof (of Theorem 3.3.1): Let the space X and constants M, m, δ, τ be as defined in Theorem 3.1.1 . (0) Let xk (t), x(t) be solutions of Eq. (3.3.1) with initial conditions xk and x(0) ¯ ¯ ¯ ¯ (0) respectively and let xk → x(0) . ¯ ¯ Define Fk , F : X → X as under ¯ ¯ [Fk x] (t) = xk + ¯ ¯ [F x] (t) = x(0) +
t0 (0) t t0 t

1 d(Fk x, x) → 0, 1−α

as k → ∞

f (s, x(s)) ds, x ∈ X ¯ ¯ f (s, x(s)) ds, x ∈ X ¯ ¯

We have ||Fk x(t) − x(0) || ≤ ||¯k − x(0) || + mτ ¯ ¯ x ¯
(0)

< δ for k large ( as xk → x(0) ) ¯ ¯

(0)

Thus Fk maps X to itself and so is the case for F (already proved in section 1). Also xk , x are fixed points of Fk and F , respectively with the same contraction ¯ ¯ (0) ¯ constant α = τ m < 1. Also, Fk x → F x for x ∈ X, as xk → x(0) in X. Using ¯ ¯ ¯ ¯ Lemma 3.3.2, we get that the fixed points xk of Fk and x of F also converge. ¯ ¯ This proves the theorem.

96

CHAPTER 3. INITIAL VALUE PROBLEMS

3.4

Cauchy-Peano Existence Theorem

We shall prove an existence theorem for IVP given by Eq. (3.1.1) with the assumptions of only continuity and boundedness of f . For this purpose, we need Schauder’s fixed point theorem - Theorem 2.4.2. We also need the following theorem, called Ascoli-Arzela theorem ( refer Rudin [ 8 ] ), which gives the relative compactness for a class F of continuous functions. Definition 3.4.1 A class F of continuous functions is said to be equicontinous, if given > 0, ∃ δ > 0 such that ||f (t) − f (t )|| < ( δ is independent of f ∈ F ). Definition 3.4.2 F is said to be uniformly bounded if ∃ M independent of f ∈ F such that ||f (t)|| ≤ M, ∀ t and ∀ f ∈ F Theorem 3.4.1 (Ascoli-Arzela theorem) Let F be a class of continuous functions defined over some interval J. Then F is relatively compact iff F is equicontinous and uniformly bounded. Theorem 3.4.2 (Cauchy-Peano theorem) Let E be as defined in Theorem 3.1.1 and (t0 , x0 ) ∈ E. Assume that f : E → n is continuous. Then the initial value problem given by Eq. (3.1.1) has a solution. Proof : Let the space X and constants m, τ, δ, be as defined in Theorem 3.1.1 Recall that X is a complete metric space of all continuous mappings from J = [t0 − τ, t0 + τ ] into Ω0 = {x ∈ Ω : ||¯ − x0 || ≤ δ} and hence is a closed, x ¯ convex bounded subset of the space C(J). As before, define F : X → X as
t

whenever |t − t | < δ for all f ∈ F

[F x(t)] = x0 + ¯ ¯
t0

f (s, x(s)) ds, t ∈ J ¯

Define F = {F x : x ∈ X} ¯ ¯ F is a collection of continuous functions defined on J. F is equicontinous, since
t2

||F x(t1 ) − F x(t2 )|| ¯ ¯

n

t1

||f (s, x(s)) || ¯

n

ds

≤ m|t1 − t2 | ∀ x ∈ X ¯ Further, F is uniformly bounded as
t

||F x(t)|| ¯

n

≤ ||¯0 || x ≤ ||¯0 || x

n

+
t0

||f (s, x(s))|| ¯

n

ds

n

+ mτ ∀ x ∈ X ¯

3.5. EXERCISES

97

Hence by Ascoli-Arzela theorem, F is relatively compact. To apply Schauder’s fixed point theorem, it remains to be shown that F : X → X is continuous.We first observe the following. f : J × Ω0 → n is continuous on a compact set and hence it is uniformly continuous. This implies that given > 0, ∃ δ( ) > 0 such that ||f (s, x(s)) − f (s, y(s))|| ¯ ¯ and this δ depends only on . We have
t
n

<

τ

whenever ||x(s) − y(s)||

n

||F x(t) − F y (t)|| ¯ ¯ This gives

n

t0

||f (s, x(s)) − f (s, y(s))|| ¯ ¯

n

ds

t0 +τ

||¯ − y|| < δ ⇒ ||F x − F y || ≤ x ¯ ¯ ¯ ≤ τ

t0

||f (s, x(s)) − f (s, y(s))|| ¯ ¯

n

ds

τ=

That is, F : X → X is continuous. Thus F : X → X satisfies all assumptions and hence has a fixed point x ∈ X. That is ¯
t

x(t) = x0 + ¯ ¯
t0

f (s, x(s))ds, t ∈ J ¯

This proves that Eq. (3.1.1) has a solution. However, this solution need not be unique. In Example 1.2.1, the existence of a solution of the initial value problem dx dt f (x) = f (x), x(0) = 0 ¯ √ 2 x, x ≥ 0 = 0, x≤0

is guaranteed by Cauchy-Peano theorem. This solution is not unique, as already seen before. For more on existence and uniqueness of IVP refer Amann [1], Arnold [2], Hartman [3] and Mattheij and Molenaar [5].

3.5

Exercises

1. Find the regions (in n × n ) of existence and uniquness of solutions of the IVP associated with the following differential equations.

98 (a)

CHAPTER 3. INITIAL VALUE PROBLEMS √ dx 3 = xt dt dx (b) = |x| dt  √ x x≥0 dx  (c) =  √ dt − −x x ≤ 0   2x t = 0 dx  t = (d)  dt  0 x=0

2. (a) Let x(t) be a solution of the differential equation dx = (t2 − x2 ) sin x + x2 cos x dt which vanishes for t = t0 . Show that x(t) ≡ 0. (b) Find the solution of the IVP dx dt x(t0 ) 3. Show that the IVP d2 x = f (t, x(t)) dt2 x(t0 ) = x01 , x(t0 ) = x02 ˙ is equivalent to the integral equation
t1

= x|x| = 0

x(t) = x01 + (t − t0 )x02 +

t0

(t − s)f (s, x(s))ds, t ∈ J

where J is the interval of existence of solution. 4. Let x(t) and y(t) be two solutions of the nonhomogeneous linear differential equation dn x dn−1 x dx + an−1 (t) n−1 + · · · + a1 (t) + a0 (t)x = b(t) n dt dt dt in the interval J around the initial point t0 . Show that u(t0 )exp(−2k|t − t0 |) ≤ u(t) ≤ u(t0 )exp(k|t − t0 |) where u(t) is given by
n−1

u(t) =
j=0

x(j) (t) − y (j) (t)

3.5. EXERCISES dj x (j) dj x sup|aj (t)|. , y = j and k = 1 + j dt dt j=1 t∈J
n−1

99

x(j) (t) =

5. (a) Show that 0 < x0 < x(t) < π if 0 < x0 < π, where x(t) is the solution of the IVP dx dt x(0) = t sin t = x0

(b) Show that x0 < x(t) < 1 for t ≥ 0 and 0 < x(t) < x1 for t ≤ 0, if x(t) is the solution of the IVP dx = x − x2 dt x(0) = x0 , 0 < x0 < 1 6. Consider the IVP dx dt x(0) = xp = 1

Find the soultion of this problem for different values of p. Find how the maximum interval Imax of existence depends on p. 7. Let x(t) be the solution of Eq. (3.1.1), where x → f (t, x) is Lipschitz continuous from to for all t ∈ . Let x1 (t), x2 (t) be two continuous function on , such that dx1 ≤ f (t, x1 (t)) dt x1 (t0 ) ≤ x0 and dx2 ≥ f (t, x2 (t)) dt x2 (t0 ) ≥ x0 Show that x1 (t) ≤ x(t) ≤ x2 (t) dx1 dt 1 x( ) ¯ π x2 , x3 1 ) π ∀t ≥ t0 8. Find the maximal interval of existence of solution for the IVP = − dx2 x1 dx3 =− , =1 dt x3 dt

= (0, 1,

1983. 10. [3 ] Philip Hartman. Ordinary Differential Equations. Consider the following perturbed IVP (corresponding to the above problem) dy = ax. [6 ] A. Springer-Verlag. Pazy. y(0) = 1 + dt Show that |x(t) − y(t)| ≤ | |e|a|t Graphically. [7 ] Lawrence Perko. Halsted Press. Arnold. for more than one initial guess x0 (t). 1991. Bose. Tata McGraw Hill Publications. M. Watter de Gruyter. Mattheij and J.Verlag. Springer . 1990. Molenaar. 1996. Ordinary Differential Equations : Introduction to Nonlinear Analysis. x(0) = 1 dt the Picard iterates xn (t) converge to the unique solution eat . I. INITIAL VALUE PROBLEMS dx = ax. show that the above estimate is very accurate for a > 0. [4 ] Mohan C Joshi and R. . Geometrical Methods in the Theory of Ordinary Differential Equations. 1974. 1982. Semigroup of Linear Operators and Applications to Partial Differential Equations. Birkauser. [5 ] R. but very inaccurate for a < 0. John Wiley.Verlag. 1983. M. Ordinary Differential Equations in Theory and Practice. Springer . Real and Complex Analysis. Some Topics in Nonlinear Functional Analysis. 1985. [8 ] Walter Rudin.100 9. Differential Equations and Dynamical Systems. References [1 ] Herbet Amann. K. Show that for the IVP CHAPTER 3. [2 ] V.

it helps to study the homogeneous system dx = A(t)x(t) dt (4.1) This is a special case of Eq.1.1. with entries aij (t) and bi (t) respectively. To study the solution of Eq.2).Chapter 4 Linear System And Transition Matrix The focus of this chapter is on the linear system of differential equations.1 Linear System In this chapter we shall be concerned with the properties of the linear system of the form dx = A(t)x(t) + b(t) dt (4. Some of the illustrative examples.1.1. The existence and uniqueness of global solution of Eq.1) of Chapter 3.1. 101 . as continuous functions of t. introduced earlier in Chapter 1.1. wherein the concept of transition matrix plays a crucial role.1) with the initial condition x(t0 ) = x0 (4. its properties and the method of computation. (4.1). (4. 4. (3. We give the definition of transition matrix. (4. Here A(t) is an n×n matrix and b(t) is a vector in n . are recalled for demonstrating the methodology involved in computation.2) We shall first prove the existence of n linearly independent solutions of Eq.1.1(a)) follows from Theorem 3.1.1.

· · · .. + + ··· + cn x1 (t0 ) (n) + cn x2 (t0 ) . . · · · .4) Eq. (4. . . This is a contradiction. Proof : We consider n initial value problems dx = A(t)x(t). .3) + ··· + ··· . x(2) (t). x(2) (t).3) implies that c1 x1 (t0 ) + c2 x1 (t0 ) (1) (2) c1 x2 (t0 ) + c2 x2 (t0 ) .102 CHAPTER 4. Then x(1) (t). x(n) (t) are linearly independent in the interval J. dt x(t0 ) = e(i) . t∈J 1≤i≤n where e(i) are the unit vectors (0. (4. · · · ..2 There exists n linearly independent solutions of Eq. . . .1. This gives us the following theorem.1. The Wronskian W (t). Theorem 4. . · · · .4) is zero). cn (not all zero) such that ¯ c1 x(1) (t) + c2 x2 (t) + · · · + cn x(n) (t) = 0 Eq. (1) (1) (2) (1) (2) x1 (t) x2 (t) . = 0 (4. + ··· (1) (2) c1 xn (t0 ) + c2 xn (t0 ) (1) (2) (n) (4. · · · . · · · . LINEAR SYSTEM AND TRANSITION MATRIX Definition 4. · · · . c2 . . x2 . Hence the set x(1) (t).1. · · · .1. .1 Let x(1) (t).1.1.1. c2 . let x(1) (t). it follows that ci = 0 for 1 ≤ i ≤ n (as RHS of Eq. 0) . This implies that there exist n constants c1 . cn . 1 . + + cn xn (t0 ) (n) = 0 = 0 . xn (i) (i) (i) . x(n) (t) be linearly dependent.1 Let W (t0 ) = 0 for some t0 ∈ J ⊆ .1. 1 ≤ i ≤ n.1. · · · . x(2) (t). As the determinant W (t0 ) of this system is nonzero. xn (t) (n) (n) (n) ··· (2) . Proof : If possible. . Theorem 4. is defined by x1 (t) x1 (t) · · · W (t) = x2 (t) x2 (t) · · · . xn (t) xn (t) · · · Here x(i) = x1 . . (4. . (4. x(n) (t) is linearly independent. x(n) (t) be n-vector valued functions.4) is a linear system of equations in n-unknowns c1 .2) in J.

1. Consider the function y(t) defined as It is clear that + (0) x2 x(2) (t) +···+ (0) xn x(n) (t). · · · . (4. . . x(1) (t). x(2) (t). . LINEAR SYSTEM 103 By Theorem 3. Let (1) (2) (n) x1 (t) x1 (t) · · · x1 (t) x2 (t) W (t) = . (1 ≤ i ≤ n). .1.1. 0 0 ··· 0 0 . x2 (t) . . .1. x(n) (t). xn y(t) = (0) x1 x(1) (t) (0) (0) (0) .3 The dimension of the solution space X of the linear homogeneous system given by Eq. . xn (t) Then 1 0 ··· 0 1 ··· . .2) is n. · · · . (4. . .x(0) = x(0) n (0) (0) (0) (0) (0) = Thus y(t) is a solution of the initial value problem dx = A(t)x(t). the above set of n initial value problems has n unique solutions x(1) (t). . Theorem 4..1.1.1. .4. y(t0 ) = x1 x(1) (t0 ) + x2 x(2) (t0 ) + · · · + x(0) x(n) (t0 ) n = x1 e(1) + x2 e(2) + · · · + x(0) e(n) n x1 . The following theorem gives the dimension of X.1. . Let X be the vector space of solutions of Eq.2 that Eq. x(t0 ) = x(0) dt . Let us assume that x(t0 ) = x(0) . . xn (t) (2) (2) ··· . . . 1 (1) (1) x2 (t) . x2 . x(n) (t) are linearly independent on J. Proof : We have already proved in Theorem 4.2).2) has n linearly independent solutions x(1) (t). (4. x(n) (t) with x(i) (t0 ) = e(i) .1.1. with initial conditions x(i) (t0 ) = e(i) . dy = A(t)y(t) dt Also. with x(0) = x1 . . . xn (t) (n) (n) ··· W (t0 ) = = 1 By Theorem 4. (4. · · · . · · · .1. · · · .2). Let x(t) be any general solution of Eq.

Theorem 4. (2) . xi−1 (t) . Then we have the Abel’s formula t W (t) = W (t0 ) exp t0 T r [A(s)] ds (4. xn (t) (1) xi+1 (t) .5) where T r [A(t)] is defined as n T r[A(t)] = i=1 aii (t) Proof : We have x1 (t) x1 (t) · · · W (t) = x2 (t) x2 (t) · · · . xn (t) (n) (n) (n) .4 Let x(1) (t). x(n) (t) of Eq. x(2) (t). . · · · . .1.1. xn (t) (n) . . The following theorem gives the relation.. it follows that n x(t) = y(t) = i=1 xi x(i) (t) (0) That is.2) and hence X has dimension n. . it helps to know the relation between W (t) and the matrix A(t).104 CHAPTER 4. . . (4. . xn (t) xn (t) · · · This gives x1 (t) x1 (t) . ··· . x1 (t) x1 (t) . . (4. (n) xi (n) (n) (n) (n) xi−1 (t) .1.2) and let t0 ∈ J.6) (t) (t) xi+1 (t) . . (1) xi (1) (1) ··· ··· ··· . ¯ ¯ ¯ · · · . . x(t) is spanned by n linearly independent solutions x(1) (t). . dW (t) = dt n i=1 (1) (1) ··· ··· . . .. ¯ At times. (4. . . LINEAR SYSTEM AND TRANSITION MATRIX By uniqueness theorem of IVP.. .1.1. (1) (1) (2) (1) (2) x1 (t) x2 (t) . x(n) (t) be solutions of Eq.

2). . . . (2) ··· . . dW (t) = dt n i=1 (1) x1 (t) . .1. xn (t) (n) (n) (n) aii (t)xj (t) . ··· That is dW = dt which gives n aii (t)W (t) = W (t)T r[A(t)] i=1 t W (t) = W (t0 ) exp t0 T r[A(t)]dt Corollary 4. . .1.2). . x1 (t) . . (n) xn (t) (n) (n) aij (t)xj (t) · · · . . aij (t)xj (t) . .1. xn (t) (2) (2) . xn (t) (1) (1) aii (t)xj (t) · · · .1 Cosider the nth order linear differential equation in J ⊆ dn x(t) dn−1 x(t) dx(t) + an−1 (t) + · · · + a1 (t) + a0 x(t) = 0 n dt dt dt (4. · · · . . . (4. . the second by ai2 (t) and so on except the ith row and substract their sum from the ith row.1.1. . 0 ≤ i ≤ n − 1 as continuous functions of t ∈ J. . then W (t) = 0 on J. . . . (1) xn (t) ··· Multiply the first row by ai1 (t). . . We get x1 (t) . (k) 105 (t) = dxi (t) = dt (k) n aij (t)xj (t). (4. LINEAR SYSTEM In view of Eq.1 If x(1) (t). . n j=1 i=1 (1) (1) dW (t) = dt n ··· .7) with ai (t). . . Example 4.1. (4. we have xi . x(n) (t) are linearly independent solutions of ¯ ¯ Eq. n j=1 x1 (t) . . . . . .6) gives x1 (t) . aii (t)xj (t) .4. .1. j=1 (k) 1≤k≤n and hence Eq.

.2. .7). x(1) (t) = on t . one can show that Eq. 1 −an−1   x(t) dx dt . .2 Transition Matrix dx = A(t)x(t) dt In the previous section we proved that the homogeneous linear system (4. .1.2.1.106 CHAPTER 4.3 that Eq. dtn Remark 4.1. 0 −a1 ··· 1 .8)    A(t) =       . .8) has n linearly independent solutions x1 (t). x2 (t). x(2) (t) = 1 but W (t) ≡ 0.7) is equivalent to the linear first order homogeneous system dx = A(t)x(t) dt where  0 0 . xn (t) (n) (n) (1) x1 (t) x2 (t) · · · Here xk (t) denotes (n) dn xk (t) . xn (t) and its Wronskian is defined as ¯ ¯ ¯ x1 (t) x1 (t) W (t) = x1 (t) . LINEAR SYSTEM AND TRANSITION MATRIX Proceeding as in Corollary 3. .1. . ..1 The converse of Theorem 4. 0 −a0 1 0 .   Recall that A(t) is the companion matrix of Eq. . (4.1. t2 t are linearly independent functions defined 4. It is now clear from Theorem 4. .1. . (n) (2) (1) ··· ··· ··· ··· xn (t) xn (t) xn (t) .           (4. . 0 −a2 ··· ··· . . . .1) . as we see from the following example. ··· ··· 0 0 .1. (n) (2) (1)     x(t) =      dn−1 x(t) dtn−1 x2 (t) x2 (t) x2 (t) . · · · .1. .1 is not true. (4. (4. .

2. A ¯ b 2 Mn×n = i=1 a(i) ¯ 2 where A = a(1) . Definition 4. x(n) (t) Then. (4. Note that Mn×n is a Hilbert space ( finite dimensional) with inner product and norm defined as n n (A. . A(t)x(2) (t). dt Φ(to ) = I (4. tf ) ⊂ (t0 . B)Mn×n = i=1 a(i) . · · · . A(t)x(n) (t) 107 (4.4. ¯(i) . That is. x(2) (t). Mn×n ]. ¯(n) .2.4) Here the solution Φ(t) of the above differential equation lies in the space Mn×n of n × n matrices. · · · . · · · . a(n) and B = ¯(1) . dt dt dt A(t)x(1) (t). is the solution of the matrix initial value problem dΦ(t) = A(t)Φ(t).2. · · · .5) in the space C[J.1 A nonsingular matrix Ψ(t) is called a fundamental matrix if it satisfies the matrix differential equation given by Eq.2.3) Definition 4. transition matrix Φ(t. ··· . t0 ) = I.2.2. Define a nonsingular matrix Ψ(t) as Ψ(t) = x(1) (t). a(2) .4) is equivalent to the solvability of the Volterra integral equation t φ(t) = I + t0 A(s)φ(s)ds.2 A nonsingular matrix Φ(t. ∞) (4. t ∈ J = (t0 . the solvability of Eq.2. x(n) .3).3) with the initial condition Φ(t0 . ¯ ¯ ¯ b b b Proceeding as before. TRANSITION MATRIX has n linearly independent solutions x(1) .2.2. (4.2. t0 ). Its unique solution φ(t) is given by ∞ t φ(t) = I + n=1 t0 An (s)ds . we have dΨ(t) dt = = That is dΨ = A(t)Ψ(t) dt dx(n) (t) dx(1) (t) dx(2) (t) . t0 ) is called principal fundamental matrix or transition matrix or evolution matrix if it satisfies the matrix differential equation given by Eq.2) (4. (4. x(2) . · · · . ¯(2) .

A3 = −A.1) iff there exists a nonsingular constant matrix C such that Ψ2 (t) = Ψ1 (t)C. (4. Theorem 4.2.2. t0 ) = I+ t0 Ads + t0 A2 dτ ds + · · · (t − t0 )2 = I + A(t − t0 ) + A2 +··· 2! = exp (A(t − t0 )) 0 1 −1 0 A2 = −I. (4.2.108 CHAPTER 4. · · · .2.1) ( refer Brockett [2]). we have φ(t) = I + t0 t0 ··· t0 [A(σ1 )A(σ2 ) · · · A(σn ) dσn · · · dσ1 ].1 Let A(t) be a constant matrix A. .5) or Eq.1 Let Ψ1 (t) be a fundamental matrix of the system given by Eq. This gives Example 4. A2k = (−1)k I.2. A2k+1 = (−1)k A.6) Eq. (4. Equivat σ1 t0 t A(σ1 )dσ1 + t0 A(σ1 )A(σ2 ) dσ2 dσ1 + · · · (4. (4. (4.1) is of the form t t s t0 Φ(t.2. Then the transition matrix of the homogeneous system given by Eq.2.2. (4. Example 4. Then Ψ2 (t) is a fundamental matrix of Eq. A5 = A.2.2 Let A = exp (A(t − t0 )) = I + A(t − t0 ) − I A(t − t0 )3 (t − t0 )2 − 2! 3! (t − t0 )4 (t − t0 )5 + +A +··· 4! 5! (t − t0 )2 (t − t0 )4 + +··· 2! 4! (t − t0 )3 +··· +A (t − t0 ) − 3! cos (t − t0 ) sin (t − t0 ) − sin (t − t0 ) cos (t − t0 ) = I 1− = The following theorem implies that two different fundamental matrices corresponding to the same system differ by a multiple of a constant nonsingular matrix. A4 = I.6) is called the Peano-Baker series for the transition matrix Φ(t) of the linear differential equation given by Eq.2. LINEAR SYSTEM AND TRANSITION MATRIX t σ1 t0 σn−1 where An (t) = lentely.1).

t0 )x0 x(t0 ) = x0 . Then dΨ1 (t) dΨ2 (t) = C = AΨ1 (t) C = AΨ2 (t) dt dt 109 This implies that Ψ2 (t) is a fundamental matrix of Eq. given by x(t) = Ψ(t)Ψ−1 (t0 )¯0 ¯ x where Ψ(t) is a fundamental matrix of Eq. t0 ) = Ψ(t)Ψ−1 (t0 ) Corollary 4. (4. let Ψ2 (t) be another fundamental matrix of Eq.1). dt has a solution x(t).2.2.2 The initial value problem dx = A(t)x(t).4.1) is given by Φ(t. t0 ) of Eq.1).2. (4. (4. Set C(t) = Ψ−1 (t)Ψ2 (t) 1 This gives Ψ2 (t) and dΨ2 (t) dt Hence A(t)Ψ2 (t) = A(t)Ψ1 (t)C(t) + Ψ1 (t) dC(t) dt dC(t) = A(t)Ψ2 (t) + Ψ1 (t) dt Ψ1 (t) and hence dC(t) =0 dt = dΨ1 (t) dC(t) C(t) + Ψ1 (t) dt dt = Ψ1 (t)C(t) This implies that dC(t) =0 dt Thus C is a constant matrix. TRANSITION MATRIX Proof : Let Ψ2 (t) = Ψ1 (t)C.1). Corollary 4.2. the transition matrix Φ(t. (4.2.2.2. (4.2.1 If Ψ(t) is a fundamental matrix of Eq.1). Ψ2 (t) are invertible. Conversely. That is x(t) = Φ(t. Also C is invertible as Ψ1 (t).

x(n) with {x(i) }n linearly independent. it follows that every solution x(t) of Eq. Ax(2) .2 Φ(t. LINEAR SYSTEM AND TRANSITION MATRIX Proof : We have dΨ(t) = A(t)Ψ(t) dt This gives dx(1) (t) dx(2) (t) dx(n) (t) = Ax(1) . (4. Then. x(2) .1) and as its solution space is of dimension n. Remark 4. we get ¯ This gives x(t) = Ψ(t)Ψ−1 (t0 )x0 = Φ(t. . t0 ) = Φ(t. . (4.2.1 If Ψ1 (t) and Ψ2 (t) are two fundamental matrices of the systems dx dx = A1 (t)x(t) and = A2 (t)x(t). . Ax(n) . cn ). t0 ) satisfies the following properties (i) Φ(t.110 CHAPTER 4.2. . ··· dt dt dt where Ψ(t) = x(1) . . c2 . · · · . i=1 This implies that dxi 1≤i≤n = Axi . Ψ1 (t) = dt dt Ψ2 (t). t0 ). ¯ Since x(t0 ) = x0 .2.1) is given by x(t) = c1 x(1) (t) + · · · + cn x(n) (t) That is x(t) = Ψ(t)c where c = (c1 . . dt dΨ1 (t) −1 Ψ1 (t) and dt Remark 4. respectively (A1 = A2 ). . t0 )x0 c = Ψ−1 (t0 )x0 We also note the following facts. dt Thus we have n linearly independent solutions of Eq. t1 )Φ(t1 .2. This follows from the fact that A1 (t) = A2 (t) = dΨ2 (t) −1 Ψ2 (t).

then Ψ−1 (t) is a fundamental matrix of the adjoint system. t0 ) = Ψ(t)Ψ−1 (t0 ).2. 111 We shall say a linear homogeneous differential equation in p(t) is adjoint equa¯ d¯ x tion associated with the given equation = A(t)¯(t). t) (iii) Φ(t.2 The adjoint equation associated with d¯ x = A(t)¯(t) x dt is d¯ p = −A(t) p(t) ¯ dt d [(¯(t). we get dΨ−1 (t) dΨ(t) −1 Ψ (t) + Ψ(t) =0 dt dt and hence A(t) + Ψ(t) This gives dΨ−1 (t) =0 dt dΨ−1 (t) = −Ψ−1 (t)A(t) dt .3 If Ψ(t) is a fundamental matrix of the system given by Eq. Theorem 4. p(t)) = constant x ¯ That is Eq. p(t))] = x dt d¯(t) p d¯ x . (4.2.2.2.2. provided that for any x dt initial data. ¯ ¯ dt dt (4. t0 . just observe that Φ(t.4. t1 ∈ J To realise this. p(t)) ¯ ¯ Theorem 4. (4.1).7) is the adjoint equation associated with Eq. This gives the following theorem. Proof : We have Ψ(t)Ψ−1 (t) = I. TRANSITION MATRIX (ii) Φ−1 (t. Differentiating this equality.1). p(t)) + x(t). t) = I for all t. −A(t) p(t) x ¯ ¯ ¯ = 0 This gives (¯(t).7) Proof : We have = (A(t)¯(t).2. p(t) + x(t). (4. t0 ) = Φ(t0 . the scalar product (x(t). = ([A(t) − A(t)] x(t). p(t)) is constant for all t.2.

3 If Ψ(t) is a fundamental matrix of Eq.8). then the transition matrix is orthogonal [Φ(t.8). (4.3 If A(t) is skew symmetric. Corollary 4.9) (4. t0 ) = A(t) Φ(t. t0 )φ(t. (4. t0 )] Φ (t. Plugging this representation of x(t) in Eq. t0 )v(t) which satisfies Eq.1) then χ(t) is fundamental matrix of its adjoint system given by Eq.8) . We shall now give the method of variation of parameter to solve the nonhomogeneous system dx = A(t)x(t) + b(t) dt We set x(t) = Φ(t. (4. This follows from the fact that (i) d φ (t. t)b(t) dv(t) = A(t)Φ(t. t0 ) = 0 and dt (ii) φ(t0 .2.1 gives us the following corollary. that is. t0 ) A (t) + A(t) φ(t.2. t0 ) v and hence dv(t) dt = Φ−1 (t.2.2. t0 ) = I. t0 ) v(t) + Φ(t. LINEAR SYSTEM AND TRANSITION MATRIX dΨ−1 (t) dt = −A(t) Ψ−1 (t) Theorem 4. we get ¯ dv(t) d Φ(t. t0 )¯(t) + b(t) v dt (4.7) iff χ(t) Ψ(t) = C where C is a constant nonsingular matrix. t0 ) v(t) + b(t) dt dt This gives A(t)Φ(t.2. t0 ) = I . t0 )¯(t) + Φ(t. v(t) to be determined. − [A(t)] = A(t).2.2. t0 )b(t) = Φ(t0 . Remark 4.112 and hence CHAPTER 4.2. t0 ) = φ (t.2. (4.

t0 ) = cos (t − t0 ) sin (t − t0 ) − sin (t − t0 ) cos (t − t0 ) t By the method of variation of parameter we have x(t) = x(0) + where x(0) = x(t0 ). s)b(s)ds t0 Φ(t.2. t0 )x0 + t0 Φ(t. b(t) = t 0 A= The transition matrix for this system is given by Φ(t. t0 )v(t0 ) + t0 Φ(t. s)b(s)ds Example 4. TRANSITION MATRIX Integrating this equation.4. t0 )v(t). s)b(s)ds This gives us the variation of parameter formula t x(t) = Φ(t.2. t0 )Φ(t0 . s)b(s)ds =      t t0 t t0 s cos (t − s)ds −s sin (t − s)ds     = This gives the solution of the   (0)  x1 x1 (t) =   (0) x2 (t) x2 t0 sin (t − t0 ) + 1 − cos (t − t0 ) t0 cos (t − t0 ) − 1 + sin (t − t0 ) For more on this section refer Agarwal and Gupta [1]. non-homogeneous system as    t0 sin (t − t0 ) + 1 − cos (t − t0 )    + t0 cos (t − t0 ) − 1 + sin (t − t0 )  . t t0 Φ(t. we get t 113 v(t) = v(t0 ) + t0 Φ(t0 . s)b(s)ds As x(t) = Φ(t.3 Solve dx = Ax(t) + b(t) dt 0 1 −1 0 . it follows x(t0 ) = v(t0 ) and t x(t) = Φ(t.

3.114 CHAPTER 4.1). T )e−Rt = P −1 (t) . the associated transition matrix φ(t. Proof : As φ(T. 0)e−R(t+T ) = φ(t + T.1b) We shall make use of the following lemma. t∈ Let us first consider a homogenous periodic system (4. noindent The Peano-Baker series representation of the transition matrix φ(t. LINEAR SYSTEM AND TRANSITION MATRIX 4. then ∃ a matrix R such that C = eR . t0 ) of the system given by Eq. 0)e−RT e−Rt = φ(t + T. If A(t + T ) = A(t). We now define a nonsingular P (t). We observe that Eq. t0 ) is also periodic of period φ(t + T. it follows that φ(t. through its inverse P −1 (t) as P −1 (t) = φ(t. 0) = eRT (redefine R in Lemma 4. 0) is a nonsingular matrix.3. t0 ). t0 ) = P −1 (t)eR(t−t0 ) P (t0 ) where P (t) is a nonsingular periodic matrix function and R is a fixed matrix.3.3 Periodic Linear System d¯ x = A(t)¯(t). T )φ(T. t0 + T ) = φ(t.1 Let C be any nonsingular matrix.1 (Floquet-Lyapunov). which we shall state without proof (refer Coddington and Levinson [3] for further details). (4. Theorem 4.1a) (4.1) has the form φ(t.1. 0)e−Rt Then P −1 (t + T ) is satisfies the relationship P −1 (t + T ) = φ(t + T.1) is a non-autonomous system. t0 ) in view of the periodicity of A(t). there exists a matrix R such that φ(T. t ∈ x dt A(t + T ) = A(t). Lemma 4.3. by Lemma 4.3.3. yet it is possible to get a simpler representition of the transition matrix φ(t.3.3. t0 ) gives t t σ1 φ(t. (4. t0 ) = I + t0 A(σ1 ) dσ1 + t0 t t0 A(σ1 ) t0 A(σ2 )dσ2 dσ1 + · · · As T : t+T t0 +T A(σ) dσ = A(σ) dσ.

3. PERIODIC LINEAR SYSTEM That is P (t) is periodic.2) (4. 0)φ(0. Also we have φ(t.3. It is a pleasant surprise to observe that if we make the transformation x(t) = P −1 (t)¯(t) ¯ z then z (t) satisfies the autonomous system ¯ d¯ z = Rz(t) dt This can be seen as follows. t) = e−Rt P (t) Eq.3. d −1 P (t) eR(t−t0 ) P (t0 ) + P −1 (t) ReR(t−t0 ) P (t0 ) dt = A(t)P −1 (t) eR(t−t0 ) P (t0 ) which gives P −1 (t)R = A(t)P −1 (t) − Operating Eq. (4.2) imply that φ(t. we have dt 115 (4.4. t0 ) = P −1 (t)eRt e−Rt0 P (t0 ) = P −1 (t)eR(t−t0 ) P (t0 ) This proves the theorem. Since dφ = A(t)φ(t).3) (4.5) d −1 P (t) z (t) ¯ dt Using transformation given by Eq. t0 ) = φ(t. (4.3) we get d¯ x d −1 d¯ z = P (t) z (t) + P −1 (t) ¯ dt dt dt .3.3. 0) = P −1 (t)eRt .3.5) on z (t) we get ¯ P −1 (t)R¯(t) = A(t)P −1 (t)¯(t) − z z = A(t)¯(t) − x d −1 P (t) z (t) ¯ dt (4.4) d P −1 (t)eR(t−t0 ) P (t0 ) = A(t) P −1 (t)eR(t−t0 ) P (t0 ) dt Equivalently.3. φ(0.3. (4.6) d −1 P (t) dt (4.3.

3.3. (4. refer Brocket[2] for a proof of this theorem. 0) = cos t − sin t sin t cos t e−2t 0 0 1 = P −1 (t)eRt where P (t) = φ(t.7a) (4.3. One observes that φ(t.1 Consider the following periodic system d¯ x dt A(t) A(t + 2π) One can verify that φ(t) =   e−2t cos t − sin t e −2t = A(t)¯(t) x = −2 cos2 t 1 − sin 2t −1 − sin 2t −2 sin2 t = A(t) sin t cos t   is a fundamental matrix of this system.7b) cos t − sin t sin t cos t is periodic of period 2π and R = . Example 4.3. ¯ + T ) = ¯ b(t b(t).6) reduces to P −1 (t)R¯(t) = A(t)¯(t) − z x = P −1 (t) This gives d¯ z dt d¯ x d¯ z + P −1 (t) dt dt d¯ z = R¯(t) z dt which is the desired result. LINEAR SYSTEM AND TRANSITION MATRIX and hence Eq. 0) = −2 0 0 0 Let us now examine the non-homogeneous initial value periodic system d¯ x = A(t)¯(t) + ¯ x b(t) dt x(t0 ) = x0 ¯ ¯ where A(t + T ) = A(t).116 CHAPTER 4. (4. We state the following theorem regarding the solvability of Eq. (4.7).3.

t0 ) [x0 − xp (t0 )] ¯ ¯ ¯ ¯ d¯ p p = −A (t)¯(t) dt (¯ is orthogonal to p iff b ¯ t0 +T t0 (4.3. (4. PERIODIC LINEAR SYSTEM 117 Theorem 4.7) can be written as x(t) = xp (t) + φ(t.3. ¯ ¯ b(σ) dσ = 0. Then the solution of the non-homogeneous system given by Eq.8) iff ¯ is orthogonal to the solution p(t) of the adjoint equation b(t) ¯ (4.9) p(σ).2 Consider the harmonic oscillator problem with periodic external force b(t) (refer Example 1.2 Let φ(t.10) with x1 (t) = x(t).4. x2 (t) = x(t).3. ω 2 = 2 dt ma This is equivalent to  dx1  dt    dx2 dt ⇐⇒      = 0 −ω 2 1 0 x1 x2 = 0 b(t) d¯ x dt = A¯(t) + ¯ x b(t) (4. t0 ) =  1  0 ω − sin ω(t − t0 ) cos ω(t − t0 ) 0 ω  cos ω(t − t0 ) −ω sin ω(t − t0 )  1 sin ω(t − t0 )  ω  cos ω(t − t0 ) Hence the solution of the homogeneous equation corresponding to Eq.1) λ d2 x + ω 2 x(t) = b(t).3.10) is given by xH (t) = x0 cos ω(t − t0 ) + (1) (1)  =  x0 sin ω(t − t0 ) . (4.3. ˙ The eigenvalues of A are ±ωi and hence the transition matrix corresponding to A is given by   1 0  cos ω(t − t0 ) sin ω(t − t0 ) 1 0    φ(t.3.3.3. t0 ) be the transition matrix generated by A(t). ω (2) (2) −x0 ω sin ω(t − t0 ) + x0 cos ω(t − t0 ) .3.) Example 4.

118 This gives CHAPTER 4. This follows from the fact n n (4.3. b(t) satisfies the orthogonality condition    cos ω(t − t0 ) ω sin ω(t − t0 ) 0 t0 +T    dt = 0 1 t0 − ω sin ω(t − t0 ) cos ω(t − t0 ) b(t) t0 +T The transition matrix of the adjoint system is   cos ω(t − t0 ) ω sin ω(t − t0 )   1 − ω sin ω(t − t0 ) cos ω(t − t0 ) sin ω(t − t0 ) b(t)d(t) = 2π .1) and S = P QP −1 . [1] If P and Q are linear transformation on exp (S) = P exp (Q)P −1 .4 Computation of Transition Matrix We shall discuss some basic properties of the linear transformation exp A in order to facilitate its computaton. We note that exp (A(t − t0 )) is the representation of the transition matrix of the system dx = Ax(t) dt where A is independent of t. That is. LINEAR SYSTEM AND TRANSITION MATRIX xH (t) = xH (0) cos ω(t − t0 ) + d¯ p = dt xH (0) ˙ sin ω(t − t0 ) ω (4. We note the following properties. then exp S = lim n→∞ k=0 (P QP −1 )k k! n = P n→∞ lim k=0 Qk k! P −1 = P exp (Q)P −1 .4. ω t0 cos ω(t − t0 ) b(t)d(t) = 0 where T is the time period 4. (4.11) The adjoint system corresponding to Eq.3.10) is given by 0 ω2 −1 0 p(t) ¯ This implies that b(t) satisfies the conditon t0 +T t0 Hence the non-homogeneous equation will have a solution iff p(t) is orthogonal ¯ to ¯ b(t).

B commutes with I and B 2 = 0. t0 ) = exp (A(t − t0 )) = P diag[ exp (λj (t − t0 )) ] P −1 Here.1 If A = a b 0 a . then 1 b(t − t0 ) 0 1 exp (A(t − t0 )) = exp (a(t − t0 )) This follows from the fact that A = aI + B. we have a −b b a This gives us exp (A) as ∞ k = Re(λk ) −Im(λk ) Im(λk ) Re(λk )) exp(A) = k=0   Re Im λk k! λk k! −Im Re λk k! λk k!   = Re [exp (λ)] −Im [exp (λ)] Im [exp (λ)] Re [exp (λ)] cos b − sin b sin b cos b = exp (a) .4. Using property[2]. B = 0 b 0 0 Also.4. then Φ(t. [2] If P and Q are linear transformation on n 119 which commute then exp (P + Q) = exp (P ) exp (Q) Example 4. COMPUTATION OF TRANSITION MATRIX This implies that if P −1 AP = diag[λj ].4. we get the required representation as follows. By induction. then exp (A) = exp (a) Denote λ = a + ib. diag[λj ] represents the diagonal matrix with entries λj . exp (A(t − t0 ) ) = exp (a(t − t0 )I ) exp B(t − t0 ) = exp (a(t − t0 ) )[I + B(t − t0 )] 1 b(t − t0 ) = exp (a(t − t0 ) ) 0 1 cos b − sin b sin b cos b [3] If A = a −b b a .

λ2 .2) λ1 u(1) .2) by λn and subtract from Eq. (4. · · · .4. . un forms a basis for n . we have ¯ ¯ AP = = A¯(1) . . we get α1 (λ1 − λn )¯(1) + α2 (λ1 − λn )¯(2) + · · · + αn−1 (λn−1 − λn )¯(n−1) = ¯ u u u 0 By induction u(1) . = P  = P diag [λj ] . · · · .4. λn . . u(n) is nonsingular.2) we get αn λn u(n) = 0 which gives αn = 0 as ¯ λn u(n) = 0. . . .4.4. (4. . 1 ≤ i ≤ ¯ ¯ n − 1. · · · . u(n−1) are linear independent and hence αi = 0. · · · . . A¯(n) u u (4. We claim that u(i) . The matrix P = u(1) . · · · . we have the following theorem. the result is true for k = 1 as u(1) = 0 is linearly independent. λn } of an n × n matrix A is real and distinct.. λ2 .3) (4. Further.3). Putting these values in Eq. . t0 ) is given by   exp (λ1 (t − t0 )) 0 ··· 0       0 exp (λ2 (t − t0 )) · · · 0     −1 φ(t. . t0 ) = P  P   . Hence the matrix P = ¯ ¯ u(1) . LINEAR SYSTEM AND TRANSITION MATRIX In a more general setting. u(i) = ¯ λi = λj u ¯ ¯ 0. ¯ Thus u(1) · · · u(n) are linear independent by induction. ¯ ¯ Obviously.4. α1 u(1) + α2 u(2) + · · · + αn u(n) = ¯ αi ∈ ¯ ¯ ¯ Operating by A we get α1 λ1 u(1) + α2 λ2 u(2) + · · · + αn λn u(n) = ¯ ¯ ¯ ¯ 0 Multiply Eq. then any set of corresponding eigenvectors u1 . u(n) is invertible and P −1 AP = diag [λ1 . . . u(n) are linearly independent. Let 0. So let ¯ the result be true for k ≤ n − 1.     0 0 0 exp λn (t − t0 ) Proof : We have A¯(i) = λi u(i) .4. We proceed by induction. · · · . We shall show that it is also true for all k = n. λn u(n) ¯ ¯   λ1   .   . · · · .. . Theorem 4. (4.120 CHAPTER 4. · · · . λn ] and hence the transition matrix φ(t.1 If the set of eigenvalues {λ1 .

u(n) .4. λj = aj − ibj with corresponding eigenvectors w (j) = u(j) ± i v (j) . 1 ≤ j ≤ n. COMPUTATION OF TRANSITION MATRIX This implies that P −1 AP = diag [λj ] Proceeding as before. 0).2 −1 −3 0 2 λ1 = −1. 1). t0 ) has the representation. t0 ) = P exp diag  aj −bj bj aj  P −1 cos (bj (t − t0 )) sin (bj (t − t0 )) cos (bj (t − t0 ))  = P diag (exp (aj (t − t0 )))  − sin (bj (t − t0 ))  P −1 . . 0 0 eλ2 (t−t0 ) .4. A= P = 1 −1 0 1 . ··· 0 0 eλn (t−t0 )    −1 P  121 Example 4. 0 ··· ··· .4. . v (n) is invertible and ¯ a j bj with 2 × 2 blocks along the diagonal . . (−1. we state the following theorem concerning the computation of the transition matrix.. λ2 = 2. v (1) .2 Let A be a 2n × 2n matrix with 2n distinct complex eigenvalues λj = aj + ibj . The tranP −1 AP = diag −bj aj siton matrix φ(t. . φ(t. u(n) . P −1 = 1 1 0 1 This gives Φ(t − t0 ) = exp (A(t − t0 )) = P =    exp (−(t − t0 )) 0 0 exp (−2(t − t0 ))   P −1   exp (−(t − t0 )) exp (−(t − t0 )) − exp (2(t − t0 )) 0 exp (2(t − t0 )) If eigenvalues are complex. t0 ) is given by eA(t−t0 ) = P exp ( diag [λj (t − t0 )]) P −1   =P   eλ1 (t−t0 ) 0 . . v (1) . the transition matrix φ(t. · · · . Theorem 4.4. A pair of corresponding eigenvectors is (1. Then u(1) . · · · . v (n) is a basis for 2n and the matrix P = u(1) .

w (2) = (0.4 = 2 ± i. 0. LINEAR SYSTEM AND TRANSITION MATRIX   1 −1 0 0  1 1 0 0   Example 4. 1) P =  v (1) ¯ 0 1 0 0 0 1 0 0 −1 1 0 0 u(1) ¯ 0 0 1 0 v (2) ¯  0 0   1  1 u(2) ¯ P −1 P −1 AP 1  0 =   0 0  1  0 =   0 0  1  1 =   0 0  0 0 0 0   1 −1  0 1  0 0 0 0   2 −1  1 2  Φ(t. 0. 0) . A corresponding pair of eigenvectors is w (1) = (i. . 1 + i. λ3. 1.122 CHAPTER 4.3 A =   0 0 3 −2  0 0 1 1 A has eigenvalues λ1. t0 ) =    exp (t) sin t  P   0   0  exp (t) cos t − exp (t) sin t exp (t) cos t 0 0 0 0 0 0     −1 P  exp (2t) cos t − exp (2t) sin t    exp (2t) sin t exp (2t) cos t 0 0 exp (2t) (cos t + sin t) 2 exp (2t) sin t 0 0 −2 exp (2t) sin t exp (2t) (cos t − sin t)           = exp (t) cos t − exp (t) sin t    exp (t) sin t exp (t) cos t     0 0   0 0 For multiple eigenvalues of A we proceed as follows.2 = 1 ± i.4.

.2 and Theorem 4. where P −1 SP = diag [λj ]. . . .1 let λ be an eigenvalue of n×n matrix A of multiplicity m ≤ n. t0 ) = P diag [exp (λj (t − t0 ))] P −1 × 1 + N (t − t0 ) + · · · + N k−1 If λ is of multiplicity n. Theorem 4.4. m. Definition 4. v (n) ] invertible and A = S + N. . t0 ) is computed as follows Φ(t.4. The transition matrix φ(t. λn repeated according to their multiplicity.3 Let A be a real n × n matrix with eigenvalues λ1 . . Then for k = 1 .4.2 A matrix N is said to be nilpotent of order k if N k−1 = 0 and N k = 0. COMPUTATION OF TRANSITION MATRIX 123 Definition 4. . .4. then S = diag[λ]. S = [I]4×4 and N = A − S. For proofs of Theorem 4.  0 −2 −1 −1  1 2 1 1   Example 4. any non zero solution v of [(A − λI)k ]v = 0 is called a generalized eigenvector of A.4 A =   0 1 1 0  0 0 0 1 λ = 1 is of multiplicity 4. . . . Then there exist a basis of generalized eigenvectors v (1) . the matrix N = A − S is nilpotent of order k ≤ n and S and N commute. v (2) .3.4.  N  −1 −2 −1 −1  1 1 1 1   =   1 2 1 1  0 1 0 0   −1 −1 −1 −1  0 0 0 0   =   1 1 1 1  0 0 0 0  (t − t0 )k−1 (k − 1)! N2 N3 = 0 Φ(t.4. v (n) with P = [v (1) .4. v (2) . refer Perko[4]. .4. Further. . t0 ) = exp (t − t0 ) I + N (t − t0 ) + N 2 (t − t0 )2 2! . .

−2 ).4. λ2 = λ3 = 2 v (1) = (1. 1 ). v (3) . we get  P  1 0 0 =  1 0 1  −2 1 0  0 0 0 =  0 0 0  −1 1 0  0 exp (2t) 0 exp (t)  P −1  1 0 0 = 2 0 1  −1 1 0  N exp (t) 0 Φ(t. 0) = P  0  =       0  P −1 [1 + N t] 0 exp (2t) 0 exp (2t) 0 exp (t) − exp (2t) 0       − exp (2t) + (2 − t) exp (2t) t exp (2t) exp (2t) . obtained by solving (A − 2I)2 v = 0. is given by   1 0 0 0 =  1 0 0 v −2 0 0 That is v (3) = (0. 0.5 A =  −1 2 0  1 1 2 Here λ1 = 1.124 = CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX exp (t − t0 ) ×   1 − [(t − t0 ) −[2(t − t0 ) −[(t − t0 ) −[(t − t0 )   + + + +    (t − t0 )2 (t − t0 )2 (t − t0 )2 (t − t0 )2   ] ] ] ]    2 2 2 2       (t − t0 ) 1 + (t − t0 ) (t − t0 ) (t − t0 )        2  2 2 2 (t − t0 )  (t − t0 ) (t − t0 )  (t − t0 ) (t − t0 ) + 1+     2 2 2 2   0 0 0 1   1 0 0 Example 4. 1. 1. v (2) = (0. 0) So.

4) 2 dt dt λ .1. passing through the initial point x(t0 ) = x0 .4). x(t0 ) = x2 as ˙ x(t) = ea(t−t0 ) x(t0 ) cos b(t − t0 ) + k x(t0 ) ˙ sin b(t − t0 ) + sin b(t − t0 ) 2b b (ii) Eigenvalues of A are real and distinct if k 2 − 4ω 2 > 0 and are given by k 1 −k ± c − ± k 2 − 4ω 2 = . 2 2 2 .4.4. c = k 2 − 4ω 2 . (1) (2) (4. k is the resistance and ω 2 = ma This is equivalent to the following non-homogenous linear system d¯ x dt 0 1 −ω 2 −k = A¯ + ¯ x b(t) = 0 −ω 2 1 −k 0 F (t) x+ ¯ A= . which is modelled by the second order differential equaiton d2 x dx + ω 2 x = F (t) +k (4. t0 ) =  ea(t−t0 )  = k 1 sin b(t − t0 )  cos b(t − t0 ) + 2b sin b(t − t0 ) b a(t−t0 )  e   ω2 k sin b(t − t0 ) cos b(t − t0 ) − sin b(t − t0 ) − b 2b        1 k − 2 0 b    cos b(t − t0 ) − sin b(t − t0 )  1   k cos b(t − t0 ) 2b sin b(t − t0 ) 0 1 b    ω 2 − (k 2 /4) .4.4. COMPUTATION OF TRANSITION MATRIX 125 Example 4. This gives the solution of the homogeneous equation corresponding to Eq.6 Let us consider the problem of mechanical oscillations discussed in Example 1.3. ¯= b 0 F (t) (i) Eigenvalues of A are complex if k 2 − 4ω 2 < 0 and are given by − k i k ± 4ω 2 − k 2 = a ± bi a = − and b = 2 2 2 Hence the transition matrix is given by φ(t.4.

126 CHAPTER 4. the transition matrix corresponding to A is given by φ(t. LINEAR SYSTEM AND TRANSITION MATRIX So. t0 ) = −k + c (t − t0 )  1 1 2  e     −k + c −k − c   2 2 0   1 k+c  2c c    ×   −k + c 1  − 2c c    =          0 −k − c 2 e (t − t0 )        −k + c 2 e −k + c 2 k+c 2c (t − t0 ) (t − t0 ) −k − c 2 e −k − c 2 (t − t0 ) −k + c 2 e −k − c 2 e       (t − t0 )     ×  −k + c 2c      1  − c (t − t0 ) −k + c 1 2  e c  − −k − c 1 2 e c  (t − t0 )  1 c                   =                 −k + c k + c 2 e  2c + −k − c −k + c 2 e 2c −k + c  c2 − k 2 2 e  4c  + −k − c k −c 2 e 4c 2 2 (t − t0 )    (t − t0 )    (t − t0 )  −k + c e  2c −k + c 2 + (t − t0 ) (t − t0 )    −k − c k+c 2 e c (t − t0 )                                      .

COMPUTATION OF TRANSITION MATRIX 127 Hence the solution of the homogeneous equation passing through the initial point is given by x(t) =  −k + c k + c 2 x(t0 )  e 2c −k + c 1 2 + x(t0 )  e ˙ c  (t − t0 ) −k − c −k + c 2 + e 2c (t − t0 )       −k − c (t − t0 ) 1 2 − e c (t − t0 ) =   −c k   k + c c (t − t0 ) −k + c − (t − t0 )  + e 2 e 2 e 2 x(t0 )  2c 2c  k (t − t0 ) x(t0 ) ˙ +e 2 − c c 1 (t − t0 ) 1 − (t − t0 ) − e 2 e2 c c k+c 1 x(t0 ) + x(t0 ) ˙ 2c c   (t − t0 )      k c − (t − t0 ) (t − t0 ) = e 2 e2 −   k c −c (t − t0 ) (t − t0 ) (t − t0 ) αe 2 .4. β= x(t0 ) − 2c c 2c c −k −k (iii) If k 2 − 4ω 2 = 0. = e 2 + βe 2 −  k −c (t − t0 ) (t − t0 ) e 2 +e 2  −k + c 1 x(t0 ) − x(t0 )  ˙ 2c c α= k+c x(t0 ) ˙ −k + c x(t0 ) ˙ x(t0 ) + .4. So. the 2 2 transition matrix is given by −k   e 2 φ(t − t0 ) =     (t − t0 ) 0 −k e 2 0 (t − t0 )    [I + N (t − t0 )]  . then the eigenvalues of A are repeated . .

LINEAR SYSTEM AND TRANSITION MATRIX where N is the nilpotent matrix given by k  −2  = A−  0    =   This gives φ(t.128 CHAPTER 4. t0 ) =  −k (t − t0 )   e 2     0   k 2 1 0  N −ω 2    k  − 2    k  − 2  0 −k e 2 (t − t0 )                 =     k (t − t0 )  1 + 2 (t − t0 )  ×  k −ω 2 (t − t0 ) 1 − (t − t0 ) 2 −k k 1 + (t − t0 ) e 2 2 −k −ω 2 (t − t0 )e 2 (t − t0 ) −k (t − t0 )e 2 k 1 − (t − t0 ) e 2 (t − t0 ) −k 2 (t − t0 )        (t − t0 )   Hence. the solution of the homogeneous system is given by −k (t − t0 ) k = x(t0 ) 1 + (t − t0 ) e 2 2 −k (t − t0 ) +x(t0 )e 2 ˙ (t − t0 ) x(t) .

(t − t0 ) β= k x(t0 ) + x(t0 ) ˙ 2 Example 4. (0. ±ωi. 2. B =   0 1  0 0  0 0   0  1 To compute the transition matrix of the above problem. x2 . where S = P diag[Bj ]P −1  1 0 1  0 0 0 =   0 1 0 3 − 2 ω 0 2ω  0 1 0  3ω 2 0 0 =   6ω 0 0 0 −2ω 0  0 0 0 0 ω  0 0 0  2  0 0 0 0 0 −ω 0  0 2ω   4  0  0 4 0 2 0   0 −ω  0 ω   −3 1 0 0 ω  2 0 ω 1 0   2 0 −ω  0 0 . discussed earlier in Chapters 1 and 2.7 Let us consider the linearized satellite problem. x3 . we note that A has eigenvalues 0.3. This gives     2 4 0 0 1 0 1 0 ω 2   0 0 0 ω  0   . 0. P −1 =  0 − ω 1  P = 2  −3  0 1 0 2  0 0 −ω  3 1 − 2 ω 0 −2ω 0 0 0 0 ω and hence the nilpotent matrix N = A − S. 1. ω.4. 0.4. We follow the procedure described in Theorem 4. 0. x4 ). d¯ x = A¯(t) + B u(t) x ¯ dt where x = (x1 . 0. − 2 ω). 2ω). 0. COMPUTATION OF TRANSITION MATRIX −k = e 2 (t − t0 ) −k (t − t0 ) x(t0 ) + (t − t0 )e 2 k x(t0 ) + x(t0 ) ˙ 2 −k + β(t − t0 )e 2 (t − t0 ) 129 −k = αe 2 where α = x(t0 ). The generalized eigenvectors corresponding to the eigenvalue λ = 0 3 are (1. ¯ ¯  0 1 0  3ω 2 0 0 A=  0 0 0 0 −2ω 0   0 0  1 2ω  .4. 0). u2 ). 0) and the eigenvectors corresponding to the complex eigenvalues λ = ±iω are (1. 0. (0.4. u = (u1 .

LINEAR SYSTEM AND TRANSITION MATRIX 0  0 N =A−S =  −6ω 0  0 0 0 0  0 0 0 0   0 −3  0 0 N is nilpotent matrix of order 2. t0 ) =                    =               P 1 0 0 0               0 1 0 0     0 0 cos ω(t − t0 ) sin ω(t − t0 )   0 0 − sin ω(t − t0 ) cos ω(t − t0 )     −1 P [I + N (t − t0 )]               − cos ω(t − t0 ))     2 sin ω(t − t0 )     4  sin ω(t − t0 )  ω   −3 + 4 cos ω(t − t0 ) 2 ω (1 4 − 3 cos ω(t − t0 ) 3ω sin ω(t − t0 ) 6 sin ω(t − t0 ) 6ω(−1 + cos ω(t − t0 ))  1 0 0 1 0 sin ω(t − t0 ) ω cos ω(t − t0 ) 0 0 −2 (1 − cos ω(t − t0 )) 1 ω −2 sin ω(t − t0 ) 0 0  0    0  ×   −6ω(t − t0 )   0       0 1 −3(t − t0 )    0 0 1 . the transition matrix φ(t.130 That is CHAPTER 4. t0 ) is given by φ(t. So.

1) where A is an autonomous matrix. (4.5 Euler’s Linear System 1 d¯ x = A¯ x dt t The equation of the form (4.4. we get the Euler’s system 1 d¯ x = dt t = 2 3 = − x1 + x2 t t 0 1 −2 3 x ¯ (4.5.5.4) . (4.3) Using the notation x1 = x and x2 = t dx1 dt dx2 dt So. we get dt 1 x2 t (4. so that Eq.5.2) and hence that of Eq.1 Solve t2 d2 x dx − 2t + 2x = 0 dt2 dt dx .5. We make a change variable of the form t = es in the independent variable.5.5.1) Example 4.1) reduces to a linear sytem of the form d¯ x = A¯ x dt (4.2) One can now make use of the linear system theory to compute the solution of Eq. EULER’S LINEAR SYSTEM  4 − 3 cos ω(t − t0 ) 3ω sin ω(t − t0 ) [6(−ω)(t − t0 ) + 6 sin ω(t − t0 ))] [−6ω + 6ω cos ω(t − t0 )] sin ω(t − t0 ) ω cos ω(t − t0 ) − 2 ω − 0 0 1  2(1 − cos ω(t − t0 ))  ω    2 sin ω(t − t0 )      − [3(t − t0 )    −   4 sin ω(t − t0 )]   ω     − [3   − 4 cos ω(t − t0 )] 131            =           2 cos ω(t − t0 ) ω −2 sin ω(t − t0 ) 0 4. is called Euler’s system. (4.5.5.5.

5.5) 1 2 .5. we get that Eq. N = A − S =  0 1 −1 1   (4.5.5. x3 = t2 2 . (4.5. (4.7) . 2es − 2e2s −es + 2e2s Hence the solution of Eq. x2 = t .2 Solve   t3 dx d3 x +t −x =0 dt3 dt (4. 1 −1 2 We use Theorem 4.6) is dt dt equivalent to the Euler’s system  0 1 0 d¯ x 1 ¯ =  0 1 1 x dt t 1 −1 2   0 1 0 A =  0 1 1  has 1 as eigenvalues of multiplicity 3.5. A =  0 1 1  x dt 1 −1 2  −1 1 0 0 1  S = I. LINEAR SYSTEM AND TRANSITION MATRIX Making the change of variable t = es we get d¯ x = A¯(s) x ds 0 1 1 has eigenvalues 1.5.3 to compute a fundamental matrix for the system   0 1 0 d¯ x = A¯(s). (4.2 with eigenvectors −2 3 1 Hence a fundamental matrix of the system given by Eq.5.4) is given by  c0 (2t − t2 ) + c1 (2t − t2 ) c0 (−t + t ) + c1 (−t + t ) 2 2 x(t) =  ¯ Example 4.8) (4.5) is where A = φ(s) = = 2 −1 −1 1 2es − e2s −es + e2s es 0 0 e2s 1 1 1 2 . (4.4.6) dx d2 x Using the notation x1 = x.132 CHAPTER 4.

     sin ωt .5.6 Exercises 0 1 −1 −2δ  .5.4. If A = δ −δt cos ωt + sin ωt  e ω  eAt =   1 − e−δt sin ωt ω √ where ω = 1 − δ 2 .6.7) is given by    c1 1 − lnt + (lnt)2 lnt − (lnt)2 (lnt)2       2 2 2   (lnt) 1 − (lnt) lnt + (lnt)   c2  x(t) = t  ¯     c3 lnt −lnt 1 + lnt This gives the solution of Eq. δ ≤ 1. (4. EXERCISES N is nilpotent of order 3 and hence φ(s) = exp(s) I + N s + N 2 s2   −1 1 1 0 0 0 = exp(s)  0 1 0  + s  0 1 −1 0 0 1   1 −1 1 + s2  1 −1 1  0 0 0  1 − s + s 2 s − s2 s2   s2 1 − s2 s + s2 = exp(s)    s −s 1+s   0 1  1 133  Hence the solution of the Euler’s system Eq. (4.6) as x(t) = c1 t 1 − lnt + (lnt)2 + c2 lnt − (lnt)2 + c3 t(lnt)2 = c1 t + tlnt(c2 − c1 ) + t(lnt)2 (c1 − c2 + c3 ) or x(t) = d1 t + d2 t lnt + d3 t (lnt)2      4. show that 1 −δt e sin ωt ω e−δt cos ωt − δ ω 1.

. Verify that φ(t) =   e−2t cos t − sin t e −2t sin t cos t   is a fundamental matrix of the system   −2 cos2 t −1 − sin 2t d¯  x  x(t) = ¯ dt 1 − sin 2t −2 sin t . Find two linearly independent solutions of the system d¯ x = A(t)¯(t) x dt where  0 1   . ··· .        0 ··· 1      0 ··· λ 2. . .   . . compute the Wronskian of the linearly independent solutions. 0 1 0 ··· 0 0 1 ··· . 1 0                0 0 0 ··· 0 0 0 ··· (c) eAt = eλt I + tB + 1 t2 2 B + · · · + tn−1 B n−1 2! n! 3. ··· 0 0 . A= . . (b) (λI)B = B(λ)I.134 CHAPTER 4.  A(t) =  4. t=0 1  1 − 2 − t t Also. . B =  . . . . Let A and B be two  λ 1    0 λ     .    0 0   0 0 Show that (a) B n = 0. . LINEAR SYSTEM AND TRANSITION MATRIX n × n matrices given by   0 ··· 0       1 ··· 0       . 5. Compute the transition matrix of the system   0 et d¯  x  x(t) ¯ = dt 0 0 by using Peano-Baker series. . . . . . .  .

s)] = φ(t. Solve the non-homogeneous system d¯ x b(t). EXERCISES 6. Solve the initial value problem d¯ x dt = A¯(t) x x(0) = (1. Show that the transition matrices corresponding to autonomous systems described by two similar matrices are necessarily similar. s)] −1   −1 0 0     0 0 −2    0 1 2  = φ(s. e−2t ). Compute the transition matrix of the following system   2 1 0    x  d¯  =  1 3 1  x(t) ¯ dt   0 1 2 8. t0 )φ(t0 . t) = I . t) (b) [φ(t. t0 ) of the system described by Problem 7. t) ¯ where    1  (a) A =    0   2  −1    0  (b) A =    0   0  0 −1 0 0 0 0 0   0 0     0 −1    1 0 0  −1 0 9. s) (c) φ(t.4. 10. (a) [φ(t.6. = A(t)¯(t) + ¯ x dt x(0) = 0 ¯ 135 where a(t) is an given in Problem 5 and ¯ = (1. Verify the following properties of the transition matrix φ(t. b(t) 7.

1970 [3 ] Earl A Coddington and Norman Levinson. Theory of Ordinary Differential Equations.136 CHAPTER 4. 1955 [4 ] Lawrence Perko. 1991 [2 ] Roger W Brockett. Essentials of Ordinary Differential Equatoins. Springer Verlag. McGraw Hill. Finite Dimensional Linear System. 1991 . LINEAR SYSTEM AND TRANSITION MATRIX 11. Inc. (a) Find two distinct fundamental matrices of the system d¯ x = dt 1 1 0 2 x(t) ¯ (b) Find a fundamental matrix of the adjoint system of the above.McGraw Hill Book Company. References [1 ] Ravi P Agarwal and Ramesh C Gupta. John Wiley and Sons. Differential Equations and Dynamical Systems.

Chapter 5 Stability Analysis In this chapter we utilize the theory of transition matrix discussed in Chapter 4. 1 ≤ j ≤ k.1 .1. Let ¯ ¯ ¯ ¯ P = [¯1 .1. vk . vn . . v ¯ ¯ ¯ ¯ ¯ Then P is invertible and   P −1 AP =    B1 B2 . k + 1 ≤ j ≤ n. vk+1 .... This gives rise to the notion of saddle point. are discussed in detail. uk+1 . and complex eigenvalues λj = aj ± ibj . vk .4. Br 137      (5. . We also sketch phase portraits.4. stable and unstable node. 5.1) for large values of t.. vn .4. known as Jordan canonical form of a matrix... . (4.1. un ]. Theorem 5... Various notions of stability including the one due to Lyapunov are introduced and examined leading to some interesting theorems. uk+1 ..3 can be completely described in terms of the following theorem.1 Asymptotic Behaviour of Linear System Theorem 4..1 ( Jordan canonical form) Let A be a real matrix with real eigenvalues λj . focus and centre. vk+1 . .. around the equilibrium points in two dimensions. .1) . Phase portraits. to study boundedness property of solutions of both linear and nonlinear systems along with the stability of equilibrium points. vector fields and solution curves of some illustrative examples. un } of 2n−k where v ¯ ¯ ¯ ¯ ¯ vj are generalized eigenvectors of A corresponding to real eigenvalues λ j (1 ≤ ¯ j ≤ k) and wj are generalized eigenvectors of A corresponding to complex ¯ eigenvalues λj (k + 1 ≤ j ≤ n) with uj = Re(wj ) and vj = Im(wj ). Then there exists a basis {¯1 . which can be used in investigating the behaviour of solution of a linear system of the form given by Eq...

.. . then B = λI + N and exp B(t − t0 ) = exp λ(t − t0 ) exp N (t − t0 ) = exp λ(t − t0 ) ×  2 1 t − t0 (t−t0 ) 2!     0 1 t − t0     ..  . I2 = 1 0 0 1 The elementary Jordan blocks B = Bj .138 CHAPTER 5... . STABILITY ANALYSIS = 1. 0 λ (∗) (∗∗) for λ = a ± ib as a pair of complex eigenvalues of A.1. r are either of the form      for λ as one of the real eigenvalues of A or of the form   D I2 0   D I2   B=  .2) is of the ¯ form x(t) = φ(t.2) Theorem 5.    0 0   0 0 ··· ··· ··· (t−t0 )m−1 (m−1)! m−2 ··· ···    (t−t0 )  (m−2)!      ···    (t − t0 )    1  .. 2.  . .  .. (a) If Bj = B is an m × m matrix of the form (*).2 The solution x(t) of the system given by Eq. t0 )¯0 x where φ(t. This gives us the solvability of the linear system d¯ x = A¯(t) x dt x(t0 ) = x0 ¯ ¯ (5. I2  0 D with D= a b −b a . (5. j  0 λ 1  λ 1  B= .  . t0 ) = φ(t−t0 ) = P diag [exp Bj (t − t0 )] P −1 is the transition matrix of the system. .1.1. 1 .

2) is bounded iff the real parts of the multiple eigenvalues of the matrix A are negative and real parts of the simple eigenvalues of A are non positive. Corollary 5. (5. then An + pn−1 An−1 + · · · + p0 I = 0.5. Theorem 5..2 Every solution of the linear system given by Eq.    0 0 0 0 2 m−1 R (t−t0 ) 2! ··· ··· R (t−t0 ) (m−1)! R(t − t0 ) .1. (5. if det(λI − A) = λn + pn−1 λn−1 + · · · + p0 . Let a = max aj and p = max pj 1≤j≤k 1 ≤j ≤l where R =   cos b(t − t0 ) sin b(t − t0 ) − sin b(t − t0 ) cos b(t − t0 ) ··· ···   R 0     R (t−t0 ) (m−2)!         R(t − t0 )  R m−2  Then by Corollary (5. Corollary 5. Theorem 5.1.  . (5.3 Let p(λ) = det(λI − A).1 Each component of the solution of the linear system Eq.1.1. Let multiplicity of each eigenvalues λj be pj where λj = aj + ibj .1.1.1 gives us the following well-known theorem called Cayley-Hamilton theorem. it follows that ∃ t1 ≥ 0 such that φ(t) = exp (At) ≤ c exp (at) tp where c is some suitable constant.1.  . Corollary 5. That is. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM 139 (b) If Bj = B is 2m × 2m matrix of the form (**) and λ = a ± ib is a complex pair of eigenvalues.1. where λ = a ± ib is a pair of eigenvalues of the matrix A and l is smaller than the dimension of the space. .2) is a linear combination of functions of the form (t − t0 )l exp a(t − t0 ) cos b(t − t0 ) or (t − t0 )l exp a(t − t0 ) sin b(t − t0 ). A ∈ n .1).1. then exp B(t − t0 ) = exp a(t − t0 ) ×  R R(t − t0 )     0 R     .1.3 Every solution of Eq. Then p(A) = 0. ∀ t ≥ t1 .2) tends to zero as t → ∞ iff the real parts of the eigenvalues of the matrix A are negative.

(5. t0 )¯0 + ¯ x t0 φ(t. t∈I (5.1. the solvability of Eq.1.1 and Theorem 5.1. t ∈ I (5. (5. Let φ(t. (5. tf ] and let t u(t) ≤ p(t) + Then we have t t0 q(s)u(s)ds. we have u(t) ≤ p(t) + r(t) and hence r(t) = q(t)u(t) ≤ p(t)q(t) + r(t)q(t) ˙ This gives r(t) − r(t)q(t) ≤ p(t)q(t) ˙ and hence multiplying both side by exp − d r(t) exp − dt which implies that t t τ t t t0 q(s)ds .2).1. s)B(s)¯(s)ds x (5.1.3.1.1.1. ˙ By assumption. refer Gantmacher[5] We shall now investigate the following perturbed linear system d¯ x = A¯ + B(t)¯ x x dt (5.5) t u(t) ≤ p(t) + Proof : Let p(τ )q(τ ) exp t0 t τ q(s)ds dτ. p(t).3) where B(t) is a continuous matrix on the interval [ t0 . (5.3) is given by t x(t) = φ(t.1. STABILITY ANALYSIS For the proof of the Theorem 5.4) We first prove the following variation of Gronwall’s lemma.2) and Eq. ∞) . t0 ) = exp (A(t − t0 )) be the transition matrix of the the linear system given by Eq.1 (Generalized Gronwall’s inequality) Let u(t). Lemma 5. we get t q(s)ds t0 ≤ exp − q(s)ds p(t)q(t) t0 exp − t0 q(s)ds r(t) ≤ t0 p(τ )q(τ ) exp − q(s)ds dτ t0 . We wish to examine the relationship between the boundedness of solutions of Eq.140 CHAPTER 5. q(t) be nonnegative continuous functions in the interval I = [t0 .3).1. By variation of parameter formula.1.6) r(t) = t0 q(s)u(s)ds Then r(t0 ) = 0 and r(t) = q(t)u(t).

1. Then all solutions of Eq. (5.1.8) Proof : that As all solutions of Eq.2) are bounded.1.1. ∞). (5.4 Let all solutions of Eq. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM This gives t t0 t 141 r(t) ≤ = p(τ )q(τ ) exp t0 t τ t q(s)ds + t0 q(s)ds dτ q(τ )p(τ ) exp t0 τ t q(s)ds dτ t and hence u(t) ≤ p(t) + q(τ )p(τ ) exp t0 τ q(s)ds dτ Corollary 5.4).4 If p(t) is nondecreasing on I.1. then t u(t) ≤ p(t) exp q(s)ds t0 (5.3) are also bounded provided ∞ t0 B(t) ≤ ∞ (5. (5.1. (5.7) Proof : As p(t) is nondecreasing. it follows that ∃ c such exp (At) ≤ c ∀ t ∈ I = [t0 .1. t From Eq.1. (5.2) be bounded on [t0 .5.1. (5.6) gives t t u(t) ≤ p(t) 1 + = p(t) 1 − = p(t) exp q(τ ) exp t0 t t0 q(s)ds dτ τ t d exp dτ t q(s)ds dτ τ q(s)ds t0 We are now ready to state and prove the boundedness of solutions of the perturbed system given by Eq.9) ≤ c x0 + c ¯ B(s) x(s) ds ¯ .1.1. we have x(t) ¯ ≤ exp (A(t − t0 )) t t0 x0 + ¯ t0 B(s) exp (A(t − s)) x(s) ds ¯ (5.3). Theorem 5.1. ∞).

2) be such that they tend to zero as t → ∞.3) vis-a-vis asymptotic convergence of solutions of Eq.3) are also asymptotically convergent to zero.1. STABILITY ANALYSIS Applying the generalized Gronwall’s inequality given by Eq. One can show that the solutions of Eq. The proof is similar to that of Theorem 5.4) we get.10) remains bounded.1.2.1. (5. provided B(t) → 0 as t → ∞ (5. (5.1.7) to the above inequality we have t x(t) ¯ ≤ c x0 ¯ exp c t0 B(s) ds ∀t∈I <∞ ≤ c x0 exp (M ) ¯ where M = exp c t0 ∞ B(s) ds This implies that Eq.11) Proof : By Corollary 5.10) ¯ b(s) ds < ∞. (5. (5.1.9) in the proof of Theorem 5.142 CHAPTER 5. (5. The inequality given by Eq.2).3.1.1. (5.1. t x(t) ≤ c exp (−δ(t − t0 )) x0 + c ¯ ¯ t0 exp (−δ(t − s)) B(s) x(s) ds ¯ . it follows that all eigenvalues of A have negative real parts and hence there exist constants c and δ > 0 such that exp (At) ≤ c exp (−δt) ∀ t ≥ t0 Using this inequality in Eq.1.1.1. Remark 5.3 will be replaced by the following inequality t x(t) ≤ c x0 + c ¯ ¯ ¯ b(s) ds t0 The following theorem gives asymptotic convergence of solutions of Eq.1.1.3) has bounded solutions.1. Then all solutions of Eq.1. (5.1. (5.2) has bounded solutions and (b) ∞ t0 (5. (5. (5.5 Let all solutions of Eq.1.1 We examine the boundedness of solutions of the non-homogeneous system d¯ x = A¯(t) + ¯ x b(t) dt under the assumption that (a) Eq. Theorem 5.

1. ¯ Let u(t) = exp (δt) ||¯(t)|| and C denote the first two terms of RHS of the above x inequality: t1 C = c exp (δt0 ) ||¯0 || + c x Then we get. ¯ This proves the theorem. then the solutions of non-homogeneous system given by Eq. t0 exp (δs) ||B(s)|| ||¯(s)||ds x t u(t) ≤ C + c c1 By Gronwall’s lemma we get t1 u(s)ds .2) has solutions asymptoticaly convergent to zero.1.2 If Eq.11). (5. . Remark 5. (5.1. we choose c1 such that cc1 < δ and hence exp (t(c c1 − δ)) → 0 as t → ∞. This gives that x(t) → 0 as t → ∞.5.1. ∃ t1 ≥ t0 > 0 such that B(t) ≤ c1 for t ≥ t1 and hence we get t1 143 x(t) exp (δt) ¯ ≤ c exp (δt0 ) x0 + c ¯ t exp (δs) B(s) t0 x(s) ds ¯ + t1 B(s) exp (δs) x(s) ds ¯ t1 ≤ c exp (δt0 ) x0 + c ¯ t exp (δs) B(s) t0 x(s) ds ¯ + c t1 c1 exp (δs) x(s) ds.10) satisfies the same property if ¯ b(t) → 0 as t → ∞.1. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM In view of Eq. t ∈ I t u(t) ≤ C exp ( c c1 ds) t1 = C exp (c c1 (t − t1 )) ∀ t ∈ I This gives ||¯(t)|| x ≤ exp (−δt)C exp (c c1 (t − t1 )) = C exp(t(c c1 − δ) − c c1 t1 ). ∀t ∈ I As c1 is arbitrary. given any c1 > 0. (5.

1. Proof : We have d ||¯(t)||2 x dt This implies that ||¯(t)||2 ≤ ||¯(0)||2 exp x x If ∞ t0 t t0 lim t→∞ ≤ (A(t) + [A(t)] x(t). Let M (t) be the largest eigenvalue of A(t) + [A(t)] . then Eq. x(t)) + (¯(t).12) has bounded (b) If t0 M (s)ds = −∞.12) = (A(t)¯(t).1.144 CHAPTER 5. STABILITY ANALYSIS We now consider the boundedness and stability of solution of non-autonomous linear system d¯ x = A(t)¯(t) x dt Observe that.1. (5. (5. x(t) + x(t). t ∈ I t t0 (b). x(t)) x x ¯ Differentiating the above relation.12) has solutions asymptotically converging to zero. We have the following theorem Theorem 5.1. If M (s)ds = −∞ then we get Example 5. we get d ||¯(t)||2 x dt = d¯(t) x d¯(t) x . M (s)ds < ∞. x(t) ¯ ¯ ≤ M (t) ||¯(t)||2 x M (s)ds . then we get (a). A(t)¯(t)) x ¯ x x = (A(t) + [A(t)] )¯(t). ∞ (a) If ∞ t0 M (s)ds is bounded then Eq. x(t) x ¯ A(t) + [A(t)] is a symmetric matrix and hence its eigenvalues are real.1 d¯ x = [A + B(t)] x(t) ¯ dt A= 0 1 −1 0 0 and B(t) = 0 0 2a at + b . ¯ ¯ dt dt (5.6 solutions.1. ||¯(t)||2 = (¯(t).

(5. − sin t cos t The perturbed system has the fundamental matrix as a sin t − (at + b) cos t a cos t + (at + b) sin t (at + b) sin t (at + b) cos t which is unbounded. if all solution of Eq. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM 145 One can compute that the fundamental matrix of the unperturbed system is cos t sin t which is bounded.13) are also bounded. 1  (1 + t)2 A(t) = −t2 2 and this gives (1 + t)2   2 t2   (1 + t)2 and A(t) + A(t) = −1 0 ∞ 0  0  −2 M (s)ds = 2.1.5. (5.5 the non-autonomous system d¯ x = A(t)¯(t) x dt has bounded solutions.1. We note that Example 5.1. Hence by Theorem 5.1.1.2  M (t) = t 0 ||B(s)||ds → ∞ as t → ∞.1. then all solutions of Eq.1.12).12) is such that (a) t lim inf t→∞ t0 T rA(s)ds > −∞ or T rA(s) = 0 (5.1. (5.12) are bounded.1. We shall now consider boundedness and asymptotic stability of the perturbed system of the form d¯ x = [A(t) + B(t)] x(t) ¯ dt (5.7 Under Assumptions 1. . (5.14) (b) The perturbation matrix B(t) satisfies the condition ∞ t0 ||B(t)||dt < ∞ Theorem 5.1. Assumptions 1 The matrix A(t) corresponding to the system given by Eq.13) corresponding to the non-autonomous system given by Eq.

Estimating Eq. to get the following integral represntation of the solution x(t).1. We now apply the variation of parameter formula to the IVP corresponding to Eq. we get t ||¯(t)|| ≤ ||ψ(t)|| c1 + c x t0 ||B(s)||||¯(s)||ds x .15). (5.146 CHAPTER 5. c0 = c||x0 ||.1. by Abel’s lemma (Theorem 4. (5.1. the boundedness of solutions implies that ψ(t) is bounded.1.1.16). ¯ x(t) = ψ(t)ψ −1 (t0 )x0 + ¯ ¯ Define c = max supt≥t0 ||ψ(t)||.12). (5. (5. supt≥t0 ||ψ −1 (t)|| .1.12) are asymptotically convergent to zero. so is the case for all soluiton for Eq. Let c be defined as before.1. we get that ||ψ −1 (t)|| → 0 as t −→ ∞.13).16) gives t t t0 ψ(t)ψ −1 (s)B(s)x(s)ds (5.15) In view of Assumptions 1(a). (5.12) will imply that ||ψ(t)|| −→ 0 as t −→ ∞. (5. ¯ Then Eq. if all solution of system Eq. STABILITY ANALYSIS Proof : Let ψ(t) be a fundamental matrix of the system Eq. Using this fact in Eq.1.1.4). it follows that ||ψ −1 (t)|| is bounded.1.16) ||¯(t)|| ≤ c0 + c2 x and hence Gronwall’s inequality gives x ||¯(t)|| ≤ c0 exp c2 This proves the theorem. t0 ||B(s)||||¯(s)||ds x t t0 ||B(s)||ds <∞ Theorem 5.13). Also. (5. we have t det[ψ(t)] = det[ψ(t0 )]exp( t0 T rA(s)ds) and hence ψ −1 (t) = adjψ(t) adjψ(t) = t detψ(t) det[ψ(t0 )]exp( t0 T rA(s)ds) (5. Proof : Asymptotic convergence of solution of Eq.1. Then. (5.1.8 Under Assumptions 1 .

Then (a) boundedness of solutions of Eq. The boundedness and asymptotic convergence of the perturbed system Eq. (5. Let Assumption 1(b) hold. Then (a) boundedness of solutions of Eq.1.1.9 Let Assumptions 2 hold.1.13) is preserved under Assumption 1(b). (5.13) and . (5. Theorem 5.1. (5.1. we can dispatch with the Assumption 1(a).1.1. (5.17) ||¯ b(s)||ds < ∞ Theorem 5.12) implies boundedness of solutions of Eq. (5.17).1.5.1.17) and (b) asymptotic convergence of solutions of Eq. If the non-autonomous system under consideration is periodic.1.1.10 Let A(t) be periodic. The above inequality gives ¯ x ||¯(t)|| ≤ c1 + c ||ψ(t)|| Put ||y(t)|| = t t0 147 ||B(s)||||ψ(s)|| ||x(s)|| ||ψ(s)|| ds ||¯(t)|| x and apply Gronwall’s inequality to get ||ψ(t)|| t ||y(t)|| ≤ c1 exp thereby implying that t t0 c2 ||B(s)||ds x ||¯(t)|| ≤ c1 ||ψ(t)|| exp t0 c2 ||B(s)||ds −→ 0 as t −→ ∞ These theorems immediately lead to the following result corresponding to the asymptotic stability of the non-homogeneous system d¯ x = A(t)¯(t) + ¯ x b(t) dt Assumptions 2 (a) Same as Assumption 1(a) (b) The function ¯ is such that b(t) t t0 (5.12) will imply boundedness of solutions of Eq. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM where c1 = ||ψ −1 (t0 )x0 ||.12) to zero will imply asymptotic convergence of solutions to zero of the non-homogeneous equation Eq. (5.

1.19) gives t ||¯(t)|| ≤ d1 d2 + d2 d2 x and hence t0 ||B(s)||||¯(s)||ds x t t0 ||¯(t)|| ≤ d1 d2 exp d2 d2 x ||B(s)||ds This proves the boundedness of solutions of Eq.18) gives x(t) ≤ d1 eRt + d2 ¯ t t0 eR(t−s) B(s) x(s) ds (5. sup ||P −1 (t)|| t≥t0 t≥t0 Eq. (5. (5. (5. This inequality immediately gives ||¯(t)|| ≤ d1 d3 e−αt + d2 d3 x t t0 e−α(t−s) ||B(s)||||¯(s)||ds x . t0 ) = P −1 (t)eR(t−t0 ) P (t0 ) where P is periodic nonsingular matrix and R is a fixed martix. (5.1.12) is given by φ(t. Boundedness of solutions of Eq. Proof : By Floquet-Lyapunov theorem .19) ¯ where d1 = d||e−Rt0 P (t0 )x0 ||. ∞).1. (5.1. (5. STABILITY ANALYSIS (b) asymptotic convergence of solutions of Eq.1.1.1.1.3.12) implies that ||eRt || ≤ d2 and hence Eq.18) P −1 (t) P (t) is periodic and nonsingular and hence det P (t) = 0 and hence both ||P (t)|| and ||P −1 (t)|| are bounded on [t0 .1.13).12) are asymptotically convergent to zero. we must have ||eRt || ≤ d3 e−αt for some positive constant d3 and α. (5. (5.13) we get x(t) = P −1 (t)eR(t−t0 ) P (t0 )x0 ¯ ¯ t + t0 P −1 (t)eRt e−Rs P (s)B(s)¯(s)ds x This gives x(t) ¯ ≤ + t0 P −1 (t) t eRt e−Rt0 P (t0 )¯0 x eR(t−s) P (s) B(s) x(s) ds ¯ (5.13).148 CHAPTER 5.1. Let d = max sup ||P (t)||.12) implies asymptotic convergence of solution of Eq.Theorem 4.1. If solutions of Eq. By variation of parameter formula applied to the IVP corresponding to Eq. the transition matrix of the periodic system Eq.1. (5.

This solution is said to be stable if for a given > 0. t0 . STABILITY ANALYSIS . x0 + ∆¯0 ) − x(t. x0 )|| < x x ¯ x ¯ ¯ Definition 5. x0 )|| → 0 as t → ∞ x x ¯ x ¯ ¯ (5.2. (5. t0 . t0 .2) can be restated as.2. t0 ||B(s)||ds −→ 0 as t −→ ∞ 5. ∞).2. t0 . (5.2) (5.2. x0 + ∆¯0 ) − x(t. indicating its dependence on t and also the initial point t 0 and initial value x0 .3 x(t. . t0 . (5.3) ∀ t ≥ t 1 ≥ t0 > 0. x). x0 ) is said to be asymptotically stable if it is stable ¯ ¯ and ∃ δ0 > 0 such that ||∆¯0 || < δ0 ⇒ ||¯(t. t0 .Formal Approach We shall formally define various concepts of stability of solution of the initial value problem d¯ x = f (t. t0 . given µ > 0.2.2. x0 ) denote the solution of the initial value problem ¯ ¯ Eq. ∃ δ > 0 such ¯ that ||∆¯0 || < δ ⇒ ||¯(t.1) has a global solution on [t0 .5. there exist δ and T (µ) such that ||∆¯0 || < δ ⇒ ||¯(t.FORMAL APPROACH and hence t 149 ||¯(t)|| ≤ d1 d3 e−αt exp d2 d3 x This proves the theorem. x0 + ∆¯0 ) − x(t.2 Stability Analysis .1).1 For the linear system d¯ x = A(t)x(t) dt stability is equivalent to the boundedness of the solution. x0 )|| < µ ∀ t ≥ t0 + T (µ) x x ¯ x ¯ ¯ Definition 5. x0 ) is said to be uniformly stable if for any ¯ ¯ ∃ δ = δ( ) > 0 such that for any solution x(1) (t.2 x(t. x(t0 ) = x0 ¯ ¯ ¯ dt We assume that Eq. x1 ) of ¯ ¯ d¯ x = f (t. x(t0 ) = x1 ¯ ¯ ¯ dt ||¯(t1 ) − x(1) (t1 )|| < δ ⇒ ||¯(t) − x(1) (t)|| < x ¯ x ¯ Theorem 5. (5.2.2.2.1) To be more explicit.2. t0 . t0 . t0 . Eq. x).1 Let x(t. Definition 5.2.

2 Let φ(t) be a fundamental matrix of Eq. t0 )|| = max ||φ(j) || = max ||φ(t. let the solution of the linear system be stable. Then. x0 )|| = ||φ(t. t0 .3). x0 )|| = ||φ(t. ¯ ¯ Conversely. Then. Then Eq.1 Let φ(t) be a fundamental matrix of the system given by Eq.3) are uniformly stable iff there exists a constant c > 0 such that ||φ(t)φ−1 (t1 )|| ≤ c ∀ t 0 ≤ t1 ≤ t < ∞ (5. we have ||¯(1) (t) − x(t)|| x ¯ Let > 0 be given.4) Proof : Let x(t) = x(t.2. x0 + ∆¯0 ) − x(t.2. t0 .3) is bounded. t0 .2. STABILITY ANALYSIS Proof : Let φ(t. t0 )¯0 || ≤ x ¯ x δ That is. Then ¯ 2 ||φ(t. x0 ) is stable. (5.3).3). Corollary 5. t0 )∆¯0 || ≤ ||φ(t. t0 . Then for any ¯ ¯ ¯ t1 ≥ t0 we have x(t) = φ(t)φ−1 (t1 )¯(t1 ). t0 . (5. (5. t0 )∆¯0 )|| = x δ ||φ(t. ∃ δ > 0 such that ||¯(t. x Choose ∆¯0 = x δ (j) e .2.3) has asymptotically stable solutions iff ||φ(t)|| → 0 as t → ∞. it follows that the solution x(t. For the uniform stability of the linear system we have the following theorem Theorem 5.2. t0 ) e(j) || < ¯ 2 > But φ(t. (5.2. given 0. the solution of the linear system given by Eq. Let x(1) (t) = φ(t)φ−1 (t1 )¯(1) (t1 ) be ¯ x ¯ x any other solution. (5. Hence we ¯ have 2 ||φ(t. (5.2.2. t0 ) is bounded. t0 )∆¯0 || < x ¯ x ¯ ¯ x whenever ||∆¯0 || < δ.2. Then all solutions of Eq. t0 . x0 )|| = ||φ(t. t0 )¯(j) || ≤ e 1≤j≤n 1≤j≤n δ This gives 2 ||¯0 || x ||¯(t. As Eq. t0 ).2. t0 ) be the transition matrix. ||¯(1) (t1 ) − x(t1 )|| ≤ x ¯ c = δ( ) > 0 imply that ||¯(1) (t) − x(t)|| < x ¯ ≤ ||φ(t)φ−1 (t1 )||||¯(1) (t1 ) − x(t1 )|| x ¯ ≤ c||¯(1) (t1 ) − x(t1 )|| ∀ t0 ≤ t1 ≤ t < ∞ x ¯ . x0 + ∆¯0 ) − x(t.4) is satisfied. t0 . t0 )|| ||∆¯0 || x ¯ x ¯ ¯ x x If φ(t. t0 ) e(j) is the j th column vector φ(j) of the matrix φ(t. ||¯(t. (5.150 CHAPTER 5. x0 ) be a solution of Eq.

dx =t dt t2 This has a solution of the form x(t) = x(t0 ) + . ∀ t ≥ t1 . (5.1.2.2 Every solution of the system dx = p(t)x(t) dt is of the form t x(t) = x(t0 ) exp t0 p(s)ds t t0 Hence x(t) ≡ 0 is asymptotically stable iff p(s)ds → −∞ as t → ∞. x(t)) dt (5. which is of the type dx = Ax(t) + g(t. it is difficult to give the stability criterion.2. Hence given > 0. δ (t1 ) = e(j) .3) uniformly stable and hence zero solution is uniformly stable. boundedness and stability are distinct concepts. for nonlinear systems. (5.1). Instead. we choose x ¯ (1) ∀ t ≥ t1 . let Eq.4) and the theorem. STABILITY ANALYSIS . Example 5. However.2.1 Consider.FORMAL APPROACH and hence x(t) is uniformly stable. (5.2. Then ¯ 2 δ < . t ≥ t1 δ This proves Eq. we examine a special form of Eq.2.2. it follows that ||φ(t)φ−1 (t1 )|| = max ||φ(j) (t)|| < 1≤j≤n 2 . Hence. ¯ 151 Conversley. all solutions are of the form x(t) = x(t0 ) not bounded. 2 ||φ(t)φ−1 (t1 )¯(1) (t1 )|| = ||φ(j) || x Here φ(j) (t) is the j th column of φ(t)φ−1 (t1 ).1).2. Whereas for dt which are uniformly stable but not asymptotically stable. For the nonlinear initial value problem given by Eq. Example 5. ∃ δ( ) such that t1 ≥ t0 and ||¯(1) (t)|| < δ ⇒ ||¯(1) (t)|| < x x Thus we have ||φ(t)φ−1 (t1 )¯(1) (t1 )|| < x As in Theorem 5. (5. This is a stable solution but 2 dx = 0.5) .2.5.2.

Then Eq.6) on g(t.5) with 0 initial condition.10) Since x0 and m are at our disposal. Then the trivial solution of Eq.7) Since real(λ) < 0. Then in x view of continuity of the function x(t).9) We assume that the initial condition x(t0 ) = x0 is such that ||¯0 || < d. implies that ||x(t)|| < d and Eq. (5.2.9) gives that ||g(t. there exists t1 > t0 such that ||x(t)|| < d for all t ∈ [t0 . (5. it follows that there exists c. x) satisfies Eq.7) gives t x(t) ≤ c exp (−δ(t − t0 )) x0 + c t0 exp (−δ(t − s)) g(s. ( λ is any eigenvalue of A).10) implies that ||x(t)|| ≤ d c . (5. Eq.2.2. t ∈ I ||x(t)|| exp (δt) ≤ c ||x0 || exp (δt0 ) exp (c m(t − t0 )). (5. Hence t ∈ I.6).2.5) with x(t0 ) = x0 is equivalent to the solvability of the following equation t x(t) = exp (A(t − t0 )) x0 + t0 exp (A(t − s)) g(s. x(s)) ds(5. (5. t1 ] = I. ∃ d > 0 such that ||g(t. (5. x(t))|| ≤ m||x(t))|| for all t ∈ I.8). we choose m and x0 in such a way that ¯ c m < δ and x0 ≤ min d. .6). we get t x(t) exp (δt) ≤ c exp (δt0 ) x0 + c m By Gronwall’s lemma. δ > 0 such that || exp (At)|| ≤ c exp (−δt) ∀ t ≥ 0.2.2. we get that t0 exp (δs) ||x(s)||ds.2. (5. t ∈ I That is ||x(t)|| ≤ c ||x0 || exp ((c m − δ)(t − t0 )). it follows that g(t. x) → 0 as x → 0 x (5. Proof : The solvability of Eq. (5. 0) = 0 and hence 0 is a trivial solution of Eq.3 Suppose the real part of eigenvalues of the matrix A are negative and the function g(t.5) is asympotically stable. We have the following stability theorem for such a kind of system. x) implies that for given m > 0.2.2. t ∈ I (5.2.2. In view of this fact. Theorem 5.6) In view of Eq.2.152 g(t. (5.2. x(s))ds (5. (5. x) satisfies the following condition CHAPTER 5. v)|| ≤ m||v|| whenever ||v|| ≤ d (5.2.8) Also the condition given by Eq.2. STABILITY ANALYSIS g(t.2. Using this inequality in Eq.

A = 0 g − l 1 k − m g ˙ and F (x) = F (θ. − f (θ) l 1 2 k k2 g A has eigenvalues − ± − 2 2m 4m l if k. x0 ) < ∀ t ∈ [t0 .2. Further Eq.11) is asymptotically stable.10) implies that given > 0. STABILITY ANALYSIS .11) is equivalent to dx = Ax + F (x) dt where ˙ x = (θ. g and l are positive and Fx = g l both of which have negative real parts θ3 −.3 is applicable and hence the main pendulum problem given by Eq. This gives us the asymptotic stability of the zero solution of Eq.2. ∃ η = min d. such that c x0 < η ⇒ x(t.11) Eq.4) the solution x(t) can be extended for t ∈ [t0 . θ).2. This gives g ¨ k ˙ g θ + θ + θ + f (θ) = 0 m l l where f (θ) → 0 as θ → 0. θ (5.5). . θ) = 0.2. (5. (5. . we get 3! f (θ) → 0 as θ → 0 θ sin θ = θ + f (θ). Example 5. 3! x →0 ≤ M |θ|3 → 0 as Thus Theorem 5.2.3 The motion of a simple pendulum with damping is given by ¨ k ˙ g θ + θ + sin θ = 0 m l Approximating sin θ by θ − θ3 + .2..5. m. ∞).2. (5. . (5.2.FORMAL APPROACH 153 for t ∈ I and this bound d does not depend on t and hence by continuation theorem (Theorem 3. ∞) and also ||x(t)|| → 0 as t → ∞. t0 .2...

2) through the linear system with d¯ x = A¯ x dt (5. (5.1) n For the phase portrait. we shall concentrate on the autonomous system Definition 5. .3. (5.2) Theorem 5.3.those points where f (x) = 0.3. of the system We shall be interested in the stability of the system around equilibrium points .2).3 Phase Portrait Analysis dx = f (x(t)) dt (5.154 CHAPTER 5. Theorem 5.2.4)) f (¯) = f (x0 ) + f (x0 )(¯ − x0 ) + g(¯) x ¯ x ¯ x where g(¯) x → 0 as x − x0 ¯ ¯ x − x0 → 0.3.3. ¯ x ¯ We shall now study Eq. then ¯ x ¯ we have (by Eq. phase space will be assumed to be two dimensional so that its geometry can be studied clearly.4 give the relationship between the stability of the trivial solution of the nonlinear system of the form Eq. Then the trivial solution of Eq.2. (5.2) with that of the linear system given by Eq.3. (5. We restate this relationship in terms of the following theorem. For details analysis concerning this section refer Agarwal and Gupta[1]. (2. (5. If x0 is an isolated equilibrium point of f (¯) and if f is differentiable at x0 .3.6).3 and 5. (5. we get that Eq.3). 5.3.3.3. (5.3. (5.2.3) (5. Theorem 5. (5. Translating the origin to ¯ x x ¯ x x0 .5).5) is unstable. (5. Also.4 Assume that A possesses at least one eigenvalue with positive real part and g(t. ¯ ¯ As f (x0 ) = 0.1) is equivalent to the system ¯ d¯ x = A¯ + g(¯) x x dt g(¯) x → 0 as x → 0 and A = f (x0 ).3) is asymptotically stable.1 The set of solution curves in the phase space given by Eq.2.3.1) is called the phase portrait of the system. so is the equilibrium point zero of the nonlinear system given by Eq. we have f (¯) = f (x0 )(¯ − x0 ) + g(¯). STABILITY ANALYSIS In a similar way we have the following theorem for the unstability of zero solution of Eq.2.1 (a) If the equilibrium point zero of the linear system given by Eq. x) satisfies Eq.3.2.

Hence the phase portrait in 2 is as under.2).1 We investigate the phase portrait of the linear system given by −1 0 Eq. (5. at each point x ∈ 2 . that is each point c ∈ 2 moves to the point x(t) ∈ 2 after a time t. (5. (5. (5. so is the unstability of the equilibrium point zero of Eq. Recall that this is called dx = f (x) = Ax. (c) If the equilibrium point zero Eq.3. c = x0 = x(t0 ). So the vector field is as given below. k = c1 2 c2 (c = (c1 .3. Since dt x(t) is tangent to the vectors in the vector field f (x) = Ax. . The solution of the linear 0 2 system is given by x(t) = exp (−t) 0 0 exp (2t) c.3) is unstable. Example 5.2) can be stable.3.3.3. the solution curve a vector field.5.3) around the origin with A = . c2 )). (5.3) is stable then the zero solution of Eq. The solution curves of the above system are of the form x2 = k .3.1: Phase portrait of Example 5. asymptotically stable or unstable.1 The function f (x) = Ax defines a mapping on 2 .3.3. x1 2 The solution x(t) defines a motion along these curves. PHASE PORTRAIT ANALYSIS 155 (b) If the equilibrium point zero of the Eq. x2 4 2 -3 -2 -1 -2 1 2 3 x1 -4 Figure 5.3.

2: Vector field of Example 5. . (−1.2 Consider now the coupled system   −1 −3 dx   x = Ax = dt 0 2 The eigenvalues of the matrix A are −1 and 2 and the pair of corresponding eigenvectors is (1. P −1 AP =   −1 0 0 2  =D ¯ So. 1). STABILITY ANALYSIS Figure 5.3.156 CHAPTER 5.3. P =  1 −1 0 1   . we obtain the decoupled system dy = Dy dt The phase portraits in y1 − y2 plane and x1 − x2 plane are as under. under the transformation y = P −1 x.3. 0).1 Example 5. P −1 =   1 1 0 1   .

3. the phase portrait of a linear system given by Eq. PHASE PORTRAIT ANALYSIS y2 4 157 2 -3 -2 -1 -2 1 2 3 y1 -4 Figure 5.3.3.3: Phase portrait in y1 − y2 plane x2 3 2 1 x1 -4 -2 -1 -2 -3 2 4 Figure 5.3.1.3) is best studied by an equivalent linear system dx = Bx dt (5.4) where the matrix B = P −1 AP has one of the following form ( by appealing to Theorem 5.5. (5.4: Phase portrait in x1 − x2 plane In general. B= λ 0 1 λ .3.1) B= λ 0 0 µ . or B= a −b b a .

If λ > 0 > µ.4) is said to have a saddle point at 0. The phase portraits are as under: Case 2: B = . λ ≤ µ < 0 or B = . STABILITY ANALYSIS The solution of the initial value problem given by Eq.4) with x(0) = x 0 is x(t) = or x(t) = exp (at) exp (λt) 0 0 exp (µt) x0 .5: Vector field with arrows reversed λ 0 λ 1 . λ<0<µ Case 1: B = 0 µ The phase portrait is of the form given in Figure 5. Figure 5.3. x(t) = exp (λt) cos bt − sin bt sin bt cos bt 1 t 0 1 x0 x0 Different cases will have different phase portraits. (5. λ<0 0 µ 0 λ In this case the origin is referred to as a stable node ( unstable if λ ≥ µ > 0 ). (5.158 CHAPTER 5. then the arrows in the vector field are reversed as we see in the following Figure. In this case the system given by Eq.3.3. λ 0 .3.3.

the phase portrait is as under. PHASE PORTRAIT ANALYSIS x2 6 4 2 -3 -2 -1 -2 -4 -6 1 2 3 x1 159 Figure 5.5 -3 -2 -1 -0.5 -1 -1. . λ < 0.3.3.5.5 1 0.3.7: Phase portrait with λ < µ < 0 (stable node) When B = λ 0 1 λ .5 1 2 3 x1 Figure 5.6: Phase portrait with λ = µ < 0 (stable node) x2 1.

the origin is a stable focus with trajectories as spiral converging to zero.3.9: Stable focus .160 CHAPTER 5. Case 3: B = x2 10 5 x1 -10 -5 -5 -10 5 Figure 5.3. If a > 0. we have unstable focus. STABILITY ANALYSIS x2 2 1 x1 -4 -2 -1 -2 2 4 Figure 5.8: Origin as stable node a −b b a If a < 0.

5.11: Origin as centre Example 5.3.10: Unstable focus 0 −b b 0 In this case the system given by Eq.3. (5.5 -1 -0.5 -5 -2.5 Figure 5.3.5 -1 Figure 5. The phase portrait is as under.5 1 x1 -0. x dt A= 0 −4 1 0 . Case 4: B = x2 1 0.3.4) is said to have a centre at the origin.5 0.3.5 -5 -10 2.3 d¯ x = A¯. PHASE PORTRAIT ANALYSIS x2 10 5 x1 161 -10 -7.5 5 7.

. then Eq. (a) If δ < 0. P −1 = 0 1 2 CHAPTER 5.3).3. STABILITY ANALYSIS 0 1 . (c) If δ > 0 and τ 2 − 4δ < 0 and τ = 0. We assume that det A = 0. node. It is stable if τ < 0 and unstable if τ > 0. (5. ¯ ¯ It can be seen that x1 2 + 4x2 2 = c1 2 + 4c2 2 . so that ¯ is the only equilibrium point.12: Phase portrait with origin as centre The above discussion can be described in the following theorem. (5. then Eq. (5. 0 Theorem 5.3.3.3) has a centre at the origin.3) has a focus at the origin.3. focus or centre at the origin of the linear system. P = 2 0 0 1 .3) has a saddle point at the origin. then Eq.3.3) has a node at the origin. B = P −1 AP = 0 −2 2 0 x(t) ¯ = P = cos 2t − sin 2t sin 2t cos 2t 1 2 P −1 c ¯ c ¯ cos 2t −2 sin 2t sin 2t cos 2t c = x(0).3.2 Let δ = detA and τ = traceA and consider the linear system given by Eq. it is stable if τ < 0 and unstable if τ > 0. (5. giving the saddle point. (d) If δ > 0 and τ = 0 then Eq. (b) If δ > 0 and τ 2 − 4δ ≥ 0.162 A has eigenvalues λ = ±2i.3. (5. x2 2 1 x1 -4 -2 -1 -2 2 4 Figure 5.

it implies that the equilibrium k a point ( . ¯ dt 2 −w −k δ = detA = w 2 > 0 and τ = traceA = −k < 0.3.3. we have origin as a stable node. For more on phase portrait analysis refer Perko [9].4 Consider the following linearized Predator-prey model (linearized k a around the equilibrium point ( . 5. then the ¯ ¯ energy will continually decrease until it finally assumes its minimum.3. S). )). One of the attractions of the Lyapunov method is its ¯ ¯ appeal to geometric intution as we see in the following two examples.2. we look for a scalar valued Lyapunov function V (¯) x ˙ x of the state which is (a) positive (b) with V (¯) < 0(¯ = xe ) and (c) V (¯) = x ¯ x ˙ x V (¯) = 0 for x = xe . ) is a centre.3. dE(¯) x of the energy of a physical system is negative If the rate of change dt for every possible state x except for a single equilibrium state xe . (1. (ii) If k 2 − 4w2 ≥ 0.4. .2).1 dx d2 x +k + w2 x = 0 dt2 dt λ k is the resistance and w 2 = . and A =  where x = x. ¯ Using the notation of the above theorem (Theorem 5. we get δ = det A = ak > 0 and τ = trace A = 0. From mathematical perspective.5 Consider the problem of mechanical oscillations discussed in Example 1. we have origin as a stable focus. As δ > 0 and τ = 0. given by Eq. ma This is equivalent to the following system d¯ x = A¯ x dt   0 1 dx . λ c Example 5. A =  x  dt   aλ 0 c Example 5. with F representing fish population and S shark population.5. LYAPUNOV STABILITY 163 where x = (F. (i) If τ 2 − 4δ = k 2 − 4w2 < 0.5) λ c   kc 0 −  λ  d¯ x   = A¯(t).4 Lyapunov Stability The principle idea in the Lyapunov method is the following physical reasoning.

case 1 of a harmonic oscillator given by a second order differential equation. d2 x λ + ω 2 x = 0. STABILITY ANALYSIS Example 5. Then x 1 2 ˙ x V (¯(t)) = (x(t). The trajectory of the above system is given by R(t) = R0 e−(t−t0 ) .1. ω 2 = dt2 ma In terms of a first order system we have d¯ x = dt 0 −ω 2 1 0 x.4.2 Consider a slight modification of the earlier example x˙1 x˙2 = x2 − ax1 = −x1 − ax2 with a > 0.3. We define V (¯) = (x2 + x2 ). x2 ) ¯ ¯ The trajectories of the above system in the phase-plane are ellipses about the origin. x2 ) = ω 2 x2 + x2 = V (¯) x x 1 2 ˙ The derivative V and V along the trajectory of the above system is given by dV x(t) ¯ ˙ x V (¯) = dt = (¯(t).164 CHAPTER 5. x = x1 ˙ V (¯)) x ∂V ∂V + x˙2 ∂x1 ∂x2 = 2ω 2 x1 x2 + 2x2 x˙2 = 0 So energy remains constant and hence the system is conservative. The energy of the system is given by E(¯) = E(x1 .4. x = (x1 .1 Let us recall Example 1. R(t) = x2 (t) + x2 (t) 1 2 . Example 5. V (¯)) ˙ x 2 = −2a(x1 + x2 )2 2 = −2aV (¯) x < 0 if x = 0 ¯ = 0 if x = 0 ¯ This implies that V (¯) is constantly decreasing along any solution of the above x system.

t x and t with V (t. Let V (¯) be a x measure of the distance of the state x from the orgin in the state space resulting ¯ in to V (¯) > 0 for x = 0 and V (0) = 0.4. d¯ x = f (t.4.1) dt where f : I × n → n with origin as an equilibrium point (f (t.4. V (¯(t)) < 0. x(t)) ≤ −r(||¯||) < 0 ∀ (t. Then x(t) → 0. Suppose this distance is continuously x ¯ ˙ x decreasing as t → ∞. x) ∈ I × ¯ x ¯ n (5. ¯ We now investigate the role of Lyapunov function to ensure the stability of the equilibrium state of the non-autonomous nonlinear system.3) .1: Trajectory along V (¯) = constant x Hence the trajectory of the above system cross the boundary of every region V (¯) = constant from the outside towards inside.4.4. that is. LYAPUNOV STABILITY x2 165 x1 V(x)=constant Figure 5. there exist a continuous. 0) = 0 and ¯ (i) V (t. x) : I × n → n . ¯ called Lyapunov function. such that it is continuous and has continuous partial derivatives w. (t. x) ∈ I × n ¯ ¯ (5. Assumptions 1: There exists a scalar valued function V (t. that is. This leads to the following x intersting geometric interpretation of the Lyapunov function. non-decreasing ¯ function α such that α(0) = 0 and for all t and all x = 0.2) (ii) there exists a continuous scalar valued function r such that r(0) = 0 and ˙ x the derivative V (¯) of V (¯) along the motion x(t) satisfies the following x ¯ ˙ V (t. x(t)). we have ¯ 0 < α(||¯||) ≤ V (t. x) is positive definite.5. 0) = 0).r. x) x ¯ (5.

¯ ≤ β ∗ (t0 . x(t)) ≤ 0 ∀ (t. x0 )|| < x ¯ whenever ||x0 || < δ for t ≥ t0 . x0 )) ≥ α(||¯(t. The above inequality gives ||¯(t.2). we have α( ) > β(δ) ≥ V (t0 . x0 ) x ¯ ¯ ¯ forall t ≥ t0 and ||x0 || < δ. Let ||x0 || < δ. We get δ( ) (independent of t0 ) such that β(δ) < α( ).6) Since α is non-decreasing and positive. ¯ 0. This proves the stability of the equilibrium state ¯ (b) Let (i).4. ||x0 ||) < α( ) if ||x0 || < δ ¯ ¯ Using (5. x0 ) ≥ V (t. for the second inequality. x0 )||) ¯ ¯ ¯ x ¯ Here.4) and (5. STABILITY ANALYSIS ˙ V (t.4.(ii) and (iii) hold.4. t0 . we get α(||¯||) ≤ V (t.(ii)(a) and (iii) hold (c) asymptotically stable if (i). (ii)(a) and (iii) hold and assume that > 0 is arbitary.4. the above inequality implies that ||¯(t.4. Then ¯ proceeding as in part(a). x(t)) ≤ β(||¯||) ∀ t ∈ I ¯ x (5. This implies the uniform stability of the zero solution the equilibriun point.4. x(t. t0 . ||¯||) be the maximum of V (t. x) be a Lyapunov function of the system (5. Proof : (a) Let β ∗ (t.166 (ii) (a) CHAPTER 5. Let > 0 be x ¯ y x arbitary.1). ¯ 0 Then the equilibrium state ¯ is (a) stable if Assumptions (i) and (ii) hold (b) uniformly stable if Assumptions (i).4) (iii) there exists a continuous. we can choose x ¯ δ(t0 .4.6). (5. As β ∗ (t0 . x) ∈ I × ¯ ¯ n (5. y ) for ||¯|| ≤ ||¯||.4. non-decreasing function β such that β(0) = 0 and V (t. t0 . t0 . x0 )) ≤ V (t0 . ) such that β ∗ (t0 .1 Let V (t. ||¯||) is a continuous function of x.5) Theorem 5. for t ≥ t0 and ||x0 || < δ ¯ . t0 . x(t. x0 ) < α( ) ¯ (5. we have used Assumption 1(iii). x0 )|| < x ¯ where δ is independent of t0 .

(5. (5. t0 . Define T (µ. then ||¯(t. t0 . we find λ(µ) > 0 such that β(λ) < α(µ).4. x0 )) ¯ ¯ t1 = V (t0 .4. t0 . we have ||x2 || = ||x(t2 .4. t0 . ∃ T (µ.5. c2 (µ. x0 ))dτ ¯ r(||¯(τ )||)dτ x ≤ V (t0 . we have ¯ ||x0 || ≤ r ⇒ ||¯(t. t0 . given µ > 0. LYAPUNOV STABILITY 167 (c) For asymptotic stability of the zero solution. t1 ].2) gives x ¯ 0<λ ≤ V (t1 . r).4.3) and Eq. Hence Eq. x0 )|| = λ for some t = t2 in [t0 . t0 . As β(µ) is continuous at 0. Also r(||¯||) is a continuous function and hence it assumes x its minimum on a compact set λ ≤ ||¯|| ≤ c1 (r) and let this minimum be x c2 (µ. x(t. t1 ]. t1 = x ¯ t0 + T. ¯ x ¯ Equivallently. x0 )|| < µ for t ≥ t0 + T ¯ x ¯ Let functions α(µ) and β(µ) be as defined in Eq. Take any positive constant c1 and find r > 0 such that β(r) < α(c1 ). we only need to show that ∃ r > 0 such that ||x0 || ≤ r ⇒ ||¯(t. If not. t0 . r) We first claim that ||¯(t. t0 )c2 ≤ β(r) − T c2 = 0 which is a contradiction.5). x0 )|| → 0 uniformly with t → ∞. r) = β(r) > 0. r) such that ||x0 || ≤ r ⇒ ||¯(t. x0 ) − t0 ≤ V (t0 . Hence. x0 )|| > λ on [t0 . x0 )|| = λ ¯ ¯ . x0 )|| < c1 ∀ t ≥ t0 ¯ x ¯ ¯ Let 0 < µ ≤ ||x0 ||. x0 ) − (t1 . Then by part (b). (5. x0 ) + t0 t1 ˙ x V (¯(τ. Let ||x0 || ≤ r. t0 .

un ]. λ2 . r) ≥ 0 x ¯ ∀ t ≥ t0 The real weakness of Lyapunov method is that no general technique is known for the construction of the Lyapunov function V (t. v2 . x).1.   . there exists a basis {v1 . t2 . · · · . · · · .1. k + 1 ≤ j ≤ n. · · · . λk and the remaining are complex eigenvalues λj = aj ± ibj . x(t.. vk+2 . vk+1 . t0 . However. x2 )) x ¯ ¯ ≤ V (t2 . Then by Theorem 5. un } of ¯ ¯ ¯ ¯ ¯ ¯ ¯ 2n−k where vj (1 ≤ j ≤ k) are eigenvectors corresponding real eigenvalues ¯ and uj ± ivj (k + 1 ≤ j ≤ n) are the complex eigenvectors. vk+1 . vk+2 . x2 ) ≤ β(λ) < α(µ) This implies that ||¯(t. t0 . we get the reduced system ¯ ¯ d¯ y = P −1 AP y = D¯ ¯ y dt .   .168 Therefore CHAPTER 5. If P = [¯1 . · · · . vk . v2 . vn . then P is nonsingular and P −1 AP v ¯ ¯ ¯ ¯ ¯ ¯ = D. vk . Using the transformation x = P y .. Bn with Bi = ai −bi as a Jordan block corresponding to the eigenvalue ai ± bi a i ibi . where D is block diagonal matrix of the form   λ1   λ2     . vn . x0 )||) ≤ V (t. x0 )|| < µ ∀ t ≥ t0 + T (µ. STABILITY ANALYSIS α(||¯(t. we shall describe ¯ the method of construction of Lyapunov function for asymptotically stable linear autonomuous system d¯ x = A¯ x dt We assume that A has k distinct real eigenvalues λ1 . · · · .     λk D=    Bk+1     .

−2 ± 2i.9) will hold. y (t) ¯ ¯ (5. (5. Example 5. (5.7). y) y ¯ ¯ (5. y(t) + B y (t). y (t) y ¯ ¯ ¯ = B = In order to ensure that BD + D B y (t).4.7) where B is a suitable positive definite matrix. We seek Lyapunov function V of the forrm V (¯) = (B y .3 Construct a Lyapunov function for the system   −1 −1 0 0 d¯  0 −1 0 x  0  x = ¯ 0 0 −3 −2  dt 0 0 1 −1 this system are −1 ± i. (5. ¯ ¯ dt dt = (BD¯(t).4. we need to solve the matrix equation given by Eq. LYAPUNOV STABILITY 169 For asymptotic stability.4.9) where Q is a given positive definite matrix. Eq. to get a Lyapunov function which is positive definite with (5.4. if we assume that BD + D B = −Q Thus.5. y (t)) + (B y (t). y (t)) y ¯ dt (5.4.4.4.10) dV as negative dt definite.10) for B with a given positive definite matrix Q (say Q = I).8) dV is negative definite. we get dV dt d¯ y dy .4. Differentiating Eq. we require all diagonal elements of D to be negative.   0 0 0  1 0 0   with P −1 =    0 −1 −1 0 1 0  The eigenvalues of One gets  1  0 P =  0 0 Then 1 0 0 0  0 0 0 1 0 0   0 0 1  0 −1 −1  −1 1 0 0  −1 −1 0 0   P −1 AP = D =   0 0 −2 −2  0 0 −2 −2 . D¯(t)) y ¯ ¯ y = (BD¯(t). we shall require that dt dV = − (Q¯(t). y (t)) + D B y (t).4.

170 CHAPTER 5. y2 = x2 . (5. Solve B to satisfy the matrix equation ¯ ¯ A B + B A = −I . y3 = x4 . B a positive definite matrix. We construct a Lyapunov function for Eq.11) A has eigenvalues with negative real part and g(¯) satisfies the growth condition x of the form ||g(¯)|| ≤ k||¯|| ∀ x x x ¯ with small k. x) .4 Consider the nonlinear system of the form d¯ x = A¯ + g(¯) x x dt (5.4.4. y4 = −x3 − x4 V (¯) = x x2 x2 (x3 + x4 )2 x2 1 + 2 + + 4 2 4 4 4 Example 5. V (¯) x = (B x.4.11) in the following way. STABILITY ANALYSIS We make the transformation x = P y to get the reduced system ¯ ¯ d¯ y = D¯ y dt Take Q = I and solve the matrix equation BD + DT B = −I to get  1  2   0  B=  0   0 0 1 2 0 0 0 0 1 4 0 0   0    0    1 4  This gives the Lyapunov function V (¯) as y V (¯) = y where 1 2 1 2 1 2 1 2 y + y + y + y 2 1 2 2 4 3 4 4   0 0 0 1 0 0   0 0 1  0 −1 −1 We have Hence 1  0 −1 y=P x= ¯ ¯  0 0 y1 = x1 .

4. there exist functions αk (t). show that for A ∈ n×n . EXERCISES We have dV dt ˙ ˙ ˙ ¯ ¯ ¯ = x Bx + x Bx ¯ = x A + [g(¯)] ¯ x B x + x B(A¯ + g(¯)) ¯ ¯ x x 171 ¯ x x ¯ ¯ x = x (A B + BA)¯ + [g(¯)] B x + x B[g(¯)] = −¯ x + [g(x)] B x + x B[g(¯)] x ¯ ¯ ¯ x Estimating the second term within the square bracket. 3.   where A(t) =  0   0  −t 0 −t 0 2 −1 −e −2t − cos t t 2       −t2 e−3t 2.5. Amann [3].5. x dt dV Thus the Lyapunov function V (¯) is such that V is positive definite with x dt negative definite and hence the system Eq.5 Exercises d¯ x = A(t)¯ x dt   −t 0 e      1 0 . ∞) .11) is asymptotically stable. (5. Lakshmikantham et al[6. By using the Cayley-Hamilton theorem (Theorem 5. Show that all solutions of the following differential equations are bounded in [ t0 . Borckett [4]. α−2k||B|| > 0.     2 −t cos t 1. 5. 1 ≤ k ≤ n − 1 such that eAt = n−1 k=0 αk (t)Ak . Show that all the solution of the linear system tend to zero as t → ∞.1.3). we get ||[g(¯)] B x + x B[g(¯)]|| x ¯ ¯ x ≤ 2||g(¯)||||B||||¯|| x x ≤ 2k||B||||¯||2 x dV ≤ −||¯||2 [1−2k||B||] = −α(||x||) where α(r) = αr 2 . Hence For more on stability results refer Aggarwal and Vidyasagar [2].7] and Mattheij and Molenaar [8].

Test the stability. +c + 1+ dt2 dt 1 + t4 (ii) 4.   0 1 d¯  x x (i) = ¯ dt −1 0 (ii)  0 1 0  d¯  x  0 1 = 0 dt  −1 −6 −5  d¯  x  = 0 1 1 dt  1 3 1   1 2 0   x ¯      x ¯  (iii)    d¯  x  (iv) =  1 1 −3  x ¯ dt   1 −5 −3 5. asymptotic stability or unstability for the trivial solution of the following linear systems. asymptotic stability or unstability of trivial solution of each of the following system (i) dx1 dt dx2 dt dx3 dt = −2x1 + x2 + 3x3 + 8x2 + x3 1 2 = −6x2 − 5x3 + 7x4 3 = −2x3 + x4 + x2 + x3 1 2 3 1 −1 −1  . Test the stability. c > 0. STABILITY ANALYSIS d2 x dx 1 x = 0.172 (i) CHAPTER 5. +c + 1+ dt2 dt 1 + t2 dx 1 d2 x x = 0. c > 0.

show that origin is asymptotically stable but not (ii) If a(t) = − t+1 uniformly stable. dx1 dt dx2 dt dx1 (ii) dt dx2 dt dx1 (iii) dt dx2 dt dx1 (iv) dt dx2 dt (i) = −2x1 + x2 = −5x1 − 6x2 = 4x1 + x2 = 3x1 + 6x2 = x2 = 2x1 − x2 = x1 + x2 = 3x1 − x2 dx = a(t)x3 dt Find its solution. p > 1 and p is integer dt Show that 7. Discuss the stability or unstability of the origin of the following linear sytem and sketch the phase portraits. Consider the nonlinear system . EXERCISES (ii) dx1 dt dx2 dt dx3 dt = 2x1 + x2 − x2 − x2 1 2 = x1 + 3x2 − x3 sin x3 1 = x2 + 2x3 + x2 + x2 1 2 173 6. (i) Show that origin is uniformly asymptotically stable if a(t) < 0. 1 (iii) If a(t) = − . show that the solution of the corresponding IVP (t + 1)2 approaches a constant value as t → ∞ but origin is uniformly stable. 8. 1 .5.5. Consider the nonlinear system dx = −xp .

McGraw Hill Book Company. 1990 [4 ] Roser W Brockett. 1960 [6 ] V. Ordinary Differential Equations in Theory and Practice. The Theory of Matrices Vols. Inc. Dowden. SpringerVerlag. John Wiley. Method of Variation of Parameters for Dynamic Sytems.. 9. 1991 . S. Find Lyapunov functions around the origin for the following system. Show that the origin is unstable equilibrium point of the system d2 x − dt2 dx dt 2 sgn(x) + x = 0 ˙ References [1 ] Ravi Agarwal and Ramesh C Gupta. 1991 [2 ] J K Aggarwal and M. . Hutchinson and Ross. Stability Analysis of Nonlinear Systems. Lakshimantham and S. R. Chelsea Publishing Company. Differential Equations and Dynamical Systems.. 1989 [8 ] R M M Mattheij and J Molenaar. Walter de Gruyter. 1998 [7 ] V. STABILITY ANALYSIS (i) If p is odd. then origin is unstable. dx = −x dt dx (ii) = −x(1 − x) dt d¯ x 0 1 (iii) = −1 −a dt (i) x.174 CHAPTER 5. Ordinary Differential Equations: Introduction to Nonlinear Analysis. Deo. Marcel Dekker. Stability Analysis.A Martynyuk. a ≥ 0 ¯ 10. Gorden and Breach Science Pub. G. Gantmacher. then origin is uniformly stable and (ii) If p is even. John Wiley and Sons. I and II. Inc. Vidyasagar. 1977 [3 ] Herbert Aman. Lakshmikantham. Finite Dimensional Linear Systems. 1970 [5 ] F. Essentials of Ordinary Differential Equations. Leela and A. 1996 [9 ] Lawrence Perko. (Editors) Nonlinear Systems.

We mainly focus on Legendre.1.1 Consider the geometric power series k=0 k=0 tk = 1 + t + t 2 + · · · The partial sums sn (t) = 1 + t + · · · + tn satisfy the relation tsn (t) = t + t2 + · · · + tn + tn+1 and hence sn (t) = 1 − tn+1 .1. Hermite and Bessel differential equations. We also deduce various interesting properties of Legendre and Hermite polynomials as well as Bessel functions. c2 . A series which does not converge is said to be a divergent series. Example 6. c1 .Chapter 6 Series Solution This chapter lays emphasis on the classical theory of series solution for ODE by using power series expansion around ordinary and singular points. 1−t 175 . · · · are constants.1 Preliminaries ∞ A power series in powers of (t − t0 ) is an infinite series of the form ck (t − t0 )k = c0 + c1 (t − t0 ) + c2 (t − t0 )2 + · · · (6. This limit f (t) is denoted as the sum of the series at the ∞ point t. ∞ k=0 n Recall that a series ck (t − t0 )k is convergent at a point t if the limit of the partial sums sn (t) = ck (t − t0 )k exists. 6. 1−t This gives f (t) = lim sn (t) = 1 for |t| < 1.1) k=0 where t is a variable and c0 .

∞ k≥1 Then the orginal series k=0 ∞ ak converges. (i) Comparision test ∞ (a) Let a series k=0 ak of real numbers be given and let there exists a ∞ convergent series k=0 bk of nonnegative real numbers such that |ak | ≤ bk . SERIES SOLUTION ck (t − t0 )k is said to converge absolutely Definition 6.1. One can show that if the series converges absolutely. (b) Let a series k=0 ak of real numbers be given and let there exists a ∞ divergent series k=0 dk of nonnegative real numbers such that |ak | ≥ dk .176 ∞ CHAPTER 6. .1 A power series ∞ k=0 at a point t if the series k=0 |ck (t − t0 )k | converges. then it also converges. ∞ k≥1 Then the orginal series k=0 ak diverges. k=0 (b) The series (c) No conclusion if L = 1. (ii) Ratio test ∞ Let the series k=0 ak ( with an = 0) be such that lim ∞ k→∞ ak+1 = L. We have the following tests for checking the convergence or divergence of a series of real numbers. ak diverges if L > 1. However. ak (a) The series k=0 ∞ ak converges absolutely if L < 1. the converse is not true.

∞ Definition 6. in turn. (6. the series ∞ k=0 t − t0 t1 − t 0 converges and hence by comparision test the series k=0 ck (t − t0 )k converges absolutely. (a) The series k=0 ∞ ak converges absolutely of L < 1. (6. it follows that the partial sums sn (t1 ) converge and hence sn (t1 ) is Cauchy.6. . then r is called the radius of convergence. k k=1 it diverges and t = −1.2) implies that |ck (t − t0 )k | = ck (t1 − t0 )k t − t0 t1 − t 0 k k k (6. the radius of convergence of this series is 1. ak diverges if L > 1.1. (6. At t = 1. Proof : Since the series given by Eq. Theorem 6. This implies that sn+1 (t1 ) − sn (t1 ) → 0 as n → ∞ This. it converges. Thus. then it converges absolutely for every t for which |t − t0 | < |t1 − t0 |. implies that ck (t1 − t0 )k −→ 0 as k → ∞.1) converges at a point t = t1 . k=0 (b) The series (c) No conclusion if L = 1. That is. ∞ k t The power series converges for |t| < 1 and diverges for |t| > 1.1 If the power series given by Eq.2 If the series k=0 ck (t − t0 )k converges absolutely for |t − t0 | < r and diverges for |t − t0 | > r. |ck (t1 − t0 )k | ≤ M for all k ≥ 1 Eq.1.3) If |t−t0 | < |t1 −t0 |.1. (6.1) are bounded at t = t1 .1.1.2) t − t0 t1 − t 0 ≤ M ∞ (6. Hence the elements ck (t1 − t0 )k of the series given by Eq.1.1) converges for t = t1 . PRELIMINARIES (iii) Root test ∞ 177 If the series k=0 ak is such that lim ∞ k→∞ (|ak |) k 1 = L.1.1.1.

(i) The function f (t) in Eq. (6.1. if L = 0. 1 Then.1. (iii) If {|ck | k } does not converge. If this sequence is not bounded.4) One can easily show the following.1) converges for all t. term by term differentiation and integration.1.2 (Radius of convergence) (i) Suppose that the sequence {|ck | k } converges . That is. multiplication. ∞ ∞ ∞ If f (t) = k=0 ∞ ck (t − t0 )k in |t − t0 | < r1 and dk (t − t0 )k in |t − t0 | < r2 g(t) = k=0 ∞ Then f (t) ± g(t) = k=0 (ck ± dk )(t − t0 )k in |t − t0 | < r = min(r1 . ∞ Then the sum of the series is a function f (t) of t and we write f (t) = k=0 ck (t − t0 )k (6.1) may be determined from the coefficients of the series as follows.1. the radius of convergence r of the power series is .4) is continuous at t = t0 . then ck = dk for all k ≥ 0. (i) Addition and Subtraction Two power series k=0 ck (t − t0 )k and k=0 dk (t − t0 )k can be added and subtracted in the common radius of convergence. then r = 1 where l = l 1 sup{|ck | k }. We can carry out the standard operations on power series with ease . L (ii) If L = 0. then r = 0 and hence the series is convergent only for t = t0 . . SERIES SOLUTION The radius of convergence of the power series given by Eq. (ii) The same function f (t) can not be represented by two different power series with the same centre. then r = ∞ and hence the series Eq. Let L denotes its limit. 1 1 ∞ Let k=0 ck (t − t0 )k be a power series with non zero radius of convergence r. r2 ). (6. if ∞ ∞ f (t) = k=0 ck (t − t0 )k = k=0 dk (t − t0 )k in a disk: |t − t0 | < r. (6. Theorem 6.1.addition and subtraction. but it is bounded.178 CHAPTER 6.

6.1. PRELIMINARIES (ii) Multiplication

179

If f (t) =
k=0 ∞

ck (t − t0 )k in |t − t0 | < r1 and dk (t − t0 )k in |t − t0 | < r2

g(t) =
k=0

Then h(t) = f (t)g(t) defined within the radius of convergence of each series and

h(t) =
k=0 k

ck (t − t0 )k

where ck =
m=0

cm dk−m . That is

h(t) = c0 d0 + (c0 d1 + c1 d0 ) t + (c0 d2 + c1 d1 + c2 d0 ) t2 + · · · The series converges absolutely within the radius of convergence of each series. (iii) Differentiation

Let f (t) =
k=0

ck (t − t0 )k in |t − t0 | < r

Then df = f˙ = dt (iv) Integration
∞ ∞ k=1

ck k(t − t0 )k−1 in |t − t0 | < r

Let f (t) =
k=0

ck (t − t0 )k in |t − t0 | < r ck (t − t0 )k+1 in |t − t0 | < r k+1

Then

f (t)dt =
k=0

Definition 6.1.3 We shall say that a function f (t) is analytic at t = t0 if it

can be expanded as a sum of a power series of the form
k=0

ck (t − t0 )k with

a radius of convergence r. It is clear that if f (t) is analytic at t0 , then ck = f (k) (t0 ) , k = 0, 1, 2, · · · ( f k (t0 ) denotes the k th derivative of f at t0 ). k!

180

CHAPTER 6. SERIES SOLUTION

Example 6.1.2 We have the binomial expansion for (1 − t)−k for a fixed positive number k to give us (1 − t)−k = 1 + kt + We denote by ur = Then (1 − t)−k = k(k + 1)...(k + r − 1) r!

k(k + 1) 2 k(k + 1)...(k + r − 1) r t + t + · · · for |t| < 1 2! r!

ur t r ,
r=0

|t| < 1

As a power series can be differentiated term by term in its interval of convergence, we have

k(1 − t)−k−1 = This gives k(1 − t)−k

rur tr−1 ,
r=1 ∞

|t| < 1

1 = rur tr−1 , (1 − t) r=1

|t| < 1

Again, using the power series expansion for
∞ ∞

1 = 1 + t + t2 + · · · , we get 1−t

k
r=0

tr
r=0

ur t r

=
r=1

rur tr−1

Using the product formula for LHS, we get
∞ r ∞

k
r=0

tr
l=0

ul =
r=0

(r + 1)ur+1 tr

Uniqueness of power series representation gives
r

(r + 1)ur+1 = k
l=0

ul

(6.1.5a)

where ur = k(k + 1).....(k + r − 1) , k is a fixed integer r! (6.1.5b)

We now make an attempt to define the concept of uniform convergence. To define this concept, we consider the following series whose terms are functions ∞ {fk (t)}k=0

k=0

fk (t) = f0 (t) + f1 (t) + f2 (t) + · · ·

(6.1.6)

Note that for fk (t) = ck (t − t0 )k , we get the power series.

6.2. LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS

181

Definition 6.1.4 We shall say that the series given by Eq. (6.1.6) with sum f (t) in an interval I ⊂ is uniformly convergent if for every > 0, we can find N = N (t), not depending on t, such that |f (t) − sn (t)| <

for all n ≥ N ( )

where sn (t) = f0 (t) + f1 (t) + · · · + fn (t). Theorem 6.1.3 A power series
k=0

ck (t − t0 )k with nonzero radius of conver-

gennt is uniformly convergent in every closed interval |t − t0 | ≤ p of radius p < r. Proof : For |t − t0 | ≤ p and any positive integers n and l we have cn+1 (t − t0 )n+1 + · · · + cn+l (t − t0 )n+l

≤ |cn+1 | pn+1 + · · · + |cn+l | pn+l

(6.1.7)

The series
k=0

ck (t − t0 )k converges absolutely if |t − t0 | ≤ p < r ( by Theorem > 0, ∃ N ( ) such that

6.1.1) and hence by Cauchy convergence, given |cn+1 | pn+1 + · · · + |cn+p | pn+l < Eq. (6.1.7) gives cn+1 (t − t0 )n+1 + · · · + cn+p (t − t0 )n+l ≤ |cn+1 | pn+1 + · · · + |cn+p | pn+l <
n

for n ≥ N ( ), l = 1, 2, ..

for n ≥ N ( ), |t − t0 | ≤ p < r for n ≥ N ( )

This implies that, given and all t where sn (t) =

> 0, ∃ N ( ) such that |f (t) − sn (t)| <

k=0

ck (t − t0 )k . This proves the uniform convergence of

the power series inside the interval |t − t0 | ≤ p < r. Example 6.1.3 The geometric series 1 + t + t2 + · · · is uniformly convergent in the interval |t| ≤ p < 1. It is not uniformly convergent in the whole interval |t| < 1.

6.2

Linear System with Analytic Coefficients
n

We now revisit the non-autonomous system in

d¯ x = A(t)¯ + g(t) x dt x(0) = x0 ¯ ¯

(6.2.1)

182

CHAPTER 6. SERIES SOLUTION

where the matrix A(t) is analytic at t = 0 and hence has the power series representation

A(t) =
k=0

A k tk

(6.2.2)

in its interval of convergence |t| < r. Here each Ak is n × n matrix. So, the homogeneous system corresponding to Eq. (6.2.1) is given by d¯ x dt x(0) ¯
∞ ∞

=
k=0

A k tk x ¯

(6.2.3)

= x0 ¯

We shall look for analytic solution of Eq. (6.2.3), which is of the form x(t) = ¯ ck tk . The vector coefficients ck are to be determined. The point t = 0 is ¯ ¯
k=0

called ordinary point of the above system.

The following theorem gives the analytic solution of the system given by Eq. (6.2.3) Theorem 6.2.1 The homogeneous system given by Eq. (6.2.3) has analytic

solution x(t) = ¯
k=0

ck tk in the interval of convergence |t| < r. This solution x(t) ¯ ¯

is uniquely determined by the initial vector x0 . ¯ Proof : Let x(t) = ¯
k=0

ck tk , where the vector coefficient ck are yet to deter¯ ¯

mined. Plugging this representation in Eq. (6.2.3) we get
∞ ∞ ∞

k¯k t c
k=1

k−1

=
k=0 ∞

Ak t
k

k k=0

c k tk ¯

=
k=0 l=0

Ak−l cl ¯

tk

Equivalently, we have
∞ ∞ k

(k + 1)¯k+1 t = c
k=0 k=0 l=0

k

Ak−l cl ¯

tk

(6.2.4)

Uniqeness fo power series in the interval |t| < r, gives
k

(k + 1)¯k+1 = c
l=0

¯ Ak−l cl

(6.2.5)

6.2. LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS and hence
k

183

(k + 1) ck+1 ≤ ¯

Ak−l
l=0

cl ¯

(6.2.6)

By Theorem 6.1.3, the power series
k=0

Ak tk converges absolutely and uniformly

in the interval |t| ≤ p < r and hence the terms Ak pk must be uniformly bounded. That is, ∃ M such that Ak pk ≤ M, k≥0 (6.2.7)

Using the above inequality in Eq. (6.2.6), we get
k

(k + 1) ck+1 ≤ ¯

M
l=0

cl ¯ pk−l

Put dl = pl cl , then the above inequality becomes ¯
k

(k + 1)dk+1 ≤ M p Using Eq. (6.2.8) inductively, we get d1 d2 ≤ M pd0 Mp (d0 + d1 ) ≤ 2 1 ≤ M p + M 2 p2 d0 2

dl
l=0

(6.2.8)

. . . dk ≤ M p(M p + 1)(M p + 2) + · · · + (M p + k − 1) d0 k! (6.2.9)

To claim that Eq. (6.2.8) holds for all k, we need to prove this inequality by induction. So, let this inequality be true for all r < k. By Eq. (6.2.8), we have
r

(r + 1)dr+1

≤ M p c0 ¯ ≤ M p c0 ¯

dl
l=0 r

l=0

M p(M p + 1)(M p + 2) + · · · + (M p + l − 1) l!

Using the notation of Example 6.1.2, set ul = M p(M p + 1)(M p + 2) + · · · + (M p + l − 1) l!

2.2. SERIES SOLUTION r (r + 1)dr+1 ≤ M p c0 ¯ = ul l=0 c0 (r + 1)ur+1 ¯ This gives (r + 1)dr+1 That is dr+1 ≤ M p(M p + 1)(M p + 2) + · · · + (M p + r) c0 ¯ (r + 1)! ≤ (r + 1) M p(M p + 1)(M p + 2) + · · · + (M p + r) c0 ¯ (r + 1)! This proves the induction.3).10) provided |t| ≤ p < r. ¯ ¯ . (6. (6. (6.2.3). This proves the existence of analytic solution of the system given by Eq. it follows that ck = ¯ This gives ∞ dk M p(M p + 1)(M p + 2) + · · · + (M p + k − 1) ≤ c0 ¯ k p k!pk x ¯ = k=0 c k tk ¯ ∞ ≤ That is c0 × ¯ k=0 M p(M p + 1)(M p + 2) + · · · + (M p + k − 1) k! |t| p k x ≤ ¯ c0 ¯ |t| 1− p Mp (6. then z (t) = x(t) − y (t) is the solution of the initial value problem ¯ ¯ ¯ d¯ z = A(t)¯(t) z dt z (0) = ¯ ¯ 0 Since c0 = 0. ¯ ¯ 0. For if x(t) and y (t) ¯ ¯ ¯ are two different solutions of the initial value problem given by Eq. we get CHAPTER 6.5). Hence.1. This solution is uniquely determined by the initial value x0 . it follows that z = ¯ thereby implying that x = y .184 Hence using Eq.

We substitute ∞ k=0 ∞ k=2 ck tk .3 4! (n + 2)(n − 1) (n + 4)(n − 3) c1 .2.11) to get ∞ ∞ ∞ (1 − t ) 2 k=2 k(k − 1)ck t k−2 − 2t kck t k=1 k−1 + n(n + 1) k=1 c k tk = 0 Equivalently.4 (n + 4)(n + 2)(n − 3)(n − 1) = c1 5! . 2.12) This gives us c2 c3 = − = − n(n + 1) (n + 3)(n − 2) (n + 3)(n + 1)n(n − 2) c0 . dx = dt ∞ kck tk−1 .2. we have ∞ k=0 [(k + 2)(k + 1)ck+2 + ck [n(n + 1) − (k − 1)k − 2k]] tk = 0 The uniqueness of series representaion in |t| < 1 gives ck+2 [n(n + 1) − (k − 1)k − 2k] ck (k + 2)(k + 1) (n − k)(n + k + 1) = − ck .11) dt2 dt dx By substituting x1 = x. x2 = .1 that ∞ Eq.2.2.2.11) has a unique analytic solution of the form determine the coefficients ck .2. LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS Example 6. c5 = − c3 3! 5. this differential equation is equivalent to the dt following linear homogeneous system (1 − t2 ) d¯ x = A(t)¯(t) x dt where  A(t) =   0 −n(n + 1) 1 − t2 1 2t 1 − t2    As A(t) is analytic in the interval |t| < 1. 1. · · · (k + 2)(k + 1) = − (6. k = 0. k=1 d2 x = dt2 k(k − 1)ck tk−2 in Eq. (6.6.2.1 ( Legendre Differential Equation) 185 d2 x dx − 2t + n(n + 1)x = 0 (6. We shall now x(t) = k=0 c k tk . it follows by Theorem 6. (6. c4 = − c2 = c0 2! 4.

SERIES SOLUTION and hence the coefficients ck ’s are inductively defined. we get ck = − This gives cn−2 = − n(n − 1) cn 2(2n − 1) (k + 2)(k + 1) ck+2 (k ≤ n − 2) (n − k)(n + k + 1) x(t) We fix cn and then compute the lower order coefficients.3. writing ck in terms of ck+2 .12).12) is zero for k = n. It is standard to choose 2n! 1. (6. In Eq. These polynomials multiplied by some constants.2. (6.. So.2.. Some of the Legendre polynowhere M = or 2 2 mials are given as below .11) as Pn (t) = (−1)m (2n − 2m)! tn−2m 2n m!(n − m)!(n − 2m)! m=0 M n−1 n whichever is an integer.186 CHAPTER 6. we have = c0 x1 (t) + c1 x2 (t) n(n + 1) 2 (n − 2)n(n + 1)(n + 3) 4 x1 (t) = 1 − t + t −··· 2! 4! (n − 1)(n + 2) 3 (n − 3)(n − 1)(n + 2)(n + 4) 5 t + t −··· x2 (t) = t − 3! 5! Since RHS of Eq. x1 (t) reduces to a polynomial of degree n and if n is odd. the same is true for x2 (t). (6. This defines 2 (n!)2 n! cn−2 = − = − = − In general cn−2m = (−1)m (2n − 2m)! 2n m!(n − m)!(n − 2m)! n(n − 1) (2n)! 2(2n − 1) 2n (n!)2 n(n − 1)2n(2n − 1)(2n − 2)! 2(2n − 1)2n (n − 1)!n(n − 1)(n − 2)! (2n − 2)! 2n (n − 1)!(n − 2)! So we get a solution of Legendre differential equation given by Eq.(2n − 1) cn = 1 when n = 0 and n = for n = 0.5. are called Legendre polynomials. we have 0 = cn+2 = cn+4 = · · · and hence if n is even.2.

14) are analytic in (−∞. we normalize cn and put it equal to 2n and then compute the coefficients in descending order.2.14)) is called Hermite differential equation. which is denoted by Hn (t).2. As in the previous example. P1 (t) = t.2.6. one encounters the following deifferential equation d2 u + λ − t2 u(t) = 0 (6.2. This gives us the solution. we “ −t2 2 put u(t) = e x(t). P0 (t) = 1. It is given by Hn (t) = (−1)m n! (2t)n−2m m!(n − 2m)! m=0 M We shall discuss the properties of the Hermite polynomials in section 4. P5 (t) = (63t − 70t + 15t) P4 (t) = 8 8 We shall discuss the properties of Legendre differential equation in section 4.2. (6. (6. (6.2.13) becomes dx d2 x − 2t + (λ − 1) x = 0 dt2 dt We take λ = 2n + 1 and get the equation d2 x dx − 2t + 2nx = 0 (6. ∞). As we look for the solution u ∈ L2 (−∞. (6.14) dt2 dt Eq. ∞).15) is a polynomial. dt dt ck satisfy the recurrence relation ck+2 = = 2k + 1 − (2n + 1) ck (k + 2)(k + 1) 2(k − n) ck (k + 2)(k + 1) (6. As the coefficient of the differential equation given by Eq.2 (Hermite Differential Equation) In the study of the linear harmonic oscillator in quantum mechanics.3. . So Eq. we get a series solution of the form ∞ x(t) = k=0 ck tk for − ∞ < t < ∞ (6. P3 (t) = (5t3 − 3t) 2 2 1 1 4 2 5 3 (35t − 30t + 3). LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS 187 1 2 1 (3t − 1).2. (6. P2 (t) = Example 6.14) we see that Plugging the representation for x(t).13) dt2 over the interval”(−∞.13) (or alternatively Eq.16) It is clear that cn+2 = 0 and hence the solution given by Eq. the Hermite polynomial. ∞). (6.2.2.15) dx2 dx and 2 in Eq.2.2.

1 The differential equation given by Eq. where n is a positive integer. Proof : We introduce a new dependent variable y (t) = tµ x(t).3.3) We prove the following theorem. we can assume that t0 = 0 and A(t) has series representation of the form ∞ (6.3. provided µ is an ¯ eigenvalue of A0 and no other eigenvalue of the form µ + n exists for A0 .1) then reduces to the equation of the form d¯ x 1 = dt t A k tk k=0 x ¯ (6. SERIES SOLUTION 6. giving the existence of a series solution of the equation given by Eq. (6. (6. we get ¯ .3 Linear System with Regular Singularities A system of the form 1 d¯ x A(t)¯ = x dt t − t0 where A(t) is analytic at t0 .3.3) Theorem 6. This gives ¯ ¯ d¯ y dt = = 1 [A(t) − µI] y(t) ¯ t ∞ 1 Ak tk − µI y (t) ¯ t k=0 ∞ Assuming that y (t) = ¯ k=0 ck tk . ∞ Eq. is said to have a regular singularity at t0 .3.1) A(t) = k=0 A k tk (6.3) has a series ∞ solution of the form x(t) = tµ ¯ k=0 ck tk in the interval |t| < r.3.188 CHAPTER 6.3.3. (6.2) in the interval of the convergence |t| < r. As before.

2. 3.4) provides k (k + 1)¯k+1 = [A0 − (µ + k + 1)I] c Using Eq. n−1 (6. it follows that there exists a positive quantity such that A0 − (µ + k)I > . 2. This gives us the recurrence relations c (A0 − µI)¯0 = 0 c (A0 − (µ + I)) c1 = −A1 c0 ¯ ¯ .5) and (ii) we get (k + 1) ck+1 ≤ ¯ Equivalently. · · · k So. (k + 1) ck+1 pk+1 ≤ ¯ M −1 (k + 1) l=0 Ak+1−l cl ¯ M (k + 1) k l=0 cl ¯ pk+1−l k (k + 1) l=0 c l pl ¯ . k = 1. = − An−l cl ¯ l=0 Since |(A0 − (µ + n)I)¯n | = 0 for all n ≥ 1. .4) (A0 − (µ + n)I)¯n c .3. · · · . .5) In view of property (ii). . and p < r we have Ak pk ≤ M. · · · (6. the recurrence relation given by Eq.3. (6.3.3. . 1.6. c1 . the above relations iteratively define c ∞ c0 . · · · We shall now show that the series x(t) = tµ ¯ ¯ ¯ ck tk converges. for some positive numbers M . cn . ¯ k=0 As in Section 6.2. LINEAR SYSTEM WITH REGULAR SINGULARITIES ∞ ∞ k 189 ck kt = ¯ k=1 k=0 l=0 k Ak−l cl ¯ − µ¯k tk .3. (6. We first observe the following (i) |A0 − (µ + k)I| = 0 for all positive integer k (ii) lim A0 − (µ + k) I k→∞ k (A0 − µI) k = lim − I k→∞ k k =n >0 k = 0.

190

CHAPTER 6. SERIES SOLUTION

Proceeding as in Theorem 6.2.1, we have the relation
M M

ck ≤ ¯ for all k ≥ 1.

+1 ··· k!pk

M

+k−1

c0 ¯

Applying this estimate to the series expansion for x(t), we get ¯

x(t) ¯


k=0

c k tk ¯
∞ M M

≤ |t|µ c0 ¯ =

k=0

+1 ··· k!

M

+k−1

|t| p

k

|t|µ c0 ¯ |t| 1− p

M

in the interval of uniform convergence |t| ≤ p < r. This proves the theorem. Example 6.3.1 (Bessel’s Equation) d2 x dx +t + (t2 − µ2 )x(t) = 0 2 dt dt µ2 d2 x 1 dx + (1 − 2 )x(t) = 0 ⇐⇒ 2 + dt t dt t We use the substitution x1 = x(t), x2 = tx(t) to get ˙ t2 dx1 dt dx2 dt = = = 1 x2 t dx1 d 2 x1 +t 2 dt dt dx1 µ2 dx1 − − (t − )x1 (t) dt dt t (6.3.6)

1 = − (t2 − µ2 )x1 (t) t This reduces Eq. (6.3.6) to d¯ x dt = 1 t  0 µ2 − t 2 1 0  x ¯

=

1 A 0 + A 2 t2 x ¯ t

6.3. LINEAR SYSTEM WITH REGULAR SINGULARITIES where A0 =

191

0 0 0 1 . It is clear that A is analytic in , A2 = −1 0 µ2 0 (−∞, ∞) and A0 has eigenvalues ±µ. Case 1: µ is not an integer ∞ Then Eq. (6.3.6) has solution of the form x(t) = tµ k=0 ck tk in (−∞, ∞). ¯ For the eigenvalue µ > 0, plugging the representation
∞ ∞

x(t) d2 x dt2

=
m=0 ∞

cm xm+µ ,

dx = cm (m + µ)tm+µ−1 dt m=0

=
m=0

cm (m + µ)(m + µ − 1)tm+µ−2

in Eq. (6.3.6) we get the recurrence relation c2m c2m+1 This gives c2 c4 . . . c2m c2m+1 (−1)m c0 , + 1)(µ + 2) · · · (µ + m) m = 1, 2, · · · = 0, m = 0, 1, 2, 3, · · · = 22m m!(µ

= −

1 c2m−2 , m = 1, 2 · · · 22 m(µ + m)

= 0, m = 0, 1, 2, · · ·

= −

c0 22 (µ + 1) c2 c0 = − 2 = 4 2 2(µ + 2) 2 2!(µ + 1)(µ + 2)

(6.3.7)

Now define the gamma function Γ(µ) as Γ(µ) =
0

e−t tµ−1 dt, µ > 0

which is a convergent integral. Γ(µ) satisfies the property that Γ(µ + 1) = µΓ(µ) 1 and Γ(n + 1) = n! If we put c0 = µ Γ(µ + 1), then c2m is given by 2 c2m = = (−1)m 22m+µ m!(µ + 1)(µ + 2) · · · (µ + m)Γ(µ + 1) (−1)m 22m+µ m!Γ(µ + m + 1)

192

CHAPTER 6. SERIES SOLUTION

This gives us the first solution Jµ (t), called Bessl’s function of order µ of the differential equation given by Eq. (6.3.6), as Jµ (t) = tµ (−1)m t2m 22m+µ m!Γ(µ + m + 1) m=0

Extending gamma function for µ < 0, we get the representation for the second solution J−µ (t) as

J−µ (t) = t−µ
m=0

22m−µ m!Γ(−µ

(−1)m t2m + m + 1)

Jµ (t) and Jµ (t) are linearly independent. Case 2: µ = n integer We have Jn (t) = tn One can easily derive that J−n (t) = (−1)n Jn (t) This gives J0 (t) = (−1)m t2m 22m (m!)2 m=0 =1−
∞ ∞

(−1)m t2m 22m+n m!(n + m) m=0

t2 t4 t6 + 4 − 6 +··· 22 (1!)2 2 (2!)2 2 (3!)2

J1 (t)

=

(−1)m t2m+1 22m+1 (k!)(k + 1)! m=0 =t− t5 t7 t3 + 5 − 7 +··· 23 (1!)(2!) 2 (2!)(3!) 2 (3!)(4!)

J0 looks similar to cosine function and J1 looks similar to sine function. The zeros of these functions are not completely regularly spaced and also oscillations are damped as we see in the following graphics.

6.3. LINEAR SYSTEM WITH REGULAR SINGULARITIES

193

1 0.8 0.6 0.4 0.2 5 -0.2 -0.4 10 15 20 t J1 J0

Figure 6.3.1: Sketch of J0 and J1 Example 6.3.2 (Bessel’s equation of order zero) Consider the equation x d2 x dx + +x =0 dt2 dt

This is equivalent to the first order system of the form 1 x d¯ =− dt t where x1 = x, x2 = t dx . dt

0 −t2

1 0

x ¯

(6.3.8)

As we have seen before, a solution of this system is given by x1 (t) x2 (t) =
k=0 ∞

(−1)k t2k 4k (k!)2 (−1)k t2k 4k (k!)2

(6.3.9) (6.3.10)

=
k=0

To get another solution y (t) linearly independent of x(t) and satisfying Eq. ¯ ¯ (6.3.8), we set   1 x1 . y (t) = ψ(t)¯(t) where ψ(t) =  ¯ z 0 x2

194 This gives

CHAPTER 6. SERIES SOLUTION

d¯ y dφ d¯ z = z (t) + ψ(t) ¯ = A(t)ψ(t)¯(t) x dt dt dt      1 1 1 x1 x2 x2 0    t t  dz =  ¯ ¯ z +   z (t)   dt 0 x2 0 −tx1 −t −tx1  0

That is

Solving for

dz we get dt

 and hence 

1 x1 0 x2

 dz =  dt  t x1 x2 t x2

0

0

−t 0 0 0 

 z (t) ¯

So, we need to solve

dz  =  dt dz1 dt dz2 dt

  z (t) ¯ (6.3.11a) (6.3.11b)

x1 x2 ˙ z1 = − z1 x2 x2 t = − z1 x2 = t

Integrating Eq. (6.3.11(a)) - Eq. (6.3.11(b)) we get z1 = c1 , z2 = −c1 x2 dt x2 (t) 2 t

Hence y (t) = ψ(t)¯(t) is given by ¯ z  c tdt 1 − c 1 x1  x2 x2 (t) 2  y (t) =  ¯  t −c1 x2 2 (t) dt x2 That is y1 (t) = c1 − c 1 x1 x2

    

−x2 /x1 ˙ dt x2 2

= −c1 x1 = −c1 x1

x2 ˙ dt x2 x2 1 1 dt tx2 1

Eq. (6. a = 2 . it follows that the second solution y (t).3. linearly independent of x(t).6. LINEAR SYSTEM WITH REGULAR SINGULARITIES y2 (t) = −c1 x2 t dt x2 2 195 Using the represntations of Eq.9) .3. is ¯ given by t2 y(t) = x(t)lnt + x(t) 1 + + · · · 4 To be more precise. it is given by Y0 (t) = 2 t (−1)m hm t2m J0 (t) ln + r + π 2 22m (m!)2 m=1 ∞ .3.10). Hence. b = r − ln2 π 1 where r is the Euler’s constant limn→∞ 1 + 1 + · · · + m − lnn is called the 2 Bessel’s function of the second kind (order zero). (6. one can show that y(t) = J0 (t)lnt + where hm = 1 + 1 2 (−1)m hm t2m 22m (m!)2 m=1 ∞ +···+ 1 m . The function Y0 (t) defined as Y0 (t) = ay(t) + bJ0 (t). we get x1 (t) x2 (t) This gives x1 dt tx2 1 = x1 dt t t2 = x1 lnt + x1 ( + · · · ) 4 1+ t2 2 t2 +··· 4 t4 t2 +··· = − + 2 16 = 1− +··· and x2 t 1 + t4 + · · · t dt = 4x2 dt x2 t4 2 −2x2 = + x2 lnt + · · · t2 2 Hence.

1) has sequence of eigenvalues {n(n + 1)} with {Pn (t)} corresponding eigenfunctions and hence they are orthogonal. it follows that J(x.4. y)|−1 = a0 y dy dx −x dt dt with a0 (t) = (1 − t2 ). (6.4.196 CHAPTER 6. SERIES SOLUTION 6.1) gives 1 −1 1 1 (yLx − xL∗ y) dt = J(x. As a0 (1) = a0 (−1) = 0. for the sake of completeness we prove this result directly.4.1. However. of Chapter 7 to claim that the eigenfunctions of L form a complete orthonormal set. y ∈ D(L) = D(L∗ ) That is L∗ = L and D(L∗ ) = D(L). We appeal to Theorem 7.1 that Eq. y)|1 = 0 and hence −1 1 −1 (yLx − xL∗ y) dt = 0 ∀x.4. y) 1 −1 −1 where J(x. this is eigenvalue problem of the form Lx = λx where L is the second order differential operator Lx = (1 − t2 ) dx d d2 x − 2t = dt2 dt dt (1 − t2 ) dx dt (6.4.1 The set of Legendre polymonials {Pn (t)} satisfying Eq.3) (6.4.1) In terms of the notation of Chapter 6.4.1) are orthogonal set of polynomials in L2 [−1.4) (6.5. 1] with D(L) as the set of all functions having second order derivatives.2. 1].5) . Then Green’s formula (to be done in Section 7.4 Properties of Legendre.4.2) We can think of L as an operator defined on L2 [−1. Hermite and Bessel Functions d2 x dx + n(n + 1)x(t) = 0 − 2t 2 dt dt Legendre polynomial Pn (t) is a solution of the differential equation (1 − t2 ) (6. We have already proved in Example 6. Theorem 6. Proof : We have d dt d dt dPn = −n(n + 1)Pn (t) dt dPm (1 − t2 ) = −m(m + 1)Pm (t) dt (1 − t2 ) (6. (6.

4) by Pm (t) and Eq. 1. That is Pm ⊥ Pn . · · · 2n n! dtn 1 n n! 2 M (6.4. we get 1 −1 1 [m(m + 1) − n(n + 1)] Pm (t)Pn (t)dt d dt (1 − t2 ) dPn dt Pm (t) − d dt (1 − t2 ) 1 −1 = −1 dPm Pn (t) dt dt = ˙ ˙ (1 − t2 ) Pn (t)Pm (t) − Pm (t)Pn (t) 1 − = 0 −1 ˙ ˙ ˙ ˙ (1 − t2 ) Pn (t)Pm (t) − Pm (t)Pn (t) Hence (Pm . u) is said to be a generating function of this sequence if ∞ F (t.2 The Legendre polynomials Pn (t) satisfy the Rodrigues formula Pn (t) = Proof : We have Pn (t) = n 1 dn (t2 − 1)n . HERMITE AND BESSEL FUNCTIONS197 Multiplying Eq.4.4.5) by Pn (t) and substracting the equations and integrating. it follows that n dt (n − 2k)! Pn (t) = 1 dn 2n n! dtn M k=0 M k=0 (−1)k n! 2n−2k t k!(n − k)! It is now clear that the sum of (t2 − 1)n and hence (−1)k n! 2n−2k t is the binomial expansion k!(n − k)! 1 dn 2 (t − 1)n 2n n! dtn Pn (t) = Definition 6. (6.4.4. u) = n=0 fn (t)un (6.4.6) k=0 (−1)k n! (2n − 2k)! n−2k t k!(n − k)! (n − 2k)! Since d (2n − 2k)! n−2k t2n−2k = t . n = 0.7) . Pn ) = 1 −1 Pm (t)Pn (t)dt = 0.1 Let {fn (t)} be a sequence of functions in an interval I. PROPERTIES OF LEGENDRE.6. A function F (t. Theorem 6. 2. (6.4.

3.4.(2n − 2) 1.....9) Proof : Differentiating the relation given by Eq.. (6. (6. we have ∞ (1 − 2tu + u ) 2 −1 2 = n=0 Pn (t)un 1 (6.(2n − 3) (2t)n − (2t)n−2 + · · · 2. .. we get ∞ (t − u)(1 − 2tu + u2 )− 2 = This gives 3 nPn (t)un−1 n=1 ∞ (t − u)(1 − 2tu + u2 )− 2 = (1 − 2tu + u2 ) 1 nPn (t)un−1 n=0 .....2.8) Proof : Let |t| ≤ r (arbitary positive number) and |u| < (1 + r 2 ) 2 − r.4 The Legendre polynomials satisfy the following recurrence relation (n + 1)Pn+1 (t) = (2n + 1)tPn (t) − nPn−1 (t).. 2....4. n = 1...3.4.. we get [1 − u(2t − u)] −1 2 13 2 1 u (2t − u)2 + · · · = 1 + u(2t − u) + 2 24 1...8) with respect to u.4... Theorem 6...(2n − 1) n u (2t − u)n + · · · + 1.3 For Legendre polynomials {Pn (t)}.4 = Pn (t) Theorem 6... Then |2tu − u2 | ≤ 2|t||u| + |u2 | 1 ≤ 2r(1 + r2 ) 2 − 2r2 + (1 + r2 ) + r2 − 2r(1 + r2 ) 2 = 1 1 1 So expanding (1 − 2tu + u2 )− 2 in a binomial series..3.4.(2n) 2. SERIES SOLUTION We have the following theorem giving the generating function for the sequence {Pn (t)} of Legendre polynomials..(2n − 1) n n(n − 1) n−2 n(n − 1)(n − 2)(n − 3) n−4 = t + t +··· t − n! (2n − 1)2 (2n − 1)(2n − 3)2..4..4...3..198 CHAPTER 6.(2n) The coefficient of un in this expression is 1.(2n − 1) 1.

6.4. PROPERTIES OF LEGENDRE, HERMITE AND BESSEL FUNCTIONS199 That is (t − u) Equivalently,
∞ n=0 ∞ ∞

∞ n=1

Pn (t)un = (1 − 2tu + u2 )

nPn (t)un−1
n=0

tPn (t)un −

Pn (t)un+1
n=0 ∞

= +

n=1 ∞

nPn (t)un−1 − 2 nPn (t)un+1

ntPn (t)un
n=1

n=0

Rewriting, we get
∞ n=1 ∞ ∞

tPn (t)un −

Pn−1 (t)un
n=1 ∞ n

= +

n=1 ∞

(n + 1)Pn+1 (t)u − 2 (n − 1)Pn−1 (t)un

ntPn (t)un
n=1

n=1

Comparing the coefficient of un , we get (n + 1)Pn+1 (t) = (2n + 1)tPn (t) − nPn−1 (t)

Corollary 6.4.1 Pn
2 1

=
−1

2 Pn (t)dt =

2 2n + 1

(6.4.10)

Proof : Recurrence relation gives (2n + 1)tPn (t) (2n − 1)tPn−1 (t) These two equations give
1

= (n + 1)Pn+1 (t) + nPn−1 (t) = nPn (t) + (n − 1)Pn−2 (t)

0 =
−1

[tPn (t)Pn−1 (t) − tPn−1 (t)Pn (t)] dt
1

=

(n + 1) (2n + 1) − (n − 1) (2n − 1)

Pn+1 (t)Pn−1 (t) +
−1 1 −1

n (2n + 1) n (2n − 1)

1 −1 1 −1

2 Pn−1 (t)dt

Pn−2 (t)Pn (t) −

2 Pn (t)dt

200 This implies that
1 −1 2 Pn (t)dt =

CHAPTER 6. SERIES SOLUTION

(2n − 1) (2n + 1)

1 −1

2 Pn−1 (t)dt

and hence
1 −1 2 Pn (t)dt

= =

(2n − 1) (2n − 3) 2 ··· (2n + 1) (2n − 1) 1 2 (2n + 1)

1 −1

2 P0 (t)dt

We now examine the properties of the Hermite polynomials. Recall that Hermite polynomial Hn (t) satisfy the differential equation d2 x dx − 2t + 2nx = 0 dt2 dt and is given by

Hn (t) =
k=0

(−1)n n! (2t)n−2k k!(n − 2k)!

We have the following theorem concerning the orthogonality of {Hn (t)} in 2 L2 (−∞, ∞) with respect to the weight function e−t . Theorem 6.4.5 The Hermite Polynomials Hn (t) are orthogonal set of polyno2 mials in the space L2 (−∞, ∞) with respect to the weight function e−t . Theorem 6.4.6 For Hermite polynomials Hn (t), we have the following formulae. (i) Rodgriues formula Hn (t) = (−1)et (ii) Generating function e2tu−u =
n=0
2 2

dn −t2 e dtn un n!

Hn (t)

(iii) Recurrence relation Hn+1 (t) = 2tHn (t) − 2nHn−1 (t), H0 = 1, H1 = 2t We now enunciate the properties of the Bessel’s function Jn (t).

6.4. PROPERTIES OF LEGENDRE, HERMITE AND BESSEL FUNCTIONS201 Theorem 6.4.7 For each fixed nonnegative integer n, the sequence of Bessel functions Jn (km t), where km are the zeros of Jn (k), form an orthogonal set in L2 [0, 1] with respect to the weight function t. That is
1

tJn (kl t)Jn (km t)dt = 0, l = m
0

(6.4.11)

Proof : The Bessel function Jn (t) satisfies the differential equation (refer Eq. (6.3.6)) ¨ ˙ t2 Jn (t) + tJn (t) + (t2 − n2 )Jn (t) = 0 Set t = ks, then the above equation reduces to the differential equation n2 d ˙ + k 2 s Jn (ks) = 0 sJn (ks) + − ds s Let km be the zeros of the Bessel function Jn (k), then we have d n2 2 ˙ sJn (kl s) + − + kl s Jn (kl s) ds s and d n2 2 sJ˙n (km s) + − + km s Jn (km s) = 0 ds s Eq. (6.4.13) - Eq. (6.4.14) give
2 2 (kl − km ) 1 1

(6.4.12)

= 0

(6.4.13)

(6.4.14)

sJn (kl s)Jn (km s)ds
0

=
0

= 0,

d d ˙ ˙ sJn (kl s) Jn (km s) − sJn (km s) Jn (kl s) ds ds ds k l = km

using the fact that Jn (kl ) = 0 = Jn (km ).

Theorem 6.4.8 The generating function for the sequence {Jn (t)} of Bessel functions is given by exp 1 1 t u− 2 u

=
n=−∞

Jn (t)un

(6.4.15)

and hence J−n (t) = (−1)n Jn (t).

202 Proof : Expanding exp exp
1 2t

CHAPTER 6. SERIES SOLUTION u−
1 u

in powers of u, we get 1 exp tu 2

1 1 t u− 2 u

=

exp

1 2

−t u (−t)l 2l l! ul

=
k=0 ∞

(tu)k 2k k!

l=0

=
−∞

cn (t)un
2k+n

t 2 = Jn (t) The coefficient cn (t) of un in the above expression is k!(k + n)! k=0 and hence we get the generating function relation given by Eq. (6.4.15) for Jn (t). 1 If we replace u by − in Eq. (6.4.15), we get v

(−1)k

1 1 exp( (v − )t) = 2 v =

Jn (t)(−1)n v −n
n=−∞ ∞

J−n (t)v n
n=−∞

Hence, it follows that

J−n (t) = (−1)n Jn (t)

Theorem 6.4.9 The Bessel functions Jµ (t) satisfy the following recurrence relations Jµ−1 (t) + Jµ+1 (t) Jµ−1 (t) − Jµ+1 (t) Proof : We have

2µ Jµ (t) t = 2Jµ (t) =

(6.4.16a) (6.4.16b)

Jµ (t) = tµ
k=0

22k+µ (k!)Γ(µ

(−1)k t2k + k + 1)

Multiplying Jµ (t) by tµ and pulling t2µ under the summation, we have

tµ Jµ (t) =
k=0

(−1)k t2k+2µ 22k+µ (k!)Γ(µ + k + 1)

6.4. PROPERTIES OF LEGENDRE, HERMITE AND BESSEL FUNCTIONS203 Differentiating the above expression we have d µ (t Jµ (t)) dt

=
k=0

(−1)k 2(k + µ)t2k+2µ−1 22k+µ (k!)Γ(µ + k + 1)

= tµ tµ−1
k=0

(−1)k t2k 22k+µ−1 (k!)Γ(µ + k)

= tµ Jµ−1 (t) Thus, we have d µ (t Jµ (t)) = tµ Jµ−1 (t) dt

d −µ t Jµ (t) = −t−µ Jµ+1 (t) dt Expanding the LHS of the above expressions, we have ˙ µtµ−1 Jµ + tµ Jµ µ−1 µ ˙ −µt Jµ + t Jµ = tµ Jµ−1 = −tµ Jµ+1 2µ Jµ (t) t ˙ = 2Jµ (t)

and

Adding and substracting the above relations, we get Jµ−1 + Jµ+1 Jµ−1 − Jµ+1 =

Corollary 6.4.2 From the definition of Jµ (t), we have J 1 (t) = 2 2 sin t, J− 1 (t) = 2 πt 2 cos t πt

and hence the above recurrence relations give J 3 (t) = 2 J− 3 (t) = 2 2 πt 2 πt sin t − cos t t − cos t − sin t t

We have the following graphics for the above functions.

5 0.4. refer Agarwal and Gupta [1]. .25 5 -0. Brown and Churchill [2]. SERIES SOLUTION 1 J-1\2 0.5 0.204 CHAPTER 6.25 5 -0.25 10 15 20 t J1\2 Figure 6.25 -0.3: Sketch of J3/2 (t) and J−3/2 (t) For more details on varioius topics in this chapter.2: Sketch of J−1/2 (t) and J1/2 (t) J3\2 0.75 -1 10 15 20 t Figure 6.4.75 0.5 J-3\2 -0. Hochstadt [3] and Kreyzig [4].

x(π) = 0 ¨ ˙ (ii) t(2 − t)¨ − 6(t − 1)x − 4x = 0.3. g : Show that f and g are analytic at t = 0.   sin t .. Show that the general solution of the Chebyshev differential equation (1 − t2 )¨ − tx + a2 x = 0 x ˙ is given by x(t) = c0 1 + + c1 t + (−a)2 (22 − a2 ).. t = 0  t . t=0   t2    1 .... 2 t=0 1. Let the functions f.((2n − 2)2 − a2 ) 2n t (2n)! n=1 (1 − a2 )(32 − a2 )..6.. x(0) = 1.. f (t) =   1... Show that the two linearly independent solutions of the differential equation t(1 − t)¨ + (1 − t)x − x = 0 x ˙ . Show that the two linearly independent solutions of the following differential equation 1 1 ˙ t2 x + t(t − )x + x = 0 ¨ 2 2 are given by ∞ x1 (t) = |t| x2 (t) = |t| (−1)n n=0 ∞ 1 2 (2t)n (2n + 1)(2n − 1). x(0) = 0 ¨ ˙ ˙ (iv) x − (sin t)x = 0. t=0 g(t) = 2.. x(0) = 1.5 Exercises → be defined as follows  1 − cos t  .((2n − 1)2 − a2 ) 2n+1 t (2n − 1)! n=1 ∞ ∞ 4..1 tn n! (−1)n n=0 5.5. x(1) = 1. EXERCISES 205 6. x(π) = 1.. Find the series solution of the following IVPs (i) x + tx − 2x = 0.. x(1) = 0 x ˙ ˙ 3.. x(0) = 0 ¨ ˙ ˙ (iii) x + et x + (1 + t2 )x = 0.

n is even 2..5.3. Show that Jn (klm t) = 0 1 2 J (klm ) 2 n+1 10. Using Rolle’s theorem show that between two consucative zeros of Jn (t).. 7. µ ∈ (i) (ii) (iii) tµ Jµ−1 (t)dt = tµ Jµ (t) + c t−µ Jµ+1 (t)dt = −t−µ Jµ (t) + c Jµ+1 (t)dt = Jµ−1 (t)dt − 2Jµ (t) 2 1 2 tJn (klm t)dt = . 1993 [2 ] James Ward Brown and Rulev Churchill. A Modern Approach. McGraw Hill....n References [1 ] Ravi P. 1993 [3 ] Harry Hochstadt.. Show that p(t) is orthogonal to all Legendre polynomials of degree strictly less than n. 8. Using Rodrigues formula.. John Wiley and Sons . Inc. Dover Publications.. Let p(t) be a polynomial of degree n. there exists precisely one zero of Jn+1 (t).n − 1 . n is odd   Pn (0) =   (−1)n/2 1.. Essentials of Ordinary differential Equations.. Gupta.((n − 1)2 + 1) (n!)2 n=1 tn ∞ ∞ = x1 (t)ln|t| + 2 n × k=1 k−2 k((k − 1)2 + 1) 6. Agarwal and Ramesh C. Advanced Engineering Mathematics. Prove the following identities for Bessel functions Jµ (t).2.. McGraw Hill Book Company. Fourier Series and Boundary Value Problems..4. Inc.. show that  0.5..2. 1999 .206 are given by x1 (t) x2 (t) = 1+ CHAPTER 6.. Inc.((n − 1)2 + 1) n t (n!)2 n=1 1. 9.. SERIES SOLUTION 1.. Differential Equations... 1964 [4 ] Erwin Kreyszig.

we shall limit ourselves to those Bi s which are linear combinations of x and its derivative.1 Introduction Lx = f (7. So the most general form of the boundary 207 . is more informative. Numerous examples are given for illustration. its properties and construction. b] Since L is of second order. 2 (7. subject to certain boundary conditions. Nonlinear boundary value problem is also the topic of our investigation. we give eigenfunction expansion technique for computing the solution of boundary value problems. Throughout this chapter. at times. i = 1. 2 dt dt a0 (t) = 0 on [a. The main tool for solvability analysis of such problems is the concept of Green’s function.1) We shall be concerned with the solution of the ordinary differential equation over the interval a ≤ t ≤ b. Towards the end. there will be two boundary conditions of the form Bi (x) = Ci . Hence there is a detailed discussion involving the definition of Green’s function through Dirac-delta function. 7.1.1. We use both fixed point theory and monotone operators for this analysis. which. This transforms the differential equation in to integral equation.2) where Ci s are given constants.Chapter 7 Boundary Value Problems In this chapter we mainly deal with boundary value problems involving second order differential operators. Here L is a second order linear differential operator of the form L = a0 (t) d d2 + a1 (t) + a2 (t). Bi s are functions of the unknown variable x.

α22 .1.3) where the coefficients αij . (7.5) This denotes the deflection of a string stretched under a unit tension at fixed end points 0 and 1 and subjected to force distribution f (t).1. To ensure that Eq.4) If α12 = β11 = β12 = α21 = β21 = β22 = 0. dt2 It is clear that we have homogeneous unmixed boundary conditions. B2 (x) ≡ x(1).1. βij are given real numbers. x(a) = x(b) ˙ ˙ Example 7.1 d2 x dt2 x(0) = f (t) = x(1) = 0 (7. then the temperature distribution x(t) satisfies − d dt K dx dt = f (t) .1. we have Lx = f. β12 ) and (α21 .208 CHAPTER 7. (7.2 Consider a rod of length unity and unit cross-section. BOUNDARY VALUE PROBLEMS condition Eq. (7. Writing Eq.6) (7.2) will be written as B1 (x) B2 (x) ≡ α11 x(a) + α12 x(a) + β11 x(b) + β12 x(b) = C1 ˙ ˙ ≡ α21 x(a) + α22 x(a) + β21 x(b) + β22 x(b) = C2 ˙ ˙ (7.3) then reduces to α11 x(a) + α12 x(a) ˙ β21 x(b) + β22 x(b) ˙ = C1 = C2 (7. If β11 = β12 = α21 = α22 = 0. If C1 = C2 = 0 . (7. Example 7.1.1.1.1. the boundary conditions are called unmixed.1. Suppose there is a heat source of density f (t) in the interior. β21 .6) in the form Eq. B1 (x) ≡ x(0). we have initial conditions x(a) = C1 x(a) = C2 ˙ The boundary conditions are periodic if they are of the form x(a) = x(b). (7. α12 .1).1. we say that boundary conditions are homogeneous. β11 . β22 ) are linearly independent . Eq.1. B1 (x) = 0 = B2 (x) where L ≡ d2 . We assume that the rod is homogeneous and the flow of heat is in the direction of t only.3) gives two distinct boundary conditions we assume that the row vectors (α11 .

2. B2 (x) ≡ x(1). B2 (x) ≡ x(0). b] with domain D. B1 (x) = 0 = B2 (x) where L ≡ m d2 . using Definition 2. By domain D(L). we mean a subspace of L2 [a. Lx) = (L∗ y.1. Then.2 Let L be the differential operator acting on the Hilbert space L2 [a. which is a function of t.1) D(L∗ ) will be obtained in the following steps: (y. ˙ dt2 This is an initial-value problem. The force on the particle is along the line and the particle starts from rest at the origin.2. with initial conditions x(0) = 0. ADJOINT OF A DIFFERENTIAL OPERATOR 209 Here. we shall denote by L2 [a. B1 (x) ≡ x(0).2 Adjoint of a Differential Operator As before. b] be a second order differential operator. b] the Hillbert space of all square integrable real valued functions on [a.6 the adjoint operator L∗ : L2 [a. b] → L2 [a.2. K is the thermal conductivity of the rod. 7. Definition 7. b] which consists of functions which have piecewise-continuous derivatives of second order and which satisfy homogeneous boundary conditions B1 (x) = 0 = B2 (x) Definition 7. b]. Example 7.3 Consider the motion of a particle along a straight line. we have homogeneous unmixed boundary conditions.2.7. Here again. Lx) = a b yLxdt y(a0 x + a1 x + a2 x) dt ¨ ˙ a = . dt We can write this in our standard form as m Lx = f. b] → L2 [a. B1 (x) = 0 = B2 (x) where L ≡ − d dt K d dt .2. If the end points of the rod are kept at zero temperature we have the boundary value problem Lx = f.1 Let L : L2 [a. B1 (x) ≡ x(0). b] is defined implicitly by the equation (y. Then the displacement x(t) satisfies Newton’s law d2 x = f (t) dt2 dx (0) = 0. x). x ∈ D(L) and y ∈ D(L∗ ) b (7.

y) = a0 (y x − yx) ˙ ˙ a . y) ∗ a Definition 7.3) Definition 7.5 L is said to be formally self-adjoint if L∗ = L. So the formal adjoint of the operator L = a0 is L∗ = a 0 d2 d + a1 + a2 2 dt dt d2 d + (2a0 − a1 ) ˙ + (¨0 − a1 + a2 ) a ˙ 2 dt dt Definition 7. It is clear that L is formally self-adjoint if a0 (t) = a1 (t). So. a formally ˙ self-adjoint operator L can be written in the form L= J(x. BOUNDARY VALUE PROBLEMS On the RHS. a second order differential operator (7.2. Lx) = [a0 (y x − xy) + xy(a1 − a0 )]a ˙ ˙ ˙ b b + a x d2 d (a0 y) − (a1 y) + a2 y dt2 dt dt Applying the definition of L∗ we see that (L∗ y. (7.2. (7.210 CHAPTER 7.3) is called the formal adjoint of L and is denoted by L∗ .2) we get the Green’s formula b a b (yLx − xL y)dt = J(x.2.2.2.3 The operator given by Eq. From Eq.4 The expression J(x. y) = a0 (y x − xy) + (a1 − a0 )xy is called ˙ ˙ ˙ the bilinear concomitant of x and y. x) b = a x d d2 (a0 y) − (a1 y) + a2 y 2 dt dt b dt (7. Thus we see that L∗ consists of two parts. integrating by parts we get (y.2) + [a0 (y x − xy) + xy(a1 − a0 )]a ˙ ˙ ˙ d2 d (a0 y) − (a1 y) + a2 y dt2 dt appearing in the integrand and some boundary terms. y) is then given by a0 (y x − yx) ˙ ˙ and the Green’s formula reduces to b a b d dt a0 (t) d dt + a2 (t) (yLx − xLy) dt = J(x.2.2.

6 Consider the boundary value problem B1 (x) = 0 = B2 (x) Its adjoint boundary value problem is L∗ x = f ∗ B1 (x) ∗ = 0 = B2 (x) Lx = f .2. Example 7.2 Consider the second order differential operator L = a0 d2 d + a1 (t) + a2 (t) dt2 dt with boundary conditions B1 (x) ≡ x(0) = 0. We know that ˙ L∗ = a 0 d2 d + (2a0 − a1 ) + (¨0 − a1 + a2 ) ˙ a ˙ dt2 dt . ADJOINT OF A DIFFERENTIAL OPERATOR 211 We say that y ∈ D ∗ (domain of L∗ ) if it is twice differentiable and if J(x. y)|0 . y).2.1 If we consider L∗ as an operator on the Hilbert space L2 [a.2. a ≤ t ≤ b (7. b] with domain D∗ . D ∗ = D.5) Boundary value problem given by Eq.2. y) = (y x − yx) |1 ˙ ˙ 0 = x(1)y(1) − y(1)x(1) − x(0)y(0) + y(0)x(0) ˙ ˙ ˙ ˙ = x(1)y(1) − x(1)y(1) ˙ ˙ ∗ ∗ So J(x. Definition 7. These are called adjoint boundary conditions. dt2 1 To determine adjoint boundary conditions. J(x.2. (7. (7. y)|1 = 0 for all x ∈ D iff B1 (y) ≡ y(1) = 0.4) is said to be self-adjoint. y)|a = 0 ∗ for all x ∈ D only if y satisfies two homogeneous boundary conditions B1 (y) = ∗ B2 (y) = 0.2.4) (7.1 L = d2 with boundary conditions dt2 B1 (x) ≡ x(0) = 0 B2 (x) ≡ x(0) = 0 ˙ L is formally self-adjoint and so L∗ = d2 . y)|b = 0 a b for every x ∈ D. one can show that L is self-adjoint as an unbounded operator if the boundary value problem given by Eq. B2 (y) ≡ y(1) = 0. Hence the boundary value problem is not self-adjoint. B2 (x) ≡ x(1) = 0. if L = L∗ and D = D∗ Remark 7. we need to find J(x.2. ˙ 0 Thus we see that though the operator is formally self-adjoint. Example 7. it can be seen that J(x.2.2.4) is self-adjoint in the sense of the above definition (Refer Dunford and Schwartz [3]). From the definition of J(x.7.

3. That is.3.1 Any solution of the equation LT = δ(t − s) is said to be a fundamental solution for the operator L.3.3. (7. y)|1 0 ∗ B1 = 0 for all x ∈ D iff ≡ y(0) = 0 ≡ y(1) − ˙ [a1 (1) − a0 (1)]y(1) ˙ =0 a0 (1) ∗ B2 Thus.2) . the solution T of Eq. y)|1 . y)|1 0 = a0 (y x − xy) + (a1 − a0 )xy|1 ˙ ˙ ˙ 0 = a0 (1) [y(1)x(1) − x(1)y(1)] + [a1 (1) − a0 (1)] x(1)y(1) ˙ ˙ ˙ −a0 (0) [y(0)x(0) − x(0)y(0)] + [a1 (0) − a0 (0)]x(0)y(0) ˙ ˙ ˙ = [a1 (1) − a0 (1)]x(1)y(1) − a0 (1)y(1)x(1) − a0 (0)y(0)x(0) ˙ ˙ ˙ J(x. 0 J(x. the inverse of L.3. The kernel of this operator will turn out to be the Green’s function of the system Eq. the boundary value problem is self-adjoint.3.2) must be interpreted in the sense of a distribution.212 CHAPTER 7. BOUNDARY VALUE PROBLEMS L will be formally self-adjoint iff a0 = a1 . (7.1) Let the boundary value problem be given by We want to find L−1 . But we will show that the solution of Eq. which we now define.1).3. then D ∗ = D. However. if L is formally delf-adjoint. we see that the adjoint boundary conditions differ from the original one and D∗ = D. 7. and use it to solve the above equation. (7. ˙ To determine the adjoint boundary conditions we first determine J(x.2) is a weak solution. (7. First we define fundamental solution for the operator L. this inverse operator is an integral operator. Since RHS of Eq. As we will see later.3 Green’s Function Lx = f B1 (x) = 0 = B2 (x) (7. Definition 7. (7.2) is a distribution.

2) in the following way:   ys (t). then integrating Eq.2).3. We now show that T (t. t>s (7. t < s d2 T dt = 1. GREEN’S FUNCTION We prove this assertion by writing Eq. (7.2) is given by   ys (t).2) formally as LT = 0.7.3.3. That is (T.3. T (t. (7.3) t<s By the theory of differential equaitons. φ) = φ(s) for every test function φ.3. Assume that T is continuous at t = s. (7. t = s 213 We will determine ys (t) by appropriate matching conditions at t = s.3.(7.3) is a weak solution of Eq. s) =  x(t). By definition 2 dk L∗ φ = (−1)k k (an−k φ) dt k=0 . We get a particular solution of Eq.2) we get s+ Let x(t) be an arbitrary solution of the homogeneous equation and let ys (t) be any other solution which is to be determined. (7. s) defined by Eq. such a solution ys (t) exists and is determined by x(t) and a0 (t).3. dt2 lim →0 a0 (t) s− where a0 is the coefficient of a0 (s) 1 a0 (s) d2 T occuring on the expression for LT . dt2 dT dt − dT dt =1 t=s− t=s+ or ys (s) = x(s) + ˙ ˙ where x(t) is an arbitrary solution of the homogeneous equation and ys (t) is another solution of the homogeneous equation satisfying ys (s) ys (s) ˙ = x(s) = x(s) + ˙ 1 a0 (s) Thus a solution T of (7. L∗ φ) = (δ. t > s T =  x(t). That is.3.

φ) That is We are now in a position to define the Green’s function of Eq. s) for the above sustem is the solution of the auxiliary problem d2 g dt2 = δ(t − s) g(0. (7. 0 ≤ t < s. L∗ φ) = φ(s) = (δ(t − s). = 0 = B2 (g) dg dt − dg dt s<t≤b = t=s− (7. Definition 7.3.3. L∗ φ) = ∞ −∞ T (t)L∗ φ(t)dt Integrating by parts twice and using boundary conditions. L∗ φ) = φ(s) + s− ∞ φLxdt + −∞ s+ φLys dt Since both x and ys satisfy the homogeneous equation LT = 0.(7.1). alternatively we can define the Green’s function g(t.3.214 CHAPTER 7.3. we get (T.1) is a fundamental solution of Lg = δ(t − s) LT = δ(t − s) which satisfies additional boundary condition B1 (g) = 0 = B2 (g). s) = 0 = g(1.3. a ≤ t < s.2 The Green’s function g(t. we get (T. s) = 0 = g(1. s) That is d2 g dt2 = 0. s) of Eq. BOUNDARY VALUE PROBLEMS (T. s) as a solution of the boundary value problem Lg B1 (g) = 0.1 Consider the string problem d2 x dt2 x(0) = f (t) = x(1) = 0 Green’s function g(t. a0 (s) Example 7.4) g is continuous at t = s and t=s+ 1 . In view of what we have just proved. s) . s < t ≤ 1 g(0.

s) at t = s gives  0≤t≤s  At. s ≤ t . dg dt − dg dt = 1. g(t. A = s − 1 s−1 So g(t. t=s− 215 t=s+ Integrating the above equation and imposing boundary conditions at 0 and 1. s) =  A(1 − t)s  . we get  0≤t<s  At. 0 ≤ t ≤ s  Finally. s ≤ t ≤ 1 Example 7. the jump condition on the derivatives dg at t = s gives dt s(t − 1).2 Consider the initial-value problem representing motion of a particle in a straight line m d2 x dt2 x(0) = f (t). s) =   t(s − 1). s) is continuous at t = s. s) =  C(t − 1). s) =  At + B. s≤t≤1  (1 − s) As − A = 1. s) ˙ m g is continuous at t = s and dg dt − dg dt = =s− t=s+ 1 . g(t. s < t < ∞ dt2 g(0.   g(t. s) = 0 = g(0.3. 0≤t<∞ = x(0) = 0 ˙ Green’s function for the above system is given by d2 g = 0. GREEN’S FUNCTION g(t. 0 ≤ t < s. s < t ≤ 1 The continuity of g(t.3. m Integrating and using the boundary conditions we get  t≤s  0.7.

s≤t m L∗ h = δ(t − s) ∗ = 0 = B2 (h) (7. In particular. s) for all s. Proof : We have Lg B1 (g) and ∗ B1 (h) = 0 = B2 (g) = δ(t − s) (7.7) by g(t.  g(t.6) by h(t. then the Green’s function g(t. (7. η b .3. If the system is self-adjoint.3.3. then we know that h = g and hence g(s. s) is defined as the solution of the system ∗ B1 (h) 1 Finally the jump condition for the derivative gives A = .3. s) and the Green’s function g(t. h)|a = 0 and hence h(s. η) = g(η.1) is self-adjoint. (7. h)|b = h(s. s). s) = g(s.7) ∗ = 0 = B2 (h) Multiply Eq. η) = g(η. h) left hand side of the above equation we get and substituting this in the J(g.3.3 The adjoint Green’s function h(t. g(t. Then we get b a (hLg − gL∗ h)dt = b b a b h(t. η)δ(t − s)dt − b a a g(t. m Hence  t≤s  0.216 CHAPTER 7. s) is symmetric. t). subtract and integrate from t = a to t = b.1 Theadjoint Green’s function h(t. if the original boundary value problem given by Eq. s ≤ t Definition 7.3. Eq.5) Theorem 7. s) =  A(t − s). s)δ(t − η)dt Using Green’s formula a (hLg − gL∗ h) = J(g. (7. s) satisfy the relation h(t.3. So  t≤s  0. s) for all s. η). s) a Since h satisfies the adjoint boundary condition J(g. s) =  (t − s)  .6) L∗ h = δ(t − η) (7. η.3. η) − g(η. BOUNDARY VALUE PROBLEMS Using continuity of g at t = s we get B = −As.

using the above mentioned lemma. (7. we have t− b ˙ x(t) = a gt (t.1).8) as b t− b x(t) = a g(t. η) be a continuous function of two variables in a region α(t) ≤ η ≤ β(t). Then it has a solution given by b x(t) = a g(t.1 (Leibnitz) Let f (t.9) ˙ represents a continuous function of t.3. g is symmetric. (7. then the derivative of the integral in Eq. t+) = gt (t−. We can write Eq. . s)f (s)ds + t+ g(t. GREEN’S FUNCTION That is. t−)f (t) − gt (t.3. s). c ≤ t ≤ d. 217 Now we shall show how Green’s function can be used to obtain the solution of the boundary value problem by constructing the inverse of the corresponding differential operator.9) ∂t is given by the generalized Leibnitz formula β(t) g(t) = ˙ α(t) ∂ ˙ f (t. the last bracketed expression vanishes. β(t))β(t) − f (t. s) be the Green’s function for the boundary value problem given by Eq. Also t− b x(t) = ¨ a gtt (t. and ˙ ∂f as well as f are continuous.3. s)f (s)ds + gt (t. where α(t) and β(t) are continuous functions. Lemma 7. a ≤ t ≤ b. s) in the triangular regions ABC and ABD respectively (Refer Fig. c≤t≤d (7. t) and gt (t.8) We shall use the following lemma which is stated without proof.1). s)f (s)ds = a g(t. Then the integral β(t) g(t) = α(t) f (t. t+)f (t)] Due to the continuity of g(t.2.2 Suppose g(t. B1 (x) = 0 = B2 (x). t) due to the continuity of g(t. η)dη + f (t.3. Theorem 7. s)f (s)ds Then. s)f (s)ds (7. 7. s) is a Green’s function for the boundary value problem Lx = f. t−) = gt (t+. s)f (s)ds + t+ gt (t. s)f (s)ds + t+ gtt (t.7. t+)f (t) We note that gt (t.3. s)f (s)ds + [g(t. η)dη.3.3. If the derivatives α(t) and β(t) exist. α(t))α(t) ˙ ∂t Proof (Proof of the main theorem): Let g(t. (7. t−)f (t) − g(t.3.

t−) − gt (t. t+)] a0 (t) = 0 + 0 + f (t) = f (t) b 1 a0 (t) a0 (t) Hence x(t) = a g(t. s) + a1 (t)gt (t. Thus. Theorem 7. (7. substituting the expressioins for x(t) and x(t) in Lx ≡ a0 (t)¨ + a1 (t)x + ¨ ˙ x ˙ a2 (t)x.1).3. we have t− Lx = a b [a0 (t)gtt (t.8) satisfies B1 (x) = 0 = B2 (x). BOUNDARY VALUE PROBLEMS S C B t=s A D t Figure 7. s)f (s)ds satisfies the given differential equation.3. s) + a2 (t)g(t. s)] ds t+ + + f (t) [gt (t. s)] ds [a0 (t)gtt (t. . In exactly the same manner one can show that x given by Eq.1: Integration in a triangular region So.3. s) + a1 (t)gt (t.3 The operator L and the integral operator G defined as b Gf = a g(t. s) + a2 (t)g(t. (7.3. s)f (s)ds are inverses of each other.8) is the solution of the boundary value problem given by Eq.218 CHAPTER 7. (7.3. x(t) given by Eq.

we have b J(x. t)f (t)dt − J(x(t).11) Multiply Eq. s))|a g(t. (7. g(t.4 If g(t.10) Theorem 7. g(t. t)) a b or b x(t) = a b J(x(s). s)) a can be calculated explicitly by the given boundary conditions. subtract and integrate from t = a to t = b. to get b a b b (hLx − xL∗ h)dt = a f (t)h(t.3. s)f (t)dt − x(s) b x(s) = a h(t. s)) a Proof : We recall that the adjoint Green’s function h(t. s). g(s.3.3.7. then it has a solution x(t) given by b b x(t) = a g(t.3.3. Eq. Green’s function also enables us to get a solution of the boundary value problem with non-homogeneous boundary conditions: Lx = f B1 (x) = α. s)f (s)ds − J(x(s).3. s)dt − a δ(t − s)x(t)dt Using Green’s formula in the LHS. t) we obtain b b x(s) = a g(s. s)f (t)dt − J(x. (7. s) satisfies the system ∗ B1 (h) L∗ h = δ(t − s) ∗ = 0 = B2 (h) (7. s) = g(s. An alternative method of solution of Eq. . h)|b = a That is b a h(t. Hence the result. s)f (s)ds − J(x(s). GREEN’S FUNCTION 219 Proof : It is clear from the definition of L and G that LGf = f and GLx = x. (7.10).11) by x(t) . h) a Using the fact that h(t.3. g(t.10) is as follows. (7. s) is a Green’s function of Eq.10) by h(t. B2 (x) = β (7.3.

s)f (s)ds + c1 x1 (t) + c2 x2 (t) g(t.3. α and β and is unique. x(1) = β It is a symmetric function and satisfies the boundary conditions g(t. We write the solution of the system as b x(t) = a g(t. b x(t) = a g(t. g(t.3. s) is given by   t(s − 1). Since a completely homogeneous system has only the trivial solution. s) is the Green’s function for the system with homogeneous boundary conditions and c1 and c2 are constants to be determined.3. corresponding to Eq. s))f (s)ds a = α = β c1 B2 (x1 ) + c2 B2 (x2 ) + Therefore. BOUNDARY VALUE PROBLEMS Let x1 (t) and x2 (t) be non-trivial solutions of the homogeneous system. As in Example 7.220 CHAPTER 7. s)f (s)ds + β α x1 (t) + x2 (t) B2 (x1 ) B1 (x2 ) It is clear that this solution depends continuously on f. t ≥ s t≥s = (s − 1). From the above g(t. respectively.1. Applying boundary conditions. s)|t=1 . (7. s) =  s(t − 1). we have b c2 B1 (x2 ) + c1 B1 (x1 ) + a b B1 (g(t. s)|t=0 = 0 = g(t. s))f (s)ds B2 (g(t. s) ˙ g(t. its Green’s function g(t. s)|t=0 ˙ =   s − 1. Example 7. s)|t=1 = s ˙ . t ≤ s g(t. t ≤ s  s.3 Let us compute the solution of the non-homogeneous boundary value problem d2 x dt2 x(0) = f (t) = α.10) satisfying B1 (x1 ) = 0 and B2 (x2 ) = 0. we must have B1 (x2 ) = 0.

Theorem 7. (7. s)f (t)dt − α(s − 1) + βs = 0 s 1 = 0 t(s − 1)f (t)dt + s s(t − 1)f (t)dt + α(1 − s) + βs or x(t) = 0 t 1 s(t − 1)f (s)ds + t t(s − 1)f (s)ds + α(1 − t) + βt 7. We need to prove the existence in two parts.4. s)f (t)dt − J(x. g) we get J(x.4 Construction of Green’s Function and Modified Green’s Function Lx = f B1 (x) = α. CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION221 Using the formula for J(x. g)|1 0 g(t. g)|1 0 = [xg − xg]0 ˙ ˙ = [x(1)g(1) − x(1)g(1)] − [x(0)g(0) − x(0)g(0)] ˙ ˙ ˙ ˙ = x(0)g(0) − x(1)g(1) ˙ ˙ = αg(0) − β g(1) ˙ ˙ 1 = α(s − 1) − βs Hence.7. For.4.1) which has only the trivial solution. . if there are two Green’s functions g1 and g2 then (g1 − g2 ) is a non-trivial solution of the completely homogeneous equation Eq.4. the Green’s function for the boundary value problem exists and is unique. first for unmixed boundary conditions and then for initial conditions.1) which is a contradiction to the hypothesis. B2 (x) = β We shall consider the boundary value problem We prove the existence of Green’s function of the above system where the boundary conditions are of unmixed and initial type only.1 For the restricted case of unmixed and initial boundary conditions of the completely homogeneous system Lx = 0. B1 (x) = 0 = B2 (x) (7. the solution x(s) of the above boundary value problem is given by 1 x(s) = 0 1 g(t.3. Proof : Uniqueness of the solution is immediate.4. by Theorem 7.4. by construction.

s) where W (x1 .) Similarly. Case 2 Boundary conditions are of initial type x(a) = 0 = x(a) ˙ . x(a) = α12 . Such a solution exists. s<t≤b 1 a0 (s) Solving the above of equations for A and B we get A B = = x2 (s) a0 (s)W (x1 . s) = This proves the existence of Green’s function for the boundary value problem. for one can take x1 (t) to be the unique solution of Lx = 0. This follows from the theory ˙ of differential equation (Refer Coddington and Levinson[1]. s)      x1 (s)x2 (t) . x1 and x2 are linearly independent. BOUNDARY VALUE PROBLEMS Case 1 Unmixed boundary conditions are given by B1 (x) B2 (x) = α11 x(a) + α21 x(a) = 0 ˙ = β21 x(b) + β22 x(b) = 0 ˙ Let x1 (t) be a non-trivial solution of Lx = 0 satisfying boundary condition at t = a and x2 (t) be a non-trivial solution satisfying boundary condition at t = b. a0 (s)W (x1 . x2 . s) is the Wronskian of x1 and x2 and is non-zero since x1 and x2 are linearly independent. x2 . a≤t<s    a0 (s)W (x1 . s) x1 (s) a0 (s)W (x1 . otherwise the completely homogeneous system will have a non-trivial solution. x2 . we have the existence of x2 (t). The Green’s function is then of the form   Ax1 (t). x2 . a ≤ t < s Continuity and jump conditions of g(t.222 CHAPTER 7. x(a) = −α11 . So  x1 (t)x2 (s)   . s) give Ax1 (s) − Bx2 (s) = 0 −Ax1 (s) − B x2 (s) = ˙ ˙  Bx2 (t). s) s<t≤b g(t. x2 .

 a0 (s)W (x1 . x2 are two linearly independent solutions of Lx = 0. s) x1 (s) a0 (s)W (x1 . s) a≤t<s t>s where x1 . s) =  x2 (t)x1 (s) − x1 (t)x2 s  .4. Continuity and jump conditions on g(t. x2 . s) Example 7. s) −1 1−s −1 = −1 x1 (s) x2 (s) . s) −1 s x1 (s) = a0 (s)W (x1 . s) = 1 = x1 (s) ˙ A = B = x2 (s) ˙ x2 (s) 1−s = a0 (s)W (x1 . CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION223 Green’s function is of the form  0. Then s W (x1 . x2 . Ax1 (s) + Bx2 (s) Ax1 (s) + B x2 (s) ˙ ˙ Solving for A and B. x2 .  g(t. x2 .4. x2 .1 Let us compute the Green’s function for the string problem d2 x = f (t) dt2 x(0) = 0 = x(1) We take x1 (t) = t x2 (t) = 1 − t. we get A = B This gives     0. x2 . s) =  Ax1 (t) + Bx2 (t). a≤t≤s t≥s = x2 (s) a0 (s)W (x1 . s) give = 0 = 1 a0 (s) g(t.7.

x2 . t ≤ s  s(t − 1). x2 (t) = k(s) x2 (s) − 1 k(s) t 1 dx k(s) x1 (s) W (x1 . dt Then 1 s W (x1 . t ≥ s Example 7. x2 . t≤s =1 We take x1 (t) = 1. t≥s m Assume that the ends of the rod are kept at 0◦ temperature x(0) = 0 = x(1) We look for two linearly independent solutions x1 (t) and x2 (t) of the homogeneous equation dx d − k =0 dt dt which satisfy the boundary conditions at 0 and 1 respectively.2 Consider the initial-value problem describing the motion of a particle in a straight line m d2 x = f (t) dt2 x(0) = 0 = x(0) ˙ x2 (t) = t as two linearly independent solutions of the d2 x homogeneous differential equation 2 = 0. BOUNDARY VALUE PROBLEMS g(t.3 Consider the problem of heat distribution in a rod of unit length and unit cross-section and with thermal conductivity k(t). s) = 1 k(s) . s) =    0.4.4. Example 7. s) =   t(s − 1).224 So CHAPTER 7. then the governing equation is − d dt k dx dt = f (t)  (t − s)  . So we write t x1 (t) = 0 ds . s) = 0 1 Hence g(t. If the heat source has density f (t).

CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION225 1 x1 (s) + x2 (s) =− =− k(s) Hence   x1 (t)x2 (s)      k(s)  k −1 (s)ds k(s) 0 k(s) 1 0 k −1 (t)dt k(s) . t≥s k −1 (t)dt  −1  x1 (t)x2 (s).4. But J0 (t) and Y0 (t) fail to satisfy the 5 boundary conditions at a and b.4 In the study of heat flow in a cylindrical shell.4. the homogeneous Bessel’s equation d dt t dg dt + tg = 0 t dg dt + tg = δ(t − s). So we define two other linearly independent solutions 1 Z0 (t) = J0 (t)Y0 (a) − Y0 (t)J0 (a) 2 Z0 (t) = J0 (t)Y0 (b) − Y0 (t)J0 (b) .7. t ≤ s 1 0 Example 7. s) =   x (s)x (t)  1  2    k(s) 1 = 0 k −1 (t)dt .5772 is the so called Euler constant which is the limit of 1 + 1 + = 2 · · · + 1 − lns as s approaches infinity. t≤s g(t. t ≥ s has two linearly independent solutions J0 (t) and Y0 (t) J0 (t) = and Y0 (t) = 2 (t/2)4 2 lnt 2 (t/2)2 − (1 + 1/2) γ+ J0 (t) + π 2 π (1!)2 π (2!)2 2 (t/2)6 + (1 + 1/2 + 1/3) − · · · π (3!)2 (−1)n t2n 22n (n!)2 n=0 ∞ where γ ∼ 0. a ≤ t ≤ b  x2 (t)x1 (s). one comes across the Bessel’s equation of the type d dt with boundary conditions g = 0 at t = a and t = b As we know.

respectively. t≤s   C  1 2  Z0 (s)Z0 (t)   .226 Obviously now CHAPTER 7. x2 . t) = W [J0 . which dt satisfy the boundary condition at 0 and 1. t][J0 (a)Y0 (b) − Y0 (a)J0 (b)] 2 [J0 (a)Y0 (b) − Y0 (a)J0 (b)] = πt So g(t. π Example 7. t≥s λ sin λ .4. √ √ x1 (t) = sin λt.5 Consider the boundary value problem − d2 x − λx dt2 x(0) = f (t) = 0 = x(1) To find the Green’s function for this problem we first look for two linearly inded2 x pendent solutions of the completely homogeneous problem − 2 −λx = 0. t≤s    λ sin λ √ √   sin λ(1 − t) sin λs    √ √ . x2 (t) = sin λ(1 − t) sin W (x1 . s) = √ √   sin λt sin λ(1 − s)  √ √ . BOUNDARY VALUE PROBLEMS 1 2 Z0 (a) = 0 = Z0 (b). t≥s C where C= 2 [J0 (a)Y0 (b) − Y0 (a)J0 (b). Y0 . It can be shown that 1 2 W (Z0 . Z0 . t) = √ λt sin √ λ(1 − t) √ √ √ √ λ cos λt − λ cos λ(1 − t) √ √ √ λ sin λ(1 − t) cos λt = √ √ + sin λt cos λ(1 − t) = √ √ λ sin λ sin √ √ λ(1 − t + t) λ = So g(t. s) =  1 2  Z0 (t)Z0 (s)  .

which is not possible as discussed earlier. These equations will have solutions if the determinant   B1 (x1 ) B1 (x2 )   B2 (x1 ) B2 (x2 ) If the determinant vanishes. we get the desired Green’s function.4. or B1 (x2 ) = B2 (x2 ) = 0 which implies that either x1 (t) or x2 (t) is a non-trivial solution of the completely homogeneous system which is a contradiction. (7. s)a0 (s) where A and B are constants which will be determined so that g(t. the Green’s function may be found in a straignt forward way. s) is continuous and its derivative has a jump of magni1 . The other possiblility for the determinant to vanish is that there exists a constant c such that B1 (x2 ) + cB1 (x1 ) = 0 B2 (x2 ) + cB2 (x1 ) = 0 These equations imply that B1 (x2 + cx1 ) = B2 (x2 + cx1 ) = 0.2) (7.2). x2 . we have proved the existence of Green’s function for the system Lx = f.4.4. Let x1 (t) and x2 (t) be any two linearly independent solutions of Lx = 0. B1 (x) = 0 = B2 (x) has only the trivial solution. B1 (x) = 0 = B2 (x) under the assumption that the completely homogeneous system Lx = 0. we have B1 (g) ≡ AB1 (x1 ) + BB1 (x2 ) + B1 (r) = 0 where r(x) stands for (7. either B1 (x1 ) = B2 (x1 ) = 0. Now consider the boundary conditions.3) B1 (g) ≡ AB1 (x1 ) + BB1 (x2 ) + B1 (r) = 0 x1 (t)x2 (s)H(s − t) + x2 (t)x1 (s)H(t − s) W (x1 . CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION227 In general. Since they are linear and tude a0 (s) homogeneous. s)a0 (s) does not vanish. What happens when the completely homogeneous system has a non-trivial solution ? . x2 .3) for A and B. It is clear that g(t.Eq. (7.4.4. The boundary conditions give us two linear equations which will determine the values of A and B.7. So far. s) satisfies the given boundary conditions. Then write g(t. So by solving Eq. where the boundary conditions are mixed. s) = Ax1 (t) + Bx2 (t) + x1 (t)x2 (s)H(s − t) + x2 (t)x1 (s)H(t − s) W (x1 .

(7.4. Example 7. dt2 has a non-trivial solution g = t. a ≤ t ≤ b.7 Consider the self-adjoint system δ(t − s)dt = 0.5) .4. (ii) The inhomogeneous system Ly = f .4. Since 0 δ(t − s)dt = 0.4) will have no solution.6 Consider the self-adjoint system d2 g = δ(t − s) dt2 g(0) = 0. a ≤ t ≤ b. So it follows by alternative theorem that Eq. BOUNDARY VALUE PROBLEMS We state the following Alternative theorem which answers the question raised above Consider the related systems: (i) The compeletely homogeneous system Lx = 0. ∗ ∗ B1 (z) = 0 = B2 (z) B1 (x) = 0 = B2 (x) B1 (y) = 0 = B2 (y) Theorem 7. (iii) The adjoint homogeneous system L∗ z = 0. d2 g = δ(t − s).4) will have a solution iff 1 0 1 (7.228 CHAPTER 7. Eq. then (iii) has also the trivial solution and (ii) has one and only one solution. Example 7.4. g(1) = g(1) ˙ The completely homogeneous system d2 g = 0.4.4. 0 ≤ t ≤ 1 dt2 g(0) = 0 = g(1) ˙ (7. refer Stakgold [5].4) g(0) = 0. g(1) = g(1) ˙ That is. a ≤ t ≤ b. the ordinary Green’s function does not exist. (iii) will also have non-trivial solution and (ii) has solution iff b f (t)z(t) = 0 for every z which is a solution of (iii).2 (a) If (i) has only the trivial solution x(t) = 0 in a ≤ t ≤ b.4. (7. (b) If (i) has a non-trivial solution. a For a proof of this.

dt2 1 0 g(0) = 0 = g(1) ˙ ˙ has a non-trivial solution g = const. where x1 (t) is normalized x2 (t)dt = 1 . (7. However. 1 It now follows that the Green’s function which is a solution of the system Lg = δ(t − s). b B1 (g) = 0 = B2 (g) will not exist for a Green’s function one has to resort to modified Green’s function by the addition of new source density of strength x1 (s). Modified Green’s Function Assume that the self-adjoint system Lx = f . b a B1 (x) = 0 = B2 (x) is such that the completely homogeneous system B1 (x) = 0 = B2 (x) has non-trivial solution of the form x1 (t). for δ(t − s)ds = 0. in order to compute the It is clear that the solution of the above system exists for b a [δ(t − s) − x1 (t)x1 (s)] x1 (t)dt = 0 The method of construction of gM is now similar to that of ordinary Green’s function which we have already discussed. s) B1 (gM ) = δ(t − s) − x1 (t)x1 (s) = 0 = B2 (gM ) x1 (t)δ(t − s)dt = 0. (7.4. a ≤ t ≤ b. s)x1 (t)dt = 0 for every s a .4. That is.4.4. a ≤ t ≤ b. CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION229 The completely homogeneous system d2 g = 0.3 The modified Green’s function gM (t. Hence Eq.7. Lx = 0.1 Modified Green’s function is the solution of the system LgM (t. Definition 7. the ordinary Green’s function can not be constructed for Eq. s) is symmetric if b gM (t. Theorem 7.5).5) has no solution.4.

s1 )) and LgM (t. (7. (7. s1 )) − gM (t. s2 )) (7. s)) = 0 = B2 (gM (t. BOUNDARY VALUE PROBLEMS Proof : Consider gM (t. s1 )dt But b a x1 (t)gM (t. we have b 0 = gM (s1 . s)dt = 0 for every s. s2 )dt x1 (s1 ) a b − x1 (t)gM (t. s1 )dt x1 (s2 ) a a Using Green’s formula and boundary conditions for gM (t. s2 ) and Eq. s1 )δ(t − s2 )dt + x1 (t)gM (t. s2 )) = 0 = B2 (gM (t.4.7) by gM (t.4. s1 ) and gM (t. s) we can alternatively define the modified Green’s function as the solution of the system: LgM (t. satisfying (i) B1 (gM (t. s2 ) wich satisfy the systems: LgM (t. s2 ) = gM (s2 . s ≤ b . gM (t.7) Multiply Eq.4. s1 ). s)) (ii) gM (t.6) by gM (t. Then we get b a b [gM (t. s1 ) − x1 (s1 ) b x1 (t)gM (t. s1 ) = δ(t − s1 ) − x1 (t)x1 (s1 ) (7. s2 )(LgM (t. s2 )δ(t − s1 )dt − gM (t. s1 ) and subtract and integrate from a to b. s1 )) = 0 = B2 (gM (t. s1 )(LgM (t. Hence gM (s1 . s) is continuous at t = s a ≤ t. s) = δ(t − s) − x1 (t)x1 (s).4.6) B1 (gM (t. s2 )dt a +x1 (s2 ) a x1 (t)gM (t.230 CHAPTER 7. s2 ) − gM (s2 . s2 ) = δ(t − s2 ) − x1 (t)x1 (s2 ) B1 (gM (t. s2 ). s1 ) On imposing the conditions of symmetry on gM (t. s2 ))] dt b = a b gM (t.

(7. s)dt = 0 for every s For computational purposes.4. then Eq. fo both x and gM satisfy zero boundary conditions and there fore we have b x(s) = a gM (t. s)f (t)dt + Cx1 (s) . CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION231 (iii) dgM dgM 1 |t=s+ − |t=s− = dt dt a0 (s) and b a (iv) x1 (t)gM (t.8) has a solution x(t) given by x(t) = a gM (t. s)f (t)dt − x(s) + Cx1 (s) The LHS vanishes. s) be the modified Green’s function of the self-adjoint system Lx = f .4. a ≤ t ≤ b.4 Let gM (t.4. it reduces to b x1 (t)x(t)dt J(x.8) having x1 (t) as the only normalized solution of the completely homogeneous system. b B1 (x) = 0 = B2 (x) (7. We have Lx = f .4.4. We finally have the following theorem which gives a solution of the self-adjoint system by using the modified Green’s function. (7. s)f (s)ds + Cx1 (t).8) is clear by alternative b theorem. s) will be used. If a b f (t)x1 (t)dt = 0. the above defining property of gM (t. s)f (t)dt − b a a x(t)δ(t − s)dt +x1 (s) Using Green’s formula. Proof : The existence of a solution of Eq. for a f (t)x1 (t)dt = 0. Theorem 7. and LgM B1 (gM ) = 0 = B2 (gM ) = δ(t − s) − x1 (t)x1 (s) B1 (x) = 0 = B2 (x) By our standard procedure we get b a b b (gM Lx − xLgM ) = a gM (t. gM ) = a gM (t.7.

s) sin 2πsds + C (7.4. we get    A− t2 2. s) is continuous at t = s dgM dgM |t=s+ − |t=s− = 1 dt dt 1 (7.9) x(0) = 0 = x(1) ˙ ˙ The completely homogeneous system has a non-trivial solution x = const.4. s) we solve the system d2 gM = δ(t − s) − 1 dt2 gM (0) = 0 = gM (1) ˙ ˙ gM (t. s)f (s)ds + Cx1 (t) Example 7.232 or CHAPTER 7.4. 0≤t≤1 (7. 1 Since given by 0 sin 2πtdt = 0. t<s t>s gM (t. s)dt = 0 for every s 0 Integrating Eq. by Theorem 7.4.8 Consider the system d2 x dt2 = sin 2πt. BOUNDARY VALUE PROBLEMS b x(t) = a gM (t.11) gM (t. s) = C +t− Continuity at t = s implies that A− or s2 s2 =C +s− 2 2 C =A−s .10) To compute gM (t. t2 2.4.4. the above system has a solution x(t) b x(t) = a gM (t.(7.11) and using boundary conditions.4.

. We get the following important theorem.7. t≤s t≥s t2 )dt = 0 2 − s3 3 The last defining condition gives s 0 A−s+t− 1 s (A − t2 )dt + 2 1 2 (A − s + t − That is. giving eigenfunctions of L as a basis for L2 [a. EIGENFUNCTION EXPANSION . s) =    s − s 2 − 1 − s + t − t2 . b] with domain D define as D = {x ∈ L2 [a.  3  gM (t. x ∈ L2 [a. t≥s 7. which gives A= s− Hence. Such a self-adjoint system is called Sturm-Liouville system. b]. t ≥ s 2 3 2 =   s−  t− 1 3 − s2 +t2 2 . 0≤t≤b Consider the self-adjoint system (7. We assume that the operator L acts on L2 [a. b].STURM-LIOUVILLE SYSTEM 233 So gM (t.2) with the assumption that λ = 0 is not an eigenvalue of L.5.5.5 Eigenfunction Expansion . b]. As − s 6 + A−s+ − 1 6 − (A − s)s + s2 1 − 2 3 s2 2 = 0.5. b] : x. we have the following symmetric modified Green’s function  2 2  t≤s s − s2 − 1 − t2 . B1 (x) = 0 = B2 (x)} ˙ ¨ We consider the related eigenvalue problem Lφ = λφ. t≤s 1 3 − s2 +t2 2 .1) = 0 = B2 (x) where a and b are finite constants and f ∈ L2 [a. t2 2. s) =    A− t2 2.Sturm-Liouville System Lx − λx B1 (x) = f. B1 φ = 0 = B2 φ (7.

φn ). φ2 .5. (7. b]. Proof : As zero is not the eigenvalue of L. Theorem 7. · · · } of L.3) in the space L2 [a. So. For. then the self-adjoint system Eq.5.5. .1 The self-adjoint system Eq.(7. b].4) in the space L2 [a. (7. g 2 (t. Let K be the integral operator generated by g(t. λ2 . Here (.2 If λ is not any one of the eigenvalues {λ1 .5. b]. φn .) is the innerproduct in L2 (a.3) is equivalent to the following operator equation [Kx] = 1 1 x = µx. (7.5. s) is a compact operator on L2 [a.2) has discrete eigenvalues {λ1 . φn . · · · } and the corresponding eigenfunctions {φ1 .2) is equivalent to the solvability of the integral equation b x(t) = λ a g(t. · · · .5.1) has one and only one solution x(t) ∈ L 2 [a.6 that eigenvalues of K are discrete and the eigenfunctions of K form a basis for L2 [a. s)x(s)ds = 0. there exists Green’s function g such that the solvability of Eq. it follows from Theorem 2. if x = 0 exists satisfying b g(t. then the boundary value problem a Lw = x. s)x(s)ds = 0. it follows that eigenvalues of L are discrete and the corresponding eigenfunctions {φ1 . s)x(s)ds Then Eq. b]. We note that µ = 0 is not an eigenvalue of K.234 CHAPTER 7.5. · · · } form a basis for L2 [a.5. This implies that x = Lw = 0. λn . BOUNDARY VALUE PROBLEMS Theorem 7. s)x(s)ds (7. · · · .2) is self-adjoint. φ2 . K is also self-adjoint as the system Eq. b]. . As K and L are inverses of each other. s) b [Kx](t) = a g(t. λ2 . B1 (w) = 0 = B2 (w) b has a unique solution w = a contradiction.5. · · · . (7. s)dtds < ∞ and hence the operator K generated by g(t. b]. · · · } form a compelete orthonormal set in L2 [a.2.. b] given by ∞ fn φn (t) x(t) = λn − λ n=1 where fn is given by fn = (f. b]. We observe that a a a b b g(t. µ = λ λ (7.

we get ∞ n=1 ∞ ∞ xn (λn − λ)φn (t) = fn φn (t) n=1 That is n=1 [fn − xn (λn − λ)] φn (t) = 0. (7.6) Operating on Eq. φn ) = a x(t)φn (t) dt (7. b] and hence we can write ∞ b f (t) = n=1 fn φn .4. (7.5.5.5.5.2.4. φn ) = a f (t)φn (t) dt (7. This remark follows by Theorem 7. (7. · · · } is a complete orthonormal set in L2 [a. Plugging in the value of xn in Eq. Eq. Completeness of {φ1 .5. EIGENFUNCTION EXPANSION .7) f (y)φn (y)dy a φn (t) λn . φ2 .5) and ∞ b x(t) = n=1 xn φn .5.5. {φ1 . fn = (f. (7. φ2 .7. uniqueness of the solution of Eq. xn = (x.5. B1 (x) = 0 = B2 (x) then the above system has a unique solution x(t) given by x(t) = fn φn (t) = λ n=1 n n=1 ∞ ∞ b (7.5.STURM-LIOUVILLE SYSTEM 235 Proof : Since λ is not any one of the eigenvalues.1 If zero is not one of the eigenvalues of the boundary value problem Lx = f .1) follows by Theorem 7.1 If λ = λj for some j.1) has infinitely many solutions if fj = 0 and no solution if fj = 0. Corollary 7.6) we get x(t) = fn φn (t) (λn − λ) n=1 ∞ Remark 7.5. · · · } implies that fn − xn (λn − λ) = 0 and hence xn = fn /(λn − λ).2.6) by (L − λI) we get ∞ ∞ (L − λI)x(t) = n=1 xn (L − λI)φn (t) = n=1 xn (λn − λ)φn (t) Since (L − λI)x = f.

we see that the above equation has infinitely many 2 solutions.2 Consider the boundary value problem − t→0 1 d t dt t dt dt = f (t) lim t dx = 0. tan k = 2k. By considering the curves y = tan x and y = 2x. φ(0) = 0. The associated eigen− 1 d t dt t dφ dt + λφ = 0 value problem is That is.5. then the above equation becomes tan k = 2k. For λ > 0. √ This is a Bessel’s equation√ zero order in the variable λt and has the solution of √ φ(t) = AJ0 ( λt) + BY0 ( λt). It can be √ easily seen that there are no negative eigenvalues. y) = 0 tx(t)y(t) dt and hence its eigenvalues are real. The corresponding normalized eigenfunctions are − cos 2kn These eigenfunctions are orthogonal and form a complete set. BOUNDARY VALUE PROBLEMS Example 7. Boundary condition φ(0) = 0 gives c2 = 0 and hence φ(t) = √ c1 sin λt.5. The second boundary condition implies that λ satisfies the equation √ √ tan λ = 2 λ Let λ = k 2 . x(1) = x(1) ˙ dt2 2 The associated eigenvalue problem is − 1 d2 φ ˙ = λφ. the solution x(t) of the above system is given by x(t) = 4 sin kn t 2 kn (cos 2kn )2 n=1 ∞ 1 f (y) sin kn y dy 0 Example 7.1 Consider the self adjoint-system − d2 x 1 = f (t). φ(1) = φ(1) dt2 2 We take up different cases. dφ 1 dφ + + λφ = 0 dt t dt . x(1) = 0 dt We first make the above system self-adjoint by defining a new inner product 1 (x. φ(t) = c 1 sin λt √ +c2 cos λt. So eigenvalues are given by λ = kn where kn is a non-zero solution of 2 sin kn t. Hence. by Theorem 7. depending upon the algebraic sign of λ.5.2.236 CHAPTER 7. But the first boundary condition gives B = 0. x(0) = 0.

6 Nonlinear Boundary Value Problem In this section we shall be concerned with two point boundary value problem of the form Lx = f (t.5.1) (7. the zeros of the Bessel’s function J0 (t) gives the eigenvalues of the above √ boundary value problem. λ) is the solution of the system Lg − λg = δ(t − s). x) is measurable for all values of x ∈ . Then the Green’s ∞ φn (t)φn (s) function for Eq. · · · } forming a complete set in L2 [a. (7. s. The second boundary condition gives J0 ( λ) = 0. √ {J0 ( λn t)} forms a complete set and hence by Theorem 7. φ(t) = AJ0 ( λt).6.1) is given by g(t.2 for f (t) = δ(t − s) (formally) we get g(t. b]. NONLINEAR BOUNDARY VALUE PROBLEM 237 √ √ Hence. s.3 Let {λ1 .2) where L is the second order differential operator as defined before and B1 (x).6. x) is continuous for almost all t ∈ [a. b] B1 (x) = α. λ) = δn φn (t) . · · · } be a discrete set of eigenvlaues of L with eigenfunctions {φ1 . λ) = provided λ is not λn − λ n=1 one of the eigenvalues.5. b]. λ) = That is g(t. (ii) t → f (t. Thus. Here f : [a.5. B2 (x) are boundary conditions on x. 7.7. Proof : Green’s function g(t. x(t)).5. B1 (g) = 0 = B2 (g).6. B2 (x) = β (7. b] × → is nonlinear function satisfying Caratheodory conditons: (i) x → f (t. λ2 . Applying Theorem 7. where δn = λ −λ n=1 n ∞ ∞ b a δ(t − s)φn (t) dt = φn (t) φn (t)φn (s) λn − λ n=1 For more on this section refer Bose and Joshi [2].2 √ ∞ fn φn (t) J0 ( λn t) √ x(t) = = λn λ J0 ( λn t) n=1 n=1 n ∞ 1 2 f (y)J0 ( λn y) dy 0 Theorem 7. The corresponding eigenfunctions are φ n = J0 ( λn t). . s. t ∈ [a. s. φ2 .

3) in the space X = L2 [a.238 CHAPTER 7.6.6. (7.6. (7. defined as below b [Kx](t) = a g(t.6. Eq. g(t. s)x(s)ds (7. Then the solvability of the nonlinear boundary value problem given by Eq. s))| a . x2 ∈ (7. BOUNDARY VALUE PROBLEMS Let g(t. b]. x1 ) − f (t.10) .6. s)f (s.1) has a unique solution x∗ (t) which is obtained as a limit of the Picard iterates xn (t).6.6.5).3) where y(t) is a known function given by y(t) = −J(x(s). x) satisfies growth condition of the form |f (t. b). (7. Proof : Solvability of Eq. x) is Lipschitz continuous . a ∈ L2 (a. b > 0 (7. defined iteratively as 1 xn (t) = T xn−1 (t) = 0 g(t. s) is a Hilbert Schmidt kernel .7) (iii) x → f (t.6.3) is equivalent to the solvability of the operator equation x = KN x + y where K. xn−1 (s))ds + y(s) (7. ∀x1 .there exists constant γ > 0 such that |f (t.1) is equivalent to the solvability of the integral equation b x(t) = a g(t. s)dtds < ∞ (7.6.5) [N x] (t) = f (t.6. That is k2 = a b a b g 2 (t.6. Then Eq. respectively. (7. (7.9) with x0 (t) = y(t).6.8) Theorem 7.1) or Eq. We shall prove the existence of solution of Eq. x2 )| ≤ γ|x1 − x2 | ∀t ∈ [a. x(t)) Assumptions I (i) g(t.6.6. (7.6. s) be the Green’s function.4). corresponding to the operator L with given zero boundary conditions.6) (ii) f (t. (7. N are as defined by Eq. s)f (s. b] by using operator theoretic approach in terms oprators K and N . x)| ≤ a(t) + b|x|.6. J being the bilinear concomitant of x and g. (7. x(s))ds + y(t) b (7.4) (7.1 Let Assumptions I hold and let kγ < 1.

11) (7.6. This proves that x∗ (t) is a solution of Eq. we find that the Green’s function g(t.1 Let us consider the nonlinear boundary value problem d2 x dt2 x(0) = γ sin x(t) = α. s) =  s(t − 1).12) This gives Following Example 7. s) for the above problem is given by   t(s − 1).1) where x∗ (t) = lim xn (t).1. x2 ∈ X ∀x ∈ X (y ∈ X fixed) X ≤ γ x 1 − x2 X ∀x1 . Thus. 90 3( 10) . t ≤ s g(t.8 that K is a compact operator on X. x(1) = β (7. it follows that T is a contraction on the space X and hence it has a fixed point x∗ and this point is approximated by the iterates xn = T xn−1 . x0 = y. Further Assumption I(iii) implies that N is Lipschitz continuous. N x1 − N x2 Let T x = KN x + y. we have n→∞ b xn (t) = and y(t) + a g(t. that is. Also. xn−1 (s))ds x0 (t) = y(t) This proves the theorem. t ≥ s 1 1 g 2 (t.6. x2 ∈ X ≤ kγ x1 − x2 If α = kγ < 1. s)f (s. it follows that N is a bounded continuous operator on X ( refer Joshi and Bose [4] ). Then T x1 − T x2 = ≤ K(N x1 − N x2 ) K N x 1 − N x2 ∀x1 .6.7.3. inview of Assumption I(ii).2. Example 7. (7. s)dtds = 0 0 1 1 and hence k = K = √ .6.6. NONLINEAR BOUNDARY VALUE PROBLEM 239 We have already proved in Example 2.

6.2 Consider the following boundary value problem: d2 x + γx + sin x dt2 x(0) = α.6. s) sin xn−1 (s)ds 0 If f is monotone increasing. (7.4) to get a solvability results for Eq. (iii) x → f (t. 1 √ < 1.6. Refer Joshi and Bose [4] for the proof of this theorem.6. inverse of the differential operator Lx = − .13) where y(t) = α(1 − t) + βt.11) is equivalent to the solvability of the integral equation 1 x(t) = γ 0 g(t.1). (7.2 Under Assumptions II. Assumptions II (i) [g(t.6.6. defined.14) (7. (ii) Assumption I(ii) holds. (7. s)[γx(s) + sin x(s)]ds = α(1 − t) + βt 1 Observe that [Kx](t) = 0 g(t.11) has a unique If we assume that γ 3( 10) solution x∗ (t). x1 ) − f (t. x2 ∈ Theorem 7. Eq. s)] is symmetric positive semidefinite ( K is monotone) and Hilbert Schmidt. x2 )) (x1 − x2 ) ≥ c(x1 − x2 )2 x1 . Solvability of Eq. s) sin x(s)ds + y(t) (7. it follows that Eq. (7. Example 7.6. x(1) = 0 = β (7.240 CHAPTER 7. g(t. BOUNDARY VALUE PROBLEMS Solvability of Eq. (7.6.1) has a unique solution. x) is strongly monotone: ∃ c > 0 such that (f (t. iteratively as 1 xn (t) = α(1 − t) + βt + γ with x0 (t) = α(1 − t) + βt.6. which is the limit of xn (t). s)x(s)ds is positive semi definite as K is the d2 x which is positive semi definite dt2 (monotone).6.14) is equivalent to the solvability of the integral equation 1 x(t) + 0 g(t. It is clear that f (x) = γ sin x which satisfies Assumptions I(i) and (ii).15) Assume that γ > 1. we may resort to the concept of monotone operators (see Section 2.

s)f (s)ds of the system Lx = f. we get 241 . Consider the operator L with boundary conditions x(0) = x(1).7 Exercises 1.1 whereas Theorem 7. w(t) > 0 ˙ w(t) 4.6. x2 ∈ As g and f satisfy all Assumptions II. ∂f (t. ˙ ˙ dt2 3.6. B1 (x) = 0 = B2 (x) depends continuously on f . dt2 d2 with boundary conditions x(0) = x(1). 2.1 If γ is such that γ > 3( 10) − 1 > 1 then Eq. 7. x) = γ + cos x ≥ (γ − 1) > 0 for all x ∈ ∂x Hence. EXERCISES As f (t. x2 )) (x1 − x2 ) ≥ (α − 1)(x1 − x2 )2 ∀x1 .6. x(1) = 0. Using Green’s function technique. y) = 0 w(t)x(t)y(t)dt 1 0 5.14) has a unique solution.14) is not solvable through Theorem 7.2 is applicable.6.7. ˙ ˙ Find L∗ and adjoint boundary conditions. Show that the differential operator Lx = − is formally self adjoint if the inner product is defined by 1 (x. x) is a strongly monotone function: (f (t. x(1) = 0. ˙ ˙ 1 (a0 (t)x) + a2 (t)x. Prove that the solution x(t) = g(t. (7. x) = γx + sin x. x → f (t. (7.6. x1 ) − f (t.7. show that the solution x(t) of the boundary value problem d2 x + tx = 1. Find the modified symmetric Green’s function for the system L = d2 . √ Remark 7. Find the Green’s function for the operator L given by (a) L = (t + 1) (b) L = d2 with boundary conditions x(0) = 0 = x(1). it follows that Eq. x(0) = x(0) = 0 ˙ dt2 . 6. dt2 −1 ≤ t ≤ 1 with boundary conditions x(−1) = x(1) and x(−1) = x(1).

Joshi. x(0) = x(1). Halsted Press. Parts I and II. Vol I and II.. 1968 . x(1) = 0 ˙ dt2 d2 x + x = sin t. Bose. BOUNDARY VALUE PROBLEMS satisfies the Volterra integral equation t x(t) = 0 s(s − t)x(s)ds + 1 2π 2 Prove that the converse of the above is also true. Linear Operators. Dunford and J. McGraw Hill Book Co. The Macmillan Co. K.. Inc.. Joshi and R. 1968 [5 ] I. 1966 [4 ] M. Levinson. Some Topic in Nonlinear Functional Analysis. x(0) = 1 ˙ (b) dt2 d2 x (c) + x = et . 7. Boundary Vlaue Problems of Mathematical Physics. Tata McGraw Hill Publ. K. co. 1984 [3 ] N. Interscinece Publ. x(0) = 0. Schwartz. x(0) = x(1) ˙ ˙ dt2 (a) 8. A. Obtain the solution of the following boundary value problems through eigenfunction expansion. Coddington and N. Bose and M. C. x(0) = 0 = x(1) dt2 d2 x (b) = x + sin x.242 CHAPTER 7. 1955 [2 ] R... Theory of Ordinary Differential Equations. Methods of Mathematical Physics.. x(1) = 2. Compare this approach with that of the initial value problem. Ltd. Stakgold.T. x(0) = 0 = x(1) ˙ ˙ dt2 (a) References [1 ] E. d2 x + x = t. Compute the first few Picard iterates for the following nonlinear boundary value problems d2 x = x + x3 . C.

¯ ¯ x(tf ) = xf ¯ ¯ 243 (8.1. The typical controllability problem involves the determination of the control vector u(t) such that the state vector x(t) has the desired properties.1.1 Controllability d¯ x = A(t)¯(t) + B(t)¯(t) x u dt x(t) = x0 ¯ ¯ We shall be interested in the linear system of the form (8. (8. modelled by ODE.Chapter 8 Control Theory Of Linear Systems Control theory deals with the maneuver of the state or trajectory of the system.1a) (8.1) satisfies ¯ ¯ the condition x(t0 ) = x0 . In this chapter we show how operator theoretic approach in function spaces is useful to discuss the above mentioned properties.1.1. Some of the important properties of such systems are controllability. (8.1b) where A(t) is n × n matrix. ¯ ¯ Definition 8.1. We substantiate our theory by concrete examples.1. there exist a ¯ ¯ control u(t) so that the correspondoing trajectory x(t) of Eq. 8.2) . observability and optimality.1) is said to be controllable if given any initial state x0 and any final state xf in n .1 The linear system given by Eq. u(t) ¯ ¯ ¯ is called control or input vector and x(t) the corresponding trajectory or state ¯ of the system. We assume ¯ ¯ that the entries of the matrices A(t). B(t) are continuous so that the above system has a unique solution x(t) for a given input u(t) . B(t) is n × m matrix. u(t) ∈ m and x(t) ∈ n .

τ )B(τ )¯ (τ )dτ u (8. τ )B(τ )¯ (τ )dτ = zf = xf − φ(tf . t0 )¯0 u ¯ ¯ x L¯ = zf u ¯ This is equivalent to solving . in view of (8. τ ∈ [t0 . ¯ Assume that φ(t.1.1. CONTROL THEORY OF LINEAR SYSTEMS The control u(t) said to steer the trajectory from the initial state x 0 to the final ¯ ¯ state xf . (8. it follows that φ(t.1) is equivalent to finding u(t) such that ¯ tf xf = x(tf ) = φ(tf . τ ) ≤ M for all t. The controllability x ¯ problem of Eq. Eq.1. tf ]. t0 )¯0 ] x x Proof : Let (¯0 . xf ) be any pair of vectors in n × n . tf ]. t0 )¯0 + ¯ x t0 φ(t. tf ) dτ (8.244 CHAPTER 8. So controllability of Eq. τ )B(τ )B (τ )φ (t.6) is nonsingular.1) is controllable iff the controllability Grammian tf W (t0 . tf x(tf ) − φ(tf . the controllability problem reduces to showing that operator L is surjective. τ )B(τ )¯ (τ )dτ u Equivalently. τ ) is continuous.1.5) [L¯] = u t0 φ(tf . (8. τ ) → φ(t. τ )B(τ )¯ (τ )dτ u Then. tf ) by (8. t) [W (t0 . Then by the variation of parameter formula.1. tf )] ¯ −1 [¯f − φ(tf . A control u(t) steering the system from the initial state x 0 to ¯ ¯ the final state xf is given by ¯ u(t) = B (t)φ (tf .1.3) As (t.1. t0 )¯0 = ¯ x φ(tf .1.4) Let us define a linear operator L : X = L2 ([t0 .1.1) reduces to the determination of u ∈ X such that ¯ tf t0 φ(tf . τ )B(τ )¯ (τ )dτ u t0 m (8.1) is equivalent to the following integral equation t x(t) = φ(t. tf ) = t0 φ(tf . Theorem 8. (8. t0 )¯0 + ¯ x ¯ t0 φ(tf . (8.1.4).1 The system given by Eq. t0 ) is the transition matrix of the above system.

where L is given by Eq.1. L is onto. which is proved as under. This is determined as follows. R(L) = n . let the Eq. t)¯ ¯ (8. by Theorem 2.7) To claim that L is onto. Equiva⊥ ⊥ lently.1. Thus N (LL∗ ) = {¯ ¯ 0. (8.1. t) [W (t0 . ¯ ¯ Conversely. t) dt α ¯ = W (t0 .8) If Eq. For u ∈ X and α ∈ n . α) u ¯ n = = α.8. (8.8) is solvable. Let L∗ : n → X be its adjoint. it is sufficient to prove that LL∗ : n → n is onto (equivalently nonsingular). Assume that LL∗ α = 0. ¯ ¯ Computing LL∗ we get [LL∗ ](¯ ) = L B (t)φ (tf . CONTROLLABILITY 245 in the space X for a given zf . This implies LL is also 1-1.2. τ ) is bounded. find α ∈ n such that ¯ ¯ LL∗ α = zf ¯ ¯ (8. t)¯ α α tf = t0 φ(tf . tf ) is nonsingular. tf )¯ α If the controllability Grammian W (t0 . L∗ α = ¯ L∗ 1-1 implies that α = 0. N (L∗ ) = [R(L)] = [ n ] = {0}. t)B(t)B (t)φ (tf . tf )] [¯f − φ(tf .1.5). we have ¯ ¯ tf (L¯.1. ¯ We first note that L is a bounded linear operator as φ(t. t0 )¯0 ] x x which will steer the state from the initial state x0 to the final state xf .1. tf )] −1 zf ¯ −1 = B (t)φ (tf . That is. α) = (L∗ α. This gives 0 = (LL∗ α. ¯ t0 tf φ(tf . τ )B(τ )¯ (τ )dτ u n (α. then a control u which solves L¯ = zf is given by ¯ u ¯ u = L∗ α. L∗ α) = L∗ α 2 and hence ¯ ¯ ¯ ¯ 0}. τ )¯ . ¯ . ∗ ∗ Hence L is 1-1. That is.1.1) be controllable. τ )B(τ )¯(τ )) ¯ u t0 tf n dτ dτ = t0 B (τ )φ (tf . given any zf ∈ n . Then. φ(tf . we have a control u(t) ¯ given by u(t) = L∗ α ¯ = L∗ (LL∗ )−1 zf ¯ = L∗ [W (t0 . L∗ α)X u ¯ This implies that α [L∗ α](t) = B (t)φ (tf . (8. u(τ ) α ¯ n = (¯.

t)] = A(t)W (t0 . (ii) W (t0 . → n is 1-1 and hence nonsingular . τ )dτ = B(t)B (t) + t0 t + t0 φ(tf . τ )B(τ )B (τ )φ (tf .1. This proves The controllability Grammian W (t0 . tf )φ (tf . α) α ¯ n = t0 tf φ(tf . τ )B(τ )B (τ )φ (t. τ )B(τ )B (τ )φ (tf . . τ )¯ α t0 2 n dτ = ≥ 0 for all α ∈ ¯ n . Further. Theorem 8. tf ). τ )B(τ )B (τ )φ(t. τ )dτ t0 t = B(t)B (t) + t0 t d φ(tf .2 (i) W (t0 . t) Proof : We have tf W (t0 . t)A (t) + B(t)B (t) dt W (t0 . t0 ) = 0 (iii) W (t0 . t) + W (t0 . tf ) = W (t0 . τ )B(τ )B (τ )φ (t. t) + W (t0 . tf ) is symmetric and positive semidefinite. τ )B(τ )B (τ )φ(t. τ )A (t)dτ = A(t)W (t0 . tf ) = t0 φ(tf . τ ) dτ (i) It is clear that W (t0 . t0 ) = 0 follows from the definition. α α ¯ B (τ )φ (tf . τ )¯ . τ ) dτ dt A(t)φ(tf . tf ) = LL∗ : the result. t) + φ(tf . tf )¯ .246 CHAPTER 8. tf ) = W (t0 . t)) = dt d dt t φ(tf . (ii) Using Leibnitz rule. tf ) has some intersting properties. t)W (t. CONTROL THEORY OF LINEAR SYSTEMS n Thus W (t0 . t)A (t) + B(t)B (t) Bounadry condition W (t0 . tf ) satisfies functional equation W (t0 . which we now describe. tf ) satisfies the linear differential equation d [W (t0 . we get d (W (t0 . tf (W (t0 .

11) in a Taylor series arround t = tf . This is a very tedious task. t) Remark 8. An−1 B The next theorem. Proof : Assume the rank of C = n.10b) (8. tf ) = t0 φ(tf . (8.1. prove that N (W (t0 . one needs to verify the invertibility of the Grammian matrix W (t0 . CONTROLLABILITY (iii) To get the functional equation we express W (t0 . tf ). if A(t) and B(t) are time invariant matrics A.11) Expanding the LHS of Eq. t) = W (t0 . (8.1 implies that to check the controllability of the linear system given by Eq. (8. t)W (t.1.1. τ )dτ φ (tf .1.1 Theorem 8. t) t φ(t.10a) (8. in this direction. Theorem 8.3 The linear autonomous system d¯ x = A¯(t) + B u(t) x ¯ dt x(t0 ) = x0 ¯ ¯ (8. then the controllability of the linear system is obtained in terms of the rank of the following controllability matrix C = B. ¯ α implies that 0 = W (t0 . tf )¯ α 2 tf = t0 tf B φ (tf . τ )dτ tf + t φ(tf . tf )φ (tf .1).9) is controllable iff the controllability matrix C given by Eq.1. · · · .1.1.9) has rank n.1.8. tf )¯ = 0. τ )B(τ )B (τ )φ (tf .1.1. τ )B(τ )B (τ )φ (tf . τ )¯ 2 dτ α B eA (tf −τ ) = t0 α 2 dτ ¯ This implies that B eA (tf −τ ) α = 0 for all t ∈ [t0 .1. tf )) = {¯ Let α ∈ n be such that W (t0 . However. tf ] ¯ (8. is due to Kalman [4]. It follows that N (C ) = {¯ We shall 0}. we get that 0 = B α = B (A )¯ = B (A )2 α = · · · B (A )n−1 α ¯ α ¯ ¯ . t) + φ(tf . t) + φ(tf . τ )B(τ )B (τ )φ (t. B. This 0}. AB. tf ) as follows t 247 W (t0 . τ )dτ tf = W (t0 .

τ )B(τ ) k=0 n−1 αk (tf − τ )B (AT )k α dτ ¯ (8.1. tf )¯ is given by α tf n−1 W (t0 .1. where  0 1 0 0 −2w  0 0 0 0     3w2  A=   0   0   0 2w     0 1    0 0       1 0     B=    0 0      0 1 . Conversely.1. let N (W (t0 . τ ) = αk (tf − τ )Ak and hence W (t0 . This is given by the control system of the form given by Eq.248 CHAPTER 8. Assume that α ∈ N (C ). We have α W (t0 .1. This implies ¯ Example 8.10).1. tf )) = {¯ We shall show that N (C ) = {¯ 0}. τ )BB φ (t.1 Let us recall Example 2.5 which gives the linearized motion of a satellite of unit mass orbiting around the earth.13) In view of Eq.12) φ(tf . tf )) = {0}. CONTROL THEORY OF LINEAR SYSTEMS This implies that α ∈ N (C ) = {¯ ¯ 0}. This implies that ¯ ¯ B (A )k α = 0. (8. tf )¯ = t0 tf 1≤k ≤n−1 (8. τ ) = eA(tf −τ ) . it follows by Cayley-Hamilton theorem (refer problem number 4 in Section 5. that there exist functions αk (tf − τ ) such that n−1 k=0 φ(tf .12).3.5). ¯ This proves the theorem. we have ¯ that α ∈ N (W (t0 . 0}. τ )¯ dτ α As φ(tf . tf )¯ = α t0 φ(tf . k=0 αk (tf − τ )B (A )k α = 0. (8.

8. A2 B1 . . is given by  0 1 0 −w 2 C1 has rank 3. A2 B. A3 B2 ]. C = B. It is given by  0 0 1 0 2w 1 0 0 −w 2 249 2w 0 0 −4w 2    1 0 0  C=   0 0 0   0 1 −2w −w 2 0 0 2w3 0 3 −2w 0   −2w     2  −4w   0  One can verify that rank of C is 4 and hence the linearized motion of the satellite is controllable. A3 B1 ]. So. A3 B . as  0 0 2w 0 0 −4w2 0     0 2w  C2 =    0 1   1 0 C2 has rank 4.    1 0  C1 =    0 0   0 −2w −w2 0 0 2w3           and this gives control- −2w 0 On the other hand u1 = 0 reduces B to B2 = [0 0 0 1] lability matrix C2 = [B2 . the above analysis means that we can maneuver the satellite just with radial rocket thrust. CONTROLLABILITY For this system. one can easily compute the controllability matrix. In terms of practical importance of satellite in motion. we see that the loss of radial thrust does not destroy controllability where as loss of tangential thrust does.   −2w3     −4w2    0 Since u1 was radial thrust and u2 was trangential thrust. AB. What happens when one of the controls or thrusts becomes inoperative ? For this purpose set u2 = 0 and hence B reduces to B1 = [0 1 0 0] . AB1 .1. AB2 . A2 B2 . the controllability matrix C1 = [B1 . It is intersting to ask the following question.

3b) → X = L2 ([t0 .2.1b) Consider the input-output system of the form The questions concerning observability relate to the problem of determining the values of the state vector x(t). t0 )¯( t0 ) + y1 (t) ¯ x ¯ where y1 (t) is known quantity of the form ¯ tf (8. The following theorem gives the observability of the system given by Eq.3a) (8.2. from Eq. tf ] of time. in terms of the nonsingularity of the matrix M (t0 . Let φ(t.2.250 CHAPTER 8. (8.3) is observable on I = [t 0 .4) [L¯0 ](t) = C(t)φ(t. tf ) which is defined by tf M (t0 . (8. tf ) = t0 φ (t. t0 )dt (8. tf ].1a) (8. A(t) ∈ n×n . (8. based on the output y (t). n ) as (8. Theorem 8.3). (8. tf ). t0 )¯0 = H(t)¯0 x x x where H : t → H(t) is a matrix function which is continuous in t.2. then. Observability of the system Eq. t0 )C (t)C(t)φ(t.2 Observability x d¯ = A(t)¯(t) + B(t)¯(t) x x dt y (t) = C(t)¯(t) ¯ x (8. we need only the homogeneous ¯ ¯ system d¯ x = A¯(t) x dt y(t) = C(t)¯(t) ¯ x in place of Eq.2.2.3).5) .2. Then the output vector y(t) can be expresssed as y (t) = C(t)φ(t. As in the previous section.2. t0 ) be the transition matrix of the above ¯ system. (8.2.3).2. B(t) ∈ n×m are assumed to be continuous functions of t.1 We shall say that Eq. τ )B(τ )u(τ )dτ Thus.2) it follows that if we are concerned about the determination of x(t0 ). Definition 8.2.2.2) y1 (t) = t0 C(t)φ(t.2.2. tf ] if y (t) = 0 on I implies that x(t) = 0 on I.1 For the homogeneous system given by Eq. it is possible to determine the initial state x(t0 ) within an additive constant vector which lies ¯ with null space of M (t0 . knowing only the output vector y(t) over some ¯ ¯ interval I = [t0 . reduces to proving that L is one-one.1). CONTROL THEORY OF LINEAR SYSTEMS 8. ¯ ¯ We define a linear operator L : n (8. (8.2.

if L is one-one then so is L∗ L and hence R(L∗ L) = [N (L∗ L)]⊥ = n .6) by L∗ .2.3) is observable.4).8. Proof : If the output vector y (t) is known.2. x(t0 ) is deter¯ mined up to an additive constant lying in the null space of L∗ L = M (t0 .2.2. (ii) M (t0 . t0 )M (t. ¯ (8. (8. t0 )C (t)y(t)dt t0 (8. (8. t1 ) − M (t.5) is called the observability Grammian. That is. tf ) = M (t0 . Conversely.2. tf ). t0 )¯(t0 ) for ¯ ¯ x x(t0 ) is equivalent to solving the operator equation ¯ L¯(t0 ) = y x ¯ (8. t1 )φ(t.7) is given by [L∗ u] = tf φ (t. t1 ) = 0 (iii) M (t0 . tf ) = L∗ L. tf ) is nonsingular. If RHS of Eq. M (t0 . tf ) defined by Eq. t) + φ (t.8) Observe that M (t0 . tf ) satisfies the following matrix differential equation d [M (t.9) . For such an observable system.2. then it is clear that L is one-one and hence the system given by Eq.2.7) is known.2. tf ) is nonsingular. (8.2.2. Eq. tf ) satisfies the following properties (i) M (t0 . tf )] ¯ This proves the theorem. t0 ) −1 tf φ (t. solving y (t) = C(t)φ(t. This implies that L∗ L is nonsingular. t1 )] = −A (t)M (t.6) where L is defined by Eq. t0 )C (t)u(t)dt t0 (8.2.2. OBSERVABILITY 251 Hence x(t0 ) is uniquely determined if M (t0 . If L∗ L is invertible. t1 )A(t) − C (t)C(t) dt M (t1 .3) is observable ifff M (t0 . Theorem 8.2 The observability Grammian M (t0 . (8. the initial state x(t0 ) is given by ¯ x(t0 ) = (L∗ L)−1 L∗ y = [M (t0 . we get [L∗ L] x(t0 ) = L∗ y ¯ ¯ where L∗ : X → n (8. tf ) is symmetric and positive semidefinite. (8. Premultiplying Eq. tf ) satisfies the functional equation M (t0 .

10) is of rank n.10) Theorem 8. CA. y1 = r. CA2 . So. We assume that we can measure only the distance r of the satellite from the centre of the force and the angle θ. x3 .3) is observable iff the observable matrix O given by Eq. x2 . the observability matrix O is given by O = [C.2.1.1. the defining state-output equation of the satellite is given by d¯ x = A¯(t) x dt y (t) = Cx(t) ¯ where  0 1 0 0     3w2  A=   0   0  x = (x1 .2. y2 ). Let us denote by O the observability matrix O = [C. So.2. See Brocket [1] for more details. We also have the following theorem giving the necessary and sufficient contition which is easily verifiable for observability. .1. CONTROL THEORY OF LINEAR SYSTEMS The proof of the above theorem follows along the lines of the proof of the Theorem 8.2. y = (y1 . · · · .2. CA.2. CA3 ]  1 0 0 0  0 0 1 0   0 1 0 0   0 0 0 1 =   3w2 0 0 2w   0 −2w 0 0   0 −w2 0 0 −6w3 0 0 −4w2               0 0 2w     0 0 1    −2w 0 0  1 0 0 0  C= 0 0 1 0 Rank of C is 4 and hence the above state -output system is observable. CAn−1 ] (8.3 The system given by Eq.252 CHAPTER 8. (8. (8. consider the satellite orbiting around the Earth. x4 ). Example 8.1 As a continuation of Example 8. y2 = θ being the radial and angle ¯ ¯ measuments.

1) where x ∈ X is the state of the dynamical system ¯ d¯ x dt x(t0 ) ¯ = A(t)¯(t) + B(t)¯(t) x u = x0 ¯ (8.8.3. if y1 is not measured.2a) (8. v (t)) + (R¯(t). C1 A.     =      0 −4w2 8. (i) Unconstrained optimal control problem: Find the control u∗ ∈ U = L2 ([t0 . x(t))] dt u ¯ x ¯ (8. However. C2 A3 ]  0 0 0 −6w3 0 0 −2w 0 1 0 0 0 1 0           which is of rank 4. Thus.3 Optimal Control In this section we shall be concerned with the optimization problem of two types. tf ]. ¯ tional J(u) which is given by tf m ) that minimizes the cost func- J(¯) = u t0 [(¯(t). the state of the satellite is known from angle measurements alone but this is not so for radial measurements. we might be tempted to measure y1 only not y2 . C1 A2 .3.3. C1 A3 ] =    3w2   0 = [0 0 1 0 0] and = [C2 .2b) . OPTIMAL CONTROL 253 To minimize the measurement. C2 A2 . C2 A. This gives C1 = [1 0 0 0 0] and  1 0 1 0 −w2 0 0 0   0     0 2w    0 0  O1 which is of rank 3. we get C2 O2    0  = [C1 .3.

1) from the initial state u ¯ ¯ x0 to the final state xf }. Using the transition matrix φ(t. CONTROL THEORY OF LINEAR SYSTEMS corresponding to the given control u(t). t x(t) = φ(t. we obtain J (¯) = u + K ∗ T K u + K ∗ T K y0 u ¯ ¯ ¯ and J (u) = I + K ∗ T K ∀ u ∈ U ¯ (8. t [K u](t) = ¯ t0 φ(t. u)U + (T x. R is a positive definite symmetric ¯ matrix and X = L2 ([t0 .3. n ).3.3.8) . (8. we obtain the state x(t) of the linear ¯ system given by Eq.3. t0 )¯0 .4) We now define a bounded linear operator K from U to X as follows.3. Using the concept of the derivative of a functional (refer section 2. ¯ u where J(¯) and Uc are given as follows u tf J(¯) = u t0 (¯(t).3. determined by Eq.3.2) through the integral equation.3). (8. the cost functional J on U .6) (8. (8. (ii) Constrained optimal control system: Find the control u∗ that minimizes J(¯) over the constrained set Uc ⊆ U.3. τ )B(τ )u(τ )dτ (8.3) Uc = {¯ ∈ U : u steers the state x(t) of Eq. τ )B(τ )¯ (τ )dτ u Set y0 (t) = φ(t. t0 ). x ∈ X is given by ¯ x ¯ x = y0 + K u ¯ ¯ ¯ So. x)X u u ¯ ¯ ¯ where T is a positive definite operator induced by R on X.3. is given by J(¯) = (¯. tf ]. then in abstract notation.1).5) Theorem 8. u(t)) dt u ¯ (8.1 J is a strictly convex functional on the space U and hence it has a unique minimizer u∗ ∈ U of J(u) and is given by ¯ u∗ = − [I + K ∗ T K] ¯ −1 [K ∗ T K ∗y0 ] ¯ (8. ¯ ¯ We shall show that unconstrained problem is a problem of minimization of a strictly convex functional defined on the Hilbert space U .3. t0 )¯0 + ¯ x t0 φ(t.7) (8.254 CHAPTER 8.

3. (8. t)] BT Ku∗(τ )dτ = T K y0(t) ¯ dt d [φ(τ.2 that J is strictly and hence it follows that J has unique minimizer u∗ ∈ U and is given by the critical point of J (refer Joshi and Kannan [5] ¯ Chapter1).3.3. In the above equation we assume that the system Eq.10a) (8. t)B(T K u∗ )(τ )dτ + ¯ t tf (T K)(¯0 (τ ))dτ = 0 y (8. We now address ourself to second problem pertaining to constrained optimal control. we have u∗ + K ∗ T K u∗ = −K ∗ T K y0 ¯ ¯ ¯ Also. v ) = (¯. OPTIMAL CONTROL Proof : We have J (¯) u J (¯) u = u + K ∗ T K u + K ∗ T K y0 ¯ ¯ ¯ = I + K ∗T K 255 for all u ∈ U.3.3.10b) with u∗ (tf ) = 0. we get the equivalent integro-differential equation dt d¯∗ u = −BT Ku∗(t) − dt tf t φ(τ. Hence.8) which is equivalent to the equation u∗ (t) + ¯ t tf φ(τ. v )U v ¯ v ¯ ¯ ¯ = > v ¯ v ¯ 2 2 + (T K v.(8. t)] = −φ(τ.3. (8. As J (u) is positive definite.3.9) Differentiating Eq. We have the following theorem in this direction. gives u∗ = −[I + K ∗ T K]−1 (K ∗ T K y0) ¯ ¯ Remark 8. .3. t)A. v )U + (K ∗ T K v. t)ABT Ku∗ (τ )dτ = T K y0(t) ¯ (8. ¯ Note that J (u) : U → U is positive definite for all u ∈ U. as (J (u)¯. it follows by Remark 2. the invertibility of [I + K ∗ T K] (refer Joshi and Bose [3]).8.1 One can obtain integro-differential equation analog of Eq.3. K v )X ¯ ¯ (for v = 0) ¯ in view of positive definiteness of T .9) we get d¯∗ u − BT Ku∗(t) + dt As tf t d [φ(τ.2) is autonomous.

Solvability of Eq. L∗ α)U = (L¯. u∗ )U + u∗ u ¯ ¯ − u∗ ¯ 2 U 2 U . t)¯0 tf tf (8.1.3.2 Let W (t0 . (8. u∗ )U u ¯ u u u ¯ (8.14) where α solves Eq. Subtracting the above two equations.3.11) (8. Eq.3.3. tf )¯ = xf − φ(tf . [¯f − φ(tf . we have L¯ − L¯∗ = 0 u u This gives (L¯∗ .3. (8.3.2). α)U = (¯. if W (t0 . tf ) be the controllability Grammian of Eq.5).12) Then u∗ ∈ Uc . If u∗ is any control of the form ¯ u∗ (t) = B (t)φ (tf .3.256 CHAPTER 8. L∗ α)U = (¯. u − u∗ )U u ¯ ¯ ¯ = = u ¯ u ¯ 2 U 2 U − 2(¯. tf ) [¯f − φ(tf .3.14) gives ¯ ¯ u∗ ¯ 2 U = (¯∗ . u∗ )U u ¯ = (¯∗ . tf ) is nonsingular. τ )B(τ )¯(τ )dτ = L¯ u u t0 tf t0 φ(tf .3. then ¯ ¯ t0 u∗ (t) 2 dt ≤ u(t) 2 dt ¯ (8. (8.3. t0 )¯0 = ¯ x φ(tf . Hence.12). t0 )¯0 = ¯ x = xf − φ(tf . controllability vector belonging to Uc .11) is equivalent to defining u∗ = L∗ α.15) Eq. if u is any.15) gives 0 ≤ u − u∗ ¯ ¯ 2 U = (¯ − u∗ . then tf t0 u∗ (t) 2 dt = W −1 (t0 . τ )B(τ )u∗ (τ )dτ = L¯∗ u where L is defined by Eq. α)U = (L¯. t0 )] x0 . (8. Further.3. (8. α)U u ¯ u ¯ (8. (8.13) t0 Moreover.12) is equivalent to the ¯ solvability of the following equation LL∗ α = zf ¯ ¯ Also Eq.3. (8. t)¯ ¯ α where α satisfies the equation ¯ α ¯ x W (t0 . CONTROL THEORY OF LINEAR SYSTEMS Theorem 8. t0 )] x0 ¯ x ¯ x ¯ tf Proof : We have zf ¯ zf ¯ = xf − φ(tf .

3.16) i is the current out of the source. α) = (W (t0 . α) α ¯ x x x = (¯f − φ(tf . u(t)) u ¯ t0 m dt Further.3. OPTIMAL CONTROL That is u∗ ¯ or tf t0 2 U 257 ≤ u ¯ 2 U tf (¯∗ (t). u∗ (t)) u ¯ m dt ≤ (¯(t). if LL∗ = W (t0 .1. r i v c Figure 8.1 Consider the following electrical network.16) with the constraint that v(0) = v0 and v(tf ) = v1 .3. L∗ α) ¯ ¯ ¯ ¯ = (LL∗ α. We wish to find the current i(t) such that the voltage v(0) = v0 and v(tf ) = v1 and the energy disspipated is minimum. given by Figure 8.1: Electrical network Its equation is given by d (cv(t)) = i(t). t)¯0 ]) Example 8. tf ) is invertible. W −1 (t0 . t)x0 .3. we want to minimize 0 f i2 (t)dt subject to the output voltage satisfying the equation Eq. tf ) [¯f − φ(tf .3. The energy dissipated in the resister in the t time interval [0. tf ] is 0 f ri2 (t)dt . (8. c = Capacitance dt (8.3. Thus. t . tf )¯ . we get u∗ ¯ 2 = L∗ α ¯ 2 = (L∗ α.8.

That is (*) is controllable iff (**) is observable. .258 CHAPTER 8. tf ) = 2 and hence the c c optimal current i∗ (t) = (v1 − v0 ) and hence is constant on the interval [0.               ¯= b         0 0 0 . CONTROL THEORY OF LINEAR SYSTEMS tf For this problem. tf ]. 1                3. which ¯ will steer the trajectory from the initial state ¯ at t = 0 to the final state 0 (1. . (∗) (∗∗) . Find the control u∗ (t) with minimum L2 . 1) at time tf = 1. 8. . . Prove that the systems d¯ x = A(t)¯(t) + B u(t) x ¯ dt and d¯ z = −A (t)¯(t) z dt w(t) = B z (t) ¯ ¯ are dual of each other. B =  where A(t) =  1 0 −1 1. Show that the system is controllable. ··· ··· 1 −αn 0 0 0           A=                  . . 0 −α1 0 −α2 0 −α3 1 0 0 0 1 0 ··· ··· ··· .norm. Show that the linear system given by the pair {A.4 Exercises d¯ x = A(t)¯(t) + B u(t) x ¯ dt     1 et ¯ 0  . ¯ is controllable for all b} values of αi ’s arising the definition of A. . tf For more on control theory refer Brockett [1] and Gopal[3]. the controllability Grammian W (t0 . 2.  0 0 0 .

y(tf ) = y1 ).1. ˙2 ˙2 8. AB. Obtain controllable conditions analogous to Theorem 8. 6. b 5. Show that. if the time invariant system d¯ x = A(t)¯(t) + B u(t) x ¯ dt is controllable.1 for the following matrix differential equation dX(t) = A(t)X(t) + X(t)B(t) + C(t)U (t)D(t) dt 7. Show that the differential equation dy d2 y + u(t) + y(t) = 0 2 dt dt is controllable ( in the sense that given any y0 . y1 .4. Consider the system d¯ x = g(t) [A¯(t) + B u(t)] x ¯ dt with g continuous and bounded (0 < α ≤ g(t) ≤ β < ∞) . which minimizes the functional ¯ tf φ(¯) = u t0 (¯(t).8. then the above system is controllable. An−1 B has rank n. 259 Show that if C = B. ˙ ˙ ˙ ˙ 2 provided y0 + y0 and y1 + y1 are nonzero. y(t0 ) = y0 . then there exists a matrix C such that d¯ x b = (A + BC)¯(t) + ¯i v(t) x dt is also controllable where ¯i is any nonzero column vector of B. y0 and y1 . y(tf ) = y1 . u(t)) dt u ¯ for steering the initial state x0 to the final state xf . EXERCISES 4. · · · . there exists ˙ ˙ a control u(t) such that y(t0 ) = y0 . Show that there exists an initial state for the adjoint system d¯ p = −A (t)¯(t) p dt such that the control u(t). is given by ¯ ¯ u(t) = −B (t)¯(t) ¯ p .

. Joshi and Ramendra K. Modern Control System Theory.260 CHAPTER 8. 1985 [4 ] R. Inc. Mathematical Description of Linear Dynamical Systems.. 1970 [2 ]M. Bose. Brockett. Kalman. pp152-192. Joshi and Kannan M. Finite Dimensional Linear Systems. John Wiley and Sons. Narosa Publishing House. Halsted Press. CONTROL THEORY OF LINEAR SYSTEMS References [1 ] Roger W. 1989 [3 ] Mohan C. Some topic in Nonlinear Functional Analysis. E. SIAM J Control. Optimization -Theory and Practice. Gopal. Pvt. 1963 [5 ] Mohan C. Moudgalya. Ltd. 2004 . Vol I. Wiley Eastern Ltd.

148 asymptotically stable. 9. 149–151. 132. 96. 230–232. 176 comparision test. 236 Boundary conditions. 34 Eigenfunction Expansion. 112. 259 Analytic Coefficients. 216. 147. 27 differential operator. 125 eigenvalues. 104 adjoint Green’s function. 67. 127. 69. 111. 21. 156. 153– 155. 214. 39–41. 237. 118.Index Abel’s formula. 12. 181 Ascoli-Arzela theorem. 20. 56. 26 fixed point. 154 Euler’s system. 71 Dirac-delta function. 215. 1. 249 convex functional. 142. 11. 243 controllability. 97 asymptotic convergence. 50. 254 differential equation. 233 eigenfunction expansion. 6. 244–246. 237 Boundary value problem. 240. 55. 137. 219–221. 173 autonomous system. 139. 63. 21. 71. 73 direction field. 76. 97. 237 Boundary condition. 20. 9. 207–209. 71. 195. 34. 156. 207. 212. 171. 54. 240 Dirac-Delta Function. 3. 239 . 256. 217. 244. 219 adjoint operator. 63–65. 212. 196. 234. 226. 168 equilibrium points. 207. 56. 258 controllability matrix. 120–123. 137–139. 216–222. 177 continuous linear operators. 96. 65 Bessel’s function. 48. 144. 254 bounded solutions. 222 boundary conditions. 77. 209. 68. 200. 242 Eigenvalues. 210 adjoint system. 145 carrying capacity. 211. 115. 209–211. 237. 135 Banach space. 211. 142. 238. 214. 4. 168. 211 boundary value problem. 166. 234 eigenvectors. 131–133 First Order Equations. 168. 80. 2. 154. 48. 81. 4 Cauchy sequence. 97 Comparision test. 241 boundary value problems. 132. 129. 136. 239 compact set. 224. 45. 234–237. 225. 245. 68 Controllability. 75. 242 bounded linear operator. 55. 227. 26. 49 contraction mapping principle. 247 autonomous systems. 241 Boundary Value Problem. 248 compact operator. 247. 247. 97 261 Cayley-Hamilton theorem. 256 controllability Grammian. 39 Cauchy-Peano theorem. 171. 1.

201. 7 linear system. 193. 165. 175 initial condition. 244. 96. 253 orbit. 111. 76 Lyapunov function. 2. 49 Jordan canonical form. 123. 10. 240. 4. 219–223. 91. 112. 55. 75. 132. 66. 241 Gronwall’s inequality. 43. 62. 37. 106. 209. 56 orthonormal set. 212. 197. 248 neural network. 107. 22 observability. 88 mechanical oscillations. 24 neural solution. 163. 222–225. 166. 54. 144. 39. 67. 57. 62 non-autonomous system. 187 linear ODE. 164. 56. 125 metric space. 137 Generating function. 21. 25 Newton’s algorithm. 63 linearly dependent. 234. 142. 3. 187 Hilbert Schmidt kernel. 102. 41–43. 117. 147 Gronwall’s lemma. 19 orthogonal matrix. 196. 147 nonlinear boundary value problem. 38. 152 Initial Value Problem. 242. 70. 145. 187 Hermite Differential Equation. 47. 56. 226–231. 154. 233. 234. 185 Legendre differential equation. 120. 64. 211. 93. 9 Harmonic Oscillator. 150 generalized eigenvector. 239 nonlinear boundary value problems. 21. 62. 79 inner product space. 38–41. 238. 46 Green’s function. 23. 94. 23. 200 generating function. 68 numerical methods. 254 Hilbert spaces. 63. 250–252 optimization. 212.262 fixed point theorem. 94. 202 gradient. 107. 186. 229. 227 Lipschitz continuity. 13 harmonic oscillator. 101. 114. 61. 54 . 70 motion of a satellite. 20. 43. 234 integro-differential equation. 148 formal adjoint. 96 modified Green’s function. 2. 137 INDEX Legendre Differential Equation. 235 Parseval’s equality. 80. 6 isomorphism. 140. 150. 39. 38. 79. 63. 242 normed space. 207. 187 Hermite differential equation. 101– 107. 59 Gram-Schmidt orthonormalisation. 3. 79. 210 fundamental matrix. 146. 39. 146. 255 Isocline. 138– 140. 2. 237–239. 49. 149. 214– 217. 102 linearly independent. 162. 107. 254 integral operator. 140. 44. 50. 238. 73. 10. 231 monotone operator. 218. 84 locally integrable. 234. 95. 163 maximal monotone. 59. 22. 81. 41 integral equation. 123 generalized eigenvectors. 53. 21. 254. 98. 157. 168–171 Lyapunov Stability. 50 infinite series. 59. 107–109. 238 Hilbert space. 145. 143 growth rate. 69. 93. 129. 97 Floquet-Lyapunov theorem. 134. 60. 110. 80. 24. 258 linearization.

85 stable focus. 134. 244. 114. 106. 148. 244 vector field. 12. 72. 9–11 Ratio test. 154 Picard iterates. 154. 2. 113. 81–83. 125–127. 1. 100. 101. 158. 177 Schauder’s theorem. 222 263 . 141. 158 vector space. 188 Rodrigues formula. 121. 123. 238. 107–109. 41. 140. 114 perturbed system. 68 Schwartz inequality. 23. 73. 129. 156. 165. 3. 2. 155. 164. 176 regular singularity. 160 variation of parameter. 161 phase portraits.INDEX Peano-Baker series. 134. 254 unstable focus. 159. 148. 2. 48 singular distribution. 147 phase portrait. 73 trajectory. 120. 254 predator-prey system. 135. 138. 19. 233 test functions. 258 transition matrix. 56. 158 Sturm-Liouville system. 145. 140. 15. 8. 112–114. 146. 244. 5. 118. 74 solution curve. 8. 158. 243. 250. 160 stable node. 102. 11. 130. 73 Wronskian. 4. 112. 14. 134 Periodic Linear System. 197 Root test. 37–39. 108. 155. 1. 242 positive definite operator. 117. 150. 173 phase space.