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 interdisciplinary 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, proofreading 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 DiracDelta 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 CauchyPeano 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  SturmLiouville 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 1dimension, 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 nonautonomous 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 (PredatorPrey Model)
Predatorprey 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 nonexistent, 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 phaseplane 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
aλ
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 +
aλ
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
2π
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 predatorprey 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 predatorprey 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
+ω
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
+ω
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
2π
ω
and amplitude α. ω is called
its frequency.
We can easily draw plots of phasespace (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 =
√
4ω
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 =
(ω
2
−β
2
)
2
+ k
2
β
2
1/2
and
sinφ =
kβ
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
(ω
2
−β
2
)
→∞ as ω →β.
For k = 0, the response is given by
x
p
=
F
0
2ω
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: InputResponse 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)(
dθ
dt
2
)
2
= −
k
r
2
(t)
+u
1
(t)
¸
r
d
2
θ
dt
2
+ 2
dθ
dt
dr
dt
= u
2
(t)
This gives a pair of second order diﬀerential equations
d
2
r
dt
2
= r(t)(
dθ
dt
)
2
−
k
r
2
(t)
+u
1
(t) (1.3.12a)
d
2
θ
dt
2
= −
2
r(t)
dθ
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
+ω
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
(Π
i
l=j+1
a
l
)b
j
20CHAPTER1. INTRODUCTIONTOORDINARY DIFFERENTIAL EQUATIONS
A k
th
order diﬀerence equation in 1dimension 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 wellknown onestep 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 gridpoints) 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 socalled explicit, onestep 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 onestep numerical methods is RungeKutta 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 RungeKutta formula
x(t
i+1
) = x(t
i
) +h
m
¸
j=1
β
j
k
j
(1.5.9)
k
j
= f(t
i
+ρ
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 RungeKutta 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 RungeKutta 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 RungeKutta 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: RungeKutta 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), 9871000, 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 Diracdelta 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 nonempty 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
+α
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 ntuples, 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
αx
a
≤ x
b
≤ βx
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)[ ≤ xy (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
= [2y
n
−x
2
+ 2y
m
−x
2
+ −2x +y
n
+y
m

2
]
≤ [2y
n
−x
2
+ 2y
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
√
2π
,
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
2π
+
∞
¸
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 GramSchmidt 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
+α
2
x
2
) = α
1
Tx
1
+α
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 ≤ kx ∀ 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
[α
ij
[
¸
x
1
Thus T is a bounded linear operator on '
n
, equipped with 1norm. 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 ≤ kf ∀ 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 selfadjoint 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 semideﬁnite 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 viceversa.
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 2norm

¯
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 selfadjoint 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 selfadjoint 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 selfadjoint 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 selfadjoint 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 selfadjoint operator and hence
the eigenvectors of A form a basis.
Let
Aφ
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 selfadjoint 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 ≤ kx −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 selfadjoint 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 semideﬁnite. 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 1dimensional function φ(α) = f(d
(k)
+αd
(k)
) to get
α
(k)
Step 4 : Set x
(k+1)
= x
(k)
+α
(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
dα
[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
+α
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
 ≤ Ax
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
≤ Ax + 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
+α
12
x
2
+. . . +α
1n
xn = b
1
α
21
x
1
+α
22
x
2
+. . . +α
2n
xn = b
2
.
.
.
α
n1
x
1
+α
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
[α
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 ≤ αx − 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
) ≥ cx
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. DIRACDELTA FUNCTION 71
2.5 DiracDelta Function
During the course of our discussion on boundary value problems, we shall en
counter Diracdelta 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 Diracdelta 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. DIRACDELTA 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 Diracdelta 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 Diracdelta 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. DIRACDELTA 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
∞
−∞
[φ
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 1norm 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, 557566, 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, 341346, 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 supnorm 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) +bx
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
≤ mt
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 CauchyPeano 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 AscoliArzela 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 (AscoliArzela 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 (CauchyPeano 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 AscoliArzela 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 CauchyPeano 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, SpringerVerlag, 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 nvector 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 nunknowns 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
dΨ
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
dσ
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
dσ
1
+ (4.2.6)
Eq. (4.2.5) or Eq. (4.2.6) is called the PeanoBaker 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
dΨ
2
(t)
dt
=
dΨ
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
dΨ
2
(t)
dt
=
dΨ
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) =
dΨ
1
(t)
dt
Ψ
−1
1
(t) and
A
2
(t) =
dΨ
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)
dΨ
−1
(t)
dt
= 0
and hence
A(t) + Ψ(t)
dΨ
−1
(t)
dt
= 0
This gives
dΨ
−1
(t)
dt
= −Ψ
−1
(t)A(t)
112 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX
and hence
¸
dΨ
−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 nonhomogeneous 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 nonautonomous 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 PeanoBaker 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
dσ
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 (FloquetLyapunov). 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
dφ
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 nonhomogeneous 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 nonhomogeneous 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
+ω
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 nonhomogeneous 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
2π
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)
+α
2
¯ u
(2)
+ +α
n
¯ u
(n)
=
¯
0, α
i
∈ ' (4.4.2)
Operating by A we get
α
1
λ
1
¯ u
(1)
+α
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
(λ
1
−λ
n
)¯ u
(1)
+α
2
(λ
1
−λ
n
)¯ u
(2)
+ +α
n−1
(λ
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
+ω
2
x = F(t) (4.4.4)
k is the resistance and ω
2
=
λ
ma
.
This is equivalent to the following nonhomogenous 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
4ω
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
3ω
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
3ω
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 PeanoBaker 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 nonhomogeneous 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 wellknown theorem called CayleyHamilton
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
dτ
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
dτ
=
t
t0
q(τ)p(τ) exp
t
τ
q(s)ds
dτ
and hence
u(t) ≤ p(t) +
t
t0
q(τ)p(τ) exp
t
τ
q(s)ds
dτ
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
dτ
= p(t)
¸
1 −
t
t0
d
dτ
exp
t
τ
q(s)ds
dτ
= 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 nonhomogeneous
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) visavis 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 nonhomogeneous 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 nonautonomous
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 nonautonomous 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 nonautonomous 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 nonhomogeneous 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 nonhomogeneous equa
tion Eq. (5.1.17).
If the nonautonomous 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 FloquetLyapunov 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 Predatorprey 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
aλ
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
+ω
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 phaseplane 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 nonautonomous 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, nondecreasing
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, nondecreasing 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 nondecreasing 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 CayleyHamilton 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 nonautonomous 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
=
dφ
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) =
2µ
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
2µ
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
=
2µ
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
J1\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
J3\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 Diracdelta 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
(α
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 crosssection. 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 initialvalue 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 piecewisecontinuous 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 selfadjoint if L
∗
= L.
It is clear that L is formally selfadjoint if ˙ a
0
(t) = a
1
(t). So, a formally
selfadjoint 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 selfadjoint, 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 selfadjoint as an unbounded operator
if the boundary value problem given by Eq. (7.2.4) is selfadjoint 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 selfadjoint 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 selfadjoint. D
∗
= D.
Hence the boundary value problem is not selfadjoint.
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 selfadjoint 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 delfadjoint, then D
∗
= D. That is, the
boundary value problem is selfadjoint.
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 initialvalue 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 selfadjoint, 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 selfadjoint, 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 nonhomogeneous 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 nontrivial 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 nonhomogeneous 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 nontrivial 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 nontrivial solution of Lx = 0 satisfying boundary condition
at t = a and x
2
(t) be a nontrivial 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 nontrivial 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 nonzero 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 initialvalue 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 crosssection 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 nontrivial 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 nontrivial 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 nontrivial solution, (iii) will also have nontrivial 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 selfadjoint 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 nontrivial 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 selfadjoint 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 nontrivial 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 selfadjoint 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 nontrivial 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 selfadjoint
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 selfadjoint
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 nontrivial 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  STURMLIOUVILLE 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  SturmLiouville
System
Consider the selfadjoint 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 selfadjoint system
is called SturmLiouville 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 selfadjoint 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 selfadjoint as the system
Eq. (7.5.2) is selfadjoint. 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 selfadjoint 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  STURMLIOUVILLE 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
(λ
n
−λ)φ
n
(t)
Since (L −λI)x = f, we get
∞
¸
n=1
x
n
(λ
n
−λ)φ
n
(t) =
∞
¸
n=1
f
n
φ
n
(t)
That is
∞
¸
n=1
[f
n
−x
n
(λ
n
−λ)] φ
n
(t) = 0. Completeness of ¦φ
1
, φ
2
, ¦ implies
that f
n
−x
n
(λ
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
(λ
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 adjointsystem
−
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 nonzero 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 selfadjoint 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
dφ
dt
+λφ = 0
That is,
dφ
dt
+
1
t
dφ
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
≤ γx
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

≤ kγx
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
2π
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
dτ
=
t
f
t0
B
¯
(τ)φ
¯
(t
f
, τ)¯ α, ¯ u(τ)
J
n
dτ
= (¯ 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 11. This implies LL
∗
is also 11, 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
∗
11 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 11 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
)φ
¯
(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)
dτ
(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
dτ
=
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)
dτ
= 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
)φ
¯
(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
dτ
=
t
f
t0
B
¯
e
A
(t
f
−τ)
¯ α
2
dτ
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
dτ
As φ(t
f
, τ) = e
A(t
f
−τ)
, it follows by CayleyHamilton 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 inputoutput 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 oneone.
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 oneone and hence the system given
by Eq. (8.2.3) is observable.
Conversely, if L is oneone 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 stateoutput 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 integrodiﬀerential 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 integrodiﬀerential 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, pp152192, 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
AscoliArzela 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
CauchyPeano theorem, 96, 97
CayleyHamilton 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
DiracDelta Function, 71
Diracdelta 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
FloquetLyapunov 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
GramSchmidt 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
integrodiﬀerential 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
nonautonomous 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
PeanoBaker 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
predatorprey 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
SturmLiouville 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 interdisciplinary area involving of science and engineering.
In this connection. Our special thanks to Asmit Pany for type setting the manuscript in LATEX.Acknowledgements This book could not have been written without direct and indirect assistance from various institutional bodies within IIT Bombay. Mohan C. M K Ramaprasad and Soumya Mohanty. we would like to acknowlege the help rendred by Amiya K Pani. proofreading the manuscript etc. Secondly. Mumbai . 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. this author would like to thank his wife for patiently supporting him in implementing the project. Finally. Joshi Powai. we would like to acknowledge the ﬁnancial support received from the Curriculum Development Program of the Institute. Several colleagues and students helped us from time to time in many ways giving correct references. Firstly.
. . . . . . . 2. . . . . 2 Mathematical Preliminaries 2. . . . . . . . .1 Linear System . . . . . . .6 Techniques for Solving First Order Equations . .3 Lipschitz Continuity and Diﬀerentiability . . . . . . . . . . .2 Transition Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Dependence on Initial Condition . . . . . . . . . . . 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . 1. . . . . . . 79 . . 2. . . . . . . . . . . . . . . . . . . . . .1 Existence and Uniqueness Theorems 3. . . . . . . . . . . . 1. . . . . . . . . . . . . . . . . . . . . . . . .3 Higher Order ODE . . . . . .6 Exercises . . . . . . . . . . . 4. . . . . . . . . . . .5 Euler’s Linear System . . . . . . . . 3 Initial Value Problems 3. . .4 Diﬀerence Equations . . . . . . . . . . . . . . . . . . . . . . . 1. . . . . . . . . 4. . . . . . . . . . . . . 97 .4 Computation of Transition Matrix . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . .1 Finite and Inﬁnite Dimensional Spaces . . 101 101 106 114 118 131 133 4 Linear System And Transition Matrix 4. . . . . . . . . .4 Fixed Point Theorems and Monotone Operators 2. .5 DiracDelta Function .2 Linear and Nonlinear Operators . . . . . . . . . . . . . . . . . . .6 Exercises . References . . . . . . . . . . . . . . . . . 1 1 7 11 19 21 26 34 36 37 37 47 57 63 71 76 78 79 . . . . . . . . . . . . . . .4 CauchyPeano Existence Theorem . . . 100 . . . . .7 Exercises . . . . . . . 3. . . . . 1. . . . . . . . . . . 4. .2 The Maximal Interval of Existence . . . . . . . . . . . . . . 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 First Order Ordinary Diﬀerential Equations . . . . . . 96 . . . . . . . . . 89 . . . . . . . . . . . . . .2 Classiﬁcation of ODE . . . . . .Contents 1 Introduction To Ordinary Diﬀerential Equations 1. . . . . . . . . . . . . . . References . . .3 Periodic Linear System . . . . . . . . . . . 3. . . . . . . 1. . . . . . . . . . . . . .5 Exercises . . . 2. . . . . 1. . 93 . . . . . . . . . . . . . . . . 4. . . . . . . .5 Discretization . . . . . . . . . . . . . . . . .
. . . . . . . . . . . .Formal Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 5 Stability Analysis 5. . 7. . . . . . . . . . .3 Green’s Function . . . . . . . . . .1 Preliminaries . . Systems . . . . . . . . . . . . . . . . . . .6 Nonlinear Boundary Value Problem . . . . . . . . . 7. . . . . . . . . . . . . . . . . . . . . . . . . 6. . . . . . . . . . . . . . . . . 5. . . . . . . . . . 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. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .CONTENTS References . . . . . . . . . . .SturmLiouville System . . . . . . . References . . . . . . . . . .4 Lyapunov Stability . . . . . . . . . . . . . . . . .4 Construction of Green’s Function and Modiﬁed Green’s Function 7. . . . . . . . . .1 Controllability . . . . References . . . . . . . . . . . . . . . . . . . . . . . .3 Phase Portrait Analysis . . . . . . . . . . . . . . . . 6. . . 5. .5 Exercises .2 Stability Analysis . . .5 Exercises . . . . . 8 Control Theory Of Linear 8. . . . . . . . . . . . . . . . . . References . . . . . . . . . . . .4 Properties of Legendre. . . . .7 Exercises . . . . . . . . . . . . . . . . . . . . .1 Introduction . .4 Exercises . . . . . 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Observability . . . . . . . . . . . . . . . 7. . . . . 7. . . . . 7 Boundary Value Problems 7. . . . . . . . . . Hermite and Bessel Functions 6. . . . .2 Linear System with Analytic Coeﬃcients . . . . . . . . . . . . .1 Asymptotic Behaviour of Linear System 5. . . . . . 8. . . . . . . . . . . . . . . . . . . . . .3 Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . .3 Linear System with Regular Singularities . . . 7. . . . . . . . . . . 6.2 Adjoint of a Diﬀerential Operator . . . 8. . . . . . . . . . . . . . . . . . .5 Eigenfunction Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . 5. . . . . . . . . . . . . . . . . . . . . . . . . .
1.1). . . . direction ﬁeld. we recall standard methods of solving some classes of ﬁrst order ODE. 1. xn )) We shall refer to f as vector ﬁeld on Ω. The value x(t) ∈ Ω ⊂ n is sometimes referred as the state of the system described by the ordinary diﬀerential equation (ODE) . x). x2 .Eq. .1. . x) dt (1. 1 . . We also give. . x2 . along with appropriate examples of practical importance. . . Towards the end. f2 (x1 . that is. xn ) = (f1 (x1 . . I a subinterval of . xn ). An interesting concept of neural solution of ODE is also touched upon.(1.1) where f : I × Ω → n . xn ).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 .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 . . . A vector ﬁeld may also depend on an additional parameter t. f = f (t. . The main thrust of our study is the following diﬀerential equation d¯ x = f (t. The reader is also familiarized with diﬀerence equations and numerical solutions of ODE through the process of discretization. . x2 . fn (x1 . all related notions like orbit or trajectory. The set Ω is then called the state space or phase space. x2 . . isocline and solution curve of IVP.
(1.1(b)) Deﬁnition 1.2) Proof : If x(t) is a solution of Eq.1.1) with the solvability of the corresponding integral equation. x(t) is a ¯ solution the IVP . x(s))ds. x)).1.1.1). .2 A function x(t) is said to be a solution of the IVP . t] for a ﬁxed t. x0 ) for x(t). f (t.1) is given by dx (1. it is desirable to indicate the dependence of the solution x(t) on its initial value x0 . (1. (1. x(s))ds ¯ This gives x(t) − x(t0 ) = ¯ ¯ and hence we have Eq. At times. t0 .1. if there exists an interval J ⊂ I. x0 ) : t ∈ J} in the state space Ω. If t → f (t. Proposition 1.1) is tangent to the vector ﬁeld whereas the solution curve is tangent to the direction ﬁeld at any point.1. by direct integration on the interval ¯ [t0 .1. (1.1. (1. Deﬁnition 1.) is deﬁned on the entire line and x(t) satisﬁes ¯ Eq.1 Assume that f : I × Ω → n is continuous. t ∈ J ¯ (1. x(t)) dt x(t0 ) = x0 (1. Then we use the notation x(t.1). x(t.1.1. The following proposition gives the equivalence of solvability of IVP .Eq.3 The orbit or the trajectory of the ODE. (1. this solution is sometimes referred as a local solution.1.(1. containing t0 such that x(t) is diﬀerentiable on J with x(t) ∈ Ω for all t ∈ J and x(t) satisﬁes Eq.1.Eq.1). is the set {x(t.1. (1. x0 )) : t ∈ J} ⊂ I × Ω.4 The direction ﬁeld of ODE . (1. t0 .Eq.1) iﬀ x(t) is the solution of the integral equation ¯ t x(t) = x0 + ¯ ¯ t0 f (s.1 The Initial Value Problem (IVP) corresponding to Eq. (1.1.1) only on J. In Chapter ¯ 3. Whereas the solution curve is the set ¯ {(t. x(s))ds ¯ t0 .1).Eq.1.1) on . ¯ Deﬁnition 1. It is clear from the above deﬁnitions that the orbit of the ODE . is the vector ﬁeld (1.1.1. t f (s.2CHAPTER 1.Eq.(1. we get t t0 d¯ x ds t ds = t0 f (s. (1.1.Eq. we shall discuss the local and global solvability of IVP in detail.1). t0 .1.2).1. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Deﬁnition 1.1. ¯ As x(t) needs to satisfy Eq. then x(t) is said to be a global solution of the IVP.1(a)) = f (t.
If the initial population N (t0 ) = N0 .1.5 The set of points in Ω where f (t.1. is called the set of ¯ equilibrium points of ODE . x(t)) ¯ t0 Deﬁnition 1. as population grows suﬃciently large.1.1. If we assume the form of R(N ) as equal to a − cN (a. we get the following initial value problem dN = R(t0 )N (1. (1. respectively. t ∈ Although the above model may predict the population N (t) in the initial stages. it begins to interact with its environment and also other species and consequently growth rate R(t) diminishes.1.1. Example 1.1). ¯ ¯ t f (s. we get the following IVP dN dt N (t0 ) = N (a − cN ) = N0 (1. . c positive constant).3(a)) dt N (t0 ) = N0 (1.Eq.1.(1.1 (Population Dynamics Model) We investigate the variation of the population N (t) of a species in a ﬁxed time span.1.1.4b) This is known as logistic equation with a as the growth rate without environmental inﬂuences and c representing the eﬀect of increase population density.3(b)) The solution of the above IVP is given by N (t) = N0 exp(R(t − t0 )). then growth rate is constant and is given by.2) holds. An important index of such investigation is growth rate per unit time.1. (1. R(t) = R(t0 ) = b − d where b and d are birth and death rates. In fact.4a) (1. denoted by dN 1 (1.3) R(t) = N (t) dt If we only assume that the population of species changes due to birth and death. we realise that no population can grow exponentially for ever. It is clear that at these points the orbit or trajectory is a singleton. x(t) given by Eq. x) = 0. FIRST ORDER ORDINARY DIFFERENTIAL EQUATIONS 3 Conversely.2) is diﬀerentiable and ¯ diﬀerentiating this equation we get d d¯ x = dt dt Also we have x(t0 ) = x0 . x(s))ds = f (t. if Eq.1.1.
we can actually carry out the computation of a and C from the given data.4CHAPTER 1. C) between the known discrete population data Nd (tm ) and theoretical data N (tm ) : M φ(a. 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).3.5) The solution curve of the above population model is as under. C = a/c is called the carrying capacity of the environment. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS The equilibrium points are 0 and a/c.1.1). C=a\c 8 6 Nt 4 2 No 2 4 6 8 10 t Figure 1.1.1: 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. So the above solution has the representation N (t) = (1.1. C) = m=1 [N (tm ) − Nd (tm )]2 (1.6) By using one of the algorithms of the unconstrained optimization techniques (refer Algorith 2. .
a is any constant The solution passing through (t0 . x0 ) in (t.1.5 2 1.1.1.5 3 2.5 1 0.5 2 1 1 2 t Figure 1.7) −t2 ) 0 (1.8) The following graphics demonstrates the solution curve and the direction ﬁeld of the above ODE. xt 3.2: Solution curves x(t) . x) plane is given by x(t) = x0 e− 2 (t 1 2 1 2 5 (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 .1. FIRST ORDER ORDINARY DIFFERENTIAL EQUATIONS Example 1.
t0 .6CHAPTER 1. Thus the uniqueness theorem. 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. The family of isoclines is then the family of curves f (t.3 We wish to ﬁnd the direction ﬁeld of the following ODE dx dt = 2√t + x dx 1− dt 1+ Solving for dx .6 Isocline is a curve in t − x plane through the direction ﬁeld along which p = f (t. to be dealt with in the Chapter 3. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Figure 1. Deﬁnition 1.1. To sketch the direction ﬁeld for a given ODE in 1dimension. no two solution curves of the IVP can intersect.2.1. x0 ). a device called isocline can be used. will be of utmost importance to us.1) has a unique solution x(t. Example 1.1. 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 .3: Direction ﬁeld of the ODE If IVP corresponding to Eq.1. x) = p in t − x plane. As we have seen through Example 1.1. x0 ) passing through (t0 . x) is constant. (1.
4: Isoclines with solution curves 1. −1 ≤ p < 1 (1.33 1 p=0.5 3 2 p=0. . x) = A(t)x + b(t) (1.2. A linear ODE is called homogeneous if b(t) ≡ 0 and if A(t) = A (constant matrix).1.1.2. 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 = .1. So we get the following graphics.9) Thus isoclines are all lines of slope 1 with −1 ≤ p < 1.1. x0 ) is given by √ √ (x − t) + t + x = (x0 − t0 ) + t0 + x0 (1. x) dt f (t.1) The ODE is called linear if f has the form (1.2) where A(t) ∈ n×n and b(t) ∈ n for all t.2 1 1 2 3 4 t 2 3 p=1 p=0 Figure 1. Linear ODE will be discussed in detail in Chapter 4.2.2 Classiﬁcation of ODE dx = f (t. The solution of the ODE passing through (t0 .10) xt 4 p=0. then we call it a linear ODE with constant coeﬃcients.
(1.1) is called periodic if f (t + T.2.2. A nonautonomous ODE .1 dx = dt 0 1 −1 0 x+ 0 2 cos t f (t. Hence the trajectory x(t) of Eq.Eq. Because x is periodic. x(t + s) is a solution of Eq. That is x (t ± nT.2. x0 ) for all t0 ∈ I This implies that the solution is completely determined by the initial state x 0 at ¯ t = 0. x0 ) = x(t. x0 ) = x(t − t0 . we have the following theorem regarding the periodic solution of the periodic ODE. sin t + t cos t) However. x) has a period T and the solution x(t) has period ˙ S with S = T. Proof : Assume that f (t.2. Example 1. x) for some T > 0 (1. t0 .3) on the interval I = (a.3) = f (x) dt For autonomous ODE we have the following proposition (refer Mattheij and Molenaar [9]). x) . (1. x) = f (t. b − s). (1.1 If x(t) is a solution of Eq. That is f (t + S.4) The smallest of such T is called its period.2.2 If a solution of a periodic ODE is periodic.8CHAPTER 1.2.2. Proposition 1.2. (1. f is periodic with period 2π but its solution (to be discussed in Chapter 4) is given by x(t) = (t − sin t. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS The ODE . t0 .1) is called autonomous if f does not depend on t. it follows that the form of the solution is not aﬀected if we shift t0 to t0 ± nT. b(t) = (0. n ∈ ℵ (set of natural numbers). Then it has the same period as the vector ﬁeld. x) = Ax + b(t). x0 ) It is not necessary that a solution of periodic ODE is periodic as we see in the following example.3) on the interval (a − s. x) will also be ¯ periodic with period S as well T. (1. b). 2 cos t).2.(1. its derivative x(t) = f (t. Proposition 1. 0.4).2. t0 ± nT. x) = f (t. then for any s ∈ .2. From Eq.3) satisﬁes the following property x(t. That is ODE has the form dx (1.
we have. Thus dS = S(−k + λF ) dt For the growth rate of ﬁsh.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. For example. the growth rate of 1 dS sharks is modelled as −k without the ﬁsh population and −k + λF with S dt ﬁsh population F. x) = f (t. f (¯) = (F (a − cS). a contradiction. S = 0. S). F = 0. x) 9 We choose n such that 0 < S − nT < T. T periodicity of f implies that f (t + S − nT. S(−k + λF )) x The equilibrium points are given by ˆ ˆ ˆ ˆ F = k/λ.Volterra model for the predatorprey system dF = F (a − cS) (1. Thus dF = F (a − cS) dt Thus. So. what is called the famous Lotka . Hence T = S and the solution x(t) has ¯ the same period as the vector ﬁeld. we note that it will decrease with the existence of sharks and will ﬂourish on small plankton (ﬂoating organism in sea) without sharks.(1.2. However. we shall focus on sharks and small ﬁsh in sea. It implies that f is periodic with a period smaller than T. S = a/c. If the food resource of sharks (ﬁsh) is nonexistent. CLASSIFICATION OF ODE Also.1.2. We shall have the discussion of the phaseplane analysis in the Chapter 5.2.2. if we linearize the system given by Eq. sharks population exponentially decays and is increased with the existence of ﬁsh.4) around the equilibrium point 2 (1.2.4(b)) . Example 1.2 (PredatorPrey Model) Predatorprey model represents the interaction between two species in an environment.
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. We have 1 dS = −k + λF S dt This gives t S(t) ln = −k(t − t0 ) + λ F (τ )dτ (1.6b) 11 10.2.10CHAPTER 1. The following graphw ics depict the ﬁsh and shark population in t − F/S plane. (1.2.5 2 4 6 8 10 Ft 12 t St Figure 1. the solution is periodic with period 2π = 2π(ak)−1/2 . S = a/c. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS ˆ ˆ F = k/λ.5) The solution of the IVP corresponding to Eq.2.4) is obtained as follows.2.7) S(t0 t0 .3 for solution analysis of this linear system).2.3 for linearization) dx = Ax(t) dt ˆ ˆ x = (F − F . we get the following linear system (refer Section 2. (1.5 10 9.2. Thus.1: Fish and shark population with time The average population of the general predatorprey system given by Eq.6a) (1. S − S) kc 0 − λ A = aλ 0 c (1. (1.2.
2. (1. . × → is a given mapping. x1 (t) = x(t). We deﬁne a vector x(t) with components xi (t). n by xi (t) = x(i−1) (t). . . no matter what the solution trajectory is. x(n−1) (t)) (1.3 Higher Order ODE Assume that x(t) is a scalar valued n−times continuously diﬀerentiable function dk x on an interval I ⊂ . 2 ≤ i ≤ n (1.1) where g : I × × × . .3. dt We shall be interested in the following higher order ordinary diﬀerential equation x(n) (t) = g(t. S(tf ) = S(t0 ) and F (tf ) = F (t0 ) and hence Eq. That is.2) .2. x(t).8(b)) So in predatorprey system. It is also interesting to analyse the men’s inﬂuence on the above ecosystem if we ﬁsh both predator and prey. . c c 1.8(a)) S(τ )dτ = a/c t0 (1. HIGHER ORDER ODE 11 In view of the information obtained from the linearized model.3. Let us denote the derivative k by x(k) (t). Then. .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. .2. x(1) (t). the average population remains around the equilibrium point.1.3. we can assume that both S and F are periodic of period T = tf − f0 . we have dF dt dS dt This is equivalent to dF dt dS dt a = a − σ1 . i = 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). . .
. . (1.. .Eq. . (1. g(t. x3 . . + a0 (t)x(t) = b(t) (1. 0 0 (1.. 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 .3. x1 .3.1 (Mechanical Oscillations) A particle of mass m is attached by a spring to a ﬁxed point as given in the following diagram. 0. 0 −a0 1 0 0 −a1 0 1 0 −a2 . Example 1.3. .2) . b(t)) The matrix A(t) is called the companion matrix. x) by ¯ f (t. −an−1 b(t) = (0.1) is equivalent to the following ﬁrst order ODE dx = f (t. . . . . .12CHAPTER 1. .Eq. .3) where ai (t)(0 ≤ i ≤ n − 1) and b(t) are given functions. ... . ... x) = (x2 . . xn )) ¯ Then the higher order ODE . .4) A(t) = .3. .3. . Then in view of Eq.3) we have dx1 = x2 (t) dt dx2 = x3 (t) dt .3. . x) dt A linear higher order ODE has the form x(n) (t) + an−1 (t)x(n−1) (t) + . (1. . 1 . INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS and the vector ﬁeld f (t.
7) Eq. The external force applied to the particle is F (t). (1.3.1. we get the following second order equation modelling the spring problem d2 x dx λ +k + ω 2 x = F (t).3.3.3.9) (1. ω 2 = dt2 ma (1.7) 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 = 0.3.3.5) F = Equivalently. ω (1.8) = λ 0 dt − ma It is easy to see the solution of the above system is given by (refer Section 4.3. By equilibrium of forces we have mF = m By Hook’s Law.10) x(t) = −x0 ω sin ω(t − t0 ) + x0 cos ω(t − t0 ) ˙ .3. ω 2 = (1.1: 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. we get T = dx d2 x + T + mk dt2 dt λx and hence a λ dx d2 x + x+k dt2 ma dt (1.4) x(t) = x0 cos ω(t − t0 ) + x0 sin ω(t − t0 ). HIGHER ORDER ODE 13 F(t) a x m Figure 1.
5 0 0. ω is called ω its frequency.2: Phase space of the linear system 5 2 1 t 0 0 x’t 1 5 2 1 0.3. (1. We can easily draw plots of phasespace (normalized ω = 1) and also solution curve of the ODE . INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Equivalently.3.3.5 1 xt Figure 1. 2 1 1 0.11) .14CHAPTER 1.5 1 2 Figure 1.3.Eq.8).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. we have x(t) = α cos(ωt + β) 2π This is a simple harmonic motion with period and amplitude α.5 1 0.
5 1 10 15 20 25 Figure 1.4) x(t) −kt )(A cos bt + B sin bt) = exp( 2 √ 4ω 2 − k 2 b = 2 15 xt 1 0.4: Solution curve for Case2(i) The oscillation curves for the other cases are as under.3.1.3.5: Solution surve for case2(ii) .5 1 1.5 t 5 0. HIGHER ORDER ODE The solution curve x(t) is a damped oscillation given by (refer Section 4.3. (ii) (k 2 − 4ω 2 ) = 0 x(t) = A exp( −kt −kt ) + Bt exp( ) 2 2 xt 1 0.5 t 10 0.5 20 30 40 50 60 Figure 1.
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 1.3. F (t)) The solution x(t) is given by x = xp + xc where xc is the solution of the homogeneous equation with F (t) = 0.6: Solution curve for case2(iii) Case 3: Eﬀect of resistance with external force dx(t) = dt 0 −ω 2 1 −k x + F (t). We have seen that xc → 0 as t → ∞.16CHAPTER 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. 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 . 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 . F (t) = (0.
as we see in the following graph. (ω − β 2 ) For k = 0. 4 2 Input t 14 Response 2 2 4 6 8 10 12 Figure 1.3.3. 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 .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. As k → 0 we have F0 xp → 2 → ∞ as ω → β.1. 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.3.7: InputResponse curve Example 1.3.
then one can show that Eq.12) reduces to dx1 dt dx2 dt dx3 dt dx4 dt x3 ˙ x4 + ω + ωt.3.12b) ] If u1 = 0 = u2 .12) has a solution given by r(t) = σ.14) = f (¯. r = x2 .3. u(t) ∈ 2 x ¯ dt .3. x4 = σ(θ − ω) ˙ This gives r = x1 + σ. x(t) ∈ 4 .12a) (1. θ(t) = ωt (σ. u).3.3. x3 = σ(θ − ωt). 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)( (1.13) Eq. Equating forces in normal and tangential direction on the orbiting satellite.18CHAPTER 1. (1.3. (1.13) 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. (1. x2 = r.3. 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 . Make the following change of variables: ˙ x1 = r − σ. θ = ˙ So Eq. ω are constant and σ 3 ω 2 = k). 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.
x3 . f4 given by f1 (x1 . (1. Then the ﬁrst order diﬀerence equation has the form xi+1 = fi (xi ). i ∈ ℵ where Ai ∈ n×n and bi ∈ n . we can deﬁne orbit or trajectory and solution curve as discrete subsets of n and n+1 . DIFFERENCE EQUATIONS Here f is a vector function with components f1 . 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. u1 . u2 ) = x2 f1 (x1 . . tk .4. u1 . (1.14) in terms of the control vector u(t) ∈ 2 . (1. i ∈ ℵ Its solution is given by i−1 xi = (Πi−1 (aj ))x0 + j=0 j=0 (Πi l=j+1 al )bj . . the state vector x(t) may not depend on t continuously. (1. f3 . 1. x2 . Rather. x(t) takes values at discrete set of points {t1 . u1 .4. . x2 . . f2 . xi ∈ Ω. x(t) ∈ 4 of linearized equation corresponding to Eq. x2 . Ω ⊆ n .1) is a constant solution x such that x = fi (x). x4 . i ∈ ℵ Eq.1(b)) is called the IVP corresponding to a diﬀerence equation. (1. u2 ) = (x1 + σ)( f1 (x1 . · · · } In such a situation we use diﬀerence quotients instead of diﬀerential quotients and that leads to diﬀerence equations. . . u2 ) = x4 f1 (x1 . A stationary point or equilibrium point of the diﬀerence equation Eq. x3 . x2 . i = 1. As for ODE.4 Diﬀerence Equations In some models (as we shall see subsequently).1(a)) If in addition x0 = z 0 (a given vector in n ) (1.4. x4 .1(b)) then Eq. x3 .1) is called linear if it has the form xi+1 (t) = Ai xi + bi . For the scalar case n = 1. respectively in an analogous way.4.1.4. ¯ (1.4. t2 . x3 .1(a)) . x4 . 2 . x4 .4. we have the linear diﬀerence equation xi+1 = ai xi + bi . u1 . Suppose there exists a sequence of vector ﬁelds fi (x) : Ω → n .Eq.3.
. c¯ ]. . (1. . c2 . . i (1. with Ai = A for all i. bi = .4. . .5) If the initial value is x0 then d = [d1 .4.3) is given by n xm = i=1 di λm ci . . . . i ≥ 0 where xi .4. .3) we get (A − rI)c = 0 (1. m = 1.4. . then the general solution of Eq. cn . Plugging this representation in Eq. A linear k th order diﬀerence equation is given by k xi+1 = j=1 aij xi−j+1 + bi . λn with linearly independent eigenvectors c1 . . . . 0 ai+k−1. . (1. .3) Ai = xi = ¯ 0 . . . xi+1−k ). . 0 ¯ . . . λ2 . . . . . (1. 2 . . i = k − 1.2) For the homogeneous case bi = 0. . (1.4. . . .3) is ¯ ¯ n completely given by Eq. dn ] is given by d = C −1 (x0 ). . . . .20CHAPTER 1. .4. . . Here C = [c1 . k. This can be written as the ﬁrst order system xi+1 = Ai xi + bi . .1 This is an eigenvalue problem. 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.k−1 ai+k−1.4) (1. So the solution to the IVP corresponding to Eq. . . 1 (1.k ai+k−1.4. xi+1 . . If A has distinct eigenvalues λ1 . . · · · .4. .5). INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS A k th order diﬀerence equation in 1dimension is given by xi+1 = gi (xi . . 0 bi+k 0 . . . xi+k−1 1 . . . ··· ··· 0 .
N1 (t).6(b)) Let N (t) = (N0 (t). DISCRETIZATION 21 Example 1. 0 ≤ i ≤ M − 1 (1. am ] is a constant vector. . a = [a1 .4. . which are used to solve ODE. . (1. Then we have N0 (t + Ni+1 (t + t) = b0 N0 (t) + b1 N1 (t) + . {λi }M i=1 {φi }M i=1 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 .4. .7) If we denote N m = N (m t). . + bM NM (t). (1. . This is also based on discretization but uses neural network approach (refer Haykin [6]). respectively. the constant vector a is given by a = CN 0 where the nonsingular matrix C is given by C = [φ1 .7) can be viewed as a diﬀerence equation of the form N m+1 = AN m (1. . φ2 .8) This is of the type given by Eq. λ2 .6(a)) (1.4.4. . . .1 (Discrete One Species Population Dynamics Model) We wish to measure population changes in a species with one year age distribution. di the corresponding birth and death rates. If the initial population N 0 is known. respectively (0 ≤ i ≤ M ). . ··· bM 0 ··· ··· 1 − dM −1 (1. then Eq.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.6) becomes N (t + where b0 1 − d0 0 . We also introduce to reader the concept of neural solution of ODE. .4. . (1. However. λM . . Let Ni (t) denote the number of ith year old species and let bi . .4.4. NM (t)).5. . . 0 t) = AN (t) ··· ··· . ( t = 1) t) = (1 − di )Ni (t). Then Eq.5 Discretization In this section we shall describe some wellknown onestep numerical methods. . . we shall not be concerned with the stability and convergence aspects of these methods. .1. .4. .
we get j=1 m x(ti+1 ) = x(ti ) + h j=1 βj f (tij . x(ti )). . h = xi+1 − xi (1. x(ti+1 )) (1. x(ti )))] 2 (1. x(ti ) + hf (ti . x(s))ds by a general quadrature formula βj f (tij . 0 ≤ i ≤ N.1) corresponding to the scalar ﬁeld f (t. t1 . x(ti+1 ))] (1.4) 2 which is again implicit. x(tij )) (1.2) and Eq. x(s))ds is approximated by the trapezoidal rule. Combining Eq.22CHAPTER 1. x) dt (1. .7) .5. tN be a set of time points (called gridpoints) wherein we would like to approximate the solution values x(ti ). (1. we get the Euler’s forward method x(ti+1 ) = x(ti ) + hf (ti .5. (1. x(ti )) + f (ti+1 . x(s))ds (1. we 1 x(ti+1 ) = x(ti ) + h [f (ti .5. x(ti )) + f (ti+1 .5. On the other hand.3) which has to be solved implicitly for xti+1 .5. Integrating the above equation on [ti . .4) to eliminate x(ti+1 ).5) An important class of onestep numerical methods is RungeKutta method. . we get the Heun’s method (refer Mattheij and Molenaar[9]) 1 x(ti+1 ) = x(ti ) + h [f (ti . INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS It is interesting to observe that sometimes neural solution is a better way of approximating the solution of ODE. if f is approximated by its values at ti+1 we get Euler’s backward method x(ti+1 ) = x(ti ) + hf (ti+1 .5.6) Approximating the integral m ti f (s. Let us begin with scalar ordinary diﬀerential equation dx = f (t. Consider the integral equation T x(T ) = x(t) + t ti+1 f (s.5. ti+1 If the integral get ti f (s. ti + 1] we get ti+1 x(ti+1 ) = x(ti ) + ti f (s.5.5.2) which is the socalled explicit. Let t0 . onestep method. x(tij )). x(s))ds If f is approximated by its values at ti . x) : I × → .
we obtain a similar Runge . W ) is the output of a feed forward neural network (refer Haykin [ 6 ]) with one input unit t and network weight vector W .5. l ≥ j. x(ti ) + hk1 ) 2 2 1 1 = f (ti + h.9) is explicit if rjl = 0.8). x(s))ds to get m x(tij ) = x(ti ) + h l=1 rjl f (til .5. x(ti + 1) being replaced by vectors x(ti ). we get the Heun’s formula given by Eq. Also. with scalars x(ti ). ti+1 ].5.5. ρ2 = 1. x(til )) (1. (1. (1. r22 = 0 and β1 = 1/2 = β2 . x(ti ) + hk2 ) 2 2 = f (ti + h.8) Combining Eq. we get the RungeKutta formula m x(ti+1 ) = x(ti ) + h j=1 βj k j m (1. x) as a vector ﬁeld. DISCRETIZATION 23 where h = ti+1 − ti and tij = ti + ρj h(0 ≤ ρj ≤ 1) are nodes on the interval [ti .10) 6 3 3 6 k1 k2 k3 k4 = f (ti . to ﬁnd them we apply another quadrature formula for tij the integral ti f (s. (1.5. As x(tij ) are unknowns.5. x(ti ) + hk3 ) In case we have f (t. For a given input t.9) 1 1 1 1 x(ti+1 ) = x(ti ) + h k1 + k2 + k3 + k4 (1. x(ti+1 ). W ) (1.9) rjl kl ) l=1 kj = f (ti + ρj h. (1. and scalars ki being replaced by vectors k i (1 ≤ i ≤ 4). x(ti )) 1 1 = f (ti + h.1. it is easy to get the following classical RungeKutta method from Eq. (1.7) and Eq. k output of the network is N = i=1 vi σ(zi ) where zi = wi t − ui . r21 = 1. If ρ1 = 0. (1.5. We now brieﬂy discuss the concept of neural solution of the IVP associated with Eq. r11 = 0 = r12 .5.Kutta method. vi denotes the weight from .5.5. x(ti ) + h Eq. wi denotes the weight from the input unit t to the hidden unit i.5.11) where N (t. A neural solution (refer Logaris et al [8]) xN (t) can be written as xN (t) = x0 + tN (t.1).5).
1] = 1 2 1 1 + e−t (1.5.5. (1.1 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 . it may not be exactly equal to f (tj .5. xN (tj )) dt 2 One can use any unconstrained optimization algorithms (refer Joshi and Kannan [ 7 ]) to minimize φ w.r. (1. As xN (t) is assumed to be a trial solution. we compare these solutions with the exact solution. vi . σ(z) is the sigmoid transfer function σ(z) = 1 1 + e−z and k is the number of hidden neurons.t. Example 1. For this we discretize the interval [t0 . wi to get the neutral solution xN (t) given by Eq. Further. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS w1 w2 x w3 uk k v1 v2 v3 N Figure 1. t ∈ [0.5. (1. 0 ≤ j ≤ m. . wherein we compute numerical solutions by Euler’s and RungeKutta methods and also neural solution.1) in some sense.24CHAPTER 1.1: Neural network with one input the hidden unit i to the output and ui denotes the bias of the unit i. xN (tj )) at point tj . . the neural network parameters to obtain their optimal ∗ ∗ values u∗ .5. vi and ui in such a manner that xN (t) satisﬁes Eq.11) i This methodology will be clear from the following example. < tm = tf . We ﬁrst determine the network parameters wi . tf ] as t0 < ti < .13) .12) The exact solution of this ODE (Bernoulli’s diﬀerential equation) is given by x(t) = (Refer Section 6 of this chapter).5.
0.3. v3 . (1.00002 0.00025 0.64566 0.5.00097 0.71095 0. DISCRETIZATION 25 For neural solution we take three data points tj = 0. The following tables give the distinction between the three types of approximate solutions. 1 and three hidden neurons with a single hidden layer network.57469 0.525. We note that the neural solution is a continuous one.14). We can compare the values of this solution at a discrete set of points with the ones obtained by Euler and RungeKutta method ( refer Conte and deBoor [3]).59913 0.66948 0.00044 0. w2 . v2 .68997 0.57444 0.1 ) to compute the optimal weight vector W . Table 1.00129 0.62315 0.5. 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 . The neural solution is given by Eq.59869 0. u1 .00069 0. u2 .00199 0.71294 0.5.73340 0. The neural solution is of the form xN (t) = 1 vi +t 2 1 + e(−wi t+ui ) i=1 3 (1.62246 0.00163 0. v1 .14) The weight vector W = (w1 .54983 0. w3 .1.52498 0.69161 0.00010 0.54994 0.1: Euler solution and actual solution Euler’s method Actual solution Absolute diﬀerence 0.66819 0.64663 0.00234 .5.5. 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 2.73106 0.
00062 0.68997 0.59869 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) (1.57444 0.73106 0.71095 0.62246 0.52498 0.59869 0.05919−9 5.5.64302 0.00264 0.2) .62048 0.54983 0. we get the equivalent equation in normal form dx + P (t)x = Q(t) dt (1.64162−9 4.54973 0.68997 0.59741 0.2: RungeKutta solution and actual solution Runge Kutta method 0.30339−9 2. b(t) and c(t) are continuous functions in a given interval with a(t) = 0.64566 0.00011 0.15537−9 1.28708−9 9.12147−8 1.64566 0.68997 0.71095 0.52511 0.54983 0.00352 0.003231 1.62246 0.00127 0.54983 0.66503 0.00316 0.68650 0. Dividing by a(t).52498 0.66819 0.70743 0.59869 0.62246 0.57444 0.57383 0.00013 0.71095 0.66819 0.5.00348 0.73106 Actual solution 0.72783 0.6.6.00198 0.34665−8 1.66819 0.59648−9 7.64566 0.52498 0.3: Neural solution and actual solution Neural Network sol Actual sol Absolute Diﬀerence 0.1) where a(t).84916−8 Table 1.57444 0.58997−8 1.73106 Absolute diﬀerence 1. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Table 1.26CHAPTER 1.
x)dt + N (t.6. we have M (t.1.6. ∂t If Eq.6.6.6.6. x)dx = ∂u ∂u dt + dx ∂t ∂x ∂u =N ∂x (1.6. (1. intergrating the above wquation we get the solution u(t. x)dt + N (t.6. x)) Hence. e Q(t)dt + c (1. x)dx = 0 or M (t. let Eq. The following theorem gives a test for the exactness of a diﬀerential equation. . x) ∈ D ⊆ 2 ( with no hole in D). x) such that ∂u = M.4). Since M and N are continuously diﬀerentiable in D. (1. x) + N (t. we have ∂ ∂u ∂M = ∂x ∂x ∂t ∂ ∂u ∂N = .5) holds.6.6) holds. then the diﬀerential equation M (t. x) dt dx (1. (1.6) and If the diﬀerential equation is exact. x) + N (t. x)dt + N (t. that ∂x∂t ∂t∂x ∂x ∂t Conversly. x) dx dt = 0 = M (t. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS This equation is solved by multiplying Eq.5) = d(u(t. Theorem 1.1 If M and N are continuously diﬀerentiable function of (t. (1. x)dx = 0 is exact iﬀ Proof : ∂M ∂N = in D ∂x ∂t (1. x) = C of the exact equation given by Eq.4) is called exact if there is a diﬀerentiable function u(t.3) Exact Equations The diﬀerential equation M (t. it follows ∂t ∂t ∂x ∂ 2u ∂M ∂N ∂2u = and hence = in D.6.2) 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.
y)dτ + N (τ. y)dy] M (τ. This gives (t+ t. y)dτ + N (τ. x) = lim [M (τ1 . y)dτ (as y = x and dy = 0) t = = M (τ1 . (1. y)dy] where the integral on the RHS is a line integral along a path joining (t0 . x) ∂x u(t. x). INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS y (t. x) − u(t. x) t→0 t . x0 ) Then.6.x) = M (t.6.6) holds. x) = u0 + (t0 .x) t+ t [M (τ. u(t. x) is given by (t. x)] = M (t.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 u(t0 . x) − u(t. x0 ) and (t.x) u(t + t.x0 ) [M (τ.28CHAPTER 1. = u0 ∂u = N (t. x) t.x) (t+ ∆t .x 0) τ Figure 1. It can be shown that this line integral is independent of path if Eq.x) (t 0 . t < τ1 < t + (by meanvalue theorem for integrals). x) = (t. Hence ∂u u(t + = lim t→0 ∂t t t. x).
xdy − ydx = (x2 + y 2 )dx is not exact.6. x). = ( − a) b dt b dt (1. Such a function is called an integrating factor.6. b ∈ dt We introduce the change of variable X = at + bx.6. x) ∂x 29 It may happen that the equation M (t. wherein substitutions suggest themselves. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS Similarly. For example. 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.7) and hence Eq. we have ∂u = N (t. (1. it becomes exact.8) . x)dt + N (t. x)dx = 0 is not exact but after multiplying both sides of this equation by a function µ(t. 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. a. as we see below [A] dx = f (at + bx).1. which gives x= 1 dx 1 dX (X − at). [B] dx x = f( ) dt t (1.6.
k except when ad − bc = 0. is called Bernoulli’s equation. In that case we have cx + dy = m(ax + by) and hence Eq. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS in which RHS depends only on the ratio Introduce the change of variable u = x = ut. a.6. d. [C] Consider the equation dx =f dt at + bx + p ct + dx + q . Then Eq. We substitute T = t − h.9) = f (u) = f (u) − u t x .10) can always be solved for h. Assume that m = 1. t x . (1. X = x − k.6. q. (1.6.30CHAPTER 1.10) Eq. (1. ∈ (1.8) becomes u+t or du dt du dt which can be easily solved.6. c. Thus t dx du =u+t dt dt in which RHS depends only on the ratio of the linear expression.6.11) in which the exponent is not necessarily an integer.9) becomes dx at + bx + p =f dt m(at + bx) + q which is of the type [A]. b.6. 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]. [D] The equation dx + P (t)x = Q(t)xm dt (1. (1. Then introduce the change of variable . p.
Then x = X ( 1−m ) .x0 ) (τ − y 2 )dτ + (τ + y 2 )dy y (t. x) C2 τ C1 (0.1. we go along the path C1 . x) In the integrand on RHS.2: Path from (0. x) as (t.1 Solve (t + x2 )dx + (x − t2 )dt = 0 M (t. Hence the integral become t x t3 x3 (−τ 2 )dτ + (t + y 2 )dy = − + tx + 3 3 0 0 .x) u(t. So we write the solution u(t.11) becomes 1 31 m dx dX 1 = X ( 1−m ) . x) = x − t2 .6. x) = u0 + (t0 .0) Figure 1. 0) to (t. x) = t + x2 ∂M ∂N =1= ∂x ∂t and hence this equation is exact. C2 . Hence Eq. N (t.6. Example 1. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS X = x1−m .6.6. (1. 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.
INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS Hence.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. the solution is x3 −t2 + tx + =C 3 3 dx t+x−1 = dt t + 4x + 2 Example 1. dT T + 4X We now make substitution X = T U.6.6. Hence t = T + 2.2 Solve Solve h + k − 1 = 0 = h + 4k + 2.32CHAPTER 1. 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 . This gives h = 2. x = dX T +X X − 1 and = . k = −1. 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.
put t2 = u. x(0) = dt 2 1 We make the change of variable X = and get an equivalent linear equation x dX + X = 1. TECHNIQUES FOR SOLVING FIRST ORDER EQUATIONS To compute I = (−3t3 )e− 2 t dt + c.6. 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].1. 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. .4 Solve the IVP dx 1 = x − x2 . 2tdt = du. Burghe and Borrie[2] and Haberman[5]. For more on real life problems giving rise to mathematical models generated by ordinary diﬀerential equations.
Consider the diﬀerential equation dx dt f (t. (0. Solve following initial value problems. 2). (b) Sketch the vector ﬁeld. (b) Solve the above equation. t ≥ 0 given by = (x2 + 1) cos at = x1 + x2 (a) Find the stationary points. wherever possible. (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). 2.7 Exercises dx dt 2 1. x) = (0. x) = f (t.34CHAPTER 1. Sketch the direction ﬁeld for the following diﬀerential equations. 3. dt (a) Determine the equilibrium points. Consider the diﬀerential equation dx = (x2 − 1)tp . (c) Find solution passing through (0. x) = (0. 0) (a) Sketch the direction ﬁeld of the above equation. (t. x(t)) 2tx . Consider the linear system in dx1 dt dx2 dt (b) Sketch the vector ﬁeld in 2 p ≥ 0. 0). y0 ) = (0. (t. 0) and (x0 . 2 . INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS 1. 0) and (0. 4. . 0) 2 t + x2 = 0. −2). 5. (c) Find the number of solutions passing through (0.
Let x(t) denote the number of uninfected individuals in the population at time t. Find the ﬁmily of curves orthogonal to the family of cardiods r = λ(1 + cos θ) in polar cordinates (r. 10. If we assume that the infection spreads to all those susceptible. (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 7.1. 11. solve the following exact equations which pass through the given point. 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(−1) = −2. π (a) [2t + exp(t) sin x] dt + exp(t) cos xdx = 0. In a model of epidemics. Test if the following equations are exact. . θ). Give a mathematical model for the above problem and solve it. EXERCISES dx + x(t + 1) = t. Using the line integrals.7. a single infected individual is introduced in to a community containing n individuals susceptible to the desease. (a) sin t sin2 xdt − (cos t cos 2x tan x + cos t tan x) dx = 0. If so. then x(t) decreases from x(0) = n to x(0) = 0. Solve the following Bernoulli equations. x(−1) = −1 (b) (1 + t2 ) tan−1 t dt a dx − 2tx = a(t3 − t). x(0) = 2 π π (b) (cos t cos x) dx − (sin t sin x) dx = 0. dt dx + x = t. solve them. x(0) = (c) 2(1 − t2 ) dt 3 (a) t2 6. x(1) = 1 t2 + x 2 t + x2 8. x( ) = 2 3 (t + x) t−x (c) dt − 2 dx = 0. 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.
[5 ] Richard Haberman. 1965. 2004. 9(5). Neural Networks. I. [6 ] S. Fotiadis. Ellis Horwood Ltd. 1996. Elements of Ordinary Diﬀerneial Equations.M.36CHAPTER 1. INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS References [1 ] Martin Braun. Springer Verlag. E. Burghe and M S Borrie.Theory and Practice. Neural Networks. 1983. 1994.. Courtney S Coleman and Donald Drew. Ordinary Diﬀerential Equations in Theory and Practice.An Algorithmic Approach. Haykin. [2 ] D. [3 ] S. Population Dynamics and Traﬃc Flow. [8 ] I. Macmillan College Publishing Company. Narosa Publishing House Pvt. Ltd. McGraw Hill Book Company. Artiﬁcial Neural Networks for Solving Ordinary and Partial Diﬀerential Equations. Boston. Mathematical Models: Mechanical Vibrations. [7 ] Mohan C Joshi and Kannan S Moudgalya.. 1981.M Mattheij and J Molenaar. Logaris. N. Diﬀerential Equations Model. A. 9871000. Elementary Numerical Analysis. McGraw Hill Book Co. Prentice Hall. . [9 ] R. John Wiley and Sons. Optimization . 1981. Likas and D. D. 1998. 1977. [4 ] Michael Golomb and Merrill Shanks. Modelling with Diﬀerential Equation. Conte and C deBoor. IEEE Trans.
β ∈ and x. y ∈ X) and scalar multiplication αx (α ∈ . This approach is essential for introducing Green’s function in Chapter 7.1.1 Let X be a nonempty set which is closed with respect to binary operation x + y (x. For any α. y. X is said to be a vector space if the following holds. diﬀerentiability and compactness . 2. We close by deﬁning Diracdelta function as a distribution function.are given a fair treatment.Lipscntiz continuity. monotonicity. x ∈ X). orthormal system and Fourier series. For any x. 1. Parseval’s equality. 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 . We start with vector spaces and go on to deﬁne normed and inner product spaces and then dwell on notions like Schwartz inequality.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. z ∈ X (a) (x + y) + z = x + (y + z) (b) (x + y) = (y + x) (d) ∃ − x ∈ X such that x + (−x) = 0 2. Bounded linear and nonlinear operators in function spaces along with their much needed properties . Bessel’s inequality.1 Finite and Inﬁnite Dimensional Spaces Deﬁnition 2.
x). Deﬁnition 2.4 A metric space X is a nonempty set with a metrix d : X × X → deﬁned on it. This is denoted by x ⊥ y. x ∈ X In a normed space X. 2. (x. α.5 A sequence {xk } in a metrix space (X. y) = d(y.1. l → ∞. y ∈ X x 2 In an inner product space. y) = x − y ∀ x. y. y) + d(y. y) in a vector space X is a function on X × X with values in such that the following holds. y. (x. x) → 0 as k → ∞. this metric d(x. In a metric space. y) = 0. x) ≥ 0 forall x ∈ X and equality holds iﬀ x = ¯ 0 2. without the vector space structure. 1.1.1.. y) = 0 iﬀ x = y [P2 ] d(x. Cauchy convergence. y) satisﬁes the following properties. y ∈ X 3.2 An inner product (x. β ∈ . . y) ≥ 0 for all x. Deﬁnition 2. Such a set is called a metric space. (αx + βy. x. z ∈ X It is possible to deﬁne a metric d(x. In an inner product space X.38 CHAPTER 2. y) satisfying properties [P1 ] − [P3 ] in any set without having a vector space structure. z). x is said to be orthogonal to y if (x. y ∈ X [P3 ] d(x. 3.3. MATHEMATICAL PRELIMINARIES Deﬁnition 2. x ) → as k. Deﬁnition 2. x) for all x. y ∈ X and d(x.1. y) (distance between two vectors) is deﬁned as d(x. as we see below. z) = α(x. d) is said to be convergent to x ∈ X if d(xk . {xk } is said to be Cauchy if d(xk . 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.1. y ∈ X In view of Deﬁnition 2. z) ≤ d(x. y) = (y. z ∈ X The vector space X with an inner product deﬁned is called inner product space. x) for all x. the induced norm is deﬁned by = (x. z) for all x. z) + β(y. we can easily deﬁne the concepts of convergence. completeness etc. 1. the induced metric d(x. [P1 ] d(x. which satisﬁes the properties [P1 ] − [P3 ].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(b)) In terms of the matrix notation we can represent (x. xn ). . Deﬁnition 2. . .1. x2 . d) converges. xn } in a vector space X is said to be a basis for X if 1. xn } in a vector space X is x ¯ ¯ said to be linearly dependent (l. whereas a complete inner product space is called a Hilbert space.i. .1(a)) = x2 + x2 + . . . . d) is said to be complete if everyCauchy sequence in (X.1.1(b)) and hence it is a Hilbert space. x2 . y as column vectors and x . (2. y) = y x = x y if we treat x. xn yn . A complete normed space is called a Banach space. y2 .). y n as row vectors.8 A set of vectors S = {x1 . . . y) = x1 y1 + x2 y2 + . .1. yn ) (2.1.2. . . ∞ In a normed space X. · · · .7 A set S of vectors {¯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. Example 2. L[S] = {α1 x1 + α2 x2 + · · · + αn xn . if every ﬁnite subset of S is . FINITE AND INFINITE DIMENSIONAL SPACES 39 Deﬁnition 2. Deﬁnition 2. and 2. S is .1 In the space product are deﬁned as x 2 n of ntuples. . + x2 . x = (x1 . The number of elements in a basis is unique (refer Bapat[1]). . . . x = (x1 . One can show that is complete with respect to the norm induced by the inner product deﬁned by Eq. xn ) 1 2 n (x.i. αi ∈ equals X. If the set S consists of inﬁnite elements of the spaces X. A vector space X which is not ﬁnite dimensional.i.1. . the Euclidean norm and inner (2. . y = (y1 . y) as (x. x2 .1.1. + αn xn = 0. . . Deﬁnition 2.1. converges to s ∈ X. is called inﬁnite dimensional space. .) if ∃ scalars αi ∈ (not all zero) such that α1 x1 + α2 x2 + . it is possible to deﬁne inﬁnite sum i=1 xi . Then S is said to be . }.i. x2 .d. Otherwise. this set is called linearly independent ( . .1. called the linear span of S. We say ∞ n that s = i=1 xi iﬀ the sequence of partial sums sn = i=1 xi ∈ X.6 A metrix space (X.
6 0. + xn  x ∞ .2(b)) One can show that the space C[0. 1] deﬁned as fn (t) = tn .40 CHAPTER 2.8 0. one can deﬁne norm and inner product as under f 2 2 1 = 0 1 f 2 (t)dt (2. g) = 0 f (t)g(t)dt (2.t. 1] of all real valued continuous functions.10 Two diﬀerent norms x a and x b in a normed space X are said to be equivalent if ∃ positive constants α. To see this consider the sequence {f n (t)} of continuous functions on [0.1 It is possible to deﬁne other norms in the space below x 1 = x1  + x2  + . 1] w. . as we see = max1≤i≤n {xi } Deﬁnition 2.r.1.r.1: A sequence which is Cauchy but not complete fn − fm 2 2 1 = 0 (tn − tm )2 dt → 0 as n.2 In the space C[0. 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.1.4 0. deﬁned earlier. 1] 1 0.4 0. .1.1.8 1 Figure 2. Example 2.1.t. the above norm is not complete and hence it is not a Banach space. x 1 and x ∞ . MATHEMATICAL PRELIMINARIES n Remark 2. m → ∞ .2 0. t ∈ [0.1.2 f1 f2 f3 f4 0.6 0.2(a)) (f.
t = 1 0.1. Parallelogram law: x+y 2 2 = x 2 + y 2 (2. Remark 2.1 A normed space X is an inner product space iﬀ the norm in X satisﬁes the parallelogram law given by Eq. 1] is complete with respect to f ∞ norm and hence it is a Banach space. b L2 [a.1] t ∈ [0.2 n is ﬁnite dimensional.1.1] f (t). we have the following inequality (x. b] → .1. Refer Goﬀman and Pedrick [5] for the proof of the above Theorem. Example 2. where as C[a.1.t.4 X = C[0. Example 2.3 Let L2 [a. we have the following identities. y) ≤ x y (2. 1] Take f = t.2.1. However.1. (2. the above norm.2).2 (Schwartz inequality) Let X be an inner product space. 1.5) Equality holds iﬀ x and y are l.1.r. (2. g = 1. ∈ X. 1] with respect to the above norm can never be made a Hilbert space. 1] with f = sup f (t). 1] is not complete w. Then L2 [a. a f 2 (t)dt < ∞} Let the norm and inner product be deﬁned as in Eq. y. ¯ .d.4). b] denote the space of all square integrable functions. Theorem 2. if f ∞ is deﬁned as f ∞ = supt∈[0. b) and L2 [a. Theorem 2. then C[0.1.4) Parallelogram law is relevant in view of the following theorem. FINITE AND INFINITE DIMENSIONAL SPACES and hence {fn } is Cauchy but fn (t) converges to f (t) deﬁned as f (t) = 1.1. Pythagorus identity: x ⊥ y iﬀ x + y 2. One can check that parallegram law is not satisﬁed and hence the Banach space C[0. b] = {f : [a.1. t = 1 41 which is not a continuous function and hence the space is C[0. Then for all x. f ∈ X t∈(0.1.3) + x−y 2 = 2[ x 2 + y 2] (2. b] are inﬁnite dimensional spaces In an inner product space. b] is a Hilbert space.
y)y. n→∞ x − yn = d .d. (2. In this theorem M ⊥ represent the spaces of all elements orthogonal to M.1.5) is true. Then ∃ a unique element z ∈ M such that x − z = inf y∈M (2.6) x − (x. This proves the ¯ ¯ theorem. 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. y = 0. Let M be a closed subspace of a Hilbert space X. So. (2. e)2 This gives Eq.1. {yn } is Cauchy and since M is a closed subspace of X. the above inequality reduces to ¯ (x.1. e)e) = (x. MATHEMATICAL PRELIMINARIES Proof : If any of the elements x. it is complete and hence y n → z ∈ M. we are through.1. Also. (2. y)y = 0 x−y Proof : Let d = inf y∈M x − y . y is zero. y)y) = 0 This implies that and hence x. x − (x.. (2. we wish to obtain z ∈ M which is closest to M. x − (x. y are l. On the other hand if x. e)e.6).42 CHAPTER 2. Given an element x ∈ X. Then ∃ a sequence {¯n } ∈ M such that y x − yn → d By parallelogram law.1. e) ≤ x for all x ∈ X So. then Eq.1. x) − (x. assume that x.d. Theorem 2. we get (x − (x.5).5). By normalizing y as y/ y = e. We have 0 ≤ (x − (x. It is clear that x − z = lim Uniqueness of z is easy to prove. if equality holds in Eq.3 Suppose x ∈ X and M a closed subspace of X. y are l. We have the following theorem in this direction. it suﬃces to show Eq. This gives us the following theorem.
eβ ) = 0. w ∈ M ⊥ This representation is written as X = M ⊕ M⊥ X is called the direct sum of M and M ⊥ . Theorem 2. α = β and eα = 1 ∀ eα ∈ S.12 Let S be an orthonormal set in a Hilbert space X.7) and further x − ¯ α∈A (x.1. Deﬁnition 2. Then the following are equivalent 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. That is. eα )2 ≤ x 2 (2.2. eα )2 (2. FINITE AND INFINITE DIMENSIONAL SPACES 43 Theorem 2.1. The following theorem presents the most important property of a complete orthonormal set. Then.1.1.4 Let M be a closed subspace of X. Theorem 2.8) . Then z ∈ M is the point of minimum distance from x ∈ X iﬀ x − z ∈ M ⊥ . x ⊥ eα for every α ⇒ x = 0.11 A set S ⊂ X is called an orthonormal set if eα . z ∈ M. (3) Every vector x ∈ X has Fourier series expansion x = (4) Every vector x ∈ X satisﬁes the Parseval’s equality x 2 (x. we have α∈A (x. This.1.1.7) is referred as Bessel’s inequality.5 Let {eα }α∈A be a collection of orthonormal elements in X. eα )eα . The inequality Eq.1. Deﬁnition 2. in turn. α∈A = α∈A (x. We have the following property for orthonormal sets in X.1.6 Let X be a Hilbert space and let S = {eα }α∈A be an orthonormal set in X. eβ ∈ S ⇒ (eα . (2. eα )eα ⊥ eβ ∀ β ∈ A. for all x ∈ X.1. (2) 0 is only vector which is orthogonal to every eα ∈ S. implies that every x ∈ X has a unique representation x = z + w. (1) S = {eα }α∈A is complete.
. √ . This contradicts the completeness of S. π] we have the Parseval’s relation π π f 2 (t)dt = (f. (2. then Parseval’s equation Eq. MATHEMATICAL PRELIMINARIES Proof : (1) ⇒ (2) Suppose (2) is not true. π] will be proved in Chapter 7. eα )eα (3) ⇒ (4) We have x 2 = (x.1. Example 2. eα )eα is orthogonal to eβ for every β. cos nt)2 + (f. 2.. . Deﬁning e = x/ x .. π π π π 2π is a complete orthonormal set and every f ∈ L2 [−π. √ . it follows that it x− α∈A must be the zero vector 0 and hence x= α∈A (x. √ .8) gives e 2 = α∈A (e. (x. then ∃ x = 0 such that x ⊥ eα ∀α ∈ A. k = 0. the collection of functions S= 1 cos t cos 2t sin t sin 2t √ .. . it is properly contained in an orthonormal set ˆ ˆ S and let e ∈ S − S.1.5 In X = L2 [−π. . 1. We also note that for any f ∈ L2 [−π. √ . eα )2 = 0 contradicting the fact that e is a unit vector. sin nt)2 ] 2π π n=1 ∞ .44 CHAPTER 2. π] has a Fourier series expansion ∞ a0 f (t) = + [ak cos kt + bk sin kt] (2.. x) = ( α∈A (x.. π].1. But by (2). eα )2 (x. we get an orthonormal set S ∪ {e} which properly contains S. 1)2 1 + [(f. (2) ⇒ (3) (x. −π The fact that the above set S forms an orthonormal basis in L2 [−π. eβ )eβ ) β∈A = α∈A (4) ⇒ (1) If S = {eα }α∈A is not complete.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 .
y2 . . . xn }.1. . FINITE AND INFINITE DIMENSIONAL SPACES 45 The Fourier series representation given by Eq. x ¯ ¯ We ﬁrst obtain orthogonal vectors {¯1 . . f (t) = 1.9) will be used while discussing the solution of boundary value problems. . . i + 2. . . . It is clear that L[¯1 .2. .1. yj ) y ¯ i≤j ≤i−1 This is continued inductively for i + 1. . . . .i. xn ] = L[¯1 . collection {¯1 .2 y2 . x2 .2: Sketch of ﬁrst few terms of the Fourier series We now give the following GramSchmidt procedure which helps in producing an orthonormal collection {e1 . . 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. n. . . . e1 . π].1. . . . yn ] x ¯ ¯ y ¯ ¯ . (¯j . . . αi.j = = x1 ¯ . inductively y ¯ ¯ y1 ¯ yi ¯ where αi. yj ) x ¯ . .1.1 y1 − αi.i−1 yi−1 ¯ ¯ ¯ ¯ (¯i . = xi − αi. . .2) ∞ 4 sin(2n + 1)t f (t) = π n=0 2n + 1 ft 1 0. y2 . x2 . .5 1 2 3 t 1 Figure 2. en } out of the .5 3 2 1 0. . t ≥ 0 −1. . (2. yn }. . .
(1. MATHEMATICAL PRELIMINARIES for each n. p3 . α1 = 0. is an orthonormal basis for L 2 [−1. P3 (t) = (5t3 − 3t). en }. p1 . t) 3 = .46 CHAPTER 2. 2 2 1 4 2 P4 (t) = (35t − 30t + 3). 1) 1 α0 = = . we get the orthonormal collection {¯. t) =0 (t. ¯ ¯ ¯ Example 2. P1 (t). The higher degree Legendre polynomials are inductively obtained.1. α0 = α1 = and hence p2 (t) p3 (t) = t2 − 1 3 (t3 . . t) (t2 . α2 = 0 (t. p2 . 8 . 1) 3 = t3 − α0 p0 (t) − α1 p1 (t) − α2 p2 (t). which are obtained by GramSchmidt orthonormalisation procedure. . p4 is given below. . . p2 (t)) = . the set of all polynomials {P0 (t). (1. 1) = 0. α3 = 0 α2 = (p2 (t). 1]. (1. Computation for p0 . 1) 1 = . α0 = α1 = (t3 . Normalizing yi . 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). e2 . . . P2 (t) = (3t2 − 1).}. 1]. p1 (t) = t = t2 − α0 p0 (t) − α1 p1 (t). 1 1 given by P0 (t) = 1. 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). called Legendre polynomials. (t4 . ¯ e ¯ ¯ ei = y i / y i . P1 (t) = t. . 1 ≤ i ≤ 4. 1) 5 4 6 (t . 1) (t3 .6 In the space L2 [−1. p0 (t) p2 (x) = 1.
α2 ∈ Otherwise.2. Pi ) = −1 f (t)Pi (t)dt. 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.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. we have the Legendre series representation ∞ 1 47 f (t) = i=0 ai Pi (t).2.2 Linear and Nonlinear Operators In this section. x2 ∈ X . i≥0 For f (t) = sin πt we have the following graphical representation of ﬁrst few terms of the Legendre series. LINEAR AND NONLINEAR OPERATORS For f ∈ L2 [−1. ∀ x 1 . Deﬁnition 2. unless otherwise stated. ∀α1 .2. 2. Deﬁnition 2. to be discussed in Chapter 6. the spaces under consideration will be normed spaces.1.1 T : X → Y is said to be linear if T (α1 x1 + α2 x2 ) = α1 T x1 + α2 T x2 . T is called nonlinear. ai = (f. ft 4 2 3 2 1 1 2 3 t 2 4 Figure 2.2.
we shall be encountering a linear operator of the form 1 [T f ](t) = 0 k(t. (2. . This inequality implies that 1 0 T f (t)2 dt ≤ ( 1 0 k(t. 1] is square integrable : 1 1 1 0 0 Assume that k : [0. .x = ¯ x (2. s)f (s)ds. Example 2. . f ∈ L2 [0. Example 2.1 Let X = [T x]i = j=1 n n and T : n → n be deﬁned by a matrix (αij )n×n n αij xj .. 0 1 0 k(t.2. equipped with the Euclidean norm x 2 = i=1 x2 i .2. x = (x1 . s)f (s)2 ds k(t. Y ) will denote the space of all bounded linear operators from X to Y.2.2. B(X. s) ∈ L 2 [0. . 1] × [0. s)2 ds 1 0 f 2 (s)ds (by the application of Schwartz inequality to f (s) and k(. Also Tx ≤ max i 1 n j=1 Thus T is a bounded linear operator on n . xn ) ∈ T is obviously linear.2. But on all norms are equivalent and hence T is also a bounded linear operator on n 1/2 αij  x 1 n n .2 For a discussion on boundary value problems. 1]). s)dtds < ∞.1) is said to be a bounded operator. Y ) is also normed space with T deﬁned by T = sup{ Tx 0} . MATHEMATICAL PRELIMINARIES One can show that a linear operator T is continuous on X iﬀ ∃ a constant k > 0 such that (2.1) Tx ≤ k x ∀ x ∈ X A linear T satisfying Eq.2) B(X) will denote the space of bounded linear operator from X into itself. One can show that B(X.48 CHAPTER 2. equipped with 1norm. 1] → We have T f (t)2 ≤ ≤ k 2 (t. . s)2 dtds)( 1 0 f 2 (s)ds) .
However. . Deﬁnition 2. + f(¯n )f n e ¯ e e Thus S ∗ is a basis for X ∗ and S ∗ consists of n . . + = 1. The spaces X and Y are then called isomorphic. then one can deﬁne an isomorphism between X and Y by deﬁning a linear operator which maps basis to basis. Then we get Tf ≤ k f ∀ f ∈ L2 [0. 1] (lp) the space of pth power absolutely integrable functions (pth power absolutely summable sequences). 1]). ) of all continuous linear operators from X to will play a signiﬁcant role throughout the solvability analysis of ordinary diﬀerential equations. . p p q From this result it follows that Lp (lp) spaces are reﬂexive spaces. 1] = L2 [0. Example 2. . . vectors. it follows that every ﬁnite dimensional space of dimension n is isomorphic to n . Deﬁne a set S ∗ = {f 1 . T ∈ B(L2 [0. i = j 1. 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. (Refer Royden [10]). . One can show that X ⊆ X ∗∗ . Example 2. . e2 .4 We denote by Lp [0.2. This space is called the conjugate space or dual space of X and is denoted by X ∗ .2. if X = X ∗∗ . en } be an orthonormal basis for n . 1](l2 = l2 ). 1] That is. f n } ⊂ X ∗ as follows: f j (ei ) = δij = 0.i. if f ∈ X ∗ . . ∗ we have L∗ [0. then X is said to be reﬂexive. one can write j=1 f = f (¯1 )f1 + f(¯2 )f 2 + . The normed space B(X. 1](lp∗ ) = Lq [0. f 2 . 2 . So X ∗ is equal n (up to an isomorphism). 1 1 One can show that L∗ [0. If X and Y are ﬁnite dimensional spaces of the same dimension. i = j It is clear that {f j }n are . Then X ∗ = n (up to an isomorphism).2. . Elements of this space are called continuous linear functionals or simply functionals. With this notion. LINEAR AND NONLINEAR OPERATORS Denote by k 2 = 1 1 0 0 49 k(t.3 A one to one and onto linear mapping from X to Y is called isomorphism.i. Also.3 Let X = n .2. s)2 dtds. Let S = {e1 . 1 < p < ∞. In particular. As claimed before. . To see this we proceed as follows. 1](lq).2.
2.50 CHAPTER 2.2. Deﬁnition 2. then corresponding to T ∈ B(X. of T as follows: [T ∗ f ][x] = f (T x). T y) = 0 for ¯ all y ∈ X. Reversing the above argument. This is done as follows by making use of Theorem 2. Let M = R(T ). MATHEMATICAL PRELIMINARIES Deﬁnition 2. Y ). If T ∈ B(X. R(T ) = {y ∈ Y : y = T x. ¯ ¯ For T ∈ B(X. Then N (T ∗ ) = R(T )⊥ and R(T ∗ )⊥ = N (T ). T ∗ y)X = (T x. v ∈ R(T )⊥ = N (T ∗ ) (2. y ∈ Y. By Theorem 2. X) as : (x.1.2 Let X and Y be Hilbert spaces.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. x ∈ X} We have the following relationship between N (T ∗ ) and R(T ) and viceversa. the subspaces N (T ) and R(T ) are deﬁned as follows N (T ) = {x ∈ X : T x = 0}.2. T is said to be positive semideﬁnite if (T x. This theorem can be extended to get the following result. we get the orthogonal projector P from X to M = R(T ). Y ). y)Y ∀ x ∈ X. if X is a Hilbert space. Y ). then R(T ) = N (T ∗ )⊥ and R(T ∗ ) = N (T )⊥ . y) = 0 for all y ∈ X. Theorem 2. We say that T ∈ B(X) is selfadjoint if T ∗ = T.1.4. Theorem 2. f ∈ X ∗ . one can show that R(T )⊥ ⊂ N (T ∗ ) and hence the ﬁrst part of the result.5 Let X be Hilbert space and T ∈ B(X) be a self adjoint operator.2.3) . Assume that x ∈ N (T ∗ ). We deﬁne the conjugate operator T ∗ : Y ∗ → X ∗ . we get R(T ) = N (T ∗ )⊥ .2.1 If R(T ) is closed.4 we can write z ∈ X as follows: z = u + v. However. That is. u ∈ R(T ) = M. then T ∗ x = 0 and hence (T ∗ x. R(T ∗ ) = N (T )⊥ For operators T ∈ B(X) with closed range in a Hilbert space. Proof : It is suﬃcient to prove the ﬁrst one. x ∈ X One can show that T ∗ is also bounded and T ∗ = T . if T ∗ T is invertible. x ∈ R(T )⊥ and hence N (T ∗ ) ⊂ R(T )⊥ . Y ) we deﬁne T ∗ ∈ B(Y. This implies that (x.1 Let X and Y be Hilbert spaces and let T ∈ B(X. Remark 2.2. Y ).
Let the ﬁrst quantity be represented by variables x1 . . .2. 2. B = (βij ) This implies that A∗ = B = (βij ) = (αji ) = A Example 2.2. x2 . xn .2. . Then A∗ = (βij ) where βij = αji . m. LINEAR AND NONLINEAR OPERATORS v ∈ N (T ∗ ) implies that T ∗ (z − T x) = 0 and hence T ∗ z = T ∗ T x.6 If X = n and A = (aij ) is a m × n matrix with (A A)m×m invertible. Data ﬁtting model is used in measuring relationship between two physical quantities. then Eq. .2. We try to ﬁt the data by a linear function F (x) given by F (¯) = a0 + a1 x1 + · · · + an xn x (2. we get n n ( i=1 j=1 αji yj )xi (Ax. the second quantity by y and the functional relationship be given by y = f (x). . 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. .2. The problem of data ﬁtting is to recover the function f (x) from the given data fk . .5) gives the element u ∈ R(A) nearest to R(A) from z ∈ n . k = 1. . y) = ( i=1 j=1 βij yj )xi = (x. xn ) We assume that the observed data fk approximate the value f (x(k) ). as we see below.2. That is A∗ . y) = i=1 j=1 n n αij xj yi αji xi yj j=1 i=1 n n = = As βij = αji . . . . x = (x1 .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. giving the vector x ∈ n . Example 2. We have n n (2. By). (2. .5 Let X = n and A be the matrix (αij ).6) . reduces to the transpose A of the matrix A.2.5) (Ax.
. 1]. . a0 . . . . . . . .r. . . .52 CHAPTER 2. . z = (f1 .2 From the point of view of applications to diﬀerential equations. (2. Remark 2. fn ) ¯ ¯ 1 x(1) x(2) · · · (1) (2) x2 ··· 1 x2 A= . the 2norm m ¯ d 2 = k=1 ¯ d2 . . d on the whole space L2 [0. k = 1. 1]} dt ∗ D(T ) is called the space AC[0. 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] of all absolutely continuous functions on [0. (2. .6). an ). dm ) k From the linear model given by Eq. . . . 1] : f˙ = df exists a.5). . 1]. 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. the vector u is given by u∗ = A(A A)−1 A z ¯ ¯ if (A A) is invertible. an ).t. a1 . with f˙ ∈ L2 [0. However.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 . d = (d1 . 1]. By using Eq. . 1] to a sequence T fn(t) = π cos nπt which is not convergent.2. It can be shown that D(T ) is dense in L2 [0. ··· .2. .2. (1) (2) 1 x m xm · · · x(n) (n) x2 . a1 . f2 . MATHEMATICAL PRELIMINARIES We shall compute parameters ak so that the deviations dk = fk − F (x(k) . m are as small as possible. . . . . d2 . . xm (n) . .
g(0) = g(1) = 0}. By Riesz representation theorem (refer Bose and Joshi [2]). LINEAR AND NONLINEAR OPERATORS 53 Deﬁnition 2. . g ∈ L2 . Hence (T f.2. y∗ ) ∀ x ∈ D(T ) The adjoint T ∗ of T is deﬁned on D(T ∗ ) by T ∗ y = y ∗ . y ∈ D(T ∗ ) T is said to be selfadjoint iﬀ D(T ∗ ) = D(T ) and T ∗ = T. g) = (f. ∀ x ∈ D(T ).2. d2 f ¨ Example 2. Also one can show that T is closed. y) = (x. 1] and T f (t) = −f(t) = − 2 . T ∗ y). 1] with f (0) = f (1) = 0} One can show that D(T ) is dense in L2 [0.2. 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.6 Let X be a Hilbert space and T : D(T ) → X be such that D(T ) is dense in X. there exists a unique y ∗ ∈ X such that (T x. T is a selfadjoint operator and is densely deﬁned. y) = (x. 1] and 1 1 (T f. In other words the adjoint T ∗ of T for densely deﬁned operators is given by (T x. That is if xn → x in X = L2 [0. So. 1]} : f.2. f ∈ L2 [0. where D(T ) is dt deﬁned as under ˙ ¨ D(T ) = {f ∈ L2 [0. Then we deﬁne D(T ∗ ) to consist of y ∈ X for which the linear functional f (x) = (T x. f ∈ D(T ) g f (t)¨(t)dt g 0 = − f (t)¨(t)dt g 0 if g(0) = g(1) = 0. 1] and T xn → y in X. Then x ∈ D(T ) and y = T x.7 Let X = L2 [0. g ˙ ¨ Thus T ∗ = T and D(T ) = D(T ∗ ). T g) where T ∗ g = −¨ and D(T ∗ ) = {g ∈ X : g. y). x ∈ D(T ) ⊂ X is continuous in D(T ).
.2. Refer Bose and Joshi [2] for the proof of this theorem. 1] → L2[0.7 T ∈ B(X. Y ) is compact. X) is compact. 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.2. This proves the → αj . 1]. . . . (2) Limit of a sequence of compact operator is compact. s)]2 dtds < ∞. as deﬁned in Example 2. 1]. Let {φi (t)} be a complete orthonormal set in L2 [0. Theorem 2.2. Y ) is compact iﬀ T ∗ ∈ B(Y. Deﬁnition 2. Hence ∞ ∞ k(t. 1] 1 with 0 1 ∞ [k(t. MATHEMATICAL PRELIMINARIES Deﬁnition 2. (n) N Consider now a bounded sequence {xn } ⊂ X. y N ]. {T xn} has a convergent x ¯ subsequence in Y . T is compact iﬀ for every bounded sequence {¯n } ⊂ X.j=1 is a complete orthonormal set in L2 ([0. s)φi (t)φj (s)dtds. Equivalently.2: 1 [T f ](t) = 0 k(t.2. .8 Consider the operator T : L2 [0. y 2 . s) = i=1 j=1 aij φi (t)φj (s) where aij = 1 1 0 0 k(t.2.8 T ∈ B(X. Theorem 2. 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. then T xn = j=1 αj y j for some scalars αj .3 Every ﬁnite rank operator T ∈ B(X. X and Y are assumed to be Hilbert spaces. then the set {φi (t)φj (s)}∞ i. s)f (s)ds. Proof : As R(T ) is ﬁnite dimensional. 1] × [0. } such that αj (n ) αj k y j (nk ) → y = α1 y1 + α2 y2 + .4 (1) The composition of a compact operator with a bounded operator is compact. we have R(T ) = L[y 1 .54 CHAPTER 2. Example 2. 1]). f ∈ L2 [0. (3) T ∈ B(X. By Parseval’s equality. Y ) is said to be compact if it maps bounded sets B ⊂ X onto relatively compact set T (B) (T (B) is compact). + αN yN .2. s)]2 dtds = i=1 j=1 (aij )2 . . we have 1 0 0 1 ∞ ∞ [k(t. 1 ≤ j ≤ N. Y ) is called ﬁnite rank operator if the range of T is ﬁnite dimensional.
refer Bose and Joshi[2] for the proof. Also Tn → T in B(X) and hence by the previous theorem T is compact. s) = i.j aij φi (t)φj (s). Then the eigenvectors of T form a complete orthonormal basis for X.2.2. 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. In particular. giving the solvability of the operator equation involving compact operators. 2. Theorem 2. LINEAR AND NONLINEAR OPERATORS Further.6 Let T be a selfadjoint compact operator on a Hilbert space. . Either both (H) and (H ∗ ) have only trivial solutions or both have nontrivial solutions. called Fredholm Alternative. s) − 1 1 0 0 ∞ i=1 j=1 ∞ = aij φi (t)φj (s) dtds 2 2 i. n Now deﬁne Kn (t. 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.j=n+1 [aij φi (t)φj (s)] dtds →0 = i. We now state the following theorem which gives orthonormal basis for a Hilbert space X arising out of eigenvectors of T ∈ B(X). s)f (s)ds is a ﬁnite rank operator and hence it is compact.5 1. we have 1 0 0 1 55 n n k(t. Then Tn : X → X deﬁned by 1 (Tn f )(t) = 0 Kn (t. The necessary and suﬃcient condition for (N H) to have a nontrivial solution is that f be orthogonal to every solution of (H ∗ ).j=n+1 ∞ ∞ aij 2 since i=1 j=1 aij 2 < ∞.2.2. We have the following theorem.
7) is equivalent to the solvability of the equation t x(t) = x(t0 ) + t0 f (s. refer Joshi and Bose [6]. Example 2. Then A ∈ B( n ) is compact selfadjoint operator and hence the eigenvectors of A form a basis.2.9 Let X = n and A : n → n be given by a symmetric matrix (aij )n×n . i=1 Let P = [φi . x(t)). .9 F : X → Y is said to be continuous if xn → x in X ⇒ F xn → F x ∀ x ∈ X. MATHEMATICAL PRELIMINARIES Corollary 2. Example 2. However. λn ].3 For nonlinear operators continuity is not equivalent to boundedness. t ∈ I = [t0 .2. We now examine the properties of a nonlinear operator F : X → Y.1.8) .2. we get P −1 AP = D.56 CHAPTER 2. . .2. Thus A is similar to a diagonal matrix. (2. . x(s))ds. . the solvability of Eq. . . (2. Then P is an orthogonal matrix (P P = I) and AP = [Aφi . . This is called diagonalisation procedure and will be used in Chapter 4 while computing transition matrix of a given system of diﬀerential equations. .1. it is merely nonsingular. . .2. φn ].2. Let Aφi = λi φi . . Aφn ] = [λ1 φ1 . X and Y are normed spaces. φ2 .10 Consider the following ODE dx = f (t. tf ] dt with f : I × → continuous and bounded. it may be pointed out A can be diagonalised even if A is not selfadjoint provided it has distinct eigenvalues. .1 Let T be a compact and selfadjoint operator in a Hilbert space X.7) Then by Proposition 1. t ∈ I = [t0 . Then T is positive semideﬁnite (positive deﬁnite) iﬀ all eigenvalues of T are nonegative (positive).2. Compact nonlinear operators are deﬁned in the same way as compact linear operators. This gives us P −1 AP = D. λn φn ] = P D where D = diag[λ1 . Deﬁnition 2. . tf ] (2. But P is no longer an orthogonal matrix. F is said to be bounded if it maps bounded sets in X into bounded sets in Y.2. 1 ≤ i ≤ n where {φi }n are orthonormal. . Assume that the eigenvectors corresponding to distinct eigenvalues are linearly independent and hence deﬁning P as before. Remark 2.
x)∈I× ≤ mt1 − t2 (m = f (s. x. 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 . one can show that F is bounded on X. One then denotes by C[t0 . it follows by bounded convergence theorem. x) is deﬁned on I × Ω → Ω where Ω is an open subset of n . Since (s.1) .tf ] Hence F is a continuous nonlinear operator from X = C[t0 .8) gives rise to a nonlinear operator F on C[t0 .2. x(s))ds sup (s. x) is continuous. x(s))ds We have t2 [F x](t1 ) − [F x](t2 ) ≤ t1 f (s. Deﬁnition 2. tf ]. x(s)ds Let xk → x in C[t0 . Similarly. x) → f (s. y ∈ X (2. xk (s)) − f (s.3. The above result also holds true if f (t.3. This implies that xk (s) → x(s) in for all s ∈ [t0 . tf ] deﬁned as follows t [F x](t) = x(t0 ) + t0 f (s.2. tf ] is deﬁned as x = sup t∈[t0 . Further t F xk (t) − F x(t) ≤ t0 f (s. tf ] in to itself. s)ds → 0 for all t ∈ [t0 . t1 ]. x)) For this inequality it follows that F x is continuous and hence F x ∈ C[t0 . xk (s)) − f (s. tf ]. (2. tf ] if x ∈ C[t0 .tf ] x(t) n 2.1 F : X → Y is said to be Lipschitz continuous if ∃ constant k > 0 such that F x − F y ≤ k x − y ∀ x. LIPSCHITZ CONTINUITY AND DIFFERENTIABILITY 57 The RHS of Eq.3. tf ] the space of all n − valued continuous functions and the norm of the elements x ∈ C[t0 .3 Lipschitz Continuity and Diﬀerentiability In this section too the spaces under consideration will be normed space. tf ]. that t F xk (t) − F x(t) ≤ t0 f (s.
. . Its value at h will be denoted by F (x)h.3. . Refer Joshi and Bose[ 6 ] for the proof of this theorem. . . 1.3.2 F : X → Y is said to be diﬀerentiable at x ∈ X if ∃ A ∈ B(X.1 If F (¯) ≤ K ∀ x ∈ X then F is Lipschitz continuous.58 CHAPTER 2. xn ). Corollary 2. fm (x1 . . . . Y ) such that F (x + h) − F (x) = Ah + w(x. . . Then for ¯ points x. MATHEMATICAL PRELIMINARIES Deﬁnition 2. which we want to compute) such that F (x) = A. . Let A = (aij ) be the m × n (matrix. . . . 0) Then lim 1 F (x + th) − F (x) − tAh = 0 t t→0 1 ⇔ lim  [fi (x + ej ) − fi (x) − taij ] = 0.1 Let F : n → n be given by F (x1 .2) where lim ¯ h →0 w(x.1 Let F : X → Y be diﬀerentiable at every x in X. there exists a constant τ. x2 . h) =0 h A is called the derivative of F at x and is denoted by F (x). F (¯ + τ h)h ¯ x ¯ ¯ x This immediately gives the following Corollary. F (¯ + h) − F (x) = e. ¯ x Equivalently. . writing φ(t) = F (x + th) for x. for all h ∈ X dt We have the following variant of the meanvalue theorem for F : X → Y. h ∈ X. ¯ We choose h ∈ n as j th coordinate vector ej = (0. . x + h ∈ X and e ∈ Y ∗ . 1 ≤ i ≤ m. . 0 < τ < 1. such that ¯ ¯ ¯ ¯ ¯ ¯ e. h) (2. xn )).3. . 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. x2 . Theorem 2. . .3. 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) . x ¯ Example 2. . if partial derivative are ∂xj n . . .3. we see that F has derivative F (¯) iﬀ x d ¯ [φ(t)]t=0 = F (x)h. xn ) = (f1 (x1 . x2 .
This implies that f (x) = Ax ∀ x ∈ X.3. A function f : X → is said to be convex if f (αx + (1 − α)y) ≤ αf (x) + (1 − α)f (y). x). X is any normed space. . . The derivative of f is computed as under. 0 < α < 1 for all x. ∂n fm (¯) x n to .3. h) t→0 2 t lim lim 1 2 t→0 1 (Ax. h) + (x. Ah) + t2 (Ah. So we have ∂1 f1 (¯) x ··· ∂n f1 (¯) x 59 In case f is a functional. f (x)h = = = = = lim f (x + th) − f (x) t 1 (A(x + th).2. that is. f is a mapping from f (x). 2 x∈X be given by We assume that A is a bounded selfadjoint linear operator. This derivative f of a functional f is also denoted as f .2 Let X be a Hilbert space and let f : X → f (x) = 1 (Ax.1 In dealing with minimization problem involving several variables we shall encounter convex functionals as in Example 2. . h) + t(Ah. y ∈ X. LIPSCHITZ CONTINUITY AND DIFFERENTIABILITY continuous. ∂1 f2 (¯) · · · x F (x) = . x) 2 t t→0 t→0 lim 1 t (Ax. ∂1 fm (¯) · · · x ∂n f2 (¯) x . . h) + (A∗ x. (x + th)) − (Ax.2.3. called the gradient of f at x.3. . Remark 2. h) 2 A + A∗ x. h) as A∗ = A. then f (x) = Example 2. h 2 = (Ax. From this representation it also follows that f (x) = A.
.3. Example 2. = (z.5 involving problems of least square approximations. Such functionals f may not be convex and hence there is a need to generate algorithms which will computationally capture the minimizer (local / global).3. z) − 2(Aa. we get φ(a) = −2A z + 2A Aa It is clear that A A is positive semideﬁnite.2 is strictly convex. a).2. This is the same result.3 While solving diﬀerential equations.3.1 Step 1 Step 2 Step 3 Step 4 Step 5 Step 6 : : : : : : Start with x(k) (k = 0) Set d (k) Minimize 1dimensional function φ(α) = f (d α(k) (k) = −g(k) . Let φ(a) be deﬁned as follows φ(a) = (z − Aa. MATHEMATICAL PRELIMINARIES Remark 2. Hence the functional of Example 2. 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. Otherwise set k = k + 1 or GO TO STEP 2 Compute f (x∗ ).3 (Least Square Approximation . Remark 2. We give below one such algorithm . z − Aa) = (z.Data Fitting) 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) . x∗ is the point of optimum. we shall also be estimating a set of parameters which will arise in the deﬁning model. z) + (A Aa. z) + (Aa.3. 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. Aa) Proceeding as in Example 2.which captures the local minimizer of a function f deﬁned on the entire space n .called the steepest descent algorithm . z) − 2(Aa.6).3.3. if A is positive deﬁnite. which we obtained earlier.60 CHAPTER 2. Estimating parameters requires the minimization of a functional involving these parameters.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. 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. a ∈ ¯ n Steepest Descent Algorithm 2.2.
s) is a bounded on [0. x(s)) and is given by ∂x 1 F (x)h(t) = 0 g(t. we can prove that F is a continuous mapping of C[0.3.3.2. 1]. x) is diﬀerentiable mapping from to for all t ∈ [0. 1] × ∂x As in Example 2. 1] be deﬁned by 1 [F x](t) = 0 k(t.3) Example 2. Also. .1 it follows that F is Lipschitz continuous on X = L2 [0. s)f (s. x ∈ 2 n 61 (2. as g(t.3. 1] such that ∂f  (t. 1] → C[0. x(s))ds where k and f satisfy the following properties (i) k(t. x) ≤ M for all (t. g(k) Ag (k) . s) 0 [f (s. 1] × [0. s)h(s)ds.11. LIPSCHITZ CONTINUITY AND DIFFERENTIABILITY In Example 2. We now compute the derivative of F. (ii) x → f (t. 1]. s) ∂ f (s. 1] to itself. 1] × [0. s) is continuous on [0. Thus F is diﬀerentiable at x with F (x) as a linear integral operator generated by the kernel ∂f g(t. (x + αh)(s)) − f (s. 1].3. x) ∈ [0. the iteration procedure for the functional f (x) = is given by x(k+1) = x(k) − g(k) . s) = k(t.2. x(s))] ¯ α k(t.2. x(s))h(s)ds ∂x by Lebesgue dominated convergence theorem.3. x). s) (s. g (k) g (k) 1 (Ax.4 Let F : X = C[0. it follows that F (x) ≤ M and hence by Corollary 2. [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.
u) = (f1 (¯.4) Deﬁnition 2.3. it suﬃces to compute the zeros F (x) and hence ˆ 0 = F (x) = F (x0 ) + F (x0 )(x − x0 ). u ∈ 2 x ¯ ¯ ¯ (2. f4 (¯. u)) x ¯ x ¯ x ¯ x ¯ x ¯ where x4 k f1 = x2 . STOP and set x∗ = x(k) as the computed zero of F (¯). then it follows that we have the following expansion for F (x) with F (¯0 ) ∈ B(X).5) F (x0 ) Hence we get the Newton’s algorithm to ﬁnd the zeros of F (x).3. x F (x) = F (x0 ) + F (x0 )(x − x0 ) + w(x.3. u). u). x ˆ By linearization. f4 = −2σ( + w) σ (x1 + σ) (x1 + σ) . Remark 2. This gives x = x0 + [F (x0 )] −1 (2.2 that the motion of a satellite orbiting around the earth is given by an ordinary diﬀerential equation (ODE) in 4 : d¯ x = f (¯.3. f3 (¯. x3 (t). f is given by ¯ f (¯.3. f2 (¯. MATHEMATICAL PRELIMINARIES Let F be diﬀerentiable at x0 ∈ X.3. ¯ Algorithm 2.3 If F is diﬀerentiable at x0 then the function (2.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) <∈. h → 0 as h → 0. u).2 .3. x ∈ 4 . h) where w(x.3.6) dt x(t) = (x1 (t). x4 (t)) represents the state vector and ¯ u(t) = (u1 (t). u).5 We have shown in Example 1. x2 (t). starting with an initial guess x0 . h ˆ F (x) = F (x0 ) + F (x0 )(x − x0 ) is called the linearization of F at x0 . Otherwise increment k by 1.62 CHAPTER 2. u2 (t)) is the control vector (thrusts on the satellite).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 = (x1 + σ)( + w)2 − + u1 . x Step 4 GO To Step 2 Example 2. σ (x1 + σ)2 x2 u2 σ x4 + f3 = x4 .
4.3.4. So the linearized ODE.1 Let X be a metric space and F : X → X. d) be a metric space. u) = fx (¯ ¯ x + fu (¯ ¯ u 0. respectively.7) Here σ is normalized to 1.6) is d¯ x = A¯ + B u x ¯ dt where A and B are given by Eq. Deﬁnition 2. 0)¯ 0. We begin with the following constructive ﬁxed point theorem by Banach.3.2. around x ¯ x = 0.7) and Eq. (2. u). described by Eq.4.8).3.3. u = 0 we get ¯ ¯ ˆx ¯ f (¯.3. 0) = ∂f3 ∂u1 ∂f 4 ∂u1 ∂f1 ∂u2 ∂f2 ∂u2 ∂f3 ∂u2 ∂f 4 ∂u2 1 0 = 0 0 0 1 (¯. known as Banach contraction mapping principle. (2. FIXED POINT THEOREMS AND MONOTONE OPERATORS 63 Using linearization procedure for f (¯.¯) oo 0 0 (2.2 Let (X. 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 (¯.¯) oo 0 1 0 0 0 0 −2w 0 2w 0 1 0 0 (2. 0)¯ ∂f ∂f ∂f ∂f 1 1 1 1 A = f x (¯ ¯ = 0.8) 2. F : X → X is said to be contraction if F is Lipschitz continuous with Lipschitz constant strictly less than .3.4 Fixed Point Theorems and Monotone Operators Deﬁnition 2. This theorem will be one of the important instruments in obtaining existence theorems for ordinary diﬀerential equations. (2. An element x ∈ X is said to be a ﬁxed point of the mapping F if F x = x. ∂f 1 ∂u1 ∂f2 ∂u 1 B = fu (0.5). (2. corresponding to Eq.
x∗ ) n ≤ d(xm . . xk−1 ) = d(F xk−1 . then (2. xn+1 = F xn and F is continuous. it converges to x∗ ∈ X. 1−α Proof : We have d(xk .64 CHAPTER 2. . x0 ) This gives d(x . . + αl−1 ]d(x1 .4. y ∗ ) = d(F x∗ . x0 ) ≤ αm [1 + α + α2 + . The error estimate d(xn . MATHEMATICAL PRELIMINARIES 1. 2.1) If X is a normed space.) then lim xn = x∗ and d(xn . we have d(x∗ . F xk−2 ) ≤ αd(xk−1 . xn ) αn d(x1 . x∗ ) + 1−α . x0 ).4. y ∈ X ¯ ¯ (2. y ∈ X for all x. xk−2 ) ≤ αd(x1 . y ∗ ) and hence d(x∗ . x −1 ) + . Then F has a unique ﬁxed point x∗ ∈ X. y) Fx − Fy ≤ α x − y ¯ ¯ ¯ ¯ for all x. . F y) ≤ αd(x. x0 ) αm = d(x1 . it follows that x∗ = F x ∗ . The uniqueness of the ﬁxed point x∗ follows from the fact that for two ﬁxed point x∗ and y ∗ . x0 ).1) is equivalent to the following inequality (2.1 Let F be a contraction mapping of a complete metric space X into itself. x∗ ) is given by d(x∗ . xm ) ≤ d(x . . . and {xn } a sequence deﬁned by xn+1 = F xn (n = 0. F y ∗ ) ≤ αd(x∗ . x∗ ) + d(xm . . y ∗ ) = 0. m > k ≤ d(xm .4. + d(xm+1 . As xn → x∗ .2) Theorem 2. there exists a positive constant α < 1 such that d(F x. . x0 ). 1. y ∗ ) < d(x∗ .]d(x1 . xm ) ≤ [αm + αm+1 + . l > m 1−α → 0 as m → ∞ This implies that {xn } is Cauchy in X and as X is complete. That is.4. . Moreover if x0 ∈ X is any arbitrary point. x∗ ) ≤ n→∞ αn d(x1 .
Then (I − A) is bounded and invertible with 1 (I − A)−1 ≤ (2. x∗ ) ≤ 65 αn d(x1 . Then F x1 − F x2 = Ax1 − Ax2 ≤ A x1 − x2 for all x1 . it follows that F is a contraction and hence it has a unique ﬁxed point x in the Banach space X (complete metric space). let x = (I − A)−1 y This implies that x − Ax = y. A ∈ B(X) with A < 1 and y ∈ X. y ∈ X has a unique solution.4. we obtain the solvability of the linear equation x − Ax = y (2. x2 ∈ X As α = A < 1. That is.4.1 Let A ∈ B(X) be such that A < 1.2.4. Further. Here. x0 ) 1−α As a Corollary of this theorem. ¯ This implies that 1 1− A 1 1− A y ¯ (I − A)−1 ≤ .4) 1− A Proof : We ﬁrst show that x − Ax = y. Let F x = Ax + y.4. FIXED POINT THEOREMS AND MONOTONE OPERATORS Letting m → ∞. Eq. (2. we get d(xn .4. ¯ Corollary 2.3) has a unique solution x.3) in a Banach space X. 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.
− αnn αnn αnn b1 α12 x2 α1n xn − . 1 ≤ i ≤ n. + α2n xn = b2 . Expanding Eq.4.6) Assume that αii > 0 and the matrix A is diagonally dominant: n i=j i=1 αij  < αii . (2. .4.66 CHAPTER 2. . α n 1 x1 + α n 2 x2 + .7) is equivalent to the solvability of the matrix equation x − B x = c.. + α n n xn = b n (2.4..4.7) Then the solvability of Eq.1 Solve the linear system A¯ = ¯ x b in n (2. . ¯ ¯ ¯ Hence by the above corollary. c = 22 . αn2 αnn ··· ··· ··· b1 α1n α11 α11 b α2n 2 α22 α . . Then Eq. (2.4. .6) is equivalent to x1 x2 = = . .. .5) . ..4. − α11 α11 α11 α21 x1 α2n xn b2 − −. bn 0 αnn = α n . MATHEMATICAL PRELIMINARIES Example 2.5) in terms of its component we get α11 x1 + α12 x2 + .. xn Deﬁne α21 α 22 B = − α n1 αnn 0 α12 α11 0 . . . + α1n xn = b1 α21 x1 + α22 x2 + . . . this has a unique solution x∗ and the iteration ¯ scheme x(k+1) = c + B x(k) ¯ ¯ ¯ . B < 1.n−1 xn−1 bn − −.− α22 α22 α22 (2. . x1 αn. . (2. A = (αij )n×n ..4.
x2 (k+1) .06619 0. −. = . For the diagonally dominant 24 3 1 12 A= 0 1 7 −4 ¯ = (1.00786 0. x4 ) of A¯ = ¯ ¯ x b. . F : (X.06603 (n) x4 2 0.00178 0.16895 0.02336 0.09077 0. d) be such that F N is a contraction for some positive integer N. −4) b Table 2.04901 0. x3 .05785 0. 2.16575 0.17548 0. . Corollary 2.16960 0. . as we see in the following corollary.2 Let (X.08465 0.53030 1.06073 0.00631 0.00089 It may happen that F : (X.16986 (n) x − x(n) 5. FIXED POINT THEOREMS AND MONOTONE OPERATORS 67 converges to x∗ . −.04018 0. d) be a complete metric space.04368 0.06154 (n) x2 0 0.4.03850 0. . .n (k) − x ···− x α22 α22 1 α22 n . . 1.1: Iterated solution of the matrix equation n 0 1 2 3 4 5 6 7 8 9 x1 1 0.07844 0.06769 0.06125 0. We still get a unique ﬁxed point. . d) → (X. Then F has a unique ﬁxed point.4.n−1 (k) − x ··· x αnn αnn 1 αnn n matrix 9 7 −3 −4 .16788 0.15573 0. 22 −3 −3 24 (n) (n) (n) (n) we get the following iterated solution x(n) = (x1 .00389 0.05923 0.06610 0.58333 0.16667 0. ¯ This gives us the well known Jacobi’s iteration scheme (refer Conte and deBoor [4]). x(k+1) n = bn αn1 (k) αn.06341 0.06634 0.4.03822 (n) x3 1 0.00316 0.2. d) → (X.04622 0.41667 0. x1 (k+1) = = b1 α12 (k) α2n (k) − x ···− x α11 α11 2 α11 n bn α21 (k) α2.56439 0.00758 0. x2 .37828 0.03902 0. .36364 0. d) is not a contraction but some power of it is a contraction.04319 0.02363 0.
Theorem 2.4. Another important notion which will be used in solvability analysis of diﬀerential equation is the concept of monotone operators in Hilbert spaces. the space X is assumed to be a Hilbert space. we shall encounter mappings in a normed space. Example 2. At times. 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. Let F be a continuous mapping of K into a compact subset of K. b] → C[a. F is not a contraction if (b − a) > 1. y ∈ X = C[a.b] (b − a)k x−y k! (b − a)k x−y k! ∀ x. This implies that F (x∗ ) is also a ﬁxed point of F N and uniqueness implies that F x∗ = x∗ . Thus there exist a positive integer N such that α = <1 N! N! and F N x − F N y ≤ α x − y ∀ x. 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. we have the following Schauder’s theorem. Then F has a ﬁxed point in K.2 Let F : C[a.4. That is. b] (b − a)k → 0 as k → ∞. For such operators. b]. That is F N is a contraction. it follows that ∃ positive integer N such that k! N (b − a) (b − a)N < 1.68 CHAPTER 2. 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.2 Let K be a nonempty closed convex subset of a normed space X. ∃ x∗ ∈ X such that F N x∗ = x∗ . However. . we get F N +1 (x∗ ) = F (x∗ ) and hence F N (F (x∗ )) = F (x∗ ). MATHEMATICAL PRELIMINARIES Proof : By the contraction mapping principle. Applying F to both sides. F N has a unique ﬁxed point x∗ . giving a ﬁxed point of F. which are not contractions but have compactness property. In the following.
4. x) is monotone increasing for almost all s ∈ [0. − d2 y dt2 . x1 − x2 ) ≥ c x1 − x2 2 for all x1 .9) If F is not deﬁned on the whole space. FIXED POINT THEOREMS AND MONOTONE OPERATORS Deﬁnition 2. x) = − 1 a d2 x x(t)dt dt2 = 0 ( dx 2 ) dt. x) is measurable w.8) F is called strictly monotone if the above inequality holds and is strict for x 1 = x2 .4 A multivalued mapping F : X → 2X is called monotone if (¯1 − y2 .3 F : X → X is said to be monotone if (F x1 − F x2 .r. Further.4. 1] be deﬁned by [F x](s) = f (s. Then we have 1 (F x1 − F x2 . y2 ∈ F x2 ¯ ¯ ¯ ¯ A monotone mapping F is called maximal monotone if it has no proper monotone extensions. b > 0 on f.4. x2 ∈ X 69 (2. 1]. y ) = x ¯ A is monotone as 1 − d2 x dt2 . d2 Example 2. F is called strongly monotone if ∃ constant c > 0 such that (F x1 − F x2 . x1 (s)) − f (s.4. y ∈ D(A) (Ax. Deﬁnition 2. 1] × to be such that f (s. x2 ∈ D(F ) y ¯ ¯ ¯ ¯ ¯ and y 1 ∈ F x1 .4. x for almost all s ∈ [0. it follows that F is monotone.4. x) ≤ a(s) + bx. x ∈ D(A) ¯ dt ≥ 0 Example 2. 1] be the diﬀerential operator − 2 with D(A) deﬁned dt as 2 2 D(A) = {x ∈ L [0. 1].t. (2. 1] : x(t). 1]. x(0) = x(1) = 0} ˙ ¨ A has dense domain and it is self adjoint (A¯.9) on D(F ). x1 − x2 ) = ¯ ¯ ¯ ¯ (2. x1 . x(t) ∈ L [0.4. x2 ∈ X (2.10) As the integrand in the above equation is nonnegative for all x1 .3 A : L2 [0. x2 ) ≥ 0 ∀ x1 . then we verify Eq. Let F : X = L2 [0.t. assume that the mapping x → f (s. 0 (f (s. 1]. then one can show that F maps X into itself and is continuous and bounded (refer Joshi and Bose[6]).4. 1] → L2 [0.r.4 Let f (s. x. x2 ∈ X.2. x(s)) If we assume a growth condition of the type f (s.4.4.Eq. x2 (s)))(x1 (s) − x2 (s))ds . the domain of F . x1 − x2 ) ≥ 0 for all x1 . a ∈ L2 [0. s for all x ∈ and continuous w. x) be a function deﬁned on [0. (2.8) . y(t) = x.
3 Let F : X → X be a strongly monotone continuous operator. x1 − x2 ) = (¯1 − x2 . ((I − F )¯1 − (I − F )¯2 . Then R(F ) = X. x) ¯ ¯ = ∞.5.5 Let F : X → X be a contraction. this corresponds to the condition lim f (x) = ∞ We have the following result due to Minty [9]. ¯ Deﬁnition 2. bounded and coercive.4.5 F : X → X is called coercive if For functions f deﬁned on and lim f (x) = −∞. Then x = F x + y has a unique solution x for every y ∈ X and this solution x varies ¯ ¯ ¯ ¯ ¯ ¯ continuously w. MATHEMATICAL PRELIMINARIES Example 2.4. we shall encounter operators. y. x1 − x2 ) − (F x1 − F x2 .4.4. Theorem 2. x ¯ x→∞ . Theorem 2. Then F x = y has a unique solution for any y ∈ X. Refer Joshi and Bose [6] for the proof.4. concerning strongly monotone continuous operators on X. Theorem 2. Then R(F ) = X where F = F1 + F2 .4. .r. which are only densely deﬁned. x2 ∈ X ¯ ¯ We have the following solvability result due to Browder[3].4 Let F : X → X be continuous. In view of Example 2. 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 . Moreover F −1 is Lipschitz ¯ ¯ continuous and monotone. x→−∞ x →∞ ¯ lim (F x. Then the following theorem gives the surjectivity of F = L + F0 . Then (I − F ) is strongly monotone. monotone and coercive.t. Corollary 2. While dealing with diﬀerential equations. F = L + F0 where L is densely deﬁned linear maximal monotone operator and F0 is monotone operator deﬁned on the whole space.3 Let F : X → X be a contraction on the Hilbert space X.4.5 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. we get the following Corollary.70 CHAPTER 2. For example.
we shall encounter Diracdelta function δ(x). whereas lim ρ (t) = →0 0. t = 0 such that −∞ δ(t)dt = 1 Obviously. this is not a function. 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. 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. . DIRACDELTA FUNCTION 71 2. To specify it. t = 0 ∞. t = 0 This limiting charge is denoted by δ(t).5 DiracDelta Function During the course of our discussion on boundary value problems.2. which is a symbolic function representing the charge density corresponding to a unit charge at the origin. the total charge for the densily distribution ρ (t) on the line is 1. That is δ(t) = lim ρ (t) →0 So.5. vaguely the Diracdelta function δ(t) can be deﬁned as δ(t) = ∞ 0. t = 0 ∞.
2 0.1: Charge density 10 5 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. sup f = {t : f (t) = 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. that is. To be more mathematically rigorous.5. mathematically. For.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. MATHEMATICAL PRELIMINARIES 6 5 4 3 2 1 0. one can take various types of sequences to get the limit δ(t).2 5 10 0.2 0. Deﬁnition 2. yet these sequences can behave diﬀerently at t = 0.5.4 Figure 2.4 0.4 0.5.72 CHAPTER 2.2 0.4 Figure 2.
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. ˜ ˜ Example 2. φ = φ(0).5. Example 2. t < 1 (2.1) conn=1 verges to zero in D( ). 1 2−1 t t≥1 . ∀φ ∈ D( ) It is easy to see that δ is linear and continuous and hence δ ∈ D ( ).5 A sequence of functions {fn }∞ in D ( ) is said to converge n=1 if for every φ ∈ D( ) the sequence of scalars {(fn . This space is a vector space with respect to pointwise addition and scalar multiplication. DIRACDELTA FUNCTION 73 Deﬁnition 2. The action of f ∈ D ( ) is denoted by (f. φ Deﬁnition 2.3 Deﬁne a functional δ on D( ) by δ.5. . φ) = −∞ f (t)φ(t) ∀ φ ∈ D( ) This is called a regular distribution.5.5. φ)}. φ)} converges to the limit {(f.5. (ii) sequence {φn (t)} converges to zero uniformly and for any positive integer (k) k.5. φ). Whereas.4 Dipole distribution δ is deﬁned as (δ. The space of all distributions is denoted by D ( ).1 φ(t) = 0. A distribution which is not a regular distribution is called a singular distribution.2 The sequence φ(t) n ∞ exp with φ(t) deﬁned by Eq. Deﬁnition 2. Example 2. Example 2.6 A locally integrable function f generates a distribution via the deﬁnition ∞ (f.5. One can show that f is also a distribution.5. φ) = φ (0) ∀ φ ∈ D( ) Deﬁnition 2.5.4 A continuous linear function on the space D( ) is called a distribution. the derivatives {φn (t)} also converges uniformly to zero. (2. the sequence t does not converge to n zero in D( ) as these functions do not have the same compact support.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.1) φ(t) is inﬁnite times diﬀerentiable and derivatives of all order vanish for t ≥ 1.5. Deﬁnition 2.5.5. The distribution functional δ is deﬁned as above is called the Diracdelta function.2.
Deﬁnition 2.5. φ(n) ) ∀ φ ∈ D( ) Here f (n) = dn f . φ ) ∞ = − Hence we have H = δ.5. For the proof of this fact refer Bose and Joshi [2] We can deﬁne translation. φ(−1)) = φ(0) = (δ.8 For any distribution f. 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). −φ ) = −(f. φ ) ∀ φ ∈ D( ) By repeated application of the above deﬁnition. H(t) = 1. dtn Example 2.5 The translation δ(t − a) and scale expansion δ(−t) are given by (δ(t − a). φ (t)dt = φ(0) 0 = (δ. φ(t + a)) ∀ φ ∈ D( ) 1 (f (t).5. we get (f (n) . φ(t)) = a Example 2. φ(t)) = (δ.5.6 Let H(t) be the Heaviside 0. φ) = (f. φ(t)) ∀ φ ∈ D( ) Deﬁnition 2. φ(t)) . a = 0 ∀ φ ∈ D( ) (f (at). φ(t)) = −(H.7 The translation f (t − a) and scale expansion f (at) of a distribution is given by (f (t − a). its derivative f is deﬁned as (f . H(t) generates a distribution given by (H(t). 1/2.74 CHAPTER 2. φ(t + a)) = φ(a) and (δ(−t). φ(t/a)). φ(t)) = (δ. MATHEMATICAL PRELIMINARIES Theorem 2. scale expansion. derivative of a distribution as follows.1 The Diracdelta distribution is a singular distribution. φ) = (−1)n (f. φ(t)) = (f (t).5.
5. . .8 If g(t) is inﬁnitely diﬀerentiable. . wherever it exists. φ) = ( k=0 an−k (t)f (k) . . φ) dk (an−k (t)φ(t)) dtk = (−1)k f. We assume that f is locally integrable wherever it exists. we get df φ(t)dt + dt n a2 a1 df φ(t)dt + . an where f (t) jumps by the respective amount f1 . .5. φ ) = − = − − ∞ f (t)φ (t)dt −∞ ∞ −∞ ∞ an a2 f (t)φ (t)dt − f (t)φ (t)dt f (t)φ (t)dt . . DIRACDELTA FUNCTION 75 Example 2. L∗ φ) . . .5. f2 . + an−1 (t) + an n dt dt dt dk an−k (t) k dt with coeﬃcients ak (t) inﬁnite times diﬀerentiable. then g(t)f (n) (t) can be found as follows (gf (n) (t).2. . fn . φ) = −(f. . . gφ) = (−1)n (f. Example 2. Here [f ] indicates the usual derivative of f. we have n (Lf.7 Let f (t) be diﬀerentiable except at points a1 .9 Let L be a linear diﬀerential operator given by L = a0 (t) n = k=0 dn dn−1 d + a1 (t) n−1 + . .5. Then for the distribution f. a1 Integrating by parts and noting that f (a+ ) − f (a− ) = k k (f . φ) = −∞ ∞ fk . . dt + an df φ(t)dt + dt ∞ fk φ(ak ) k=1 n Therefore (f . = (f. φ) = −∞ n [f ]φ(t)dt + k=1 fk φ(ak ) This gives f = [f ] + k=1 fk δ(t − ak ). Then the distributional derivative f is computed as follows ∞ (f . dn (gφ)) dtn Example 2. φ) = (f (n) . .
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 . 0. However. This sequence does not converge point wise except at t = 0. t ≥ 1 n Example 2. Example 2. t < 1 2 n ∞ 1 n (fn . dtk The nth order diﬀerential operator L∗ is known as the formal adjoint of L. φ) = n→∞ lim n 2 1/n φ(t)dt −1/n = φ(0) = (δ. φ) Thus the sequence of regular distribution {fn } converges in D ( ) to the delta distribution whereas the corresponding functions converge almost everywhere to zero. MATHEMATICAL PRELIMINARIES (−1)k k=0 where L∗ φ = dk (an−k (t)φ(t)). As it is locally integrable. we have ∞ ∞ (sin nt. 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.5. it generates a sequence of distributions. if it exists. Test the convergence of the following sequences of the functions in the space L2 [0. Compute the limit function. 1].11 Let fn (t) = n . (a) xn (t) = sin nπt .76 n CHAPTER 2. 2.6 Exercises 1.5.10 Consider the sequence {sin nt}n=1 .
x2 . Determine if the function F : F (x1 . 5. 1] → L2 [0. 4. x3 ) = 3 1 sin2 x1 + cos x2 cos x3 . Let L : L2 [0. ﬁnd its adjoint L∗ and show that L is a singular operator. x2 ) 2 (x1 . t2 . Find an element in M closest to cos 2t ∈ L2 [−π.6. 0) (b) G(x1 . show that L is compact. a Hilbert space ) be diﬀerentiable. is a test function. cos t . respectively. cos x1 sin x2 + x1 x2 ) (x1 . 10 cos x1 cos x2 cos x3 ) w. 1]. x ¯ 0≤t≤a t ≤ 0. x2 ) = (0. (a) F (x1 . π] be the linear span of 1. π]. 1] be deﬁned by 1 [Lx](t) = 0 (s + sin t)2 x(s)ds From the ﬁrst principle. .2. x2 x1 exp(x1 . Let A ∈ n×n and let N (A) and R(A) denote the null space and range space of A. 0) 7. x2 ) = exp(x1 x2 ) + cos x1 sin x2 . (X. x2 ) = (0. 0. Also. Linearize the following nonlinear operators around the given points. Show that n = N (A) ⊕ R(A) 3. Let M ⊂ L2 [−π. 6. x2 ) = (x1 + sin x1 cos x2 . sin x2 cos x3 . or t ≥ a exp(−(t−2 ))exp(−(t − a)−2 ). EXERCISES (b) xn (t) = tn (c) xn (t) = t sin nπt sin nπt (d) xn (t) = n 77 2. Show that [Lx](t) = 0 t x(s)ds is a bounded linear operator in L2 [0. Show that Ψ(t) = 3 . Show that f is f (¯) is monotone for all x ∈ X.r.t 1norm in 3 → 3 deﬁned as is a contraction in 8. Let f : X → convex iﬀ x → ¯ 9.
. [9 ] G. Royden. MATHEMATICAL PRELIMINARIES (b) δ (b − a) = (c) (a) g(t)δ (t) = g(0)δ (t) − g (t)δ(t). 1988. [6 ] Mohan C Joshi and Ramendra K Bose. Real Analysis. 1981. Krasnoselskii. 1964. 1993. The Solvability of Nonlinear Functional Equations. Pergaman Press. Joshi.. D. Linear Algebra and Linear Models.. DukeMath Journal 30. 1985 [7 ] M. Tata McGraw Hill Publishing Co. Topological Methods in the Theory of Nonlinear Integral Equation. Hindustan Book Agency. C. [4 ] S. 1962. 29. Show that CHAPTER 2. Elemenatry Numerical Analysis . [3 ] F. A. Some Topics in Nonlinear Functional Analysis. [10 ] H. K. 341346. 1974. Prentice Hall of India. 1963. McGraw Hill Book Co. Minty. 1996. 557566. Functional Analysis. Browder. New Age International. [5 ] Casper Goﬀman and George Pedrick. [2 ] R. First Course in Functional Analysis. Monotone (Nonlinear) Operator in a Hilbert Space. Macmillan Publications. 1984. g ∈ C ∞ d (sgnt) = 2δ(t) dt δ (b − a)δ(b − a)dt References [1 ] R B Bapat. Duke Math Journal. Wiley Eastern Ltd. Methods of Mathematical Physics. [8 ] Balmohan V Limaye. Conte and C. deBoor.78 10. E.An Algorithmic Approach. L. Bose and M. Ltd.
1). by Initial Value Problem (IVP). As deﬁned earlier (Deﬁnition 1. if ∃ an interval J ⊂ I. y) ≤ M x − y ∀ t ∈ I and ∀ x.1). (3.are the prime targets of our investigation. x(t)) = x0 (3. containing ‘t0 ’ such that x(t) is diﬀerentiable on J with x(t) ∈ Ω. x) − f (t.1a) (3. 79 (3.1. Let f : I × Ω → n be a function.Chapter 3 Initial Value Problems This chapter primarily deals with the existence and uniqueness theory of initial value problems.1.1.1) has a unique solution.1. maximal interval of existence. (3. Further assume that x → f (t. continuation of solution and continuity of solution with respect to initial condition .2) . not necessarily linear. ∀ t ∈ J and x(t) satisﬁes Eq. requiring Lipschitz continuity on f (t.1). x(t) is a solution of Eq.1.1 Let E = I × Ω ⊆ × n be an open set containing (t0 . we mean the following diﬀerential equation with initial condition dx dt x(t0 ) = f (t. x0 ) ∈ I × Ω. x) with respect to x.1. x) is Lipschitz continuous : ∃ a constant M such that f (t.1.1. let us recall what we mean by the solution of the IVP . The following theorem gives the existence of unique solution of Eq.Eq.1). The notions of . Theorem 3. (3.1. y ∈ Ω Then the IVP given by Eq.1 Existence and Uniqueness Theorems Let I be an interval in and Ω an open set containing (t0 . (3. x0 ) and let f : E → n be continuous.1.1b) Also. 3.Picard iterates.1). (3.
x0 ) and contained in E. x(s))ds.1.1.3) in some appropriate subinterval J ⊂ I. (3.4) It is clear that F x is a continuous function deﬁned on J. solvability of Eq. Further. x ¯ t F x(t) − F y(t) n ≤ t0 f (s. we have. x). t [F x](t) − x0 n = t0 t f (s.1) is equivalent to the solvability of the following integral equation t x(t) = x0 + t0 f (s. x(s))ds f (s.80 CHAPTER 3. Also. x(s))ds (3. t0 + τ ] and Ω0 = { x ∈ Ω : x − x0  ≤ δ }.1. We deﬁne a nonlinear operator F on X as follows t [F x](t) = x0 + t0 f (s. ‘m’ is ﬁnite as f is continuous on a compact set B ⊆ I × (t. x(s)) − f (s. y ∈ X . Choose two positive numbers ‘δ’ and ‘τ ’ such that M τ < 1.1. Deﬁne X ⊂ C(J) to be set of all continuous functions on J with values in Ω0 ⊆ Ω. ∀¯. x) ∈ I × ¯ n : t − t0  ≤ τ. y ∈ X. INITIAL VALUE PROBLEMS Proof : As shown in Proposition 1. 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.x)∈B n . Denote by J = [t0 − τ.1. Then X is a closed subset of C(J) and hence a complete metric space with metric induced by the supnorm of C(J). Let ¯ ¯ m = sup f (t. 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. mτ < δ and { (t. x − x0  ≤ δ } ⊆ B. t ∈ J (3. Let B denote a closed ball with centre at (t0 .
(3. x(t)) dt x(0) = x0 (3.3) and hence that of Eq.4.1.1. This proves that x∗ (t) is unique solution of Eq.1. . (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. · · · . (3. (3.1.3). x2 =x1 =x. Invoking Theorem 2. x(t)) (3. Ω0 ) such that ¯ x∗ (t) = F x∗ (t) = x0 + t f (s. .4). converge to the unique solution x∗ (t) of IVP given by Eq. x ) Lipschitz continuous with Lipschitz constant M .1.1 The iterations x(k) (t). Further. x(0) ∈ X.1. set . .1.3. Ω0 ). The iterates deﬁned by Eq.1. Proof : We shall reduce Eq. We note that x(0) (t) = x0 ∈ X = C(J. For Proceeding as in section (1.7b) . = f (t. Proof : This follows from Theorem 2. ∃ x∗ ∈ X = C(J..1.6) has a unique solution. x(1) (0) = x2 .5) are called Picard iterates.1.1. deﬁned as t x(k) (t) = x0 + t0 f (s. x3 =x2 =x1 =x.1. it follows that F has a unique ﬁxed point x ∗ ∈ X.4. assume that ai (t).6) reduces to the following IVP on I × Ω × Ω · · · Ω ⊆ I × n dx = F (t. Corollary 3.1.7a) (3.1. wherein the converging iterates are deﬁned as x(k) = F x(k−1) . . it follows that F is a contraction on the complete metric space X.1. Corollary 3. xn =xn−1 n .1. . (3.1).1). x(k−1) (s))ds (3.6a) (3.6b) x(0) = x1 .6) in to an initial value problem in this. 1 ≤ i ≤ n − 1 are continuous functions on I.2 Let Ω ⊆ be open and f : I × Ω → be continuous with x → f ( . Then nth order initial value problem given by Eq. · · · .1. x1 = x. x(n−1) (0) = x(0) n Here x(k) (t) refer to the k th derivative of x(t).1. (3. x∗ (s))ds t0 and x∗ is unique. .. the IVP given by Eq. That is.5) with x(0) (t) = x0 . EXISTENCE AND UNIQUENESS THEOREMS 81 As α = τ M < 1. (3. (3.
.1 ≤ α2 x − y2 n dx = x2 . 1 + na2 1 2 n−1 . xn . 1 ≤ i ≤ n.1.1. x3 . satisﬁes all conditions of Theorem 3. x(k−1) (s))ds x(k−1) (s) 2 = 1+ 0 ds So.1.. t∈I Hence we have. y) 2 n Thus the function F appearing in the IVP given by Eq. x2 . This in turn implies that the nth order diﬀerential equation given by Eq. bounded subset of . x(0) = 1 dt Clearly. The Picard’s iterates x(k) (t) are given by x(k) (t) t = x0 + 0 t f (s. f (t. F (t. y1 )2 + a1 (t)2 x2 − y2 2 ≤ nM 2 x1 − y1 2 + (1 + na1 (t)2 )x2 − y2 2 + · · · + 1 + nan−1 (t)2 xn − yn 2 ≤ α2 x1 − y1 2 + x2 − y2 2 + · · · + xn − yn 2 . x2 .1. x) is Lipschitz continuous as we see below. we get the following Picard iterates for the above problem.1 and hence it has a unique solution. x1 ) − a1 (t)x2 − · · · − an−1 (t)xn ) . 1 + na2 . x0 = x1 . Hence by Theorem 3. (3. 1 + na2 . It is clear that F is continuous on I × Ωn and further x → F (. · · · . . x(1) (t) = 1 + t x(2) (t) = (1 + t + t2 ) + t3 3 .1 this IVP has a unique solution. the function f (x) = x2 is Lipschitz continuous on a closed. · · · . xn ). x) − F (t. . x) = (x2 . where α2 = max nM 2 .6) has a unique solution. with ai = sup ai (t).82 CHAPTER 3. Example 3. (n) x (t) = (1 + t + t2 + · · · + tk ) + o(tk+1 ) . (3. x) − F (t. . .1. y)2 n ≤ x2 − y2 2 + x3 − y3 2 + · · · + xn − yn 2 + · · · + an−1 2 xn − yn 2 + n f (t. INITIAL VALUE PROBLEMS (0) (0) (0) Here x = (x1 . . · · · . xn ∈ n and ¯ ¯ F (t. x1 ) − f (t. F (t.7).
1 is plotted in Figure 3.5 0. The largest interval on which 1−t this solution is deﬁned in (−∞. 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).2. as envisaged in the Corollary 3.75 t Figure 3. The actual solution x(t) = 1−t 4 x3 3 x2 x1 1 2 1 0.5 0.1.25 0.3.1: First three Picard iterates . (refer Chapter 6).1.25 0. For small t . The following graph gives ﬁrst three Picard iterates.75 0.1.1. the function x(t) has the series expansion given by x(t) = 1 + t2 + t3 + · · · + tn + · · · . 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) = .
25 0.1. Remark 3.5 0.1.84 CHAPTER 3. This point will discussed in ¯ detail in section 3. x(0) = 0 dt where f (x) is given by f (x) = √ 2 x. 0. the solution is not unique.1. .2: Actual solution x(t) = 1 1−t Example 3. and hence f is not Lipschitz conx x x tinuous at ‘0’.75 Figure 3.5 0.1) without Lipschitz continuity assumptions on the nonlinear function f (t. x≤0 f (x) 2 f (x) = √ .1. INITIAL VALUE PROBLEMS xt 10 8 6 4 2 t 1 0. x). However. x > 0 0. (3.1 This example concludes that there may exist a solution of IVP of the type Eq.1. by inspection one can easily compute the following solutions of above IVP : As x1 (t) = t2 .2 dx = f (x). However it may not be unique.25 0. t>0 t≤0 ∀t∈ x2 (t) = 0 That is.75 0. Hence Theorem 3. x > 0. → ∞ as x → 0+ .4.1 is not applicable. the above IVP has at least two distinct solutions or in other words.
0) is given by the following curve in (t. x) = 0 dx dt 85 (t. EXISTENCE AND UNIQUENESS THEOREMS Example 3. x) 2tx 2 t + x2 f (t. x) = (0.3: Solution curve for the above example It may noted that the above IVP has a unique solution passing through the point away form (0. .1. Yet. 0). x)plane (x − c)2 − t2 = c2 .8) xt 10 8 6 4 2 6 4 2 2 2 4 6 t 4 Figure 3. 0) (t. 0).3.1.1. x) → f (t. 0). c is arbitrary constant The solution curve is plotted as under.1. x) is not even continuous at (0. It can be easily seen that (t. of the above diﬀerential equation.3 = f (t. (3. x) = (0. 0) The solution of the above diﬀerential equation passing through (0. there are inﬁnitely many solutions passing through (0.
t2 t6 t10 − + 2 240 34560 t9 t5 (4) + x2 (t) = t − 40 3456 (4) x1 (t) = .2 that the above second order initial value problem has a unique solution. x(t) = (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.4 CHAPTER 3. therefore. x2 (t)) ¯ d¯ x = F (¯(t)). x2 (t) . 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 . One can.1.86 Example 3. = 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). x(0) = 0 ˙ dt2 As f (x) = cos x is Lipschitz continuous. a close form solution of the above problem is diﬃcult to obtain. INITIAL VALUE PROBLEMS d2 x = cos x. However. x1 (t) = x(t). it follows by Corollary 3.1. F (x) = (x2 . cos x1 ). x2 (t) =x. resorts to Picard iterative procedure to obtain iterates x(k) (t) converging to the unique solution of the above second order IVP. x(0) = 0.
Using the notion of monotonicity of Chapter 2 (Deﬁnition 2. x1 (t) xt 3 2 1 x22 (2) (4) x24 1 2 3 t 3 2 1 1 2 3 Figure 3. x1 − x2 ) ¯ ¯ ≥c n (2) (4) ≥ 0 ∀ x 1 .1.9) It is said to be strong monotone if ∃ c > 0 such that n x 1 − x2 ¯ ¯ n ∀ x1 . EXISTENCE AND UNIQUENESS THEOREMS xt 87 4 x12 3 x14 2 1 3 2 1 1 2 3 t Figure 3.1.10) . x2 (t) There do exist examples where the nonlinear function f (t.5: Graph of iterates x2 (t).1.4: Graph of iterates x1 (t).3. x) is monotone in n if ¯ ¯ (f (t.1. x) under consider¯ ation is not Lipschitz continuous but satisﬁes other nice property like monotonicity with respect to the variable x. x2 ∈ ¯ ¯ n (3. x1 − x2 ) ¯ ¯ ¯ ¯ (f (t. x2 ∈ ¯ ¯ 2 n (3. x1 ) − f (t.1) under the monotonicity assumption. x1 ) − f (t. (3. x2 ).1. we shall say that x → f (t.4. We shall now state and prove the existence ¯ and uniquness theorem for IVP of the type given by Eq.1.3). x2 ).
2 Let f (t.1. (3. one can show that L is maximal monotone as R(L + λI) = X for λ > 0 (refer Pazy[6]). tf ] 2 Further.1.1 A function x(t) ∈ L2 [0. b > 0 (3.12) in the interval I = [0. (3.88 CHAPTER 3. 0 d¯ x dt dt n x ¯ = ¯(tf )2 n − (L¯.1. tf ] satisfying Eq.1) in ¯ almost everywhere (a. x) = 0 dt x(0) = 0 (3. tf ]. x) n ≤ a(t) + b x n .1. (3.1.1) for all t ∈ I except on a set of measure zero in I. ¯ Deﬁnition 3.11) Then there exist unique x(t) ∈ L2 [0. Theorem 3. Proof : We are interested in the solvability of the IVP of the form dx + f (t. x) = x ¯ x(tf ) ≥ 0 ∀ x ∈ L2 [0. x) x This implies that (L¯. x) = (Lx.1. in almost every where sense. tf ] is said to satisfy Eq. t f ] = I and ¯ satisﬁes Eq. tf ]. x(t)) ¯ (3. x) : [0. We now deﬁne the nonlinear operator N on X as in Example 2. tf ] (refer Joshi and Bose .1). Here L2 [0.4: [N x] = f (t. Then . tf ]. Deﬁne L as L¯ = x dt dt D(L) is dense in X and L is monotone as tf (L¯.4. (3. continuous operator on X = L2 [0.13) N is a bounded.tf ] x ¯ ¯ = 0 d¯ x .1. tf ] × n → n be continuous and x → f (t.1.1) in almost every¯ where sense. x) ¯ be strongly monotone and further it satisﬁes a growth condition of the form f (t.1. x)L2 [0.e) sense if x(t) is diﬀerentiable for all t ∈ [0. Set X = L2 [0. INITIAL VALUE PROBLEMS In the IVP given by Eq. a ∈ L2 [0. 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(t) dt tf dt n = x(tf )2 n − x(t). we assume that t0 = 0 and x0 = 0.
Hence by Theorem 2. Theorem 3.1 Let f satisﬁes Assumption 1. N is a bounded.1.1 is not applicable.2.5.1. (3. x1 (t)) − f (t.1. Let f : × Ω → n satisﬁes the assumptions of Theorem 3.1.5 dx + αx + cos x = 0 dt x(0) = 0 f (x) = αx + cos x is strongly monotone if α > 1.1) on the intervals I1 . x1 − x2 ) = ¯ ¯ ¯ ¯ ≥ c 0 (f (t. L : D(L) → X is a linear maximal monotone operator with dense domain. In this section we shall show that the solution x(t) is deﬁned on the maximal interval of existence (α. (¯1 (t) − x2 (t))) dt ¯ ¯ x ¯ tf 0 x1 (t) − x2 (t)2 dt = cx1 − x2 2 for all x1 .1. x).12). Assumption 1: Let Ω ⊆ n be an open set containing x0 .1.1.2 The Maximal Interval of Existence Theorem 3. it is strongly monotone as we have tf 89 (N x1 − N x2 .1. (3. (3.1.1.12) reduces to the solvability of the operator equation (3. tf ].2 that the above initial value problem has a unique solution x(t).4.3. y(t) are solutions of the IVP given by Eq.1 Theorem 3. it follows that Eq. However.2.1. ¯ ¯ Thus the solvability of equation Eq. the IVP of the form Eq. Hence it is follows by Theorem 3. x2 (t)).14) Lx + N x = 0 in the space X = L2 [0. If x(t).e. in almost everywhere sense in X.1. Then t0 ∈ I1 ∩ I2 and if I is any open interval around ‘t0 ’ contained in I1 ∩ I2 . (3. 3. Example 3.1) has a unique solution x(t) on some interval surrounding t0 . (a. then x(t) = y(t) ∀ t ∈ I. THE MAXIMAL INTERVAL OF EXISTENCE [4]). ∞). . I2 .14) has a unique solution and hence the unique solvability of Eq. for t) in a ﬁnite subinterval of [0. Further. β). (3.2 establish the fact that under suitable continuity and Lipschitz continuity amd monotoncity assumptions on f (t.1 and Theorem 3. continuous strongly monotone operator deﬁned on the whole space X. x2 ∈ X.
then one of the end points t∗ of I ∗ is contained in I ⊂ I1 ∩ I2 . The intervals [t0 .1. If t0 ∈ I ⊂ I1 ∩ I2 . if t ∈ I1 ∩I2 . (3. then t0 ∈ (α.1) for t ∈ Ia . (3.1. x(t∗ ) = x∗ ¯ dt (3. x(t) = xa (t).1. then by Theorem 3.1) as x a (t) is a unique solution of Eq. ∃ some open interval Ia = (a − t0 .1) Eq. Without loss of generality we can concentrate on the right maximal interval of existence.1.1. the maximal interval J is open. t ∈ Ia This is well deﬁned. As both x(t) and y(t) are continuous on I. Also let the maximal interval of existence be not open say of the form (α. 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). t∗ + a∗ ) ⊂ I.2.2.1. β) is the maximal interval of existence for IVP given by Eq.2.1).1) on (α.1.1) has a unique solution. We wish to ﬁnd out if the right maximum interval of existence [t0 .2. Proof : As in Theorem 3.1) has a unique solution.1. (3. (3.2 Let f satisﬁes Assumption 1. The following theorem is an important one in that direction. it follows that. Furthermore. x(t) = y(t) on Ia ⊂ I ⊂ I 1 ∩ I 2 . (3. (3.1) has a unique solution xa .1) has a unique solution on some interval I0 = (t∗ − a∗ .1. If I ∗ is properly contained in I. . Then by Theorem 3. ∃ a maximal interval I ∗ on which the IVP given by Eq. β). β). Then I ∗ is the largest open interval contained in I on which x(t) = y(t). Deﬁne x : I ∗ → Ω as follows. (3. Then for each x0 ∈ Ω. Theorem 3. ∞). where I1 and I2 are two intervals on which Eq. (3. β) and (α.1 it follows that x1 (t) = x2 (t).1.1) has solutions x1 (t) and x2 (t). a + t0 ) ⊂ I such that Eq. β) can be extended to [t0 . Also x(t) is a unique solution of Eq. In view of the uniqueness of solutions on Ia .2. Hence I ∗ = I and thereby implying that x(t) = y(t) ∀ t ∈ I. Deﬁne I ∗ = ∪Ia . If (α. β + a] with a > 0 as in the proof of Theorem 3. (3.90 CHAPTER 3. that is I ∗ = (α. (3. I ∗ is maximal. This will then lead to global solvability of IVP given by Eq. for. INITIAL VALUE PROBLEMS Proof : It is clear that t0 ∈ I1 ∩ I2 .1. on which Eq.2. Again by uniqueness on I0 ⊂ I ⊂ I1 ∩I2 . By deﬁnition. β] then we can extend the solution of Eq. it follows that. we deﬁne I ∗ = ∪Ia where Ia is an open interval surrounding t0 . lim∗ x(t) = lim y(t) = x∗ ¯ ∗ t→t t→t Now examine the initial value problem dx = f (t.1. t0 ] are called the right and the left maximal intervals of existence respectively.1. x(t)).1).
β) such that x(t) ∈ K. / As a Corollary of the above result. Hence. it follows that x(t1 ) − x(t2 ) → 0 which by Cauchy’s criterion of convergence in n implies that lim x(t) exists. Since f is continuous on (α. β] t ∈ [β. y(s))ds By Theorem 3. Let (α. We now examine the IVP dx = f (t. z(t).1. (3. β)×K If possible. (3.1. as y(t) satisﬁes the ¯ integral equation t y(t) = x0 + t0 f (s.1). x) < ∞ (t.2. β] by x(t). For α < t1 < t2 < β.2 with x0 ∈ Ω. Let x1 be this limit. x(s)) ds t1 ≤ m t 2 − t1 As t1 .1.2. t2 → β. / Proof : Let K be compact set in n .1. Assume that β < ∞. we immediately get the following theorem. Now we deﬁne a function v(t) as v(t) = y(t). we have t2 x(t1 ) − x(t2 ) ≤ f (s. let the solution x(t) of Eq. t ∈ (α. Then given any compact set K ⊂ Ω.1 the equation (3. The process of continuation is already described in the proof of Theorem 3. THE MAXIMAL INTERVAL OF EXISTENCE 91 Theorem 3.1) be such that x(t) ∈ K ∀ t ∈ (α. β) be the maximal interval of existence of the solution x(t) of the IVP given by Eq. if β < ∞. . Now deﬁne y(t) on (α. β] with y (β) = f (β. β].2. t ∈ (α.2) has a unique solution z(t) on some open interval (β − a. β + a) Thus y is a continuation of the solution x(t) on (α. β) with β < ∞.2. β) × K. β + a) and hence z(t) = y(t) on (β − a. contradicting the assumption that (α. t=β ˙ y(t) is diﬀerentiable on (α. y(β)).x)∈(α. ∃ t ∈ (α. β + a).2.3.1) on (α.2. It follows that m= sup f (t. β) y(t) = x1 .3.2) Then v(t) is a solution of the IVP given by Eq. ∃ a t ∈ (α. x(t)) dt x(β) = x1 ¯ ¯ (3. (3. called continuation theorem. β) such that x(t) ∈ K.3 Let Ω and f be as deﬁned in Theorem 3. it follows that x1 ∈ K ⊂ Ω. Since t→β− x(t) ∈ K and K is compact. β) and at the end point β we have z(β) = x1 = y(β). β) is the maximal interval of existence for given IVP.
1).Then the initial value problem given by Eq. x(t)) dt Taking scalar product with x(t) we get x(t). (3. f be as deﬁned in Theorem 3. Assume that ∃ a compact set K ⊂ Ω such that y∈ n y = x(t) for t ∈ [ t0 .2 Let x → f ( .1. Let [t0 . (3. Then β = ∞. dx dt = (f (t. Corollary 3. x(t)).1) has a global solution. x) n n and satisﬁes the following condition: >ρ ≤ a(t) x ∀ x n with a(t) as a continuous function.4 Let Ω.1 Suppose Ω = ∃ ρ > 0 such that f (t. Corollary 3.2.1. β) ⊂ K. . x(t) is a global solution.92 CHAPTER 3. x) be Lipschitz continuous for all x(t) ∈ Then Eq.2.1) has a global solution. we again apply the same theorem to K = x : x2 ≤ x(0)2 e t0 .2.1. That is. n .. 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. then we apply the previous theorem to K = {x : x ≤ ρ} and get the result. That is. Proof : This follows from the previous corollary. Proof : We have dx = f (t.1) has a solution x(t) on [t0 . β) be the right maximal interval of existence of solution x(t) of the IVP given by Eq. x(t) ) n ≤ f (t. If x(t) > ρ. (3. INITIAL VALUE PROBLEMS Theorem 3. ∞).2.IVP given by Eq.3 with x0 ∈ Ω. dx dt = x(t).1. (3.
we get dx .1. x(t)).2. Lemma 3.1) has a global solution. x(t)) ¯ (3. = 2 (f (t. we have g(t) ≤ c exp (kt) ∀ t ∈ [t0 .) Then. x) : I × n → n be continuous with x → −f (t. x(t))) = 0 dt As −f strongly monotone.1. t1 ] . x) strongly monotone and f (t. (3.1) dt on the initial condition x(t0 ) ∈ n .3 Let 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)}. refer Perko[ 7 ]. k are positive constants.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 . DEPENDENCE ON INITIAL CONDITION 93 ¯ ¯ Corollary 3.1). Proof : We have a solution x(t) of Eq. We ﬁrst state and prove the following lemma ¯ called Gronwall’s lemma.1.2.4. (3.3.3. For more details on this section. 0) = 0.1) satisfying dx − f (t. x(t)). (3. implies that x(t) is a global solution of Eq. t1 ] (c.3. Then the IVP given by Eq. x(t)) ≤ −2cx(t)2 3.3 Dependence on Initial Condition In this section we investigate the dependence of the solution of the diﬀerential equation d¯ x = f (t.3. x(t)) = 0 dt Taking scalar product with x(t). Hence Theorem 3. x(t) − (f (t.
1 Let f (t. G(t) ≥ g(t) and G(t) > 0 on [t0 . (3. t1 ]. 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 (3.1) depends continuously on the ¯ initial condition x0 .1.94 Proof : Let G(t) = c + k t1 t0 CHAPTER 3.1.1). 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. x(s)) − f (s. t1 ]. with respect to the ¯ ¯ initial conditions x0 and y0 .3. By second fundamental theorem of calculus ˙ G(t) = kg(t) and hence ˙ G(t) g(t) =k ≤ k ∀ t ∈ [t0 . Using Lipschitz continuity of f and Gronwall inequality.1. y ∈ X = ¯ ¯ C(J. Then. t ∈ [t0 . we have x.1. This gives Theorem 3. y (t) be unique solutions of Eq. ¯ Proof : Let x(t).2) n ≤ ¯0 − y0  x ¯ n exp (M t) ∀t ∈ J = [t0 . respectively. INITIAL VALUE PROBLEMS g(s)ds. x(s)) ds ¯ f (s. Then the ¯ solution x(t) of the diﬀerential equation (3. Using the corresponding integral equation analog we have t x(t) ¯ y(t) ¯ = x0 + ¯ t0 t f (s. tf ] (3. Denote by J the interval of existence ¯ ¯ of solutions.1.3.1. t1 ] which is the required result. t1 ] G(t) G(t) d (log G(t)) ≤ k dt Equivalently G(t) ≤ G(t0 ) exp (kt) ∀t ∈ [t0 . y(s)) ¯ ¯ n ds Borrowing our earlier notation of Theorem 3. Ω0 ⊂ Ω. x) satisfy all conditions of Theorem 3.3) .3. Ω0 ).
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
CauchyPeano 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 AscoliArzela 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 (AscoliArzela 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 (CauchyPeano 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
≤ mt1 − 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 AscoliArzela 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
<δ
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 CauchyPeano 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
= xx = 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(−2kt − t0 ) ≤ u(t) ≤ u(t0 )exp(kt − 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 supaj (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,
Verlag. SpringerVerlag. . M. I. Molenaar. [5 ] R. Mattheij and J. K. Consider the following perturbed IVP (corresponding to the above problem) dy = ax. Ordinary Diﬀerential Equations. Watter de Gruyter. Ordinary Diﬀerential Equations : Introduction to Nonlinear Analysis. [8 ] Walter Rudin. Geometrical Methods in the Theory of Ordinary Diﬀerential Equations. [7 ] Lawrence Perko. Bose. [2 ] V. 1996. Diﬀerential Equations and Dynamical Systems. 1990. [4 ] Mohan C Joshi and R. for more than one initial guess x0 (t). Springer . Arnold. John Wiley. 1983. Birkauser.100 9. y(0) = 1 + dt Show that x(t) − y(t) ≤  eat Graphically. show that the above estimate is very accurate for a > 0. Semigroup of Linear Operators and Applications to Partial Differential Equations. 1991. 1974. 1982. but very inaccurate for a < 0. Pazy. Halsted Press. Real and Complex Analysis. 1983.Verlag. Ordinary Diﬀerential Equations in Theory and Practice. M. INITIAL VALUE PROBLEMS dx = ax. 1985. 10. Springer . [3 ] Philip Hartman. x(0) = 1 dt the Picard iterates xn (t) converge to the unique solution eat . References [1 ] Herbet Amann. Tata McGraw Hill Publications. Some Topics in Nonlinear Functional Analysis. [6 ] A. Show that for the IVP CHAPTER 3.
1) of Chapter 3.1. as continuous functions of t.1. We give the deﬁnition of transition matrix. (4.1(a)) follows from Theorem 3. Here A(t) is an n×n matrix and b(t) is a vector in n . (4.1.Chapter 4 Linear System And Transition Matrix The focus of this chapter is on the linear system of diﬀerential equations. introduced earlier in Chapter 1. (4. with entries aij (t) and bi (t) respectively. (3.1. 101 .1. wherein the concept of transition matrix plays a crucial role.1. 4. it helps to study the homogeneous system dx = A(t)x(t) dt (4. To study the solution of Eq.2) We shall ﬁrst prove the existence of n linearly independent solutions of Eq.1) with the initial condition x(t0 ) = x0 (4. The existence and uniqueness of global solution of Eq.2).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. its properties and the method of computation.1) This is a special case of Eq.1. are recalled for demonstrating the methodology involved in computation.1.1).1.
. · · · .1. Theorem 4. + + ··· + cn x1 (t0 ) (n) + cn x2 (t0 ) . (4.1.1. dt x(t0 ) = e(i) . . . . Then x(1) (t). . Proof : If possible. This implies that there exist n constants c1 . (1) (1) (2) (1) (2) x1 (t) x2 (t) . x(2) (t). .2) in J.1 Let x(1) (t).4) is a linear system of equations in nunknowns c1 . x(n) (t) be linearly dependent. .1.2 There exists n linearly independent solutions of Eq. Hence the set x(1) (t). The Wronskian W (t). .3) + ··· + ··· . Proof : We consider n initial value problems dx = A(t)x(t).1. = 0 (4. xn (i) (i) (i) .4) Eq. it follows that ci = 0 for 1 ≤ i ≤ n (as RHS of Eq. + ··· (1) (2) c1 xn (t0 ) + c2 xn (t0 ) (1) (2) (n) (4.. x2 . c2 .1. · · · . · · · .1. x(2) (t). As the determinant W (t0 ) of this system is nonzero.4) is zero). x(n) (t) is linearly independent.1 Let W (t0 ) = 0 for some t0 ∈ J ⊆ . is deﬁned by x1 (t) x1 (t) · · · W (t) = x2 (t) x2 (t) · · · .3) implies that c1 x1 (t0 ) + c2 x1 (t0 ) (1) (2) c1 x2 (t0 ) + c2 x2 (t0 ) . x(n) (t) be nvector valued functions.102 CHAPTER 4. x(2) (t). · · · . xn (t) (n) (n) (n) ··· (2) . · · · . c2 . Theorem 4. . · · · . cn (not all zero) such that ¯ c1 x(1) (t) + c2 x2 (t) + · · · + cn x(n) (t) = 0 Eq. (4. . . . (4.1. xn (t) xn (t) · · · Here x(i) = x1 .1. · · · . This is a contradiction. 1 ≤ i ≤ n. + + cn xn (t0 ) (n) = 0 = 0 . . · · · . t∈J 1≤i≤n where e(i) are the unit vectors (0. 0) . · · · . This gives us the following theorem. x(n) (t) are linearly independent in the interval J. let x(1) (t). 1 . (4. . cn . LINEAR SYSTEM AND TRANSITION MATRIX Deﬁnition 4.
x(t0 ) = x(0) dt . .1. Let x(t) be any general solution of Eq. x2 (t) . . x(2) (t).1. .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). . x2 . . The following theorem gives the dimension of X. 0 0 ··· 0 0 . .1. with initial conditions x(i) (t0 ) = e(i) . dy = A(t)y(t) dt Also.2).1.1. 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 . (4. .2 that Eq. xn (t) (2) (2) ··· . the above set of n initial value problems has n unique solutions x(1) (t). . Let (1) (2) (n) x1 (t) x1 (t) · · · x1 (t) x2 (t) W (t) = . .1.1. xn (t) (n) (n) ··· W (t0 ) = = 1 By Theorem 4. x(n) (t) are linearly independent on J. · · · .2) has n linearly independent solutions x(1) (t). . Proof : We have already proved in Theorem 4. .. · · · . · · · . x(1) (t). (4. .4. x(n) (t). .3 The dimension of the solution space X of the linear homogeneous system given by Eq. . 1 (1) (1) x2 (t) . (4. xn y(t) = (0) x1 x(1) (t) (0) (0) (0) . Let us assume that x(t0 ) = x(0) .2) is n.1. . LINEAR SYSTEM 103 By Theorem 3. Consider the function y(t) deﬁned as It is clear that + (0) x2 x(2) (t) +···+ (0) xn x(n) (t). with x(0) = x1 .1. x(n) (t) with x(i) (t0 ) = e(i) . .2). Theorem 4. . (4.1. . · · · . xn (t) Then 1 0 ··· 0 1 ··· . (1 ≤ i ≤ n). · · · . Let X be the vector space of solutions of Eq.1.
.6) (t) (t) xi+1 (t) .104 CHAPTER 4.1. . . . x(t) is spanned by n linearly independent solutions x(1) (t).2) and hence X has dimension n. ¯ ¯ ¯ · · · . xn (t) (1) xi+1 (t) . (2) . . Theorem 4.2) and let t0 ∈ J. (4.1. .. . it helps to know the relation between W (t) and the matrix A(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) · · · . x(n) (t) be solutions of Eq.1. . · · · . x1 (t) x1 (t) . (4. xi−1 (t) . . .1. . it follows that n x(t) = y(t) = i=1 xi x(i) (t) (0) That is. LINEAR SYSTEM AND TRANSITION MATRIX By uniqueness theorem of IVP.. xn (t) xn (t) · · · This gives x1 (t) x1 (t) .1. . xn (t) (n) (n) (n) . ··· . x(n) (t) of Eq. (n) xi (n) (n) (n) (n) xi−1 (t) . dW (t) = dt n i=1 (1) (1) ··· ··· . (4. (1) xi (1) (1) ··· ··· ··· .4 Let x(1) (t). .. (1) (1) (2) (1) (2) x1 (t) x2 (t) . . The following theorem gives the relation. . ¯ At times. Then we have the Abel’s formula t W (t) = W (t0 ) exp t0 T r [A(s)] ds (4. . xn (t) (n) . x(2) (t).
n j=1 x1 (t) . 0 ≤ i ≤ n − 1 as continuous functions of t ∈ J. . . the second by ai2 (t) and so on except the ith row and substract their sum from the ith row. . xn (t) (n) (n) (n) aii (t)xj (t) . . dW (t) = dt n i=1 (1) x1 (t) . . (4. . . (1) xn (t) ··· Multiply the ﬁrst row by ai1 (t). we have xi . xn (t) (2) (2) .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. ··· 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. . . n j=1 i=1 (1) (1) dW (t) = dt n ··· . . (4.7) with ai (t). aij (t)xj (t) . Example 4. . . .1.1. .1. x1 (t) .4. xn (t) (1) (1) aii (t)xj (t) · · · . LINEAR SYSTEM In view of Eq. .6) gives x1 (t) . .1. We get x1 (t) .1 If x(1) (t). . . aii (t)xj (t) .2).2). . . (n) xn (t) (n) (n) aij (t)xj (t) · · · . (4. . . . . . .1. · · · . . then W (t) = 0 on J. (k) 105 (t) = dxi (t) = dt (k) n aij (t)xj (t). .1. (2) ··· . . . j=1 (k) 1≤k≤n and hence Eq.1. x(n) (t) are linearly independent solutions of ¯ ¯ Eq.
. .1) . as we see from the following example. . 0 −a1 ··· 1 . one can show that Eq. (4.1.8) A(t) = .7) is equivalent to the linear ﬁrst order homogeneous system dx = A(t)x(t) dt where 0 0 . It is now clear from Theorem 4.106 CHAPTER 4. x(2) (t) = 1 but W (t) ≡ 0.1. Recall that A(t) is the companion matrix of Eq. 0 −a2 ··· ··· .2 Transition Matrix dx = A(t)x(t) dt In the previous section we proved that the homogeneous linear system (4. (4. . .7). x2 (t). . . . dtn Remark 4. (4. · · · . LINEAR SYSTEM AND TRANSITION MATRIX Proceeding as in Corollary 3.1. 0 −a0 1 0 .1. .8) has n linearly independent solutions x1 (t)..1. .1 is not true.1.2. xn (t) (n) (n) (1) x1 (t) x2 (t) · · · Here xk (t) denotes (n) dn xk (t) . x(1) (t) = on t .2. ··· ··· 0 0 . xn (t) and its Wronskian is deﬁned as ¯ ¯ ¯ x1 (t) x1 (t) W (t) = x1 (t) . . . . . . (n) (2) (1) ··· ··· ··· ··· xn (t) xn (t) xn (t) . .3 that Eq.1 The converse of Theorem 4. (n) (2) (1) x(t) = dn−1 x(t) dtn−1 x2 (t) x2 (t) x2 (t) .1. . t2 t are linearly independent functions deﬁned 4. (4.1. 1 −an−1 x(t) dx dt .
2.3) with the initial condition Φ(t0 . TRANSITION MATRIX has n linearly independent solutions x(1) . x(n) .2. Mn×n ].4) is equivalent to the solvability of the Volterra integral equation t φ(t) = I + t0 A(s)φ(s)ds. t0 ) is called principal fundamental matrix or transition matrix or evolution matrix if it satisﬁes the matrix diﬀerential equation given by Eq. Deﬁne a nonsingular matrix Ψ(t) as Ψ(t) = x(1) (t). we have dΨ(t) dt = = That is dΨ = A(t)Ψ(t) dt dx(n) (t) dx(1) (t) dx(2) (t) . ¯(2) . · · · . A(t)x(n) (t) 107 (4. · · · .2. a(n) and B = ¯(1) .5) in the space C[J. x(n) (t) Then. ··· .1 A nonsingular matrix Ψ(t) is called a fundamental matrix if it satisﬁes the matrix diﬀerential equation given by Eq. dt dt dt A(t)x(1) (t). t0 ).4) Here the solution Φ(t) of the above diﬀerential equation lies in the space Mn×n of n × n matrices.3) Deﬁnition 4. a(2) . A(t)x(2) (t). (4. · · · . t0 ) = I. · · · . ∞) (4. (4. Deﬁnition 4. ¯(i) .2.2. ¯(n) .2. t ∈ J = (t0 . transition matrix Φ(t. ¯ ¯ ¯ b b b Proceeding as before. dt Φ(to ) = I (4. the solvability of Eq.4.2 A nonsingular matrix Φ(t. x(2) . .2.2) (4. Its unique solution φ(t) is given by ∞ t φ(t) = I + n=1 t0 An (s)ds . · · · . tf ) ⊂ (t0 . (4. Note that Mn×n is a Hilbert space ( ﬁnite dimensional) with inner product and norm deﬁned as n n (A.2. is the solution of the matrix initial value problem dΦ(t) = A(t)Φ(t). A ¯ b 2 Mn×n = i=1 a(i) ¯ 2 where A = a(1) .2.3). That is.2. B)Mn×n = i=1 a(i) . x(2) (t).
2.2.108 CHAPTER 4. Equivat σ1 t0 t A(σ1 )dσ1 + t0 A(σ1 )A(σ2 ) dσ2 dσ1 + · · · (4. · · · .1 Let A(t) be a constant matrix A. (4.1) iﬀ there exists a nonsingular constant matrix C such that Ψ2 (t) = Ψ1 (t)C.2. (4.2. Example 4. Theorem 4.2.2.2. we have φ(t) = I + t0 t0 ··· t0 [A(σ1 )A(σ2 ) · · · A(σn ) dσn · · · dσ1 ].6) Eq.6) is called the PeanoBaker series for the transition matrix Φ(t) of the linear diﬀerential equation given by Eq. Then the transition matrix of the homogeneous system given by Eq. (4. A5 = A. (4. (4. Then Ψ2 (t) is a fundamental matrix of Eq.1) ( refer Brockett [2]). A2k = (−1)k I. A4 = I.1 Let Ψ1 (t) be a fundamental matrix of the system given by Eq. A2k+1 = (−1)k A. LINEAR SYSTEM AND TRANSITION MATRIX t σ1 t0 σn−1 where An (t) = lentely.1).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.1) is of the form t t s t0 Φ(t.2.2.2 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. A3 = −A. This gives Example 4. (4.5) or Eq. .
2.2. 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. That is x(t) = Φ(t. Also C is invertible as Ψ1 (t).2. 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.1). (4. t0 )x0 x(t0 ) = x0 .2 The initial value problem dx = A(t)x(t). t0 ) = Ψ(t)Ψ−1 (t0 ) Corollary 4. Conversely. the transition matrix Φ(t.1). let Ψ2 (t) be another fundamental matrix of Eq.1 If Ψ(t) is a fundamental matrix of Eq. (4. given by x(t) = Ψ(t)Ψ−1 (t0 )¯0 ¯ x where Ψ(t) is a fundamental matrix of Eq. (4.2.4.2. dt has a solution x(t).2.2. (4. Ψ2 (t) are invertible.1).2. t0 ) of Eq.1) is given by Φ(t. (4. Corollary 4.1). TRANSITION MATRIX Proof : Let Ψ2 (t) = Ψ1 (t)C.
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).110 CHAPTER 4. Then. we get ¯ This gives x(t) = Ψ(t)Ψ−1 (t0 )x0 = Φ(t. dt Thus we have n linearly independent solutions of Eq. t1 )Φ(t1 . ¯ Since x(t0 ) = x0 . t0 ) satisﬁes the following properties (i) Φ(t. ··· dt dt dt where Ψ(t) = x(1) . respectively (A1 = A2 ). .2 Φ(t. dt dΨ1 (t) −1 Ψ1 (t) and dt Remark 4. c2 .2. . .2.2. cn ).1) is given by x(t) = c1 x(1) (t) + · · · + cn x(n) (t) That is x(t) = Ψ(t)c where c = (c1 . x(n) with {x(i) }n linearly independent. t0 ). t0 )x0 c = Ψ−1 (t0 )x0 We also note the following facts. .1) and as its solution space is of dimension n. Remark 4. x(2) . Ax(2) . 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) . it follows that every solution x(t) of Eq. Ax(n) . . i=1 This implies that dxi 1≤i≤n = Axi . (4. .2. · · · . Ψ1 (t) = dt dt Ψ2 (t). This follows from the fact that A1 (t) = A2 (t) = dΨ2 (t) −1 Ψ2 (t). t0 ) = Φ(t. (4. .
Proof : We have Ψ(t)Ψ−1 (t) = I. (4. 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 . p(t)) is constant for all t. p(t)) ¯ ¯ Theorem 4. p(t)) + x(t).4. just observe that Φ(t. t) (iii) Φ(t. the scalar product (x(t). t) = I for all t.3 If Ψ(t) is a fundamental matrix of the system given by Eq. t1 ∈ J To realise this. (4. then Ψ−1 (t) is a fundamental matrix of the adjoint system. This gives the following theorem. Theorem 4.2. Diﬀerentiating this equality. t0 .1). TRANSITION MATRIX (ii) Φ−1 (t.2 The adjoint equation associated with d¯ x = A(t)¯(t) x dt is d¯ p = −A(t) p(t) ¯ dt d [(¯(t).7) Proof : We have = (A(t)¯(t). −A(t) p(t) x ¯ ¯ ¯ = 0 This gives (¯(t). p(t))] = x dt d¯(t) p d¯ x .2. provided that for any x dt initial data.2. p(t)) = constant x ¯ That is Eq. t0 ) = Ψ(t)Ψ−1 (t0 ). 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).7) is the adjoint equation associated with Eq. = ([A(t) − A(t)] x(t).1). t0 ) = Φ(t0 .2. p(t) + x(t). ¯ ¯ dt dt (4.2.2. (4.2.
t0 ) A (t) + A(t) φ(t.8). t0 )¯(t) + b(t) v dt (4. t0 ) = 0 and dt (ii) φ(t0 .2.2. t0 )b(t) = Φ(t0 .2. t0 ) = φ (t. (4. (4. LINEAR SYSTEM AND TRANSITION MATRIX dΨ−1 (t) dt = −A(t) Ψ−1 (t) Theorem 4. t0 ) v(t) + Φ(t. Remark 4. t0 ) = A(t) Φ(t.2.1 gives us the following corollary. t0 ) v(t) + b(t) dt dt This gives A(t)Φ(t.3 If A(t) is skew symmetric. t)b(t) dv(t) = A(t)Φ(t.8). v(t) to be determined. (4.2. Corollary 4.1) then χ(t) is fundamental matrix of its adjoint system given by Eq. t0 ) = I.2. − [A(t)] = A(t). we get ¯ dv(t) d Φ(t.7) iﬀ χ(t) Ψ(t) = C where C is a constant nonsingular matrix. This follows from the fact that (i) d φ (t.8) . t0 )¯(t) + Φ(t. that is. 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.2.2. then the transition matrix is orthogonal [Φ(t. Plugging this representation of x(t) in Eq. t0 )φ(t.9) (4.2.112 and hence CHAPTER 4. t0 ) v and hence dv(t) dt = Φ−1 (t. t0 ) = I . t0 )] Φ (t. (4. t0 )v(t) which satisﬁes Eq.3 If Ψ(t) is a fundamental matrix of Eq.
2. t t0 Φ(t.4.2. t0 )x0 + t0 Φ(t. nonhomogeneous system as t0 sin (t − t0 ) + 1 − cos (t − t0 ) + t0 cos (t − t0 ) − 1 + sin (t − t0 ) . b(t) = t 0 A= The transition matrix for this system is given by Φ(t. 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]. s)b(s)ds As x(t) = Φ(t. s)b(s)ds This gives us the variation of parameter formula t x(t) = Φ(t. we get t 113 v(t) = v(t0 ) + t0 Φ(t0 . 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 ). t0 )Φ(t0 . t0 )v(t). s)b(s)ds t0 Φ(t. TRANSITION MATRIX Integrating this equation. it follows x(t0 ) = v(t0 ) and t x(t) = Φ(t.3 Solve dx = Ax(t) + b(t) dt 0 1 −1 0 . t0 )v(t0 ) + t0 Φ(t. s)b(s)ds Example 4.
through its inverse P −1 (t) as P −1 (t) = φ(t.114 CHAPTER 4. it follows that φ(t.3.1 Let C be any nonsingular matrix. yet it is possible to get a simpler representition of the transition matrix φ(t. 0) = eRT (redeﬁne R in Lemma 4. t0 ) = P −1 (t)eR(t−t0 ) P (t0 ) where P (t) is a nonsingular periodic matrix function and R is a ﬁxed matrix.1). Lemma 4. T )e−Rt = P −1 (t) . t0 ). 0)e−R(t+T ) = φ(t + T. If A(t + T ) = A(t). 0)e−RT e−Rt = φ(t + T.3.1. t ∈ x dt A(t + T ) = A(t). t0 ) is also periodic of period φ(t + T. t0 ) in view of the periodicity of A(t). by Lemma 4.3.1) has the form φ(t.3. t0 ) of the system given by Eq. there exists a matrix R such that φ(T. Proof : As φ(T.1 (FloquetLyapunov). the associated transition matrix φ(t. T )φ(T.3. t0 + T ) = φ(t. We now deﬁne a nonsingular P (t).1) is a nonautonomous system. 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. t∈ Let us ﬁrst consider a homogenous periodic system (4. Theorem 4.1a) (4. 0)e−Rt Then P −1 (t + T ) is satisﬁes the relationship P −1 (t + T ) = φ(t + T. We observe that Eq. 0) is a nonsingular matrix.3.3.1b) We shall make use of the following lemma. which we shall state without proof (refer Coddington and Levinson [3] for further details). (4. t0 ) gives t t σ1 φ(t. LINEAR SYSTEM AND TRANSITION MATRIX 4. (4. noindent The PeanoBaker series representation of the transition matrix φ(t. then ∃ a matrix R such that C = eR .3 Periodic Linear System d¯ x = A(t)¯(t).
PERIODIC LINEAR SYSTEM That is P (t) is periodic.3.5) d −1 P (t) z (t) ¯ dt Using transformation given by Eq.4.3.6) d −1 P (t) dt (4.3) we get d¯ x d −1 d¯ z = P (t) z (t) + P −1 (t) ¯ dt dt dt . t0 ) = φ(t.2) imply that φ(t. 0)φ(0. Also we have φ(t.3.3. Since dφ = A(t)φ(t).3) (4. 0) = P −1 (t)eRt . t) = e−Rt P (t) Eq. φ(0. 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. (4. (4. (4. 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.3.2) (4.5) 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 (4.3.3.3. t0 ) = P −1 (t)eRt e−Rt0 P (t0 ) = P −1 (t)eR(t−t0 ) P (t0 ) This proves the theorem.4) d P −1 (t)eR(t−t0 ) P (t0 ) = A(t) P −1 (t)eR(t−t0 ) P (t0 ) dt Equivalently. we have dt 115 (4.3.
7b) cos t − sin t sin t cos t is periodic of period 2π and R = . (4. (4.3. 0) = cos t − sin t sin t cos t e−2t 0 0 1 = P −1 (t)eRt where P (t) = φ(t. One observes that φ(t.3. (4. We state the following theorem regarding the solvability of Eq.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.116 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX and hence Eq.3. refer Brocket[2] for a proof of this theorem. 0) = −2 0 0 0 Let us now examine the nonhomogeneous initial value periodic system d¯ x = A(t)¯(t) + ¯ x b(t) dt x(t0 ) = x0 ¯ ¯ where A(t + T ) = A(t).3. Example 4.7).7a) (4.3.1 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. ¯ + T ) = ¯ b(t b(t).
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. t0 ) [x0 − xp (t0 )] ¯ ¯ ¯ ¯ d¯ p p = −A (t)¯(t) dt (¯ is orthogonal to p iﬀ b ¯ t0 +T t0 (4. ˙ 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. Then the solution of the nonhomogeneous system given by Eq.2 Consider the harmonic oscillator problem with periodic external force b(t) (refer Example 1.10) with x1 (t) = x(t).3. ¯ ¯ b(σ) dσ = 0.3. x2 (t) = x(t).3. ω (2) (2) −x0 ω sin ω(t − t0 ) + x0 cos ω(t − t0 ) .3.7) can be written as x(t) = xp (t) + φ(t. PERIODIC LINEAR SYSTEM 117 Theorem 4.3.3.3. (4.10) is given by xH (t) = x0 cos ω(t − t0 ) + (1) (1) = x0 sin ω(t − t0 ) .8) iﬀ ¯ is orthogonal to the solution p(t) of the adjoint equation b(t) ¯ (4.) Example 4.9) p(σ). 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) (4.3. (4.1) λ d2 x + ω 2 x(t) = b(t). t0 ) be the transition matrix generated by A(t).2 Let φ(t.4.
We note the following properties.1) and S = P QP −1 .10) is given by 0 ω2 −1 0 p(t) ¯ This implies that b(t) satisﬁes the conditon t0 +T t0 Hence the nonhomogeneous equation will have a solution iﬀ p(t) is orthogonal ¯ to ¯ b(t). (4.3.3. That is. 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 . 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. [1] If P and Q are linear transformation on exp (S) = P exp (Q)P −1 . 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 . ω t0 cos ω(t − t0 ) b(t)d(t) = 0 where T is the time period 4.4. This follows from the fact n n (4.4 Computation of Transition Matrix We shall discuss some basic properties of the linear transformation exp A in order to facilitate its computaton. LINEAR SYSTEM AND TRANSITION MATRIX xH (t) = xH (0) cos ω(t − t0 ) + d¯ p = dt xH (0) ˙ sin ω(t − t0 ) ω (4.11) The adjoint system corresponding to Eq.118 This gives CHAPTER 4.
1 If A = a b 0 a . [2] If P and Q are linear transformation on n 119 which commute then exp (P + Q) = exp (P ) exp (Q) Example 4.4. then Φ(t. then exp (A) = exp (a) Denote λ = a + ib. By induction. B commutes with I and B 2 = 0.4. COMPUTATION OF TRANSITION MATRIX This implies that if P −1 AP = diag[λj ]. 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) . we get the required representation as follows. 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 . Using property[2]. B = 0 b 0 0 Also.4. t0 ) = exp (A(t − t0 )) = P diag[ exp (λj (t − t0 )) ] P −1 Here. then 1 b(t − t0 ) 0 1 exp (A(t − t0 )) = exp (a(t − t0 )) This follows from the fact that A = aI + B.
. λ2 .4. . . 1 ≤ i ≤ ¯ ¯ n − 1.1 If the set of eigenvalues {λ1 . u(n) are linearly independent. t0 ) is given by exp (λ1 (t − t0 )) 0 ··· 0 0 exp (λ2 (t − t0 )) · · · 0 −1 φ(t.120 CHAPTER 4. So let ¯ the result be true for k ≤ n − 1. .4. · · · . ¯ Thus u(1) · · · u(n) are linear independent by induction. 0 0 0 exp λn (t − t0 ) Proof : We have A¯(i) = λi u(i) . A¯(n) u u (4. (4. LINEAR SYSTEM AND TRANSITION MATRIX In a more general setting. . u(n−1) are linear independent and hence αi = 0. . . Putting these values in Eq. . We claim that u(i) . Let 0. Theorem 4. · · · .. . Further. un forms a basis for n . 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. the result is true for k = 1 as u(1) = 0 is linearly independent. (4. · · · . · · · . · · · . = P = P diag [λj ] . We shall show that it is also true for all k = n. we have ¯ ¯ AP = = A¯(1) . · · · .4. · · · .4. We proceed by induction. . t0 ) = P P . .2) by λn and subtract from Eq. λn ] and hence the transition matrix φ(t. λ2 .2) λ1 u(1) .4. u(n) is nonsingular. we have the following theorem. λn } of an n × n matrix A is real and distinct. (4.3) (4. λn u(n) ¯ ¯ λ1 . . . 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) . λn .2) we get αn λn u(n) = 0 which gives αn = 0 as ¯ λn u(n) = 0. then any set of corresponding eigenvectors u1 . The matrix P = u(1) .4. u(n) is invertible and P −1 AP = diag [λ1 . . · · · . ¯ ¯ Obviously. u(i) = ¯ λi = λj u ¯ ¯ 0.3). Hence the matrix P = ¯ ¯ u(1) . .
A= P = 1 −1 0 1 . · · · . the transition matrix φ(t.2 −1 −3 0 2 λ1 = −1. 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 . Then u(1) .2 Let A be a 2n × 2n matrix with 2n distinct complex eigenvalues λj = aj + ibj . t0 ) has the representation. A pair of corresponding eigenvectors is (1. λ2 = 2. v (1) . Theorem 4. (−1. 0 ··· ··· . . . 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. 0). v (1) .4. u(n) . φ(t. The tranP −1 AP = diag −bj aj siton matrix φ(t. we state the following theorem concerning the computation of the transition matrix. ··· 0 0 eλn (t−t0 ) −1 P 121 Example 4. 0 0 eλ2 (t−t0 ) . v (n) is invertible and ¯ a j bj with 2 × 2 blocks along the diagonal .4. λj = aj − ibj with corresponding eigenvectors w (j) = u(j) ± i v (j) . COMPUTATION OF TRANSITION MATRIX This implies that P −1 AP = diag [λj ] Proceeding as before. t0 ) is given by eA(t−t0 ) = P exp ( diag [λj (t − t0 )]) P −1 =P eλ1 (t−t0 ) 0 .. u(n) . 1 ≤ j ≤ n. . .4. . 1).4. · · · . v (n) is a basis for 2n and the matrix P = u(1) .
w (2) = (0.122 CHAPTER 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. 1. 0. λ3. A corresponding pair of eigenvectors is w (1) = (i. 0) .4 = 2 ± i. . 0. 1 + i.3 A = 0 0 3 −2 0 0 1 1 A has eigenvalues λ1. LINEAR SYSTEM AND TRANSITION MATRIX 1 −1 0 0 1 1 0 0 Example 4.4.2 = 1 ± i. 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.
3. S = [I]4×4 and N = A − S. .4.1 let λ be an eigenvalue of n×n matrix A of multiplicity m ≤ n.4.2 and Theorem 4. the matrix N = A − S is nilpotent of order k ≤ n and S and N commute.4. t0 ) = exp (t − t0 ) I + N (t − t0 ) + N 2 (t − t0 )2 2! . . 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. . For proofs of Theorem 4. Theorem 4.4. . Deﬁnition 4. . The transition matrix φ(t.2 A matrix N is said to be nilpotent of order k if N k−1 = 0 and N k = 0. 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. . . . . v (2) .4. COMPUTATION OF TRANSITION MATRIX 123 Deﬁnition 4. . λn repeated according to their multiplicity. Then there exist a basis of generalized eigenvectors v (1) . v (n) ] invertible and A = S + N. then S = diag[λ].4. Then for k = 1 . v (n) with P = [v (1) . .4. 0 −2 −1 −1 1 2 1 1 Example 4.3 Let A be a real n × n matrix with eigenvalues λ1 . . Further. t0 ) is computed as follows Φ(t.4. refer Perko[4]. v (2) . t0 ) = P diag [exp (λj (t − t0 ))] P −1 × 1 + N (t − t0 ) + · · · + N k−1 If λ is of multiplicity n. where P −1 SP = diag [λj ]. m.
1 ). 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. 1.5 A = −1 2 0 1 1 2 Here λ1 = 1. obtained by solving (A − 2I)2 v = 0. 0) So.124 = CHAPTER 4. 0.4. v (3) . is given by 1 0 0 0 = 1 0 0 v −2 0 0 That is v (3) = (0. 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) . v (2) = (0. λ2 = λ3 = 2 v (1) = (1. 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. −2 ).
4). 2 2 2 . COMPUTATION OF TRANSITION MATRIX 125 Example 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. which is modelled by the second order diﬀerential equaiton d2 x dx + ω 2 x = F (t) +k (4. This gives the solution of the homogeneous equation corresponding to Eq. passing through the initial point x(t0 ) = x0 .3.6 Let us consider the problem of mechanical oscillations discussed in Example 1. ¯= 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. k is the resistance and ω 2 = ma This is equivalent to the following nonhomogenous linear system d¯ x dt 0 1 −ω 2 −k = A¯ + ¯ x b(t) = 0 −ω 2 1 −k 0 F (t) x+ ¯ A= .4) 2 dt dt λ . 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 = . c = k 2 − 4ω 2 .1.4.4.4. (1) (2) (4.
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 ) . the transition matrix corresponding to A is given by φ(t.126 CHAPTER 4. LINEAR SYSTEM AND TRANSITION MATRIX So.
= 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. . then the eigenvalues of A are repeated . β= x(t0 ) − 2c c 2c c −k −k (iii) If k 2 − 4ω 2 = 0. 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 )] . So.4. 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 .
LINEAR SYSTEM AND TRANSITION MATRIX where N is the nilpotent matrix given by k −2 = A− 0 = This gives φ(t. 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) .128 CHAPTER 4.
4. This gives 2 4 0 0 1 0 1 0 ω 2 0 0 0 ω 0 . ¯ ¯ 0 1 0 3ω 2 0 0 A= 0 0 0 0 −2ω 0 0 0 1 2ω . 0). 2.7 Let us consider the linearized satellite problem. 0. 2ω). x3 .4.4. 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 ). (0. d¯ x = A¯(t) + B u(t) x ¯ dt where x = (x1 . x4 ). (0. u = (u1 . discussed earlier in Chapters 1 and 2. The generalized eigenvectors corresponding to the eigenvalue λ = 0 3 are (1. 1. ±ωi. we note that A has eigenvalues 0.4. (t − t0 ) β= k x(t0 ) + x(t0 ) ˙ 2 Example 4. u2 ). We follow the procedure described in Theorem 4. x2 . 0. 0.3. − 2 ω). ω. 0. 0) and the eigenvectors corresponding to the complex eigenvalues λ = ±iω are (1. 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. B = 0 1 0 0 0 0 0 1 To compute the transition matrix of the above problem. 0. 0. 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 .
130 That is CHAPTER 4. So. 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 . 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. the transition matrix φ(t. t0 ) is given by φ(t.
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. (4. we get dt 1 x2 t (4.1) Example 4.1) reduces to a linear sytem of the form d¯ x = A¯ x dt (4.5.4) .5. so that Eq.5.5 Euler’s Linear System 1 d¯ x = A¯ x dt t The equation of the form (4.5. (4.4.5.3) Using the notation x1 = x and x2 = t dx1 dt dx2 dt So. is called Euler’s system.1) where A is an autonomous matrix.5.5. (4. we get the Euler’s system 1 d¯ x = dt t = 2 3 = − x1 + x2 t t 0 1 −2 3 x ¯ (4.5.1 Solve t2 d2 x dx − 2t + 2x = 0 dt2 dt dx .2) and hence that of Eq.2) One can now make use of the linear system theory to compute the solution of Eq.5. We make a change variable of the form t = es in the independent variable.
5.3 to compute a fundamental matrix for the system 0 1 0 d¯ x = A¯(s). 2es − 2e2s −es + 2e2s Hence the solution of Eq.5. (4.4.5) is where A = φ(s) = = 2 −1 −1 1 2es − e2s −es + e2s es 0 0 e2s 1 1 1 2 . we get that Eq.2 with eigenvectors −2 3 1 Hence a fundamental matrix of the system given by Eq.5) 1 2 .5. x2 = t . (4.5.132 CHAPTER 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 3.5.5.5. (4.4) is given by c0 (2t − t2 ) + c1 (2t − t2 ) c0 (−t + t ) + c1 (−t + t ) 2 2 x(t) = ¯ Example 4. N = A − S = 0 1 −1 1 (4. 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. 1 −1 2 We use Theorem 4.2 Solve t3 dx d3 x +t −x =0 dt3 dt (4.6) dx d2 x Using the notation x1 = x. x3 = t2 2 .7) .5. (4. A = 0 1 1 x dt 1 −1 2 −1 1 0 0 1 S = I.8) (4.
(4. show that 1 −δt e sin ωt ω e−δt cos ωt − δ ω 1. 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.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. If A = δ −δt cos ωt + sin ωt e ω eAt = 1 − e−δt sin ωt ω √ where ω = 1 − δ 2 .6.4. sin ωt . (4.6 Exercises 0 1 −1 −2δ .5.5. δ ≤ 1.7) 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.
. A= . . B = .134 CHAPTER 4. . Compute the transition matrix of the system 0 et d¯ x x(t) ¯ = dt 0 0 by using PeanoBaker series. . Find two linearly independent solutions of the system d¯ x = A(t)¯(t) x dt where 0 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 . . . A(t) = 4. . 0 0 0 0 Show that (a) B n = 0. . . 0 1 0 ··· 0 0 1 ··· . compute the Wronskian of the linearly independent solutions. . . ··· 0 0 . (b) (λI)B = B(λ)I. . Let A and B be two λ 1 0 λ . . LINEAR SYSTEM AND TRANSITION MATRIX n × n matrices given by 0 ··· 0 1 ··· 0 . 0 ··· 1 0 ··· λ 2. 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. 5. ··· . . . . . . t=0 1 1 − 2 − t t Also.
10.4. Show that the transition matrices corresponding to autonomous systems described by two similar matrices are necessarily similar. t0 )φ(t0 . Compute the transition matrix of the following system 2 1 0 x d¯ = 1 3 1 x(t) ¯ dt 0 1 2 8. b(t) 7. EXERCISES 6. s)] −1 −1 0 0 0 0 −2 0 1 2 = φ(s. 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. t) = I . t0 ) of the system described by Problem 7. e−2t ). Solve the initial value problem d¯ x dt = A¯(t) x x(0) = (1. t) (b) [φ(t. Verify the following properties of the transition matrix φ(t. = A(t)¯(t) + ¯ x dt x(0) = 0 ¯ 135 where a(t) is an given in Problem 5 and ¯ = (1. s)] = φ(t.6. Solve the nonhomogeneous system d¯ x b(t). s) (c) φ(t. (a) [φ(t.
1955 [4 ] Lawrence Perko. (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. 1991 [2 ] Roger W Brockett. Finite Dimensional Linear System. Theory of Ordinary Diﬀerential Equations. John Wiley and Sons. LINEAR SYSTEM AND TRANSITION MATRIX 11. 1970 [3 ] Earl A Coddington and Norman Levinson.136 CHAPTER 4. Essentials of Ordinary Diﬀerential Equatoins. Springer Verlag. 1991 . Inc. McGraw Hill.McGraw Hill Book Company. Diﬀerential Equations and Dynamical Systems. References [1 ] Ravi P Agarwal and Ramesh C Gupta.
5..1 Asymptotic Behaviour of Linear System Theorem 4. Let ¯ ¯ ¯ ¯ P = [¯1 . focus and centre.. uk+1 . known as Jordan canonical form of a matrix.. We also sketch phase portraits. vn .. k + 1 ≤ j ≤ n. vk . which can be used in investigating the behaviour of solution of a linear system of the form given by Eq. . stable and unstable node. .1 .. vk .. This gives rise to the notion of saddle point. Theorem 5.. to study boundedness property of solutions of both linear and nonlinear systems along with the stability of equilibrium points. Br 137 (5. (4. vk+1 ..4.3 can be completely described in terms of the following theorem. are discussed in detail. Various notions of stability including the one due to Lyapunov are introduced and examined leading to some interesting theorems. . un ]..4.1 ( Jordan canonical form) Let A be a real matrix with real eigenvalues λj . 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 ). . 1 ≤ j ≤ k.4. Then there exists a basis {¯1 .1) for large values of t. uk+1 . and complex eigenvalues λj = aj ± ibj .Chapter 5 Stability Analysis In this chapter we utilize the theory of transition matrix discussed in Chapter 4.. . vn ... v ¯ ¯ ¯ ¯ ¯ Then P is invertible and P −1 AP = B1 B2 . around the equilibrium points in two dimensions. vk+1 ..1. Phase portraits.1) . vector ﬁelds and solution curves of some illustrative examples.1.1.
0 λ (∗) (∗∗) for λ = a ± ib as a pair of complex eigenvalues of A.. 0 0 0 0 ··· ··· ··· (t−t0 )m−1 (m−1)! m−2 ··· ··· (t−t0 ) (m−2)! ··· (t − t0 ) 1 . 2.1. . (a) If Bj = B is an m × m matrix of the form (*). .2 The solution x(t) of the system given by Eq. .. .138 CHAPTER 5. (5. . 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. j 0 λ 1 λ 1 B= . I2 0 D with D= a b −b a . t0 )¯0 x where φ(t.1.. I2 = 1 0 0 1 The elementary Jordan blocks B = Bj . 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 .1. t0 ) = φ(t−t0 ) = P diag [exp Bj (t − t0 )] P −1 is the transition matrix of the system. This gives us the solvability of the linear system d¯ x = A¯(t) x dt x(t0 ) = x0 ¯ ¯ (5. . STABILITY ANALYSIS = 1. . ..2) Theorem 5. 1 .. .
1. .2) tends to zero as t → ∞ iﬀ the real parts of the eigenvalues of the matrix A are negative.1. (5. Then p(A) = 0.2) 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 ). if det(λI − A) = λn + pn−1 λn−1 + · · · + p0 . A ∈ n .1 Each component of the solution of the linear system Eq. .1.3 Let p(λ) = det(λI − A).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. it follows that ∃ t1 ≥ 0 such that φ(t) = exp (At) ≤ c exp (at) tp where c is some suitable constant.1.1. That is.1 gives us the following wellknown theorem called CayleyHamilton theorem.1.2 Every solution of the linear system given by Eq.1.1). Theorem 5. Theorem 5. then An + pn−1 An−1 + · · · + p0 I = 0. Corollary 5.1. Let multiplicity of each eigenvalues λj be pj where λj = aj + ibj .3 Every solution of Eq. Corollary 5. where λ = a ± ib is a pair of eigenvalues of the matrix A and l is smaller than the dimension of the space. 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. then exp B(t − t0 ) = exp a(t − t0 ) × R R(t − t0 ) 0 R . 0 0 0 0 2 m−1 R (t−t0 ) 2! ··· ··· R (t−t0 ) (m−1)! R(t − t0 ) .. Corollary 5. . (5. ∀ t ≥ t1 . (5.5. 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.
p(t). t0 ) = exp (A(t − t0 )) be the transition matrix of the the linear system given by Eq. 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 .4) We ﬁrst prove the following variation of Gronwall’s lemma. By variation of parameter formula.1 (Generalized Gronwall’s inequality) Let u(t).1. (5. q(t) be nonnegative continuous functions in the interval I = [t0 .1. t∈I (5.3) where B(t) is a continuous matrix on the interval [ t0 .1.2).1. STABILITY ANALYSIS For the proof of the Theorem 5. Lemma 5.140 CHAPTER 5.1.1. 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 . refer Gantmacher[5] We shall now investigate the following perturbed linear system d¯ x = A¯ + B(t)¯ x x dt (5.3) is given by t x(t) = φ(t. t ∈ I (5. ˙ By assumption. t0 )¯0 + ¯ x t0 φ(t.1.1 and Theorem 5.3). tf ] and let t u(t) ≤ p(t) + Then we have t t0 q(s)u(s)ds.1. ∞) . We wish to examine the relationship between the boundedness of solutions of Eq.5) t u(t) ≤ p(t) + Proof : Let p(τ )q(τ ) exp t0 t τ q(s)ds dτ.1.1.1. (5. the solvability of Eq. (5. s)B(s)¯(s)ds x (5.6) r(t) = t0 q(s)u(s)ds Then r(t0 ) = 0 and r(t) = q(t)u(t).3. Let φ(t. (5.2) and Eq.
9) ≤ c x0 + c ¯ B(s) x(s) ds ¯ .1. (5.1. (5. we have x(t) ¯ ≤ exp (A(t − t0 )) t t0 x0 + ¯ t0 B(s) exp (A(t − s)) x(s) ds ¯ (5. (5.1. 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 5.1.1.4 Let all solutions of Eq.3) are also bounded provided ∞ t0 B(t) ≤ ∞ (5.1.1. Then all solutions of Eq. (5.2) are bounded.4).5. it follows that ∃ c such exp (At) ≤ c ∀ t ∈ I = [t0 . Theorem 5.1.1.1.4 If p(t) is nondecreasing on I.8) Proof : that As all solutions of Eq.1. ∞). (5.2) be bounded on [t0 . t From Eq. then t u(t) ≤ p(t) exp q(s)ds t0 (5.1. ∞).7) Proof : As p(t) is nondecreasing.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.3). (5.
1. (5.1 We examine the boundedness of solutions of the nonhomogeneous system d¯ x = A¯(t) + ¯ x b(t) dt under the assumption that (a) Eq.1. (5.1. (5.3. One can show that the solutions of Eq.2) be such that they tend to zero as t → ∞. (5.1.1. 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.1.2). (5.2) has bounded solutions and (b) ∞ t0 (5. provided B(t) → 0 as t → ∞ (5. (5. Then all solutions of Eq.3) visavis asymptotic convergence of solutions of Eq.4) we get.1. (5. STABILITY ANALYSIS Applying the generalized Gronwall’s inequality given by Eq.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.1.2.1.1. t x(t) ≤ c exp (−δ(t − t0 )) x0 + c ¯ ¯ t0 exp (−δ(t − s)) B(s) x(s) ds ¯ .1.1.1.3 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.3) are also asymptotically convergent to zero. (5. The proof is similar to that of Theorem 5.1. (5.9) in the proof of Theorem 5. The inequality given by Eq. (5.1.10) ¯ b(s) ds < ∞.142 CHAPTER 5.3) has bounded solutions.5 Let all solutions of Eq. Theorem 5.10) remains bounded.1.11) Proof : By Corollary 5. Remark 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.11). given any c1 > 0. ∃ 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. t0 exp (δs) B(s) ¯(s)ds x t u(t) ≤ C + c c1 By Gronwall’s lemma we get t1 u(s)ds . 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 ).2) has solutions asymptoticaly convergent to zero.10) satisﬁes the same property if ¯ b(t) → 0 as t → ∞. ¯ This proves the theorem. This gives that x(t) → 0 as t → ∞.1.5. (5. (5. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM In view of Eq.1. then the solutions of nonhomogeneous system given by Eq.2 If Eq.1. we choose c1 such that cc1 < δ and hence exp (t(c c1 − δ)) → 0 as t → ∞. Remark 5. . (5. ∀t ∈ I As c1 is arbitrary.1.1.
(5. x(t) x ¯ A(t) + [A(t)] is a symmetric matrix and hence its eigenvalues are real. ∞ (a) If ∞ t0 M (s)ds is bounded then Eq.6 solutions. STABILITY ANALYSIS We now consider the boundedness and stability of solution of nonautonomous linear system d¯ x = A(t)¯(t) x dt Observe that. x(t) ¯ ¯ ≤ M (t) ¯(t)2 x M (s)ds .1. x(t)) x x ¯ Diﬀerentiating the above relation. If M (s)ds = −∞ then we get Example 5.12) has solutions asymptotically converging to zero. then Eq.144 CHAPTER 5. We have the following theorem Theorem 5. x(t) + x(t). M (s)ds < ∞.1. Let M (t) be the largest eigenvalue of A(t) + [A(t)] .12) has bounded (b) If t0 M (s)ds = −∞. ¯(t)2 = (¯(t).1 d¯ x = [A + B(t)] x(t) ¯ dt A= 0 1 −1 0 0 and B(t) = 0 0 2a at + b .12) = (A(t)¯(t). (5. 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). A(t)¯(t)) x ¯ x x = (A(t) + [A(t)] )¯(t). t ∈ I t t0 (b). then we get (a).1.1. ¯ ¯ dt dt (5. x(t)) + (¯(t). we get d ¯(t)2 x dt = d¯(t) x d¯(t) x .1.
2 M (t) = t 0 B(s)ds → ∞ as t → ∞. Assumptions 1 The matrix A(t) corresponding to the system given by Eq. 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 = 2.1.13) corresponding to the nonautonomous system given by Eq.12).14) (b) The perturbation matrix B(t) satisﬁes the condition ∞ t0 B(t)dt < ∞ Theorem 5.1.1. (5.1. We note that Example 5. 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. (5. 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.12) are bounded.12) is such that (a) t lim inf t→∞ t0 T rA(s)ds > −∞ or T rA(s) = 0 (5.5.5 the nonautonomous system d¯ x = A(t)¯(t) x dt has bounded solutions. (5. if all solution of Eq. − 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.1.1.1.1. then all solutions of Eq.13) are also bounded. .7 Under Assumptions 1. (5.1.1. Hence by Theorem 5.
¯ x(t) = ψ(t)ψ −1 (t0 )x0 + ¯ ¯ Deﬁne c = max supt≥t0 ψ(t). Let c be deﬁned as before. (5. so is the case for all soluiton for Eq. (5.1. the boundedness of solutions implies that ψ(t) is bounded. t0 B(s)¯(s)ds x t t0 B(s)ds <∞ Theorem 5. by Abel’s lemma (Theorem 4.13).1. STABILITY ANALYSIS Proof : Let ψ(t) be a fundamental matrix of the system Eq. c0 = cx0 . Then. Also.12) will imply that ψ(t) −→ 0 as t −→ ∞.1.15) In view of Assumptions 1(a).1. (5.13). Using this fact in Eq.1. it follows that ψ −1 (t) is bounded.146 CHAPTER 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. (5.12) are asymptotically convergent to zero.1.8 Under Assumptions 1 .15). (5.16).1. we get that ψ −1 (t) → 0 as t −→ ∞.1.12).16) ¯(t) ≤ c0 + c2 x and hence Gronwall’s inequality gives x ¯(t) ≤ c0 exp c2 This proves the theorem. Proof : Asymptotic convergence of solution of Eq. We now apply the variation of parameter formula to the IVP corresponding to Eq.4).1.1. to get the following integral represntation of the solution x(t). Estimating Eq. if all solution of system Eq.1. ¯ Then Eq. (5. (5.16) gives t t t0 ψ(t)ψ −1 (s)B(s)x(s)ds (5.1. (5. we get t ¯(t) ≤ ψ(t) c1 + c x t0 B(s)¯(s)ds x . supt≥t0 ψ −1 (t) .
1. (5.5.1. Then (a) boundedness of solutions of Eq. ASYMPTOTIC BEHAVIOUR OF LINEAR SYSTEM where c1 = ψ −1 (t0 )x0 .1. Theorem 5. Then (a) boundedness of solutions of Eq. (5. (5.9 Let Assumptions 2 hold.1. (5. (5.12) will imply boundedness of solutions of Eq.1.17) ¯ b(s)ds < ∞ Theorem 5.13) is preserved under Assumption 1(b). (5. we can dispatch with the Assumption 1(a).13) and . (5.17) and (b) asymptotic convergence of solutions of Eq.1.1.17).1.1.1. Let Assumption 1(b) hold. 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 nonhomogeneous 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. The boundedness and asymptotic convergence of the perturbed system Eq.12) implies boundedness of solutions of Eq.12) to zero will imply asymptotic convergence of solutions to zero of the nonhomogeneous equation Eq.1. If the nonautonomous system under consideration is periodic.10 Let A(t) be periodic.
18) gives x(t) ≤ d1 eRt + d2 ¯ t t0 eR(t−s) B(s) x(s) ds (5.148 CHAPTER 5. Boundedness of solutions of Eq. (5.1.13). STABILITY ANALYSIS (b) asymptotic convergence of solutions of Eq. t0 ) = P −1 (t)eR(t−t0 ) P (t0 ) where P is periodic nonsingular matrix and R is a ﬁxed martix.1. (5. sup P −1 (t) t≥t0 t≥t0 Eq. (5. we must have eRt  ≤ d3 e−αt for some positive constant d3 and α.1.1.19) 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.1. (5.1. (5.1. Proof : By FloquetLyapunov theorem .12) implies that eRt  ≤ d2 and hence Eq. (5.1.13) 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. ∞).19) ¯ where d1 = de−Rt0 P (t0 )x0 .1. By variation of parameter formula applied to the IVP corresponding to Eq. Let d = max sup P (t). This inequality immediately gives ¯(t) ≤ d1 d3 e−αt + d2 d3 x t t0 e−α(t−s) B(s)¯(s)ds x .1.12) are asymptotically convergent to zero.3.1. (5.12) is given by φ(t.1.13). (5. (5.12) implies asymptotic convergence of solution of Eq.Theorem 4. the transition matrix of the periodic system 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 . If solutions of Eq.
2. x0 ) < x x ¯ x ¯ ¯ Deﬁnition 5. x).2. t0 . x(t0 ) = x1 ¯ ¯ ¯ dt ¯(t1 ) − x(1) (t1 ) < δ ⇒ ¯(t) − x(1) (t) < x ¯ x ¯ Theorem 5.2. .1) has a global solution on [t0 . Deﬁnition 5.2.FORMAL APPROACH and hence t 149 ¯(t) ≤ d1 d3 e−αt exp d2 d3 x This proves the theorem. t0 . (5. Eq. x0 + ∆¯0 ) − x(t. ∃ δ > 0 such ¯ that ∆¯0  < δ ⇒ ¯(t. 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.1 For the linear system d¯ x = A(t)x(t) dt stability is equivalent to the boundedness of the solution.3) ∀ t ≥ t 1 ≥ t0 > 0.2. This solution is said to be stable if for a given > 0.2. there exist δ and T (µ) such that ∆¯0  < δ ⇒ ¯(t. t0 . t0 . x0 ) → 0 as t → ∞ x x ¯ x ¯ ¯ (5. (5. t0 .2) can be restated as.1).2. x). indicating its dependence on t and also the initial point t 0 and initial value x0 . x0 ) denote the solution of the initial value problem ¯ ¯ Eq.2. x(t0 ) = x0 ¯ ¯ ¯ dt We assume that Eq. (5.Formal Approach We shall formally deﬁne various concepts of stability of solution of the initial value problem d¯ x = f (t.2.2 x(t.2) (5.2.2. t0 .5. STABILITY ANALYSIS . x0 + ∆¯0 ) − x(t. t0 B(s)ds −→ 0 as t −→ ∞ 5.1 Let x(t. t0 .1) To be more explicit. t0 . x0 ) is said to be asymptotically stable if it is stable ¯ ¯ and ∃ δ0 > 0 such that ∆¯0  < δ0 ⇒ ¯(t.2 Stability Analysis . (5. ∞).3 x(t. t0 . given µ > 0. x1 ) of ¯ ¯ d¯ x = f (t. x0 + ∆¯0 ) − x(t. t0 .
x0 ) = φ(t. Let x(1) (t) = φ(t)φ−1 (t1 )¯(1) (t1 ) be ¯ x ¯ x any other solution.2. Then all solutions of Eq. For the uniform stability of the linear system we have the following theorem Theorem 5.3). ¯(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 ¯ .4) is satisﬁed. Hence we ¯ have 2 φ(t. (5. Then ¯ 2 φ(t.2. (5. t0 )∆¯0  ≤ φ(t. (5.3) is bounded. t0 .2.2. it follows that the solution x(t. x0 + ∆¯0 ) − x(t. (5. t0 ) ∆¯0  x ¯ x ¯ ¯ x x If φ(t. x0 ) be a solution of Eq.2. t0 ) = max φ(j)  = max φ(t. the solution of the linear system given by Eq. Then for any ¯ ¯ ¯ t1 ≥ t0 we have x(t) = φ(t)φ−1 (t1 )¯(t1 ). STABILITY ANALYSIS Proof : Let φ(t. given 0. t0 . x Choose ∆¯0 = x δ (j) e . ¯ ¯ Conversely. Then.2. t0 )¯0  ≤ x ¯ x δ That is.1 Let φ(t) be a fundamental matrix of the system given by Eq. x0 ) = φ(t. ∃ δ > 0 such that ¯(t. t0 ) is bounded.3) has asymptotically stable solutions iﬀ φ(t) → 0 as t → ∞. Corollary 5. t0 ). t0 .2. As Eq. t0 )¯(j)  ≤ e 1≤j≤n 1≤j≤n δ This gives 2 ¯0  x ¯(t. (5. t0 ) be the transition matrix. let the solution of the linear system be stable. t0 .3) are uniformly stable iﬀ there exists a constant c > 0 such that φ(t)φ−1 (t1 ) ≤ c ∀ t 0 ≤ t1 ≤ t < ∞ (5. Then.150 CHAPTER 5.2.2 Let φ(t) be a fundamental matrix of Eq. t0 .4) Proof : Let x(t) = x(t. (5. t0 )∆¯0 ) = x δ φ(t. x0 ) = φ(t. we have ¯(1) (t) − x(t) x ¯ Let > 0 be given.3). t0 ) e(j) is the j th column vector φ(j) of the matrix φ(t. t0 )∆¯0  < x ¯ x ¯ ¯ x whenever ∆¯0  < δ. (5. Then Eq. x0 + ∆¯0 ) − x(t. t0 .2.2. t0 ) e(j)  < ¯ 2 > But φ(t. ¯(t.3). x0 ) is stable. t0 .
Whereas for dt which are uniformly stable but not asymptotically stable. This is a stable solution but 2 dx = 0.3) uniformly stable and hence zero solution is uniformly stable. we choose x ¯ (1) ∀ t ≥ t1 .1). boundedness and stability are distinct concepts. However. Instead. ¯ 151 Conversley.1. (5. all solutions are of the form x(t) = x(t0 ) not bounded.2. x(t)) dt (5. we examine a special form of Eq. dx =t dt t2 This has a solution of the form x(t) = x(t0 ) + . ∀ t ≥ t1 .1).4) and the theorem. Example 5.2. (5.2.FORMAL APPROACH and hence x(t) is uniformly stable. Example 5. t ≥ t1 δ This proves Eq. Hence. Hence given > 0.5. STABILITY ANALYSIS . it follows that φ(t)φ−1 (t1 ) = max φ(j) (t) < 1≤j≤n 2 . let Eq. it is diﬃcult to give the stability criterion.5) . which is of the type dx = Ax(t) + g(t. for nonlinear systems.2. ∃ 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. Then ¯ 2 δ < . 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 ). For the nonlinear initial value problem given by Eq.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 → ∞.2.2. (5.2. δ (t1 ) = e(j) . (5.1 Consider.2.2.
we get that t0 exp (δs) x(s)ds.8) Also the condition given by Eq. x(s)) ds(5.7) gives t x(t) ≤ c exp (−δ(t − t0 )) x0 + c t0 exp (−δ(t − s)) g(s. Then in x view of continuity of the function x(t).6) on g(t. Using this inequality in Eq. (5. δ > 0 such that  exp (At) ≤ c exp (−δt) ∀ t ≥ 0.2. (5.9) gives that g(t.2. Proof : The solvability of Eq.6) In view of Eq.152 g(t. 0) = 0 and hence 0 is a trivial solution of Eq. (5.2.8).7) Since real(λ) < 0.5) is asympotically stable. Then the trivial solution of Eq. (5.9) We assume that the initial condition x(t0 ) = x0 is such that ¯0  < d.2. t ∈ I That is x(t) ≤ c x0  exp ((c m − δ)(t − t0 )).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.10) Since x0 and m are at our disposal.2. In view of this fact. ∃ d > 0 such that g(t.2.3 Suppose the real part of eigenvalues of the matrix A are negative and the function g(t. it follows that g(t. (5. x) satisﬁes Eq. Hence t ∈ I. implies that x(t) < d and Eq. t ∈ I (5. we get t x(t) exp (δt) ≤ c exp (δt0 ) x0 + c m By Gronwall’s lemma.2. (5. ( λ is any eigenvalue of A). Eq. We have the following stability theorem for such a kind of system.10) implies that x(t) ≤ d c .2. (5. Theorem 5. . v) ≤ mv whenever v ≤ d (5. (5. we choose m and x0 in such a way that ¯ c m < δ and x0 ≤ min d. STABILITY ANALYSIS g(t.2. (5. x) implies that for given m > 0. t ∈ I x(t) exp (δt) ≤ c x0  exp (δt0 ) exp (c m(t − t0 )).2. x) → 0 as x → 0 x (5. x) satisﬁes the following condition CHAPTER 5.5) with 0 initial condition.2.2.2.6).2.2. there exists t1 > t0 such that x(t) < d for all t ∈ [t0 . x(t)) ≤ mx(t)) for all t ∈ I. it follows that there exists c. (5.6).2. Then Eq. t1 ] = I. x(s))ds (5.
g and l are positive and Fx = g l both of which have negative real parts θ3 −. Further Eq.5).11) is asymptotically stable. ∞) and also x(t) → 0 as t → ∞. (5.. Example 5.11) Eq. ∞)..2. (5.2.10) implies that given > 0.2.3 is applicable and hence the main pendulum problem given by Eq. (5.4) the solution x(t) can be extended for t ∈ [t0 . θ (5. θ). 3! x →0 ≤ M θ3 → 0 as Thus Theorem 5.2. A = 0 g − l 1 k − m g ˙ and F (x) = F (θ. STABILITY ANALYSIS . we get 3! f (θ) → 0 as θ → 0 θ sin θ = θ + f (θ). − f (θ) l 1 2 k k2 g A has eigenvalues − ± − 2 2m 4m l if k.2.2. θ) = 0.3 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 + . t0 .FORMAL APPROACH 153 for t ∈ I and this bound d does not depend on t and hence by continuation theorem (Theorem 3. ∃ η = min d.2. . x0 ) < ∀ t ∈ [t0 . (5. This gives g ¨ k ˙ g θ + θ + θ + f (θ) = 0 m l l where f (θ) → 0 as θ → 0. . such that c x0 < η ⇒ x(t.5.2. . m.11) is equivalent to dx = Ax + F (x) dt where ˙ x = (θ.2. This gives us the asymptotic stability of the zero solution of Eq..
Also.154 CHAPTER 5.3. Theorem 5.1) is called the phase portrait of the system.2. we get that Eq. For details analysis concerning this section refer Agarwal and Gupta[1].4)) f (¯) = f (x0 ) + f (x0 )(¯ − x0 ) + g(¯) x ¯ x ¯ x where g(¯) x → 0 as x − x0 ¯ ¯ x − x0 → 0. (5.2.3. .2. (5. phase space will be assumed to be two dimensional so that its geometry can be studied clearly.3.2). so is the equilibrium point zero of the nonlinear system given by Eq.1 The set of solution curves in the phase space given by Eq. (5. we shall concentrate on the autonomous system Deﬁnition 5. (5.4 give the relationship between the stability of the trivial solution of the nonlinear system of the form Eq. (5. Theorem 5.3. of the system We shall be interested in the stability of the system around equilibrium points .1) is equivalent to the system ¯ d¯ x = A¯ + g(¯) x x dt g(¯) x → 0 as x → 0 and A = f (x0 ). Translating the origin to ¯ x x ¯ x x0 .3).3.2.2) through the linear system with d¯ x = A¯ x dt (5.3. If x0 is an isolated equilibrium point of f (¯) and if f is diﬀerentiable at x0 .2.4 Assume that A possesses at least one eigenvalue with positive real part and g(t. (5. we have f (¯) = f (x0 )(¯ − x0 ) + g(¯). ¯ x ¯ We shall now study Eq.3) (5. (2.3.1) n For the phase portrait. then ¯ x ¯ we have (by Eq.3) is asymptotically stable. 5. (5.those points where f (x) = 0.3. Then the trivial solution of Eq. We restate this relationship in terms of the following theorem.3 and 5.2. STABILITY ANALYSIS In a similar way we have the following theorem for the unstability of zero solution of Eq. (5.3. ¯ ¯ As f (x0 ) = 0. (5.5) is unstable.3.6). (5.5).2) with that of the linear system given by Eq. x) satisﬁes Eq.3.3.2) Theorem 5.3 Phase Portrait Analysis dx = f (x(t)) dt (5.3.1 (a) If the equilibrium point zero of the linear system given by Eq.
5. c2 )). (5. asymptotically stable or unstable.3) is unstable.3.3) is stable then the zero solution of Eq.3.3.3. (5. c = x0 = x(t0 ). so is the unstability of the equilibrium point zero of Eq.1 We investigate the phase portrait of the linear system given by −1 0 Eq. x2 4 2 3 2 1 2 1 2 3 x1 4 Figure 5. the solution curve a vector ﬁeld. The solution of the linear 0 2 system is given by x(t) = exp (−t) 0 0 exp (2t) c. . Example 5. (c) If the equilibrium point zero Eq. (5.3. Since dt x(t) is tangent to the vectors in the vector ﬁeld f (x) = Ax.1 The function f (x) = Ax deﬁnes a mapping on 2 . at each point x ∈ 2 . (5. x1 2 The solution x(t) deﬁnes a motion along these curves.2). Hence the phase portrait in 2 is as under. PHASE PORTRAIT ANALYSIS 155 (b) If the equilibrium point zero of the Eq. k = c1 2 c2 (c = (c1 . The solution curves of the above system are of the form x2 = k .3.3.2) can be stable.3. Recall that this is called dx = f (x) = Ax.3) around the origin with A = .1: Phase portrait of Example 5. (5. that is each point c ∈ 2 moves to the point x(t) ∈ 2 after a time t. So the vector ﬁeld is as given below.3.
STABILITY ANALYSIS Figure 5. P −1 = 1 1 0 1 . . P −1 AP = −1 0 0 2 =D ¯ So.2: Vector ﬁeld of Example 5. (−1.156 CHAPTER 5.1 Example 5.2 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 −1 0 1 . 1).3. 0). under the transformation y = P −1 x.3. we obtain the decoupled system dy = Dy dt The phase portraits in y1 − y2 plane and x1 − x2 plane are as under.3.
3.3.4) where the matrix B = P −1 AP has one of the following form ( by appealing to Theorem 5. (5.4: Phase portrait in x1 − x2 plane In general.1.3.3: Phase portrait in y1 − y2 plane x2 3 2 1 x1 4 2 1 2 3 2 4 Figure 5. 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 5.3.5.3) is best studied by an equivalent linear system dx = Bx dt (5. B= λ 0 1 λ . or B= a −b b a .3.1) B= λ 0 0 µ .
then the arrows in the vector ﬁeld are reversed as we see in the following Figure. If λ > 0 > µ. The phase portraits are as under: Case 2: B = . λ 0 .5: Vector ﬁeld with arrows reversed λ 0 λ 1 . (5. In this case the system given by Eq.3. STABILITY ANALYSIS The solution of the initial value problem given by Eq. (5.4) is said to have a saddle point at 0.3. Figure 5.3.3. λ<0 0 µ 0 λ In this case the origin is referred to as a stable node ( unstable if λ ≥ µ > 0 ). 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. λ ≤ µ < 0 or B = .3. λ<0<µ Case 1: B = 0 µ The phase portrait is of the form given in Figure 5.4) with x(0) = x 0 is x(t) = or x(t) = exp (at) exp (λt) 0 0 exp (µt) x0 .158 CHAPTER 5.
.5 1 0.3. PHASE PORTRAIT ANALYSIS x2 6 4 2 3 2 1 2 4 6 1 2 3 x1 159 Figure 5.3. λ < 0.5.7: Phase portrait with λ < µ < 0 (stable node) When B = λ 0 1 λ .5 1 2 3 x1 Figure 5.5 3 2 1 0.3. the phase portrait is as under.6: Phase portrait with λ = µ < 0 (stable node) x2 1.5 1 1.
Case 3: B = x2 10 5 x1 10 5 5 10 5 Figure 5.9: Stable focus . If a > 0. the origin is a stable focus with trajectories as spiral converging to zero. we have unstable focus.3. 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 < 0.3.160 CHAPTER 5.
5 5 7. The phase portrait is as under.3.4) is said to have a centre at the origin.5 0.5 5 10 2. (5.5.3. PHASE PORTRAIT ANALYSIS x2 10 5 x1 161 10 7.5 5 2.5 1 Figure 5.5 1 x1 0.11: Origin as centre Example 5.10: Unstable focus 0 −b b 0 In this case the system given by Eq.3 d¯ x = A¯. x dt A= 0 −4 1 0 .3.3.5 1 0. Case 4: B = x2 1 0.5 Figure 5.3.
3. (c) If δ > 0 and τ 2 − 4δ < 0 and τ = 0. it is stable if τ < 0 and unstable if τ > 0. then Eq. ¯ ¯ It can be seen that x1 2 + 4x2 2 = c1 2 + 4c2 2 . then Eq. STABILITY ANALYSIS 0 1 . P −1 = 0 1 2 CHAPTER 5. It is stable if τ < 0 and unstable if τ > 0. (5.3) has a node at the origin. (5. 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).2 Let δ = detA and τ = traceA and consider the linear system given by Eq. (5.3. (b) If δ > 0 and τ 2 − 4δ ≥ 0. (a) If δ < 0. node. (d) If δ > 0 and τ = 0 then Eq.12: Phase portrait with origin as centre The above discussion can be described in the following theorem. P = 2 0 0 1 . then Eq.3. (5.3. x2 2 1 x1 4 2 1 2 2 4 Figure 5.3).3) has a centre at the origin.162 A has eigenvalues λ = ±2i. . focus or centre at the origin of the linear system. We assume that det A = 0. so that ¯ is the only equilibrium point.3.3.3) has a saddle point at the origin. 0 Theorem 5. giving the saddle point.3. (5.3) has a focus at the origin.
4 Consider the following linearized Predatorprey model (linearized k a around the equilibrium point ( . One of the attractions of the Lyapunov method is its ¯ ¯ appeal to geometric intution as we see in the following two examples. For more on phase portrait analysis refer Perko [9].4. ¯ Using the notation of the above theorem (Theorem 5. LYAPUNOV STABILITY 163 where x = (F. ma This is equivalent to the following system d¯ x = A¯ x dt 0 1 dx . (1.2). S). it implies that the equilibrium k a point ( .4 Lyapunov Stability The principle idea in the Lyapunov method is the following physical reasoning. we have origin as a stable focus. we have origin as a stable node.1 dx d2 x +k + w2 x = 0 dt2 dt λ k is the resistance and w 2 = . ¯ dt 2 −w −k δ = detA = w 2 > 0 and τ = traceA = −k < 0.5) λ c kc 0 − λ d¯ x = A¯(t). with F representing ﬁsh population and S shark population.3. As δ > 0 and τ = 0. we get δ = det A = ak > 0 and τ = trace A = 0. (i) If τ 2 − 4δ = k 2 − 4w2 < 0. then the ¯ ¯ energy will continually decrease until it ﬁnally assumes its minimum. 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 .5.3.2. A = x dt aλ 0 c Example 5.3. (ii) If k 2 − 4w2 ≥ 0. ) is a centre. From mathematical perspective. 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 . )). 5. . and A = where x = x. λ c Example 5.3.5 Consider the problem of mechanical oscillations discussed in Example 1. given by Eq.
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). x2 ) ¯ ¯ The trajectories of the above system in the phaseplane are ellipses about the origin.164 CHAPTER 5. Then x 1 2 ˙ x V (¯(t)) = (x(t).1.4. Example 5. R(t) = x2 (t) + x2 (t) 1 2 . 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.1 Let us recall Example 1. d2 x λ + ω 2 x = 0.4. The trajectory of the above system is given by R(t) = R0 e−(t−t0 ) .2 Consider a slight modiﬁcation of the earlier example x˙1 x˙2 = x2 − ax1 = −x1 − ax2 with a > 0. We deﬁne V (¯) = (x2 + x2 ). 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. x = (x1 . case 1 of a harmonic oscillator given by a second order diﬀerential equation. STABILITY ANALYSIS Example 5.3. The energy of the system is given by E(¯) = E(x1 . ω 2 = dt2 ma In terms of a ﬁrst order system we have d¯ x = dt 0 −ω 2 1 0 x.
that is. ¯ We now investigate the role of Lyapunov function to ensure the stability of the equilibrium state of the nonautonomous nonlinear system.1) dt where f : I × n → n with origin as an equilibrium point (f (t. (t.4. Assumptions 1: There exists a scalar valued function V (t. x(t)). that is.4.5. ¯ called Lyapunov function. Suppose this distance is continuously x ¯ ˙ x decreasing as t → ∞. such that it is continuous and has continuous partial derivatives w. V (¯(t)) < 0. x) : I × n → n . 0) = 0). x) is positive deﬁnite.4.4. x) x ¯ (5.3) . there exist a continuous.4. x) ∈ I × n ¯ ¯ (5. 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. we have ¯ 0 < α(¯) ≤ V (t. Then x(t) → 0. d¯ x = f (t.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. This leads to the following x intersting geometric interpretation of the Lyapunov function. 0) = 0 and ¯ (i) V (t. x) ∈ I × ¯ x ¯ n (5. nondecreasing ¯ function α such that α(0) = 0 and for all t and all x = 0.t x and t with V (t.r. x(t)) ≤ −r(¯) < 0 ∀ (t.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. LYAPUNOV STABILITY x2 165 x1 V(x)=constant Figure 5.
4. x0 )) ≤ V (t0 . As β ∗ (t0 .4. This implies the uniform stability of the zero solution the equilibriun point.166 (ii) (a) CHAPTER 5. We get δ( ) (independent of t0 ) such that β(δ) < α( ). we get α(¯) ≤ V (t. t0 . x(t)) ≤ β(¯) ∀ t ∈ I ¯ x (5. for the second inequality. x(t.(ii)(a) and (iii) hold (c) asymptotically stable if (i). Then ¯ proceeding as in part(a).4. x0 ) < α( ) if x0  < δ ¯ ¯ Using (5. x0 ) ≥ V (t. STABILITY ANALYSIS ˙ V (t. ¯ 0. x0 ) < x ¯ whenever x0  < δ for t ≥ t0 .(ii) and (iii) hold. x0 )) ¯ ¯ ¯ x ¯ Here. Let > 0 be x ¯ y x arbitary. x) ∈ I × ¯ ¯ n (5. t0 .4.2). ¯ 0 Then the equilibrium state ¯ is (a) stable if Assumptions (i) and (ii) hold (b) uniformly stable if Assumptions (i). Proof : (a) Let β ∗ (t.4) and (5. (5. the above inequality implies that ¯(t.5) Theorem 5. t0 .4. t0 . for t ≥ t0 and x0  < δ ¯ . ) such that β ∗ (t0 .4) (iii) there exists a continuous. nondecreasing function β such that β(0) = 0 and V (t.6). Let x0  < δ. x(t)) ≤ 0 ∀ (t. we have α( ) > β(δ) ≥ V (t0 .1). x0 )) ≥ α(¯(t. x) be a Lyapunov function of the system (5. we have used Assumption 1(iii). t0 . we can choose x ¯ δ(t0 . x0 ) < α( ) ¯ (5. ¯) is a continuous function of x.6) Since α is nondecreasing and positive.1 Let V (t. y ) for ¯ ≤ ¯.4. x(t. x0 ) x ¯ ¯ ¯ forall t ≥ t0 and x0  < δ.4. The above inequality gives ¯(t. ¯ ≤ β ∗ (t0 . ¯) be the maximum of V (t. (ii)(a) and (iii) hold and assume that > 0 is arbitary. This proves the stability of the equilibrium state ¯ (b) Let (i).4. x0 ) < x ¯ where δ is independent of t0 .
t0 .4. t1 ]. x0 ) < µ for t ≥ t0 + T ¯ x ¯ Let functions α(µ) and β(µ) be as deﬁned in Eq. we only need to show that ∃ r > 0 such that x0  ≤ r ⇒ ¯(t. Let x0  ≤ r. (5. r). t0 . t1 = x ¯ t0 + T. t0 )c2 ≤ β(r) − T c2 = 0 which is a contradiction. LYAPUNOV STABILITY 167 (c) For asymptotic stability of the zero solution. Deﬁne T (µ. x0 ))dτ ¯ r(¯(τ ))dτ x ≤ V (t0 . r) such that x0  ≤ r ⇒ ¯(t.4. t0 . Take any positive constant c1 and ﬁnd r > 0 such that β(r) < α(c1 ). Hence Eq. x0 ) > λ on [t0 . we ﬁnd λ(µ) > 0 such that β(λ) < α(µ). given µ > 0.2) gives x ¯ 0<λ ≤ V (t1 . x(t. r) We ﬁrst claim that ¯(t. we have x2  = x(t2 . x0 ) → 0 uniformly with t → ∞.5. we have ¯ x0  ≤ r ⇒ ¯(t. t0 . t1 ].4. ¯ x ¯ Equivallently. t0 . x0 ) − t0 ≤ V (t0 . (5. x0 )) ¯ ¯ t1 = V (t0 .3) and Eq. (5. x0 ) = λ for some t = t2 in [t0 . t0 . Then by part (b). As β(µ) is continuous at 0.5). Hence. t0 . then ¯(t. c2 (µ. x0 ) = λ ¯ ¯ . If not. x0 ) + t0 t1 ˙ x V (¯(τ.4. r) = β(r) > 0. x0 ) < c1 ∀ t ≥ t0 ¯ x ¯ ¯ Let 0 < µ ≤ x0 . ∃ T (µ. 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. t0 . x0 ) − (t1 .
. Using the transformation x = P y . there exists a basis {v1 . . vk+1 . vk+2 . x(t. v2 . Then by Theorem 5. x2 ) ≤ β(λ) < α(µ) This implies that ¯(t.. x). · · · . λ2 . vk+1 . t0 . we get the reduced system ¯ ¯ d¯ y = P −1 AP y = D¯ ¯ y dt . λk and the remaining are complex eigenvalues λj = aj ± ibj . · · · . vn . · · · . un ].1. · · · . 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. x2 )) x ¯ ¯ ≤ V (t2 . · · · . where D is block diagonal matrix of the form λ1 λ2 . t2 . v2 . vk . However.168 Therefore CHAPTER 5. 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 . STABILITY ANALYSIS α(¯(t. k + 1 ≤ j ≤ n. x0 ) < µ ∀ t ≥ t0 + T (µ. vk+2 . 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. λk D= Bk+1 . vk . . x0 )) ≤ V (t. vn . t0 . Bn with Bi = ai −bi as a Jordan block corresponding to the eigenvalue ai ± bi a i ibi .1. then P is nonsingular and P −1 AP v ¯ ¯ ¯ ¯ ¯ ¯ = D. If P = [¯1 .
we require all diagonal elements of D to be negative. we need to solve the matrix equation given by Eq.4.3 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. We seek Lyapunov function V of the forrm V (¯) = (B y . y (t)) y ¯ dt (5. 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 . (5. we get dV dt d¯ y dy .4. y (t)) + D B y (t). y (t)) + (B y (t). y (t) y ¯ ¯ ¯ = B = In order to ensure that BD + D B y (t). Diﬀerentiating Eq. Example 5. (5.9) where Q is a given positive deﬁnite matrix.4. y (t) ¯ ¯ (5.4. (5.4. y) y ¯ ¯ (5. D¯(t)) y ¯ ¯ y = (BD¯(t).5.9) will hold. Eq.4.10) dV as negative dt deﬁnite.7).4. LYAPUNOV STABILITY 169 For asymptotic stability. y(t) + B y (t). we shall require that dt dV = − (Q¯(t). −2 ± 2i. to get a Lyapunov function which is positive deﬁnite with (5. if we assume that BD + D B = −Q Thus.10) for B with a given positive deﬁnite matrix Q (say Q = I). ¯ ¯ dt dt = (BD¯(t).4.7) where B is a suitable positive deﬁnite matrix.8) dV is negative deﬁnite.4.
4. V (¯) x = (B x. y2 = x2 . y3 = x4 . 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 .170 CHAPTER 5.11) in the following way.4. Solve B to satisfy the matrix equation ¯ ¯ A B + B A = −I .4. x) . We construct a Lyapunov function for Eq. y4 = −x3 − x4 V (¯) = x x2 x2 (x3 + x4 )2 x2 1 + 2 + + 4 2 4 4 4 Example 5. (5. B a positive deﬁnite matrix.4 Consider the nonlinear system of the form d¯ x = A¯ + g(¯) x x dt (5.11) 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.
5 Exercises d¯ x = A(t)¯ x dt −t 0 e 1 0 . 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. Show that all solutions of the following diﬀerential equations are bounded in [ t0 . show that for A ∈ n×n . (5. 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). Amann [3]. Hence For more on stability results refer Aggarwal and Vidyasagar [2]. 3. 5. Borckett [4]. α−2kB > 0. 1 ≤ k ≤ n − 1 such that eAt = n−1 k=0 αk (t)Ak .1.11) is asymptotically stable.5. ∞) . Show that all the solution of the linear system tend to zero as t → ∞. where A(t) = 0 0 −t 0 −t 0 2 −1 −e −2t − cos t t 2 −t2 e−3t 2. we get [g(¯)] B x + x B[g(¯)] x ¯ ¯ x ≤ 2g(¯)B¯ x x ≤ 2kB¯2 x dV ≤ −¯2 [1−2kB] = −α(x) where α(r) = αr 2 .4. 2 −t cos t 1. By using the CayleyHamilton theorem (Theorem 5.7] and Mattheij and Molenaar [8]. there exist functions αk (t). Lakshmikantham et al[6.5.
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. c > 0. +c + 1+ dt2 dt 1 + t4 (ii) 4. STABILITY ANALYSIS d2 x dx 1 x = 0. c > 0. Test the stability. Test the stability. +c + 1+ dt2 dt 1 + t2 dx 1 d2 x x = 0.172 (i) CHAPTER 5. 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 . asymptotic stability or unstability for the trivial solution of the following linear systems.
1 . 1 (iii) If a(t) = − . (i) Show that origin is uniformly asymptotically stable if a(t) < 0. p > 1 and p is integer dt Show that 7. Consider the nonlinear system . 8.5. Discuss the stability or unstability of the origin of the following linear sytem and sketch the phase portraits. show that origin is asymptotically stable but not (ii) If a(t) = − t+1 uniformly stable.5. Consider the nonlinear system dx = −xp . 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. show that the solution of the corresponding IVP (t + 1)2 approaches a constant value as t → ∞ but origin is uniformly stable. 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.
Ordinary Diﬀerential Equations: Introduction to Nonlinear Analysis. dx = −x dt dx (ii) = −x(1 − x) dt d¯ x 0 1 (iii) = −1 −a dt (i) x. 1977 [3 ] Herbert Aman. 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.A Martynyuk. John Wiley.. Diﬀerential Equations and Dynamical Systems. then origin is unstable. R. . STABILITY ANALYSIS (i) If p is odd. 1998 [7 ] V. Stability Analysis. 1970 [5 ] F. 1991 [2 ] J K Aggarwal and M. Vidyasagar. 1960 [6 ] V. 9. Walter de Gruyter. Leela and A. Stability Analysis of Nonlinear Systems. Marcel Dekker. S. 1991 . Ordinary Diﬀerential Equations in Theory and Practice. John Wiley and Sons. Chelsea Publishing Company. Hutchinson and Ross.174 CHAPTER 5. Dowden. Essentials of Ordinary Diﬀerential Equations. 1996 [9 ] Lawrence Perko. 1990 [4 ] Roser W Brockett. Inc. Finite Dimensional Linear Systems. G. Lakshmikantham. I and II. 1989 [8 ] R M M Mattheij and J Molenaar. McGraw Hill Book Company.. Gorden and Breach Science Pub. (Editors) Nonlinear Systems. Inc. Method of Variation of Parameters for Dynamic Sytems. Find Lyapunov functions around the origin for the following system. Lakshimantham and S. a ≥ 0 ¯ 10. SpringerVerlag. Deo. The Theory of Matrices Vols. then origin is uniformly stable and (ii) If p is even. Gantmacher.
∞ 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. A series which does not converge is said to be a divergent series. We mainly focus on Legendre. Example 6. 1−t 175 . 1−t This gives f (t) = lim sn (t) = 1 for t < 1.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) k=0 where t is a variable and c0 .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. Hermite and Bessel diﬀerential equations.1. c2 .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 . · · · are constants. We also deduce various interesting properties of Legendre and Hermite polynomials as well as Bessel functions.1. This limit f (t) is denoted as the sum of the series at the ∞ point t. c1 . 6.
One can show that if the series converges absolutely. . k=0 (b) The series (c) No conclusion if L = 1. SERIES SOLUTION ck (t − t0 )k is said to converge absolutely Deﬁnition 6.1 A power series ∞ k=0 at a point t if the series k=0 ck (t − t0 )k  converges.176 ∞ CHAPTER 6. then it also converges. ak diverges if L > 1. (ii) Ratio test ∞ Let the series k=0 ak ( with an = 0) be such that lim ∞ k→∞ ak+1 = L. (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 . ak (a) The series k=0 ∞ ak converges absolutely if L < 1. the converse is not true. However. ∞ k≥1 Then the orginal series k=0 ∞ ak converges. (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 .1. We have the following tests for checking the convergence or divergence of a series of real numbers. ∞ k≥1 Then the orginal series k=0 ak diverges.
(6. ak diverges if L > 1.1. . ∞ k t The power series converges for t < 1 and diverges for t > 1.1) converges at a point t = t1 . 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.1.1 If the power series given by Eq. then r is called the radius of convergence. Thus.1) are bounded at t = t1 . Proof : Since the series given by Eq.1. Hence the elements ck (t1 − t0 )k of the series given by Eq. it follows that the partial sums sn (t1 ) converge and hence sn (t1 ) is Cauchy.1. (6.1. it converges. (a) The series k=0 ∞ ak converges absolutely of L < 1. ∞ Deﬁnition 6.2) implies that ck (t − t0 )k  = ck (t1 − t0 )k t − t0 t1 − t 0 k k k (6. PRELIMINARIES (iii) Root test ∞ 177 If the series k=0 ak is such that lim ∞ k→∞ (ak ) k 1 = L.1) converges for t = t1 . (6. the radius of convergence of this series is 1.6. k=0 (b) The series (c) No conclusion if L = 1. implies that ck (t1 − t0 )k −→ 0 as k → ∞.1.3) If t−t0  < t1 −t0 .1. Theorem 6.1. ck (t1 − t0 )k  ≤ M for all k ≥ 1 Eq. in turn. then it converges absolutely for every t for which t − t0  < t1 − t0 . This implies that sn+1 (t1 ) − sn (t1 ) → 0 as n → ∞ This. At t = 1. (6.1. k k=1 it diverges and t = −1. That is.2 If the series k=0 ck (t − t0 )k converges absolutely for t − t0  < r and diverges for t − t0  > r.2) t − t0 t1 − t 0 ≤ M ∞ (6.
but it is bounded. That is. Let L denotes its limit. .addition and subtraction. then r = ∞ and hence the series Eq. 1 Then. then ck = dk for all k ≥ 0.1. Theorem 6.4) is continuous at t = t0 .2 (Radius of convergence) (i) Suppose that the sequence {ck  k } converges . SERIES SOLUTION The radius of convergence of the power series given by Eq. (6.1.1. (6. then r = 0 and hence the series is convergent only for t = t0 . (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 .1. if ∞ ∞ f (t) = k=0 ck (t − t0 )k = k=0 dk (t − t0 )k in a disk: t − t0  < r. ∞ ∞ ∞ 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 . if L = 0. ∞ 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. (i) The function f (t) in Eq. the radius of convergence r of the power series is . multiplication.4) One can easily show the following. term by term diﬀerentiation and integration.178 CHAPTER 6. 1 1 ∞ Let k=0 ck (t − t0 )k be a power series with non zero radius of convergence r. L (ii) If L = 0. r2 ).1. If this sequence is not bounded.1) may be determined from the coeﬃcients of the series as follows. (iii) If {ck  k } does not converge. (6.1) converges for all t. (ii) The same function f (t) can not be represented by two diﬀerent power series with the same centre.
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 nonautonomous 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!
(6. we get CHAPTER 6.184 Hence using Eq. 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. This solution is uniquely determined by the initial value x0 . 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. ¯ ¯ . (6. it follows that z = ¯ thereby implying that x = y .1. For if x(t) and y (t) ¯ ¯ ¯ are two diﬀerent solutions of the initial value problem given by Eq.2.2.2. 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 (6.5). This proves the existence of analytic solution of the system given by Eq. ¯ ¯ 0.3). Hence.3). (6.10) provided t ≤ p < r.
c5 = − c3 3! 5.2.2. c4 = − c2 = c0 2! 4. dx = dt ∞ kck tk−1 .2.6. 1.11) dt2 dt dx By substituting x1 = x.12) This gives us c2 c3 = − = − n(n + 1) (n + 3)(n − 2) (n + 3)(n + 1)n(n − 2) c0 . (6. (6. k = 0.2.11) 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.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 .4 (n + 4)(n + 2)(n − 3)(n − 1) = c1 5! . LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS Example 6.2. x2 = . We shall now x(t) = k=0 c k tk .3 4! (n + 2)(n − 1) (n + 4)(n − 3) c1 . 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 ( Legendre Diﬀerential Equation) 185 d2 x dx − 2t + n(n + 1)x = 0 (6. We substitute ∞ k=0 ∞ k=2 ck tk .1 that ∞ Eq. 2. · · · (k + 2)(k + 1) = − (6.2. it follows by Theorem 6. k=1 d2 x = dt2 k(k − 1)ck tk−2 in Eq.11) has a unique analytic solution of the form determine the coeﬃcients ck .
186 CHAPTER 6. So.(2n − 1) cn = 1 when n = 0 and n = for n = 0.12) is zero for k = n. the same is true for x2 (t). (6.. In Eq.. x1 (t) reduces to a polynomial of degree n and if n is odd. 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.5. writing ck in terms of ck+2 . (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. SERIES SOLUTION and hence the coeﬃcients ck ’s are inductively deﬁned. (6. are called Legendre polynomials.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.3.2. we have 0 = cn+2 = cn+4 = · · · and hence if n is even.2. These polynomials multiplied by some constants. 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.2. Some of the Legendre polynowhere M = or 2 2 mials are given as below .12). It is standard to choose 2n! 1.
we “ −t2 2 put u(t) = e x(t). (6. 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. P0 (t) = 1.16) It is clear that cn+2 = 0 and hence the solution given by Eq. one encounters the following deiﬀerential equation d2 u + λ − t2 u(t) = 0 (6. P3 (t) = (5t3 − 3t) 2 2 1 1 4 2 5 3 (35t − 30t + 3). As in the previous example.2. which is denoted by Hn (t).2.13) dt2 over the interval”(−∞.15) dx2 dx and 2 in Eq. we get a series solution of the form ∞ x(t) = k=0 ck tk for − ∞ < t < ∞ (6. the Hermite polynomial.2.14) are analytic in (−∞. As the coeﬃcient of the diﬀerential equation given by Eq. 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. (6. (6. P2 (t) = Example 6.2.13) (or alternatively Eq.2.2. (6. So Eq.6.14) we see that Plugging the representation for x(t). ∞).2.2. LINEAR SYSTEM WITH ANALYTIC COEFFICIENTS 187 1 2 1 (3t − 1). As we look for the solution u ∈ L2 (−∞. . we normalize cn and put it equal to 2n and then compute the coeﬃcients in descending order.2 (Hermite Diﬀerential Equation) In the study of the linear harmonic oscillator in quantum mechanics. (6. ∞).3. This gives us the solution.2. (6.14)) is called Hermite diﬀerential equation.2. P1 (t) = t. ∞).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 (6.15) is a polynomial. P5 (t) = (63t − 70t + 15t) P4 (t) = 8 8 We shall discuss the properties of Legendre diﬀerential equation in section 4.2.14) dt2 dt Eq.
3) Theorem 6. we get ¯ . provided µ is an ¯ eigenvalue of A0 and no other eigenvalue of the form µ + n exists for A0 .1 The diﬀerential equation given by Eq. is said to have a regular singularity at t0 . 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 .3.1) then reduces to the equation of the form d¯ x 1 = dt t A k tk k=0 x ¯ (6.3.188 CHAPTER 6. (6. (6. giving the existence of a series solution of the equation given by Eq. where n is a positive integer. As before.3.1) A(t) = k=0 A k tk (6.3. (6.3. we can assume that t0 = 0 and A(t) has series representation of the form ∞ (6. ∞ Eq.3.3) We prove the following theorem. Proof : We introduce a new dependent variable y (t) = tµ x(t).2) in the interval of the convergence t < r.3) has a series ∞ solution of the form x(t) = tµ ¯ k=0 ck tk in the interval t < r. SERIES SOLUTION 6.3.3 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. . . · · · .4) (A0 − (µ + n)I)¯n c . This gives us the recurrence relations c (A0 − µI)¯0 = 0 c (A0 − (µ + I)) c1 = −A1 c0 ¯ ¯ . 2. . the recurrence relation given by Eq. cn . · · · We shall now show that the series x(t) = tµ ¯ ¯ ¯ ck tk converges.4) provides k (k + 1)¯k+1 = [A0 − (µ + k + 1)I] c Using Eq.3. 3.3.5) and (ii) we get (k + 1) ck+1 ≤ ¯ Equivalently. it follows that there exists a positive quantity such that A0 − (µ + k)I > . 2.3. 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. for some positive numbers M . (6. LINEAR SYSTEM WITH REGULAR SINGULARITIES ∞ ∞ k 189 ck kt = ¯ k=1 k=0 l=0 k Ak−l cl ¯ − µ¯k tk . the above relations iteratively deﬁne c ∞ c0 . . · · · (6. (6.5) In view of property (ii).6. · · · 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 ¯ . 1. = − An−l cl ¯ l=0 Since (A0 − (µ + n)I)¯n  = 0 for all n ≥ 1. and p < r we have Ak pk ≤ M.2. c1 . n−1 (6. k = 1. ¯ k=0 As in Section 6.3.
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) ¯
≤
tµ
k=0
c k tk ¯
∞ M M
≤ tµ c0 ¯ =
k=0
+1 ··· k!
M
+k−1
t p
k
tµ 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
3. it follows that the second solution y (t). a = 2 . 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).3. (6.10). Hence. linearly independent of x(t).3.Eq.6.9) . (6. 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 ∞ . LINEAR SYSTEM WITH REGULAR SINGULARITIES y2 (t) = −c1 x2 t dt x2 2 195 Using the represntations of Eq. 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. 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).
2) We can think of L as an operator deﬁned on L2 [−1. y ∈ D(L) = D(L∗ ) That is L∗ = L and D(L∗ ) = D(L). We have already proved in Example 6. of Chapter 7 to claim that the eigenfunctions of L form a complete orthonormal set.4) (6. 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. y) 1 −1 −1 where J(x.1) gives 1 −1 1 1 (yLx − xL∗ y) dt = J(x. 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. it follows that J(x.1 The set of Legendre polymonials {Pn (t)} satisfying Eq.196 CHAPTER 6.4. for the sake of completeness we prove this result directly. 1]. y)1 = 0 and hence −1 1 −1 (yLx − xL∗ y) dt = 0 ∀x. We appeal to Theorem 7.4.2.3) (6. y)−1 = a0 y dy dx −x dt dt with a0 (t) = (1 − t2 ).1 that Eq.1) are orthogonal set of polynomials in L2 [−1.5. (6. Theorem 6.1) has sequence of eigenvalues {n(n + 1)} with {Pn (t)} corresponding eigenfunctions and hence they are orthogonal. (6.4. 1] with D(L) as the set of all functions having second order derivatives.4.4. Then Green’s formula (to be done in Section 7.4.4. As a0 (1) = a0 (−1) = 0.4 Properties of Legendre.4. SERIES SOLUTION 6.1. 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 (6.1) In terms of the notation of Chapter 6. However.5) .
· · · 2n n! dtn 1 n n! 2 M (6. Pn ) = 1 −1 Pm (t)Pn (t)dt = 0.4. 1. u) = n=0 fn (t)un (6.4. HERMITE AND BESSEL FUNCTIONS197 Multiplying Eq.1 Let {fn (t)} be a sequence of functions in an interval I. (6. 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 .7) . (6.4. Theorem 6.5) by Pn (t) and substracting the equations and integrating.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 . A function F (t. PROPERTIES OF LEGENDRE. n = 0.4. 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. That is Pm ⊥ Pn .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 .6.4.4) by Pm (t) and Eq.4. 2.
..4 The Legendre polynomials satisfy the following recurrence relation (n + 1)Pn+1 (t) = (2n + 1)tPn (t) − nPn−1 (t).(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.. 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 .4. (6..4.9) Proof : Diﬀerentiating the relation given by Eq.2.(2n − 1) 1.(2n − 2) 1. SERIES SOLUTION We have the following theorem giving the generating function for the sequence {Pn (t)} of Legendre polynomials..3.(2n − 3) (2t)n − (2t)n−2 + · · · 2.... we get [1 − u(2t − u)] −1 2 13 2 1 u (2t − u)2 + · · · = 1 + u(2t − u) + 2 24 1.4...(2n) 2.. ...8) Proof : Let t ≤ r (arbitary positive number) and u < (1 + r 2 ) 2 − r....3. Then 2tu − u2  ≤ 2tu + 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.8) with respect to u. Theorem 6.4...4.(2n − 1) n u (2t − u)n + · · · + 1.3.......4. (6..198 CHAPTER 6..4..4 = Pn (t) Theorem 6. n = 1. 2.(2n) The coeﬃcient of un in this expression is 1.3 For Legendre polynomials {Pn (t)}.....3.. we have ∞ (1 − 2tu + u ) 2 −1 2 = n=0 Pn (t)un 1 (6.
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.
204 CHAPTER 6.5 0. .25 5 0.75 0.4. Brown and Churchill [2].25 10 15 20 t J1\2 Figure 6.5 J3\2 0. refer Agarwal and Gupta [1]. SERIES SOLUTION 1 J1\2 0.25 0. Hochstadt [3] and Kreyzig [4].2: Sketch of J−1/2 (t) and J1/2 (t) J3\2 0.4.5 0.3: Sketch of J3/2 (t) and J−3/2 (t) For more details on varioius topics in this chapter.75 1 10 15 20 t Figure 6.25 5 0.
x(0) = 1.5 Exercises → be deﬁned as follows 1 − cos 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 ). Let the functions f. Find the series solution of the following IVPs (i) x + tx − 2x = 0. 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).. g : Show that f and g are analytic at t = 0. x(π) = 0 ¨ ˙ (ii) t(2 − t)¨ − 6(t − 1)x − 4x = 0... x(π) = 1. sin t .. x(0) = 0 ¨ ˙ ˙ (iii) x + et x + (1 + t2 )x = 0.....3. x(1) = 1. Show that the two linearly independent solutions of the diﬀerential equation t(1 − t)¨ + (1 − t)x − x = 0 x ˙ ... t=0 g(t) = 2.5. x(1) = 0 x ˙ ˙ 3..6. EXERCISES 205 6..((2n − 1)2 − a2 ) 2n+1 t (2n − 1)! n=1 ∞ ∞ 4. x(0) = 1. 2 t=0 1.((2n − 2)2 − a2 ) 2n t (2n)! n=1 (1 − a2 )(32 − a2 ).. t=0 t2 1 . t = 0 t .1 tn n! (−1)n n=0 5. f (t) = 1.. x(0) = 0 ¨ ˙ ˙ (iv) x − (sin t)x = 0..
n is odd Pn (0) = (−1)n/2 1. Diﬀerential Equations. Inc.((n − 1)2 + 1) n t (n!)2 n=1 1. 1993 [3 ] Harry Hochstadt.5. 1999 . McGraw Hill.n − 1 ...2...4.. Show that p(t) is orthogonal to all Legendre polynomials of degree strictly less than n..5. 7. 9. John Wiley and Sons . Inc. Using Rolle’s theorem show that between two consucative zeros of Jn (t).. show that 0... Dover Publications. McGraw Hill Book Company.2.. Agarwal and Ramesh C. Advanced Engineering Mathematics.3. Fourier Series and Boundary Value Problems. 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.((n − 1)2 + 1) (n!)2 n=1 tn ∞ ∞ = x1 (t)lnt + 2 n × k=1 k−2 k((k − 1)2 + 1) 6. 1964 [4 ] Erwin Kreyszig. Using Rodrigues formula. Let p(t) be a polynomial of degree n... 8..206 are given by x1 (t) x2 (t) = 1+ CHAPTER 6. n is even 2.... Show that Jn (klm t) = 0 1 2 J (klm ) 2 n+1 10... Inc. Prove the following identities for Bessel functions Jµ (t). Gupta.n References [1 ] Ravi P. there exists precisely one zero of Jn+1 (t). A Modern Approach. SERIES SOLUTION 1.. Essentials of Ordinary diﬀerential Equations.
This transforms the diﬀerential equation in to integral equation. 2 dt dt a0 (t) = 0 on [a.Chapter 7 Boundary Value Problems In this chapter we mainly deal with boundary value problems involving second order diﬀerential operators. Nonlinear boundary value problem is also the topic of our investigation. Towards the end.1) We shall be concerned with the solution of the ordinary diﬀerential equation over the interval a ≤ t ≤ b. We use both ﬁxed point theory and monotone operators for this analysis. b] Since L is of second order.1. there will be two boundary conditions of the form Bi (x) = Ci . Throughout this chapter. 7. The main tool for solvability analysis of such problems is the concept of Green’s function. Hence there is a detailed discussion involving the deﬁnition of Green’s function through Diracdelta function.1 Introduction Lx = f (7.2) where Ci s are given constants. its properties and construction. Here L is a second order linear diﬀerential operator of the form L = a0 (t) d d2 + a1 (t) + a2 (t). we shall limit ourselves to those Bi s which are linear combinations of x and its derivative. So the most general form of the boundary 207 . is more informative. which. i = 1. we give eigenfunction expansion technique for computing the solution of boundary value problems. 2 (7. Bi s are functions of the unknown variable x.1. subject to certain boundary conditions. Numerous examples are given for illustration. at times.
(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. 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). BOUNDARY VALUE PROBLEMS condition Eq.2 Consider a rod of length unity and unit crosssection. dt2 It is clear that we have homogeneous unmixed boundary conditions. To ensure that Eq. x(a) = x(b) ˙ ˙ Example 7. Suppose there is a heat source of density f (t) in the interior.1. β12 ) and (α21 .1. Eq. (7.1. If C1 = C2 = 0 . we have Lx = f.6) (7. (7.1. β22 ) are linearly independent . β21 . Example 7. α12 . α22 .1 d2 x dt2 x(0) = f (t) = x(1) = 0 (7.1. B2 (x) ≡ x(1).4) If α12 = β11 = β12 = α21 = β21 = β22 = 0. β11 . Writing Eq. If β11 = β12 = α21 = α22 = 0. then the temperature distribution x(t) satisﬁes − d dt K dx dt = f (t) .1.6) in the form Eq. We assume that the rod is homogeneous and the ﬂow of heat is in the direction of t only.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). the boundary conditions are called unmixed. B1 (x) = 0 = B2 (x) where L ≡ d2 .1. (7. (7. βij are given real numbers.1).1.208 CHAPTER 7.3) then reduces to α11 x(a) + α12 x(a) ˙ β21 x(b) + β22 x(b) ˙ = C1 = C2 (7. we say that boundary conditions are homogeneous.1.3) gives two distinct boundary conditions we assume that the row vectors (α11 .1. B1 (x) ≡ x(0).3) where the coeﬃcients αij .1.
2 Adjoint of a Diﬀerential Operator As before. b] which consists of functions which have piecewisecontinuous derivatives of second order and which satisfy homogeneous boundary conditions B1 (x) = 0 = B2 (x) Deﬁnition 7. we have homogeneous unmixed boundary conditions. K is the thermal conductivity of the rod. If the end points of the rod are kept at zero temperature we have the boundary value problem Lx = f.2. B1 (x) = 0 = B2 (x) where L ≡ m d2 . ADJOINT OF A DIFFERENTIAL OPERATOR 209 Here. b].1. b] be a second order diﬀerential operator.6 the adjoint operator L∗ : L2 [a.1) D(L∗ ) will be obtained in the following steps: (y. The force on the particle is along the line and the particle starts from rest at the origin. using Deﬁnition 2. B2 (x) ≡ x(1). ˙ dt2 This is an initialvalue problem. Lx) = (L∗ y. 7. Deﬁnition 7. b] with domain D. dt We can write this in our standard form as m Lx = f.2. b] is deﬁned implicitly by the equation (y. B1 (x) ≡ x(0). we mean a subspace of L2 [a. b] → L2 [a. b] → L2 [a. B1 (x) ≡ x(0). x). x ∈ D(L) and y ∈ D(L∗ ) b (7. Then.2. Example 7.2 Let L be the diﬀerential operator acting on the Hilbert space L2 [a. we shall denote by L2 [a.2.3 Consider the motion of a particle along a straight line. Lx) = a b yLxdt y(a0 x + a1 x + a2 x) dt ¨ ˙ a = . which is a function of t. B2 (x) ≡ x(0). B1 (x) = 0 = B2 (x) where L ≡ − d dt K d dt . Then the displacement x(t) satisﬁes Newton’s law d2 x = f (t) dt2 dx (0) = 0. with initial conditions x(0) = 0.2. By domain D(L). b] the Hillbert space of all square integrable real valued functions on [a.7. Here again.1 Let L : L2 [a.
y) ∗ a Deﬁnition 7. 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. From Eq. y) = a0 (y x − xy) + (a1 − a0 )xy is called ˙ ˙ ˙ the bilinear concomitant of x and y. It is clear that L is formally selfadjoint if a0 (t) = a1 (t). 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. So.210 CHAPTER 7. y) = a0 (y x − yx) ˙ ˙ a .5 L is said to be formally selfadjoint if L∗ = L. BOUNDARY VALUE PROBLEMS On the RHS.3 The operator given by Eq. 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 7.3) Deﬁnition 7. a second order diﬀerential operator (7.2) + [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.2. (7.2.3) is called the formal adjoint of L and is denoted by L∗ .2. (7.2. a formally ˙ selfadjoint operator L can be written in the form L= J(x. Thus we see that L∗ consists of two parts. integrating by parts we get (y.2.2.2.4 The expression J(x.2) we get the Green’s formula b a b (yLx − xL y)dt = J(x. x) b = a x d d2 (a0 y) − (a1 y) + a2 y 2 dt dt b dt (7.
2. (7. (7. B2 (x) ≡ x(1) = 0. 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)a = 0 ∗ for all x ∈ D only if y satisﬁes two homogeneous boundary conditions B1 (y) = ∗ B2 (y) = 0. if L = L∗ and D = D∗ Remark 7. ˙ 0 Thus we see that though the operator is formally selfadjoint.4) (7. y)1 = 0 for all x ∈ D iﬀ B1 (y) ≡ y(1) = 0.2. From the deﬁnition of J(x.1 L = d2 with boundary conditions dt2 B1 (x) ≡ x(0) = 0 B2 (x) ≡ x(0) = 0 ˙ L is formally selfadjoint and so L∗ = d2 .1 If we consider L∗ as an operator on the Hilbert space L2 [a. J(x. one can show that L is selfadjoint as an unbounded operator if the boundary value problem given by Eq. We know that ˙ L∗ = a 0 d2 d + (2a0 − a1 ) + (¨0 − a1 + a2 ) ˙ a ˙ dt2 dt . y)0 . D ∗ = D.4) is said to be selfadjoint. y)b = 0 a b for every x ∈ D. dt2 1 To determine adjoint boundary conditions. Example 7.4) is selfadjoint in the sense of the above deﬁnition (Refer Dunford and Schwartz [3]). B2 (y) ≡ y(1) = 0.2.2. a ≤ t ≤ b (7. Deﬁnition 7.6 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 . b] with domain D∗ . it can be seen that J(x.2.7.5) Boundary value problem given by Eq.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. y).2. ADJOINT OF A DIFFERENTIAL OPERATOR 211 We say that y ∈ D ∗ (domain of L∗ ) if it is twice diﬀerentiable and if J(x.2. we need to ﬁnd J(x. Hence the boundary value problem is not selfadjoint.2. These are called adjoint boundary conditions.
if L is formally delfadjoint.3 Green’s Function Lx = f B1 (x) = 0 = B2 (x) (7. y)1 0 = a0 (y x − xy) + (a1 − a0 )xy1 ˙ ˙ ˙ 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. then D ∗ = D. the solution T of Eq.2) must be interpreted in the sense of a distribution. the inverse of L.3.3. this inverse operator is an integral operator. (7.3.3. 7.2) is a distribution. we see that the adjoint boundary conditions diﬀer from the original one and D∗ = D.3. The kernel of this operator will turn out to be the Green’s function of the system Eq. That is. (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. Since RHS of Eq. But we will show that the solution of Eq. First we deﬁne fundamental solution for the operator L. BOUNDARY VALUE PROBLEMS L will be formally selfadjoint iﬀ a0 = a1 . and use it to solve the above equation. y)1 . ˙ To determine the adjoint boundary conditions we ﬁrst determine J(x. which we now deﬁne.3. 0 J(x.2) is a weak solution. (7. (7.1) Let the boundary value problem be given by We want to ﬁnd L−1 .1 Any solution of the equation LT = δ(t − s) is said to be a fundamental solution for the operator L. the boundary value problem is selfadjoint.3.212 CHAPTER 7.2) . (7. As we will see later.1). Deﬁnition 7. However.
L∗ φ) = (δ. T (t.2) formally as LT = 0. By deﬁnition 2 dk L∗ φ = (−1)k k (an−k φ) dt k=0 .(7. t = s 213 We will determine ys (t) by appropriate matching conditions at t = s. φ) = φ(s) for every test function φ.3.2) is given by ys (t). then integrating Eq. We now show that T (t. That is. 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. GREEN’S FUNCTION We prove this assertion by writing Eq. Assume that T is continuous at t = s. (7. We get a particular solution of Eq.7. That is (T. t < s d2 T dt = 1.3. t > s T = x(t).3.3) is a weak solution of Eq.2) in the following way: ys (t).3) t<s By the theory of diﬀerential equaitons. (7. s) = x(t).2).3.3. s) deﬁned by Eq. (7.3.3. such a solution ys (t) exists and is determined by x(t) and a0 (t).3.2) 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. t>s (7.
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 = 0.3. In view of what we have just proved. s) as a solution of the boundary value problem Lg B1 (g) = 0. L∗ φ) = φ(s) = (δ(t − s). s) That is d2 g dt2 = 0.1) is a fundamental solution of Lg = δ(t − s) LT = δ(t − s) which satisﬁes additional boundary condition B1 (g) = 0 = B2 (g). we get (T. (7. alternatively we can deﬁne the Green’s function g(t.3.3.(7. s < t ≤ 1 g(0.3.214 CHAPTER 7. BOUNDARY VALUE PROBLEMS (T. a0 (s) Example 7. we get (T.2 The Green’s function g(t. s) = 0 = g(1.1). s) . 0 ≤ t < s. s) for the above sustem is the solution of the auxiliary problem d2 g dt2 = δ(t − s) g(0. = 0 = B2 (g) dg dt − dg dt s<t≤b = t=s− (7. φ) That is We are now in a position to deﬁne the Green’s function of Eq. s) = 0 = g(1.1 Consider the string problem d2 x dt2 x(0) = f (t) = x(1) = 0 Green’s function g(t. s) of Eq.4) g is continuous at t = s and t=s+ 1 . Deﬁnition 7. a ≤ t < s.
s ≤ t . s < t ≤ 1 The continuity of g(t. s ≤ t ≤ 1 Example 7. s) = A(1 − t)s . s < t < ∞ dt2 g(0. g(t. s) ˙ m g is continuous at t = s and dg dt − dg dt = =s− t=s+ 1 .7.3. s) = At + B. m Integrating and using the boundary conditions we get t≤s 0. dg dt − dg dt = 1. 0≤t<∞ = x(0) = 0 ˙ Green’s function for the above system is given by d2 g = 0.3. 0 ≤ t < s.2 Consider the initialvalue problem representing motion of a particle in a straight line m d2 x dt2 x(0) = f (t). g(t. s) at t = s gives 0≤t≤s At. s) is continuous at t = s. s) = t(s − 1). GREEN’S FUNCTION g(t. t=s− 215 t=s+ Integrating the above equation and imposing boundary conditions at 0 and 1. the jump condition on the derivatives dg at t = s gives dt s(t − 1). s) = 0 = g(0. s) = C(t − 1). s≤t≤1 (1 − s) As − A = 1. g(t. we get 0≤t<s At. 0 ≤ t ≤ s Finally. A = s − 1 s−1 So g(t.
t).3. s) = (t − s) . h)a = 0 and hence h(s. In particular. s ≤ t Deﬁnition 7. s) = g(s. Then we get b a (hLg − gL∗ h)dt = b b a b h(t. So t≤s 0.3. s) for all s. If the system is selfadjoint. h) left hand side of the above equation we get and substituting this in the J(g.1 Theadjoint Green’s function h(t. η).3. BOUNDARY VALUE PROBLEMS Using continuity of g at t = s we get B = −As. s) = A(t − s). if the original boundary value problem given by Eq. s≤t m L∗ h = δ(t − s) ∗ = 0 = B2 (h) (7. g(t. s).3. s) satisfy the relation h(t.3. η b .3. η) − g(η. η. Eq. η) = g(η. m Hence t≤s 0. (7. subtract and integrate from t = a to t = b. s) is deﬁned as the solution of the system ∗ B1 (h) 1 Finally the jump condition for the derivative gives A = . s)δ(t − η)dt Using Green’s formula a (hLg − gL∗ h) = J(g.5) Theorem 7.1) is selfadjoint.6) by h(t. then we know that h = g and hence g(s. s) for all s.7) ∗ = 0 = B2 (h) Multiply Eq. h)b = h(s.6) L∗ h = δ(t − η) (7. s) is symmetric. Proof : We have Lg B1 (g) and ∗ B1 (h) = 0 = B2 (g) = δ(t − s) (7. (7. s) and the Green’s function g(t. η) = g(η. g(t.216 CHAPTER 7.3.7) by g(t. s) a Since h satisﬁes the adjoint boundary condition J(g.3 The adjoint Green’s function h(t. η)δ(t − s)dt − b a a g(t.3. then the Green’s function g(t. (7.
(7.3. . t−) = gt (t+.3. t−)f (t) − gt (t. If the derivatives α(t) and β(t) exist. c≤t≤d (7. using the above mentioned lemma. s) in the triangular regions ABC and ABD respectively (Refer Fig. η) be a continuous function of two variables in a region α(t) ≤ η ≤ β(t).2. we have t− b ˙ x(t) = a gt (t. η)dη.8) as b t− b x(t) = a g(t.8) We shall use the following lemma which is stated without proof. s)f (s)ds + [g(t. (7.3.3. t+)f (t)] Due to the continuity of g(t. t+) = gt (t−. where α(t) and β(t) are continuous functions. t−)f (t) − g(t. s)f (s)ds + gt (t. a ≤ t ≤ b.2 Suppose g(t. 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. then the derivative of the integral in Eq. α(t))α(t) ˙ ∂t Proof (Proof of the main theorem): Let g(t.1 (Leibnitz) Let f (t.3. s)f (s)ds + t+ g(t. 7. the last bracketed expression vanishes. t+)f (t) We note that gt (t. s)f (s)ds = a g(t. g is symmetric.1). (7.3. t) due to the continuity of g(t. η)dη + f (t. s) be the Green’s function for the boundary value problem given by Eq. and ˙ ∂f as well as f are continuous. s). s)f (s)ds + t+ gt (t.3.1). t) and gt (t. Then it has a solution given by b x(t) = a g(t. s)f (s)ds (7. We can write Eq. s)f (s)ds + t+ gtt (t.7. c ≤ t ≤ d. B1 (x) = 0 = B2 (x). β(t))β(t) − f (t. Then the integral β(t) g(t) = α(t) f (t.3. s) is a Green’s function for the boundary value problem Lx = f. GREEN’S FUNCTION That is. Theorem 7. Lemma 7.9) ∂t is given by the generalized Leibnitz formula β(t) g(t) = ˙ α(t) ∂ ˙ f (t.9) ˙ represents a continuous function of t. Also t− b x(t) = ¨ a gtt (t. s)f (s)ds Then.
Thus.3. In exactly the same manner one can show that x given by Eq. (7. t+)] a0 (t) = 0 + 0 + f (t) = f (t) b 1 a0 (t) a0 (t) Hence x(t) = a g(t.3.8) is the solution of the boundary value problem given by Eq.3 The operator L and the integral operator G deﬁned as b Gf = a g(t. s)] ds [a0 (t)gtt (t.218 CHAPTER 7. s)] ds t+ + + f (t) [gt (t. . t−) − gt (t.8) satisﬁes B1 (x) = 0 = B2 (x). BOUNDARY VALUE PROBLEMS S C B t=s A D t Figure 7.3. s)f (s)ds satisﬁes the given diﬀerential equation. (7. s) + a2 (t)g(t.3. substituting the expressioins for x(t) and x(t) in Lx ≡ a0 (t)¨ + a1 (t)x + ¨ ˙ x ˙ a2 (t)x. x(t) given by Eq.1: Integration in a triangular region So. (7.3. Theorem 7. we have t− Lx = a b [a0 (t)gtt (t. s)f (s)ds are inverses of each other.1). s) + a1 (t)gt (t. s) + a1 (t)gt (t. s) + a2 (t)g(t.
An alternative method of solution of Eq. s)f (s)ds − J(x(s). to get b a b b (hLx − xL∗ h)dt = a f (t)h(t. t)f (t)dt − J(x(t). s)f (t)dt − x(s) b x(s) = a h(t. s). B2 (x) = β (7. t) we obtain b b x(s) = a g(s. Green’s function also enables us to get a solution of the boundary value problem with nonhomogeneous boundary conditions: Lx = f B1 (x) = α. s)dt − a δ(t − s)x(t)dt Using Green’s formula in the LHS.3. g(t.3. t)) a b or b x(t) = a b J(x(s). s) satisﬁes the system ∗ B1 (h) L∗ h = δ(t − s) ∗ = 0 = B2 (h) (7. s) = g(s.10) is as follows.3.11) by x(t) . g(t.10) Theorem 7.10). Hence the result. s)f (t)dt − J(x. Eq. (7. s) is a Green’s function of Eq. h)b = a That is b a h(t. then it has a solution x(t) given by b b x(t) = a g(t.3. h) a Using the fact that h(t.3.4 If g(t. we have b J(x.7. s)f (s)ds − J(x(s). g(s. .3.10) by h(t. (7. s)) a can be calculated explicitly by the given boundary conditions.3. s)) a Proof : We recall that the adjoint Green’s function h(t. g(t. s))a g(t. (7.3.11) Multiply Eq. GREEN’S FUNCTION 219 Proof : It is clear from the deﬁnition of L and G that LGf = f and GLx = x. subtract and integrate from t = a to t = b. (7.
s) = s(t − 1). α and β and is unique. corresponding to Eq. Example 7. we must have B1 (x2 ) = 0. t ≤ s s. s)t=0 ˙ = s − 1. s) ˙ g(t. s))f (s)ds B2 (g(t. BOUNDARY VALUE PROBLEMS Let x1 (t) and x2 (t) be nontrivial solutions of the homogeneous system. s) is the Green’s function for the system with homogeneous boundary conditions and c1 and c2 are constants to be determined. t ≥ s t≥s = (s − 1). As in Example 7. s)t=1 . s)f (s)ds + β α x1 (t) + x2 (t) B2 (x1 ) B1 (x2 ) It is clear that this solution depends continuously on f. From the above g(t. We write the solution of the system as b x(t) = a g(t.3.3. we have b c2 B1 (x2 ) + c1 B1 (x1 ) + a b B1 (g(t. s))f (s)ds a = α = β c1 B2 (x1 ) + c2 B2 (x2 ) + Therefore. b x(t) = a g(t. s)t=1 = s ˙ .10) satisfying B1 (x1 ) = 0 and B2 (x2 ) = 0. x(1) = β It is a symmetric function and satisﬁes the boundary conditions g(t. s)f (s)ds + c1 x1 (t) + c2 x2 (t) g(t. Applying boundary conditions. Since a completely homogeneous system has only the trivial solution. (7. s)t=0 = 0 = g(t.3.1. s) is given by t(s − 1). g(t. its Green’s function g(t. t ≤ s g(t.3 Let us compute the solution of the nonhomogeneous boundary value problem d2 x dt2 x(0) = f (t) = α. respectively.220 CHAPTER 7.
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. Proof : Uniqueness of the solution is immediate.1) which has only the trivial solution. We need to prove the existence in two parts. by construction. B1 (x) = 0 = B2 (x) (7.4.4.1 For the restricted case of unmixed and initial boundary conditions of the completely homogeneous system Lx = 0. s)f (t)dt − J(x. (7. g)1 0 g(t.4. Theorem 7. by Theorem 7. . the solution x(s) of the above boundary value problem is given by 1 x(s) = 0 1 g(t. 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. ﬁrst for unmixed boundary conditions and then for initial conditions. For.3. if there are two Green’s functions g1 and g2 then (g1 − g2 ) is a nontrivial solution of the completely homogeneous equation Eq. CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION221 Using the formula for J(x.4 Construction of Green’s Function and Modiﬁed Green’s Function Lx = f B1 (x) = α.4. the Green’s function for the boundary value problem exists and is unique.7. 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) which is a contradiction to the hypothesis. g) we get J(x.4.
This follows from the theory ˙ of diﬀerential equation (Refer Coddington and Levinson[1]. otherwise the completely homogeneous system will have a nontrivial solution. s) is the Wronskian of x1 and x2 and is nonzero since x1 and x2 are linearly independent. s) x1 (s)x2 (t) . The Green’s function is then of the form Ax1 (t). 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) give Ax1 (s) − Bx2 (s) = 0 −Ax1 (s) − B x2 (s) = ˙ ˙ Bx2 (t).) Similarly. s) = This proves the existence of Green’s function for the boundary value problem. a0 (s)W (x1 . we have the existence of x2 (t). x2 . x2 . x1 and x2 are linearly independent.222 CHAPTER 7. a≤t<s a0 (s)W (x1 . for one can take x1 (t) to be the unique solution of Lx = 0. a ≤ t < s Continuity and jump conditions of g(t. s) s<t≤b g(t. s) where W (x1 . So x1 (t)x2 (s) . x(a) = α12 . x(a) = −α11 . 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 nontrivial solution of Lx = 0 satisfying boundary condition at t = a and x2 (t) be a nontrivial solution satisfying boundary condition at t = b. x2 . Case 2 Boundary conditions are of initial type x(a) = 0 = x(a) ˙ . s) x1 (s) a0 (s)W (x1 . x2 . Such a solution exists. x2 .
s) a≤t<s t>s where x1 . Then s W (x1 . s) −1 s x1 (s) = a0 (s)W (x1 . x2 . x2 . x2 . we get A = B This gives 0.4. s) −1 1−s −1 = −1 x1 (s) x2 (s) . CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION223 Green’s function is of the form 0. x2 . a≤t≤s t≥s = x2 (s) a0 (s)W (x1 .7.4. Ax1 (s) + Bx2 (s) Ax1 (s) + B x2 (s) ˙ ˙ Solving for A and B. s) = x2 (t)x1 (s) − x1 (t)x2 s . x2 . a0 (s)W (x1 . g(t. Continuity and jump conditions on g(t. 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. x2 are two linearly independent solutions of Lx = 0. s) give = 0 = 1 a0 (s) g(t. s) x1 (s) a0 (s)W (x1 . s) = 1 = x1 (s) ˙ A = B = x2 (s) ˙ x2 (s) 1−s = a0 (s)W (x1 . s) Example 7. s) = Ax1 (t) + Bx2 (t).
t ≤ s s(t − 1).4. t≤s =1 We take x1 (t) = 1. So we write t x1 (t) = 0 ds . If the heat source has density f (t).3 Consider the problem of heat distribution in a rod of unit length and unit crosssection and with thermal conductivity k(t).2 Consider the initialvalue 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. dt Then 1 s W (x1 . x2 (t) = k(s) x2 (s) − 1 k(s) t 1 dx k(s) x1 (s) W (x1 . then the governing equation is − d dt k dx dt = f (t) (t − s) . x2 .4. s) = t(s − 1).224 So CHAPTER 7. x2 . 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. s) = 1 k(s) . BOUNDARY VALUE PROBLEMS g(t. s) = 0. Example 7. s) = 0 1 Hence g(t. t ≥ s Example 7.
5772 is the so called Euler constant which is the limit of 1 + 1 + = 2 · · · + 1 − lns as s approaches inﬁnity.4.4 In the study of heat ﬂow in a cylindrical shell. 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) . But J0 (t) and Y0 (t) fail to satisfy the 5 boundary conditions at a and b. 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). t ≤ s 1 0 Example 7. the homogeneous Bessel’s equation d dt t dg dt + tg = 0 t dg dt + tg = δ(t − s). s) = x (s)x (t) 1 2 k(s) 1 = 0 k −1 (t)dt . 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) . 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.4.7. t≥s k −1 (t)dt −1 x1 (t)x2 (s). t≤s g(t.
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][J0 (a)Y0 (b) − Y0 (a)J0 (b)] 2 [J0 (a)Y0 (b) − Y0 (a)J0 (b)] = πt So g(t. s) = √ √ sin λt sin λ(1 − s) √ √ . Y0 . π Example 7. t) = W [J0 . respectively. √ √ x1 (t) = sin λt.4. x2 (t) = sin λ(1 − t) sin W (x1 . Z0 . t≤s C 1 2 Z0 (s)Z0 (t) . 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. t≥s C where C= 2 [J0 (a)Y0 (b) − Y0 (a)J0 (b). It can be shown that 1 2 W (Z0 . x2 . BOUNDARY VALUE PROBLEMS 1 2 Z0 (a) = 0 = Z0 (b). which dt satisfy the boundary condition at 0 and 1. t≥s λ sin λ . s) = 1 2 Z0 (t)Z0 (s) .226 Obviously now CHAPTER 7. t≤s λ sin λ √ √ sin λ(1 − t) sin λs √ √ .
x2 . The boundary conditions give us two linear equations which will determine the values of A and B. s) = Ax1 (t) + Bx2 (t) + x1 (t)x2 (s)H(s − t) + x2 (t)x1 (s)H(t − s) W (x1 . s)a0 (s) does not vanish. 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 ) = 0.4. Let x1 (t) and x2 (t) be any two linearly independent solutions of Lx = 0.4. s) is continuous and its derivative has a jump of magni1 . s)a0 (s) where A and B are constants which will be determined so that g(t. or B1 (x2 ) = B2 (x2 ) = 0 which implies that either x1 (t) or x2 (t) is a nontrivial solution of the completely homogeneous system which is a contradiction.2).4.3) for A and B. B1 (x) = 0 = B2 (x) has only the trivial solution. (7. It is clear that g(t. either B1 (x1 ) = B2 (x1 ) = 0.3) 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 . What happens when the completely homogeneous system has a nontrivial solution ? .Eq. x2 . So far. Then write g(t. Since they are linear and tude a0 (s) homogeneous. CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION227 In general. where the boundary conditions are mixed. So by solving Eq. These equations will have solutions if the determinant B1 (x1 ) B1 (x2 ) B2 (x1 ) B2 (x2 ) If the determinant vanishes. B1 (x) = 0 = B2 (x) under the assumption that the completely homogeneous system Lx = 0. we get the desired Green’s function. Now consider the boundary conditions. which is not possible as discussed earlier. we have proved the existence of Green’s function for the system Lx = f.4.7. we have B1 (g) ≡ AB1 (x1 ) + BB1 (x2 ) + B1 (r) = 0 where r(x) stands for (7. s) satisﬁes the given boundary conditions. (7.4. the Green’s function may be found in a straignt forward way.2) (7.
refer Stakgold [5]. Example 7. (7. ∗ ∗ B1 (z) = 0 = B2 (z) B1 (x) = 0 = B2 (x) B1 (y) = 0 = B2 (y) Theorem 7. then (iii) has also the trivial solution and (ii) has one and only one solution. So it follows by alternative theorem that Eq.4) g(0) = 0. g(1) = g(1) ˙ That is. a ≤ t ≤ b.4. Since 0 δ(t − s)dt = 0.6 Consider the selfadjoint system d2 g = δ(t − s) dt2 g(0) = 0. d2 g = δ(t − s). Eq. dt2 has a nontrivial solution g = t. the ordinary Green’s function does not exist. a For a proof of this.4) will have a solution iﬀ 1 0 1 (7.228 CHAPTER 7.5) .4.4.7 Consider the selfadjoint system δ(t − s)dt = 0. (ii) The inhomogeneous system Ly = f .4. (iii) will also have nontrivial solution and (ii) has solution iﬀ b f (t)z(t) = 0 for every z which is a solution of (iii). a ≤ t ≤ b. (7.2 (a) If (i) has only the trivial solution x(t) = 0 in a ≤ t ≤ b.4.4) will have no solution.4. a ≤ t ≤ b. Example 7. 0 ≤ t ≤ 1 dt2 g(0) = 0 = g(1) ˙ (7. (iii) The adjoint homogeneous system L∗ z = 0.4. 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. (b) If (i) has a nontrivial solution. g(1) = g(1) ˙ The completely homogeneous system d2 g = 0.
dt2 1 0 g(0) = 0 = g(1) ˙ ˙ has a nontrivial solution g = const. s) B1 (gM ) = δ(t − s) − x1 (t)x1 (s) = 0 = B2 (gM ) x1 (t)δ(t − s)dt = 0. for δ(t − s)ds = 0. However.4. a ≤ t ≤ b. Deﬁnition 7. Theorem 7. Lx = 0. CONSTRUCTION OF GREEN’S FUNCTION AND MODIFIED GREEN’S FUNCTION229 The completely homogeneous system d2 g = 0.4. (7. Hence Eq. the ordinary Green’s function can not be constructed for Eq. s) is symmetric if b gM (t.4. Modiﬁed Green’s Function Assume that the selfadjoint system Lx = f . That is. (7. 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). 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. b a B1 (x) = 0 = B2 (x) is such that the completely homogeneous system B1 (x) = 0 = B2 (x) has nontrivial solution of the form x1 (t).7. a ≤ t ≤ b.3 The modiﬁed Green’s function gM (t. where x1 (t) is normalized x2 (t)dt = 1 .5).4.1 Modiﬁed Green’s function is the solution of the system LgM (t. s)x1 (t)dt = 0 for every s a .4. 1 It now follows that the Green’s function which is a solution of the system Lg = δ(t − s).5) has no solution.
s2 )) = 0 = B2 (gM (t. s1 )δ(t − s2 )dt + x1 (t)gM (t. s)) = 0 = B2 (gM (t. we have b 0 = gM (s1 . s) we can alternatively deﬁne the modiﬁed Green’s function as the solution of the system: LgM (t. satisfying (i) B1 (gM (t. s1 ) and subtract and integrate from a to b. BOUNDARY VALUE PROBLEMS Proof : Consider gM (t. s2 ) = δ(t − s2 ) − x1 (t)x1 (s2 ) B1 (gM (t. s)dt = 0 for every s. Hence gM (s1 . s1 )dt x1 (s2 ) a a Using Green’s formula and boundary conditions for gM (t. s2 ) and Eq. s2 )dt a +x1 (s2 ) a x1 (t)gM (t. s)) (ii) gM (t. s2 )(LgM (t.230 CHAPTER 7. s) = δ(t − s) − x1 (t)x1 (s). s1 )(LgM (t. s2 ) − gM (s2 . s2 )δ(t − s1 )dt − gM (t. s1 )) − gM (t. gM (t. s2 ). s2 )dt x1 (s1 ) a b − x1 (t)gM (t.4. s1 ) − x1 (s1 ) b x1 (t)gM (t.6) by gM (t. s2 )) (7. s1 )dt But b a x1 (t)gM (t. s ≤ b .4.6) B1 (gM (t. s1 )) = 0 = B2 (gM (t. s1 ). s2 ) wich satisfy the systems: LgM (t. (7. s1 ) On imposing the conditions of symmetry on gM (t. s1 ) = δ(t − s1 ) − x1 (t)x1 (s1 ) (7.7) Multiply Eq.4. s) is continuous at t = s a ≤ t. (7. s1 )) and LgM (t. s2 ))] dt b = a b gM (t. Then we get b a b [gM (t. s2 ) = gM (s2 . s1 ) and gM (t.7) by gM (t.4.
If a b f (t)x1 (t)dt = 0. gM ) = a gM (t.4. b B1 (x) = 0 = B2 (x) (7. for a f (t)x1 (t)dt = 0.8) is clear by alternative b theorem. s)f (t)dt + Cx1 (s) . s) be the modiﬁed Green’s function of the selfadjoint system Lx = f . (7. s)dt = 0 for every s For computational purposes. fo both x and gM satisfy zero boundary conditions and there fore we have b x(s) = a gM (t. We ﬁnally have the following theorem which gives a solution of the selfadjoint system by using the modiﬁed Green’s function. then Eq.4 Let gM (t. a ≤ t ≤ b.8) has a solution x(t) given by x(t) = a gM (t.4.4. We have Lx = f . Theorem 7.7. it reduces to b x1 (t)x(t)dt J(x. s)f (s)ds + Cx1 (t). 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. s)f (t)dt − x(s) + Cx1 (s) The LHS vanishes.4. Proof : The existence of a solution of Eq.8) having x1 (t) as the only normalized solution of the completely homogeneous system.4. (7. s) will be used. 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. s)f (t)dt − b a a x(t)δ(t − s)dt +x1 (s) Using Green’s formula. the above deﬁning property of gM (t.
t2 2.4. s)dt = 0 for every s 0 Integrating Eq.9) x(0) = 0 = x(1) ˙ ˙ The completely homogeneous system has a nontrivial solution x = const.10) To compute gM (t.4.4.11) and using boundary conditions. the above system has a solution x(t) b x(t) = a gM (t. t<s t>s gM (t.4.4.4.232 or CHAPTER 7. 0≤t≤1 (7. s) = C +t− Continuity at t = s implies that A− or s2 s2 =C +s− 2 2 C =A−s . we get A− t2 2.4. by Theorem 7. 1 Since given by 0 sin 2πtdt = 0.8 Consider the system d2 x dt2 = sin 2πt. s) is continuous at t = s dgM dgM t=s+ − t=s− = 1 dt dt 1 (7.11) gM (t. BOUNDARY VALUE PROBLEMS b x(t) = a gM (t. s)f (s)ds + Cx1 (t) Example 7. s) we solve the system d2 gM = δ(t − s) − 1 dt2 gM (0) = 0 = gM (1) ˙ ˙ gM (t. s) sin 2πsds + C (7.(7.
which gives A= s− Hence. B1 φ = 0 = B2 φ (7.5. We assume that the operator L acts on L2 [a. x ∈ L2 [a. b] : x.2) with the assumption that λ = 0 is not an eigenvalue of L. b].5.1) = 0 = B2 (x) where a and b are ﬁnite constants and f ∈ L2 [a. s) = A− t2 2. 3 gM (t.SturmLiouville System Lx − λx B1 (x) = f. s) = s − s 2 − 1 − s + t − t2 . . 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.STURMLIOUVILLE SYSTEM 233 So gM (t. Such a selfadjoint system is called SturmLiouville system.5. EIGENFUNCTION EXPANSION . t≤s 1 3 − s2 +t2 2 . We get the following important theorem. As − s 6 + A−s+ − 1 6 − (A − s)s + s2 1 − 2 3 s2 2 = 0. t2 2. 0≤t≤b Consider the selfadjoint system (7. b].5 Eigenfunction Expansion . t ≥ s 2 3 2 = s− t− 1 3 − s2 +t2 2 . giving eigenfunctions of L as a basis for L2 [a. 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.7. t≥s 7. b] with domain D deﬁne as D = {x ∈ L2 [a. b].
2) is equivalent to the solvability of the integral equation b x(t) = λ a g(t. then the boundary value problem a Lw = x. (7.5.2 If λ is not any one of the eigenvalues {λ1 . s)x(s)ds = 0. . (7.5. · · · } and the corresponding eigenfunctions {φ1 . b]. Here (. · · · } form a basis for L2 [a. φn . φn ). then the selfadjoint system Eq. For. if x = 0 exists satisfying b g(t.2.) is the innerproduct in L2 (a.2) has discrete eigenvalues {λ1 .5.3) in the space L2 [a. b]. b]. Proof : As zero is not the eigenvalue of L. s)x(s)ds Then Eq. · · · .2) is selfadjoint. BOUNDARY VALUE PROBLEMS Theorem 7. it follows from Theorem 2. it follows that eigenvalues of L are discrete and the corresponding eigenfunctions {φ1 .1 The selfadjoint system Eq.5. · · · } form a compelete orthonormal set in L2 [a. We observe that a a a b b g(t.5. So. Theorem 7.1) has one and only one solution x(t) ∈ L 2 [a.5.234 CHAPTER 7. (7. b]. b]. As K and L are inverses of each other. g 2 (t. b]. λ2 . φ2 . there exists Green’s function g such that the solvability of Eq. This implies that x = Lw = 0.4) in the space L2 [a. (7. s)x(s)ds = 0. b].5. b] given by ∞ fn φn (t) x(t) = λn − λ n=1 where fn is given by fn = (f.6 that eigenvalues of K are discrete and the eigenfunctions of K form a basis for L2 [a. s)dtds < ∞ and hence the operator K generated by g(t.3) is equivalent to the following operator equation [Kx] = 1 1 x = µx. · · · } of L.5. Let K be the integral operator generated by g(t. · · · . φn . λn . s) b [Kx](t) = a g(t.. s) is a compact operator on L2 [a. · · · . s)x(s)ds (7. φ2 . We note that µ = 0 is not an eigenvalue of K. K is also selfadjoint as the system Eq. λ2 . µ = λ λ (7.(7.5. . B1 (w) = 0 = B2 (w) b has a unique solution w = a contradiction.
(7. Plugging in the value of xn in Eq. fn = (f. xn = (x.1) has inﬁnitely many solutions if fj = 0 and no solution if fj = 0. uniqueness of the solution of Eq.5) and ∞ b x(t) = n=1 xn φn . (7.5. we get ∞ n=1 ∞ ∞ xn (λn − λ)φn (t) = fn φn (t) n=1 That is n=1 [fn − xn (λn − λ)] φn (t) = 0.5.4.5. EIGENFUNCTION EXPANSION .5. This remark follows by Theorem 7.6) Operating on Eq.7.2.2. {φ1 . · · · } implies that fn − xn (λn − λ) = 0 and hence xn = fn /(λn − λ). φ2 . 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 (7.1) follows by Theorem 7.6) we get x(t) = fn φn (t) (λn − λ) n=1 ∞ Remark 7. Eq. (7.5. φ2 .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. (7.5.7) f (y)φn (y)dy a φn (t) λn .STURMLIOUVILLE 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 If λ = λj for some j. Corollary 7.5.5. φn ) = a x(t)φn (t) dt (7.4.5.5. φn ) = a f (t)φn (t) dt (7. b] and hence we can write ∞ b f (t) = n=1 fn φn . · · · } is a complete orthonormal set in L2 [a. Completeness of {φ1 .
It can be √ easily seen that there are no negative eigenvalues. dφ 1 dφ + + λφ = 0 dt t dt . φ(t) = c 1 sin λt √ +c2 cos λt. x(1) = 0 dt We ﬁrst make the above system selfadjoint by deﬁning a new inner product 1 (x. x(1) = x(1) ˙ dt2 2 The associated eigenvalue problem is − 1 d2 φ ˙ = λφ.5. For λ > 0. by Theorem 7. Boundary condition φ(0) = 0 gives c2 = 0 and hence φ(t) = √ c1 sin λt. The second boundary condition implies that λ satisﬁes the equation √ √ tan λ = 2 λ Let λ = k 2 . By considering the curves y = tan x and y = 2x. φ(1) = φ(1) dt2 2 We take up diﬀerent cases. tan k = 2k. BOUNDARY VALUE PROBLEMS Example 7. √ This is a Bessel’s equation√ zero order in the variable λt and has the solution of √ φ(t) = AJ0 ( λt) + BY0 ( λt). But the ﬁrst boundary condition gives B = 0.2. y) = 0 tx(t)y(t) dt and hence its eigenvalues are real. we see that the above equation has inﬁnitely many 2 solutions. The corresponding normalized eigenfunctions are − cos 2kn These eigenfunctions are orthogonal and form a complete set. φ(0) = 0. depending upon the algebraic sign of λ.5. then the above equation becomes tan k = 2k.236 CHAPTER 7. 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 7.5.2 Consider the boundary value problem − t→0 1 d t dt t dt dt = f (t) lim t dx = 0. x(0) = 0. Hence. So eigenvalues are given by λ = kn where kn is a nonzero solution of 2 sin kn t.1 Consider the self adjointsystem − d2 x 1 = f (t). The associated eigen− 1 d t dt t dφ dt + λφ = 0 value problem is That is.
5. φ2 . s. λ) = δn φn (t) . b]. x) is measurable for all values of x ∈ . 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]. s.5. b] × → is nonlinear function satisfying Caratheodory conditons: (i) x → f (t. B2 (x) are boundary conditions on x.1) (7.6. λ) = That is g(t. Here f : [a.2) where L is the second order diﬀerential operator as deﬁned before and B1 (x). t ∈ [a. · · · } forming a complete set in L2 [a. · · · } be a discrete set of eigenvlaues of L with eigenfunctions {φ1 . √ {J0 ( λn t)} forms a complete set and hence by Theorem 7. s.7.2 for f (t) = δ(t − s) (formally) we get g(t. The second boundary condition gives J0 ( λ) = 0. Proof : Green’s function g(t.1) is given by g(t. λ2 . the zeros of the Bessel’s function J0 (t) gives the eigenvalues of the above √ boundary value problem.5.2 √ ∞ 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 7.6 Nonlinear Boundary Value Problem In this section we shall be concerned with two point boundary value problem of the form Lx = f (t. λ) is the solution of the system Lg − λg = δ(t − s). . Applying Theorem 7. Thus. b]. φ(t) = AJ0 ( λt).6. x) is continuous for almost all t ∈ [a.3 Let {λ1 . NONLINEAR BOUNDARY VALUE PROBLEM 237 √ √ Hence. Then the Green’s ∞ φn (t)φn (s) function for Eq. λ) = provided λ is not λn − λ n=1 one of the eigenvalues. (ii) t → f (t. B1 (g) = 0 = B2 (g). The corresponding eigenfunctions are φ n = J0 ( λn t). (7. 7. s. B2 (x) = β (7. x(t)).6.5. b] B1 (x) = α.
BOUNDARY VALUE PROBLEMS Let g(t.6. x) ≤ a(t) + bx. g(t. s)) a . s) be the Green’s function.6.3) where y(t) is a known function given by y(t) = −J(x(s).3) in the space X = L2 [a. s)f (s.5). x(t)) Assumptions I (i) g(t. (7. ∀x1 .10) .6.8) Theorem 7. x2 ) ≤ γx1 − x2  ∀t ∈ [a. a ∈ L2 (a.6. x2 ∈ (7.5) [N x] (t) = f (t. N are as deﬁned by Eq.1) is equivalent to the solvability of the integral equation b x(t) = a g(t.6. x1 ) − f (t.6.9) with x0 (t) = y(t). Then the solvability of the nonlinear boundary value problem given by Eq. (7.238 CHAPTER 7. (7.6.7) (iii) x → f (t. Proof : Solvability of Eq. deﬁned as below b [Kx](t) = a g(t.3) is equivalent to the solvability of the operator equation x = KN x + y where K. Eq.6. That is k2 = a b a b g 2 (t.1 Let Assumptions I hold and let kγ < 1.6. J being the bilinear concomitant of x and g.there exists constant γ > 0 such that f (t. s)dtds < ∞ (7. (7. deﬁned iteratively as 1 xn (t) = T xn−1 (t) = 0 g(t.1) or Eq. (7. s)f (s. We shall prove the existence of solution of Eq. corresponding to the operator L with given zero boundary conditions.6. x) is Lipschitz continuous . b > 0 (7.6. b).4).1) has a unique solution x∗ (t) which is obtained as a limit of the Picard iterates xn (t). b]. respectively. s) is a Hilbert Schmidt kernel . b] by using operator theoretic approach in terms oprators K and N .4) (7. Then Eq. (7. (7.6. (7. xn−1 (s))ds + y(s) (7. s)x(s)ds (7.6. x) satisﬁes growth condition of the form f (t. x(s))ds + y(t) b (7.6.6) (ii) f (t.6.6.
Then T x1 − T x2 = ≤ K(N x1 − N x2 ) K N x 1 − N x2 ∀x1 . 90 3( 10) .1 Let us consider the nonlinear boundary value problem d2 x dt2 x(0) = γ sin x(t) = α.6. x2 ∈ X ≤ kγ x1 − x2 If α = kγ < 1.6.12) This gives Following Example 7.6. Further Assumption I(iii) implies that N is Lipschitz continuous.7. Also. t ≥ s 1 1 g 2 (t. inview of Assumption I(ii).2. (7.6. x2 ∈ X ∀x ∈ X (y ∈ X ﬁxed) X ≤ γ x 1 − x2 X ∀x1 . s) for the above problem is given by t(s − 1). t ≤ s g(t.11) (7.1) where x∗ (t) = lim xn (t). s)dtds = 0 0 1 1 and hence k = K = √ .3. xn−1 (s))ds x0 (t) = y(t) This proves the theorem.1.6. NONLINEAR BOUNDARY VALUE PROBLEM 239 We have already proved in Example 2. x0 = y. s)f (s. N x1 − N x2 Let T x = KN x + y. s) = s(t − 1). Example 7. we ﬁnd that the Green’s function g(t. that is. it follows that N is a bounded continuous operator on X ( refer Joshi and Bose [4] ). we have n→∞ b xn (t) = and y(t) + a g(t. Thus. 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 .8 that K is a compact operator on X. This proves that x∗ (t) is a solution of Eq. x(1) = β (7.
6.6. which is the limit of xn (t). x1 ) − f (t. we may resort to the concept of monotone operators (see Section 2. x) is strongly monotone: ∃ c > 0 such that (f (t. s)x(s)ds is positive semi deﬁnite as K is the d2 x which is positive semi deﬁnite dt2 (monotone). deﬁned. iteratively as 1 xn (t) = α(1 − t) + βt + γ with x0 (t) = α(1 − t) + βt. inverse of the diﬀerential operator Lx = − .11) is equivalent to the solvability of the integral equation 1 x(t) = γ 0 g(t.6.2 Consider the following boundary value problem: d2 x + γx + sin x dt2 x(0) = α. Assumptions II (i) [g(t.6. Solvability of Eq.6. it follows that Eq. 1 √ < 1. x2 )) (x1 − x2 ) ≥ c(x1 − x2 )2 x1 .14) (7.14) is equivalent to the solvability of the integral equation 1 x(t) + 0 g(t. s) sin xn−1 (s)ds 0 If f is monotone increasing. (7. s)] is symmetric positive semideﬁnite ( K is monotone) and Hilbert Schmidt. Example 7. (7. s) sin x(s)ds + y(t) (7. (iii) x → f (t.6.1) has a unique solution.240 CHAPTER 7. g(t.6.2 Under Assumptions II. It is clear that f (x) = γ sin x which satisﬁes Assumptions I(i) and (ii). (ii) Assumption I(ii) holds. x(1) = 0 = β (7.4) to get a solvability results for Eq. x2 ∈ Theorem 7. Eq. (7. Refer Joshi and Bose [4] for the proof of this theorem.6. s)[γx(s) + sin x(s)]ds = α(1 − t) + βt 1 Observe that [Kx](t) = 0 g(t.13) where y(t) = α(1 − t) + βt.6.11) has a unique If we assume that γ 3( 10) solution x∗ (t).6. BOUNDARY VALUE PROBLEMS Solvability of Eq.15) Assume that γ > 1.1). (7. (7.
2.2 is applicable. show that the solution x(t) of the boundary value problem d2 x + tx = 1. ˙ ˙ dt2 3. x) is a strongly monotone function: (f (t. we get 241 . w(t) > 0 ˙ w(t) 4.6. EXERCISES As f (t. x) = γx + sin x. Using Green’s function technique. 6. x1 ) − f (t. x(1) = 0.14) has a unique solution. x2 )) (x1 − x2 ) ≥ (α − 1)(x1 − x2 )2 ∀x1 . x(1) = 0. √ Remark 7.1 whereas Theorem 7. B1 (x) = 0 = B2 (x) depends continuously on f . s)f (s)ds of the system Lx = f.7.6.1 If γ is such that γ > 3( 10) − 1 > 1 then Eq. (7. (7.14) is not solvable through Theorem 7. x) = γ + cos x ≥ (γ − 1) > 0 for all x ∈ ∂x Hence. Prove that the solution x(t) = g(t.6. ˙ ˙ 1 (a0 (t)x) + a2 (t)x.6. x → f (t. Find the modiﬁed symmetric Green’s function for the system L = d2 . y) = 0 w(t)x(t)y(t)dt 1 0 5.7 Exercises 1.7. dt2 d2 with boundary conditions x(0) = x(1). it follows that Eq. Consider the operator L with boundary conditions x(0) = x(1). 7. dt2 −1 ≤ t ≤ 1 with boundary conditions x(−1) = x(1) and x(−1) = x(1). x(0) = x(0) = 0 ˙ dt2 . x2 ∈ As g and f satisfy all Assumptions II. ˙ ˙ Find L∗ and adjoint boundary conditions. ∂f (t. 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).6. Show that the diﬀerential operator Lx = − is formally self adjoint if the inner product is deﬁned by 1 (x.
x(0) = 1 ˙ (b) dt2 d2 x (c) + x = et . Stakgold.. x(0) = 0. 7.242 CHAPTER 7. Bose and M. Compute the ﬁrst few Picard iterates for the following nonlinear boundary value problems d2 x = x + x3 . Bose. d2 x + x = t.. Linear Operators. x(0) = 0 = x(1) ˙ ˙ dt2 (a) References [1 ] E. 1984 [3 ] N. Interscinece Publ. Coddington and N. The Macmillan Co. Tata McGraw Hill Publ. 1966 [4 ] M. Dunford and J. 1955 [2 ] R. Obtain the solution of the following boundary value problems through eigenfunction expansion. C. Joshi and R. x(0) = x(1).. Schwartz. Ltd. x(0) = x(1) ˙ ˙ dt2 (a) 8. x(1) = 0 ˙ dt2 d2 x + x = sin t. K. Inc. 1968 [5 ] I. Halsted Press. x(0) = 0 = x(1) dt2 d2 x (b) = x + sin x.. Some Topic in Nonlinear Functional Analysis.. Theory of Ordinary Diﬀerential Equations.. Methods of Mathematical Physics. co. Parts I and II. Compare this approach with that of the initial value problem. x(1) = 2. Joshi.T. Levinson. K. Boundary Vlaue Problems of Mathematical Physics. C. A. 1968 . 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. McGraw Hill Book Co. Vol I and II.
(8. We substantiate our theory by concrete examples.1. B(t) is n × m matrix.1. ¯ ¯ x(tf ) = xf ¯ ¯ 243 (8.1.1. modelled by ODE.1.1 The linear system given by Eq. Some of the important properties of such systems are controllability. ¯ ¯ Deﬁnition 8.2) .1b) where A(t) is n × n matrix.Chapter 8 Control Theory Of Linear Systems Control theory deals with the maneuver of the state or trajectory of the system. u(t) ¯ ¯ ¯ is called control or input vector and x(t) the corresponding trajectory or state ¯ of the system.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. u(t) ∈ m and x(t) ∈ n . In this chapter we show how operator theoretic approach in function spaces is useful to discuss the above mentioned properties. B(t) are continuous so that the above system has a unique solution x(t) for a given input u(t) . there exist a ¯ ¯ control u(t) so that the correspondoing trajectory x(t) of Eq. We assume ¯ ¯ that the entries of the matrices A(t). observability and optimality.1. (8.1a) (8. 8.1) is said to be controllable if given any initial state x0 and any ﬁnal state xf in n .1) satisﬁes ¯ ¯ the condition x(t0 ) = x0 . The typical controllability problem involves the determination of the control vector u(t) such that the state vector x(t) has the desired properties.
Theorem 8. t) [W (t0 . τ )B(τ )¯ (τ )dτ u Then.1 The system given by Eq.1. (8.3) As (t. t0 )¯0 u ¯ ¯ x L¯ = zf u ¯ This is equivalent to solving .1) reduces to the determination of u ∈ X such that ¯ tf t0 φ(tf . tf ]. τ )B(τ )¯ (τ )dτ u Equivalently.1) is equivalent to the following integral equation t x(t) = φ(t. τ )B(τ )¯ (τ )dτ u t0 m (8. in view of (8.1.1. tf ) = t0 φ(tf . tf ) dτ (8. So controllability of Eq. tf x(tf ) − φ(tf . The controllability x ¯ problem of Eq. τ )B(τ )B (τ )φ (t. it follows that φ(t. t0 )¯0 = ¯ x φ(tf . Then by the variation of parameter formula. ¯ Assume that φ(t.244 CHAPTER 8. t0 )¯0 ] x x Proof : Let (¯0 . τ ) ≤ M for all t.1. τ ) is continuous. (8. 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 . τ ∈ [t0 . the controllability problem reduces to showing that operator L is surjective.1) is controllable iﬀ the controllability Grammian tf W (t0 .4).1.5) [L¯] = u t0 φ(tf . τ )B(τ )¯ (τ )dτ u (8. t0 )¯0 + ¯ x ¯ t0 φ(tf . 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 .1. τ )B(τ )¯ (τ )dτ = zf = xf − φ(tf .1. t0 ) is the transition matrix of the above system. (8. tf )] ¯ −1 [¯f − φ(tf .1.1) is equivalent to ﬁnding u(t) such that ¯ tf xf = x(tf ) = φ(tf . τ ) → φ(t. (8.1. tf ]. Eq.4) Let us deﬁne a linear operator L : X = L2 ([t0 . tf ) by (8.6) is nonsingular.1. xf ) be any pair of vectors in n × n . t0 )¯0 + ¯ x t0 φ(t.
by Theorem 2. t) [W (t0 . it is suﬃcient to prove that LL∗ : n → n is onto (equivalently nonsingular). τ )B(τ )¯ (τ )dτ u n (α. α) u ¯ n = = α. ¯ ¯ Computing LL∗ we get [LL∗ ](¯ ) = L B (t)φ (tf . For u ∈ X and α ∈ n . CONTROLLABILITY 245 in the space X for a given zf . tf )] [¯f − φ(tf . (8. t)B(t)B (t)φ (tf .8. u(τ ) α ¯ n = (¯. This is determined as follows.8) is solvable. ¯ ¯ Conversely. That is.1. where L is given by Eq. L∗ α) = L∗ α 2 and hence ¯ ¯ ¯ ¯ 0}. ﬁnd α ∈ n such that ¯ ¯ LL∗ α = zf ¯ ¯ (8. φ(tf .1. we have ¯ ¯ tf (L¯. This gives 0 = (LL∗ α. then a control u which solves L¯ = zf is given by ¯ u ¯ u = L∗ α.5). ¯ . let the Eq. τ ) is bounded. (8.1. τ )B(τ )¯(τ )) ¯ u t0 tf n dτ dτ = t0 B (τ )φ (tf . N (L∗ ) = [R(L)] = [ n ] = {0}. Then. tf ) is nonsingular. t) dt α ¯ = W (t0 . Let L∗ : n → X be its adjoint. L∗ α)X u ¯ This implies that α [L∗ α](t) = B (t)φ (tf .2.1. (8. That is. ∗ ∗ Hence L is 11. tf )¯ α If the controllability Grammian W (t0 .1) be controllable. given any zf ∈ n . R(L) = n . This implies LL is also 11. we have a control u(t) ¯ given by u(t) = L∗ α ¯ = L∗ (LL∗ )−1 zf ¯ = L∗ [W (t0 . t)¯ ¯ (8. ¯ We ﬁrst note that L is a bounded linear operator as φ(t. Equiva⊥ ⊥ lently.1. L is onto.1. Assume that LL∗ α = 0. α) = (L∗ α. L∗ α = ¯ L∗ 11 implies that α = 0.8) If Eq.1. t)¯ α α tf = t0 φ(tf . ¯ t0 tf φ(tf . t0 )¯0 ] x x which will steer the state from the initial state x0 to the ﬁnal state xf . τ )¯ .7) To claim that L is onto. Thus N (LL∗ ) = {¯ ¯ 0. tf )] −1 zf ¯ −1 = B (t)φ (tf . which is proved as under.
t)] = A(t)W (t0 . . (ii) W (t0 . τ ) dτ dt A(t)φ(tf . τ )B(τ )B (τ )φ(t. τ )¯ α t0 2 n dτ = ≥ 0 for all α ∈ ¯ n . τ )A (t)dτ = A(t)W (t0 . Theorem 8. tf ). tf ) = W (t0 .246 CHAPTER 8. tf ) = W (t0 . τ )B(τ )B (τ )φ (tf . τ )B(τ )B (τ )φ (tf . tf ) has some intersting properties. α) α ¯ n = t0 tf φ(tf . t) + W (t0 . τ )dτ = B(t)B (t) + t0 t + t0 φ(tf . tf ) satisﬁes the linear diﬀerential equation d [W (t0 . This proves The controllability Grammian W (t0 . τ )¯ . t)W (t. (ii) Using Leibnitz rule. t0 ) = 0 (iii) W (t0 . t)A (t) + B(t)B (t) dt W (t0 . t) + φ(tf . CONTROL THEORY OF LINEAR SYSTEMS n Thus W (t0 .2 (i) W (t0 . τ )B(τ )B (τ )φ (t. tf )φ (tf . tf ) is symmetric and positive semideﬁnite. τ )B(τ )B (τ )φ(t. α α ¯ B (τ )φ (tf . t)) = dt d dt t φ(tf . tf ) satisﬁes functional equation W (t0 . tf ) = LL∗ : the result. t) + W (t0 . tf )¯ .1. tf ) = t0 φ(tf . τ ) dτ (i) It is clear that W (t0 . τ )dτ t0 t = B(t)B (t) + t0 t d φ(tf . → n is 11 and hence nonsingular . tf (W (t0 . t) Proof : We have tf W (t0 . τ )B(τ )B (τ )φ (t. t0 ) = 0 follows from the deﬁnition. Further. we get d (W (t0 . t)A (t) + B(t)B (t) Bounadry condition W (t0 . which we now describe.
t) = W (t0 . tf ) as follows t 247 W (t0 . tf ).1.1). Theorem 8. tf ) = t0 φ(tf .1 Theorem 8. tf )φ (tf . τ )dτ tf = W (t0 . in this direction. then the controllability of the linear system is obtained in terms of the rank of the following controllability matrix C = B.1. t) + φ(tf . we get that 0 = B α = B (A )¯ = B (A )2 α = · · · B (A )n−1 α ¯ α ¯ ¯ . An−1 B The next theorem.1.10a) (8. This is a very tedious task.11) Expanding the LHS of Eq.8. (8. t) + φ(tf .1. τ )dτ φ (tf .9) is controllable iﬀ the controllability matrix C given by Eq. if A(t) and B(t) are time invariant matrics A. CONTROLLABILITY (iii) To get the functional equation we express W (t0 . 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 .1.1. (8. τ )B(τ )B (τ )φ (tf .1. It follows that N (C ) = {¯ We shall 0}.9) has rank n. t)W (t. ¯ α implies that 0 = W (t0 . one needs to verify the invertibility of the Grammian matrix W (t0 .1. tf ] ¯ (8. tf )) = {¯ Let α ∈ n be such that W (t0 . τ )B(τ )B (τ )φ (t.11) in a Taylor series arround t = tf . However. Proof : Assume the rank of C = n.1 implies that to check the controllability of the linear system given by Eq.3 The linear autonomous system d¯ x = A¯(t) + B u(t) x ¯ dt x(t0 ) = x0 ¯ ¯ (8.1. prove that N (W (t0 .1. (8. · · · .10b) (8. tf )¯ α 2 tf = t0 tf B φ (tf . This 0}. AB. is due to Kalman [4]. tf )¯ = 0.1. τ )B(τ )B (τ )φ (tf . t) Remark 8. B. τ )dτ tf + t φ(tf . t) t φ(t.
tf )) = {¯ We shall show that N (C ) = {¯ 0}.1. tf )¯ is given by α tf n−1 W (t0 . k=0 αk (tf − τ )B (A )k α = 0. tf )) = {0}. we have ¯ that α ∈ N (W (t0 . CONTROL THEORY OF LINEAR SYSTEMS This implies that α ∈ N (C ) = {¯ ¯ 0}.12). (8. τ )BB φ (t.10). τ )B(τ ) k=0 n−1 αk (tf − τ )B (AT )k α dτ ¯ (8.1. τ ) = αk (tf − τ )Ak and hence W (t0 .1.248 CHAPTER 8. τ )¯ dτ α As φ(tf . Assume that α ∈ N (C ). This implies that ¯ ¯ B (A )k α = 0. We have α W (t0 .5 which gives the linearized motion of a satellite of unit mass orbiting around the earth. τ ) = eA(tf −τ ) . tf )¯ = α t0 φ(tf . let N (W (t0 . it follows by CayleyHamilton theorem (refer problem number 4 in Section 5. 0}. Conversely. tf )¯ = t0 tf 1≤k ≤n−1 (8.13) In view of Eq.5).1. This is given by the control system of the form given by Eq.3.12) φ(tf . This implies ¯ Example 8. that there exist functions αk (tf − τ ) such that n−1 k=0 φ(tf . (8.1 Let us recall Example 2.1. ¯ This proves the theorem. 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 .
we see that the loss of radial thrust does not destroy controllability where as loss of tangential thrust does. 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 . A2 B2 . −2w3 −4w2 0 Since u1 was radial thrust and u2 was trangential thrust. A3 B . the controllability matrix C1 = [B1 . one can easily compute the controllability matrix. C = B. as 0 0 2w 0 0 −4w2 0 0 2w C2 = 0 1 1 0 C2 has rank 4. A3 B2 ]. A2 B. 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] . the above analysis means that we can maneuver the satellite just with radial rocket thrust. AB. In terms of practical importance of satellite in motion. It is intersting to ask the following question. 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.1. . A2 B1 . is given by 0 1 0 −w 2 C1 has rank 3. So. CONTROLLABILITY For this system.8. AB1 . A3 B1 ]. AB2 .
2. (8. (8.2.1b) Consider the inputoutput system of the form The questions concerning observability relate to the problem of determining the values of the state vector x(t). Deﬁnition 8. tf ) which is deﬁned by tf M (t0 .3b) → X = L2 ([t0 . knowing only the output vector y(t) over some ¯ ¯ interval I = [t0 .3a) (8.1).2.3). tf ] of time. CONTROL THEORY OF LINEAR SYSTEMS 8.1a) (8. tf ]. (8. A(t) ∈ n×n .2. τ )B(τ )u(τ )dτ Thus. tf ] if y (t) = 0 on I implies that x(t) = 0 on I. Then the output vector y(t) can be expresssed as y (t) = C(t)φ(t.3) is observable on I = [t 0 .2. Theorem 8.2.2. from Eq. t0 )¯0 = H(t)¯0 x x x where H : t → H(t) is a matrix function which is continuous in t. in terms of the nonsingularity of the matrix M (t0 . Observability of the system Eq.2. t0 ) be the transition matrix of the above ¯ system. then. tf ) = t0 φ (t.3). The following theorem gives the observability of the system given by Eq. reduces to proving that L is oneone.2 Observability x d¯ = A(t)¯(t) + B(t)¯(t) x x dt y (t) = C(t)¯(t) ¯ x (8.2.2.2. (8.250 CHAPTER 8. based on the output y (t).2) y1 (t) = t0 C(t)φ(t.1 We shall say that Eq. Let φ(t. we need only the homogeneous ¯ ¯ system d¯ x = A¯(t) x dt y(t) = C(t)¯(t) ¯ x in place of Eq.2. tf ).5) . t0 )¯( t0 ) + y1 (t) ¯ x ¯ where y1 (t) is known quantity of the form ¯ tf (8.1 For the homogeneous system given by Eq. t0 )dt (8.2) it follows that if we are concerned about the determination of x(t0 ). As in the previous section. n ) as (8. t0 )C (t)C(t)φ(t. (8. B(t) ∈ n×m are assumed to be continuous functions of t.2. (8.4) [L¯0 ](t) = C(t)φ(t.3).2. ¯ ¯ We deﬁne a linear operator L : n (8. it is possible to determine the initial state x(t0 ) within an additive constant vector which lies ¯ with null space of M (t0 .2.
If L∗ L is invertible. t1 )A(t) − C (t)C(t) dt M (t1 .2.7) is given by [L∗ u] = tf φ (t.2.2. This implies that L∗ L is nonsingular.2.2. tf ) = M (t0 . Premultiplying Eq. (8. tf ). M (t0 .6) by L∗ . tf ) is symmetric and positive semideﬁnite. x(t0 ) is deter¯ mined up to an additive constant lying in the null space of L∗ L = M (t0 . t1 ) − M (t. (ii) M (t0 . Conversely. (8. tf ) satisﬁes the following properties (i) M (t0 . t1 )] = −A (t)M (t. tf ) = L∗ L. Eq. the initial state x(t0 ) is given by ¯ x(t0 ) = (L∗ L)−1 L∗ y = [M (t0 .2 The observability Grammian M (t0 . t0 )¯(t0 ) for ¯ ¯ x x(t0 ) is equivalent to solving the operator equation ¯ L¯(t0 ) = y x ¯ (8. tf ) satisﬁes the following matrix diﬀerential equation d [M (t.2. t1 ) = 0 (iii) M (t0 .2. t0 )C (t)u(t)dt t0 (8. t) + φ (t. then it is clear that L is oneone and hence the system given by Eq. tf ) is nonsingular.2. t0 )C (t)y(t)dt t0 (8. Theorem 8.2.2.5) is called the observability Grammian. tf ) is nonsingular. tf )] ¯ This proves the theorem. Proof : If the output vector y (t) is known. tf ) satisﬁes the functional equation M (t0 .6) where L is deﬁned by Eq. tf ) deﬁned by Eq. t1 )φ(t. t0 )M (t.8. (8.7) is known.2. (8. ¯ (8. (8. OBSERVABILITY 251 Hence x(t0 ) is uniquely determined if M (t0 . solving y (t) = C(t)φ(t.2. If RHS of Eq. For such an observable system. we get [L∗ L] x(t0 ) = L∗ y ¯ ¯ where L∗ : X → n (8.3) is observable. t0 ) −1 tf φ (t.9) .8) Observe that M (t0 .4). if L is oneone then so is L∗ L and hence R(L∗ L) = [N (L∗ L)]⊥ = n . That is.3) is observable iﬀf M (t0 .
1.10) Theorem 8. Example 8. y = (y1 . x2 . consider the satellite orbiting around the Earth.2.1 As a continuation of Example 8.10) is of rank n. CAn−1 ] (8. We assume that we can measure only the distance r of the satellite from the centre of the force and the angle θ.3) is observable iﬀ the observable matrix O given by Eq. 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. y2 = θ being the radial and angle ¯ ¯ measuments. (8.2. CONTROL THEORY OF LINEAR SYSTEMS The proof of the above theorem follows along the lines of the proof of the Theorem 8. See Brocket [1] for more details. (8.252 CHAPTER 8. the observability matrix O is given by O = [C.2. the deﬁning stateoutput 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 . So.2. CA. We also have the following theorem giving the necessary and suﬃcient contition which is easily veriﬁable for observability. y1 = r. So.2. CA2 . · · · .2.1.3 The system given by Eq. Let us denote by O the observability matrix O = [C. x4 ). y2 ). x3 . CA.1. .
¯ tional J(u) which is given by tf m ) that minimizes the cost func J(¯) = u t0 [(¯(t). C2 A. C2 A3 ] 0 0 0 −6w3 0 0 −2w 0 1 0 0 0 1 0 which is of rank 4. 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. tf ].8. (i) Unconstrained optimal control problem: Find the control u∗ ∈ U = L2 ([t0 .3. if y1 is not measured. C1 A.2a) (8.3.3.3.3 Optimal Control In this section we shall be concerned with the optimization problem of two types. C1 A2 . the state of the satellite is known from angle measurements alone but this is not so for radial measurements. C2 A2 .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. Thus. OPTIMAL CONTROL 253 To minimize the measurement. = 0 −4w2 8. v (t)) + (R¯(t).2b) . we get C2 O2 0 = [C1 . C1 A3 ] = 3w2 0 = [0 0 1 0 0] and = [C2 . However. x(t))] dt u ¯ x ¯ (8. we might be tempted to measure y1 only not y2 .
n ).3. determined by Eq.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 ] ¯ (8.1) from the initial state u ¯ ¯ x0 to the ﬁnal state xf }. (8.3.6) (8.3.8) . t x(t) = φ(t. R is a positive deﬁnite symmetric ¯ matrix and X = L2 ([t0 .3) Uc = {¯ ∈ U : u steers the state x(t) of Eq. CONTROL THEORY OF LINEAR SYSTEMS corresponding to the given control u(t). (8. Using the transition matrix φ(t. the cost functional J on U .254 CHAPTER 8. ¯ ¯ We shall show that unconstrained problem is a problem of minimization of a strictly convex functional deﬁned on the Hilbert space U . (ii) Constrained optimal control system: Find the control u∗ that minimizes J(¯) over the constrained set Uc ⊆ U. u)U + (T x. tf ]. is given by J(¯) = (¯.3.3. then in abstract notation.7) (8. we obtain the state x(t) of the linear ¯ system given by Eq. we obtain J (¯) = u + K ∗ T K u + K ∗ T K y0 u ¯ ¯ ¯ and J (u) = I + K ∗ T K ∀ u ∈ U ¯ (8.2) through the integral equation.3).3.1).3. ¯ u where J(¯) and Uc are given as follows u tf J(¯) = u t0 (¯(t).5) Theorem 8. t0 )¯0 . (8. u(t)) dt u ¯ (8.3. τ )B(τ )u(τ )dτ (8.3. x ∈ X is given by ¯ x ¯ x = y0 + K u ¯ ¯ ¯ So. t0 ). τ )B(τ )¯ (τ )dτ u Set y0 (t) = φ(t. Using the concept of the derivative of a functional (refer section 2.3.4) We now deﬁne a bounded linear operator K from U to X as follows. t [K u](t) = ¯ t0 φ(t. t0 )¯0 + ¯ x t0 φ(t. x)X u u ¯ ¯ ¯ where T is a positive deﬁnite operator induced by R on X.
1 One can obtain integrodiﬀerential equation analog of Eq.8. We have the following theorem in this direction.2) is autonomous. we get the equivalent integrodiﬀerential equation dt d¯∗ u = −BT Ku∗(t) − dt tf t φ(τ. t)ABT Ku∗ (τ )dτ = T K y0(t) ¯ (8. t)A.3.(8. v )U + (K ∗ T K v.3. K v )X ¯ ¯ (for v = 0) ¯ in view of positive deﬁniteness of T . (8.8) which is equivalent to the equation u∗ (t) + ¯ t tf φ(τ. (8.3. 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. v ) = (¯. we have u∗ + K ∗ T K u∗ = −K ∗ T K y0 ¯ ¯ ¯ Also.3. We now address ourself to second problem pertaining to constrained optimal control.3.9) we get d¯∗ u − BT Ku∗(t) + dt As tf t d [φ(τ. the invertibility of [I + K ∗ T K] (refer Joshi and Bose [3]). as (J (u)¯.3. t)] BT Ku∗(τ )dτ = T K y0(t) ¯ dt d [φ(τ. it follows by Remark 2. As J (u) is positive deﬁnite. t)] = −φ(τ. ¯ Note that J (u) : U → U is positive deﬁnite for all u ∈ U.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). v )U v ¯ v ¯ ¯ ¯ = > v ¯ v ¯ 2 2 + (T K v. Hence. t)B(T K u∗ )(τ )dτ + ¯ t tf (T K)(¯0 (τ ))dτ = 0 y (8. . gives u∗ = −[I + K ∗ T K]−1 (K ∗ T K y0) ¯ ¯ Remark 8.9) Diﬀerentiating Eq.10a) (8. In the above equation we assume that the system Eq.3.3.10b) with u∗ (tf ) = 0.
α)U u ¯ u ¯ (8. (8.14) where α solves Eq.13) t0 Moreover. tf ) is nonsingular.3. Hence. tf ) be the controllability Grammian of Eq. u∗ )U + u∗ u ¯ ¯ − u∗ ¯ 2 U 2 U . L∗ α)U = (L¯. u∗ )U u ¯ u u u ¯ (8.11) is equivalent to deﬁning u∗ = L∗ α. t0 )] x0 ¯ x ¯ x ¯ tf Proof : We have zf ¯ zf ¯ = xf − φ(tf . u∗ )U u ¯ = (¯∗ .15) gives 0 ≤ u − u∗ ¯ ¯ 2 U = (¯ − u∗ .3.15) Eq.2 Let W (t0 .1.3. (8. L∗ α)U = (¯. t)¯ ¯ α where α satisﬁes the equation ¯ α ¯ x W (t0 . then ¯ ¯ t0 u∗ (t) 2 dt ≤ u(t) 2 dt ¯ (8. then tf t0 u∗ (t) 2 dt = W −1 (t0 . t0 )¯0 = ¯ x = xf − φ(tf . if W (t0 . Eq.3. controllability vector belonging to Uc . we have L¯ − L¯∗ = 0 u u This gives (L¯∗ . t0 )¯0 = ¯ x φ(tf .3.3. (8. If u∗ is any control of the form ¯ u∗ (t) = B (t)φ (tf . α)U = (¯. [¯f − φ(tf .3. t)¯0 tf tf (8.3. τ )B(τ )¯(τ )dτ = L¯ u u t0 tf t0 φ(tf .12). (8. t0 )] x0 .5). (8.3.12) is equivalent to the ¯ solvability of the following equation LL∗ α = zf ¯ ¯ Also Eq.14) gives ¯ ¯ u∗ ¯ 2 U = (¯∗ . Subtracting the above two equations. α)U = (L¯. CONTROL THEORY OF LINEAR SYSTEMS Theorem 8. τ )B(τ )u∗ (τ )dτ = L¯∗ u where L is deﬁned by Eq. (8. Further.2). Solvability of Eq. tf )¯ = xf − φ(tf .3. (8.3. u − u∗ )U u ¯ ¯ ¯ = = u ¯ u ¯ 2 U 2 U − 2(¯. tf ) [¯f − φ(tf .11) (8.256 CHAPTER 8.3.12) Then u∗ ∈ Uc . if u is any.
L∗ α) ¯ ¯ ¯ ¯ = (LL∗ α. The energy dissipated in the resister in the t time interval [0.1: Electrical network Its equation is given by d (cv(t)) = i(t). tf ) [¯f − φ(tf .3. Thus. tf ) is invertible. r i v c Figure 8. t)¯0 ]) Example 8.1 Consider the following electrical network. we get u∗ ¯ 2 = L∗ α ¯ 2 = (L∗ α.16) i is the current out of the source. 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. α) α ¯ x x x = (¯f − φ(tf . (8.3. tf ] is 0 f ri2 (t)dt .3.8.1.3. W −1 (t0 . t . tf )¯ . t)x0 . given by Figure 8.16) with the constraint that v(0) = v0 and v(tf ) = v1 . u∗ (t)) u ¯ m dt ≤ (¯(t). c = Capacitance dt (8. if LL∗ = W (t0 . α) = (W (t0 .3. OPTIMAL CONTROL That is u∗ ¯ or tf t0 2 U 257 ≤ u ¯ 2 U tf (¯∗ (t). we want to minimize 0 f i2 (t)dt subject to the output voltage satisfying the equation Eq.3. u(t)) u ¯ t0 m dt Further.
CONTROL THEORY OF LINEAR SYSTEMS tf For this problem. ··· ··· 1 −αn 0 0 0 A= .norm. B = where A(t) = 1 0 −1 1. 1) at time tf = 1. . Show that the linear system given by the pair {A. . Show that the system is controllable. 2. 0 0 0 . 0 −α1 0 −α2 0 −α3 1 0 0 0 1 0 ··· ··· ··· .4 Exercises d¯ x = A(t)¯(t) + B u(t) x ¯ dt 1 et ¯ 0 . tf ]. the controllability Grammian W (t0 . 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. . 1 3. ¯= b 0 0 0 . (∗) (∗∗) . tf ) = 2 and hence the c c optimal current i∗ (t) = (v1 − v0 ) and hence is constant on the interval [0. tf For more on control theory refer Brockett [1] and Gopal[3]. . ¯ is controllable for all b} values of αi ’s arising the deﬁnition of A. Find the control u∗ (t) with minimum L2 .258 CHAPTER 8. . which ¯ will steer the trajectory from the initial state ¯ at t = 0 to the ﬁnal state 0 (1. . That is (*) is controllable iﬀ (**) is observable. 8.
y0 and y1 .8. EXERCISES 4. · · · . 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. 259 Show that if C = B. AB.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. y(tf ) = y1 . Show that there exists an initial state for the adjoint system d¯ p = −A (t)¯(t) p dt such that the control u(t). An−1 B has rank n. ˙2 ˙2 8.1. 6. Obtain controllable conditions analogous to Theorem 8. is given by ¯ ¯ u(t) = −B (t)¯(t) ¯ p . y(tf ) = y1 ). then the above system is controllable. Show that. which minimizes the functional ¯ tf φ(¯) = u t0 (¯(t). y(t0 ) = y0 . u(t)) dt u ¯ for steering the initial state x0 to the ﬁnal state xf . ˙ ˙ ˙ ˙ 2 provided y0 + y0 and y1 + y1 are nonzero. b 5.4. Consider the system d¯ x = g(t) [A¯(t) + B u(t)] x ¯ dt with g continuous and bounded (0 < α ≤ g(t) ≤ β < ∞) . y1 . 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 . there exists ˙ ˙ a control u(t) such that y(t0 ) = y0 . if the time invariant system d¯ x = A(t)¯(t) + B u(t) x ¯ dt is controllable.
Brockett. Ltd.. 1985 [4 ] R. Gopal. Moudgalya.. 2004 . Mathematical Description of Linear Dynamical Systems. Joshi and Kannan M. 1989 [3 ] Mohan C. Narosa Publishing House. Modern Control System Theory. 1963 [5 ] Mohan C. Pvt. Halsted Press. Inc. John Wiley and Sons. Kalman. Some topic in Nonlinear Functional Analysis. CONTROL THEORY OF LINEAR SYSTEMS References [1 ] Roger W. 1970 [2 ]M. E. pp152192. Vol I.260 CHAPTER 8. Optimization Theory and Practice. Finite Dimensional Linear Systems. SIAM J Control. Wiley Eastern Ltd. Joshi and Ramendra K. Bose.
20. 96. 239 . 48. 80. 77. 12. 75. 26. 81. 39–41. 148 asymptotically stable. 256 controllability Grammian. 237 Boundary condition. 97. 237. 156. 234. 166. 209–211. 97 asymptotic convergence. 195. 242 bounded linear operator. 153– 155. 132. 137. 131–133 First Order Equations. 237 Boundary value problem. 142. 135 Banach space. 21. 219–221. 27 diﬀerential operator. 210 adjoint system. 132. 1. 219 adjoint operator. 56. 71 Diracdelta function. 144. 244–246. 34. 9. 216. 196. 200. 68. 154 Euler’s system. 214. 239 compact set. 127. 240. 96. 63–65. 21. 248 compact operator. 63. 215. 1. 45. 234–237. 69. 147. 212. 149–151. 254 bounded solutions. 97 261 CayleyHamilton theorem. 173 autonomous system. 225. 227. 216–222. 71. 139. 207–209. 9. 226. 4. 115. 171. 241 boundary value problems. 4 Cauchy sequence. 73 direction ﬁeld. 97 Comparision test. 217. 241 Boundary Value Problem. 154. 259 Analytic Coeﬃcients. 254 diﬀerential equation. 209. 207. 177 continuous linear operators. 20. 76. 54. 168. 234 eigenvectors. 233 eigenfunction expansion. 26 ﬁxed point. 67. 247. 247. 125 eigenvalues. 112. 230–232. 256. 68 Controllability. 39 CauchyPeano theorem. 49 contraction mapping principle. 156. 48. 50. 145 carrying capacity. 56. 120–123. 111. 137–139. 242 Eigenvalues. 222 boundary conditions. 258 controllability matrix. 211 boundary value problem. 6. 249 convex functional. 207. 237. 224. 142. 118. 171. 247 autonomous systems. 104 adjoint Green’s function. 211. 236 Boundary conditions. 176 comparision test. 181 AscoliArzela theorem. 3. 243 controllability. 129. 2. 245. 34 Eigenfunction Expansion. 168 equilibrium points.Index Abel’s formula. 240 DiracDelta Function. 55. 11. 214. 244. 211. 238. 136. 168. 55. 65 Bessel’s function. 212. 71.
241 Gronwall’s inequality. 211. 207. 107. 70. 24 neural solution. 76 Lyapunov function. 235 Parseval’s equality. 43. 242. 63. 38–41. 59. 10. 142. 210 fundamental matrix. 107–109. 227 Lipschitz continuity. 193. 38. 81. 187 Hilbert Schmidt kernel. 37. 9 Harmonic Oscillator. 258 linearization. 25 Newton’s algorithm. 238 Hilbert space. 209. 145. 50 inﬁnite series. 49 Jordan canonical form. 2. 239 nonlinear boundary value problems. 43. 22. 107. 244. 200 generating function. 107. 39. 187 linear ODE. 125 metric space. 95. 238.262 ﬁxed point theorem. 56. 114. 53. 79 inner product space. 69. 23. 132. 93. 21. 201. 154. 255 Isocline. 75. 102. 165. 54. 93. 96. 222–225. 242 normed space. 10. 134. 138– 140. 23. 226–231. 162. 231 monotone operator. 63. 68 numerical methods. 56. 96 modiﬁed Green’s function. 62. 238. 61. 147 nonlinear boundary value problem. 39. 129. 44. 20. 197. 163 maximal monotone. 140. 234. 111. 110. 166. 234 integrodiﬀerential equation. 123. 62 nonautonomous system. 186. 55. 22 observability. 185 Legendre diﬀerential equation. 145. 70 motion of a satellite. 140. 212. 41–43. 13 harmonic oscillator. 54 . 214– 217. 102 linearly independent. 21. 248 neural network. 19 orthogonal matrix. 94. 4. 67. 143 growth rate. 3. 80. 64. 63 linearly dependent. 229. 21. 144. 112. 148 formal adjoint. 163. 66. 234. 41 integral equation. 38. 234. 202 gradient. 94. 84 locally integrable. 101. 7 linear system. 253 orbit. 88 mechanical oscillations. 187 Hermite Diﬀerential Equation. 59 GramSchmidt orthonormalisation. 2. 24. 56 orthonormal set. 150. 49. 62. 80. 219–223. 168–171 Lyapunov Stability. 97 FloquetLyapunov theorem. 137 Generating function. 254 Hilbert spaces. 117. 254 integral operator. 3. 146. 250–252 optimization. 218. 39. 237–239. 101– 107. 6 isomorphism. 157. 196. 46 Green’s function. 254. 73. 98. 212. 60. 2. 79. 57. 164. 152 Initial Value Problem. 106. 175 initial condition. 123 generalized eigenvectors. 187 Hermite diﬀerential equation. 233. 240. 149. 47. 59. 150 generalized eigenvector. 147 Gronwall’s lemma. 146. 137 INDEX Legendre Diﬀerential Equation. 50. 91. 79. 120.
150. 158 vector space. 56. 23. 188 Rodrigues formula. 138. 113. 41. 222 263 . 244. 134. 135. 158. 68 Schwartz inequality. 112. 258 transition matrix. 48 singular distribution. 5. 85 stable focus. 140. 130. 197 Root test. 244. 12. 3. 2. 160 stable node. 120. 161 phase portraits. 243. 102. 2. 158. 8. 14. 1. 73 Wronskian. 121. 101. 147 phase portrait. 134. 134 Periodic Linear System. 106. 173 phase space. 155. 155. 9–11 Ratio test. 238. 148. 154. 37–39. 244 vector ﬁeld. 114 perturbed system. 107–109. 242 positive deﬁnite operator. 2. 160 variation of parameter. 15. 112–114. 129. 1. 140. 123. 141.INDEX PeanoBaker series. 4. 233 test functions. 154 Picard iterates. 114. 72. 108. 81–83. 73 trajectory. 100. 250. 19. 156. 158 SturmLiouville system. 125–127. 159. 74 solution curve. 73. 146. 176 regular singularity. 164. 145. 148. 117. 118. 177 Schauder’s theorem. 165. 8. 254 predatorprey system. 11. 254 unstable focus.