Chapter 3

REVIEW OF BASIC CONCEPTS

In this chapter, we will briefly review some basic concepts of linear systems: differential
equations, dynamical systems, flows and fundamental theory. In the following, D.E. and
D.S. will be used to denote the Differential Equation and Dynamical Systems, respectively.

3.1 D.E., D.S. and Flows

Consider the D.E.
 
02
ẋ = A x where A= 1  and x = (x1 x2 ) ∈ R2 . (3.1)
0 −
2
This is an autonomous system given in 1st-order, standard vector form (system). The solu-
tion of (3.1), given by
(
x1 = u1 e2t , u1 = x1 (0)
(3.2)
− 21 t
x2 = u2 e , u2 = x2 (0) ,

is trivial since the system is uncoupled. Here, the initial condition u = (u1 , u2 ) is a
given point in R2 . Geometrically, this means that when t = t0 the curve is required to
pass through the given point. From a geometrical viewpoint, regard the solution (3.2) as
specifying a curve in R2 , i.e.

x : R → R2 , x(t) = (x1 (t), x2 (t)) (3.3)

and the R.H.S. (Right-Hand-Side) of (3.1), A x , is the tangent vector to the curve. The
initial condition requires the curve to pass through (u1 , u2 ) .
The map (i.e. function) A : R2 → R2 , (or x → A x) can be considered as a vector
field on R2 . This means that to each point x in the plane we assign the vector A x . For
purpose of visualization, we picture A x as a vector “based at x”; that is, we assign to x
the directed line segment from x to x + A x . For example, for x = (1, 2) , then at (1, 2)

39
AM 9524 Chapter 3 40

we picture an arrow pointing from (1, 2) to (1, 2) + (2, − 1) = (3, 1) . Thus, we attach to
each point x in the plane an arrow with tail at x and headed at x + A x and obtain the
picture Fig 3.1.
x2

x1

1
Fig. 3.1 Visualization of a vector field A x = (2 x1 , − x2 ) .
2

x2

x1

 
2 0
Fig. 3.2 Some solution curves (“phase portrait”) to ẋ = A x , A = .
0 − 1/2

Solving the differential equation (3.1) with initial condition (u1 , u2 ) at t = 0 means
finding in the plane a curve x(t) that satisfies (3.1) and passes through the point u =
(u1 , u2 ) when t = 0 . A few solution curves are sketched in Fig. 3.2. The trivial solution
(x1 (t), x2 (t)) = (0, 0) is also considered a “curve”. The family of all solution curves as
AM 9524 Chapter 3 41

subsets of R2 is called the phase portrait of equation (3.1).

Now let us consider the differential equation (3.1) as a dynamical system. This means
that the independent variable t is interpreted as time and the solution curve x(t) could be
thought of, for example, as the path of a particle moving in the plane R2 . As time proceeds
the particle moves along the solution curve x(t) that satisfies the initial condition u = x(0) .
At any later time t > 0 the particle will be in another position x(t) and at an earlier time
t < 0 , the particle was at a position x(t) . To indicate the dependence of the position on
t and u we denote it by φt (u). Since A x is independent of t (for system (3.1)) we may
write
1
φt (u) = (u1 e2t , u2 e− 2 t ) . (3.4)

We can imagine particles placed at each point of the given place, all moving simultane-
ously. The solution curves are spoken of, in this context, as trajectories or orbits. For each
t ∈ R , we have a transformation assigning to each point u ∈ R2 another point φt (u). This
transformation, denoted by φt : R2 → R2 , is clearly a linear transformation, since

φt (u + v) = φt (u) + φt (v)

φt (λu) = λ φt (u) ∀ λ ∈ R and u, v ∈ R2 . (3.5)

As time proceeds, every point of the plane moves simultaneously along the trajectories
passing through it. In this way, the collection of all these transformations {φt } forms a
one-parameter family of transformations called the flow or dynamical system of D.E. (3.1).
The above discussion and results can be generated to Rn :

ẋ = f (x) (f linear)

φ t : U → Rn (3.6)

{φt } flow, etc. are demonstrated in Fig. 3.3.

It is clear that

φ0 = id. (id. means identity, i.e. Rn )

and it can be proved that

φt+s = φt φs . (3.7)
AM 9524 Chapter 3 42

I
( ) R
a b
φ ( , x 0)
n
R

U )

x0
(

(a)

n
R

φt (U)

(b)

Fig. 3.3 A solution curve and flow: (a) the solution curve φt (x0 ); (b) the flow φt .

Hence, {φt } forms a one-parameter group of transformations.

3.2 Homogeneous Linear Systems

The equation (3.1) considered above is a simple, uncoupled system. More typical of the
general case is the following example of a coupled system:
 
5 3
ẋ = A x , A= . (3.8)
−6 −4

The idea here is to change the variables so that (3.8) becomes uncoupled and then the
previous work applies. First, find the eigenvalues of A: the characteristic polynomial of
AM 9524 Chapter 3 43

system (3.8) is given by

 
λ − 5 −3
det [ λ I − A ] = det = λ2 − λ − 2 = (λ − 2)(λ + 1) = 0 (3.9)
6 λ+4

which, in turn, results in the eigenvalues

λ1 = 2 and λ2 = − 1 .

The associated eigenvectors can be easily found from the equation [ λi I − A ] f = 0 for
i = 1, 2 . Thus, the vector f 1 = (1, − 1) ∈ R2 is a basis for the eigenspace belonging to λ1 ,
and similarly, f 2 = (− 1, 2) is a basis for the eigenspace belonging to λ2 . Hence, we can
write the transformation

(x1 , x2 ) = (y1 − y2 , − y1 + 2 y2) , (3.10)

i.e. y1 and y2 are the coordinates of x in the new basis.

In the new basis, the original D.E. (3.8) can be described in terms of y as the following
equation:  
2 0
ẏ = B y , B= . (3.11)
0 −1
The system is decoupled, and from the results given in the previous subsection we know that
the solution of (3.11) is
y = (v1 e2t , v2 e− t ) . (3.12)

The flow of the system (3.11) is shown in Fig. 3.4.

By straightforward manipulations one can go back to the original coordinates:

x1 (t) = (2 u1 + u2 ) e2t − (u1 + u2 ) e− t

x2 (t) = − (2 u1 + u2 ) e2t + 2 (u1 + u2 ) e− t (3.13)

where (u1 , u2) = (x1 (0) , x2 (0)) . The phase portrait of the original system is shown in
Fig. 3.5, where the line L1 given by y1 = 0 is described by 2 x1 + x2 = 0 in the original
coordinates, solved directly from equation (3.10). Similarly one can find the line L2 .
The idea which emerged in this example can be generalized:
AM 9524 Chapter 3 44

y2

y1

Fig. 3.4 Flow of system (3.11) in the y coordinates.

x2

x1

L 2 : y2 = 0
L 1 : y1 = 0

Fig. 3.5 Flow of system (3.8) in the x coordinates.

Theorem 3.1. Let A be a linear operator on Rn . Then a solution of

ẋ = A x , x(0) = k ∈ Rn (3.14)

is given by
x(t) = et A k (3.15)

and there are no other solutions.

 
a 0
Example 3.1. A = (a, b = constants) .
b a
AM 9524 Chapter 3 45

 
tA ta 1 0
Since e =e ,
tb 1

Hence,
x1 (t) = et a k1

x2 (t) = et a (t b k1 + k2 ) . (3.16)

It’s useful to employ the two-dimensional case to see that we can get useful
 information

a11 a12
about the solutions of ẋ = A x from just the eigenvalues of A . Let A = ,
a21 a22
then the characteristic polynomial is given by

λ2 − p λ + q = 0, (3.17)

and so we have eigenvalues

1 p 1 √
λ1 = (p − p2 − 4 q) = (p − ∆)
2 2
1 p 1 √
λ2 = (p + p2 − 4 q) = (p + ∆) (3.18)
2 2
where

p = a11 + a22 trace

q = a11 a22 − a12 a21 determinant

∆ = p2 − 4 q discriminant . (3.19)

Based on the eigenvalues λ1 and λ2 (or the coefficients aij ), there exist seven cases,
considered below.

Case I: Saddle. when λ1 and λ2 are both real and have opposite signs, i.e. λ1 < 0 <
λ2 (q < 0) .
A suitable change of the coordinates x = P y puts the equation ẋ = A x in the form
 
λ1 0
ẏ = B y where B = P A P =
−1
. Thus, the phase portrait of the new system
0 λ2
looks like Fig. 3.4, and the phase portrait of the original system is depicted in Fig. 3.6.

Case II: Node. when λ1 and λ2 are distinct, non-zero and have same signs, i.e. either (i)
stable node if λ1 < λ2 < 0 (p < 0) or (ii) unstable node if 0 < λ1 < λ2 (p > 0) . This
case is demonstrated in Fig. 3.7.
AM 9524 Chapter 3 46

Fig. 3.6 Case I: saddle.

(i) (ii)
Fig. 3.7 Case II: node: (i) stable node; (ii) unstable node.

Case III: Focus (spiral). when λ1 and λ2 are a pair of complex conjugates, i.e. λ1,2 =
1 1√
a ± i b where a = p and b = − ∆ > 0 (p =
6 0 , ∆ < 0) . Then one can change the
2 2
coordinates to obtain the equivalent system
 
a b
ẏ = B y , B= (3.20)
−b a
and
 
tB ta cos tb sin tb
e =e . (3.21)
− sin tb cos tb
Thus, the solution is given by

y = et B k = et a (k1 cos tb + k2 sin tb , k2 cos tb − k1 sin tb) . (3.22)

There are two subcases: (i) stable focus if Re(λ’s) < 0, i.e. p < 0 ; (ii) unstable focus if
Re(λ’s) > 0, i.e. p > 0 . This case is shown in Fig. 3.8. If the matrix B in equation (3.20)
AM 9524 Chapter 3 47

 
a −b
is chosen as B = , then how should you change Fig. 3.8?
b a

(i) (ii)
Fig. 3.8 Case III: focus: (i) stable focus; (ii) unstable focus.

The three cases discussed above are robust cases since a small perturbation of the
system (in terms of changing aij ) does not change the type of trajectories (or stability
which will be discussed in the next chapter). The next four cases are non-robust cases.

Case IV: when one of λ1 and λ2 is equal to zero. Without loss of generality, assume that
λ1 = 0 . Then we have either (i) λ2 = p < 0 or (ii) λ2 = p > 0 . This case is shown in Fig.
3.9.

(i) (ii)
Fig. 3.9 Case IV: λ1 = 0 : (i) λ2 < 0 ; (ii) λ2 > 0 .

1
Case V: when λ1 = λ2 = p (∆ = 0) .
2
(i) a12 = a21 = 0 (∆ = 0 ⇐⇒ a11 = a22 ) . In this case, A is of simple structure and it
can be diagonalized. There are two subcases: (a) degenerate stable node if p < 0 ; (b)
AM 9524 Chapter 3 48

degenerate unstable node if p > 0 . A degenerate node is also called an improper node.
This subcase is depicted in Fig. 3.10.

(a) (b)
Fig. 3.10 Case V(i): (a) degenerate stable node; (b) degenerate unstable node.

(ii) a12 a21 6= 0 . In this case, A does not have a simple structure and cannot be diago-
nalized. There is a change of coordinates x = P y giving the equivalent equation
 
λ 1
ẏ = B y , B = P AP =
−1
(3.23)
0 λ

and the solutions look like Fig. 3.11.

(a) (b)
Fig. 3.11 Case V(ii): a12 a21 6= 0 : (a) p < 0 ; (b) p > 0 .

√
Case VI: Center. when λ1,2 = ± i ω where ω = q (p = 0 , q > 0). Now the solutions of
the system are periodic with the same period. To see this, change coordinates to get
 
0 b
ẏ = B y , B= . (3.24)
−b 0
AM 9524 Chapter 3 49

We know that  
tB cos tb sin tb
e = . (3.25)
− sin tb cos tb
2π
Hence, y(t + ) = y(t) . So in the y coordinates the phase portrait consists of concentric
b
circles, which will be ellipses in x coordinates (see Fig. 3.12).

Fig. 3.12 Case VI: center.

Case VII: when λ1 = λ2 = 0 ( p = q = 0). This is shown in Fig. 3.13.

Fig. 3.13 Case VII: λ1 = λ2 = 0 .

The four non-robust cases (IV)–(VII) represent borders between the robust cases; that
is, any change no matter how small in the components of A can cause them to change to
another case. (This is referred as structural instability.)
All the seven cases discussed above can be summarized in a single graph (see Fig. 3.14).
Note that case (V) is not sensitive regarding the case of stability, only in the case of real to
 AM 9524 Chapter 3 50

complex. Thus, the only border line cases in terms of stability, are along the p-axis, and
positive q-axis. So, (asymptotic) stability is in the second quadrant only.

q
Case VI:
Center
Case V: Re(λ’s) = 0 Case V:
∆= 0 ∆= 0
λ 1 = λ2 < 0 Case III: Focus Case III: Focus λ 1 = λ2 > 0
Re(λ’s) < 0 Re(λ’s) > 0
Case II: Node Case II: Node
λ 1< λ 2 < 0 0 < λ1 < λ2

p
Case VII:
Case IV: λ 1= 0 λ 1= λ 2 = 0
Case IV: λ 1 = 0
(λ 2 < 0) (λ 2 > 0)
Case I:
Saddle
λ 1< 0 < λ2

Fig. 3.14 Seven different cases for classifying qualitative behavior of a 2-dim. linear D.E. system

Remark. The solution x = et A k can be interpreted as a flow:

φt (x) = et A x

φ t : Rn → R n , φt = et A (3.26)

where φt satisfies φt+s = et A+s A = et A es A = φt φs . This is an example of an actual flow.

It should be pointed out that the above classification of seven cases may be given in
different versions. For example (cf. Hirsch & Smale), one can use a four case classification:
(1) A has real eigenvalues of opposite signs (Saddle); (2) All eigenvalues of A have negative
real parts (Sink); (3) All eigenvalues of A have positive real parts (Source); and (4) The
eigenvalues are pure imaginary (Center).
AM 9524 Chapter 3 51

Now let’s consider an example to see how to obtain those trajectories shown in Figs 3.6
– 3.13, given in the original coordinates x . Choose the saddle case, for simplicity, which has
non-zero real eigenvalues of opposite signs. From the transformation
  x = P y , the solution
λ1 0
x = et A k and the relation A = P B P −1 where B = , we can obtain
0 λ2
 λ t 
e 1 0
x = P P −1 k
0 eλ2 t
 λ t 
e 1 0
= (v 1 v 2 ) P −1 k
0 eλ2 t
= (v 1 eλ1 t v 2 eλ2 t ) C

= C1 eλ1 t v 1 + C2 eλ2 t v 2 (3.27)

 
C1
where C = P k = −1
is an arbitrary vector. Thus, solutions with C2 = 0 , are
  C 2

λ1 t α
C1 v 1 e = C1 eλ1 t , or
β

x1 = C1 α eλ1 t and x2 = C1 β eλ1 t . (3.28)

x2 β
Thus, = , or x2 = (const.) x1 which represents a straight line through the origin.
x1 α
Similarly, for C1 = 0 , we have another straight line. Note that the slopes of these two
lines are actually the eigenvectors corresponding to the eigenvalues λ1 and λ2 , respectively.
Since λ1 < 0 and λ2 > 0 for the saddle case, x1 and x2 shrink along the line C2 = 0 while
they grow along the line C1 = 0 (see Fig. 3.15(a)). Then the adjacent trajectories can be
easily obtained as in Fig. 3.15(b).

3.3 Fundamental Theory

A “state” of physical system is information about it, e.g. (x, v) is the phase space of
Physics. Let S be the state (phase) space of a system. Then the dynamics is to describe
the passage in time of all points of the given space S. Mathematically, we may think of S
as a Euclidean space Rn or an open subset of Euclidean space.

Definition 3.1. A Dynamical System is a C 1 map

R×S → S (3.29)
AM 9524 Chapter 3 52

x2 x2
C 2= 0 C 2= 0

x1 x1

C1 = 0 C 1= 0

(a) (b)
Fig. 3.15 Trajectories of saddle case: (a) two particular trajectories; (b) phase portrait.

where S is an open set of Rn . R×S can be think of an enlarged phase space. (Notation C 1
means continuously differentiable.) It is assumed that the map is continuously differentiable
or at least continuous and continuously differentiable in t.
Writing φ(t, x) = φt (x) , then the map R × S → S is defined by (t, x) → xt , i.e.
the map
φt : S → S (3.30)

takes x into xt ∀ t ∈ R . This map is called flow and it satisfies

(i) φ0 = id. (i.e. φ0 : S → S is the identity.)
(ii) The composition φt ◦ φs = φt+s for each t, s ∈ R.
Note that the definition implies that the map φt : S → S is C 1 for each t and has a C 1
inverse φ− t ( take s = − t in (ii)).

Example 3.2.  
0 1
ẋ = A x A= (3.31)
−1 0
where x = (x1 x2 ) and equation (3.31) can be written as

ẋ1 = x2
⇐⇒ ẍ1 + x1 = 0 (3.32)
ẋ2 = − x1
AM 9524 Chapter 3 53

which yields solution x = ( r cos (t − t0 ) , − r sin (t − t0 ) ) . The graph of the solution is

shown in Fig. 3. 16.

t x2

x1
x2

x1
(a) (b)

Fig. 3.16 The graph of D.E. (3.31): (a) a solution curve; (b) phase portrait.

A D.S. φt on S in general gives rise to a D.E. on S , that is, a vector field (V.F.) on
S , f : S → Rn , where S ⊂ Rn . Given φt , define f by
d
f (x) = φt (x)|t = 0 ; (3.33)
dt
thus ∀ x ∈ S , f (x) is a vector in Rn which we think of as the tangent vector to the curve

t → φt (x)

at t = 0. Thus every D.S. gives rise to a D.E.

Equation (3.33) can be written as a more conventional form: if φt : S → S is a D.S.
and x ∈ S , let x(t) = φt (x) , and f : S → Rn be defined as in (3.33). Then on may write
(3.33) as
ẋ = f (x) . (3.34)

Therefore x(t) and φt (x) is the solution curve of (3.34) satisfying the initial condition
x(0) = x . There is converse process: given a D.E. one has associated to it, an object which
would be a flow (D.S.) if it were defined for all t. This is the subject of the fundamental
theory.
AM 9524 Chapter 3 54

Remark. The equation (3.34) is called an autonomous equation, which means that the
function f does not depend on time explicitly. One can also consider a C 1 map f :
I × S → Rn where I is a real interval. The equation in that case is

ẋ = f (t, x) (3.35)

and is called non-autonomous. This course deals with both the autonomous and non-
autonomous systems, but the emphasis is on the autonomous case.

Example 3.3. Consider the Van der Pol’s equation:

ẍ + α φ(x) ẋ + x = β p(t) (3.36)

where α and β are constants. This equation can be rewritten as

   
ẋ1 x2
ẋ = = ≡ f (t, x) (3.37)
ẋ2 − x1 − α φ(x1 ) x2 + β p(t)
by introducing x1 = x and x2 = ẋ. This equation is non-autonomous if β 6= 0 and it is
autonomous when β = 0.
In general, the solution of the D.E. (3.35) can be defined as a differentiable function

φ : I → Rn (3.38)
satisfying
(i) (t, φ(t)) ⊂ R × Rn ∀ t ∈ I;
dφ
(ii) = f (t, φ(t)) ∀ t ∈ I.
dt
Note that I can be the whole real line R as in the solutions of the D.E. (3.1).

IVP (Initial Value Problem): D.E. (3.35) plus initial condition

=⇒
ẋ = f (t, x) x(t0 ) = x0 where (t, x0 ) ⊂ I × Rn (3.39)

and f is defined as C 1 map f ; I × S → Rn .

A solution of (3.39) is φ satisfying (3.39) and also φ(t0 ) = x0 . In order to show dependence
of φ on x0 , we write
φ(t, x0 ) ≡ φt (x0 ) = φt (x0 ) (3.40)
AM 9524 Chapter 3 55

for a solution of (3.39), which is called “t-advance mapping” value of solutions at time t
through initial vector x0 . It is usually to fix t0 = 0 , but, more generally t0 can vary too.
It should be noted that, theoretically, a non-autonomous system can always be trans-
ferred to an autonomous system by introducing an additional state variable xn+1 = t , and
therefore, ẋn+1 = 1. For Van der Pol’s equation (3.36) can be written as
   
ẋ1 x2
ẋ =  ẋ2  =  − x1 − α φ(x1 ) x2 + β p(x3 )  ≡ f (x) (3.41)
ẋ3 1
where x3 = t. Dynamical systems are mainly autonomous.
So far when we solve a D.E. we always assume or imply that solutions exist (even if we
can’t “get” them). However, the following simple example
1 2
ẋ2 + x = −2 (3.42)
2
does not have any real solutions. So if for a general D.E. our solution techniques (e.g.
perturbation approaches, direct integration methods, etc.) fail, how do we know whether
there is actually anything else to look for? The following fundamental questions:
(1) When does the solution φ(t, x0 ) exist?
(2) Is the solution unique?
(3) Is φ(t, x0 ) continuous on the initial condition x0 )?
are raised and they lead us to the most important results of the fundamental theory:

Theorem 3.2 (Existence and Uniqueness). Let U ⊂ Rn be an open subset of Euclidean

space; and let f : U → Rn be a C 1 map, and x0 ∈ U . Then there exists some constant
a > 0 and a unique solution
x : (− a, a) → U (3.43)
of the D.E.
ẋ = f (x) (3.44)
satisfying the initial condition
x(0) = x0 . (3.45)

Note: In fact all we need for f is that the function f (x) satisfies the Lipschitz condition:

k f (x) − f (y) k ≤ m k x − y k ∀ x, y ∈ Rn (3.46)

AM 9524 Chapter 3 56

where m ≥ 0 is called a Lipschitz constant for f , and f is then a Lipschitz function. Here,
the notation k • k denotes a norm. The Lipschitz condition turns out to be a less restrictive
requirement than the existence of a derivative.

3.3.1. Existence Theorem. Let U ⊂ Rn consisting of all x such that

kx− ck ≤ a > 0 (3.47)

where c is a constant vector defining a point in U . Let f (x) be a continuous vector function
in U satisfying the D.E. (3.44) for 0 ≤ t ≤ b. We seek a solution such that x(0) = c for
0 ≤ t ≤ b. It should be noted that this “forward” problem also characterizes the problem in
negative t by just changing the sign of f .
In general, by integrating the D.E. (3.44) we have
Z t
x(t) = c + f (x(τ )) dτ (3.48)
0

on 0 ≤ t ≤ b. It is helpful to think of the RHS of the equation as a transform, T {x(t)} that

produces a new function on the LHS of the equation. The solution of the system of D.E.’s
is a fixed point of this transform. Thus, for any x(t) , let
Z t
T {x(t)} = c + f (x(τ )) dτ = y(t) , (3.49)
0

then x(t) is a solution if y(t) = x(t).

Now, since f (x) is continuous in U , it is bounded in U , thus

k f (x) k ≤ K ∀ x ∈ U. (3.50)
Therefore,

k y − c k = k T {x} − c k
Z t
= f (x(τ )) dτ
0
Z t
≤ k f (x(τ )) k dτ
0
≤ Kt (3.51)

for any x, y ∈ U .
AM 9524 Chapter 3 57

If we choose b (recall that 0 ≤ t ≤ b) such that

a
b= , (3.52)
K

then, the above inequality (3.51) implies that k y − c k ≤ a in addition to the requirement
k x − c k ≤ a to define U . That is, both input and output functions are members of U at
a a
least for 0 ≤ t ≤ . Consequently, U is mapped onto itself by T for 0 ≤ t ≤ (see Fig.
K K
3.17).

a
x slope K = -
b
c+a

c x(t)

c-a

0 b t

Fig. 3.17 Geometric explanation of the existence theorem.

Fixed points of T clearly have this property that they are confined to U in that
interval of t .
If T has no fixed points in U , there is no solution;
If T has one fixed point in U , there is one solution;
etc.
If we choose some function x(0) (t) with appropriate behavior in U for t ∈ [0, b] then T
induces a sequence
x(0) (t) , x(1) (t) , . . . , x(n) (t) , . . .
which may converge to a solution (fixed point) of the original system. The following the-
orem guarantees the convergence of such a sequence. This method is called successive
approximation.
AM 9524 Chapter 3 58

Theorem 3.3. Let x(0) (t) be a continuous function in U for 0 ≤ t ≤ b. For k = 0, 1, . . .

define Z t
(k+1) (k)
x (t) = T {x (t)} = c + f (x(k) (τ )) dτ . (3.53)
0

If f (x) is a Lipschitz function, then the sequence converges to a solution of the initial
value problem.

Proof: First note that

Z t
(1) (0) (0)
x −x =c−x + f (x(0) (τ )) dτ . (3.54)
0

Thus, we have
Z t
(1) (0) (0)
x −x ≤ c−x + f (x(0) (τ )) dτ
0
Z t
≤ c − x(0) + f (x(0) (τ )) dτ
0
≤ a+Kt

≤ 2a (3.55)

a a
where K t ≤ a due to b = , (i.e. K ≡ ) and 0 ≤ t ≤ b .
K b
Next, for k = 1 ,
Z t Z t
(2) (1) (1)
x −x = f (x (τ )) dτ − f (x(0) (τ )) dτ
0 0
Z t
f (x(1) (τ )) − f (x(0) (τ )) dτ
 
=
0
Z t
≤ f (x(1) (τ )) − f (x(0) (τ )) dτ
0
Z t
≤ m x(1) (τ ) − x(0) (τ ) dτ
0
≤ 2 a (m t) (3.56)

on 0 ≤ t ≤ b . Finally, by induction, it follows that

(m t)k (m b)k
x(k+1) (τ ) − x(k) (τ ) ≤ 2a ≤ 2a → 0 as k → ∞. (3.57)
k! k!

This shows that the sequence

AM 9524 Chapter 3 59

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

uniformly (independence of time t) converges and the limit exists, hence, a solution exists.
Thus, we have established under very general conditions that a solution must exist for the
D.E. (3.44) even if we can’t compute one.

3.3.2. Uniqueness Theorem. It is very important that only one solution passes
through any particular point in phase space (uniqueness). This turns out to be a result of
the Lipschitz condition. Uniqueness ultimately restricts very strongly what is possible and
impossible in terms of solutions.

Proof of Uniqueness: Suppose, we have a solution x(t) (from the part of existence the-
orem), generated from the sequence x(0) (t) , x(1) (t) , . . . , and assume that y(t) is another
solution on the same interval 0 ≤ t ≤ b. We want to prove that y(t) = x(t).
From
x(0) (t) = c
Z t
(1)
x (t) = c + f (x(0) (τ )) dτ
0
..
.
Z t
(n)
x (t) = c + f (x(n−1) (τ )) dτ
Z0 t
y(t) = c + f (y(τ )) dτ , (3.58)
0

we have
Z t
(0)
x (t) − y(t) = f (y(τ )) dτ ≤ K t ≤ a,
0
Z t
(1)
f (x(0) (τ )) − f (y(τ )) dτ
 
x (t) − y(t) =
0
Z t
≤ f (x(0) (τ )) − f (y(τ )) dτ
Z0 t
≤ m x(0) (τ ) − y(τ ) dτ
0
≤ a (m t) . (3.59)

Again, by induction, one may obtain

(m t)k (m b)k
x(k) (t) − y(t) ≤a ≤a → 0 as k → ∞. (3.60)
k! k!
AM 9524 Chapter 3 60

This implies that the sequence uniformly converges to y(t). But since x(t) is generated
from this same sequence (i.e. the sequence also uniformly converges to x(t) ), therefore,
y(t) = x(t). This proves that the solution is unique.
Let us see how the “iteration scheme” used in the proof in this section can be applied
to the following very simple example:

Example 3.4.
ẋ = x. (3.61)

We know that the solution, , x(t) , of the D.E. satisfying x(0) = x0 is given by x(t) = x0 et ).
Set x(0) (t) = x0 , then by using formula (3.58) one obtains

x(0) (t) = x0
Z t Z t
(1) (0)
x (t) = x0 + x (τ ) dτ = x0 + x0 dτ = x0 + x0 t = x0 (1 + t)
0 0
t t
t2
Z Z
(2) (1)
x (t) = x0 + x (τ ) dτ = x0 + x0 (1 + τ ) dτ = x0 (1 + t + )
0 0 2!
..
.
t k
t2 tk ti
Z X
(k) (k−1)
x (t) = x0 + x (τ ) dτ = x0 (1 + t + + · · · + ) = x0 . (3.62)
0 2! k! i=0
i!

As k → ∞, x(k) (t) converges to

∞ i
X t
x0 = x0 et = x(t), (3.63)
i=0
i!

which is, of course, the solution of the original D.E. ẋ = x.

Note: The Lipschitz condition definition (3.46), and the existence and uniqueness theory
given above are applicable for autonomous systems only. However, they can be easily ex-
tended to non-autonomous systems. For example, the equations can be modified as

k f (t, x) − f (t, y) k ≤ m k x − y k ∀ (t, x), (t, y) ∈ R × Rn . (3.64)

for non-autonomous systems. The proof of existence and uniqueness theorem for non-
autonomous systems can be found from, for example, Hirsch and Small (1974, Chapter
15).
AM 9524 Chapter 3 61

Also note that the theorem is local. Furthermore, in order for the theorem to be
physically relevant, it needs to be complemented with the property that the solution x(t)
depends continuously on the initial condition.

Remarks and Examples.

R1) An immediate implication is that cases such as shown in Fig. 3.18 are excluded by
the theorem.
φ ψ

(a) (b)
Fig. 3.18 Non-unique solutions: (a) two solutions cross; (b) a solution crosses itself.

Moreover, if a solution curve φ satisfies φ(t1 + w) = φ(t1 ) , then the curve must close
up as shown in Fig. 3.19.

φ (t 1 + w) = φ (t 1 )

Fig. 3.19 A closed solution curve.

R2) Consider the IVP (Initial Value Problem)

ẋ = 1 + x2 , x(0) = 0 (3.65)

which may have solutions

AM 9524 Chapter 3 62

π π
x = tan t , − ≤t≤
2 2
1 1
x = tan t , − ≤t≤ (3.66)
2 2

and, strictly speaking, they should be considered different. But this is artificial and inconve-
nient. To avoid this, we introduce the concept of maximal solutions, i.e. roughly speaking,
solutions which hold over the largest possible interval. For the above example, the general
solution is given by the function

x = tan(t − c) , c = constant. (3.67)

Such a function cannot be extended over an interval larger than

π π
c− <t<c+ (3.68)
2 2
π
since x(t) → ± ∞ as c ± .
2
R3) Note that globality does not require the solution to be defined for all t. Consider
the following example

ẋ = x2 in R with x(0) ≡ x0 > 0 (3.69)

1 1
The solution of this D.E. is x = , defined for − ∞ < t < .
1/x0 − t x0

R4), For the linear system ẋ = A x , the solution

x(t, x0 ) = et A x0 (3.70)

is global. (This is NOT true for nonlinear systems.)

Certain solutions play a special role, i.e. those which lie in the linear subspaces spanned
by the eigenvectors of A. They are invariant under et A . For example, given an eigen-
vector, v j , of A associated with eigenvalue λ1 , then x(t, C v j ) = C v j eλ t .
It’s traditional to divide these subspaces into three classes:
(i) The stable subspace, E s = span{v 1 v 2 . . . v ns } , where v j are eigenvectors of A
with negative real parts, i.e. Re(v j ) < 0.
(ii) The unstable subspace, E u = span{u1 u2 . . . unu } with Re(uj ) > 0.
AM 9524 Chapter 3 63

(iii) The center subspace, E c = span{w1 w2 . . . wnc } with Re(w j ) = 0.

Note on nomenclature: if the eigenvalues of A have:
(a) Re < 0 , then et A is called a contraction;
(b) Re > 0 , then et A is called an expansion.
If Re 6= 0 , then et A is called hyperbolic.

To end this chapter, we use the following example to demonstrate how to discuss the
uniqueness of solutions of a D.E.

Example 3.5. Consider the D.E

dx √
= −K x , x≥0 (K > 0 is a constant) (3.71)
dt
a) Find a value of x for which the D.E. does not have unique solutions. What are these
solutions?

b) Why does this not contradict the uniqueness theorem?

c) The D.E. represents the rate at which the depth of liquid in a bucket decreases due to
a small leak. Explain how non-uniqueness pertains to this physical problem.

Solution:
(a) Consider the functions
1 2
x(t) ≡ 0 and x(t) = K (t + C)2 (t ≤ − C) (3.72)
4
where C is an arbitrary constant. When x = 0 , for any given initial point (t0 , 0) in
the (t , x) plane, one can choose C = − t0 so that there are two solutions given by (3.72)
satisfying the D.E. and passing through the same point (t0 , 0) . Thus, the solution x = 0
is not unique. The solutions of the D.E. (3.71) can be solved as follows:
dx √
= −K x
dt
=⇒ x−1/2 dx = − K dt

=⇒ 2 x1/2 = − K (t + C) (t ≤ − C since x ≥ 0 and K > 0)

1
=⇒ x(t) = K 2 (t + C)2 (t ≤ − C). (3.73)
4
AM 9524 Chapter 3 64

The solutions given by (3.72) are shown in Fig. 3.20. Note that only half of the parabola
1
are the family of the solutions x(t) = K 2 (t + C)2 (t ≤ − C). It can be proved that
4
the solution x(t) ≡ 0 is a singular solution, which is the envelope of the family of the
1
solutions K 2 (t + C)2 .
4
x

t
0

Fig. 3.20 Solutions of D.E (3.72).

√
(b) Because it does not satisfy the Lipschitz condition at x = 0. Here, f (x) = − K x,
so, for arbitrary x1 and x2 in the neighborhood of x = 0 , we have
k f (x1 ) − f (x2 ) k = | f (x1 ) − f (x2 )|
√ √
= |− K x1 + K x2 |
√ √
= |K| | x1 − x2 |
√ √
x1 − x2
= |K| |x1 − x2 |
x1 − x2
1
= |K| √ √ |x1 − x2 | . (3.74)
x1 + x2
Want to the Lipschitz condition to be satisfied, there must exist a Lipschitz constant (finite)
m ≥ 0 such that
| f (x1 ) − f (x2 )| ≤ m | x1 − x2 | ∀ x1 , x2 ≥ 0. (3.75)
|K|
It is clear that such an m does not exist because √ √ → ∞ when x1 , x2 → 0.
x1 + x2
(b) One cannot determine when a formerly full, empty bucket began to leak. Empty
bucket corresponds to the solution x = 0. Since this solution is not unique, so one cannot
determine which branch represents the solution corresponding to the non-empty bucket.