Lecture 15

Linear systems and the Fundamental Matrix (cont’d)

Theorem 4 (Dimension of the solution space S): Let

x′ = A(t)x (1)

be a linear homogeneous DE in Rn , where the n × n matrix A(t) is continuous on an interval

I. Then the solution space S of this system, that is, the linear vector space of all solutions to
(1), is an n-dimensional subspace of the vector space of functions C 1 (I, Rn ), i.e., the space of
all continuously differentiable functions which map the interval t ∈ I to Rn .

Proof: Let ek , k = 1, 2, · · · n, denote the “standard basis” in Rn :

     
1 0 0
     
 0   1  0
     
 
e1 =   , e2 =   , · · · en = 
 ..   ..  ..
. (2)
 .   .  .
 
 
     
0 0 1

Let t0 ∈ I. From Theorem 1 (Existence-Uniqueness for linear homogeneous ODEs in Rn , there

exists a set of n solutions, x1 (t), ..., xn (t) to the linear system in (1) that satisfy the following
initial conditions:
x1 (t0 ) = e1 , x2 (t0 ) = e2 , ··· xn (t0 ) = en . (3)

From Theorem 3 proved earlier, this set of solutions is linearly independent for each t ∈ I and
hence forms a linearly independent set in the solution space S.
We now prove that an arbitrary solution x(t) of (1) is a linear combination of these solutions
xk (t). Given such an arbitrary solution x(t), define

x0 = x(t0 ) . (4)

Since x0 ∈ Rn , there exist constants c1 , c2 , · · · , cn , not all zero, such that

x0 = c1 e1 + c2 e2 + · · · cn en . (5)

142
Now use these constants to define a new function, y : I → Rn ,

y(t) = c1 x1 (t) + c2 x2 (t) + · · · xn (t) . (6)

By the Principle of Superposition, y(t) is a solution of the DE in (1) which, by construction –

see (3) and (5) – satisfies
y(t0 ) = x0 . (7)

From Eq. (4), the solutions x(t) and y(t) satisfy the same initial condition. Therefore, from
the uniqueness part of Theorem 1 (Existence-Uniqueness for linear homogeneous DEs),

x(t) = y(t) for all t ∈ I . (8)

From (6), it follows that our arbitrary solution x(t) may be written as

x(t) = c1 x1 (t) + c2 x2 (t) + · · · + cn xn (t) , (9)

i.e., a linear combination of the solutions xk (t). Therefore the set of solutions,

{x1 (t), x2 (t), · · · , xn (t)} , (10)

is a basis for the solution space S, implying that the dimension of S is n. The proof is complete.

143
Important remarks: This result finally proves a statement made earlier in this course that
the dimension of the solution space for linear second order homogeneous DEs of the form,

y ′′ + P (t)y ′ + Q(t)y = 0 , (11)

is two. (Note that t is used as the independent variable here, i.e., y = y(t).) From our earlier
discussion, the above DE may be written as a linear homogeneous DE in R2 , where

x1 (t) = y , x2 (t) = y ′ . (12)

Given a solution y(t) to (11), and the fact that x1 (t) = y, along with the fact that x′2 (t) = x1 (t),
one might wonder why the dimension of the solution y(t) isn’t 1, instead of 2. After all, if we
know x1 (t), we know y(t). The solution x2 (t), which is, in fact, x′1 (t) = y ′ (t) seems to be
redundant.
The fact, however, is that for general linear systems in x1 (t) and x2 (t), one cannot, in
general, solve for x1 (t) independently without solving for x2 (t) as well. Furthermore, in order
to isolate a particular solution of the second order DE in (11), one must prescribe the initial
conditions, y(t0 ) and y ′(t0 ) which become the initial conditions x1 (t0 ) and x2 (t0 ). The require-
ment to know two initial pieces of information is a consequence of the two-dimensionality of
the solution space S of solutions to (11).

Also recall that an nth order homogeneous linear differential equation in the function y(t)
having the form,
an (t)y (n) + an−1 (t)y (n−1) + · · · + a1 (t)y ′ + a0 (t) = 0 , (13)

can be transformed into a linear system of n first order DEs in the functions,

x1 (t) = y(t) , x2 (t) = y ′(t) , · · · , xn−1 = y (n−1) (t) . (14)

The solution x(t) is a mapping from an interval I ⊆ R to Rn . (Details in an earlier lecture.)

We have just proved that the solution space S of this linear first order system in Rn is n. We
can therefore conclude that the dimension of the solution space of the linear nth order DE in
(13) is n.

144
Special case: Constant matrix A

Recall from AMATH 250/251 that the solutions to linear systems with constant coefficients,
i.e.,
x′ = Ax , (15)

where A is a constant n × n matrix, can be constructed from a knowledge of the eigenvalues

and eigenvectors of A. If we assume a solution to (15) of the form

x(t) = veλt , (16)

where v ∈ Rn and λ ∈ R are to be determined, then substitution into (15) yields,

λveλt = Aveλt . (17)

Dividing both sides by eλt 6= 0 and rearrangement yields

Av = λv . (18)

In other words, λ is an eigenvalue of A and v is its associated λ-eigenvector (provided it exists).

Example: The linear system x′ = Ax, where the matrix A is given by

 
1 1
A= . (19)
4 1
The eigenvalues of A are λ1 = 3 and λ2 = −1. The associated eigenvectors are
   
1 1
v1 =   , v2 =  . (20)
2 −2
In this case, two linearly independent solutions of this system are
   
1 1
x1 (t) = v1 eλ1 t =   e3t , x2 (t) = v2 eλ2 t =   e−t . (21)
2 −2
The general solution of this system is therefore
 
3t −t
c1 e + c2 e
x(t) = c1 x1 (t) + c2 x2 (t) =  . (22)
3t −t
2c1 e − 2c2 e

145
The Fundamental Matrix associated with x′ = Ax

Let’s now return to the following idea used in the proof of Theorem 4 (dimensionality of the
solution space S for a linear system in Rn ). Let ek , k = 1, 2, · · · n, denote the “standard basis”
in Rn :      
1 0 0
     
0 1  0 
     
   
e1 =  ..
, e2 =  ..
 , · · · en =   .
 ..  (23)
. .  . 
   
   
     
0 0 1
Suppose that we have n solutions X1 (t), ..., Xn (t) to the linear system,

x′ = A(t)x , (24)

that satisfy the following initial conditions:

X1 (t0 ) = e1 , X2 (t0 ) = e2 , ··· Xn (t0 ) = en . (25)

Recall that these solutions are represented by 1 × n column vectors. Now put these solu-
tions/column vectors together to form the following n × n matrix,

Φ(t, t0 ) = [X1 (t) X2 (t) · · · Xn (t)]. (26)

Then this matrix, known as the fundamental matrix of the linear system in (24) at t0 ,
satisfies the following properties:

1. Identity property:

Φ(t0 , t0 ) = [X1 (t0 ) X2 (t0 ) · · · Xn (t0 )] (27)

= [e1 e2 · · · en ]

= I,

where I denotes the n × n identity matrix.

146
2. Initial value property: Let the matrix Φ(t, t0 ) operate on a constant vector a ∈ Rn as
follows:

x(t) = Φ(t, t0 )a (28)

= [X1 (t) X2 (t) · · · Xn (t)]a.

Then by the rules of matrix multiplication,

x(t) = a1 X1 (t) + a2 X2 (t) + · · · + an Xn (t), (29)

which implies that x(t) is also a solution to (24). But note that

x(t0 ) = Φ(t0 , t0 )a (30)

= Ia

= a.

In other words, the solution x(t) = Φ(t, t0 )a is the unique solution of the initial value
problem x(t0 ) = a.

We come to an important point:

The fundamental matrix Φ(t, t0 ) is a mapping of the initial condition a to the

solution vector x(t). As t varies, the point x(t) traces out a curve in Rn .

In the parlance of “Dynamical Systems Theory,” the mapping Φ(t, t0 ) is an example of

a flow operator: It defines the evolution of points – here in Rn – as determined by
solutions of the “dynamical system,”

x′ = A(t)x . (31)

147
 3. Derivative property: The n×n matrix Φ(t, t0 ) is a solution of the matrix differential
equation
d
Φ(t, t0 ) = A(t)Φ(t, t0 ). (32)
dt
(It is understood that the term on LHS is the n × n matrix of the time derivatives of the
elements of Φ(t, t0 ).)

This result is a consequence of matrix multiplication. Recall that the kth column of
Φ(t, t0 ) is the solution vector Xk (t). The time derivative of this solution vector is the
vector A(t)Xk (t). Placing each vector in the kth column on each side preserves these
relations.

Another, more formal, proof is as follows. Recall that

x(t) = Φ(t, t0 )a (33)

is the solution of the linear system IVP

x′ (t) = A(t)x , x(t0 ) = a . (34)

Now substitute the left and right hand sides of (33) into the respective sides of (34):

Φ(t, t0 )′ a = A(t)Φ(t, t0 )a. (35)

Rearranging, we have
[Φ(t, t0 )′ − A(t)Φ(t, t0 )]a = 0. (36)

Since this equation is true for all a ∈ Rn , it follows that the matrix expression in the
square brackets is zero, which proves (32).

Example: Let us return to the earlier example, namely, the linear system x′ = Ax, with
matrix A given by  
1 1
A= . (37)
4 1

148
Recall that the general solution of this system was found to be
 
3t −t
c1 e + c2 e
x(t) = c1 x1 (t) + c2 x2 (t) =  . (38)
3t −t
2c1 e − 2c2 e

We now determine the particular solutions X1 (t) and X2 (t) from which the fundamental matrix
Φ(t, 0) will be constructed:
 
1
(i) Find X1 (t) such that X1 (0) =  . This leads to the following linear system of equations
0
in c1 and c2 :

c1 + c2 = 1

2c1 − 2c2 = 0

with solution c1 = c2 = 1/2. Thus

 
1 3t 1 −t
2
e + 2
e
X1 (t) =  .
e3t − e−t

 
0
(ii) Find X2 (t) such that X2 (0) =  . This leads to the following linear system of equations
1
in c1 and c2 :

c1 + c2 = 0

2c1 − 2c2 = 1

with solution c1 = 1/4 and c2 = −1/4. Thus

 
1 3t
4
e − 41 e−t
X2 (t) =  .
1 3t 1 −t
2
e + 2
e

Therefore the fundamental matrix associated with this linear system is given by
 
1 3t 1 −t 1 3t 1 −t
e + 2e e − 4e
Φ(t, 0) =  2 4 . (39)
1 3t 1 −t
e3t − e−t 2
e + 2
e

149
You will note that Φ(0, 0) = I.

The particular solution of this linear system satisfying the initial condition,
 
1
x0 = a =   , (40)
1

is

x(t) = Φ(t, 0)a

  
1 3t 1 −t 1 3t 1 −t
e + e e − e 1
=  2 2 4 4  
1 3t
e3t − e−t 2
e + 21 e−t 1
 
3 3t
4
e + 41 e−t
=  . (41)
3 3t 1 −t
2
e − 2e

We used this result in the previous lecture to illustrate the vector Picard iteration method.

4. Multiplication property: For any t0 , t1 , t ∈ I, we have

Φ(t, t0 ) = Φ(t, t1 )Φ(t1 , t0 ). (42)

To prove this result, let a ∈ Rn . For a t0 ∈ I,

x(t) = Φ(t, t0 )a. (43)

Now, for a t1 ∈ I, define

b = x(t1 ) = Φ(t1 , t0 )a (44)

and define
y(t) = Φ(t, t1 )b. (45)

This means that y(t1 ) = b = x(t1 ). By Existence-Uniqueness, it follows that

y(t) = x(t). (46)

Therefore,
Φ(t, t1 )b = Φ(t, t0 )a, (47)

150
or
Φ(t, t1 )Φ(t1 , t0 )a = Φ(t, t0 )a. (48)

Since this result follows for any a ∈ Rn , the property in (42) follows.

5. Inverse property: If we choose t = t0 in (42) above, then

Φ(t0 , t0 ) = Φ(t0 , t1 )Φ(t1 , t0 ), (49)

But φ(t0 , t0 ) = I, the identity matrix, which implies that

Φ(t0 , t1 ) = [Φ(t1 , t0 )]−1 (inverse). (50)

If we replace t1 with t ∈ I, then

Φ(t0 , t) = [Φ(t, t0 )]−1 (inverse). (51)

151
Special case: “Linear time-invariant systems”

In the special case that the matrix A in the linear system is constant, the solutions of the DE
along with its fundamental matrix possess some additional properties. For example:

If x(t) is a solution of the linear system x′ = Ax then x(t − a) is also a solution

where a ∈ R is a constant.

To illustrate, let us return to the linear system studied in the previous lecture and defined by
the constant matrix  
1 1
A= . (52)
4 1
The general solution of this system was found to be
   
1 1
x(t) = c1   e3t + c2   e−t . (53)
2 −2
Now examine the function y(t) = x(t − a):
   
1 1
x(t − a) = c1   e3(t−a) + c2   e−(t−a) (54)
2 −2
   
1 1
= c1 e−3a   e3t + ea   e−t
2 −2
   
1 1
= d1   e3t + d2   e−t ,
2 −2
which is also a solution. We can prove this result more formally as follows.
Let y(t) = x(t − a). Then
d d
y(t) = x(t − a) (55)
dt dt
d ds
= x(s) , where s = t − a,
ds dt
′
= x (s)

= Ax(s)

= Ay(t),

and the result is proved.

152
Lecture 16

Linear time-invariant systems (cont’d)

At the end of the previous lecture, we proved the following important property of linear systems
defined by constant matrices A:

If x(t) is a solution of the linear system x′ = Ax then x(t − a) is also a solution

where a ∈ R is a constant.

The above property extends to the fundamental matrix associated with the linear system
x′ = Ax, where A is a constant matrix, as we now show.

Let x(t) be a solution to this system. For a t0 ∈ R, define

y(t) = x(t − t0 ), (56)

which implies that

y(t0 ) = x(0) . (57)

From the previous result, y(t) is also a solution. By the definition of the fundamental matrix:

y(t) = Φ(t, t0 )y(t0 ) (58)

and
x(t) = Φ(t, 0)x(0). (59)

Now replace t by t − t0 in (59):

x(t − t0 ) = Φ(t − t0 , 0)x(0). (60)

But from Eq. (58), since y(t) = x(t − t0 ),

x(t − t0 ) = Φ(t, t0 )y(t0 )

= Φ(t, t0 )x(0) . (61)

153
Comparing (60) and (61), we have the result

Φ(t, t0 ) = Φ(t − t0 , 0). (62)

In other words:

Φ(tf inish , tstart ) = Φ(tf inish − tstart , 0).

The most important quantity in a linear time-invariant system is ∆ = tf inish − tstart ,

the time interval that has elapsed during the flow, not the start and finish times.

Therefore, for a linear time-invariant system, without loss of generality, we may define

Φ(t) := Φ(t, 0), (63)

where t denotes the time elapsed from t0 = 0. This fundamental matrix satisfies some important
properties:

1. Φ(0) = Φ(0, 0) = I, the identity matrix.

2. x(t) = Φ(t)a is the unique solution to the IVP x′ = Ax, x(0) = a.

3. Φ(t) satisfies the matrix differential equation Φ′ (t) = AΦ(t).

4. Φ(t1 + t2 ) = Φ(t1 )Φ(t2 ).

5. Φ(−t) = Φ(t)−1 (inverse).

Let us return to the linear system studied earlier, with fundamental matrix given by
 
1 3t 1 −t 1 3t 1 −t
e + e e − e
Φ(t, 0) =  2 2 4 4 . (64)
3t −t 1 3t 1 −t
e −e 2
e + 2e

We leave it as an exercise for the reader to show that

1. Φ(−t)Φ(t) = I and

2. the matrix Φ(t) satisfies the matrix differential equation, Φ′ (t) = AΦ(t).

154
The exponential matrix

We now examine more closely the fundamental matrix Φ(t, t0 ) associated with a linear time-
invariant system, i.e.,
x′ = Ax, x(0) = a, (65)

where A is a constant n × n matrix.

In the one-dimensional case, the matrix A is simply a real constant a, and the linear system
degenerates to the single DE,
x′ = ax, x(0) = x0 , (66)

with solution
x(t) = x0 eat . (67)

If we replace the constant “a” with the matrix A, then one might conjecture that the solution
to the linear system in (65) is given by

x(t) = etA a. (68)

Of course, it remains to determine what “etA ” means.

Given a matrix A, we shall define the matrix exponential eA in terms of the Taylor series
of the exponential. Recall that for all x ∈ R,

1 2 1
ex = 1 + x + x + · · · xn + · · · . (69)
2! n!

In the same way, we define, for an n × n constant matrix A,

1 2 1
eA = I + A + A + · · · + An + · · · . (70)
2! n!

It still remains to determine if the above infinte series makes sense. For the moment, we’ll
simply continue, with the assumption that everything above is fine, i.e., that the above infinite
series converges. The technical points will be addressed at the end of the lecture.

155
Some properties of matrix exponentials

Some of the properties listed below follow immediately from the series definition. The others
require a little more work, which we omit here.

1. e0 = I

2. eA+B 6= eA eB unless A and A commute with respect to matrix multiplication, i.e.,

AB = BA.

3. e−A = [eA ]−1 (inverse).

Some examples:

We examine a few examples in R2 .

1. If A is a diagonal matrix, i.e.  

a 0
A= , (71)
0 b
then
   
2
a 0 a /2! 0
eA = I +  +  +··· (72)
0 b 0 b2 /2!
 
2
1 + a + a /2! + · · · 0
=  
2
0 1 + b + b /2! + · · ·
 
ea 0
=  .
b
0 e

2. The upper triangular matrix  

a b
A= . (73)
0 a
(Note: In class, we considered the special case b = 1. In these notes, we have generalized
the result.) We shall write A as
A = aI + bN, (74)

156
 where  
0 1
N= . (75)
0 0
The matrix N is nilpotent since N2 = N3 = · · · = 0. This implies that the series for the
exponential of N terminates quite rapidly, i.e.,

eN = I + N . (76)

Then

eA = eaI+bN (77)

= eaI ebN since IN = NI

= ea I [I + bN]
 
1 b
= ea  .
0 1

3. The antisymmetric matrix  

a b
A= , (78)
−b a
which we shall write as
A = aI + B, (79)

where  
0 b
B= , (80)
−b 0
Since I and B commute,
eA = eaI+B = eaI eB = ea eB . (81)

It now remains to compute eB . Note that

    
2
0 b 0 b −b 0
B2 =   =  (82)
2
−b 0 −b 0 0 −b

157
 and     
2 3
0 b −b 0 0 −b
B3 = BB2 =   = . (83)
2 3
−b 0 0 −b b 0
Combining all of the above results:
   
2 3
1−b +··· b − b /3! + · · · cos b sin b
eB =  = . (84)
3 2
−b + b /3! + · · · 1−b +··· − sin b cos b

The final result is  

cos b sin b
eA = ea  . (85)
− sin b cos b

The reader may recall that the three examples discussed above represent the three “Jordan
canonical forms” for 2 × 2 matrices. Example 1 corresponds to a diagonalizable matrix (which
also includes the case a = b), Example 2 corresponds to a nondiagonalizable matrix with equal
eigenvalues a and Example 3 corresponds to the complex conjugate pair of eigenvalues a ± bi.
We shall return to these cases as they are encountered in linear systems of ODEs.

Linear systems of ODEs and the exponential matrix

We now come to the main result of this section, which is to verify the conjecture in Eq. (68),
i.e. that the solution to the initial value problem

x′ = Ax, x(0) = a, (86)

where A is a constant n × n matrix, is given by

x(t) = etA a. (87)

First expand the exponential series for x(t) above:

t2 2 tn n
 
x(t) = I + tA + A + · · · + A + · · · a. (88)
2! n!

Because this series converges for all x ∈ Rn , we can differentiate termwise:

tn−1
 
′ 2 n
x (t) = A + tA + · · · + A + · · · a. (89)
(n − 1)!

158
Now factor out an A from the above series,
tn−1
 
′ n−1
x (t) = A I + tA + · · · + A +··· a (90)
(n − 1)!
= AetA a

= Ax(t).

Therefore x(t) in (87) satisfies the initial value problem in (86).

Some technical comments regarding the convergence of the exponential matrix

series

(These comments were not presented in this lecture. They are presented here for the interested
reader.)

We now return to the infinite series definition of the matrix exponential,

1 2 1
eA = I + A + A + · · · + An + · · · . (91)
2! n!
The goal is to show that this series converges. To do so, we need to have an idea of the “bigness”
the matrix A and its powers. As you may recall from an earlier course in linear algebra, it is
useful to define the “norm” of the matrix A in terms of how it acts on vectors x in Rn . We
define the norm of A as follows,

k A k = max k Ax k . (92)
kxk=1

Note that we are looking for the maximum value of k Ax k for x in the “unit ball” in Rn . It
follows that
k Ax k ≤ k A k, for any x such that k x k= 1. (93)

This is not restrictive, for it follows that

k Ay k ≤ k A kk y k, for all y ∈ Rn . (94)

To see this, let y be any vector in Rn . Now define

1
x= y, (95)
kyk

159
which is clearly a unit vector. Now insert this vector into Eq. (93) to give

1
k Ay k ≤ k A k, (96)
kyk

from which (94) follows. Note that

k I k= 1, (97)

since the identity matrix does not change a vector.

Let us now return to the series in (91). We consider the partial sums of this series:

1 2 1
Sn = I + A + A + · · · + An , n = 0, 1, 2, · · · . (98)
2! n!

The matrices An are all n × n matricies, implying that Sn is an n × n matrix. Let us now
examine the norm of the matrices Sn . From the property of norms,

1 2 1
k Sn k = k I + A + A + · · · + An
2! n!
1 1
≤ k I k + k A k + k A2 k + · · · + k An k (99)
2! n!

We must establish one additional property of matrix norms, namely that

k A n k ≤ k A kn . (100)

To see this, let’s go back to Eq. (94) and let y = Ax, for any x such that k x k= 1. Then

k AAx k ≤ k A kk Ax k

≤ k A k2 k x k, k x k= 1. (101)

Therefore, from Eq. (99),

1 1
k Sn k ≤ k I k + k A k + k A k2 + · · · + k A kn
2! n!
1 2 1 n
= 1 + a + a + · · · a , where a =k A k . (102)
2! n!

Note that
kSn k ≤ ea for all n ≥ 0 . (103)

160
Note that for any n ≥ 0, the quantity kSn k is a real number – it is the norm of the matrix Sn .
From (103),
lim kSn k ≤ ea = ekAk , (104)
n→∞

provided that the limit exists.

We’re almost there – a tiny amount of extra work is required. Let us return to the partial
sum in (98),
1 2 1
Sn = I + A + A + · · · + An , n = 0, 1, 2, · · · . (105)
2! n!
In order to conclude that the sequence of partial sums {Sn } is convergent, we must show that it
is a Cauchy series, i.e., that for any ǫ > 0, there exists an N > 0, such that for any n, m > N,

k Sm − Sn k < ǫ. (106)

This can be done in the same way as is done for convergent series of real numbers. Without
loss of generality, let us assume that m > n. Then
1 m 1 1
k Sm − Sn k = k A + Am+1 + · · · + An k
m! (m + 1)! n!
1 m 1 1
≤ k A k+k Am+1 k + · · · + k An k
m! (m + 1)! n!
m m+1 n
a a a
≤ + +···+ , (107)
m! (m + 1)! n!
where a =k A k. We know that the series
∞
X an
(108)
n=0
n!

converges – it converges to the value ea . This implies that the sequence of partial sums,
1 2 1
Sn = 1 + a + a · · · + an , (109)
2! n!
converges to the limit ea . An elementary theorem from analysis states that if a sequence is
convergent, i.e., if it converges to a limit, then it is a Cauchy sequence. From Eq. (107), this
implies that for any ǫ > 0, there exists an N > 0 such that

kSm − Sm k < ǫ for all m, n > N . (110)

161
This establishes that the sequence of partial sums Sn in (98) is a convergent sequence and that
the limit of this sequence S is an n × n matrix. We call this limit the exponential of matrix A
and denote it as follows,
lim Sn = S = eA . (111)
n→∞

162
Lecture 17

Linear systems and the Fundamental Matrix - continued

We now apply the results obtained earlier to examine the matrix exponential solutions of three
classes of linear systems in R2 :
 
a 0
Example 1: x′ = Ax where A= .
0 b
In the previous lecture, we found eA
to be the diagonal matrix with entries ea and eb . To
compute etA , we may simply replace a and b with at and bt so that
 
at
e 0
etA =  . (112)
0 ebt

This means that the solution to the initial value problem x(0) = a is
    
at at
e 0 a a e
etA a =   1  =  1 . (113)
bt bt
0 e a2 a2 e

Of course, the two DEs in x1 (t) and x2 (t) could easily have been solved separately. Since A is
diagonal, the two DEs are uncoupled. Nevertheless, it is still worthwhile to see the exponential
matrix applied to this problem.
 
a 1
Example 2: x′ = Ax where A= .
0 a
As in the previous lecture, we shall write A as

A = aI + N, (114)

where  
0 1
N= . (115)
0 0

163
Recall that the matrix N is nilpotent since N2 = N3 = · · · = 0. Then

etA = eatI+tN (116)

= eatI eN since IN = NI

= eat I [I + tN]
 
1 t
= eat  
0 1
 
at at
e te
=  .
at
0 e
The solution to the initial value problem x(0) = a is
    
at at at
e te a (a + a2 t)e
etA a =   1  =  1 . (117)
at at
0 e a2 a2 e
Here we see the appearance of solutions with polynomials multiplied by exponentials. This
is, of course, reminiscent of the case for second-order linear DEs with constant coefficients for
which the eigenvalues of the characteristic polynomial are equal. The matrix A above is the
Jordan canonical form for such a situation.
 
a b
Example 3: x′ = Ax where A= .
−b a
As in the previous lecture, we shall write A as

A = aI + B, (118)

where  
0 b
B= , (119)
−b 0
Since I and B commute,
eA = eatI+B = eatI etB = eat etB . (120)

The matrix eB was computed in the previous lecture. We obtain etB by replacing b with bt.
The final result is  
cos bt sin bt
etA = eat  . (121)
− sin bt cos bt

164
The solution to the initial value problem x(0) = a is
    
at at at
e cos bt e sin bt a e (a cos bt + a sin bt)
etA a =   1  =  1 2
. (122)
at at at
−e sin bt e cos bt a2 e (−a1 sin bt + a2 cos bt)
The eigenvalues of A are complex, i.e., λ = a±bi. As we know for linear second order DEs with
constant coefficients, the solutions involve exponentials multiplied by trigonometric functions.

The above three examples actually take care of all situations that would be encountered in
R2 . Why is this? Because they are the three fundamental Jordan canonical (or normal)
forms of 2 × 2 matrices into which all 2 × 2 matrices can be converted by means of a similarity
transformation.
This leads to the important question, “How do we compute matrix exponentials etA in
general?” The answer will almost always be, “Certainly not by using the series definition!”
Fortunately, we can rely upon similarity transformations to simplify the problem. Recall, from
linear algebra, that for a general n × n matrix A, we can construct a matrix C such that the
matrix B defined as
B = C−1 AC (123)

is in Jordan form. In the simplest possible case, where A has n real and distinct eigenvalues
λ1 , λ2 , · · · , λn , the matrix B will be a diagonal matrix with entries λi , i.e.,
 
λ 0 ··· 0
 1 
 0 λ2 0 0 
 
B = diag[λ1 , λ2 , · · · , λn ] =  . (124)
 ··· ··· ··· 0 
 
 
0 0 0 λn
In this case, the matrix C that accomplishes this job is the matrix formed by concatenating
the n linearly independent λk -eigenvectors vk , k = 1, 2, · · · , n, to form an n × n matrix:

C = [v2 v2 · · · vn ] (125)

In more complicated cases, B will be in block diagonal form. For example, if λ1 = λ2 = λ,

and we can find only one λ-eigenvector of A, then we can construct a matrix C so that B will

165
have the following 2 × 2 block in the upper left corner:
 
λ 1
 . (126)
0 λ
And if two eigenvalues λk and λk+1 are complex conjugates of each other, e.g., λk = a + bi,
with an appropriate selection of basis vectors (we omit a detailed discussion here), there will
be a corresponding 2 × 2 block containing the kth and k + 1st diagonal elements of B; that
block will have the form  
a b
 . (127)
−b a
The exponentiation of B is then relatively straightforward: In the simplest cases, etB will
generally have blocks that correspond to the 2 × 2 blocks produced in the first three examples.
(I use the word “generally” because there can be additional complications, e.g., triply degen-
erate eigenvalues, doubly degenerate complex eigenvalues, etc.. We avoid discussion of these
complications here.)

Question: So once you have produced the matrix etB , how do you get back to etA ?

Answer: We use the seemingly remarkable result

etA = etCBC = CetB C−1 .

−1
(128)

Let us derive this result. First of all,

B = C−1 AC implies A = CBC−1 . (129)

By definition,

etA = etCBC
−1
(130)
1 2
= I + tCBC−1 + t (CBC−1)(CBC−1 ) + · · ·
2!
1
= CC−1 + tCBC−1 + t2 CB2 C−1 + · · ·
 2! 
1 2 2
= C I + tB + t B + · · · C−1
2!
tB −1
= Ce C

166
This result implies that we can take the general matrix A, transform it to matrix B, exponen-
tiate tB, then inverse transform back to yield etA .

 
1 1
Example 4: x′ = Ax where A= .
4 1

We computed the fundamental matrix Φ(t, 0) of this system in a previous lecture. As we

now know, this fundamental matrix is nothing other than etA . Previously, we found that the
eigenvalues of A are λ1 = 3 and λ2 = −1 with associated eigenvectors
   
1 1
v1 =   , v2 =  . (131)
2 −2
These two vectors may be used to construct the matrix C:
 
1 1
C = [v1 v2 ] =  . (132)
2 −2
A little algebra gives  
1 1
2 4
C−1 =  . (133)
1
2
− 41
The matrix B then must be  
3 0
B = C−1 AC =  . (134)
0 −1
so that  
3t
e 0
etB =  . (135)
0 e−t
Then

etA = CetB C−1 (136)

   
1 1
1 1 e3t 0
=    2 4 
1 1
2 −2 0 e−t 2
−4
 
1 3t
(e + e−t ) 41 (e3t − e−t )
=  2 ,
1 3t
e3t − e−t 2
(e + e−t
)

167
which agrees with the earlier result.

We now have outlined the basic procedure to determine all solutions to linear homogeneous
systems having the form,
x′ = Ax , (137)

by computing the associated fundamental matrix, or flow operator,

Φ(t) = etA . (138)

The particular solution satisfying the initial condition,

x(0) = a , (139)

is
x(t) = Φ(t)a = etA a . (140)

For example, the solution to this initial value problem for Example 4 above is,

x(t) = etA a
  
1 3t −t 1 3t −t
(e + e ) (e − e ) a
=  2 4  1 
1 3t
e3t − e−t 2
(e + e−t ) a2
 
( 21 a1 + 14 a2 )e3t + ( 12 a1 − 41 a2 )e−t
=  . (141)
1 3t 1 −t
(a1 + 2 a2 )e + (−a1 + 2 a2 )e

168
Linear homogeneous DEs as “dynamical systems”

(Relevant section from AMATH 351 Course Notes by J. Wainwright: Section 2.2.5,
p. 85)

Dynamical systems theory is the study of evolution of systems with respect to a parameter,
usually time. Here, we view the solutions of the linear homogeneous DE,

x′ = Ax , (142)

in from a dynamical systems viewpoint, which is essentially examining solutions, or the entire
family of solutions, from a geometrical point of view. Recall that these solutions may be
expressed in the form,
x(t) = etA a (143)

where
x(0) = a . (144)

We’ll be viewing these solutions as curves in the state space Rn . These curves are called
orbits of the DE. As the time variable t varies, the point x(t) traces out a curve in Rn . And
the motion of x(t) is determined by the flow operator

Φ(t) = etA . (145)

We don’t simply have to think about the curve x(t) for positive values of t, i.e., forward
motion – we can also consider its motion for t decreasing from zero. The net result is the
orbit of the flow etA through a is defined as

γ(a) = {x ∈ Rn | x = etA for some t ∈ R} . (146)

The equilibrium point x = 0

There is one special point for this system, namely, the point x = 0 since

x(t) = 0 (147)

169
is a solution of the DE. Since
A0 = 0 , (148)

it is easy to see, from the series expansion definition of the matrix exponential, that

Φ(t)0 = etA 0 = 0 . (149)

The point x = 0 is called an equilibrium point of the DE. It is also called a fixed point of
the flow operator Φ(t). Note that the orbit of 0 is simply the point 0 itself,

γ(0) = {0} . (150)

170