Chapter 4

Systems of Linear Equations

Denote by Mn (R) the set of n ⇥ n matrices with real coefficients.

In this chapter, we are interested in studying linear systems of the form

y 0 (t) = A(t)y(t) + r(t), y(t0 ) = y0 , (4.1)

where, for each t, A(t) 2 Mn (R) and r(t) 2 Rn . If r ⌘ 0, then the system is called
homogeneous. If not, the system is called nonhomogeneous.

Example 4.0.1. Consider the third order nonhomogeneous equation y 000 2y 00 + 4y 0 5y =

def def def
sin t + t3 . Then letting y1 = y, y2 = y 0 and y3 = y 00 leads to
0 01 0 1 0 10 1 0 1
y1 y2 0 1 0 y1 0
@y20 A = @ y3 A = @0 0 1A @y2 A + @ 0 A,
y30 2y3 4y2 + 5y1 + sin t + t3 5 4 2 y3 sin t + t 3

which is of the form (4.1) with

0 1 0 1
0 1 0 0
A = @0 0 1A 2 M3 (R) and r(t) = @ 0 A 2 R3 .
5 4 2 sin t + t3

4.1 Theoretical Aspects

Theorem 4.1.1. If each entry of the functions A and r are continuous on some interval
I ⇢ R, then the IVP (4.1) has a unique solution y : I ! Rn .

The proof is left as an exercise. A consequence of Theorem 4.1.1 is that if A and r are
continuous on R, then the unique solution is globally defined on R, that is y : R ! Rn .

27
28 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

Denote by C 1 (I) the set of functions y : I ! Rn which are continuously di↵erentiable

on some given interval ⇢ R. Assume that A(t) 2 Mn (R) is continuous at each t 2 I and
consider the linear homogenous system of ODEs

y 0 = A(t)y. (4.2)

Denote the solution set of (4.2) by

def
S = {y 2 C 1 (I) : y is a solution of (4.2)}. (4.3)

Then we have the following result.

Lemma 4.1.2. S is an n-dimensional vector subspace of C 1 (I).
def
Proof. Let y1 , y2 2 S and ↵1 , ↵2 two scalars (in R or C). Letting z(t) = ↵1 y1 (t) + ↵2 y2 (t),
we get that

z 0 (t) = ↵1 y10 (t) + ↵2 y20 (t) = ↵1 A(t)y1 + ↵2 A(t)y2 = A(t)(↵1 y1 (t) + ↵2 y2 (t)) = A(t)z(t).

Hence, z 2 S. In other words, any linear combination of solutions of (4.2) is a solution.

This is what we call the principle of superposition. Moreover, y ⌘ 0 is obviously a solution
of (4.2). Hence S is vector subspace of C 1 (I).
It is left to demonstrate that S has dimension n. Fix any t0 2 I. For i = 1, . . . , n,
denote by ei the ith basis vector of the canonical basis of Rn . Denote by yi (t) the unique
solution of the IVP
y 0 (t) = A(t)y(t), y(t0 ) = ei 2 Rn
and denote
def
B = {yi (t) : i = 1, . . . , n}.
The rest of the proof is twofold. First, show that B spans the solution set (that is hBi = S)
and second, show that the elements of B are linearly independent.
By the principle of superposition, hBi ⇢ S. We now show that S ⇢ hBi. Consider any
z 2 S. Denote 0 11
z0
Bz2 C X n
B 0C
z(t0 ) = B . C = z0i ei .
@ .. A
i=1
z0n
Then, the function
n
X
def
w(t) = z0i yi (t)
i=1
satisfies
n
X n
X
w(t0 ) = z0i yi (t0 ) = z0i ei = z(t0 ).
i=1 i=1
4.1. THEORETICAL ASPECTS 29

Since z, w 2 S and z(t0 ) = w(t0 ), then by Theorem 1.2.4 (Existence and Uniqueness),
w(t) = z(t) for all t 2 I. Hence
n
X
z(t) = w(t) = z0i yi (t) 2 hBi.
i=1

Hence, S ⇢ hBi, and we conclude that S = hBi, that is B spans the solution set.
Finally, assume that there exists ↵1 , . . . , ↵n such that

↵1 y1 (t) + ↵2 y2 (t) + · · · + ↵n yn (t) = 0 2 Rn , for each t 2 I.

In particular, at t = t0 ,
0
1
↵1
n
X n
X B ↵2 C
B C
0= ↵i yi (t0 ) = ↵i ei = B . C =) ↵i = 0, for all i = 1, . . . , n,
@ .. A
i=1 i=1
↵n

which implies that the vectors yi (t) are linearly independent.

The principle of superposition allows us to think of linear combinations of solutions,

which in turn can be codified as follows. A matrix valued function Y : I ! Mn (R) is a
matrix solution to (4.2) if each column is a solution to (4.2). Observe that if Y is a matrix
solution, then
Y 0 (t) = A(t)Y (t).
Proposition 4.1.3. Let Y (t) be a matrix solution to (4.2). Then, either det Y (t) ⌘ 0 or
det Y (t) 6= 0 for all t 2 I.
Proof. If there exists t0 2 I such that det Y (t0 ) = 0, then Y (t0 ) is not invertible and hence
there exists a nonzero vector v such that Y (t0 )v = 0. Since Y (t) is a matrix solution to
def
(4.2), y1 (t) = Y (t)v is a solution to (4.2) by the principle of superposition. In particular,
it is a solution with initial value y1 (t0 ) = 0. However, y2 ⌘ 0 is a solution for (4.2)
and therefore also a solution satisfying the initial value y2 (t0 ) = 0. By the Theorem of
Existence and Uniqueness, we conclude that Y (t)v = y1 (t) = y2 (t) = 0, for all t 2 I. Hence
det Y (t) = 0 for all t 2 I.

Definition 4.1.4. A fundamental matrix solution to (4.2) is a matrix solution Y (t) of

(4.2) such that det Y (t) 6= 0.
Theorem 4.1.5 (Liouville’s formula). If Y (t) is a matrix solution to (4.2) satisfying
the initial condition Y (t0 ) = Y0 2 Mn (R), then
d
(det Y ) = (tr A(t)) det Y
dt
30 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

and hence Rt
tr A(s) ds
det Y (t) = det Y0 e t0 . (4.4)

The last equation (4.4) is typically called Liouville’s formula.

Proof. Let 2 3
R1 (t)
6 R2 (t) 7
6 7
Y (t) = 6 . 7
4 .. 5
Rn (t)
where Rj (t) denotes the j-th row of Y (t). Since the determinant is a multilinear function
of its rows, 2 3
R1 (t)
6 .. 7
n 6 . 7
d X 6d 7
(det Y (t)) = det 6
6 dt Ri (t)7
7. (4.5)
dt 6 .. 7
i=1
4 . 5
Rn (t)
Denoting Y (t) = {yi,j }ni,j=1 and A(t) = {ai,k }ni,k=1 , and writing out (4.2) in matrix form
(that is Y 0 (t) = A(t)Y (t)) gives the expression

X n
d
yi,j (t) = ai,k (t)yk,j (t)
dt
k=1

from which we can deduce that

X n
dRi
(t) = ai,i (t)Ri (t) + ai,k (t)Rk (t).
dt k=1
k6=i

Thus (4.5) becomes

=0
2 3 z 2}| 3{
R1 (t) R1 (t)
6 .. 7 6 7..
n 6 . 7 n Xn 6 7 .
d X 6 7 X 6 7
(det Y (t)) = 6 7
det 6ai,i (t)Ri (t)7 + det 6ai,k (t)Rk (t)7
6
dt 7 = (tr A(t)) det Y (t)
i=1 6 .. 7 i=1 k=1 6 .. 7
4 . 5 k6=i 4 . 5
Rn (t) Rn (t)

where the last equality follows from the multilinearity of det.

4.2. CONSTANT COEFFICIENT HOMOGENEOUS LINEAR SYSTEMS 31

Definition 4.1.6. Given n vector functions yi : I ! Rn (i = 1, . . . , n), we define their

Wronskian by
0 1
.. .. ..
B . . . C
W (t) = W [y1 , . . . , yn ](t) = det B C
def
y
@ 1 (t) y 2 (t) · · · y n (t)A . (4.6)
.. .. ..
. . .

Our goal is now clear: look for n distinct solutions y1 , . . . , yn of (4.2) and use the
Wronskian in (4.6) to show that they are linearly independent by showing that W (t0 ) 6= 0
at a given point t0 2 I. The general solution is then given by

y(t) = c1 y1 (t) + c2 y2 (t) + · · · + cn yn (t), c1 , . . . , cn 2 R.

If an initial condition y(t0 ) = y0 = (y01 , y02 , . . . , y0n ) is given, find uniquely the constants
c1 , . . . , cn 2 R as follows:

y(t0 ) = c1 y1 (t0 ) + c2 y2 (t0 ) + · · · + cn yn (t0 )

0 1 0c 1 0 11
y0
.. .. .. 1
B . . . C Bc C
B 2C
B 2C
B y0 C
B C
= @y1 (t0 ) y2 (t0 ) · · · yn (t0 )A B . C = y0 = B . C .
.. .. .. @ .. A @ .. A
. . . cn y0n

Since W (t0 ) 6= 0, then the matrix is invertible and the constants c1 , . . . , cn 2 R are uniquely
determined by 0 1 0
c1 1 1 0y1 1
.. .. .. 0
B c2 C B . . . C By2 C
B C B B 0C
B .. C = @y1 (t0 ) y2 (t0 ) · · · yn (t0 )C
A B . C.
@.A .. .. .. @ .. A
cn . . . y0n

4.2 Constant Coefficient Homogeneous Linear Systems

In this section we study properties of the linear system

y 0 = Ay (4.7)

where A 2 Mn (R).
def
We know that if n = 1, then A = 2 R and the solution to (4.7) takes the form
y(t) = e t . For general n 1, it is therefore natural to look for solutions of (4.7) in the
form
y(t) = e t u, (4.8)
32 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

where is a scalar (possibly complex) and u is an n-dimensional vector (possibly complex

valued). For (4.8) to be a solution of (4.7), one must have that

e t=
6 0
e t u = y 0 (t) = Ay(t) = Ae t u =) Au = u,

that is ( , u) is an eigenvalue-eigenvector couple of the matrix A.

Definition 4.2.1. Given a matrix A 2 Mn (C), denotes its kernel by

ker A = {u 2 Cn : Au = 0} .

To find an eigenvalue of a matrix A, we solve the characteristic polynomial defined

by
def
p( ) = det(A I) = 0, (4.9)
where I is the n ⇥ n identity matrix. If solves (4.9), then the matrix A I has zero
determinant and therefore is not invertible. This implies the existence of a vector u such
that (A I)u = 0, or equivalently u 2 ker(A I). Hence, Au = u.

Definition 4.2.2. Given a matrix A 2 Mn (R), denotes its spectrum by

(A) = { 2 C : dim ker(A I) > 0} .

In other words, (A) is the set of eigenvalues of the matrix A.

Example 4.2.3. Consider the system

✓ ◆
0 def 1 3
y = Ay, with A = . (4.10)
3 1

To find the eigenvalues of A, we solve

p( ) = det(A I)
1 3
=
3 1
= (1 )2 9
2
= 2 8
= ( + 2)( 4) = 0,
def def
and hence = 1 = 2 and = 2 = 4, that is

(A) = { 2, 4}.
4.2. CONSTANT COEFFICIENT HOMOGENEOUS LINEAR SYSTEMS 33

Let us find eigenvectors u1 and u2 associated to 1 and 2 , respectively. To find u1 = (a, b)T ,
we solve for
✓ ◆✓ ◆ ✓ ◆
3 3 a 0
(A 1 I)u1 = (A + 2I)u1 = = =) a = b.
3 3 b 0

Choosing b = 1 yields a = 1 and therefore u1 = ( 1, 1)T is an eigenvector associated to

T
1 = 2. To find the eigenvector u2 = (a, b) associated to the eigenvalue 2 , we solve for

✓ ◆✓ ◆ ✓ ◆
3 3 a 0
(A 2 I)u2 = (A 4I)u2 = = =) a = b.
3 3 b 0

Choosing b = 1 yields a = 1 and therefore u2 = (1, 1)T is an eigenvector associated to

2 = 4. Hence, by construction
✓ ◆ ✓ ◆
def 1t 2t 1 def 2t
1
y1 (t) = e u1 = e and y2 (t) = e u2 = e4t
1 1

are two solutions of (4.10). To verify that they are linearly independent, we compute their
Wronskian at t = 0:

1 1
W (0) = det(y1 (0) y2 (0)) = = 2 6= 0.
1 1

By Proposition 4.1.3, y1 (t) and y2 (t) are two linearly independent solutions of (4.10). By
Lemma 4.1.2, the solution set (recall (4.3)) S of (4.10) is two-dimensional. Hence, we
conclude that the general solution of (4.10) is given by
✓ ◆ ✓ ◆
2t 1 4t 1
y(t) = c1 y1 (t) + c2 y2 (t) = c1 e + c2 e , c1 , c2 2 R.
1 1

For instance, to find the unique solution going through the initial condition y(0) = (1, 2)T
must satisfy
✓ ◆ ✓ ◆ ✓ ◆✓ ◆ ✓ ◆
1 1 1 1 c1 1
y(0) = c1 + c2 = = ,
1 1 1 1 c2 2
and therefore
✓ ◆ ✓ ◆ !✓ ◆
1✓ ◆ ✓ ◆✓ ◆ 1 1
c1 1 1 1 1 1 1 1 2 2 1
= = = 1 1
c2 1 1 2 2 1 1 2 2 2
2

Let us do a three dimensional example.

34 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

Example 4.2.4. Consider the system

0 1
3 1 2
y 0 = Ay, with A = @ 1 5 2A .
def
(4.11)
3 3 2

To find the eigenvalues of A, we solve

3 2
p( ) = det(A I) = +6 32 = ( + 2)( 4)2 = 0,

and hence = 2 is a simple eigenvalue and = 4 is a double eigenvalue. We conclude

that (A) = { 2, 4}. Let us find the eigenvector u1 associated to = 2. To find
u1 = (a, b, c)T , we solve for
0 10 1 0 1 0 10 1 0 1
5 1 2 a 0 Gauss 5 1 2 a 0
(A + 2I)u1 = @ 1 7 2A @ b A = @0A =) @ 1 1 0 A @ b A = @0A
elimination

3 3 0 c 0 0 0 0 c 0

and therefore a = b and 0 = 5a + b 2c = 6a 2c, that is c = 3a. Choosing a = 1 yields

b = 1 and c = 3. Therefore u1 = (1, 1, 3)T is an eigenvector associated to = 2. In other
words, u1 spans the subspace ker(A 1 I), that is {u1 } is a basis of ker(A 1 I).
Let us now find a basis of the subspace ker(A 4I), that is we look for (a, b, c)T such
that
0 1 0 10 1 0 1 0 10 1 0 1
a 1 1 2 a 0 Gauss 1 1 2 a 0
@
(A 4I) b = A @ 1 1 2A @ A
b = 0 @ A elimination
=) @ 0 0 0 A @ b = 0A
A @
c 3 3 6 c 0 0 0 0 c 0

and therefore a+b 2c = 0, which yields

0 1 0 1 0 1 0 1
a b 2c 1 2
@ b A = @ b A = b @1A + c @ 0 A .
c c 0 1

Hence,
*011 0 21+
ker(A 4I) = @1A , @ 0 A
0 1
def def
and therefore u2 = (1, 1, 0)T and u3 = ( 2, 0, 1)T are two eigenvectors associated to = 4.
By construction
0 1 0 1 0 1
1 1 2
2t @ A @ A @ 0A
def def 4t def 4t
y1 (t) = e 1 , y2 (t) = e 1 and y3 (t) = e
3 0 1
4.2. CONSTANT COEFFICIENT HOMOGENEOUS LINEAR SYSTEMS 35

are three solutions of (4.10). To verify that they are linearly independent, we compute
their Wronskian at t = 0:
1 1 2
W (0) = det y1 (0) y2 (0) y3 (0) = 1 1 0 = 6 6= 0.
3 0 1
By Proposition 4.1.3, y1 (t), y2 (t) and y3 (t) are three linearly independent solutions of
(4.11). By Lemma 4.1.2, the solution set (recall (4.3)) S of (4.11) is three-dimensional.
We conclude that the general solution of (4.11) is given by
y(t) = c1 y1 (t) + c2 y2 (t) + c3 y3 (t)
0 1 0 1 0 1
1 1 2
= c1 e 2t @1A + c2 e4t @1A + c3 e4t @ 0 A , c1 , c2 , c3 2 R,
3 0 1
where the constants c1 , c2 , c3 2 R can be determined by fixing an initial condition y(t0 ) =
y0 2 R3 . For instance, the unique solution y(t) of (4.11) satisfying y(0) = (1, 0, 0)T must
satisfy
0 1 0 1 0 1 0 10 1 0 1
1 1 2 1 1 2 c1 1
y(0) = c1 @1A + c2 @1A + c3 @ 0 A = @1 1 0 A @c2 A = @0A ,
3 0 1 3 0 1 c3 0
that is 0 1 0 1 1 0 1
10 1
c1 1 1 2 1 6
@c2 A = @1 1 0A @0A = B 1C
@ 6A .
c3 3 0 1 0 1
2
Hence the unique solution y(t) of (4.11) satisfying y(0) = (1, 0, 0)T
is given by
0 1 2t 1 4t 4t
1 0 1 2t + 5 e4t
1
6e 6e + e 6e 6
B C B C
y(t) = @ 16 e 2t 16 e4t A = @ 16 e 2t 16 e4t A .
1 2t 1 4t 1 2t 1 4t
2e 2e 2e 2e

Definition 4.2.5. A n ⇥ n matrix A is called diagonalizable if there exists an invertible

matrix P such that P 1 AP is a diagonal matrix.
Lemma 4.2.6. If A 2 Mn (R) has n real and distinct eigenvalues, then it is diagonalisable.
Proof. Let 1 , . . . , n 2 (A) real and distinct, and denote by u1 , . . . , un the corresponding
n real-valued eigenvectors. Letting
0 1
.. .. ..
B. . . C
P =B C
def
u
@ 1 u 2 · · · u nA ,
.. .. ..
. . .
36 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

we get that 0 1
1 0 0 ··· 0
B0 0 ··· 0C
B 2 C
B .. C
P 1 AP = B
B0 0 3 . 0CC
B .. .. .. .. C
@. . . . 0A
0 0 ··· 0 n

is a diagonal matrix. Hence A is a diagonalisable matrix.

Proposition 4.2.7. Assume (A) = { 1 , . . . , n } with the i real and distinct. Denote
by u1 , . . . , un the corresponding n real-valued eigenvectors. Then A is diagonalisable and
the linear system y 0 = Ay has n linearly independent real-valued solutions yi (t) = e i t ui
(i = 1, . . . , n). The general solution is then given by
n
X n
X
it
y(t) = ci yi (t) = ci e ui , c1 , . . . , cn 2 R.
i=1 i=1

Theorem 4.2.8. A n ⇥ n matrix A is diagonalisable if and only if it has a set {u1 , . . . , un }

of n linearly independent eigenvectors.

Example 4.2.9. The matrix ✓ ◆

1 1
A=
0 1
has the characteristic polynomial p( ) = (1 )2 = 0 and then = 1 is the only eigenvalue.
An associated eigenvector u = (a, b)T must satisfy
✓ ◆✓ ◆ ✓ ◆ ✓ ◆
0 0 a 0 0
(A I)u = = = =) b = 0.
0 1 b b 0
✓ ◆
1
Hence, ker(A I) = h i. In other words, A does not have two linearly independent
0
eigenvectors and therefore it is not diagonalizable.

Theorem 4.2.10. If A 2 Mn (R) is a diagonalisable matrix, then the linear system y 0 = Ay

has a basis of the form
{e i t ui : i = 1, . . . , n},
where ( i , ui ) is an eigenvalue-eigenvector couple associated to A. Note that the eigenvalues
and eigenvectors may be complex valued.

Proof. By Theorem 4.2.8, there exists a set {u1 , . . . , un } of n linearly independent (possibly
complex valued) eigenvectors of A. For i = 1, . . . , n let i be the eigenvalue associated to
4.2. CONSTANT COEFFICIENT HOMOGENEOUS LINEAR SYSTEMS 37

the eigenvector ui . Denote yi (t) = e i t ui a (possibly complex valued) solution of y 0 = Ay.

Evaluating the Wronskian at t = 0 yields
0 1
.. .. ..
B. . . C
W (0) = det B @ 1u u 2 · · · u C
n A 6= 0,
.. .. ..
. . .
since the vectors are linearly independent. Hence, W (t) 6= 0 for all t 2 R and the conclusion
follows.

In case A is diagonalizable, the natural question which arises is what happens when it
has complex eigenvalues? It is known that complex eigenvalues always come in pairs (after
all they are roots of a polynomial equation). Indeed, if ↵ + i is an eigenvalue associated
to the eigenvector u + iv, then ↵ i is an eigenvalue associated to the eigenvector u iv.
Now,
def
y(t) = e(↵+i )t
(u + iv)
= y1 (t) + iy2 (t)
def
= e↵t (cos( t)u sin( t)v) + ie↵t (sin( t)u + cos( t)v)
is a complex valued solution of y 0 = Ay. However, y1 (t) = e↵t (cos( t)u sin( t)v) and
y2 (t) = e↵t (sin( t)u + cos( t)v) are two real solutions of y 0 = Ay, as
y 0 (t) = y10 (t) + iy20 (t) = Ay(t) = A(y1 (t) + iy2 (t)) = Ay1 (t) + iAy2 (t)
and therefore y10 (t) = Ay1 (t) and y20 (t) = Ay2 (t).
Lemma 4.2.11. Let u, v 2 Rn such that u 6= 0 or v 6= 0. If 6= 0, then
y1 (t) = e↵t (cos( t)u sin( t)v) and y2 (t) = e↵t (sin( t)u + cos( t)v)
are two linearly independent functions on R.
Proof. Assume the opposite, that is there exists c 6= 0 such that y1 (t) = cy2 (t) for all t 2 R.
Hence, dividing by e↵t leads to
(cos( t)u sin( t)v) = c (sin( t)u + cos( t)v) , 8 t 2 R.
In particular, at t = 0, we get that u = cv. Di↵erentiating leads to
( sin( t)u cos( t)v) = c ( cos( t)u sin( t)v) , 8 t 2 R,
and in particular at t = 0, we get
6=0
v = c u = c2 v =) (c2 + 1)u = 0 =) u = v = 0,
which is a contradiction.
38 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

Example 4.2.12. Let a, b, c 2 R such that b 6= 0. Let

0 1
a b 0
A = @ b a 0A . (4.12)
0 0 c

Then p( ) = (a )2 + b2 (c ) = 0 implies that 2 {a ± ib, c}. Denote 1 = a + ib,

2 = a ib and 3 = c. Then (A) = {a ± ib, c}. It is straightforward to verify that
def T
u3 = (0, 0, 1) is an eigenvector associated to the eigenvalue 3 = c. Now, let us find an
eigenvector w associated to the eigenvalue a + ib. We look for w = (w1 , w2 , w3 ) 2 C3 such
that 0 10 1 0 1
ib b 0 w1 0
(A (a + ib)I)w = @ b ib 0 A @ w 2 A = 0A .
@
0 0 (c a) ib w3 0
Since b 6= 0, then [(c a) ib]w3 = 0 implies that w3 = 0. Similarly ibw1 bw2 = 0
implies that w2 = iw1 . Choosing w1 = 1 implies that w2 = i. Hence, w = (1, i, 0)T is
an eigenvector associated to the eigenvalue a + ib. Now,
00 1 0 11
1 0
e (a+ib)t at
w = e (cos bt + i sin bt) @ @ A
0 +i @ 1AA
0 0
0 0 1 0 11 0 0 1 0 11
1 0 1 0
= eat @cos bt @0A sin bt @ 1AA + ieat @sin bt @0A + cos bt @ 1AA
0 0 0 0
0 1 0 1
cos bt sin bt
= eat @ sin bt A + ieat @ cos btA
0 0

Let 0 1 0 1
cos bt sin bt
y1 (t) = eat @ sin bt A y2 (t) = eat @ cos btA
def def
and
0 0
the real and the imaginary parts of e(a+ib)t w, respectively. By Lemma 4.2.11, y1 and y2
are linearly independent. We conclude that

y(t) = c1 y1 (t) + c2 y2 (t) + c3 y3 (t)

0 1 0 1 0 1
cos bt sin bt 0
= c1 eat @ sin bt A + c2 eat @ cos btA + c3 ect @0A , c1 , c2 , c3 2 R
0 0 1

is the general (real valued) solution of y 0 = Ay, where A is given in (4.12).

4.2. CONSTANT COEFFICIENT HOMOGENEOUS LINEAR SYSTEMS 39

Example 4.2.13. Consider y 0 = Ay where

0 1
5 0 1 4
B4 0 1 3C
A=B@3
C.
4 1 6A
1 0 1 2

The associated characteristic polynomial is p( ) = ( 2 + 4)( 2)( + 6), and therefore

(A) = {±2i, 2, 6}.

Tedious calculations yield that

0 1
* 8 6i +
B 19 7iC
ker(A 2iI) = B @ 12 14i A
C

10
0 1
* 8 + 6i +
B 19 + 7iC
ker(A + 2iI) = B C
@ 12 + 14i A
10
0 1
* 2 +
B 3C
ker(A 2I) = B @ 2A
C

4
0 1
* 2 +
B1C
ker(A + 6I) = B@2A .
C

0
This gives us four linearly independent complex-valued solutions:
0 1 0 1 0 1 0 1
8 6i 8 + 6i 2 2
B 19 7i C B 19 + 7i C B
2t B 3C
C B1C
e2it B C
@ 12 14i A , e
2it B C
@ 12 + 14i A , e @ 2A and e 6t B C
@ 2 A.
10 10 4 0
Taking the real and the imaginary parts of the first complex-valued solution yields two
linearly independent real-valued solutions
0 1 0 1
8 cos 2t + 6 sin 2t 6 cos 2t + 8 sin 2t
B 19 cos 2t + 7 sin 2tC B 7 cos 2t 19 sin 2t C
y1 (t) = B C B C
@ 12 cos 2t + 14 sin 2t A and y2 (t) = @ 14 cos 2t + 12 sin 2tA
10 cos 2t 10 sin 2t
40 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

while the two other solutions are given by

0 1 0 1
2 2
B 3C B1C
y3 (t) = e2t B C
@ 2A and y4 (t) = e 6t B C
@ 2 A.
4 0

We conclude that the general solution is given by

y(t) = c1 y1 (t) + c2 y2 (t) + c3 y3 (t) + c4 y4 (t), c1 , c2 , c3 , c4 2 R.

4.2.1 Non-diagonalizable matrices

If the matrix A 2 Mn (R) is not diagonalizable, then the approach based on Theorem 4.2.10
does not yield enough linearly independent solutions to construct a basis of the solution
set S of the linear system y 0 = Ay. In this case, the problem does not come from a lack of
eigenvalues (in fact by the Fundamental Theorem of Algebra, there are always n (possibly
complex) roots of the characteristic polynomial equation p( ) = 0). In fact, it comes from
a lack of standard eigenvectors. The following definitions clarify this point.

Definition 4.2.14. Let A 2 Mn (R) and let p( ) = det(A I) its associated characteristic
polynomial. Let 0 2 (A), that is such that p( 0 ) = 0.

(a) The algebraic multiplicity of 0 is defined by the largest integer m 1 such that
( m
0 ) divides the characteristic polynomial p( ).

(b) The geometric multiplicity of 0 is defined by the number dim (ker(A 0 I)).

Definition 4.2.15. Let A 2 Mn (R). If 2 (A) has algebraic multiplicity m, we say that
is defective if
dim (ker(A I)) < m, (4.13)
that is the geometric multiplicity of is strictly less than its algebraic multiplicity. We
define the defect of an eigenvalue by
def
d =m dim (ker(A I)) 0. (4.14)

In other words, the defect of an eigenvalue is the number of missing eigenvector(s)

associated to a given eigenvalue.

Example 4.2.16. Let 0 1

0 1 0 1
B0 2 1 1C
A=B
@1
C
1 0 0A
0 1 0 1
4.2. CONSTANT COEFFICIENT HOMOGENEOUS LINEAR SYSTEMS 41

which has characteristic polynomial p( ) = 3 ( + 1), that is (A) = { 1, 0}. Hence,

= 0 is an eigenvalue with algebraic multiplicity 3. However,
0 1
* 1 +
B 1C
ker(A) = B @1A
C

which implies that dim(ker(A)) = 1, that is = 0 has geometric multiplicity equal to 1.

In this case, the defect of the eigenvalue = 0 is d = 3 1 = 2.
Example 4.2.17. In Example 4.2.4, we considered the matrix A in (4.11), that is
0 1
3 1 2
A=@ 1 5 2A ,
3 3 2

with p( ) = ( + 2)( 4)2 , that is with eigenvalues 2 and 4. Hence, the algebraic
multiplicity of 4 is two. However, we showed that dim(ker(A 4I)) = 2 and so = 4 is
not defective: it has a defect equal to 0.
Definition 4.2.18. If is an eigenvalue of A 2 Mn (R) with algebraic multiplicity m, we
say that a vector u is a generalized eigenvector associated to if

(A I)m u = 0.

Clearly, an eigenvector is a generalized eigenvector, and therefore

ker(A I) ⇢ ker((A I)m ).

The following fundamental result from linear algebra is crucial for our understanding
of defective eigenvalues. Its technical linear algebra proof is omitted.
Theorem 4.2.19. Let 2 (A) with algebraic multiplicity m 1. Then

dim (ker(A I)m ) = m.

Hence, if 2 (A) has algebraic multiplicity m > 1 and is defective of defect d > 0,
then the previous result shows that is always possible to construct d generalized eigen-
vectors associated to . To present this construction, it is convenient to introduce the
following notation.
Kj = ker (A I)j , j = 0, . . . , m.
Note that K0 = ker (A I)0 = ker I = {0} and that K1 = ker (A I) corresponds to
the set of standard eigenvectors. Moreover, for all j = 1, . . . , m,

(A I) : Kj ! Kj 1,
42 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

since for all u 2 Kj , w = (A I)u satisfies (A I)j 1 w = (A I)j u = 0, that

is (A I)u 2 Kj . Note also that Kj ⇢ Kj+1 , as if u 2 Kj , then (A I)j+1 u =
(A I) (A I)j u = 0, which implies that u 2 Kj+1 . Hence, we get

{0} = K0 ⇢ K1 ⇢ · · · ⇢ Kj ⇢ Kj+1 ⇢ · · · ⇢ Km . (4.15)

In fact, we have a little more, as the following result demonstrates.

Lemma 4.2.20. If for a fixed j 2 {0, . . . , m 1}, Kj = Kj+1 , then

Kj = Kj+1 = Kj+2 = · · · = Km .

Proof. Assume that Kj = Kj+1 . We now show that Kj+1 = Kj+2 . From (4.15), we get that
Kj+1 ⇢ Kj+2 . It remains to show that Kj+2 ⇢ Kj+1 . Let u 2 Kj+2 = ker (A I)j+2 .
Let v = (A I)u. Then,

(A I)j+1 v = (A I)j+2 u = 0 =) v 2 ker (A I)j+1 = Kj+1 = Kj .

Therefore, since v 2 Kj

(A I)j+1 u = (A I)j v = 0 =) u 2 Kj+1 .

We conclude that Kj+2 ⇢ Kj+1 and then that Kj+1 = Kj+2 . The proof follows by
induction.

The conclusion of Lemma 4.2.20 is that

{0} = K0 ( K1 ( K2 ( · · · ( Kj 1 ( Kj = Kj+1 = · · · = Km .

The proof of Lemma 4.2.20 indicates the procedure to follow to construct all the gen-
eralized eigenvectors. Suppose that for a given i 1, we have computed a basis for the
subspace Ki consisting of ki < m generalized eigenvectors. Since ki 6= m, we cannot have
that Ki = Ki+1 , that is we have that Ki ( Ki+1 . Hence, there exists a vector u 2 Ki+1 \Ki .
Denote
def
w = (A I)u 2 Ki .

Then (A I)i 1 w = (A I)i u 6= 0, since u 2

/ Ki . This implies that w 2 Ki \ Ki 1. In
other words, to find a new generalized eigenvector u 2 Ki+1 \ Ki , we solve

(A I)u = w

for the vectors w 2 Ki \ Ki 1.

4.2. CONSTANT COEFFICIENT HOMOGENEOUS LINEAR SYSTEMS 43
✓ ◆
1 1
Example 4.2.21. Let A = . We saw in Example 4.2.9 that = 1 2 (A) has
0 1
defect d = 1 with (1, 0)T the only eigenvector. Hence,
⌧✓ ◆
1
{0} = K0 ( K1 = ( K2 .
0

Let us find a generalized eigenvector. Following the above procedure, we look for u 2
K2 \ K1 such that
(A I)u = w

where w 2 K1 \ K0 . The only such vector is w = (1, 0)T . Hence, we look for u = (a, b)T
such that ✓ ◆✓ ◆ ✓ ◆
0 1 a 1
(A I)u = = = w.
0 0 b 0

Hence b = 1 and u = (0, 1)T . We conclude that u = (0, 1)T is the extra missing generalized
eigenvector. Moreover,
⌧✓ ◆ ⌧✓ ◆ ✓ ◆
1 1 0
K0 = {0} ( K1 = ( K2 = , .
0 0 1

Example 4.2.22. In Example 4.2.16 we verified that = 0 is a defective eigenvalue of

0 1
0 1 0 1
B0 2 1 1C
A=B
@1
C
1 0 0A
0 1 0 1

with defect d = 2. Let us find two generalized eigenvectors. We showed also that K1 =
def
ker(A) is spanned by v1 = (1, 1, 1, 1)T 2 K1 \ K0 . Let us find u 2 K2 \ K1 such that
Au = v1 , that is
0 10 1 0 1
0 1 0 1 a 1
B0 2 1 1C B b C B 1C
C B C B
A=B @1 1 = C
0 0 A @cA @ 1 A
0 1 0 1 d 1
This leads to

b+d=1
2b c d= 1
a+b=1
44 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

and therefore 0 1 0 1 0 1 0 1
a 1 b 1 1
BbC B b C B 1 C B0C
B C=B C B C B C
@ c A @ b A = b @ 1A + @0A
d 1 b 1 1
We choose the particular solution
0 1
1
B0C
u = v2 = B C
def
@0A 2 K2 \ K1
1
Let us now find u 2 K3 \ K2 such that Au = v2 , that is
0 10 1 0 1
0 1 0 1 a 1
B0 2 1 1C B b C B0C
C B C B
A=B @1 1 = C
0 0 A @ c A @0A
0 1 0 1 d 1
This leads to
b+d=1
2b c d=0
a+b=0
and therefore 0 1 0 1 0 1 0 1
a d 1 1 1
B b C B1 dC B 1C B 1C
B C=B C = dB C + B C
@ c A @d 2A @1A @ 2A
d d 1 0
We choose the particular solution
0 1
1
B1C
u = v3 = B C
def
@ 2A 2 K3 \ K2
0
Hence
0 1 0 1 0 1 0 1 0 1 0 1
* 1 + * 1 1 + * 1 1 1 +
B 1C B 1C B C B 1C B0C
C B B 1C
K1 = B C ( K2 = B C , B0C ( K3 = ker(A3 ) = B ,@ C ,B C
@1A @ 1 A @0 A @ 1 A 0 @
A 2A
1 1 1 1 1 0
Now that we have developed a general approach to compute generalized eigenvectors,
we must find a way to construct solutions to y 0 = Ay with them. It is not surprising that
this construction involves once again the exponential.
4.2. CONSTANT COEFFICIENT HOMOGENEOUS LINEAR SYSTEMS 45

4.2.2 Exponential of a matrix

Given a matrix A 2 Mn (R), we define (formally) the matrix exponential by
1
X
def 1 k
eA = A . (4.16)
k!
k=0

Given any norm k · k on Rn , the induced matrix norm is defined by

def kAxk
kAk = sup = sup kAxk,
x2Rn \{0} kxk kxk=1

where A 2 Mn (R).
Lemma 4.2.23. For all x 2 Rn and for any A 2 Mn (R),
kAxk  kAkkxk. (4.17)
Moreover, for any A, B 2 Mn (R),
kABk  kAkkBk. (4.18)
Proof. The proof is left as an exercise.
Proposition 4.2.24. Let {Ak }k 0 ⇢ Mn (R) be such that
1
X
kAk k < 1.
k=0

Then, for each x 2 Rn , the vector series converges (in norm) and the resulting function
1
X
x 7! Ak x (4.19)
k=0

is an element of Mn (R).
Proof. Fix x 2 Rn and consider the sequence of partial sums {Sj }j 0 defined by
j
X
def
Sj = kAk xk.
k=0

By (4.17), we get that for each j 0,

j
X j
X j
X 1
X
def
Sj = kAk xk  kAk kkxk = kxk kAk k  M (x) = kxk kAk k < 1,
k=0 k=0 k=0 k=0

which shows that {Sj }j 0 is an increasing sequence of real numbers which is bounded,
hence converging. Hence, the vector series converges in norm. The linearity of the function
defined in (4.19) follows from the properties of series.
46 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

Using (4.18), we obtain that

j
X j
X
1 1
k
kA k  kAkk  ekAk , for all j 0.
k! k!
k=0 k=0

Hence,
1
X j
X
1 1
k
kA k = lim kAk k  ekAk .
k! j!1 k!
k=0 k=0
By Proposition 4.2.24, the function
1
X 1 k
x 7! A x
k!
k=0

defines a matrix which we denote by eA . Moreover, we get that

keA k  ekAk .

The fundamental property of the standard exponential function is that ex+y = ex ey .

This property cannot directly be applied to matrices. However, under some assumption,
it can.
Proposition 4.2.25. If A, B 2 Mn (R) commute, that is AB = BA, then eA+B = eA eB .
Proof. First of all, since AB = BA, we may easily show by induction that the binomial
formula
k
X k!
(A + B)k = Aj B k j
j!(k j)!
j=0

holds for commuting matrices. This being said, the partial sum (of order m) of eA+B is
given by
m
X m
X k m X
k
A+B def 1 1 X k! X 1
Sm = (A + B)k = Aj B k j
= Aj B k j
.
k! k! j!(k j)! j!(k j)!
k=0 k=0 j=0 k=0 j=0

Moreover, letting
m
X m
X
A def 1 k B def 1 k
Sm = A and Sm = B ,
k! k!
k=0 k=0
we get from the Cauchy product formula that
2m
X X
A B 1
Sm Sm = Ak 1 B k 2 .
k1 !k2 !
k=0 k1 +k2 =k
4.2. CONSTANT COEFFICIENT HOMOGENEOUS LINEAR SYSTEMS 47

However,
m
X X m XXk
1 1
Ak 1 B k 2 = Aj B k j A+B
= Sm
k1 !k2 ! j!(k j)!
k=0 k1 +k2 =k k=0 j=0

and therefore
2m
X X
A B A+B 1
Sm Sm Sm = Ak1 B k2 .
k1 !k2 !
k=m+1 k1 +k2 =k

We then get that

2m
X X
A B A+B 1
kSm Sm Sm k kAkk1 kBkk2 ,
k1 !k2 !
k=m+1 k1 +k2 =k

which yields
A B A+B kAk kBk kAk+kBk
kSm Sm Sm k  Sm Sm Sm .

Since ekAk+kBk = ekAk ekBk , we conclude that the right-hand side of the last inequality
tends to 0 as m tends to 1.

We are now ready to define the function t 7! eAt 2 Mn (R) by

1
X 1 k k
eAt = A t . (4.20)
k!
k=0

Proposition 4.2.26. The matrix valued function defined in (4.21) satisfies the following
properties:

(a) eA·0 = I.

(b) eA(s+t) = eAs eAt , for all s, t 2 R.

(c) For each t 2 R, eAt is invertible with inverse (eAt ) 1 =e At .

d At
(d) e = AeAt = eAt A.
dt
Proof. Part (a) follows by plugging t = 0 in (4.21). The proof of (b) follows by observing
that the matrices As and At commute (for any s, t 2 R) and by applying Proposition 4.2.27.
Part (c) follows by plugging s = t in (b) and then using part (a). Finally, to prove part
(d), we can show that the function eAt converges uniformly on any compact set of time,
48 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

and hence, we may interchange di↵erentiation and summation to get:

1
d At d X 1 k k
e = A t
dt dt k!
k=0
1
X d 1 k k
= A t
dt k!
k=0
1
X 1
= Ak t k 1
(k 1)!
k=1
X1
1
=A Ak 1 k 1
t
(k 1)!
k=1
X1
1 k k
=A A t
k!
k=0
At
= Ae .

d At
Similarly, we obtain that dt e = eAt A.
def
Part (d) of Proposition 4.2.28 implies that X(t) = eAt satisfies

d
X(t) = AX(t),
dt

that is X(t) = eAt is the unique fundamental matrix solution of y 0 = Ay such that X(0) = I.
For that reason, the matrix exponential eAt is called the principal fundamental matrix
solution of y 0 = Ay. The following result justifies the name principal.

Proposition 4.2.27. Let A 2 Mn (R) and y0 2 Rn . Then the general solution of y 0 = Ay

is given by
y(t) = eAt c, c 2 Rn (4.21)
while the unique solution of the initial value problem

y 0 (t) = Ay(t), y(0) = y0

is given by
y(t) = eAt y0 . (4.22)

Proof. Since eAt is a fundamental matrix solution, its columns form a basis of the solution
set S and therefore for any solution y of y 0 = Ay there exists c = (c1 , . . . , cn )T 2 Rn such
that y(t) = eAt c. This conclude that (4.22) is the general solution. In particular, the
unique solution y satisfying y(0) = y0 is the one for which y(0) = eA·0 c = Ic = c = y0 .
4.3. BUILDING THE SOLUTION SET WITH THE MATRIX EXPONENTIAL 49

Remark 4.2.28. For any two general fundamental matrix solution X(t) and Y (t), there
exists C 2 Mn (R) invertible such that X(t) = Y (t)C. For instance, if X(t) = eAt , then
eAt = Y (t)C, and in particular I = eA·0 = Y (0)C, which implies that C = Y (0) 1 . A
consequence is that given any fundamental matrix solution Y (t) the exponential matrix
eAt can be recovered from it using the formula
eAt = Y (t)Y (0) 1 . (4.23)
◆ ✓
1 3
Example 4.2.29. Consider A = as in Example 4.2.3, where we verified that
3 1
✓ ◆ ✓ 4t ◆
e 2t e
y1 (t) = 2t and y2 (t) = 4t
e e
are two linearly independent solutions of y 0 = Ay. Hence
✓ ◆
def e 2t e4t
Y (t) =
e 2t e4t
is a fundamental matrix solution. Note that Y (t) is not equal to eAt since Y (0) 6= I. To
recover eAt from Y (t) we use (4.24). Note that
✓ ◆ 1 ✓ ◆ !
1 1
1 1 1 1 1
Y (0) 1 = = = 1
2 2
1 ,
1 1 2 1 1 2 2

and hence
✓ ◆ ! ✓ ◆
1 1
At 1 e 2t e4t 2 2 1 e 2t + e4t e 2t + e4t
e = Y (t)Y (0) = =
e 2t e4t 1
2
1
2
2 e 2t + e4t e 2t + e4t
✓ ◆
1 3
is the matrix exponential function of A = .
3 1
The formulas (4.22) provides a beautiful and powerful construction of the general solu-
tion of a homogeneous linear system with constant coefficients. This being said, in practice,
computing the matrix exponential function eAt may be difficult. However, we can combine
eAt together with the notion of generalized eigenvectors as introduced in Section 4.2.1 to
build the solution set of y 0 = Ay explicitly.

4.3 Building the solution set with the matrix exponential

From (4.22), we obtain that y(t) = eAt c is a solution of y 0 = Ay for any c 2 Rn . What if
c = u is an eigenvector associated to an eigenvalue ? In this case,
! !
1 k⇣ k ⌘ X 1 k k
X1 1
X 1 X1
At 1 k k 1 k k
e u= t A u= t A u = t u= t u = e t u,
k! k! k! k!
k=0 k=0 k=0 k=0
50 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

which is already a solution that we have obtained previously in (4.8).

Now, what if c = u is a generalized eigenvector associated to an eigenvalue of algebraic
multiplicity m? In this case, there exists a j  m such that (A I)k u = 0 for all k > j.
Hence,

y(t) = eAt u
It (A I)t
=e e u
= e t Ie(A I)t
u
1
!
X 1 k
t
=e t (A I)k u
k!
k=0
j
!
X 1 k
t k
=e t (A I) u,
k!
k=0

which is a finite computation.

We are ready to summarize all the work we have done.
Let A 2 Mn (R) and let { 1 , . . . , k } the set of distinct eigenvalues of A. Denote by mi
the algebraic multiplicity of i so that m1 +m2 +· · ·+mk = n. For each i = 1, . . . , k, we use
Theorem 4.2.19 to conclude that dim ker((A mi
i I) ) = mi . Hence, for each i = 1, . . . , k,
(i) (i)
there exists Bi = {v1 , . . . , vmi } a basis for ker((A i I)mi ), that is mi linearly independent
generalized eigenvectors associated to i . We accept without proof that Bi \ Bj = {0} for
i 6= j, that is the set
m
[
def
V = Bi
i=1

consists of a set of n linearly independent generalized eigenvectors. This leads to the

following general construction of the solution set.

Proposition 4.3.1. Denote V = {u1 , . . . , un } a set of n linearly independent generalized

eigenvectors of A 2 Mn (R). For each i = 1, . . . , n, define

yi (t) = eAt ui .

Hence, the general solution of y 0 = Ay is given by

n
X n
X
y(t) = ci yi (t) = ci eAt ui .
i=1 i=1

Proof. By Lemma 4.1.2, the solution set S of y 0 = Ay is an n-dimensional vector subspace

of C 1 (I). Moreover, for each i = 1, . . . , n, yi (t) = eAt ui is a solution. The Wronskian
4.3. BUILDING THE SOLUTION SET WITH THE MATRIX EXPONENTIAL 51

satisfies 0 1
.. .. ..
B. . . C
W (0) = det Bu
@ 1 u 2 ··· un C A 6= 0,
.. .. ..
. . .
then we conclude that {y1 (t), . . . , yn (t)} is a set of linearly independent solutions of y 0 = Ay,
that is it forms a basis for the solution set S. The proof follows.
✓ ◆
2 4
Example 4.3.2. Let A = which has characteristic polynomial
1 2
2
p( ) = (2 )( 2 )+4= = 0,
that is = 0 2 (A) has algebraic multiplicity m = 2. An associated eigenvector u1 =
(a, b)T satifies ✓ ◆✓ ◆ ✓ ◆
2 4 a 0
= =) a = 2b.
1 2 b 0
def
Choosing b = 1 leads to u1 = ( 2, 1)T , that is
⌧✓ ◆
2
K1 = ker A =
1
which implies that = 0 is defective with defect d = 1. To find a generalized eigenvector
u2 = (a, b)T 2 K2 \ K1 , we solve for v such that Au2 = u1 , that is
✓ ◆✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆
2 4 a 2 a 2b 1 2 1
= =) a = 2b 1 =) u2 = = =b + .
1 2 b 1 b b 1 0
def
We choose u2 = ( 1, 0)T so that
⌧✓ ◆ ✓ ◆
2 2 1
K2 = ker A = , .
1 0
Hence, ⇢✓ ◆ ✓ ◆
2 1
V = {u1 , u2 } = ,
1 0
forms a set of two linearly independent generalized eigenvectors of A. Let
def
y1 (t) = eAt u1 = e t u1 = u1
1
X 1
X
def 1 k k 1 k k
y2 (t) = eAt u2 = e t e(A I)t
u2 = t A u2 = t A u2 = u2 + tAu2 = u2 + tu1 .
k! k!
k=0 k=0

Hence the general solution of y0 = Ay is given by

✓ ◆ ✓ ◆
2 1 2t
y(t) = c1 y1 (t) + c2 y2 (t) = c1 u1 + c2 (u2 + tu1 ) = c1 + c2 , c1 , c2 2 R.
1 t
52 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS
✓
◆
1 1
Example 4.3.3. Let A = with p( ) = ( 1)2 . We verified in Example 4.2.9 that
0 1
= 1 2 (A) is defective with defect d = 1 and that u1 = (1, 0)T is an eigenvector. In
Example 4.2.21, we showed that u2 = (0, 1)T is a generalized eigenvector. Hence, let

def
y1 (t) = eAt u1 = e t u1 = et u1
1
X
def At t (A I)t t 1 k
y2 (t) = e u2 = e e u2 = e t (A I)k u2 = et (u2 + tAu2 ) = et (u2 + tu1 ).
k!
k=0

Hence the general solution of y 0 = Ay is given by

✓ t◆ ✓ t◆
t t e te
y(t) = c1 y1 (t) + c2 y2 (t) = c1 e u1 + c2 e (u2 + tu1 ) = c1 + c2 , c1 , c2 2 R.
0 et

Example 4.3.4. Let

0 1
0 1 0 1
B0 2 1 1C
A=B
@1
C.
1 0 0A
0 1 0 1

As seen in Example 4.2.16, (A) = { 1, 0} such that 1 = 0 is an eigenvalue with algebraic

multiplicity 3 and defect d = 2. 2 = 1 has algebraic multiplicity 1. In Example 4.2.22,
we obtained

0 1 0 1 0 1 0 1 0 1 0 1
* 1 + * 1 1 + * 1 1 1 +
B 1C B 1C B C B 1C B0C
C B B 1C
K1 = B C ( K2 = B C , B0C ( K3 = ker(A3 ) = B , C,B C .
@1A @ 1 A @0A @ 1 A @0A @ 2A
1 1 1 1 1 0

def def
Denote u1 = (1, 1, 1, 1)T the (only) eigenvector associated to 1 = 0, u2 = (1, 0, 0, 1)T and
def
u3 = ( 1, 1, 2, 0)T the generalized eigenvectors associated to 1 = 0 such that Au2 = u1
4.4. NON-HOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 53

and Au3 = u2 . Define

1 0
1
B 1C
y1 (t) = eAt u1 = e 1 t u1 = e0·t u1 = u1 = B C
def
@1A
1
0 1
1+t
1
X
1 k k B t C
t A u2 = e 1 t (u2 + tAu2 ) = u2 + tu1 = B C
def
y2 (t) = eAt u2 = e 1 t @ t A
k!
k=0
1+t
0 21
1 + t + t2
2
X t2 t2 B 2 C
def 1 k k B 1 t2 C
y3 (t) = eAt u3 = e 1t
t A u3 = u3 + tAu3 + A2 u3 = u3 + tu2 + u1 = B 2 C
k!
k=0
2 2 @ 2 + t2 A
2
t + t2
def
The eigenvector associated to 2 = 1 is given by u4 = (1, 2, 1, 1)T so that
def
y4 (t) = eAt u4 = e t u4 .

The Wronskian of y1 , y2 , y3 , y4 at t = 0 satisfies

0 1 0 1
.. .. .. 1 1 1 1
B . . . C B 1 0 1 2C
W (0) = det B @u1 u2 u3 u4 A = det @
C B C = 1 6= 0,
.. .. .. 1 0 2 1A
. . . 1 1 0 1
and we conclude that {y1 (t), y2 (t), y3 (t), y4 (t)} is a set of linearly independent solutions of
y 0 = Ay, that is it forms a basis for the solution set S. Hence, the general solution of
y 0 = Ay is given by
4
X
y(t) = ci yi (t), ci 2 R
i=1
0 1 0 1 0 21 0 1
1 1+t 1 + t + t2 1
B 1C B t C B 2 C B C
B C B C B 1 t2 C t B 2C
= c1 @ A + c2 @ + c 3 B C + c4 e @ 1 A, ci 2 R.
t A
2
1 @ 2 + t2 A
2
1 1+t t + t2 1

4.4 Non-homogeneous systems with constant coefficients

In this section, we consider non-homogeneous systems with constant coefficients of the form

y 0 (t) = Ay(t) + r(t) (4.24)

54 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

where A 2 Mn (R) and r : R ! Rn is continuous. From

def
Let y1 and y2 be two solutions of (4.25), then z(t) = y2 (t) y1 (t) satisfies

z 0 (t) = y20 (t) y10 (t) = Ay2 (t) + r(t) (Ay1 (t) + r(t)) = A(y2 (t) y1 (t)) = Az(t),

that is z is a solution of y 0 = Ay. Hence, there exists c 2 Rn such that z(t) = Y (t)c for
Y (t) any fundamental matrix solution of y 0 = Ay. In other words

y2 (t) = y1 (t) + Y (t)c, c 2 Rn

solves (4.25), where Y (t)c is the general solution of the homogeneous equation y 0 = Ay
and y1 (t) is a particular solution of the non-homogeneous system (4.25). This leads to the
following result.

Proposition 4.4.1. Let {y1 , . . . , yn } be n linearly independent solution of the homogeneous

equation y 0 = Ay, and let yp (t) be a particular solution of the non-homogeneous equation
(4.25). Then the general solution of (4.25) is given by

y(t) = c1 y1 (t) + c2 y2 (t) + · · · + cn yn (t) + yp (t), ci 2 R. (4.25)

In the previous sections, we developed the theory to compute the solutions y1 , . . . , yn

of the homogeneous equation y 0 = Ay. We now turn to the question of the computation of
the particular solution yp (t) of the non-homogeneous equation (4.25).
To find a particular solution yp (t) of (4.25), consider any fundamental matrix solution
Y (t) of y 0 = Ay, and let
yp (t) = Y (t)cp (t),

where cp (t) is to be determined. Assume that yp solve (4.25) and by the product rule of
di↵erentiation

yp0 (t) = Y 0 (t)cp (t) + Y (t)c0p (t) = AY (t)cp (t) + Y (t)c0p (t)
= Ayp (t) + r(t) = AY (t)cp (t) + r(t)

which implies that

Z
Y (t)c0p (t) = r(t) =) c0p (t) = Y (t) 1
r(t) =) cp (t) = Y (t) 1
r(t) dt.

Since yp (t) = Y (t)cp (t), we conclude that

Z
def 1
yp (t) = Y (t) Y (t) r(t) dt (4.26)
4.4. NON-HOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 55

is a particular solution of the non-homogeneous equation (4.25). We conclude that the

general solution of (4.25) is given by

Z
y(t) = Y (t)c + Y (t) Y (t) 1 r(t) dt
✓ Z ◆
= Y (t) c + Y (t) 1 r(t) dt , c 2 Rn . (4.27)

In particular, letting Y (t) = eAt , the unique solution of the initial value problem

y 0 = Ay, y(0) = y0

is given by
✓ Z t ◆
y(t) = eAt y0 + e As
r(s) ds . (4.28)
0

Equation (4.29) is often called the variation of constants formula.

Example 4.4.2. Solve the IVP

✓ ◆ ✓ t◆ ✓ ◆
0 1 1 e 1
y = y+ , y(0) = .
4 1 1 0

✓ ◆
1 1
The eigenvalues of A = are 1 = 3 and 2 = 1 with associated eigenvectors
4 1
u1 = ( 12 , 1)T and u2 = ( 12 , 1)T . Hence a fundamental matrix solution is given by

!
1 3t 1 t
2e 2e
Y (t) =
e3t e t
56 CHAPTER 4. SYSTEMS OF LINEAR EQUATIONS

From (4.27), a particular solution is given by

Z
yp (t) = Y (t) Y (t) 1 r(t) dt
! ! 1✓ ◆
1 3t 1 t Z 1 3t 1 t
e e e e e t
= 2 3t 2 2 2
dt
e e t e3t e t 1
! !
1 3t
e 1
e t Z e 3t 1 e 3t ✓ t ◆
e
= 2 3t 2
t t
2
1 t dt
e e e 2e
1
! !
1 3t
e 1
e t Z e 4t + 1 e 3t
= 2 3t 2 2
dt
e e t 1 + 12 et
! !
1 3t 1 t 1 4t 1 3t
2e 2e 4e 6e
=
e3t e t t + 12 et
!
1 t 1 t t 1
8e 12 + 2 e 4
= 1 1
4e
t
6 te t + 12
!
1 t t 1
8 + 2 e 3
= 1 t+ 1
.
4 t e 3

Hence the general solution of the non-homogeneous system is

y(t) = Y (t)c + yp (t)

! ! !
1 3t 1 t 1 t t 1
2e 2e 8 + 2 e 3
= c1 + c2 + 1 1 , c1 , c2 2 R.
e3t e t 4 t e t + 3

The unique solution satisfying y(0) = (1, 0)T satisfies

! !
1 1 ✓ ◆ 11 ✓ ◆
c1 1
y(0) = Y (0)c + yp (0) = 2 2
+ 1
24
=
1 1 c2 12
0

and therefore
✓ ◆ ! 1 "✓ ◆ !# ! ! !
1 1 11 1 35 17
c1 2 2 1 24 1 2 24 12
= 1 = 1 1 = 3 .
c2 1 1 0 12 1 2 12 2

We conclude that the unique solution of the IVP is

! ! !
17 12 e3t 3 1
2e
t 1 t
8 + 2 e t 1
3
y(t) = + 1 1 .
12 e3t 2 e t 4 t e t + 3